fs/btrfs/inode.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2007 Oracle.  All rights reserved.
   4  */
   5
   6 #include <linux/kernel.h>
   7 #include <linux/bio.h>
   8 #include <linux/buffer_head.h>
   9 #include <linux/file.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/highmem.h>
  13 #include <linux/time.h>
  14 #include <linux/init.h>
  15 #include <linux/string.h>
  16 #include <linux/backing-dev.h>
  17 #include <linux/writeback.h>
  18 #include <linux/compat.h>
  19 #include <linux/xattr.h>
  20 #include <linux/posix_acl.h>
  21 #include <linux/falloc.h>
  22 #include <linux/slab.h>
  23 #include <linux/ratelimit.h>
  24 #include <linux/btrfs.h>
  25 #include <linux/blkdev.h>
  26 #include <linux/posix_acl_xattr.h>
  27 #include <linux/uio.h>
  28 #include <linux/magic.h>
  29 #include <linux/iversion.h>
  30 #include <linux/swap.h>
  31 #include <linux/sched/mm.h>
  32 #include <asm/unaligned.h>
  33 #include "misc.h"
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "ordered-data.h"
  40 #include "xattr.h"
  41 #include "tree-log.h"
  42 #include "volumes.h"
  43 #include "compression.h"
  44 #include "locking.h"
  45 #include "free-space-cache.h"
  46 #include "inode-map.h"
  47 #include "backref.h"
  48 #include "props.h"
  49 #include "qgroup.h"
  50 #include "delalloc-space.h"
  51 #include "block-group.h"
  52
  53 struct btrfs_iget_args {
  54         struct btrfs_key *location;
  55         struct btrfs_root *root;
  56 };
  57
  58 struct btrfs_dio_data {
  59         u64 reserve;
  60         u64 unsubmitted_oe_range_start;
  61         u64 unsubmitted_oe_range_end;
  62         int overwrite;
  63 };
  64
  65 static const struct inode_operations btrfs_dir_inode_operations;
  66 static const struct inode_operations btrfs_symlink_inode_operations;
  67 static const struct inode_operations btrfs_dir_ro_inode_operations;
  68 static const struct inode_operations btrfs_special_inode_operations;
  69 static const struct inode_operations btrfs_file_inode_operations;
  70 static const struct address_space_operations btrfs_aops;
  71 static const struct file_operations btrfs_dir_file_operations;
  72 static const struct extent_io_ops btrfs_extent_io_ops;
  73
  74 static struct kmem_cache *btrfs_inode_cachep;
  75 struct kmem_cache *btrfs_trans_handle_cachep;
  76 struct kmem_cache *btrfs_path_cachep;
  77 struct kmem_cache *btrfs_free_space_cachep;
  78 struct kmem_cache *btrfs_free_space_bitmap_cachep;
  79
  80 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
  81 static int btrfs_truncate(struct inode *inode, bool skip_writeback);
  82 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
  83 static noinline int cow_file_range(struct inode *inode,
  84                                    struct page *locked_page,
  85                                    u64 start, u64 end, int *page_started,
  86                                    unsigned long *nr_written, int unlock);
  87 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
  88                                        u64 orig_start, u64 block_start,
  89                                        u64 block_len, u64 orig_block_len,
  90                                        u64 ram_bytes, int compress_type,
  91                                        int type);
  92
  93 static void __endio_write_update_ordered(struct inode *inode,
  94                                          const u64 offset, const u64 bytes,
  95                                          const bool uptodate);
  96
  97 /*
  98  * Cleanup all submitted ordered extents in specified range to handle errors
  99  * from the btrfs_run_delalloc_range() callback.
 100  *
 101  * NOTE: caller must ensure that when an error happens, it can not call
 102  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 103  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 104  * to be released, which we want to happen only when finishing the ordered
 105  * extent (btrfs_finish_ordered_io()).
 106  */
 107 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 108                                                  struct page *locked_page,
 109                                                  u64 offset, u64 bytes)
 110 {
 111         unsigned long index = offset >> PAGE_SHIFT;
 112         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 113         u64 page_start = page_offset(locked_page);
 114         u64 page_end = page_start + PAGE_SIZE - 1;
 115
 116         struct page *page;
 117
 118         while (index <= end_index) {
 119                 page = find_get_page(inode->i_mapping, index);
 120                 index++;
 121                 if (!page)
 122                         continue;
 123                 ClearPagePrivate2(page);
 124                 put_page(page);
 125         }
 126
 127         /*
 128          * In case this page belongs to the delalloc range being instantiated
 129          * then skip it, since the first page of a range is going to be
 130          * properly cleaned up by the caller of run_delalloc_range
 131          */
 132         if (page_start >= offset && page_end <= (offset + bytes - 1)) {
 133                 offset += PAGE_SIZE;
 134                 bytes -= PAGE_SIZE;
 135         }
 136
 137         return __endio_write_update_ordered(inode, offset, bytes, false);
 138 }
 139
 140 static int btrfs_dirty_inode(struct inode *inode);
 141
 142 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 143 void btrfs_test_inode_set_ops(struct inode *inode)
 144 {
 145         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 146 }
 147 #endif
 148
 149 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 150                                      struct inode *inode,  struct inode *dir,
 151                                      const struct qstr *qstr)
 152 {
 153         int err;
 154
 155         err = btrfs_init_acl(trans, inode, dir);
 156         if (!err)
 157                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 158         return err;
 159 }
 160
 161 /*
 162  * this does all the hard work for inserting an inline extent into
 163  * the btree.  The caller should have done a btrfs_drop_extents so that
 164  * no overlapping inline items exist in the btree
 165  */
 166 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 167                                 struct btrfs_path *path, int extent_inserted,
 168                                 struct btrfs_root *root, struct inode *inode,
 169                                 u64 start, size_t size, size_t compressed_size,
 170                                 int compress_type,
 171                                 struct page **compressed_pages)
 172 {
 173         struct extent_buffer *leaf;
 174         struct page *page = NULL;
 175         char *kaddr;
 176         unsigned long ptr;
 177         struct btrfs_file_extent_item *ei;
 178         int ret;
 179         size_t cur_size = size;
 180         unsigned long offset;
 181
 182         ASSERT((compressed_size > 0 && compressed_pages) ||
 183                (compressed_size == 0 && !compressed_pages));
 184
 185         if (compressed_size && compressed_pages)
 186                 cur_size = compressed_size;
 187
 188         inode_add_bytes(inode, size);
 189
 190         if (!extent_inserted) {
 191                 struct btrfs_key key;
 192                 size_t datasize;
 193
 194                 key.objectid = btrfs_ino(BTRFS_I(inode));
 195                 key.offset = start;
 196                 key.type = BTRFS_EXTENT_DATA_KEY;
 197
 198                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 199                 path->leave_spinning = 1;
 200                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 201                                               datasize);
 202                 if (ret)
 203                         goto fail;
 204         }
 205         leaf = path->nodes[0];
 206         ei = btrfs_item_ptr(leaf, path->slots[0],
 207                             struct btrfs_file_extent_item);
 208         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 209         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 210         btrfs_set_file_extent_encryption(leaf, ei, 0);
 211         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 212         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 213         ptr = btrfs_file_extent_inline_start(ei);
 214
 215         if (compress_type != BTRFS_COMPRESS_NONE) {
 216                 struct page *cpage;
 217                 int i = 0;
 218                 while (compressed_size > 0) {
 219                         cpage = compressed_pages[i];
 220                         cur_size = min_t(unsigned long, compressed_size,
 221                                        PAGE_SIZE);
 222
 223                         kaddr = kmap_atomic(cpage);
 224                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 225                         kunmap_atomic(kaddr);
 226
 227                         i++;
 228                         ptr += cur_size;
 229                         compressed_size -= cur_size;
 230                 }
 231                 btrfs_set_file_extent_compression(leaf, ei,
 232                                                   compress_type);
 233         } else {
 234                 page = find_get_page(inode->i_mapping,
 235                                      start >> PAGE_SHIFT);
 236                 btrfs_set_file_extent_compression(leaf, ei, 0);
 237                 kaddr = kmap_atomic(page);
 238                 offset = offset_in_page(start);
 239                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 240                 kunmap_atomic(kaddr);
 241                 put_page(page);
 242         }
 243         btrfs_mark_buffer_dirty(leaf);
 244         btrfs_release_path(path);
 245
 246         /*
 247          * we're an inline extent, so nobody can
 248          * extend the file past i_size without locking
 249          * a page we already have locked.
 250          *
 251          * We must do any isize and inode updates
 252          * before we unlock the pages.  Otherwise we
 253          * could end up racing with unlink.
 254          */
 255         BTRFS_I(inode)->disk_i_size = inode->i_size;
 256         ret = btrfs_update_inode(trans, root, inode);
 257
 258 fail:
 259         return ret;
 260 }
 261
 262
 263 /*
 264  * conditionally insert an inline extent into the file.  This
 265  * does the checks required to make sure the data is small enough
 266  * to fit as an inline extent.
 267  */
 268 static noinline int cow_file_range_inline(struct inode *inode, u64 start,
 269                                           u64 end, size_t compressed_size,
 270                                           int compress_type,
 271                                           struct page **compressed_pages)
 272 {
 273         struct btrfs_root *root = BTRFS_I(inode)->root;
 274         struct btrfs_fs_info *fs_info = root->fs_info;
 275         struct btrfs_trans_handle *trans;
 276         u64 isize = i_size_read(inode);
 277         u64 actual_end = min(end + 1, isize);
 278         u64 inline_len = actual_end - start;
 279         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 280         u64 data_len = inline_len;
 281         int ret;
 282         struct btrfs_path *path;
 283         int extent_inserted = 0;
 284         u32 extent_item_size;
 285
 286         if (compressed_size)
 287                 data_len = compressed_size;
 288
 289         if (start > 0 ||
 290             actual_end > fs_info->sectorsize ||
 291             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 292             (!compressed_size &&
 293             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 294             end + 1 < isize ||
 295             data_len > fs_info->max_inline) {
 296                 return 1;
 297         }
 298
 299         path = btrfs_alloc_path();
 300         if (!path)
 301                 return -ENOMEM;
 302
 303         trans = btrfs_join_transaction(root);
 304         if (IS_ERR(trans)) {
 305                 btrfs_free_path(path);
 306                 return PTR_ERR(trans);
 307         }
 308         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 309
 310         if (compressed_size && compressed_pages)
 311                 extent_item_size = btrfs_file_extent_calc_inline_size(
 312                    compressed_size);
 313         else
 314                 extent_item_size = btrfs_file_extent_calc_inline_size(
 315                     inline_len);
 316
 317         ret = __btrfs_drop_extents(trans, root, inode, path,
 318                                    start, aligned_end, NULL,
 319                                    1, 1, extent_item_size, &extent_inserted);
 320         if (ret) {
 321                 btrfs_abort_transaction(trans, ret);
 322                 goto out;
 323         }
 324
 325         if (isize > actual_end)
 326                 inline_len = min_t(u64, isize, actual_end);
 327         ret = insert_inline_extent(trans, path, extent_inserted,
 328                                    root, inode, start,
 329                                    inline_len, compressed_size,
 330                                    compress_type, compressed_pages);
 331         if (ret && ret != -ENOSPC) {
 332                 btrfs_abort_transaction(trans, ret);
 333                 goto out;
 334         } else if (ret == -ENOSPC) {
 335                 ret = 1;
 336                 goto out;
 337         }
 338
 339         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 340         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 341 out:
 342         /*
 343          * Don't forget to free the reserved space, as for inlined extent
 344          * it won't count as data extent, free them directly here.
 345          * And at reserve time, it's always aligned to page size, so
 346          * just free one page here.
 347          */
 348         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 349         btrfs_free_path(path);
 350         btrfs_end_transaction(trans);
 351         return ret;
 352 }
 353
 354 struct async_extent {
 355         u64 start;
 356         u64 ram_size;
 357         u64 compressed_size;
 358         struct page **pages;
 359         unsigned long nr_pages;
 360         int compress_type;
 361         struct list_head list;
 362 };
 363
 364 struct async_chunk {
 365         struct inode *inode;
 366         struct page *locked_page;
 367         u64 start;
 368         u64 end;
 369         unsigned int write_flags;
 370         struct list_head extents;
 371         struct cgroup_subsys_state *blkcg_css;
 372         struct btrfs_work work;
 373         atomic_t *pending;
 374 };
 375
 376 struct async_cow {
 377         /* Number of chunks in flight; must be first in the structure */
 378         atomic_t num_chunks;
 379         struct async_chunk chunks[];
 380 };
 381
 382 static noinline int add_async_extent(struct async_chunk *cow,
 383                                      u64 start, u64 ram_size,
 384                                      u64 compressed_size,
 385                                      struct page **pages,
 386                                      unsigned long nr_pages,
 387                                      int compress_type)
 388 {
 389         struct async_extent *async_extent;
 390
 391         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 392         BUG_ON(!async_extent); /* -ENOMEM */
 393         async_extent->start = start;
 394         async_extent->ram_size = ram_size;
 395         async_extent->compressed_size = compressed_size;
 396         async_extent->pages = pages;
 397         async_extent->nr_pages = nr_pages;
 398         async_extent->compress_type = compress_type;
 399         list_add_tail(&async_extent->list, &cow->extents);
 400         return 0;
 401 }
 402
 403 /*
 404  * Check if the inode has flags compatible with compression
 405  */
 406 static inline bool inode_can_compress(struct inode *inode)
 407 {
 408         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW ||
 409             BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
 410                 return false;
 411         return true;
 412 }
 413
 414 /*
 415  * Check if the inode needs to be submitted to compression, based on mount
 416  * options, defragmentation, properties or heuristics.
 417  */
 418 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 419 {
 420         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 421
 422         if (!inode_can_compress(inode)) {
 423                 WARN(IS_ENABLED(CONFIG_BTRFS_DEBUG),
 424                         KERN_ERR "BTRFS: unexpected compression for ino %llu\n",
 425                         btrfs_ino(BTRFS_I(inode)));
 426                 return 0;
 427         }
 428         /* force compress */
 429         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 430                 return 1;
 431         /* defrag ioctl */
 432         if (BTRFS_I(inode)->defrag_compress)
 433                 return 1;
 434         /* bad compression ratios */
 435         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 436                 return 0;
 437         if (btrfs_test_opt(fs_info, COMPRESS) ||
 438             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 439             BTRFS_I(inode)->prop_compress)
 440                 return btrfs_compress_heuristic(inode, start, end);
 441         return 0;
 442 }
 443
 444 static inline void inode_should_defrag(struct btrfs_inode *inode,
 445                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 446 {
 447         /* If this is a small write inside eof, kick off a defrag */
 448         if (num_bytes < small_write &&
 449             (start > 0 || end + 1 < inode->disk_i_size))
 450                 btrfs_add_inode_defrag(NULL, inode);
 451 }
 452
 453 /*
 454  * we create compressed extents in two phases.  The first
 455  * phase compresses a range of pages that have already been
 456  * locked (both pages and state bits are locked).
 457  *
 458  * This is done inside an ordered work queue, and the compression
 459  * is spread across many cpus.  The actual IO submission is step
 460  * two, and the ordered work queue takes care of making sure that
 461  * happens in the same order things were put onto the queue by
 462  * writepages and friends.
 463  *
 464  * If this code finds it can't get good compression, it puts an
 465  * entry onto the work queue to write the uncompressed bytes.  This
 466  * makes sure that both compressed inodes and uncompressed inodes
 467  * are written in the same order that the flusher thread sent them
 468  * down.
 469  */
 470 static noinline int compress_file_range(struct async_chunk *async_chunk)
 471 {
 472         struct inode *inode = async_chunk->inode;
 473         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 474         u64 blocksize = fs_info->sectorsize;
 475         u64 start = async_chunk->start;
 476         u64 end = async_chunk->end;
 477         u64 actual_end;
 478         u64 i_size;
 479         int ret = 0;
 480         struct page **pages = NULL;
 481         unsigned long nr_pages;
 482         unsigned long total_compressed = 0;
 483         unsigned long total_in = 0;
 484         int i;
 485         int will_compress;
 486         int compress_type = fs_info->compress_type;
 487         int compressed_extents = 0;
 488         int redirty = 0;
 489
 490         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 491                         SZ_16K);
 492
 493         /*
 494          * We need to save i_size before now because it could change in between
 495          * us evaluating the size and assigning it.  This is because we lock and
 496          * unlock the page in truncate and fallocate, and then modify the i_size
 497          * later on.
 498          *
 499          * The barriers are to emulate READ_ONCE, remove that once i_size_read
 500          * does that for us.
 501          */
 502         barrier();
 503         i_size = i_size_read(inode);
 504         barrier();
 505         actual_end = min_t(u64, i_size, end + 1);
 506 again:
 507         will_compress = 0;
 508         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 509         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 510         nr_pages = min_t(unsigned long, nr_pages,
 511                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 512
 513         /*
 514          * we don't want to send crud past the end of i_size through
 515          * compression, that's just a waste of CPU time.  So, if the
 516          * end of the file is before the start of our current
 517          * requested range of bytes, we bail out to the uncompressed
 518          * cleanup code that can deal with all of this.
 519          *
 520          * It isn't really the fastest way to fix things, but this is a
 521          * very uncommon corner.
 522          */
 523         if (actual_end <= start)
 524                 goto cleanup_and_bail_uncompressed;
 525
 526         total_compressed = actual_end - start;
 527
 528         /*
 529          * skip compression for a small file range(<=blocksize) that
 530          * isn't an inline extent, since it doesn't save disk space at all.
 531          */
 532         if (total_compressed <= blocksize &&
 533            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 534                 goto cleanup_and_bail_uncompressed;
 535
 536         total_compressed = min_t(unsigned long, total_compressed,
 537                         BTRFS_MAX_UNCOMPRESSED);
 538         total_in = 0;
 539         ret = 0;
 540
 541         /*
 542          * we do compression for mount -o compress and when the
 543          * inode has not been flagged as nocompress.  This flag can
 544          * change at any time if we discover bad compression ratios.
 545          */
 546         if (inode_need_compress(inode, start, end)) {
 547                 WARN_ON(pages);
 548                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 549                 if (!pages) {
 550                         /* just bail out to the uncompressed code */
 551                         nr_pages = 0;
 552                         goto cont;
 553                 }
 554
 555                 if (BTRFS_I(inode)->defrag_compress)
 556                         compress_type = BTRFS_I(inode)->defrag_compress;
 557                 else if (BTRFS_I(inode)->prop_compress)
 558                         compress_type = BTRFS_I(inode)->prop_compress;
 559
 560                 /*
 561                  * we need to call clear_page_dirty_for_io on each
 562                  * page in the range.  Otherwise applications with the file
 563                  * mmap'd can wander in and change the page contents while
 564                  * we are compressing them.
 565                  *
 566                  * If the compression fails for any reason, we set the pages
 567                  * dirty again later on.
 568                  *
 569                  * Note that the remaining part is redirtied, the start pointer
 570                  * has moved, the end is the original one.
 571                  */
 572                 if (!redirty) {
 573                         extent_range_clear_dirty_for_io(inode, start, end);
 574                         redirty = 1;
 575                 }
 576
 577                 /* Compression level is applied here and only here */
 578                 ret = btrfs_compress_pages(
 579                         compress_type | (fs_info->compress_level << 4),
 580                                            inode->i_mapping, start,
 581                                            pages,
 582                                            &nr_pages,
 583                                            &total_in,
 584                                            &total_compressed);
 585
 586                 if (!ret) {
 587                         unsigned long offset = offset_in_page(total_compressed);
 588                         struct page *page = pages[nr_pages - 1];
 589                         char *kaddr;
 590
 591                         /* zero the tail end of the last page, we might be
 592                          * sending it down to disk
 593                          */
 594                         if (offset) {
 595                                 kaddr = kmap_atomic(page);
 596                                 memset(kaddr + offset, 0,
 597                                        PAGE_SIZE - offset);
 598                                 kunmap_atomic(kaddr);
 599                         }
 600                         will_compress = 1;
 601                 }
 602         }
 603 cont:
 604         if (start == 0) {
 605                 /* lets try to make an inline extent */
 606                 if (ret || total_in < actual_end) {
 607                         /* we didn't compress the entire range, try
 608                          * to make an uncompressed inline extent.
 609                          */
 610                         ret = cow_file_range_inline(inode, start, end, 0,
 611                                                     BTRFS_COMPRESS_NONE, NULL);
 612                 } else {
 613                         /* try making a compressed inline extent */
 614                         ret = cow_file_range_inline(inode, start, end,
 615                                                     total_compressed,
 616                                                     compress_type, pages);
 617                 }
 618                 if (ret <= 0) {
 619                         unsigned long clear_flags = EXTENT_DELALLOC |
 620                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 621                                 EXTENT_DO_ACCOUNTING;
 622                         unsigned long page_error_op;
 623
 624                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 625
 626                         /*
 627                          * inline extent creation worked or returned error,
 628                          * we don't need to create any more async work items.
 629                          * Unlock and free up our temp pages.
 630                          *
 631                          * We use DO_ACCOUNTING here because we need the
 632                          * delalloc_release_metadata to be done _after_ we drop
 633                          * our outstanding extent for clearing delalloc for this
 634                          * range.
 635                          */
 636                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 637                                                      clear_flags,
 638                                                      PAGE_UNLOCK |
 639                                                      PAGE_CLEAR_DIRTY |
 640                                                      PAGE_SET_WRITEBACK |
 641                                                      page_error_op |
 642                                                      PAGE_END_WRITEBACK);
 643
 644                         for (i = 0; i < nr_pages; i++) {
 645                                 WARN_ON(pages[i]->mapping);
 646                                 put_page(pages[i]);
 647                         }
 648                         kfree(pages);
 649
 650                         return 0;
 651                 }
 652         }
 653
 654         if (will_compress) {
 655                 /*
 656                  * we aren't doing an inline extent round the compressed size
 657                  * up to a block size boundary so the allocator does sane
 658                  * things
 659                  */
 660                 total_compressed = ALIGN(total_compressed, blocksize);
 661
 662                 /*
 663                  * one last check to make sure the compression is really a
 664                  * win, compare the page count read with the blocks on disk,
 665                  * compression must free at least one sector size
 666                  */
 667                 total_in = ALIGN(total_in, PAGE_SIZE);
 668                 if (total_compressed + blocksize <= total_in) {
 669                         compressed_extents++;
 670
 671                         /*
 672                          * The async work queues will take care of doing actual
 673                          * allocation on disk for these compressed pages, and
 674                          * will submit them to the elevator.
 675                          */
 676                         add_async_extent(async_chunk, start, total_in,
 677                                         total_compressed, pages, nr_pages,
 678                                         compress_type);
 679
 680                         if (start + total_in < end) {
 681                                 start += total_in;
 682                                 pages = NULL;
 683                                 cond_resched();
 684                                 goto again;
 685                         }
 686                         return compressed_extents;
 687                 }
 688         }
 689         if (pages) {
 690                 /*
 691                  * the compression code ran but failed to make things smaller,
 692                  * free any pages it allocated and our page pointer array
 693                  */
 694                 for (i = 0; i < nr_pages; i++) {
 695                         WARN_ON(pages[i]->mapping);
 696                         put_page(pages[i]);
 697                 }
 698                 kfree(pages);
 699                 pages = NULL;
 700                 total_compressed = 0;
 701                 nr_pages = 0;
 702
 703                 /* flag the file so we don't compress in the future */
 704                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 705                     !(BTRFS_I(inode)->prop_compress)) {
 706                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 707                 }
 708         }
 709 cleanup_and_bail_uncompressed:
 710         /*
 711          * No compression, but we still need to write the pages in the file
 712          * we've been given so far.  redirty the locked page if it corresponds
 713          * to our extent and set things up for the async work queue to run
 714          * cow_file_range to do the normal delalloc dance.
 715          */
 716         if (async_chunk->locked_page &&
 717             (page_offset(async_chunk->locked_page) >= start &&
 718              page_offset(async_chunk->locked_page)) <= end) {
 719                 __set_page_dirty_nobuffers(async_chunk->locked_page);
 720                 /* unlocked later on in the async handlers */
 721         }
 722
 723         if (redirty)
 724                 extent_range_redirty_for_io(inode, start, end);
 725         add_async_extent(async_chunk, start, end - start + 1, 0, NULL, 0,
 726                          BTRFS_COMPRESS_NONE);
 727         compressed_extents++;
 728
 729         return compressed_extents;
 730 }
 731
 732 static void free_async_extent_pages(struct async_extent *async_extent)
 733 {
 734         int i;
 735
 736         if (!async_extent->pages)
 737                 return;
 738
 739         for (i = 0; i < async_extent->nr_pages; i++) {
 740                 WARN_ON(async_extent->pages[i]->mapping);
 741                 put_page(async_extent->pages[i]);
 742         }
 743         kfree(async_extent->pages);
 744         async_extent->nr_pages = 0;
 745         async_extent->pages = NULL;
 746 }
 747
 748 /*
 749  * phase two of compressed writeback.  This is the ordered portion
 750  * of the code, which only gets called in the order the work was
 751  * queued.  We walk all the async extents created by compress_file_range
 752  * and send them down to the disk.
 753  */
 754 static noinline void submit_compressed_extents(struct async_chunk *async_chunk)
 755 {
 756         struct inode *inode = async_chunk->inode;
 757         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 758         struct async_extent *async_extent;
 759         u64 alloc_hint = 0;
 760         struct btrfs_key ins;
 761         struct extent_map *em;
 762         struct btrfs_root *root = BTRFS_I(inode)->root;
 763         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
 764         int ret = 0;
 765
 766 again:
 767         while (!list_empty(&async_chunk->extents)) {
 768                 async_extent = list_entry(async_chunk->extents.next,
 769                                           struct async_extent, list);
 770                 list_del(&async_extent->list);
 771
 772 retry:
 773                 lock_extent(io_tree, async_extent->start,
 774                             async_extent->start + async_extent->ram_size - 1);
 775                 /* did the compression code fall back to uncompressed IO? */
 776                 if (!async_extent->pages) {
 777                         int page_started = 0;
 778                         unsigned long nr_written = 0;
 779
 780                         /* allocate blocks */
 781                         ret = cow_file_range(inode, async_chunk->locked_page,
 782                                              async_extent->start,
 783                                              async_extent->start +
 784                                              async_extent->ram_size - 1,
 785                                              &page_started, &nr_written, 0);
 786
 787                         /* JDM XXX */
 788
 789                         /*
 790                          * if page_started, cow_file_range inserted an
 791                          * inline extent and took care of all the unlocking
 792                          * and IO for us.  Otherwise, we need to submit
 793                          * all those pages down to the drive.
 794                          */
 795                         if (!page_started && !ret)
 796                                 extent_write_locked_range(inode,
 797                                                   async_extent->start,
 798                                                   async_extent->start +
 799                                                   async_extent->ram_size - 1,
 800                                                   WB_SYNC_ALL);
 801                         else if (ret && async_chunk->locked_page)
 802                                 unlock_page(async_chunk->locked_page);
 803                         kfree(async_extent);
 804                         cond_resched();
 805                         continue;
 806                 }
 807
 808                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 809                                            async_extent->compressed_size,
 810                                            async_extent->compressed_size,
 811                                            0, alloc_hint, &ins, 1, 1);
 812                 if (ret) {
 813                         free_async_extent_pages(async_extent);
 814
 815                         if (ret == -ENOSPC) {
 816                                 unlock_extent(io_tree, async_extent->start,
 817                                               async_extent->start +
 818                                               async_extent->ram_size - 1);
 819
 820                                 /*
 821                                  * we need to redirty the pages if we decide to
 822                                  * fallback to uncompressed IO, otherwise we
 823                                  * will not submit these pages down to lower
 824                                  * layers.
 825                                  */
 826                                 extent_range_redirty_for_io(inode,
 827                                                 async_extent->start,
 828                                                 async_extent->start +
 829                                                 async_extent->ram_size - 1);
 830
 831                                 goto retry;
 832                         }
 833                         goto out_free;
 834                 }
 835                 /*
 836                  * here we're doing allocation and writeback of the
 837                  * compressed pages
 838                  */
 839                 em = create_io_em(inode, async_extent->start,
 840                                   async_extent->ram_size, /* len */
 841                                   async_extent->start, /* orig_start */
 842                                   ins.objectid, /* block_start */
 843                                   ins.offset, /* block_len */
 844                                   ins.offset, /* orig_block_len */
 845                                   async_extent->ram_size, /* ram_bytes */
 846                                   async_extent->compress_type,
 847                                   BTRFS_ORDERED_COMPRESSED);
 848                 if (IS_ERR(em))
 849                         /* ret value is not necessary due to void function */
 850                         goto out_free_reserve;
 851                 free_extent_map(em);
 852
 853                 ret = btrfs_add_ordered_extent_compress(inode,
 854                                                 async_extent->start,
 855                                                 ins.objectid,
 856                                                 async_extent->ram_size,
 857                                                 ins.offset,
 858                                                 BTRFS_ORDERED_COMPRESSED,
 859                                                 async_extent->compress_type);
 860                 if (ret) {
 861                         btrfs_drop_extent_cache(BTRFS_I(inode),
 862                                                 async_extent->start,
 863                                                 async_extent->start +
 864                                                 async_extent->ram_size - 1, 0);
 865                         goto out_free_reserve;
 866                 }
 867                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 868
 869                 /*
 870                  * clear dirty, set writeback and unlock the pages.
 871                  */
 872                 extent_clear_unlock_delalloc(inode, async_extent->start,
 873                                 async_extent->start +
 874                                 async_extent->ram_size - 1,
 875                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 876                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 877                                 PAGE_SET_WRITEBACK);
 878                 if (btrfs_submit_compressed_write(inode,
 879                                     async_extent->start,
 880                                     async_extent->ram_size,
 881                                     ins.objectid,
 882                                     ins.offset, async_extent->pages,
 883                                     async_extent->nr_pages,
 884                                     async_chunk->write_flags,
 885                                     async_chunk->blkcg_css)) {
 886                         struct page *p = async_extent->pages[0];
 887                         const u64 start = async_extent->start;
 888                         const u64 end = start + async_extent->ram_size - 1;
 889
 890                         p->mapping = inode->i_mapping;
 891                         btrfs_writepage_endio_finish_ordered(p, start, end, 0);
 892
 893                         p->mapping = NULL;
 894                         extent_clear_unlock_delalloc(inode, start, end,
 895                                                      NULL, 0,
 896                                                      PAGE_END_WRITEBACK |
 897                                                      PAGE_SET_ERROR);
 898                         free_async_extent_pages(async_extent);
 899                 }
 900                 alloc_hint = ins.objectid + ins.offset;
 901                 kfree(async_extent);
 902                 cond_resched();
 903         }
 904         return;
 905 out_free_reserve:
 906         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 907         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 908 out_free:
 909         extent_clear_unlock_delalloc(inode, async_extent->start,
 910                                      async_extent->start +
 911                                      async_extent->ram_size - 1,
 912                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 913                                      EXTENT_DELALLOC_NEW |
 914                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 915                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 916                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 917                                      PAGE_SET_ERROR);
 918         free_async_extent_pages(async_extent);
 919         kfree(async_extent);
 920         goto again;
 921 }
 922
 923 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 924                                       u64 num_bytes)
 925 {
 926         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 927         struct extent_map *em;
 928         u64 alloc_hint = 0;
 929
 930         read_lock(&em_tree->lock);
 931         em = search_extent_mapping(em_tree, start, num_bytes);
 932         if (em) {
 933                 /*
 934                  * if block start isn't an actual block number then find the
 935                  * first block in this inode and use that as a hint.  If that
 936                  * block is also bogus then just don't worry about it.
 937                  */
 938                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 939                         free_extent_map(em);
 940                         em = search_extent_mapping(em_tree, 0, 0);
 941                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 942                                 alloc_hint = em->block_start;
 943                         if (em)
 944                                 free_extent_map(em);
 945                 } else {
 946                         alloc_hint = em->block_start;
 947                         free_extent_map(em);
 948                 }
 949         }
 950         read_unlock(&em_tree->lock);
 951
 952         return alloc_hint;
 953 }
 954
 955 /*
 956  * when extent_io.c finds a delayed allocation range in the file,
 957  * the call backs end up in this code.  The basic idea is to
 958  * allocate extents on disk for the range, and create ordered data structs
 959  * in ram to track those extents.
 960  *
 961  * locked_page is the page that writepage had locked already.  We use
 962  * it to make sure we don't do extra locks or unlocks.
 963  *
 964  * *page_started is set to one if we unlock locked_page and do everything
 965  * required to start IO on it.  It may be clean and already done with
 966  * IO when we return.
 967  */
 968 static noinline int cow_file_range(struct inode *inode,
 969                                    struct page *locked_page,
 970                                    u64 start, u64 end, int *page_started,
 971                                    unsigned long *nr_written, int unlock)
 972 {
 973         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 974         struct btrfs_root *root = BTRFS_I(inode)->root;
 975         u64 alloc_hint = 0;
 976         u64 num_bytes;
 977         unsigned long ram_size;
 978         u64 cur_alloc_size = 0;
 979         u64 blocksize = fs_info->sectorsize;
 980         struct btrfs_key ins;
 981         struct extent_map *em;
 982         unsigned clear_bits;
 983         unsigned long page_ops;
 984         bool extent_reserved = false;
 985         int ret = 0;
 986
 987         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 988                 WARN_ON_ONCE(1);
 989                 ret = -EINVAL;
 990                 goto out_unlock;
 991         }
 992
 993         num_bytes = ALIGN(end - start + 1, blocksize);
 994         num_bytes = max(blocksize,  num_bytes);
 995         ASSERT(num_bytes <= btrfs_super_total_bytes(fs_info->super_copy));
 996
 997         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 998
 999         if (start == 0) {
1000                 /* lets try to make an inline extent */
1001                 ret = cow_file_range_inline(inode, start, end, 0,
1002                                             BTRFS_COMPRESS_NONE, NULL);
1003                 if (ret == 0) {
1004                         /*
1005                          * We use DO_ACCOUNTING here because we need the
1006                          * delalloc_release_metadata to be run _after_ we drop
1007                          * our outstanding extent for clearing delalloc for this
1008                          * range.
1009                          */
1010                         extent_clear_unlock_delalloc(inode, start, end, NULL,
1011                                      EXTENT_LOCKED | EXTENT_DELALLOC |
1012                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1013                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1014                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1015                                      PAGE_END_WRITEBACK);
1016                         *nr_written = *nr_written +
1017                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1018                         *page_started = 1;
1019                         goto out;
1020                 } else if (ret < 0) {
1021                         goto out_unlock;
1022                 }
1023         }
1024
1025         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1026         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1027                         start + num_bytes - 1, 0);
1028
1029         while (num_bytes > 0) {
1030                 cur_alloc_size = num_bytes;
1031                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1032                                            fs_info->sectorsize, 0, alloc_hint,
1033                                            &ins, 1, 1);
1034                 if (ret < 0)
1035                         goto out_unlock;
1036                 cur_alloc_size = ins.offset;
1037                 extent_reserved = true;
1038
1039                 ram_size = ins.offset;
1040                 em = create_io_em(inode, start, ins.offset, /* len */
1041                                   start, /* orig_start */
1042                                   ins.objectid, /* block_start */
1043                                   ins.offset, /* block_len */
1044                                   ins.offset, /* orig_block_len */
1045                                   ram_size, /* ram_bytes */
1046                                   BTRFS_COMPRESS_NONE, /* compress_type */
1047                                   BTRFS_ORDERED_REGULAR /* type */);
1048                 if (IS_ERR(em)) {
1049                         ret = PTR_ERR(em);
1050                         goto out_reserve;
1051                 }
1052                 free_extent_map(em);
1053
1054                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1055                                                ram_size, cur_alloc_size, 0);
1056                 if (ret)
1057                         goto out_drop_extent_cache;
1058
1059                 if (root->root_key.objectid ==
1060                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1061                         ret = btrfs_reloc_clone_csums(inode, start,
1062                                                       cur_alloc_size);
1063                         /*
1064                          * Only drop cache here, and process as normal.
1065                          *
1066                          * We must not allow extent_clear_unlock_delalloc()
1067                          * at out_unlock label to free meta of this ordered
1068                          * extent, as its meta should be freed by
1069                          * btrfs_finish_ordered_io().
1070                          *
1071                          * So we must continue until @start is increased to
1072                          * skip current ordered extent.
1073                          */
1074                         if (ret)
1075                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1076                                                 start + ram_size - 1, 0);
1077                 }
1078
1079                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1080
1081                 /* we're not doing compressed IO, don't unlock the first
1082                  * page (which the caller expects to stay locked), don't
1083                  * clear any dirty bits and don't set any writeback bits
1084                  *
1085                  * Do set the Private2 bit so we know this page was properly
1086                  * setup for writepage
1087                  */
1088                 page_ops = unlock ? PAGE_UNLOCK : 0;
1089                 page_ops |= PAGE_SET_PRIVATE2;
1090
1091                 extent_clear_unlock_delalloc(inode, start,
1092                                              start + ram_size - 1,
1093                                              locked_page,
1094                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1095                                              page_ops);
1096                 if (num_bytes < cur_alloc_size)
1097                         num_bytes = 0;
1098                 else
1099                         num_bytes -= cur_alloc_size;
1100                 alloc_hint = ins.objectid + ins.offset;
1101                 start += cur_alloc_size;
1102                 extent_reserved = false;
1103
1104                 /*
1105                  * btrfs_reloc_clone_csums() error, since start is increased
1106                  * extent_clear_unlock_delalloc() at out_unlock label won't
1107                  * free metadata of current ordered extent, we're OK to exit.
1108                  */
1109                 if (ret)
1110                         goto out_unlock;
1111         }
1112 out:
1113         return ret;
1114
1115 out_drop_extent_cache:
1116         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1117 out_reserve:
1118         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1119         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1120 out_unlock:
1121         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1122                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1123         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1124                 PAGE_END_WRITEBACK;
1125         /*
1126          * If we reserved an extent for our delalloc range (or a subrange) and
1127          * failed to create the respective ordered extent, then it means that
1128          * when we reserved the extent we decremented the extent's size from
1129          * the data space_info's bytes_may_use counter and incremented the
1130          * space_info's bytes_reserved counter by the same amount. We must make
1131          * sure extent_clear_unlock_delalloc() does not try to decrement again
1132          * the data space_info's bytes_may_use counter, therefore we do not pass
1133          * it the flag EXTENT_CLEAR_DATA_RESV.
1134          */
1135         if (extent_reserved) {
1136                 extent_clear_unlock_delalloc(inode, start,
1137                                              start + cur_alloc_size,
1138                                              locked_page,
1139                                              clear_bits,
1140                                              page_ops);
1141                 start += cur_alloc_size;
1142                 if (start >= end)
1143                         goto out;
1144         }
1145         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1146                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1147                                      page_ops);
1148         goto out;
1149 }
1150
1151 /*
1152  * work queue call back to started compression on a file and pages
1153  */
1154 static noinline void async_cow_start(struct btrfs_work *work)
1155 {
1156         struct async_chunk *async_chunk;
1157         int compressed_extents;
1158
1159         async_chunk = container_of(work, struct async_chunk, work);
1160
1161         compressed_extents = compress_file_range(async_chunk);
1162         if (compressed_extents == 0) {
1163                 btrfs_add_delayed_iput(async_chunk->inode);
1164                 async_chunk->inode = NULL;
1165         }
1166 }
1167
1168 /*
1169  * work queue call back to submit previously compressed pages
1170  */
1171 static noinline void async_cow_submit(struct btrfs_work *work)
1172 {
1173         struct async_chunk *async_chunk = container_of(work, struct async_chunk,
1174                                                      work);
1175         struct btrfs_fs_info *fs_info = btrfs_work_owner(work);
1176         unsigned long nr_pages;
1177
1178         nr_pages = (async_chunk->end - async_chunk->start + PAGE_SIZE) >>
1179                 PAGE_SHIFT;
1180
1181         /* atomic_sub_return implies a barrier */
1182         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1183             5 * SZ_1M)
1184                 cond_wake_up_nomb(&fs_info->async_submit_wait);
1185
1186         /*
1187          * ->inode could be NULL if async_chunk_start has failed to compress,
1188          * in which case we don't have anything to submit, yet we need to
1189          * always adjust ->async_delalloc_pages as its paired with the init
1190          * happening in cow_file_range_async
1191          */
1192         if (async_chunk->inode)
1193                 submit_compressed_extents(async_chunk);
1194 }
1195
1196 static noinline void async_cow_free(struct btrfs_work *work)
1197 {
1198         struct async_chunk *async_chunk;
1199
1200         async_chunk = container_of(work, struct async_chunk, work);
1201         if (async_chunk->inode)
1202                 btrfs_add_delayed_iput(async_chunk->inode);
1203         if (async_chunk->blkcg_css)
1204                 css_put(async_chunk->blkcg_css);
1205         /*
1206          * Since the pointer to 'pending' is at the beginning of the array of
1207          * async_chunk's, freeing it ensures the whole array has been freed.
1208          */
1209         if (atomic_dec_and_test(async_chunk->pending))
1210                 kvfree(async_chunk->pending);
1211 }
1212
1213 static int cow_file_range_async(struct inode *inode,
1214                                 struct writeback_control *wbc,
1215                                 struct page *locked_page,
1216                                 u64 start, u64 end, int *page_started,
1217                                 unsigned long *nr_written)
1218 {
1219         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1220         struct cgroup_subsys_state *blkcg_css = wbc_blkcg_css(wbc);
1221         struct async_cow *ctx;
1222         struct async_chunk *async_chunk;
1223         unsigned long nr_pages;
1224         u64 cur_end;
1225         u64 num_chunks = DIV_ROUND_UP(end - start, SZ_512K);
1226         int i;
1227         bool should_compress;
1228         unsigned nofs_flag;
1229         const unsigned int write_flags = wbc_to_write_flags(wbc);
1230
1231         unlock_extent(&BTRFS_I(inode)->io_tree, start, end);
1232
1233         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1234             !btrfs_test_opt(fs_info, FORCE_COMPRESS)) {
1235                 num_chunks = 1;
1236                 should_compress = false;
1237         } else {
1238                 should_compress = true;
1239         }
1240
1241         nofs_flag = memalloc_nofs_save();
1242         ctx = kvmalloc(struct_size(ctx, chunks, num_chunks), GFP_KERNEL);
1243         memalloc_nofs_restore(nofs_flag);
1244
1245         if (!ctx) {
1246                 unsigned clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC |
1247                         EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
1248                         EXTENT_DO_ACCOUNTING;
1249                 unsigned long page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1250                         PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
1251                         PAGE_SET_ERROR;
1252
1253                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1254                                              clear_bits, page_ops);
1255                 return -ENOMEM;
1256         }
1257
1258         async_chunk = ctx->chunks;
1259         atomic_set(&ctx->num_chunks, num_chunks);
1260
1261         for (i = 0; i < num_chunks; i++) {
1262                 if (should_compress)
1263                         cur_end = min(end, start + SZ_512K - 1);
1264                 else
1265                         cur_end = end;
1266
1267                 /*
1268                  * igrab is called higher up in the call chain, take only the
1269                  * lightweight reference for the callback lifetime
1270                  */
1271                 ihold(inode);
1272                 async_chunk[i].pending = &ctx->num_chunks;
1273                 async_chunk[i].inode = inode;
1274                 async_chunk[i].start = start;
1275                 async_chunk[i].end = cur_end;
1276                 async_chunk[i].write_flags = write_flags;
1277                 INIT_LIST_HEAD(&async_chunk[i].extents);
1278
1279                 /*
1280                  * The locked_page comes all the way from writepage and its
1281                  * the original page we were actually given.  As we spread
1282                  * this large delalloc region across multiple async_chunk
1283                  * structs, only the first struct needs a pointer to locked_page
1284                  *
1285                  * This way we don't need racey decisions about who is supposed
1286                  * to unlock it.
1287                  */
1288                 if (locked_page) {
1289                         /*
1290                          * Depending on the compressibility, the pages might or
1291                          * might not go through async.  We want all of them to
1292                          * be accounted against wbc once.  Let's do it here
1293                          * before the paths diverge.  wbc accounting is used
1294                          * only for foreign writeback detection and doesn't
1295                          * need full accuracy.  Just account the whole thing
1296                          * against the first page.
1297                          */
1298                         wbc_account_cgroup_owner(wbc, locked_page,
1299                                                  cur_end - start);
1300                         async_chunk[i].locked_page = locked_page;
1301                         locked_page = NULL;
1302                 } else {
1303                         async_chunk[i].locked_page = NULL;
1304                 }
1305
1306                 if (blkcg_css != blkcg_root_css) {
1307                         css_get(blkcg_css);
1308                         async_chunk[i].blkcg_css = blkcg_css;
1309                 } else {
1310                         async_chunk[i].blkcg_css = NULL;
1311                 }
1312
1313                 btrfs_init_work(&async_chunk[i].work, async_cow_start,
1314                                 async_cow_submit, async_cow_free);
1315
1316                 nr_pages = DIV_ROUND_UP(cur_end - start, PAGE_SIZE);
1317                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1318
1319                 btrfs_queue_work(fs_info->delalloc_workers, &async_chunk[i].work);
1320
1321                 *nr_written += nr_pages;
1322                 start = cur_end + 1;
1323         }
1324         *page_started = 1;
1325         return 0;
1326 }
1327
1328 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1329                                         u64 bytenr, u64 num_bytes)
1330 {
1331         int ret;
1332         struct btrfs_ordered_sum *sums;
1333         LIST_HEAD(list);
1334
1335         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1336                                        bytenr + num_bytes - 1, &list, 0);
1337         if (ret == 0 && list_empty(&list))
1338                 return 0;
1339
1340         while (!list_empty(&list)) {
1341                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1342                 list_del(&sums->list);
1343                 kfree(sums);
1344         }
1345         if (ret < 0)
1346                 return ret;
1347         return 1;
1348 }
1349
1350 /*
1351  * when nowcow writeback call back.  This checks for snapshots or COW copies
1352  * of the extents that exist in the file, and COWs the file as required.
1353  *
1354  * If no cow copies or snapshots exist, we write directly to the existing
1355  * blocks on disk
1356  */
1357 static noinline int run_delalloc_nocow(struct inode *inode,
1358                                        struct page *locked_page,
1359                                        const u64 start, const u64 end,
1360                                        int *page_started, int force,
1361                                        unsigned long *nr_written)
1362 {
1363         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1364         struct btrfs_root *root = BTRFS_I(inode)->root;
1365         struct btrfs_path *path;
1366         u64 cow_start = (u64)-1;
1367         u64 cur_offset = start;
1368         int ret;
1369         bool check_prev = true;
1370         const bool freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
1371         u64 ino = btrfs_ino(BTRFS_I(inode));
1372         bool nocow = false;
1373         u64 disk_bytenr = 0;
1374
1375         path = btrfs_alloc_path();
1376         if (!path) {
1377                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1378                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1379                                              EXTENT_DO_ACCOUNTING |
1380                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1381                                              PAGE_CLEAR_DIRTY |
1382                                              PAGE_SET_WRITEBACK |
1383                                              PAGE_END_WRITEBACK);
1384                 return -ENOMEM;
1385         }
1386
1387         while (1) {
1388                 struct btrfs_key found_key;
1389                 struct btrfs_file_extent_item *fi;
1390                 struct extent_buffer *leaf;
1391                 u64 extent_end;
1392                 u64 extent_offset;
1393                 u64 num_bytes = 0;
1394                 u64 disk_num_bytes;
1395                 u64 ram_bytes;
1396                 int extent_type;
1397
1398                 nocow = false;
1399
1400                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1401                                                cur_offset, 0);
1402                 if (ret < 0)
1403                         goto error;
1404
1405                 /*
1406                  * If there is no extent for our range when doing the initial
1407                  * search, then go back to the previous slot as it will be the
1408                  * one containing the search offset
1409                  */
1410                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1411                         leaf = path->nodes[0];
1412                         btrfs_item_key_to_cpu(leaf, &found_key,
1413                                               path->slots[0] - 1);
1414                         if (found_key.objectid == ino &&
1415                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1416                                 path->slots[0]--;
1417                 }
1418                 check_prev = false;
1419 next_slot:
1420                 /* Go to next leaf if we have exhausted the current one */
1421                 leaf = path->nodes[0];
1422                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1423                         ret = btrfs_next_leaf(root, path);
1424                         if (ret < 0) {
1425                                 if (cow_start != (u64)-1)
1426                                         cur_offset = cow_start;
1427                                 goto error;
1428                         }
1429                         if (ret > 0)
1430                                 break;
1431                         leaf = path->nodes[0];
1432                 }
1433
1434                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1435
1436                 /* Didn't find anything for our INO */
1437                 if (found_key.objectid > ino)
1438                         break;
1439                 /*
1440                  * Keep searching until we find an EXTENT_ITEM or there are no
1441                  * more extents for this inode
1442                  */
1443                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1444                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1445                         path->slots[0]++;
1446                         goto next_slot;
1447                 }
1448
1449                 /* Found key is not EXTENT_DATA_KEY or starts after req range */
1450                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1451                     found_key.offset > end)
1452                         break;
1453
1454                 /*
1455                  * If the found extent starts after requested offset, then
1456                  * adjust extent_end to be right before this extent begins
1457                  */
1458                 if (found_key.offset > cur_offset) {
1459                         extent_end = found_key.offset;
1460                         extent_type = 0;
1461                         goto out_check;
1462                 }
1463
1464                 /*
1465                  * Found extent which begins before our range and potentially
1466                  * intersect it
1467                  */
1468                 fi = btrfs_item_ptr(leaf, path->slots[0],
1469                                     struct btrfs_file_extent_item);
1470                 extent_type = btrfs_file_extent_type(leaf, fi);
1471
1472                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1473                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1474                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1475                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1476                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1477                         extent_end = found_key.offset +
1478                                 btrfs_file_extent_num_bytes(leaf, fi);
1479                         disk_num_bytes =
1480                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1481                         /*
1482                          * If extent we got ends before our range starts, skip
1483                          * to next extent
1484                          */
1485                         if (extent_end <= start) {
1486                                 path->slots[0]++;
1487                                 goto next_slot;
1488                         }
1489                         /* Skip holes */
1490                         if (disk_bytenr == 0)
1491                                 goto out_check;
1492                         /* Skip compressed/encrypted/encoded extents */
1493                         if (btrfs_file_extent_compression(leaf, fi) ||
1494                             btrfs_file_extent_encryption(leaf, fi) ||
1495                             btrfs_file_extent_other_encoding(leaf, fi))
1496                                 goto out_check;
1497                         /*
1498                          * If extent is created before the last volume's snapshot
1499                          * this implies the extent is shared, hence we can't do
1500                          * nocow. This is the same check as in
1501                          * btrfs_cross_ref_exist but without calling
1502                          * btrfs_search_slot.
1503                          */
1504                         if (!freespace_inode &&
1505                             btrfs_file_extent_generation(leaf, fi) <=
1506                             btrfs_root_last_snapshot(&root->root_item))
1507                                 goto out_check;
1508                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1509                                 goto out_check;
1510                         /* If extent is RO, we must COW it */
1511                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1512                                 goto out_check;
1513                         ret = btrfs_cross_ref_exist(root, ino,
1514                                                     found_key.offset -
1515                                                     extent_offset, disk_bytenr);
1516                         if (ret) {
1517                                 /*
1518                                  * ret could be -EIO if the above fails to read
1519                                  * metadata.
1520                                  */
1521                                 if (ret < 0) {
1522                                         if (cow_start != (u64)-1)
1523                                                 cur_offset = cow_start;
1524                                         goto error;
1525                                 }
1526
1527                                 WARN_ON_ONCE(freespace_inode);
1528                                 goto out_check;
1529                         }
1530                         disk_bytenr += extent_offset;
1531                         disk_bytenr += cur_offset - found_key.offset;
1532                         num_bytes = min(end + 1, extent_end) - cur_offset;
1533                         /*
1534                          * If there are pending snapshots for this root, we
1535                          * fall into common COW way
1536                          */
1537                         if (!freespace_inode && atomic_read(&root->snapshot_force_cow))
1538                                 goto out_check;
1539                         /*
1540                          * force cow if csum exists in the range.
1541                          * this ensure that csum for a given extent are
1542                          * either valid or do not exist.
1543                          */
1544                         ret = csum_exist_in_range(fs_info, disk_bytenr,
1545                                                   num_bytes);
1546                         if (ret) {
1547                                 /*
1548                                  * ret could be -EIO if the above fails to read
1549                                  * metadata.
1550                                  */
1551                                 if (ret < 0) {
1552                                         if (cow_start != (u64)-1)
1553                                                 cur_offset = cow_start;
1554                                         goto error;
1555                                 }
1556                                 WARN_ON_ONCE(freespace_inode);
1557                                 goto out_check;
1558                         }
1559                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr))
1560                                 goto out_check;
1561                         nocow = true;
1562                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1563                         extent_end = found_key.offset + ram_bytes;
1564                         extent_end = ALIGN(extent_end, fs_info->sectorsize);
1565                         /* Skip extents outside of our requested range */
1566                         if (extent_end <= start) {
1567                                 path->slots[0]++;
1568                                 goto next_slot;
1569                         }
1570                 } else {
1571                         /* If this triggers then we have a memory corruption */
1572                         BUG();
1573                 }
1574 out_check:
1575                 /*
1576                  * If nocow is false then record the beginning of the range
1577                  * that needs to be COWed
1578                  */
1579                 if (!nocow) {
1580                         if (cow_start == (u64)-1)
1581                                 cow_start = cur_offset;
1582                         cur_offset = extent_end;
1583                         if (cur_offset > end)
1584                                 break;
1585                         path->slots[0]++;
1586                         goto next_slot;
1587                 }
1588
1589                 btrfs_release_path(path);
1590
1591                 /*
1592                  * COW range from cow_start to found_key.offset - 1. As the key
1593                  * will contain the beginning of the first extent that can be
1594                  * NOCOW, following one which needs to be COW'ed
1595                  */
1596                 if (cow_start != (u64)-1) {
1597                         ret = cow_file_range(inode, locked_page,
1598                                              cow_start, found_key.offset - 1,
1599                                              page_started, nr_written, 1);
1600                         if (ret) {
1601                                 if (nocow)
1602                                         btrfs_dec_nocow_writers(fs_info,
1603                                                                 disk_bytenr);
1604                                 goto error;
1605                         }
1606                         cow_start = (u64)-1;
1607                 }
1608
1609                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1610                         u64 orig_start = found_key.offset - extent_offset;
1611                         struct extent_map *em;
1612
1613                         em = create_io_em(inode, cur_offset, num_bytes,
1614                                           orig_start,
1615                                           disk_bytenr, /* block_start */
1616                                           num_bytes, /* block_len */
1617                                           disk_num_bytes, /* orig_block_len */
1618                                           ram_bytes, BTRFS_COMPRESS_NONE,
1619                                           BTRFS_ORDERED_PREALLOC);
1620                         if (IS_ERR(em)) {
1621                                 if (nocow)
1622                                         btrfs_dec_nocow_writers(fs_info,
1623                                                                 disk_bytenr);
1624                                 ret = PTR_ERR(em);
1625                                 goto error;
1626                         }
1627                         free_extent_map(em);
1628                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1629                                                        disk_bytenr, num_bytes,
1630                                                        num_bytes,
1631                                                        BTRFS_ORDERED_PREALLOC);
1632                         if (ret) {
1633                                 btrfs_drop_extent_cache(BTRFS_I(inode),
1634                                                         cur_offset,
1635                                                         cur_offset + num_bytes - 1,
1636                                                         0);
1637                                 goto error;
1638                         }
1639                 } else {
1640                         ret = btrfs_add_ordered_extent(inode, cur_offset,
1641                                                        disk_bytenr, num_bytes,
1642                                                        num_bytes,
1643                                                        BTRFS_ORDERED_NOCOW);
1644                         if (ret)
1645                                 goto error;
1646                 }
1647
1648                 if (nocow)
1649                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1650                 nocow = false;
1651
1652                 if (root->root_key.objectid ==
1653                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1654                         /*
1655                          * Error handled later, as we must prevent
1656                          * extent_clear_unlock_delalloc() in error handler
1657                          * from freeing metadata of created ordered extent.
1658                          */
1659                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1660                                                       num_bytes);
1661
1662                 extent_clear_unlock_delalloc(inode, cur_offset,
1663                                              cur_offset + num_bytes - 1,
1664                                              locked_page, EXTENT_LOCKED |
1665                                              EXTENT_DELALLOC |
1666                                              EXTENT_CLEAR_DATA_RESV,
1667                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1668
1669                 cur_offset = extent_end;
1670
1671                 /*
1672                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1673                  * handler, as metadata for created ordered extent will only
1674                  * be freed by btrfs_finish_ordered_io().
1675                  */
1676                 if (ret)
1677                         goto error;
1678                 if (cur_offset > end)
1679                         break;
1680         }
1681         btrfs_release_path(path);
1682
1683         if (cur_offset <= end && cow_start == (u64)-1)
1684                 cow_start = cur_offset;
1685
1686         if (cow_start != (u64)-1) {
1687                 cur_offset = end;
1688                 ret = cow_file_range(inode, locked_page, cow_start, end,
1689                                      page_started, nr_written, 1);
1690                 if (ret)
1691                         goto error;
1692         }
1693
1694 error:
1695         if (nocow)
1696                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1697
1698         if (ret && cur_offset < end)
1699                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1700                                              locked_page, EXTENT_LOCKED |
1701                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1702                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1703                                              PAGE_CLEAR_DIRTY |
1704                                              PAGE_SET_WRITEBACK |
1705                                              PAGE_END_WRITEBACK);
1706         btrfs_free_path(path);
1707         return ret;
1708 }
1709
1710 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1711 {
1712
1713         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1714             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1715                 return 0;
1716
1717         /*
1718          * @defrag_bytes is a hint value, no spinlock held here,
1719          * if is not zero, it means the file is defragging.
1720          * Force cow if given extent needs to be defragged.
1721          */
1722         if (BTRFS_I(inode)->defrag_bytes &&
1723             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1724                            EXTENT_DEFRAG, 0, NULL))
1725                 return 1;
1726
1727         return 0;
1728 }
1729
1730 /*
1731  * Function to process delayed allocation (create CoW) for ranges which are
1732  * being touched for the first time.
1733  */
1734 int btrfs_run_delalloc_range(struct inode *inode, struct page *locked_page,
1735                 u64 start, u64 end, int *page_started, unsigned long *nr_written,
1736                 struct writeback_control *wbc)
1737 {
1738         int ret;
1739         int force_cow = need_force_cow(inode, start, end);
1740
1741         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1742                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1743                                          page_started, 1, nr_written);
1744         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1745                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1746                                          page_started, 0, nr_written);
1747         } else if (!inode_can_compress(inode) ||
1748                    !inode_need_compress(inode, start, end)) {
1749                 ret = cow_file_range(inode, locked_page, start, end,
1750                                       page_started, nr_written, 1);
1751         } else {
1752                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1753                         &BTRFS_I(inode)->runtime_flags);
1754                 ret = cow_file_range_async(inode, wbc, locked_page, start, end,
1755                                            page_started, nr_written);
1756         }
1757         if (ret)
1758                 btrfs_cleanup_ordered_extents(inode, locked_page, start,
1759                                               end - start + 1);
1760         return ret;
1761 }
1762
1763 void btrfs_split_delalloc_extent(struct inode *inode,
1764                                  struct extent_state *orig, u64 split)
1765 {
1766         u64 size;
1767
1768         /* not delalloc, ignore it */
1769         if (!(orig->state & EXTENT_DELALLOC))
1770                 return;
1771
1772         size = orig->end - orig->start + 1;
1773         if (size > BTRFS_MAX_EXTENT_SIZE) {
1774                 u32 num_extents;
1775                 u64 new_size;
1776
1777                 /*
1778                  * See the explanation in btrfs_merge_delalloc_extent, the same
1779                  * applies here, just in reverse.
1780                  */
1781                 new_size = orig->end - split + 1;
1782                 num_extents = count_max_extents(new_size);
1783                 new_size = split - orig->start;
1784                 num_extents += count_max_extents(new_size);
1785                 if (count_max_extents(size) >= num_extents)
1786                         return;
1787         }
1788
1789         spin_lock(&BTRFS_I(inode)->lock);
1790         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1791         spin_unlock(&BTRFS_I(inode)->lock);
1792 }
1793
1794 /*
1795  * Handle merged delayed allocation extents so we can keep track of new extents
1796  * that are just merged onto old extents, such as when we are doing sequential
1797  * writes, so we can properly account for the metadata space we'll need.
1798  */
1799 void btrfs_merge_delalloc_extent(struct inode *inode, struct extent_state *new,
1800                                  struct extent_state *other)
1801 {
1802         u64 new_size, old_size;
1803         u32 num_extents;
1804
1805         /* not delalloc, ignore it */
1806         if (!(other->state & EXTENT_DELALLOC))
1807                 return;
1808
1809         if (new->start > other->start)
1810                 new_size = new->end - other->start + 1;
1811         else
1812                 new_size = other->end - new->start + 1;
1813
1814         /* we're not bigger than the max, unreserve the space and go */
1815         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1816                 spin_lock(&BTRFS_I(inode)->lock);
1817                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1818                 spin_unlock(&BTRFS_I(inode)->lock);
1819                 return;
1820         }
1821
1822         /*
1823          * We have to add up either side to figure out how many extents were
1824          * accounted for before we merged into one big extent.  If the number of
1825          * extents we accounted for is <= the amount we need for the new range
1826          * then we can return, otherwise drop.  Think of it like this
1827          *
1828          * [ 4k][MAX_SIZE]
1829          *
1830          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1831          * need 2 outstanding extents, on one side we have 1 and the other side
1832          * we have 1 so they are == and we can return.  But in this case
1833          *
1834          * [MAX_SIZE+4k][MAX_SIZE+4k]
1835          *
1836          * Each range on their own accounts for 2 extents, but merged together
1837          * they are only 3 extents worth of accounting, so we need to drop in
1838          * this case.
1839          */
1840         old_size = other->end - other->start + 1;
1841         num_extents = count_max_extents(old_size);
1842         old_size = new->end - new->start + 1;
1843         num_extents += count_max_extents(old_size);
1844         if (count_max_extents(new_size) >= num_extents)
1845                 return;
1846
1847         spin_lock(&BTRFS_I(inode)->lock);
1848         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1849         spin_unlock(&BTRFS_I(inode)->lock);
1850 }
1851
1852 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1853                                       struct inode *inode)
1854 {
1855         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1856
1857         spin_lock(&root->delalloc_lock);
1858         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1859                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1860                               &root->delalloc_inodes);
1861                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1862                         &BTRFS_I(inode)->runtime_flags);
1863                 root->nr_delalloc_inodes++;
1864                 if (root->nr_delalloc_inodes == 1) {
1865                         spin_lock(&fs_info->delalloc_root_lock);
1866                         BUG_ON(!list_empty(&root->delalloc_root));
1867                         list_add_tail(&root->delalloc_root,
1868                                       &fs_info->delalloc_roots);
1869                         spin_unlock(&fs_info->delalloc_root_lock);
1870                 }
1871         }
1872         spin_unlock(&root->delalloc_lock);
1873 }
1874
1875
1876 void __btrfs_del_delalloc_inode(struct btrfs_root *root,
1877                                 struct btrfs_inode *inode)
1878 {
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         if (!list_empty(&inode->delalloc_inodes)) {
1882                 list_del_init(&inode->delalloc_inodes);
1883                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1884                           &inode->runtime_flags);
1885                 root->nr_delalloc_inodes--;
1886                 if (!root->nr_delalloc_inodes) {
1887                         ASSERT(list_empty(&root->delalloc_inodes));
1888                         spin_lock(&fs_info->delalloc_root_lock);
1889                         BUG_ON(list_empty(&root->delalloc_root));
1890                         list_del_init(&root->delalloc_root);
1891                         spin_unlock(&fs_info->delalloc_root_lock);
1892                 }
1893         }
1894 }
1895
1896 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1897                                      struct btrfs_inode *inode)
1898 {
1899         spin_lock(&root->delalloc_lock);
1900         __btrfs_del_delalloc_inode(root, inode);
1901         spin_unlock(&root->delalloc_lock);
1902 }
1903
1904 /*
1905  * Properly track delayed allocation bytes in the inode and to maintain the
1906  * list of inodes that have pending delalloc work to be done.
1907  */
1908 void btrfs_set_delalloc_extent(struct inode *inode, struct extent_state *state,
1909                                unsigned *bits)
1910 {
1911         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1912
1913         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1914                 WARN_ON(1);
1915         /*
1916          * set_bit and clear bit hooks normally require _irqsave/restore
1917          * but in this case, we are only testing for the DELALLOC
1918          * bit, which is only set or cleared with irqs on
1919          */
1920         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1921                 struct btrfs_root *root = BTRFS_I(inode)->root;
1922                 u64 len = state->end + 1 - state->start;
1923                 u32 num_extents = count_max_extents(len);
1924                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1925
1926                 spin_lock(&BTRFS_I(inode)->lock);
1927                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1928                 spin_unlock(&BTRFS_I(inode)->lock);
1929
1930                 /* For sanity tests */
1931                 if (btrfs_is_testing(fs_info))
1932                         return;
1933
1934                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1935                                          fs_info->delalloc_batch);
1936                 spin_lock(&BTRFS_I(inode)->lock);
1937                 BTRFS_I(inode)->delalloc_bytes += len;
1938                 if (*bits & EXTENT_DEFRAG)
1939                         BTRFS_I(inode)->defrag_bytes += len;
1940                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1941                                          &BTRFS_I(inode)->runtime_flags))
1942                         btrfs_add_delalloc_inodes(root, inode);
1943                 spin_unlock(&BTRFS_I(inode)->lock);
1944         }
1945
1946         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1947             (*bits & EXTENT_DELALLOC_NEW)) {
1948                 spin_lock(&BTRFS_I(inode)->lock);
1949                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1950                         state->start;
1951                 spin_unlock(&BTRFS_I(inode)->lock);
1952         }
1953 }
1954
1955 /*
1956  * Once a range is no longer delalloc this function ensures that proper
1957  * accounting happens.
1958  */
1959 void btrfs_clear_delalloc_extent(struct inode *vfs_inode,
1960                                  struct extent_state *state, unsigned *bits)
1961 {
1962         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
1963         struct btrfs_fs_info *fs_info = btrfs_sb(vfs_inode->i_sb);
1964         u64 len = state->end + 1 - state->start;
1965         u32 num_extents = count_max_extents(len);
1966
1967         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1968                 spin_lock(&inode->lock);
1969                 inode->defrag_bytes -= len;
1970                 spin_unlock(&inode->lock);
1971         }
1972
1973         /*
1974          * set_bit and clear bit hooks normally require _irqsave/restore
1975          * but in this case, we are only testing for the DELALLOC
1976          * bit, which is only set or cleared with irqs on
1977          */
1978         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1979                 struct btrfs_root *root = inode->root;
1980                 bool do_list = !btrfs_is_free_space_inode(inode);
1981
1982                 spin_lock(&inode->lock);
1983                 btrfs_mod_outstanding_extents(inode, -num_extents);
1984                 spin_unlock(&inode->lock);
1985
1986                 /*
1987                  * We don't reserve metadata space for space cache inodes so we
1988                  * don't need to call delalloc_release_metadata if there is an
1989                  * error.
1990                  */
1991                 if (*bits & EXTENT_CLEAR_META_RESV &&
1992                     root != fs_info->tree_root)
1993                         btrfs_delalloc_release_metadata(inode, len, false);
1994
1995                 /* For sanity tests. */
1996                 if (btrfs_is_testing(fs_info))
1997                         return;
1998
1999                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
2000                     do_list && !(state->state & EXTENT_NORESERVE) &&
2001                     (*bits & EXTENT_CLEAR_DATA_RESV))
2002                         btrfs_free_reserved_data_space_noquota(
2003                                         &inode->vfs_inode,
2004                                         state->start, len);
2005
2006                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
2007                                          fs_info->delalloc_batch);
2008                 spin_lock(&inode->lock);
2009                 inode->delalloc_bytes -= len;
2010                 if (do_list && inode->delalloc_bytes == 0 &&
2011                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
2012                                         &inode->runtime_flags))
2013                         btrfs_del_delalloc_inode(root, inode);
2014                 spin_unlock(&inode->lock);
2015         }
2016
2017         if ((state->state & EXTENT_DELALLOC_NEW) &&
2018             (*bits & EXTENT_DELALLOC_NEW)) {
2019                 spin_lock(&inode->lock);
2020                 ASSERT(inode->new_delalloc_bytes >= len);
2021                 inode->new_delalloc_bytes -= len;
2022                 spin_unlock(&inode->lock);
2023         }
2024 }
2025
2026 /*
2027  * btrfs_bio_fits_in_stripe - Checks whether the size of the given bio will fit
2028  * in a chunk's stripe. This function ensures that bios do not span a
2029  * stripe/chunk
2030  *
2031  * @page - The page we are about to add to the bio
2032  * @size - size we want to add to the bio
2033  * @bio - bio we want to ensure is smaller than a stripe
2034  * @bio_flags - flags of the bio
2035  *
2036  * return 1 if page cannot be added to the bio
2037  * return 0 if page can be added to the bio
2038  * return error otherwise
2039  */
2040 int btrfs_bio_fits_in_stripe(struct page *page, size_t size, struct bio *bio,
2041                              unsigned long bio_flags)
2042 {
2043         struct inode *inode = page->mapping->host;
2044         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2045         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
2046         u64 length = 0;
2047         u64 map_length;
2048         int ret;
2049         struct btrfs_io_geometry geom;
2050
2051         if (bio_flags & EXTENT_BIO_COMPRESSED)
2052                 return 0;
2053
2054         length = bio->bi_iter.bi_size;
2055         map_length = length;
2056         ret = btrfs_get_io_geometry(fs_info, btrfs_op(bio), logical, map_length,
2057                                     &geom);
2058         if (ret < 0)
2059                 return ret;
2060
2061         if (geom.len < length + size)
2062                 return 1;
2063         return 0;
2064 }
2065
2066 /*
2067  * in order to insert checksums into the metadata in large chunks,
2068  * we wait until bio submission time.   All the pages in the bio are
2069  * checksummed and sums are attached onto the ordered extent record.
2070  *
2071  * At IO completion time the cums attached on the ordered extent record
2072  * are inserted into the btree
2073  */
2074 static blk_status_t btrfs_submit_bio_start(void *private_data, struct bio *bio,
2075                                     u64 bio_offset)
2076 {
2077         struct inode *inode = private_data;
2078         blk_status_t ret = 0;
2079
2080         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2081         BUG_ON(ret); /* -ENOMEM */
2082         return 0;
2083 }
2084
2085 /*
2086  * extent_io.c submission hook. This does the right thing for csum calculation
2087  * on write, or reading the csums from the tree before a read.
2088  *
2089  * Rules about async/sync submit,
2090  * a) read:                             sync submit
2091  *
2092  * b) write without checksum:           sync submit
2093  *
2094  * c) write with checksum:
2095  *    c-1) if bio is issued by fsync:   sync submit
2096  *         (sync_writers != 0)
2097  *
2098  *    c-2) if root is reloc root:       sync submit
2099  *         (only in case of buffered IO)
2100  *
2101  *    c-3) otherwise:                   async submit
2102  */
2103 static blk_status_t btrfs_submit_bio_hook(struct inode *inode, struct bio *bio,
2104                                           int mirror_num,
2105                                           unsigned long bio_flags)
2106
2107 {
2108         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2109         struct btrfs_root *root = BTRFS_I(inode)->root;
2110         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
2111         blk_status_t ret = 0;
2112         int skip_sum;
2113         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
2114
2115         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
2116
2117         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
2118                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
2119
2120         if (bio_op(bio) != REQ_OP_WRITE) {
2121                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
2122                 if (ret)
2123                         goto out;
2124
2125                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
2126                         ret = btrfs_submit_compressed_read(inode, bio,
2127                                                            mirror_num,
2128                                                            bio_flags);
2129                         goto out;
2130                 } else if (!skip_sum) {
2131                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
2132                         if (ret)
2133                                 goto out;
2134                 }
2135                 goto mapit;
2136         } else if (async && !skip_sum) {
2137                 /* csum items have already been cloned */
2138                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2139                         goto mapit;
2140                 /* we're doing a write, do the async checksumming */
2141                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
2142                                           0, inode, btrfs_submit_bio_start);
2143                 goto out;
2144         } else if (!skip_sum) {
2145                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2146                 if (ret)
2147                         goto out;
2148         }
2149
2150 mapit:
2151         ret = btrfs_map_bio(fs_info, bio, mirror_num);
2152
2153 out:
2154         if (ret) {
2155                 bio->bi_status = ret;
2156                 bio_endio(bio);
2157         }
2158         return ret;
2159 }
2160
2161 /*
2162  * given a list of ordered sums record them in the inode.  This happens
2163  * at IO completion time based on sums calculated at bio submission time.
2164  */
2165 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2166                              struct inode *inode, struct list_head *list)
2167 {
2168         struct btrfs_ordered_sum *sum;
2169         int ret;
2170
2171         list_for_each_entry(sum, list, list) {
2172                 trans->adding_csums = true;
2173                 ret = btrfs_csum_file_blocks(trans,
2174                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2175                 trans->adding_csums = false;
2176                 if (ret)
2177                         return ret;
2178         }
2179         return 0;
2180 }
2181
2182 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2183                               unsigned int extra_bits,
2184                               struct extent_state **cached_state)
2185 {
2186         WARN_ON(PAGE_ALIGNED(end));
2187         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2188                                    extra_bits, cached_state);
2189 }
2190
2191 /* see btrfs_writepage_start_hook for details on why this is required */
2192 struct btrfs_writepage_fixup {
2193         struct page *page;
2194         struct btrfs_work work;
2195 };
2196
2197 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2198 {
2199         struct btrfs_writepage_fixup *fixup;
2200         struct btrfs_ordered_extent *ordered;
2201         struct extent_state *cached_state = NULL;
2202         struct extent_changeset *data_reserved = NULL;
2203         struct page *page;
2204         struct inode *inode;
2205         u64 page_start;
2206         u64 page_end;
2207         int ret;
2208
2209         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2210         page = fixup->page;
2211 again:
2212         lock_page(page);
2213         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2214                 ClearPageChecked(page);
2215                 goto out_page;
2216         }
2217
2218         inode = page->mapping->host;
2219         page_start = page_offset(page);
2220         page_end = page_offset(page) + PAGE_SIZE - 1;
2221
2222         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2223                          &cached_state);
2224
2225         /* already ordered? We're done */
2226         if (PagePrivate2(page))
2227                 goto out;
2228
2229         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2230                                         PAGE_SIZE);
2231         if (ordered) {
2232                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2233                                      page_end, &cached_state);
2234                 unlock_page(page);
2235                 btrfs_start_ordered_extent(inode, ordered, 1);
2236                 btrfs_put_ordered_extent(ordered);
2237                 goto again;
2238         }
2239
2240         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2241                                            PAGE_SIZE);
2242         if (ret) {
2243                 mapping_set_error(page->mapping, ret);
2244                 end_extent_writepage(page, ret, page_start, page_end);
2245                 ClearPageChecked(page);
2246                 goto out;
2247          }
2248
2249         ret = btrfs_set_extent_delalloc(inode, page_start, page_end, 0,
2250                                         &cached_state);
2251         if (ret) {
2252                 mapping_set_error(page->mapping, ret);
2253                 end_extent_writepage(page, ret, page_start, page_end);
2254                 ClearPageChecked(page);
2255                 goto out_reserved;
2256         }
2257
2258         ClearPageChecked(page);
2259         set_page_dirty(page);
2260 out_reserved:
2261         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2262         if (ret)
2263                 btrfs_delalloc_release_space(inode, data_reserved, page_start,
2264                                              PAGE_SIZE, true);
2265 out:
2266         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2267                              &cached_state);
2268 out_page:
2269         unlock_page(page);
2270         put_page(page);
2271         kfree(fixup);
2272         extent_changeset_free(data_reserved);
2273 }
2274
2275 /*
2276  * There are a few paths in the higher layers of the kernel that directly
2277  * set the page dirty bit without asking the filesystem if it is a
2278  * good idea.  This causes problems because we want to make sure COW
2279  * properly happens and the data=ordered rules are followed.
2280  *
2281  * In our case any range that doesn't have the ORDERED bit set
2282  * hasn't been properly setup for IO.  We kick off an async process
2283  * to fix it up.  The async helper will wait for ordered extents, set
2284  * the delalloc bit and make it safe to write the page.
2285  */
2286 int btrfs_writepage_cow_fixup(struct page *page, u64 start, u64 end)
2287 {
2288         struct inode *inode = page->mapping->host;
2289         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2290         struct btrfs_writepage_fixup *fixup;
2291
2292         /* this page is properly in the ordered list */
2293         if (TestClearPagePrivate2(page))
2294                 return 0;
2295
2296         if (PageChecked(page))
2297                 return -EAGAIN;
2298
2299         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2300         if (!fixup)
2301                 return -EAGAIN;
2302
2303         SetPageChecked(page);
2304         get_page(page);
2305         btrfs_init_work(&fixup->work, btrfs_writepage_fixup_worker, NULL, NULL);
2306         fixup->page = page;
2307         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2308         return -EBUSY;
2309 }
2310
2311 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2312                                        struct inode *inode, u64 file_pos,
2313                                        u64 disk_bytenr, u64 disk_num_bytes,
2314                                        u64 num_bytes, u64 ram_bytes,
2315                                        u8 compression, u8 encryption,
2316                                        u16 other_encoding, int extent_type)
2317 {
2318         struct btrfs_root *root = BTRFS_I(inode)->root;
2319         struct btrfs_file_extent_item *fi;
2320         struct btrfs_path *path;
2321         struct extent_buffer *leaf;
2322         struct btrfs_key ins;
2323         u64 qg_released;
2324         int extent_inserted = 0;
2325         int ret;
2326
2327         path = btrfs_alloc_path();
2328         if (!path)
2329                 return -ENOMEM;
2330
2331         /*
2332          * we may be replacing one extent in the tree with another.
2333          * The new extent is pinned in the extent map, and we don't want
2334          * to drop it from the cache until it is completely in the btree.
2335          *
2336          * So, tell btrfs_drop_extents to leave this extent in the cache.
2337          * the caller is expected to unpin it and allow it to be merged
2338          * with the others.
2339          */
2340         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2341                                    file_pos + num_bytes, NULL, 0,
2342                                    1, sizeof(*fi), &extent_inserted);
2343         if (ret)
2344                 goto out;
2345
2346         if (!extent_inserted) {
2347                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2348                 ins.offset = file_pos;
2349                 ins.type = BTRFS_EXTENT_DATA_KEY;
2350
2351                 path->leave_spinning = 1;
2352                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2353                                               sizeof(*fi));
2354                 if (ret)
2355                         goto out;
2356         }
2357         leaf = path->nodes[0];
2358         fi = btrfs_item_ptr(leaf, path->slots[0],
2359                             struct btrfs_file_extent_item);
2360         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2361         btrfs_set_file_extent_type(leaf, fi, extent_type);
2362         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2363         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2364         btrfs_set_file_extent_offset(leaf, fi, 0);
2365         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2366         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2367         btrfs_set_file_extent_compression(leaf, fi, compression);
2368         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2369         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2370
2371         btrfs_mark_buffer_dirty(leaf);
2372         btrfs_release_path(path);
2373
2374         inode_add_bytes(inode, num_bytes);
2375
2376         ins.objectid = disk_bytenr;
2377         ins.offset = disk_num_bytes;
2378         ins.type = BTRFS_EXTENT_ITEM_KEY;
2379
2380         /*
2381          * Release the reserved range from inode dirty range map, as it is
2382          * already moved into delayed_ref_head
2383          */
2384         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2385         if (ret < 0)
2386                 goto out;
2387         qg_released = ret;
2388         ret = btrfs_alloc_reserved_file_extent(trans, root,
2389                                                btrfs_ino(BTRFS_I(inode)),
2390                                                file_pos, qg_released, &ins);
2391 out:
2392         btrfs_free_path(path);
2393
2394         return ret;
2395 }
2396
2397 /* snapshot-aware defrag */
2398 struct sa_defrag_extent_backref {
2399         struct rb_node node;
2400         struct old_sa_defrag_extent *old;
2401         u64 root_id;
2402         u64 inum;
2403         u64 file_pos;
2404         u64 extent_offset;
2405         u64 num_bytes;
2406         u64 generation;
2407 };
2408
2409 struct old_sa_defrag_extent {
2410         struct list_head list;
2411         struct new_sa_defrag_extent *new;
2412
2413         u64 extent_offset;
2414         u64 bytenr;
2415         u64 offset;
2416         u64 len;
2417         int count;
2418 };
2419
2420 struct new_sa_defrag_extent {
2421         struct rb_root root;
2422         struct list_head head;
2423         struct btrfs_path *path;
2424         struct inode *inode;
2425         u64 file_pos;
2426         u64 len;
2427         u64 bytenr;
2428         u64 disk_len;
2429         u8 compress_type;
2430 };
2431
2432 static int backref_comp(struct sa_defrag_extent_backref *b1,
2433                         struct sa_defrag_extent_backref *b2)
2434 {
2435         if (b1->root_id < b2->root_id)
2436                 return -1;
2437         else if (b1->root_id > b2->root_id)
2438                 return 1;
2439
2440         if (b1->inum < b2->inum)
2441                 return -1;
2442         else if (b1->inum > b2->inum)
2443                 return 1;
2444
2445         if (b1->file_pos < b2->file_pos)
2446                 return -1;
2447         else if (b1->file_pos > b2->file_pos)
2448                 return 1;
2449
2450         /*
2451          * [------------------------------] ===> (a range of space)
2452          *     |<--->|   |<---->| =============> (fs/file tree A)
2453          * |<---------------------------->| ===> (fs/file tree B)
2454          *
2455          * A range of space can refer to two file extents in one tree while
2456          * refer to only one file extent in another tree.
2457          *
2458          * So we may process a disk offset more than one time(two extents in A)
2459          * and locate at the same extent(one extent in B), then insert two same
2460          * backrefs(both refer to the extent in B).
2461          */
2462         return 0;
2463 }
2464
2465 static void backref_insert(struct rb_root *root,
2466                            struct sa_defrag_extent_backref *backref)
2467 {
2468         struct rb_node **p = &root->rb_node;
2469         struct rb_node *parent = NULL;
2470         struct sa_defrag_extent_backref *entry;
2471         int ret;
2472
2473         while (*p) {
2474                 parent = *p;
2475                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2476
2477                 ret = backref_comp(backref, entry);
2478                 if (ret < 0)
2479                         p = &(*p)->rb_left;
2480                 else
2481                         p = &(*p)->rb_right;
2482         }
2483
2484         rb_link_node(&backref->node, parent, p);
2485         rb_insert_color(&backref->node, root);
2486 }
2487
2488 /*
2489  * Note the backref might has changed, and in this case we just return 0.
2490  */
2491 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2492                                        void *ctx)
2493 {
2494         struct btrfs_file_extent_item *extent;
2495         struct old_sa_defrag_extent *old = ctx;
2496         struct new_sa_defrag_extent *new = old->new;
2497         struct btrfs_path *path = new->path;
2498         struct btrfs_key key;
2499         struct btrfs_root *root;
2500         struct sa_defrag_extent_backref *backref;
2501         struct extent_buffer *leaf;
2502         struct inode *inode = new->inode;
2503         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2504         int slot;
2505         int ret;
2506         u64 extent_offset;
2507         u64 num_bytes;
2508
2509         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2510             inum == btrfs_ino(BTRFS_I(inode)))
2511                 return 0;
2512
2513         key.objectid = root_id;
2514         key.type = BTRFS_ROOT_ITEM_KEY;
2515         key.offset = (u64)-1;
2516
2517         root = btrfs_read_fs_root_no_name(fs_info, &key);
2518         if (IS_ERR(root)) {
2519                 if (PTR_ERR(root) == -ENOENT)
2520                         return 0;
2521                 WARN_ON(1);
2522                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2523                          inum, offset, root_id);
2524                 return PTR_ERR(root);
2525         }
2526
2527         key.objectid = inum;
2528         key.type = BTRFS_EXTENT_DATA_KEY;
2529         if (offset > (u64)-1 << 32)
2530                 key.offset = 0;
2531         else
2532                 key.offset = offset;
2533
2534         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2535         if (WARN_ON(ret < 0))
2536                 return ret;
2537         ret = 0;
2538
2539         while (1) {
2540                 cond_resched();
2541
2542                 leaf = path->nodes[0];
2543                 slot = path->slots[0];
2544
2545                 if (slot >= btrfs_header_nritems(leaf)) {
2546                         ret = btrfs_next_leaf(root, path);
2547                         if (ret < 0) {
2548                                 goto out;
2549                         } else if (ret > 0) {
2550                                 ret = 0;
2551                                 goto out;
2552                         }
2553                         continue;
2554                 }
2555
2556                 path->slots[0]++;
2557
2558                 btrfs_item_key_to_cpu(leaf, &key, slot);
2559
2560                 if (key.objectid > inum)
2561                         goto out;
2562
2563                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2564                         continue;
2565
2566                 extent = btrfs_item_ptr(leaf, slot,
2567                                         struct btrfs_file_extent_item);
2568
2569                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2570                         continue;
2571
2572                 /*
2573                  * 'offset' refers to the exact key.offset,
2574                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2575                  * (key.offset - extent_offset).
2576                  */
2577                 if (key.offset != offset)
2578                         continue;
2579
2580                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2581                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2582
2583                 if (extent_offset >= old->extent_offset + old->offset +
2584                     old->len || extent_offset + num_bytes <=
2585                     old->extent_offset + old->offset)
2586                         continue;
2587                 break;
2588         }
2589
2590         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2591         if (!backref) {
2592                 ret = -ENOENT;
2593                 goto out;
2594         }
2595
2596         backref->root_id = root_id;
2597         backref->inum = inum;
2598         backref->file_pos = offset;
2599         backref->num_bytes = num_bytes;
2600         backref->extent_offset = extent_offset;
2601         backref->generation = btrfs_file_extent_generation(leaf, extent);
2602         backref->old = old;
2603         backref_insert(&new->root, backref);
2604         old->count++;
2605 out:
2606         btrfs_release_path(path);
2607         WARN_ON(ret);
2608         return ret;
2609 }
2610
2611 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2612                                    struct new_sa_defrag_extent *new)
2613 {
2614         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2615         struct old_sa_defrag_extent *old, *tmp;
2616         int ret;
2617
2618         new->path = path;
2619
2620         list_for_each_entry_safe(old, tmp, &new->head, list) {
2621                 ret = iterate_inodes_from_logical(old->bytenr +
2622                                                   old->extent_offset, fs_info,
2623                                                   path, record_one_backref,
2624                                                   old, false);
2625                 if (ret < 0 && ret != -ENOENT)
2626                         return false;
2627
2628                 /* no backref to be processed for this extent */
2629                 if (!old->count) {
2630                         list_del(&old->list);
2631                         kfree(old);
2632                 }
2633         }
2634
2635         if (list_empty(&new->head))
2636                 return false;
2637
2638         return true;
2639 }
2640
2641 static int relink_is_mergable(struct extent_buffer *leaf,
2642                               struct btrfs_file_extent_item *fi,
2643                               struct new_sa_defrag_extent *new)
2644 {
2645         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2646                 return 0;
2647
2648         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2649                 return 0;
2650
2651         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2652                 return 0;
2653
2654         if (btrfs_file_extent_encryption(leaf, fi) ||
2655             btrfs_file_extent_other_encoding(leaf, fi))
2656                 return 0;
2657
2658         return 1;
2659 }
2660
2661 /*
2662  * Note the backref might has changed, and in this case we just return 0.
2663  */
2664 static noinline int relink_extent_backref(struct btrfs_path *path,
2665                                  struct sa_defrag_extent_backref *prev,
2666                                  struct sa_defrag_extent_backref *backref)
2667 {
2668         struct btrfs_file_extent_item *extent;
2669         struct btrfs_file_extent_item *item;
2670         struct btrfs_ordered_extent *ordered;
2671         struct btrfs_trans_handle *trans;
2672         struct btrfs_ref ref = { 0 };
2673         struct btrfs_root *root;
2674         struct btrfs_key key;
2675         struct extent_buffer *leaf;
2676         struct old_sa_defrag_extent *old = backref->old;
2677         struct new_sa_defrag_extent *new = old->new;
2678         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2679         struct inode *inode;
2680         struct extent_state *cached = NULL;
2681         int ret = 0;
2682         u64 start;
2683         u64 len;
2684         u64 lock_start;
2685         u64 lock_end;
2686         bool merge = false;
2687         int index;
2688
2689         if (prev && prev->root_id == backref->root_id &&
2690             prev->inum == backref->inum &&
2691             prev->file_pos + prev->num_bytes == backref->file_pos)
2692                 merge = true;
2693
2694         /* step 1: get root */
2695         key.objectid = backref->root_id;
2696         key.type = BTRFS_ROOT_ITEM_KEY;
2697         key.offset = (u64)-1;
2698
2699         index = srcu_read_lock(&fs_info->subvol_srcu);
2700
2701         root = btrfs_read_fs_root_no_name(fs_info, &key);
2702         if (IS_ERR(root)) {
2703                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2704                 if (PTR_ERR(root) == -ENOENT)
2705                         return 0;
2706                 return PTR_ERR(root);
2707         }
2708
2709         if (btrfs_root_readonly(root)) {
2710                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2711                 return 0;
2712         }
2713
2714         /* step 2: get inode */
2715         key.objectid = backref->inum;
2716         key.type = BTRFS_INODE_ITEM_KEY;
2717         key.offset = 0;
2718
2719         inode = btrfs_iget(fs_info->sb, &key, root);
2720         if (IS_ERR(inode)) {
2721                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2722                 return 0;
2723         }
2724
2725         srcu_read_unlock(&fs_info->subvol_srcu, index);
2726
2727         /* step 3: relink backref */
2728         lock_start = backref->file_pos;
2729         lock_end = backref->file_pos + backref->num_bytes - 1;
2730         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2731                          &cached);
2732
2733         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2734         if (ordered) {
2735                 btrfs_put_ordered_extent(ordered);
2736                 goto out_unlock;
2737         }
2738
2739         trans = btrfs_join_transaction(root);
2740         if (IS_ERR(trans)) {
2741                 ret = PTR_ERR(trans);
2742                 goto out_unlock;
2743         }
2744
2745         key.objectid = backref->inum;
2746         key.type = BTRFS_EXTENT_DATA_KEY;
2747         key.offset = backref->file_pos;
2748
2749         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2750         if (ret < 0) {
2751                 goto out_free_path;
2752         } else if (ret > 0) {
2753                 ret = 0;
2754                 goto out_free_path;
2755         }
2756
2757         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2758                                 struct btrfs_file_extent_item);
2759
2760         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2761             backref->generation)
2762                 goto out_free_path;
2763
2764         btrfs_release_path(path);
2765
2766         start = backref->file_pos;
2767         if (backref->extent_offset < old->extent_offset + old->offset)
2768                 start += old->extent_offset + old->offset -
2769                          backref->extent_offset;
2770
2771         len = min(backref->extent_offset + backref->num_bytes,
2772                   old->extent_offset + old->offset + old->len);
2773         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2774
2775         ret = btrfs_drop_extents(trans, root, inode, start,
2776                                  start + len, 1);
2777         if (ret)
2778                 goto out_free_path;
2779 again:
2780         key.objectid = btrfs_ino(BTRFS_I(inode));
2781         key.type = BTRFS_EXTENT_DATA_KEY;
2782         key.offset = start;
2783
2784         path->leave_spinning = 1;
2785         if (merge) {
2786                 struct btrfs_file_extent_item *fi;
2787                 u64 extent_len;
2788                 struct btrfs_key found_key;
2789
2790                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2791                 if (ret < 0)
2792                         goto out_free_path;
2793
2794                 path->slots[0]--;
2795                 leaf = path->nodes[0];
2796                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2797
2798                 fi = btrfs_item_ptr(leaf, path->slots[0],
2799                                     struct btrfs_file_extent_item);
2800                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2801
2802                 if (extent_len + found_key.offset == start &&
2803                     relink_is_mergable(leaf, fi, new)) {
2804                         btrfs_set_file_extent_num_bytes(leaf, fi,
2805                                                         extent_len + len);
2806                         btrfs_mark_buffer_dirty(leaf);
2807                         inode_add_bytes(inode, len);
2808
2809                         ret = 1;
2810                         goto out_free_path;
2811                 } else {
2812                         merge = false;
2813                         btrfs_release_path(path);
2814                         goto again;
2815                 }
2816         }
2817
2818         ret = btrfs_insert_empty_item(trans, root, path, &key,
2819                                         sizeof(*extent));
2820         if (ret) {
2821                 btrfs_abort_transaction(trans, ret);
2822                 goto out_free_path;
2823         }
2824
2825         leaf = path->nodes[0];
2826         item = btrfs_item_ptr(leaf, path->slots[0],
2827                                 struct btrfs_file_extent_item);
2828         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2829         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2830         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2831         btrfs_set_file_extent_num_bytes(leaf, item, len);
2832         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2833         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2834         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2835         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2836         btrfs_set_file_extent_encryption(leaf, item, 0);
2837         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2838
2839         btrfs_mark_buffer_dirty(leaf);
2840         inode_add_bytes(inode, len);
2841         btrfs_release_path(path);
2842
2843         btrfs_init_generic_ref(&ref, BTRFS_ADD_DELAYED_REF, new->bytenr,
2844                                new->disk_len, 0);
2845         btrfs_init_data_ref(&ref, backref->root_id, backref->inum,
2846                             new->file_pos);  /* start - extent_offset */
2847         ret = btrfs_inc_extent_ref(trans, &ref);
2848         if (ret) {
2849                 btrfs_abort_transaction(trans, ret);
2850                 goto out_free_path;
2851         }
2852
2853         ret = 1;
2854 out_free_path:
2855         btrfs_release_path(path);
2856         path->leave_spinning = 0;
2857         btrfs_end_transaction(trans);
2858 out_unlock:
2859         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2860                              &cached);
2861         iput(inode);
2862         return ret;
2863 }
2864
2865 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2866 {
2867         struct old_sa_defrag_extent *old, *tmp;
2868
2869         if (!new)
2870                 return;
2871
2872         list_for_each_entry_safe(old, tmp, &new->head, list) {
2873                 kfree(old);
2874         }
2875         kfree(new);
2876 }
2877
2878 static void relink_file_extents(struct new_sa_defrag_extent *new)
2879 {
2880         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2881         struct btrfs_path *path;
2882         struct sa_defrag_extent_backref *backref;
2883         struct sa_defrag_extent_backref *prev = NULL;
2884         struct rb_node *node;
2885         int ret;
2886
2887         path = btrfs_alloc_path();
2888         if (!path)
2889                 return;
2890
2891         if (!record_extent_backrefs(path, new)) {
2892                 btrfs_free_path(path);
2893                 goto out;
2894         }
2895         btrfs_release_path(path);
2896
2897         while (1) {
2898                 node = rb_first(&new->root);
2899                 if (!node)
2900                         break;
2901                 rb_erase(node, &new->root);
2902
2903                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2904
2905                 ret = relink_extent_backref(path, prev, backref);
2906                 WARN_ON(ret < 0);
2907
2908                 kfree(prev);
2909
2910                 if (ret == 1)
2911                         prev = backref;
2912                 else
2913                         prev = NULL;
2914                 cond_resched();
2915         }
2916         kfree(prev);
2917
2918         btrfs_free_path(path);
2919 out:
2920         free_sa_defrag_extent(new);
2921
2922         atomic_dec(&fs_info->defrag_running);
2923         wake_up(&fs_info->transaction_wait);
2924 }
2925
2926 static struct new_sa_defrag_extent *
2927 record_old_file_extents(struct inode *inode,
2928                         struct btrfs_ordered_extent *ordered)
2929 {
2930         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2931         struct btrfs_root *root = BTRFS_I(inode)->root;
2932         struct btrfs_path *path;
2933         struct btrfs_key key;
2934         struct old_sa_defrag_extent *old;
2935         struct new_sa_defrag_extent *new;
2936         int ret;
2937
2938         new = kmalloc(sizeof(*new), GFP_NOFS);
2939         if (!new)
2940                 return NULL;
2941
2942         new->inode = inode;
2943         new->file_pos = ordered->file_offset;
2944         new->len = ordered->len;
2945         new->bytenr = ordered->start;
2946         new->disk_len = ordered->disk_len;
2947         new->compress_type = ordered->compress_type;
2948         new->root = RB_ROOT;
2949         INIT_LIST_HEAD(&new->head);
2950
2951         path = btrfs_alloc_path();
2952         if (!path)
2953                 goto out_kfree;
2954
2955         key.objectid = btrfs_ino(BTRFS_I(inode));
2956         key.type = BTRFS_EXTENT_DATA_KEY;
2957         key.offset = new->file_pos;
2958
2959         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2960         if (ret < 0)
2961                 goto out_free_path;
2962         if (ret > 0 && path->slots[0] > 0)
2963                 path->slots[0]--;
2964
2965         /* find out all the old extents for the file range */
2966         while (1) {
2967                 struct btrfs_file_extent_item *extent;
2968                 struct extent_buffer *l;
2969                 int slot;
2970                 u64 num_bytes;
2971                 u64 offset;
2972                 u64 end;
2973                 u64 disk_bytenr;
2974                 u64 extent_offset;
2975
2976                 l = path->nodes[0];
2977                 slot = path->slots[0];
2978
2979                 if (slot >= btrfs_header_nritems(l)) {
2980                         ret = btrfs_next_leaf(root, path);
2981                         if (ret < 0)
2982                                 goto out_free_path;
2983                         else if (ret > 0)
2984                                 break;
2985                         continue;
2986                 }
2987
2988                 btrfs_item_key_to_cpu(l, &key, slot);
2989
2990                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2991                         break;
2992                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2993                         break;
2994                 if (key.offset >= new->file_pos + new->len)
2995                         break;
2996
2997                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2998
2999                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
3000                 if (key.offset + num_bytes < new->file_pos)
3001                         goto next;
3002
3003                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
3004                 if (!disk_bytenr)
3005                         goto next;
3006
3007                 extent_offset = btrfs_file_extent_offset(l, extent);
3008
3009                 old = kmalloc(sizeof(*old), GFP_NOFS);
3010                 if (!old)
3011                         goto out_free_path;
3012
3013                 offset = max(new->file_pos, key.offset);
3014                 end = min(new->file_pos + new->len, key.offset + num_bytes);
3015
3016                 old->bytenr = disk_bytenr;
3017                 old->extent_offset = extent_offset;
3018                 old->offset = offset - key.offset;
3019                 old->len = end - offset;
3020                 old->new = new;
3021                 old->count = 0;
3022                 list_add_tail(&old->list, &new->head);
3023 next:
3024                 path->slots[0]++;
3025                 cond_resched();
3026         }
3027
3028         btrfs_free_path(path);
3029         atomic_inc(&fs_info->defrag_running);
3030
3031         return new;
3032
3033 out_free_path:
3034         btrfs_free_path(path);
3035 out_kfree:
3036         free_sa_defrag_extent(new);
3037         return NULL;
3038 }
3039
3040 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
3041                                          u64 start, u64 len)
3042 {
3043         struct btrfs_block_group *cache;
3044
3045         cache = btrfs_lookup_block_group(fs_info, start);
3046         ASSERT(cache);
3047
3048         spin_lock(&cache->lock);
3049         cache->delalloc_bytes -= len;
3050         spin_unlock(&cache->lock);
3051
3052         btrfs_put_block_group(cache);
3053 }
3054
3055 /* as ordered data IO finishes, this gets called so we can finish
3056  * an ordered extent if the range of bytes in the file it covers are
3057  * fully written.
3058  */
3059 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
3060 {
3061         struct inode *inode = ordered_extent->inode;
3062         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3063         struct btrfs_root *root = BTRFS_I(inode)->root;
3064         struct btrfs_trans_handle *trans = NULL;
3065         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3066         struct extent_state *cached_state = NULL;
3067         struct new_sa_defrag_extent *new = NULL;
3068         int compress_type = 0;
3069         int ret = 0;
3070         u64 logical_len = ordered_extent->len;
3071         bool freespace_inode;
3072         bool truncated = false;
3073         bool range_locked = false;
3074         bool clear_new_delalloc_bytes = false;
3075         bool clear_reserved_extent = true;
3076
3077         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3078             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
3079             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
3080                 clear_new_delalloc_bytes = true;
3081
3082         freespace_inode = btrfs_is_free_space_inode(BTRFS_I(inode));
3083
3084         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
3085                 ret = -EIO;
3086                 goto out;
3087         }
3088
3089         btrfs_free_io_failure_record(BTRFS_I(inode),
3090                         ordered_extent->file_offset,
3091                         ordered_extent->file_offset +
3092                         ordered_extent->len - 1);
3093
3094         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
3095                 truncated = true;
3096                 logical_len = ordered_extent->truncated_len;
3097                 /* Truncated the entire extent, don't bother adding */
3098                 if (!logical_len)
3099                         goto out;
3100         }
3101
3102         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
3103                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
3104
3105                 /*
3106                  * For mwrite(mmap + memset to write) case, we still reserve
3107                  * space for NOCOW range.
3108                  * As NOCOW won't cause a new delayed ref, just free the space
3109                  */
3110                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3111                                        ordered_extent->len);
3112                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3113                 if (freespace_inode)
3114                         trans = btrfs_join_transaction_spacecache(root);
3115                 else
3116                         trans = btrfs_join_transaction(root);
3117                 if (IS_ERR(trans)) {
3118                         ret = PTR_ERR(trans);
3119                         trans = NULL;
3120                         goto out;
3121                 }
3122                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3123                 ret = btrfs_update_inode_fallback(trans, root, inode);
3124                 if (ret) /* -ENOMEM or corruption */
3125                         btrfs_abort_transaction(trans, ret);
3126                 goto out;
3127         }
3128
3129         range_locked = true;
3130         lock_extent_bits(io_tree, ordered_extent->file_offset,
3131                          ordered_extent->file_offset + ordered_extent->len - 1,
3132                          &cached_state);
3133
3134         ret = test_range_bit(io_tree, ordered_extent->file_offset,
3135                         ordered_extent->file_offset + ordered_extent->len - 1,
3136                         EXTENT_DEFRAG, 0, cached_state);
3137         if (ret) {
3138                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
3139                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
3140                         /* the inode is shared */
3141                         new = record_old_file_extents(inode, ordered_extent);
3142
3143                 clear_extent_bit(io_tree, ordered_extent->file_offset,
3144                         ordered_extent->file_offset + ordered_extent->len - 1,
3145                         EXTENT_DEFRAG, 0, 0, &cached_state);
3146         }
3147
3148         if (freespace_inode)
3149                 trans = btrfs_join_transaction_spacecache(root);
3150         else
3151                 trans = btrfs_join_transaction(root);
3152         if (IS_ERR(trans)) {
3153                 ret = PTR_ERR(trans);
3154                 trans = NULL;
3155                 goto out;
3156         }
3157
3158         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3159
3160         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3161                 compress_type = ordered_extent->compress_type;
3162         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3163                 BUG_ON(compress_type);
3164                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3165                                        ordered_extent->len);
3166                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3167                                                 ordered_extent->file_offset,
3168                                                 ordered_extent->file_offset +
3169                                                 logical_len);
3170         } else {
3171                 BUG_ON(root == fs_info->tree_root);
3172                 ret = insert_reserved_file_extent(trans, inode,
3173                                                 ordered_extent->file_offset,
3174                                                 ordered_extent->start,
3175                                                 ordered_extent->disk_len,
3176                                                 logical_len, logical_len,
3177                                                 compress_type, 0, 0,
3178                                                 BTRFS_FILE_EXTENT_REG);
3179                 if (!ret) {
3180                         clear_reserved_extent = false;
3181                         btrfs_release_delalloc_bytes(fs_info,
3182                                                      ordered_extent->start,
3183                                                      ordered_extent->disk_len);
3184                 }
3185         }
3186         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3187                            ordered_extent->file_offset, ordered_extent->len,
3188                            trans->transid);
3189         if (ret < 0) {
3190                 btrfs_abort_transaction(trans, ret);
3191                 goto out;
3192         }
3193
3194         ret = add_pending_csums(trans, inode, &ordered_extent->list);
3195         if (ret) {
3196                 btrfs_abort_transaction(trans, ret);
3197                 goto out;
3198         }
3199
3200         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3201         ret = btrfs_update_inode_fallback(trans, root, inode);
3202         if (ret) { /* -ENOMEM or corruption */
3203                 btrfs_abort_transaction(trans, ret);
3204                 goto out;
3205         }
3206         ret = 0;
3207 out:
3208         if (range_locked || clear_new_delalloc_bytes) {
3209                 unsigned int clear_bits = 0;
3210
3211                 if (range_locked)
3212                         clear_bits |= EXTENT_LOCKED;
3213                 if (clear_new_delalloc_bytes)
3214                         clear_bits |= EXTENT_DELALLOC_NEW;
3215                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3216                                  ordered_extent->file_offset,
3217                                  ordered_extent->file_offset +
3218                                  ordered_extent->len - 1,
3219                                  clear_bits,
3220                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3221                                  0, &cached_state);
3222         }
3223
3224         if (trans)
3225                 btrfs_end_transaction(trans);
3226
3227         if (ret || truncated) {
3228                 u64 start, end;
3229
3230                 if (truncated)
3231                         start = ordered_extent->file_offset + logical_len;
3232                 else
3233                         start = ordered_extent->file_offset;
3234                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3235                 clear_extent_uptodate(io_tree, start, end, NULL);
3236
3237                 /* Drop the cache for the part of the extent we didn't write. */
3238                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3239
3240                 /*
3241                  * If the ordered extent had an IOERR or something else went
3242                  * wrong we need to return the space for this ordered extent
3243                  * back to the allocator.  We only free the extent in the
3244                  * truncated case if we didn't write out the extent at all.
3245                  *
3246                  * If we made it past insert_reserved_file_extent before we
3247                  * errored out then we don't need to do this as the accounting
3248                  * has already been done.
3249                  */
3250                 if ((ret || !logical_len) &&
3251                     clear_reserved_extent &&
3252                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3253                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3254                         btrfs_free_reserved_extent(fs_info,
3255                                                    ordered_extent->start,
3256                                                    ordered_extent->disk_len, 1);
3257         }
3258
3259
3260         /*
3261          * This needs to be done to make sure anybody waiting knows we are done
3262          * updating everything for this ordered extent.
3263          */
3264         btrfs_remove_ordered_extent(inode, ordered_extent);
3265
3266         /* for snapshot-aware defrag */
3267         if (new) {
3268                 if (ret) {
3269                         free_sa_defrag_extent(new);
3270                         atomic_dec(&fs_info->defrag_running);
3271                 } else {
3272                         relink_file_extents(new);
3273                 }
3274         }
3275
3276         /* once for us */
3277         btrfs_put_ordered_extent(ordered_extent);
3278         /* once for the tree */
3279         btrfs_put_ordered_extent(ordered_extent);
3280
3281         return ret;
3282 }
3283
3284 static void finish_ordered_fn(struct btrfs_work *work)
3285 {
3286         struct btrfs_ordered_extent *ordered_extent;
3287         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3288         btrfs_finish_ordered_io(ordered_extent);
3289 }
3290
3291 void btrfs_writepage_endio_finish_ordered(struct page *page, u64 start,
3292                                           u64 end, int uptodate)
3293 {
3294         struct inode *inode = page->mapping->host;
3295         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3296         struct btrfs_ordered_extent *ordered_extent = NULL;
3297         struct btrfs_workqueue *wq;
3298
3299         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3300
3301         ClearPagePrivate2(page);
3302         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3303                                             end - start + 1, uptodate))
3304                 return;
3305
3306         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
3307                 wq = fs_info->endio_freespace_worker;
3308         else
3309                 wq = fs_info->endio_write_workers;
3310
3311         btrfs_init_work(&ordered_extent->work, finish_ordered_fn, NULL, NULL);
3312         btrfs_queue_work(wq, &ordered_extent->work);
3313 }
3314
3315 static int __readpage_endio_check(struct inode *inode,
3316                                   struct btrfs_io_bio *io_bio,
3317                                   int icsum, struct page *page,
3318                                   int pgoff, u64 start, size_t len)
3319 {
3320         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3321         SHASH_DESC_ON_STACK(shash, fs_info->csum_shash);
3322         char *kaddr;
3323         u16 csum_size = btrfs_super_csum_size(fs_info->super_copy);
3324         u8 *csum_expected;
3325         u8 csum[BTRFS_CSUM_SIZE];
3326
3327         csum_expected = ((u8 *)io_bio->csum) + icsum * csum_size;
3328
3329         kaddr = kmap_atomic(page);
3330         shash->tfm = fs_info->csum_shash;
3331
3332         crypto_shash_init(shash);
3333         crypto_shash_update(shash, kaddr + pgoff, len);
3334         crypto_shash_final(shash, csum);
3335
3336         if (memcmp(csum, csum_expected, csum_size))
3337                 goto zeroit;
3338
3339         kunmap_atomic(kaddr);
3340         return 0;
3341 zeroit:
3342         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3343                                     io_bio->mirror_num);
3344         memset(kaddr + pgoff, 1, len);
3345         flush_dcache_page(page);
3346         kunmap_atomic(kaddr);
3347         return -EIO;
3348 }
3349
3350 /*
3351  * when reads are done, we need to check csums to verify the data is correct
3352  * if there's a match, we allow the bio to finish.  If not, the code in
3353  * extent_io.c will try to find good copies for us.
3354  */
3355 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3356                                       u64 phy_offset, struct page *page,
3357                                       u64 start, u64 end, int mirror)
3358 {
3359         size_t offset = start - page_offset(page);
3360         struct inode *inode = page->mapping->host;
3361         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3362         struct btrfs_root *root = BTRFS_I(inode)->root;
3363
3364         if (PageChecked(page)) {
3365                 ClearPageChecked(page);
3366                 return 0;
3367         }
3368
3369         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3370                 return 0;
3371
3372         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3373             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3374                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3375                 return 0;
3376         }
3377
3378         phy_offset >>= inode->i_sb->s_blocksize_bits;
3379         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3380                                       start, (size_t)(end - start + 1));
3381 }
3382
3383 /*
3384  * btrfs_add_delayed_iput - perform a delayed iput on @inode
3385  *
3386  * @inode: The inode we want to perform iput on
3387  *
3388  * This function uses the generic vfs_inode::i_count to track whether we should
3389  * just decrement it (in case it's > 1) or if this is the last iput then link
3390  * the inode to the delayed iput machinery. Delayed iputs are processed at
3391  * transaction commit time/superblock commit/cleaner kthread.
3392  */
3393 void btrfs_add_delayed_iput(struct inode *inode)
3394 {
3395         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3396         struct btrfs_inode *binode = BTRFS_I(inode);
3397
3398         if (atomic_add_unless(&inode->i_count, -1, 1))
3399                 return;
3400
3401         atomic_inc(&fs_info->nr_delayed_iputs);
3402         spin_lock(&fs_info->delayed_iput_lock);
3403         ASSERT(list_empty(&binode->delayed_iput));
3404         list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3405         spin_unlock(&fs_info->delayed_iput_lock);
3406         if (!test_bit(BTRFS_FS_CLEANER_RUNNING, &fs_info->flags))
3407                 wake_up_process(fs_info->cleaner_kthread);
3408 }
3409
3410 static void run_delayed_iput_locked(struct btrfs_fs_info *fs_info,
3411                                     struct btrfs_inode *inode)
3412 {
3413         list_del_init(&inode->delayed_iput);
3414         spin_unlock(&fs_info->delayed_iput_lock);
3415         iput(&inode->vfs_inode);
3416         if (atomic_dec_and_test(&fs_info->nr_delayed_iputs))
3417                 wake_up(&fs_info->delayed_iputs_wait);
3418         spin_lock(&fs_info->delayed_iput_lock);
3419 }
3420
3421 static void btrfs_run_delayed_iput(struct btrfs_fs_info *fs_info,
3422                                    struct btrfs_inode *inode)
3423 {
3424         if (!list_empty(&inode->delayed_iput)) {
3425                 spin_lock(&fs_info->delayed_iput_lock);
3426                 if (!list_empty(&inode->delayed_iput))
3427                         run_delayed_iput_locked(fs_info, inode);
3428                 spin_unlock(&fs_info->delayed_iput_lock);
3429         }
3430 }
3431
3432 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3433 {
3434
3435         spin_lock(&fs_info->delayed_iput_lock);
3436         while (!list_empty(&fs_info->delayed_iputs)) {
3437                 struct btrfs_inode *inode;
3438
3439                 inode = list_first_entry(&fs_info->delayed_iputs,
3440                                 struct btrfs_inode, delayed_iput);
3441                 run_delayed_iput_locked(fs_info, inode);
3442         }
3443         spin_unlock(&fs_info->delayed_iput_lock);
3444 }
3445
3446 /**
3447  * btrfs_wait_on_delayed_iputs - wait on the delayed iputs to be done running
3448  * @fs_info - the fs_info for this fs
3449  * @return - EINTR if we were killed, 0 if nothing's pending
3450  *
3451  * This will wait on any delayed iputs that are currently running with KILLABLE
3452  * set.  Once they are all done running we will return, unless we are killed in
3453  * which case we return EINTR. This helps in user operations like fallocate etc
3454  * that might get blocked on the iputs.
3455  */
3456 int btrfs_wait_on_delayed_iputs(struct btrfs_fs_info *fs_info)
3457 {
3458         int ret = wait_event_killable(fs_info->delayed_iputs_wait,
3459                         atomic_read(&fs_info->nr_delayed_iputs) == 0);
3460         if (ret)
3461                 return -EINTR;
3462         return 0;
3463 }
3464
3465 /*
3466  * This creates an orphan entry for the given inode in case something goes wrong
3467  * in the middle of an unlink.
3468  */
3469 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3470                      struct btrfs_inode *inode)
3471 {
3472         int ret;
3473
3474         ret = btrfs_insert_orphan_item(trans, inode->root, btrfs_ino(inode));
3475         if (ret && ret != -EEXIST) {
3476                 btrfs_abort_transaction(trans, ret);
3477                 return ret;
3478         }
3479
3480         return 0;
3481 }
3482
3483 /*
3484  * We have done the delete so we can go ahead and remove the orphan item for
3485  * this particular inode.
3486  */
3487 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3488                             struct btrfs_inode *inode)
3489 {
3490         return btrfs_del_orphan_item(trans, inode->root, btrfs_ino(inode));
3491 }
3492
3493 /*
3494  * this cleans up any orphans that may be left on the list from the last use
3495  * of this root.
3496  */
3497 int btrfs_orphan_cleanup(struct btrfs_root *root)
3498 {
3499         struct btrfs_fs_info *fs_info = root->fs_info;
3500         struct btrfs_path *path;
3501         struct extent_buffer *leaf;
3502         struct btrfs_key key, found_key;
3503         struct btrfs_trans_handle *trans;
3504         struct inode *inode;
3505         u64 last_objectid = 0;
3506         int ret = 0, nr_unlink = 0;
3507
3508         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3509                 return 0;
3510
3511         path = btrfs_alloc_path();
3512         if (!path) {
3513                 ret = -ENOMEM;
3514                 goto out;
3515         }
3516         path->reada = READA_BACK;
3517
3518         key.objectid = BTRFS_ORPHAN_OBJECTID;
3519         key.type = BTRFS_ORPHAN_ITEM_KEY;
3520         key.offset = (u64)-1;
3521
3522         while (1) {
3523                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3524                 if (ret < 0)
3525                         goto out;
3526
3527                 /*
3528                  * if ret == 0 means we found what we were searching for, which
3529                  * is weird, but possible, so only screw with path if we didn't
3530                  * find the key and see if we have stuff that matches
3531                  */
3532                 if (ret > 0) {
3533                         ret = 0;
3534                         if (path->slots[0] == 0)
3535                                 break;
3536                         path->slots[0]--;
3537                 }
3538
3539                 /* pull out the item */
3540                 leaf = path->nodes[0];
3541                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3542
3543                 /* make sure the item matches what we want */
3544                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3545                         break;
3546                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3547                         break;
3548
3549                 /* release the path since we're done with it */
3550                 btrfs_release_path(path);
3551
3552                 /*
3553                  * this is where we are basically btrfs_lookup, without the
3554                  * crossing root thing.  we store the inode number in the
3555                  * offset of the orphan item.
3556                  */
3557
3558                 if (found_key.offset == last_objectid) {
3559                         btrfs_err(fs_info,
3560                                   "Error removing orphan entry, stopping orphan cleanup");
3561                         ret = -EINVAL;
3562                         goto out;
3563                 }
3564
3565                 last_objectid = found_key.offset;
3566
3567                 found_key.objectid = found_key.offset;
3568                 found_key.type = BTRFS_INODE_ITEM_KEY;
3569                 found_key.offset = 0;
3570                 inode = btrfs_iget(fs_info->sb, &found_key, root);
3571                 ret = PTR_ERR_OR_ZERO(inode);
3572                 if (ret && ret != -ENOENT)
3573                         goto out;
3574
3575                 if (ret == -ENOENT && root == fs_info->tree_root) {
3576                         struct btrfs_root *dead_root;
3577                         struct btrfs_fs_info *fs_info = root->fs_info;
3578                         int is_dead_root = 0;
3579
3580                         /*
3581                          * this is an orphan in the tree root. Currently these
3582                          * could come from 2 sources:
3583                          *  a) a snapshot deletion in progress
3584                          *  b) a free space cache inode
3585                          * We need to distinguish those two, as the snapshot
3586                          * orphan must not get deleted.
3587                          * find_dead_roots already ran before us, so if this
3588                          * is a snapshot deletion, we should find the root
3589                          * in the dead_roots list
3590                          */
3591                         spin_lock(&fs_info->trans_lock);
3592                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3593                                             root_list) {
3594                                 if (dead_root->root_key.objectid ==
3595                                     found_key.objectid) {
3596                                         is_dead_root = 1;
3597                                         break;
3598                                 }
3599                         }
3600                         spin_unlock(&fs_info->trans_lock);
3601                         if (is_dead_root) {
3602                                 /* prevent this orphan from being found again */
3603                                 key.offset = found_key.objectid - 1;
3604                                 continue;
3605                         }
3606
3607                 }
3608
3609                 /*
3610                  * If we have an inode with links, there are a couple of
3611                  * possibilities. Old kernels (before v3.12) used to create an
3612                  * orphan item for truncate indicating that there were possibly
3613                  * extent items past i_size that needed to be deleted. In v3.12,
3614                  * truncate was changed to update i_size in sync with the extent
3615                  * items, but the (useless) orphan item was still created. Since
3616                  * v4.18, we don't create the orphan item for truncate at all.
3617                  *
3618                  * So, this item could mean that we need to do a truncate, but
3619                  * only if this filesystem was last used on a pre-v3.12 kernel
3620                  * and was not cleanly unmounted. The odds of that are quite
3621                  * slim, and it's a pain to do the truncate now, so just delete
3622                  * the orphan item.
3623                  *
3624                  * It's also possible that this orphan item was supposed to be
3625                  * deleted but wasn't. The inode number may have been reused,
3626                  * but either way, we can delete the orphan item.
3627                  */
3628                 if (ret == -ENOENT || inode->i_nlink) {
3629                         if (!ret)
3630                                 iput(inode);
3631                         trans = btrfs_start_transaction(root, 1);
3632                         if (IS_ERR(trans)) {
3633                                 ret = PTR_ERR(trans);
3634                                 goto out;
3635                         }
3636                         btrfs_debug(fs_info, "auto deleting %Lu",
3637                                     found_key.objectid);
3638                         ret = btrfs_del_orphan_item(trans, root,
3639                                                     found_key.objectid);
3640                         btrfs_end_transaction(trans);
3641                         if (ret)
3642                                 goto out;
3643                         continue;
3644                 }
3645
3646                 nr_unlink++;
3647
3648                 /* this will do delete_inode and everything for us */
3649                 iput(inode);
3650         }
3651         /* release the path since we're done with it */
3652         btrfs_release_path(path);
3653
3654         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3655
3656         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3657                 trans = btrfs_join_transaction(root);
3658                 if (!IS_ERR(trans))
3659                         btrfs_end_transaction(trans);
3660         }
3661
3662         if (nr_unlink)
3663                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3664
3665 out:
3666         if (ret)
3667                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3668         btrfs_free_path(path);
3669         return ret;
3670 }
3671
3672 /*
3673  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3674  * don't find any xattrs, we know there can't be any acls.
3675  *
3676  * slot is the slot the inode is in, objectid is the objectid of the inode
3677  */
3678 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3679                                           int slot, u64 objectid,
3680                                           int *first_xattr_slot)
3681 {
3682         u32 nritems = btrfs_header_nritems(leaf);
3683         struct btrfs_key found_key;
3684         static u64 xattr_access = 0;
3685         static u64 xattr_default = 0;
3686         int scanned = 0;
3687
3688         if (!xattr_access) {
3689                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3690                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3691                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3692                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3693         }
3694
3695         slot++;
3696         *first_xattr_slot = -1;
3697         while (slot < nritems) {
3698                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3699
3700                 /* we found a different objectid, there must not be acls */
3701                 if (found_key.objectid != objectid)
3702                         return 0;
3703
3704                 /* we found an xattr, assume we've got an acl */
3705                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3706                         if (*first_xattr_slot == -1)
3707                                 *first_xattr_slot = slot;
3708                         if (found_key.offset == xattr_access ||
3709                             found_key.offset == xattr_default)
3710                                 return 1;
3711                 }
3712
3713                 /*
3714                  * we found a key greater than an xattr key, there can't
3715                  * be any acls later on
3716                  */
3717                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3718                         return 0;
3719
3720                 slot++;
3721                 scanned++;
3722
3723                 /*
3724                  * it goes inode, inode backrefs, xattrs, extents,
3725                  * so if there are a ton of hard links to an inode there can
3726                  * be a lot of backrefs.  Don't waste time searching too hard,
3727                  * this is just an optimization
3728                  */
3729                 if (scanned >= 8)
3730                         break;
3731         }
3732         /* we hit the end of the leaf before we found an xattr or
3733          * something larger than an xattr.  We have to assume the inode
3734          * has acls
3735          */
3736         if (*first_xattr_slot == -1)
3737                 *first_xattr_slot = slot;
3738         return 1;
3739 }
3740
3741 /*
3742  * read an inode from the btree into the in-memory inode
3743  */
3744 static int btrfs_read_locked_inode(struct inode *inode,
3745                                    struct btrfs_path *in_path)
3746 {
3747         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3748         struct btrfs_path *path = in_path;
3749         struct extent_buffer *leaf;
3750         struct btrfs_inode_item *inode_item;
3751         struct btrfs_root *root = BTRFS_I(inode)->root;
3752         struct btrfs_key location;
3753         unsigned long ptr;
3754         int maybe_acls;
3755         u32 rdev;
3756         int ret;
3757         bool filled = false;
3758         int first_xattr_slot;
3759
3760         ret = btrfs_fill_inode(inode, &rdev);
3761         if (!ret)
3762                 filled = true;
3763
3764         if (!path) {
3765                 path = btrfs_alloc_path();
3766                 if (!path)
3767                         return -ENOMEM;
3768         }
3769
3770         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3771
3772         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3773         if (ret) {
3774                 if (path != in_path)
3775                         btrfs_free_path(path);
3776                 return ret;
3777         }
3778
3779         leaf = path->nodes[0];
3780
3781         if (filled)
3782                 goto cache_index;
3783
3784         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3785                                     struct btrfs_inode_item);
3786         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3787         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3788         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3789         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3790         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3791
3792         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3793         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3794
3795         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3796         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3797
3798         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3799         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3800
3801         BTRFS_I(inode)->i_otime.tv_sec =
3802                 btrfs_timespec_sec(leaf, &inode_item->otime);
3803         BTRFS_I(inode)->i_otime.tv_nsec =
3804                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3805
3806         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3807         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3808         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3809
3810         inode_set_iversion_queried(inode,
3811                                    btrfs_inode_sequence(leaf, inode_item));
3812         inode->i_generation = BTRFS_I(inode)->generation;
3813         inode->i_rdev = 0;
3814         rdev = btrfs_inode_rdev(leaf, inode_item);
3815
3816         BTRFS_I(inode)->index_cnt = (u64)-1;
3817         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3818
3819 cache_index:
3820         /*
3821          * If we were modified in the current generation and evicted from memory
3822          * and then re-read we need to do a full sync since we don't have any
3823          * idea about which extents were modified before we were evicted from
3824          * cache.
3825          *
3826          * This is required for both inode re-read from disk and delayed inode
3827          * in delayed_nodes_tree.
3828          */
3829         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3830                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3831                         &BTRFS_I(inode)->runtime_flags);
3832
3833         /*
3834          * We don't persist the id of the transaction where an unlink operation
3835          * against the inode was last made. So here we assume the inode might
3836          * have been evicted, and therefore the exact value of last_unlink_trans
3837          * lost, and set it to last_trans to avoid metadata inconsistencies
3838          * between the inode and its parent if the inode is fsync'ed and the log
3839          * replayed. For example, in the scenario:
3840          *
3841          * touch mydir/foo
3842          * ln mydir/foo mydir/bar
3843          * sync
3844          * unlink mydir/bar
3845          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3846          * xfs_io -c fsync mydir/foo
3847          * <power failure>
3848          * mount fs, triggers fsync log replay
3849          *
3850          * We must make sure that when we fsync our inode foo we also log its
3851          * parent inode, otherwise after log replay the parent still has the
3852          * dentry with the "bar" name but our inode foo has a link count of 1
3853          * and doesn't have an inode ref with the name "bar" anymore.
3854          *
3855          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3856          * but it guarantees correctness at the expense of occasional full
3857          * transaction commits on fsync if our inode is a directory, or if our
3858          * inode is not a directory, logging its parent unnecessarily.
3859          */
3860         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3861
3862         path->slots[0]++;
3863         if (inode->i_nlink != 1 ||
3864             path->slots[0] >= btrfs_header_nritems(leaf))
3865                 goto cache_acl;
3866
3867         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3868         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3869                 goto cache_acl;
3870
3871         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3872         if (location.type == BTRFS_INODE_REF_KEY) {
3873                 struct btrfs_inode_ref *ref;
3874
3875                 ref = (struct btrfs_inode_ref *)ptr;
3876                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3877         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3878                 struct btrfs_inode_extref *extref;
3879
3880                 extref = (struct btrfs_inode_extref *)ptr;
3881                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3882                                                                      extref);
3883         }
3884 cache_acl:
3885         /*
3886          * try to precache a NULL acl entry for files that don't have
3887          * any xattrs or acls
3888          */
3889         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3890                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3891         if (first_xattr_slot != -1) {
3892                 path->slots[0] = first_xattr_slot;
3893                 ret = btrfs_load_inode_props(inode, path);
3894                 if (ret)
3895                         btrfs_err(fs_info,
3896                                   "error loading props for ino %llu (root %llu): %d",
3897                                   btrfs_ino(BTRFS_I(inode)),
3898                                   root->root_key.objectid, ret);
3899         }
3900         if (path != in_path)
3901                 btrfs_free_path(path);
3902
3903         if (!maybe_acls)
3904                 cache_no_acl(inode);
3905
3906         switch (inode->i_mode & S_IFMT) {
3907         case S_IFREG:
3908                 inode->i_mapping->a_ops = &btrfs_aops;
3909                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3910                 inode->i_fop = &btrfs_file_operations;
3911                 inode->i_op = &btrfs_file_inode_operations;
3912                 break;
3913         case S_IFDIR:
3914                 inode->i_fop = &btrfs_dir_file_operations;
3915                 inode->i_op = &btrfs_dir_inode_operations;
3916                 break;
3917         case S_IFLNK:
3918                 inode->i_op = &btrfs_symlink_inode_operations;
3919                 inode_nohighmem(inode);
3920                 inode->i_mapping->a_ops = &btrfs_aops;
3921                 break;
3922         default:
3923                 inode->i_op = &btrfs_special_inode_operations;
3924                 init_special_inode(inode, inode->i_mode, rdev);
3925                 break;
3926         }
3927
3928         btrfs_sync_inode_flags_to_i_flags(inode);
3929         return 0;
3930 }
3931
3932 /*
3933  * given a leaf and an inode, copy the inode fields into the leaf
3934  */
3935 static void fill_inode_item(struct btrfs_trans_handle *trans,
3936                             struct extent_buffer *leaf,
3937                             struct btrfs_inode_item *item,
3938                             struct inode *inode)
3939 {
3940         struct btrfs_map_token token;
3941
3942         btrfs_init_map_token(&token, leaf);
3943
3944         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3945         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3946         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3947                                    &token);
3948         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3949         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3950
3951         btrfs_set_token_timespec_sec(leaf, &item->atime,
3952                                      inode->i_atime.tv_sec, &token);
3953         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3954                                       inode->i_atime.tv_nsec, &token);
3955
3956         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3957                                      inode->i_mtime.tv_sec, &token);
3958         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3959                                       inode->i_mtime.tv_nsec, &token);
3960
3961         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3962                                      inode->i_ctime.tv_sec, &token);
3963         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3964                                       inode->i_ctime.tv_nsec, &token);
3965
3966         btrfs_set_token_timespec_sec(leaf, &item->otime,
3967                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3968         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3969                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3970
3971         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3972                                      &token);
3973         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3974                                          &token);
3975         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3976                                        &token);
3977         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3978         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3979         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3980         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3981 }
3982
3983 /*
3984  * copy everything in the in-memory inode into the btree.
3985  */
3986 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3987                                 struct btrfs_root *root, struct inode *inode)
3988 {
3989         struct btrfs_inode_item *inode_item;
3990         struct btrfs_path *path;
3991         struct extent_buffer *leaf;
3992         int ret;
3993
3994         path = btrfs_alloc_path();
3995         if (!path)
3996                 return -ENOMEM;
3997
3998         path->leave_spinning = 1;
3999         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
4000                                  1);
4001         if (ret) {
4002                 if (ret > 0)
4003                         ret = -ENOENT;
4004                 goto failed;
4005         }
4006
4007         leaf = path->nodes[0];
4008         inode_item = btrfs_item_ptr(leaf, path->slots[0],
4009                                     struct btrfs_inode_item);
4010
4011         fill_inode_item(trans, leaf, inode_item, inode);
4012         btrfs_mark_buffer_dirty(leaf);
4013         btrfs_set_inode_last_trans(trans, inode);
4014         ret = 0;
4015 failed:
4016         btrfs_free_path(path);
4017         return ret;
4018 }
4019
4020 /*
4021  * copy everything in the in-memory inode into the btree.
4022  */
4023 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
4024                                 struct btrfs_root *root, struct inode *inode)
4025 {
4026         struct btrfs_fs_info *fs_info = root->fs_info;
4027         int ret;
4028
4029         /*
4030          * If the inode is a free space inode, we can deadlock during commit
4031          * if we put it into the delayed code.
4032          *
4033          * The data relocation inode should also be directly updated
4034          * without delay
4035          */
4036         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4037             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4038             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4039                 btrfs_update_root_times(trans, root);
4040
4041                 ret = btrfs_delayed_update_inode(trans, root, inode);
4042                 if (!ret)
4043                         btrfs_set_inode_last_trans(trans, inode);
4044                 return ret;
4045         }
4046
4047         return btrfs_update_inode_item(trans, root, inode);
4048 }
4049
4050 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4051                                          struct btrfs_root *root,
4052                                          struct inode *inode)
4053 {
4054         int ret;
4055
4056         ret = btrfs_update_inode(trans, root, inode);
4057         if (ret == -ENOSPC)
4058                 return btrfs_update_inode_item(trans, root, inode);
4059         return ret;
4060 }
4061
4062 /*
4063  * unlink helper that gets used here in inode.c and in the tree logging
4064  * recovery code.  It remove a link in a directory with a given name, and
4065  * also drops the back refs in the inode to the directory
4066  */
4067 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4068                                 struct btrfs_root *root,
4069                                 struct btrfs_inode *dir,
4070                                 struct btrfs_inode *inode,
4071                                 const char *name, int name_len)
4072 {
4073         struct btrfs_fs_info *fs_info = root->fs_info;
4074         struct btrfs_path *path;
4075         int ret = 0;
4076         struct btrfs_dir_item *di;
4077         u64 index;
4078         u64 ino = btrfs_ino(inode);
4079         u64 dir_ino = btrfs_ino(dir);
4080
4081         path = btrfs_alloc_path();
4082         if (!path) {
4083                 ret = -ENOMEM;
4084                 goto out;
4085         }
4086
4087         path->leave_spinning = 1;
4088         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4089                                     name, name_len, -1);
4090         if (IS_ERR_OR_NULL(di)) {
4091                 ret = di ? PTR_ERR(di) : -ENOENT;
4092                 goto err;
4093         }
4094         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4095         if (ret)
4096                 goto err;
4097         btrfs_release_path(path);
4098
4099         /*
4100          * If we don't have dir index, we have to get it by looking up
4101          * the inode ref, since we get the inode ref, remove it directly,
4102          * it is unnecessary to do delayed deletion.
4103          *
4104          * But if we have dir index, needn't search inode ref to get it.
4105          * Since the inode ref is close to the inode item, it is better
4106          * that we delay to delete it, and just do this deletion when
4107          * we update the inode item.
4108          */
4109         if (inode->dir_index) {
4110                 ret = btrfs_delayed_delete_inode_ref(inode);
4111                 if (!ret) {
4112                         index = inode->dir_index;
4113                         goto skip_backref;
4114                 }
4115         }
4116
4117         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4118                                   dir_ino, &index);
4119         if (ret) {
4120                 btrfs_info(fs_info,
4121                         "failed to delete reference to %.*s, inode %llu parent %llu",
4122                         name_len, name, ino, dir_ino);
4123                 btrfs_abort_transaction(trans, ret);
4124                 goto err;
4125         }
4126 skip_backref:
4127         ret = btrfs_delete_delayed_dir_index(trans, dir, index);
4128         if (ret) {
4129                 btrfs_abort_transaction(trans, ret);
4130                 goto err;
4131         }
4132
4133         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4134                         dir_ino);
4135         if (ret != 0 && ret != -ENOENT) {
4136                 btrfs_abort_transaction(trans, ret);
4137                 goto err;
4138         }
4139
4140         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4141                         index);
4142         if (ret == -ENOENT)
4143                 ret = 0;
4144         else if (ret)
4145                 btrfs_abort_transaction(trans, ret);
4146
4147         /*
4148          * If we have a pending delayed iput we could end up with the final iput
4149          * being run in btrfs-cleaner context.  If we have enough of these built
4150          * up we can end up burning a lot of time in btrfs-cleaner without any
4151          * way to throttle the unlinks.  Since we're currently holding a ref on
4152          * the inode we can run the delayed iput here without any issues as the
4153          * final iput won't be done until after we drop the ref we're currently
4154          * holding.
4155          */
4156         btrfs_run_delayed_iput(fs_info, inode);
4157 err:
4158         btrfs_free_path(path);
4159         if (ret)
4160                 goto out;
4161
4162         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4163         inode_inc_iversion(&inode->vfs_inode);
4164         inode_inc_iversion(&dir->vfs_inode);
4165         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4166                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4167         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4168 out:
4169         return ret;
4170 }
4171
4172 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4173                        struct btrfs_root *root,
4174                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4175                        const char *name, int name_len)
4176 {
4177         int ret;
4178         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4179         if (!ret) {
4180                 drop_nlink(&inode->vfs_inode);
4181                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4182         }
4183         return ret;
4184 }
4185
4186 /*
4187  * helper to start transaction for unlink and rmdir.
4188  *
4189  * unlink and rmdir are special in btrfs, they do not always free space, so
4190  * if we cannot make our reservations the normal way try and see if there is
4191  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4192  * allow the unlink to occur.
4193  */
4194 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4195 {
4196         struct btrfs_root *root = BTRFS_I(dir)->root;
4197
4198         /*
4199          * 1 for the possible orphan item
4200          * 1 for the dir item
4201          * 1 for the dir index
4202          * 1 for the inode ref
4203          * 1 for the inode
4204          */
4205         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4206 }
4207
4208 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4209 {
4210         struct btrfs_root *root = BTRFS_I(dir)->root;
4211         struct btrfs_trans_handle *trans;
4212         struct inode *inode = d_inode(dentry);
4213         int ret;
4214
4215         trans = __unlink_start_trans(dir);
4216         if (IS_ERR(trans))
4217                 return PTR_ERR(trans);
4218
4219         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4220                         0);
4221
4222         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4223                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4224                         dentry->d_name.len);
4225         if (ret)
4226                 goto out;
4227
4228         if (inode->i_nlink == 0) {
4229                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4230                 if (ret)
4231                         goto out;
4232         }
4233
4234 out:
4235         btrfs_end_transaction(trans);
4236         btrfs_btree_balance_dirty(root->fs_info);
4237         return ret;
4238 }
4239
4240 static int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4241                                struct inode *dir, u64 objectid,
4242                                const char *name, int name_len)
4243 {
4244         struct btrfs_root *root = BTRFS_I(dir)->root;
4245         struct btrfs_path *path;
4246         struct extent_buffer *leaf;
4247         struct btrfs_dir_item *di;
4248         struct btrfs_key key;
4249         u64 index;
4250         int ret;
4251         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4252
4253         path = btrfs_alloc_path();
4254         if (!path)
4255                 return -ENOMEM;
4256
4257         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4258                                    name, name_len, -1);
4259         if (IS_ERR_OR_NULL(di)) {
4260                 ret = di ? PTR_ERR(di) : -ENOENT;
4261                 goto out;
4262         }
4263
4264         leaf = path->nodes[0];
4265         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4266         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4267         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4268         if (ret) {
4269                 btrfs_abort_transaction(trans, ret);
4270                 goto out;
4271         }
4272         btrfs_release_path(path);
4273
4274         ret = btrfs_del_root_ref(trans, objectid, root->root_key.objectid,
4275                                  dir_ino, &index, name, name_len);
4276         if (ret < 0) {
4277                 if (ret != -ENOENT) {
4278                         btrfs_abort_transaction(trans, ret);
4279                         goto out;
4280                 }
4281                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4282                                                  name, name_len);
4283                 if (IS_ERR_OR_NULL(di)) {
4284                         if (!di)
4285                                 ret = -ENOENT;
4286                         else
4287                                 ret = PTR_ERR(di);
4288                         btrfs_abort_transaction(trans, ret);
4289                         goto out;
4290                 }
4291
4292                 leaf = path->nodes[0];
4293                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4294                 index = key.offset;
4295         }
4296         btrfs_release_path(path);
4297
4298         ret = btrfs_delete_delayed_dir_index(trans, BTRFS_I(dir), index);
4299         if (ret) {
4300                 btrfs_abort_transaction(trans, ret);
4301                 goto out;
4302         }
4303
4304         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4305         inode_inc_iversion(dir);
4306         dir->i_mtime = dir->i_ctime = current_time(dir);
4307         ret = btrfs_update_inode_fallback(trans, root, dir);
4308         if (ret)
4309                 btrfs_abort_transaction(trans, ret);
4310 out:
4311         btrfs_free_path(path);
4312         return ret;
4313 }
4314
4315 /*
4316  * Helper to check if the subvolume references other subvolumes or if it's
4317  * default.
4318  */
4319 static noinline int may_destroy_subvol(struct btrfs_root *root)
4320 {
4321         struct btrfs_fs_info *fs_info = root->fs_info;
4322         struct btrfs_path *path;
4323         struct btrfs_dir_item *di;
4324         struct btrfs_key key;
4325         u64 dir_id;
4326         int ret;
4327
4328         path = btrfs_alloc_path();
4329         if (!path)
4330                 return -ENOMEM;
4331
4332         /* Make sure this root isn't set as the default subvol */
4333         dir_id = btrfs_super_root_dir(fs_info->super_copy);
4334         di = btrfs_lookup_dir_item(NULL, fs_info->tree_root, path,
4335                                    dir_id, "default", 7, 0);
4336         if (di && !IS_ERR(di)) {
4337                 btrfs_dir_item_key_to_cpu(path->nodes[0], di, &key);
4338                 if (key.objectid == root->root_key.objectid) {
4339                         ret = -EPERM;
4340                         btrfs_err(fs_info,
4341                                   "deleting default subvolume %llu is not allowed",
4342                                   key.objectid);
4343                         goto out;
4344                 }
4345                 btrfs_release_path(path);
4346         }
4347
4348         key.objectid = root->root_key.objectid;
4349         key.type = BTRFS_ROOT_REF_KEY;
4350         key.offset = (u64)-1;
4351
4352         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
4353         if (ret < 0)
4354                 goto out;
4355         BUG_ON(ret == 0);
4356
4357         ret = 0;
4358         if (path->slots[0] > 0) {
4359                 path->slots[0]--;
4360                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
4361                 if (key.objectid == root->root_key.objectid &&
4362                     key.type == BTRFS_ROOT_REF_KEY)
4363                         ret = -ENOTEMPTY;
4364         }
4365 out:
4366         btrfs_free_path(path);
4367         return ret;
4368 }
4369
4370 /* Delete all dentries for inodes belonging to the root */
4371 static void btrfs_prune_dentries(struct btrfs_root *root)
4372 {
4373         struct btrfs_fs_info *fs_info = root->fs_info;
4374         struct rb_node *node;
4375         struct rb_node *prev;
4376         struct btrfs_inode *entry;
4377         struct inode *inode;
4378         u64 objectid = 0;
4379
4380         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
4381                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
4382
4383         spin_lock(&root->inode_lock);
4384 again:
4385         node = root->inode_tree.rb_node;
4386         prev = NULL;
4387         while (node) {
4388                 prev = node;
4389                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4390
4391                 if (objectid < btrfs_ino(entry))
4392                         node = node->rb_left;
4393                 else if (objectid > btrfs_ino(entry))
4394                         node = node->rb_right;
4395                 else
4396                         break;
4397         }
4398         if (!node) {
4399                 while (prev) {
4400                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
4401                         if (objectid <= btrfs_ino(entry)) {
4402                                 node = prev;
4403                                 break;
4404                         }
4405                         prev = rb_next(prev);
4406                 }
4407         }
4408         while (node) {
4409                 entry = rb_entry(node, struct btrfs_inode, rb_node);
4410                 objectid = btrfs_ino(entry) + 1;
4411                 inode = igrab(&entry->vfs_inode);
4412                 if (inode) {
4413                         spin_unlock(&root->inode_lock);
4414                         if (atomic_read(&inode->i_count) > 1)
4415                                 d_prune_aliases(inode);
4416                         /*
4417                          * btrfs_drop_inode will have it removed from the inode
4418                          * cache when its usage count hits zero.
4419                          */
4420                         iput(inode);
4421                         cond_resched();
4422                         spin_lock(&root->inode_lock);
4423                         goto again;
4424                 }
4425
4426                 if (cond_resched_lock(&root->inode_lock))
4427                         goto again;
4428
4429                 node = rb_next(node);
4430         }
4431         spin_unlock(&root->inode_lock);
4432 }
4433
4434 int btrfs_delete_subvolume(struct inode *dir, struct dentry *dentry)
4435 {
4436         struct btrfs_fs_info *fs_info = btrfs_sb(dentry->d_sb);
4437         struct btrfs_root *root = BTRFS_I(dir)->root;
4438         struct inode *inode = d_inode(dentry);
4439         struct btrfs_root *dest = BTRFS_I(inode)->root;
4440         struct btrfs_trans_handle *trans;
4441         struct btrfs_block_rsv block_rsv;
4442         u64 root_flags;
4443         int ret;
4444         int err;
4445
4446         /*
4447          * Don't allow to delete a subvolume with send in progress. This is
4448          * inside the inode lock so the error handling that has to drop the bit
4449          * again is not run concurrently.
4450          */
4451         spin_lock(&dest->root_item_lock);
4452         if (dest->send_in_progress) {
4453                 spin_unlock(&dest->root_item_lock);
4454                 btrfs_warn(fs_info,
4455                            "attempt to delete subvolume %llu during send",
4456                            dest->root_key.objectid);
4457                 return -EPERM;
4458         }
4459         root_flags = btrfs_root_flags(&dest->root_item);
4460         btrfs_set_root_flags(&dest->root_item,
4461                              root_flags | BTRFS_ROOT_SUBVOL_DEAD);
4462         spin_unlock(&dest->root_item_lock);
4463
4464         down_write(&fs_info->subvol_sem);
4465
4466         err = may_destroy_subvol(dest);
4467         if (err)
4468                 goto out_up_write;
4469
4470         btrfs_init_block_rsv(&block_rsv, BTRFS_BLOCK_RSV_TEMP);
4471         /*
4472          * One for dir inode,
4473          * two for dir entries,
4474          * two for root ref/backref.
4475          */
4476         err = btrfs_subvolume_reserve_metadata(root, &block_rsv, 5, true);
4477         if (err)
4478                 goto out_up_write;
4479
4480         trans = btrfs_start_transaction(root, 0);
4481         if (IS_ERR(trans)) {
4482                 err = PTR_ERR(trans);
4483                 goto out_release;
4484         }
4485         trans->block_rsv = &block_rsv;
4486         trans->bytes_reserved = block_rsv.size;
4487
4488         btrfs_record_snapshot_destroy(trans, BTRFS_I(dir));
4489
4490         ret = btrfs_unlink_subvol(trans, dir, dest->root_key.objectid,
4491                                   dentry->d_name.name, dentry->d_name.len);
4492         if (ret) {
4493                 err = ret;
4494                 btrfs_abort_transaction(trans, ret);
4495                 goto out_end_trans;
4496         }
4497
4498         btrfs_record_root_in_trans(trans, dest);
4499
4500         memset(&dest->root_item.drop_progress, 0,
4501                 sizeof(dest->root_item.drop_progress));
4502         dest->root_item.drop_level = 0;
4503         btrfs_set_root_refs(&dest->root_item, 0);
4504
4505         if (!test_and_set_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &dest->state)) {
4506                 ret = btrfs_insert_orphan_item(trans,
4507                                         fs_info->tree_root,
4508                                         dest->root_key.objectid);
4509                 if (ret) {
4510                         btrfs_abort_transaction(trans, ret);
4511                         err = ret;
4512                         goto out_end_trans;
4513                 }
4514         }
4515
4516         ret = btrfs_uuid_tree_remove(trans, dest->root_item.uuid,
4517                                   BTRFS_UUID_KEY_SUBVOL,
4518                                   dest->root_key.objectid);
4519         if (ret && ret != -ENOENT) {
4520                 btrfs_abort_transaction(trans, ret);
4521                 err = ret;
4522                 goto out_end_trans;
4523         }
4524         if (!btrfs_is_empty_uuid(dest->root_item.received_uuid)) {
4525                 ret = btrfs_uuid_tree_remove(trans,
4526                                           dest->root_item.received_uuid,
4527                                           BTRFS_UUID_KEY_RECEIVED_SUBVOL,
4528                                           dest->root_key.objectid);
4529                 if (ret && ret != -ENOENT) {
4530                         btrfs_abort_transaction(trans, ret);
4531                         err = ret;
4532                         goto out_end_trans;
4533                 }
4534         }
4535
4536 out_end_trans:
4537         trans->block_rsv = NULL;
4538         trans->bytes_reserved = 0;
4539         ret = btrfs_end_transaction(trans);
4540         if (ret && !err)
4541                 err = ret;
4542         inode->i_flags |= S_DEAD;
4543 out_release:
4544         btrfs_subvolume_release_metadata(fs_info, &block_rsv);
4545 out_up_write:
4546         up_write(&fs_info->subvol_sem);
4547         if (err) {
4548                 spin_lock(&dest->root_item_lock);
4549                 root_flags = btrfs_root_flags(&dest->root_item);
4550                 btrfs_set_root_flags(&dest->root_item,
4551                                 root_flags & ~BTRFS_ROOT_SUBVOL_DEAD);
4552                 spin_unlock(&dest->root_item_lock);
4553         } else {
4554                 d_invalidate(dentry);
4555                 btrfs_prune_dentries(dest);
4556                 ASSERT(dest->send_in_progress == 0);
4557
4558                 /* the last ref */
4559                 if (dest->ino_cache_inode) {
4560                         iput(dest->ino_cache_inode);
4561                         dest->ino_cache_inode = NULL;
4562                 }
4563         }
4564
4565         return err;
4566 }
4567
4568 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4569 {
4570         struct inode *inode = d_inode(dentry);
4571         int err = 0;
4572         struct btrfs_root *root = BTRFS_I(dir)->root;
4573         struct btrfs_trans_handle *trans;
4574         u64 last_unlink_trans;
4575
4576         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4577                 return -ENOTEMPTY;
4578         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4579                 return btrfs_delete_subvolume(dir, dentry);
4580
4581         trans = __unlink_start_trans(dir);
4582         if (IS_ERR(trans))
4583                 return PTR_ERR(trans);
4584
4585         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4586                 err = btrfs_unlink_subvol(trans, dir,
4587                                           BTRFS_I(inode)->location.objectid,
4588                                           dentry->d_name.name,
4589                                           dentry->d_name.len);
4590                 goto out;
4591         }
4592
4593         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4594         if (err)
4595                 goto out;
4596
4597         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4598
4599         /* now the directory is empty */
4600         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4601                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4602                         dentry->d_name.len);
4603         if (!err) {
4604                 btrfs_i_size_write(BTRFS_I(inode), 0);
4605                 /*
4606                  * Propagate the last_unlink_trans value of the deleted dir to
4607                  * its parent directory. This is to prevent an unrecoverable
4608                  * log tree in the case we do something like this:
4609                  * 1) create dir foo
4610                  * 2) create snapshot under dir foo
4611                  * 3) delete the snapshot
4612                  * 4) rmdir foo
4613                  * 5) mkdir foo
4614                  * 6) fsync foo or some file inside foo
4615                  */
4616                 if (last_unlink_trans >= trans->transid)
4617                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4618         }
4619 out:
4620         btrfs_end_transaction(trans);
4621         btrfs_btree_balance_dirty(root->fs_info);
4622
4623         return err;
4624 }
4625
4626 /*
4627  * Return this if we need to call truncate_block for the last bit of the
4628  * truncate.
4629  */
4630 #define NEED_TRUNCATE_BLOCK 1
4631
4632 /*
4633  * this can truncate away extent items, csum items and directory items.
4634  * It starts at a high offset and removes keys until it can't find
4635  * any higher than new_size
4636  *
4637  * csum items that cross the new i_size are truncated to the new size
4638  * as well.
4639  *
4640  * min_type is the minimum key type to truncate down to.  If set to 0, this
4641  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4642  */
4643 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4644                                struct btrfs_root *root,
4645                                struct inode *inode,
4646                                u64 new_size, u32 min_type)
4647 {
4648         struct btrfs_fs_info *fs_info = root->fs_info;
4649         struct btrfs_path *path;
4650         struct extent_buffer *leaf;
4651         struct btrfs_file_extent_item *fi;
4652         struct btrfs_key key;
4653         struct btrfs_key found_key;
4654         u64 extent_start = 0;
4655         u64 extent_num_bytes = 0;
4656         u64 extent_offset = 0;
4657         u64 item_end = 0;
4658         u64 last_size = new_size;
4659         u32 found_type = (u8)-1;
4660         int found_extent;
4661         int del_item;
4662         int pending_del_nr = 0;
4663         int pending_del_slot = 0;
4664         int extent_type = -1;
4665         int ret;
4666         u64 ino = btrfs_ino(BTRFS_I(inode));
4667         u64 bytes_deleted = 0;
4668         bool be_nice = false;
4669         bool should_throttle = false;
4670
4671         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4672
4673         /*
4674          * for non-free space inodes and ref cows, we want to back off from
4675          * time to time
4676          */
4677         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4678             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4679                 be_nice = true;
4680
4681         path = btrfs_alloc_path();
4682         if (!path)
4683                 return -ENOMEM;
4684         path->reada = READA_BACK;
4685
4686         /*
4687          * We want to drop from the next block forward in case this new size is
4688          * not block aligned since we will be keeping the last block of the
4689          * extent just the way it is.
4690          */
4691         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4692             root == fs_info->tree_root)
4693                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4694                                         fs_info->sectorsize),
4695                                         (u64)-1, 0);
4696
4697         /*
4698          * This function is also used to drop the items in the log tree before
4699          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4700          * it is used to drop the logged items. So we shouldn't kill the delayed
4701          * items.
4702          */
4703         if (min_type == 0 && root == BTRFS_I(inode)->root)
4704                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4705
4706         key.objectid = ino;
4707         key.offset = (u64)-1;
4708         key.type = (u8)-1;
4709
4710 search_again:
4711         /*
4712          * with a 16K leaf size and 128MB extents, you can actually queue
4713          * up a huge file in a single leaf.  Most of the time that
4714          * bytes_deleted is > 0, it will be huge by the time we get here
4715          */
4716         if (be_nice && bytes_deleted > SZ_32M &&
4717             btrfs_should_end_transaction(trans)) {
4718                 ret = -EAGAIN;
4719                 goto out;
4720         }
4721
4722         path->leave_spinning = 1;
4723         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4724         if (ret < 0)
4725                 goto out;
4726
4727         if (ret > 0) {
4728                 ret = 0;
4729                 /* there are no items in the tree for us to truncate, we're
4730                  * done
4731                  */
4732                 if (path->slots[0] == 0)
4733                         goto out;
4734                 path->slots[0]--;
4735         }
4736
4737         while (1) {
4738                 fi = NULL;
4739                 leaf = path->nodes[0];
4740                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4741                 found_type = found_key.type;
4742
4743                 if (found_key.objectid != ino)
4744                         break;
4745
4746                 if (found_type < min_type)
4747                         break;
4748
4749                 item_end = found_key.offset;
4750                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4751                         fi = btrfs_item_ptr(leaf, path->slots[0],
4752                                             struct btrfs_file_extent_item);
4753                         extent_type = btrfs_file_extent_type(leaf, fi);
4754                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4755                                 item_end +=
4756                                     btrfs_file_extent_num_bytes(leaf, fi);
4757
4758                                 trace_btrfs_truncate_show_fi_regular(
4759                                         BTRFS_I(inode), leaf, fi,
4760                                         found_key.offset);
4761                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4762                                 item_end += btrfs_file_extent_ram_bytes(leaf,
4763                                                                         fi);
4764
4765                                 trace_btrfs_truncate_show_fi_inline(
4766                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4767                                         found_key.offset);
4768                         }
4769                         item_end--;
4770                 }
4771                 if (found_type > min_type) {
4772                         del_item = 1;
4773                 } else {
4774                         if (item_end < new_size)
4775                                 break;
4776                         if (found_key.offset >= new_size)
4777                                 del_item = 1;
4778                         else
4779                                 del_item = 0;
4780                 }
4781                 found_extent = 0;
4782                 /* FIXME, shrink the extent if the ref count is only 1 */
4783                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4784                         goto delete;
4785
4786                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4787                         u64 num_dec;
4788                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4789                         if (!del_item) {
4790                                 u64 orig_num_bytes =
4791                                         btrfs_file_extent_num_bytes(leaf, fi);
4792                                 extent_num_bytes = ALIGN(new_size -
4793                                                 found_key.offset,
4794                                                 fs_info->sectorsize);
4795                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4796                                                          extent_num_bytes);
4797                                 num_dec = (orig_num_bytes -
4798                                            extent_num_bytes);
4799                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4800                                              &root->state) &&
4801                                     extent_start != 0)
4802                                         inode_sub_bytes(inode, num_dec);
4803                                 btrfs_mark_buffer_dirty(leaf);
4804                         } else {
4805                                 extent_num_bytes =
4806                                         btrfs_file_extent_disk_num_bytes(leaf,
4807                                                                          fi);
4808                                 extent_offset = found_key.offset -
4809                                         btrfs_file_extent_offset(leaf, fi);
4810
4811                                 /* FIXME blocksize != 4096 */
4812                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4813                                 if (extent_start != 0) {
4814                                         found_extent = 1;
4815                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4816                                                      &root->state))
4817                                                 inode_sub_bytes(inode, num_dec);
4818                                 }
4819                         }
4820                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4821                         /*
4822                          * we can't truncate inline items that have had
4823                          * special encodings
4824                          */
4825                         if (!del_item &&
4826                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4827                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4828                             btrfs_file_extent_compression(leaf, fi) == 0) {
4829                                 u32 size = (u32)(new_size - found_key.offset);
4830
4831                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4832                                 size = btrfs_file_extent_calc_inline_size(size);
4833                                 btrfs_truncate_item(path, size, 1);
4834                         } else if (!del_item) {
4835                                 /*
4836                                  * We have to bail so the last_size is set to
4837                                  * just before this extent.
4838                                  */
4839                                 ret = NEED_TRUNCATE_BLOCK;
4840                                 break;
4841                         }
4842
4843                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4844                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4845                 }
4846 delete:
4847                 if (del_item)
4848                         last_size = found_key.offset;
4849                 else
4850                         last_size = new_size;
4851                 if (del_item) {
4852                         if (!pending_del_nr) {
4853                                 /* no pending yet, add ourselves */
4854                                 pending_del_slot = path->slots[0];
4855                                 pending_del_nr = 1;
4856                         } else if (pending_del_nr &&
4857                                    path->slots[0] + 1 == pending_del_slot) {
4858                                 /* hop on the pending chunk */
4859                                 pending_del_nr++;
4860                                 pending_del_slot = path->slots[0];
4861                         } else {
4862                                 BUG();
4863                         }
4864                 } else {
4865                         break;
4866                 }
4867                 should_throttle = false;
4868
4869                 if (found_extent &&
4870                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4871                      root == fs_info->tree_root)) {
4872                         struct btrfs_ref ref = { 0 };
4873
4874                         btrfs_set_path_blocking(path);
4875                         bytes_deleted += extent_num_bytes;
4876
4877                         btrfs_init_generic_ref(&ref, BTRFS_DROP_DELAYED_REF,
4878                                         extent_start, extent_num_bytes, 0);
4879                         ref.real_root = root->root_key.objectid;
4880                         btrfs_init_data_ref(&ref, btrfs_header_owner(leaf),
4881                                         ino, extent_offset);
4882                         ret = btrfs_free_extent(trans, &ref);
4883                         if (ret) {
4884                                 btrfs_abort_transaction(trans, ret);
4885                                 break;
4886                         }
4887                         if (be_nice) {
4888                                 if (btrfs_should_throttle_delayed_refs(trans))
4889                                         should_throttle = true;
4890                         }
4891                 }
4892
4893                 if (found_type == BTRFS_INODE_ITEM_KEY)
4894                         break;
4895
4896                 if (path->slots[0] == 0 ||
4897                     path->slots[0] != pending_del_slot ||
4898                     should_throttle) {
4899                         if (pending_del_nr) {
4900                                 ret = btrfs_del_items(trans, root, path,
4901                                                 pending_del_slot,
4902                                                 pending_del_nr);
4903                                 if (ret) {
4904                                         btrfs_abort_transaction(trans, ret);
4905                                         break;
4906                                 }
4907                                 pending_del_nr = 0;
4908                         }
4909                         btrfs_release_path(path);
4910
4911                         /*
4912                          * We can generate a lot of delayed refs, so we need to
4913                          * throttle every once and a while and make sure we're
4914                          * adding enough space to keep up with the work we are
4915                          * generating.  Since we hold a transaction here we
4916                          * can't flush, and we don't want to FLUSH_LIMIT because
4917                          * we could have generated too many delayed refs to
4918                          * actually allocate, so just bail if we're short and
4919                          * let the normal reservation dance happen higher up.
4920                          */
4921                         if (should_throttle) {
4922                                 ret = btrfs_delayed_refs_rsv_refill(fs_info,
4923                                                         BTRFS_RESERVE_NO_FLUSH);
4924                                 if (ret) {
4925                                         ret = -EAGAIN;
4926                                         break;
4927                                 }
4928                         }
4929                         goto search_again;
4930                 } else {
4931                         path->slots[0]--;
4932                 }
4933         }
4934 out:
4935         if (ret >= 0 && pending_del_nr) {
4936                 int err;
4937
4938                 err = btrfs_del_items(trans, root, path, pending_del_slot,
4939                                       pending_del_nr);
4940                 if (err) {
4941                         btrfs_abort_transaction(trans, err);
4942                         ret = err;
4943                 }
4944         }
4945         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4946                 ASSERT(last_size >= new_size);
4947                 if (!ret && last_size > new_size)
4948                         last_size = new_size;
4949                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4950         }
4951
4952         btrfs_free_path(path);
4953         return ret;
4954 }
4955
4956 /*
4957  * btrfs_truncate_block - read, zero a chunk and write a block
4958  * @inode - inode that we're zeroing
4959  * @from - the offset to start zeroing
4960  * @len - the length to zero, 0 to zero the entire range respective to the
4961  *      offset
4962  * @front - zero up to the offset instead of from the offset on
4963  *
4964  * This will find the block for the "from" offset and cow the block and zero the
4965  * part we want to zero.  This is used with truncate and hole punching.
4966  */
4967 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4968                         int front)
4969 {
4970         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4971         struct address_space *mapping = inode->i_mapping;
4972         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4973         struct btrfs_ordered_extent *ordered;
4974         struct extent_state *cached_state = NULL;
4975         struct extent_changeset *data_reserved = NULL;
4976         char *kaddr;
4977         u32 blocksize = fs_info->sectorsize;
4978         pgoff_t index = from >> PAGE_SHIFT;
4979         unsigned offset = from & (blocksize - 1);
4980         struct page *page;
4981         gfp_t mask = btrfs_alloc_write_mask(mapping);
4982         int ret = 0;
4983         u64 block_start;
4984         u64 block_end;
4985
4986         if (IS_ALIGNED(offset, blocksize) &&
4987             (!len || IS_ALIGNED(len, blocksize)))
4988                 goto out;
4989
4990         block_start = round_down(from, blocksize);
4991         block_end = block_start + blocksize - 1;
4992
4993         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4994                                            block_start, blocksize);
4995         if (ret)
4996                 goto out;
4997
4998 again:
4999         page = find_or_create_page(mapping, index, mask);
5000         if (!page) {
5001                 btrfs_delalloc_release_space(inode, data_reserved,
5002                                              block_start, blocksize, true);
5003                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5004                 ret = -ENOMEM;
5005                 goto out;
5006         }
5007
5008         if (!PageUptodate(page)) {
5009                 ret = btrfs_readpage(NULL, page);
5010                 lock_page(page);
5011                 if (page->mapping != mapping) {
5012                         unlock_page(page);
5013                         put_page(page);
5014                         goto again;
5015                 }
5016                 if (!PageUptodate(page)) {
5017                         ret = -EIO;
5018                         goto out_unlock;
5019                 }
5020         }
5021         wait_on_page_writeback(page);
5022
5023         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
5024         set_page_extent_mapped(page);
5025
5026         ordered = btrfs_lookup_ordered_extent(inode, block_start);
5027         if (ordered) {
5028                 unlock_extent_cached(io_tree, block_start, block_end,
5029                                      &cached_state);
5030                 unlock_page(page);
5031                 put_page(page);
5032                 btrfs_start_ordered_extent(inode, ordered, 1);
5033                 btrfs_put_ordered_extent(ordered);
5034                 goto again;
5035         }
5036
5037         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
5038                          EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
5039                          0, 0, &cached_state);
5040
5041         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
5042                                         &cached_state);
5043         if (ret) {
5044                 unlock_extent_cached(io_tree, block_start, block_end,
5045                                      &cached_state);
5046                 goto out_unlock;
5047         }
5048
5049         if (offset != blocksize) {
5050                 if (!len)
5051                         len = blocksize - offset;
5052                 kaddr = kmap(page);
5053                 if (front)
5054                         memset(kaddr + (block_start - page_offset(page)),
5055                                 0, offset);
5056                 else
5057                         memset(kaddr + (block_start - page_offset(page)) +  offset,
5058                                 0, len);
5059                 flush_dcache_page(page);
5060                 kunmap(page);
5061         }
5062         ClearPageChecked(page);
5063         set_page_dirty(page);
5064         unlock_extent_cached(io_tree, block_start, block_end, &cached_state);
5065
5066 out_unlock:
5067         if (ret)
5068                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
5069                                              blocksize, true);
5070         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
5071         unlock_page(page);
5072         put_page(page);
5073 out:
5074         extent_changeset_free(data_reserved);
5075         return ret;
5076 }
5077
5078 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
5079                              u64 offset, u64 len)
5080 {
5081         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5082         struct btrfs_trans_handle *trans;
5083         int ret;
5084
5085         /*
5086          * Still need to make sure the inode looks like it's been updated so
5087          * that any holes get logged if we fsync.
5088          */
5089         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
5090                 BTRFS_I(inode)->last_trans = fs_info->generation;
5091                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
5092                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
5093                 return 0;
5094         }
5095
5096         /*
5097          * 1 - for the one we're dropping
5098          * 1 - for the one we're adding
5099          * 1 - for updating the inode.
5100          */
5101         trans = btrfs_start_transaction(root, 3);
5102         if (IS_ERR(trans))
5103                 return PTR_ERR(trans);
5104
5105         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
5106         if (ret) {
5107                 btrfs_abort_transaction(trans, ret);
5108                 btrfs_end_transaction(trans);
5109                 return ret;
5110         }
5111
5112         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
5113                         offset, 0, 0, len, 0, len, 0, 0, 0);
5114         if (ret)
5115                 btrfs_abort_transaction(trans, ret);
5116         else
5117                 btrfs_update_inode(trans, root, inode);
5118         btrfs_end_transaction(trans);
5119         return ret;
5120 }
5121
5122 /*
5123  * This function puts in dummy file extents for the area we're creating a hole
5124  * for.  So if we are truncating this file to a larger size we need to insert
5125  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
5126  * the range between oldsize and size
5127  */
5128 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
5129 {
5130         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5131         struct btrfs_root *root = BTRFS_I(inode)->root;
5132         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5133         struct extent_map *em = NULL;
5134         struct extent_state *cached_state = NULL;
5135         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
5136         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
5137         u64 block_end = ALIGN(size, fs_info->sectorsize);
5138         u64 last_byte;
5139         u64 cur_offset;
5140         u64 hole_size;
5141         int err = 0;
5142
5143         /*
5144          * If our size started in the middle of a block we need to zero out the
5145          * rest of the block before we expand the i_size, otherwise we could
5146          * expose stale data.
5147          */
5148         err = btrfs_truncate_block(inode, oldsize, 0, 0);
5149         if (err)
5150                 return err;
5151
5152         if (size <= hole_start)
5153                 return 0;
5154
5155         btrfs_lock_and_flush_ordered_range(io_tree, BTRFS_I(inode), hole_start,
5156                                            block_end - 1, &cached_state);
5157         cur_offset = hole_start;
5158         while (1) {
5159                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
5160                                 block_end - cur_offset, 0);
5161                 if (IS_ERR(em)) {
5162                         err = PTR_ERR(em);
5163                         em = NULL;
5164                         break;
5165                 }
5166                 last_byte = min(extent_map_end(em), block_end);
5167                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
5168                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
5169                         struct extent_map *hole_em;
5170                         hole_size = last_byte - cur_offset;
5171
5172                         err = maybe_insert_hole(root, inode, cur_offset,
5173                                                 hole_size);
5174                         if (err)
5175                                 break;
5176                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
5177                                                 cur_offset + hole_size - 1, 0);
5178                         hole_em = alloc_extent_map();
5179                         if (!hole_em) {
5180                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
5181                                         &BTRFS_I(inode)->runtime_flags);
5182                                 goto next;
5183                         }
5184                         hole_em->start = cur_offset;
5185                         hole_em->len = hole_size;
5186                         hole_em->orig_start = cur_offset;
5187
5188                         hole_em->block_start = EXTENT_MAP_HOLE;
5189                         hole_em->block_len = 0;
5190                         hole_em->orig_block_len = 0;
5191                         hole_em->ram_bytes = hole_size;
5192                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
5193                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
5194                         hole_em->generation = fs_info->generation;
5195
5196                         while (1) {
5197                                 write_lock(&em_tree->lock);
5198                                 err = add_extent_mapping(em_tree, hole_em, 1);
5199                                 write_unlock(&em_tree->lock);
5200                                 if (err != -EEXIST)
5201                                         break;
5202                                 btrfs_drop_extent_cache(BTRFS_I(inode),
5203                                                         cur_offset,
5204                                                         cur_offset +
5205                                                         hole_size - 1, 0);
5206                         }
5207                         free_extent_map(hole_em);
5208                 }
5209 next:
5210                 free_extent_map(em);
5211                 em = NULL;
5212                 cur_offset = last_byte;
5213                 if (cur_offset >= block_end)
5214                         break;
5215         }
5216         free_extent_map(em);
5217         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state);
5218         return err;
5219 }
5220
5221 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5222 {
5223         struct btrfs_root *root = BTRFS_I(inode)->root;
5224         struct btrfs_trans_handle *trans;
5225         loff_t oldsize = i_size_read(inode);
5226         loff_t newsize = attr->ia_size;
5227         int mask = attr->ia_valid;
5228         int ret;
5229
5230         /*
5231          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5232          * special case where we need to update the times despite not having
5233          * these flags set.  For all other operations the VFS set these flags
5234          * explicitly if it wants a timestamp update.
5235          */
5236         if (newsize != oldsize) {
5237                 inode_inc_iversion(inode);
5238                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5239                         inode->i_ctime = inode->i_mtime =
5240                                 current_time(inode);
5241         }
5242
5243         if (newsize > oldsize) {
5244                 /*
5245                  * Don't do an expanding truncate while snapshotting is ongoing.
5246                  * This is to ensure the snapshot captures a fully consistent
5247                  * state of this file - if the snapshot captures this expanding
5248                  * truncation, it must capture all writes that happened before
5249                  * this truncation.
5250                  */
5251                 btrfs_wait_for_snapshot_creation(root);
5252                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5253                 if (ret) {
5254                         btrfs_end_write_no_snapshotting(root);
5255                         return ret;
5256                 }
5257
5258                 trans = btrfs_start_transaction(root, 1);
5259                 if (IS_ERR(trans)) {
5260                         btrfs_end_write_no_snapshotting(root);
5261                         return PTR_ERR(trans);
5262                 }
5263
5264                 i_size_write(inode, newsize);
5265                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5266                 pagecache_isize_extended(inode, oldsize, newsize);
5267                 ret = btrfs_update_inode(trans, root, inode);
5268                 btrfs_end_write_no_snapshotting(root);
5269                 btrfs_end_transaction(trans);
5270         } else {
5271
5272                 /*
5273                  * We're truncating a file that used to have good data down to
5274                  * zero. Make sure it gets into the ordered flush list so that
5275                  * any new writes get down to disk quickly.
5276                  */
5277                 if (newsize == 0)
5278                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5279                                 &BTRFS_I(inode)->runtime_flags);
5280
5281                 truncate_setsize(inode, newsize);
5282
5283                 /* Disable nonlocked read DIO to avoid the endless truncate */
5284                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5285                 inode_dio_wait(inode);
5286                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5287
5288                 ret = btrfs_truncate(inode, newsize == oldsize);
5289                 if (ret && inode->i_nlink) {
5290                         int err;
5291
5292                         /*
5293                          * Truncate failed, so fix up the in-memory size. We
5294                          * adjusted disk_i_size down as we removed extents, so
5295                          * wait for disk_i_size to be stable and then update the
5296                          * in-memory size to match.
5297                          */
5298                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5299                         if (err)
5300                                 return err;
5301                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5302                 }
5303         }
5304
5305         return ret;
5306 }
5307
5308 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5309 {
5310         struct inode *inode = d_inode(dentry);
5311         struct btrfs_root *root = BTRFS_I(inode)->root;
5312         int err;
5313
5314         if (btrfs_root_readonly(root))
5315                 return -EROFS;
5316
5317         err = setattr_prepare(dentry, attr);
5318         if (err)
5319                 return err;
5320
5321         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5322                 err = btrfs_setsize(inode, attr);
5323                 if (err)
5324                         return err;
5325         }
5326
5327         if (attr->ia_valid) {
5328                 setattr_copy(inode, attr);
5329                 inode_inc_iversion(inode);
5330                 err = btrfs_dirty_inode(inode);
5331
5332                 if (!err && attr->ia_valid & ATTR_MODE)
5333                         err = posix_acl_chmod(inode, inode->i_mode);
5334         }
5335
5336         return err;
5337 }
5338
5339 /*
5340  * While truncating the inode pages during eviction, we get the VFS calling
5341  * btrfs_invalidatepage() against each page of the inode. This is slow because
5342  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5343  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5344  * extent_state structures over and over, wasting lots of time.
5345  *
5346  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5347  * those expensive operations on a per page basis and do only the ordered io
5348  * finishing, while we release here the extent_map and extent_state structures,
5349  * without the excessive merging and splitting.
5350  */
5351 static void evict_inode_truncate_pages(struct inode *inode)
5352 {
5353         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5354         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5355         struct rb_node *node;
5356
5357         ASSERT(inode->i_state & I_FREEING);
5358         truncate_inode_pages_final(&inode->i_data);
5359
5360         write_lock(&map_tree->lock);
5361         while (!RB_EMPTY_ROOT(&map_tree->map.rb_root)) {
5362                 struct extent_map *em;
5363
5364                 node = rb_first_cached(&map_tree->map);
5365                 em = rb_entry(node, struct extent_map, rb_node);
5366                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5367                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5368                 remove_extent_mapping(map_tree, em);
5369                 free_extent_map(em);
5370                 if (need_resched()) {
5371                         write_unlock(&map_tree->lock);
5372                         cond_resched();
5373                         write_lock(&map_tree->lock);
5374                 }
5375         }
5376         write_unlock(&map_tree->lock);
5377
5378         /*
5379          * Keep looping until we have no more ranges in the io tree.
5380          * We can have ongoing bios started by readpages (called from readahead)
5381          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5382          * still in progress (unlocked the pages in the bio but did not yet
5383          * unlocked the ranges in the io tree). Therefore this means some
5384          * ranges can still be locked and eviction started because before
5385          * submitting those bios, which are executed by a separate task (work
5386          * queue kthread), inode references (inode->i_count) were not taken
5387          * (which would be dropped in the end io callback of each bio).
5388          * Therefore here we effectively end up waiting for those bios and
5389          * anyone else holding locked ranges without having bumped the inode's
5390          * reference count - if we don't do it, when they access the inode's
5391          * io_tree to unlock a range it may be too late, leading to an
5392          * use-after-free issue.
5393          */
5394         spin_lock(&io_tree->lock);
5395         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5396                 struct extent_state *state;
5397                 struct extent_state *cached_state = NULL;
5398                 u64 start;
5399                 u64 end;
5400                 unsigned state_flags;
5401
5402                 node = rb_first(&io_tree->state);
5403                 state = rb_entry(node, struct extent_state, rb_node);
5404                 start = state->start;
5405                 end = state->end;
5406                 state_flags = state->state;
5407                 spin_unlock(&io_tree->lock);
5408
5409                 lock_extent_bits(io_tree, start, end, &cached_state);
5410
5411                 /*
5412                  * If still has DELALLOC flag, the extent didn't reach disk,
5413                  * and its reserved space won't be freed by delayed_ref.
5414                  * So we need to free its reserved space here.
5415                  * (Refer to comment in btrfs_invalidatepage, case 2)
5416                  *
5417                  * Note, end is the bytenr of last byte, so we need + 1 here.
5418                  */
5419                 if (state_flags & EXTENT_DELALLOC)
5420                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5421
5422                 clear_extent_bit(io_tree, start, end,
5423                                  EXTENT_LOCKED | EXTENT_DELALLOC |
5424                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
5425                                  &cached_state);
5426
5427                 cond_resched();
5428                 spin_lock(&io_tree->lock);
5429         }
5430         spin_unlock(&io_tree->lock);
5431 }
5432
5433 static struct btrfs_trans_handle *evict_refill_and_join(struct btrfs_root *root,
5434                                                         struct btrfs_block_rsv *rsv)
5435 {
5436         struct btrfs_fs_info *fs_info = root->fs_info;
5437         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5438         struct btrfs_trans_handle *trans;
5439         u64 delayed_refs_extra = btrfs_calc_insert_metadata_size(fs_info, 1);
5440         int ret;
5441
5442         /*
5443          * Eviction should be taking place at some place safe because of our
5444          * delayed iputs.  However the normal flushing code will run delayed
5445          * iputs, so we cannot use FLUSH_ALL otherwise we'll deadlock.
5446          *
5447          * We reserve the delayed_refs_extra here again because we can't use
5448          * btrfs_start_transaction(root, 0) for the same deadlocky reason as
5449          * above.  We reserve our extra bit here because we generate a ton of
5450          * delayed refs activity by truncating.
5451          *
5452          * If we cannot make our reservation we'll attempt to steal from the
5453          * global reserve, because we really want to be able to free up space.
5454          */
5455         ret = btrfs_block_rsv_refill(root, rsv, rsv->size + delayed_refs_extra,
5456                                      BTRFS_RESERVE_FLUSH_EVICT);
5457         if (ret) {
5458                 /*
5459                  * Try to steal from the global reserve if there is space for
5460                  * it.
5461                  */
5462                 if (btrfs_check_space_for_delayed_refs(fs_info) ||
5463                     btrfs_block_rsv_migrate(global_rsv, rsv, rsv->size, 0)) {
5464                         btrfs_warn(fs_info,
5465                                    "could not allocate space for delete; will truncate on mount");
5466                         return ERR_PTR(-ENOSPC);
5467                 }
5468                 delayed_refs_extra = 0;
5469         }
5470
5471         trans = btrfs_join_transaction(root);
5472         if (IS_ERR(trans))
5473                 return trans;
5474
5475         if (delayed_refs_extra) {
5476                 trans->block_rsv = &fs_info->trans_block_rsv;
5477                 trans->bytes_reserved = delayed_refs_extra;
5478                 btrfs_block_rsv_migrate(rsv, trans->block_rsv,
5479                                         delayed_refs_extra, 1);
5480         }
5481         return trans;
5482 }
5483
5484 void btrfs_evict_inode(struct inode *inode)
5485 {
5486         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5487         struct btrfs_trans_handle *trans;
5488         struct btrfs_root *root = BTRFS_I(inode)->root;
5489         struct btrfs_block_rsv *rsv;
5490         int ret;
5491
5492         trace_btrfs_inode_evict(inode);
5493
5494         if (!root) {
5495                 clear_inode(inode);
5496                 return;
5497         }
5498
5499         evict_inode_truncate_pages(inode);
5500
5501         if (inode->i_nlink &&
5502             ((btrfs_root_refs(&root->root_item) != 0 &&
5503               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5504              btrfs_is_free_space_inode(BTRFS_I(inode))))
5505                 goto no_delete;
5506
5507         if (is_bad_inode(inode))
5508                 goto no_delete;
5509
5510         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5511
5512         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags))
5513                 goto no_delete;
5514
5515         if (inode->i_nlink > 0) {
5516                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5517                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5518                 goto no_delete;
5519         }
5520
5521         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5522         if (ret)
5523                 goto no_delete;
5524
5525         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5526         if (!rsv)
5527                 goto no_delete;
5528         rsv->size = btrfs_calc_metadata_size(fs_info, 1);
5529         rsv->failfast = 1;
5530
5531         btrfs_i_size_write(BTRFS_I(inode), 0);
5532
5533         while (1) {
5534                 trans = evict_refill_and_join(root, rsv);
5535                 if (IS_ERR(trans))
5536                         goto free_rsv;
5537
5538                 trans->block_rsv = rsv;
5539
5540                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5541                 trans->block_rsv = &fs_info->trans_block_rsv;
5542                 btrfs_end_transaction(trans);
5543                 btrfs_btree_balance_dirty(fs_info);
5544                 if (ret && ret != -ENOSPC && ret != -EAGAIN)
5545                         goto free_rsv;
5546                 else if (!ret)
5547                         break;
5548         }
5549
5550         /*
5551          * Errors here aren't a big deal, it just means we leave orphan items in
5552          * the tree. They will be cleaned up on the next mount. If the inode
5553          * number gets reused, cleanup deletes the orphan item without doing
5554          * anything, and unlink reuses the existing orphan item.
5555          *
5556          * If it turns out that we are dropping too many of these, we might want
5557          * to add a mechanism for retrying these after a commit.
5558          */
5559         trans = evict_refill_and_join(root, rsv);
5560         if (!IS_ERR(trans)) {
5561                 trans->block_rsv = rsv;
5562                 btrfs_orphan_del(trans, BTRFS_I(inode));
5563                 trans->block_rsv = &fs_info->trans_block_rsv;
5564                 btrfs_end_transaction(trans);
5565         }
5566
5567         if (!(root == fs_info->tree_root ||
5568               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5569                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5570
5571 free_rsv:
5572         btrfs_free_block_rsv(fs_info, rsv);
5573 no_delete:
5574         /*
5575          * If we didn't successfully delete, the orphan item will still be in
5576          * the tree and we'll retry on the next mount. Again, we might also want
5577          * to retry these periodically in the future.
5578          */
5579         btrfs_remove_delayed_node(BTRFS_I(inode));
5580         clear_inode(inode);
5581 }
5582
5583 /*
5584  * Return the key found in the dir entry in the location pointer, fill @type
5585  * with BTRFS_FT_*, and return 0.
5586  *
5587  * If no dir entries were found, returns -ENOENT.
5588  * If found a corrupted location in dir entry, returns -EUCLEAN.
5589  */
5590 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5591                                struct btrfs_key *location, u8 *type)
5592 {
5593         const char *name = dentry->d_name.name;
5594         int namelen = dentry->d_name.len;
5595         struct btrfs_dir_item *di;
5596         struct btrfs_path *path;
5597         struct btrfs_root *root = BTRFS_I(dir)->root;
5598         int ret = 0;
5599
5600         path = btrfs_alloc_path();
5601         if (!path)
5602                 return -ENOMEM;
5603
5604         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5605                         name, namelen, 0);
5606         if (IS_ERR_OR_NULL(di)) {
5607                 ret = di ? PTR_ERR(di) : -ENOENT;
5608                 goto out;
5609         }
5610
5611         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5612         if (location->type != BTRFS_INODE_ITEM_KEY &&
5613             location->type != BTRFS_ROOT_ITEM_KEY) {
5614                 ret = -EUCLEAN;
5615                 btrfs_warn(root->fs_info,
5616 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5617                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5618                            location->objectid, location->type, location->offset);
5619         }
5620         if (!ret)
5621                 *type = btrfs_dir_type(path->nodes[0], di);
5622 out:
5623         btrfs_free_path(path);
5624         return ret;
5625 }
5626
5627 /*
5628  * when we hit a tree root in a directory, the btrfs part of the inode
5629  * needs to be changed to reflect the root directory of the tree root.  This
5630  * is kind of like crossing a mount point.
5631  */
5632 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5633                                     struct inode *dir,
5634                                     struct dentry *dentry,
5635                                     struct btrfs_key *location,
5636                                     struct btrfs_root **sub_root)
5637 {
5638         struct btrfs_path *path;
5639         struct btrfs_root *new_root;
5640         struct btrfs_root_ref *ref;
5641         struct extent_buffer *leaf;
5642         struct btrfs_key key;
5643         int ret;
5644         int err = 0;
5645
5646         path = btrfs_alloc_path();
5647         if (!path) {
5648                 err = -ENOMEM;
5649                 goto out;
5650         }
5651
5652         err = -ENOENT;
5653         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5654         key.type = BTRFS_ROOT_REF_KEY;
5655         key.offset = location->objectid;
5656
5657         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5658         if (ret) {
5659                 if (ret < 0)
5660                         err = ret;
5661                 goto out;
5662         }
5663
5664         leaf = path->nodes[0];
5665         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5666         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5667             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5668                 goto out;
5669
5670         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5671                                    (unsigned long)(ref + 1),
5672                                    dentry->d_name.len);
5673         if (ret)
5674                 goto out;
5675
5676         btrfs_release_path(path);
5677
5678         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5679         if (IS_ERR(new_root)) {
5680                 err = PTR_ERR(new_root);
5681                 goto out;
5682         }
5683
5684         *sub_root = new_root;
5685         location->objectid = btrfs_root_dirid(&new_root->root_item);
5686         location->type = BTRFS_INODE_ITEM_KEY;
5687         location->offset = 0;
5688         err = 0;
5689 out:
5690         btrfs_free_path(path);
5691         return err;
5692 }
5693
5694 static void inode_tree_add(struct inode *inode)
5695 {
5696         struct btrfs_root *root = BTRFS_I(inode)->root;
5697         struct btrfs_inode *entry;
5698         struct rb_node **p;
5699         struct rb_node *parent;
5700         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5701         u64 ino = btrfs_ino(BTRFS_I(inode));
5702
5703         if (inode_unhashed(inode))
5704                 return;
5705         parent = NULL;
5706         spin_lock(&root->inode_lock);
5707         p = &root->inode_tree.rb_node;
5708         while (*p) {
5709                 parent = *p;
5710                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5711
5712                 if (ino < btrfs_ino(entry))
5713                         p = &parent->rb_left;
5714                 else if (ino > btrfs_ino(entry))
5715                         p = &parent->rb_right;
5716                 else {
5717                         WARN_ON(!(entry->vfs_inode.i_state &
5718                                   (I_WILL_FREE | I_FREEING)));
5719                         rb_replace_node(parent, new, &root->inode_tree);
5720                         RB_CLEAR_NODE(parent);
5721                         spin_unlock(&root->inode_lock);
5722                         return;
5723                 }
5724         }
5725         rb_link_node(new, parent, p);
5726         rb_insert_color(new, &root->inode_tree);
5727         spin_unlock(&root->inode_lock);
5728 }
5729
5730 static void inode_tree_del(struct inode *inode)
5731 {
5732         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5733         struct btrfs_root *root = BTRFS_I(inode)->root;
5734         int empty = 0;
5735
5736         spin_lock(&root->inode_lock);
5737         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5738                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5739                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5740                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5741         }
5742         spin_unlock(&root->inode_lock);
5743
5744         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5745                 synchronize_srcu(&fs_info->subvol_srcu);
5746                 spin_lock(&root->inode_lock);
5747                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5748                 spin_unlock(&root->inode_lock);
5749                 if (empty)
5750                         btrfs_add_dead_root(root);
5751         }
5752 }
5753
5754
5755 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5756 {
5757         struct btrfs_iget_args *args = p;
5758         inode->i_ino = args->location->objectid;
5759         memcpy(&BTRFS_I(inode)->location, args->location,
5760                sizeof(*args->location));
5761         BTRFS_I(inode)->root = args->root;
5762         return 0;
5763 }
5764
5765 static int btrfs_find_actor(struct inode *inode, void *opaque)
5766 {
5767         struct btrfs_iget_args *args = opaque;
5768         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5769                 args->root == BTRFS_I(inode)->root;
5770 }
5771
5772 static struct inode *btrfs_iget_locked(struct super_block *s,
5773                                        struct btrfs_key *location,
5774                                        struct btrfs_root *root)
5775 {
5776         struct inode *inode;
5777         struct btrfs_iget_args args;
5778         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5779
5780         args.location = location;
5781         args.root = root;
5782
5783         inode = iget5_locked(s, hashval, btrfs_find_actor,
5784                              btrfs_init_locked_inode,
5785                              (void *)&args);
5786         return inode;
5787 }
5788
5789 /*
5790  * Get an inode object given its location and corresponding root.
5791  * Path can be preallocated to prevent recursing back to iget through
5792  * allocator. NULL is also valid but may require an additional allocation
5793  * later.
5794  */
5795 struct inode *btrfs_iget_path(struct super_block *s, struct btrfs_key *location,
5796                               struct btrfs_root *root, struct btrfs_path *path)
5797 {
5798         struct inode *inode;
5799
5800         inode = btrfs_iget_locked(s, location, root);
5801         if (!inode)
5802                 return ERR_PTR(-ENOMEM);
5803
5804         if (inode->i_state & I_NEW) {
5805                 int ret;
5806
5807                 ret = btrfs_read_locked_inode(inode, path);
5808                 if (!ret) {
5809                         inode_tree_add(inode);
5810                         unlock_new_inode(inode);
5811                 } else {
5812                         iget_failed(inode);
5813                         /*
5814                          * ret > 0 can come from btrfs_search_slot called by
5815                          * btrfs_read_locked_inode, this means the inode item
5816                          * was not found.
5817                          */
5818                         if (ret > 0)
5819                                 ret = -ENOENT;
5820                         inode = ERR_PTR(ret);
5821                 }
5822         }
5823
5824         return inode;
5825 }
5826
5827 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5828                          struct btrfs_root *root)
5829 {
5830         return btrfs_iget_path(s, location, root, NULL);
5831 }
5832
5833 static struct inode *new_simple_dir(struct super_block *s,
5834                                     struct btrfs_key *key,
5835                                     struct btrfs_root *root)
5836 {
5837         struct inode *inode = new_inode(s);
5838
5839         if (!inode)
5840                 return ERR_PTR(-ENOMEM);
5841
5842         BTRFS_I(inode)->root = root;
5843         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5844         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5845
5846         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5847         inode->i_op = &btrfs_dir_ro_inode_operations;
5848         inode->i_opflags &= ~IOP_XATTR;
5849         inode->i_fop = &simple_dir_operations;
5850         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5851         inode->i_mtime = current_time(inode);
5852         inode->i_atime = inode->i_mtime;
5853         inode->i_ctime = inode->i_mtime;
5854         BTRFS_I(inode)->i_otime = inode->i_mtime;
5855
5856         return inode;
5857 }
5858
5859 static inline u8 btrfs_inode_type(struct inode *inode)
5860 {
5861         /*
5862          * Compile-time asserts that generic FT_* types still match
5863          * BTRFS_FT_* types
5864          */
5865         BUILD_BUG_ON(BTRFS_FT_UNKNOWN != FT_UNKNOWN);
5866         BUILD_BUG_ON(BTRFS_FT_REG_FILE != FT_REG_FILE);
5867         BUILD_BUG_ON(BTRFS_FT_DIR != FT_DIR);
5868         BUILD_BUG_ON(BTRFS_FT_CHRDEV != FT_CHRDEV);
5869         BUILD_BUG_ON(BTRFS_FT_BLKDEV != FT_BLKDEV);
5870         BUILD_BUG_ON(BTRFS_FT_FIFO != FT_FIFO);
5871         BUILD_BUG_ON(BTRFS_FT_SOCK != FT_SOCK);
5872         BUILD_BUG_ON(BTRFS_FT_SYMLINK != FT_SYMLINK);
5873
5874         return fs_umode_to_ftype(inode->i_mode);
5875 }
5876
5877 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5878 {
5879         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5880         struct inode *inode;
5881         struct btrfs_root *root = BTRFS_I(dir)->root;
5882         struct btrfs_root *sub_root = root;
5883         struct btrfs_key location;
5884         u8 di_type = 0;
5885         int index;
5886         int ret = 0;
5887
5888         if (dentry->d_name.len > BTRFS_NAME_LEN)
5889                 return ERR_PTR(-ENAMETOOLONG);
5890
5891         ret = btrfs_inode_by_name(dir, dentry, &location, &di_type);
5892         if (ret < 0)
5893                 return ERR_PTR(ret);
5894
5895         if (location.type == BTRFS_INODE_ITEM_KEY) {
5896                 inode = btrfs_iget(dir->i_sb, &location, root);
5897                 if (IS_ERR(inode))
5898                         return inode;
5899
5900                 /* Do extra check against inode mode with di_type */
5901                 if (btrfs_inode_type(inode) != di_type) {
5902                         btrfs_crit(fs_info,
5903 "inode mode mismatch with dir: inode mode=0%o btrfs type=%u dir type=%u",
5904                                   inode->i_mode, btrfs_inode_type(inode),
5905                                   di_type);
5906                         iput(inode);
5907                         return ERR_PTR(-EUCLEAN);
5908                 }
5909                 return inode;
5910         }
5911
5912         index = srcu_read_lock(&fs_info->subvol_srcu);
5913         ret = fixup_tree_root_location(fs_info, dir, dentry,
5914                                        &location, &sub_root);
5915         if (ret < 0) {
5916                 if (ret != -ENOENT)
5917                         inode = ERR_PTR(ret);
5918                 else
5919                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5920         } else {
5921                 inode = btrfs_iget(dir->i_sb, &location, sub_root);
5922         }
5923         srcu_read_unlock(&fs_info->subvol_srcu, index);
5924
5925         if (!IS_ERR(inode) && root != sub_root) {
5926                 down_read(&fs_info->cleanup_work_sem);
5927                 if (!sb_rdonly(inode->i_sb))
5928                         ret = btrfs_orphan_cleanup(sub_root);
5929                 up_read(&fs_info->cleanup_work_sem);
5930                 if (ret) {
5931                         iput(inode);
5932                         inode = ERR_PTR(ret);
5933                 }
5934         }
5935
5936         return inode;
5937 }
5938
5939 static int btrfs_dentry_delete(const struct dentry *dentry)
5940 {
5941         struct btrfs_root *root;
5942         struct inode *inode = d_inode(dentry);
5943
5944         if (!inode && !IS_ROOT(dentry))
5945                 inode = d_inode(dentry->d_parent);
5946
5947         if (inode) {
5948                 root = BTRFS_I(inode)->root;
5949                 if (btrfs_root_refs(&root->root_item) == 0)
5950                         return 1;
5951
5952                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5953                         return 1;
5954         }
5955         return 0;
5956 }
5957
5958 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5959                                    unsigned int flags)
5960 {
5961         struct inode *inode = btrfs_lookup_dentry(dir, dentry);
5962
5963         if (inode == ERR_PTR(-ENOENT))
5964                 inode = NULL;
5965         return d_splice_alias(inode, dentry);
5966 }
5967
5968 /*
5969  * All this infrastructure exists because dir_emit can fault, and we are holding
5970  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5971  * our information into that, and then dir_emit from the buffer.  This is
5972  * similar to what NFS does, only we don't keep the buffer around in pagecache
5973  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5974  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5975  * tree lock.
5976  */
5977 static int btrfs_opendir(struct inode *inode, struct file *file)
5978 {
5979         struct btrfs_file_private *private;
5980
5981         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5982         if (!private)
5983                 return -ENOMEM;
5984         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5985         if (!private->filldir_buf) {
5986                 kfree(private);
5987                 return -ENOMEM;
5988         }
5989         file->private_data = private;
5990         return 0;
5991 }
5992
5993 struct dir_entry {
5994         u64 ino;
5995         u64 offset;
5996         unsigned type;
5997         int name_len;
5998 };
5999
6000 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
6001 {
6002         while (entries--) {
6003                 struct dir_entry *entry = addr;
6004                 char *name = (char *)(entry + 1);
6005
6006                 ctx->pos = get_unaligned(&entry->offset);
6007                 if (!dir_emit(ctx, name, get_unaligned(&entry->name_len),
6008                                          get_unaligned(&entry->ino),
6009                                          get_unaligned(&entry->type)))
6010                         return 1;
6011                 addr += sizeof(struct dir_entry) +
6012                         get_unaligned(&entry->name_len);
6013                 ctx->pos++;
6014         }
6015         return 0;
6016 }
6017
6018 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
6019 {
6020         struct inode *inode = file_inode(file);
6021         struct btrfs_root *root = BTRFS_I(inode)->root;
6022         struct btrfs_file_private *private = file->private_data;
6023         struct btrfs_dir_item *di;
6024         struct btrfs_key key;
6025         struct btrfs_key found_key;
6026         struct btrfs_path *path;
6027         void *addr;
6028         struct list_head ins_list;
6029         struct list_head del_list;
6030         int ret;
6031         struct extent_buffer *leaf;
6032         int slot;
6033         char *name_ptr;
6034         int name_len;
6035         int entries = 0;
6036         int total_len = 0;
6037         bool put = false;
6038         struct btrfs_key location;
6039
6040         if (!dir_emit_dots(file, ctx))
6041                 return 0;
6042
6043         path = btrfs_alloc_path();
6044         if (!path)
6045                 return -ENOMEM;
6046
6047         addr = private->filldir_buf;
6048         path->reada = READA_FORWARD;
6049
6050         INIT_LIST_HEAD(&ins_list);
6051         INIT_LIST_HEAD(&del_list);
6052         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
6053
6054 again:
6055         key.type = BTRFS_DIR_INDEX_KEY;
6056         key.offset = ctx->pos;
6057         key.objectid = btrfs_ino(BTRFS_I(inode));
6058
6059         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6060         if (ret < 0)
6061                 goto err;
6062
6063         while (1) {
6064                 struct dir_entry *entry;
6065
6066                 leaf = path->nodes[0];
6067                 slot = path->slots[0];
6068                 if (slot >= btrfs_header_nritems(leaf)) {
6069                         ret = btrfs_next_leaf(root, path);
6070                         if (ret < 0)
6071                                 goto err;
6072                         else if (ret > 0)
6073                                 break;
6074                         continue;
6075                 }
6076
6077                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6078
6079                 if (found_key.objectid != key.objectid)
6080                         break;
6081                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
6082                         break;
6083                 if (found_key.offset < ctx->pos)
6084                         goto next;
6085                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
6086                         goto next;
6087                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
6088                 name_len = btrfs_dir_name_len(leaf, di);
6089                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
6090                     PAGE_SIZE) {
6091                         btrfs_release_path(path);
6092                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6093                         if (ret)
6094                                 goto nopos;
6095                         addr = private->filldir_buf;
6096                         entries = 0;
6097                         total_len = 0;
6098                         goto again;
6099                 }
6100
6101                 entry = addr;
6102                 put_unaligned(name_len, &entry->name_len);
6103                 name_ptr = (char *)(entry + 1);
6104                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
6105                                    name_len);
6106                 put_unaligned(fs_ftype_to_dtype(btrfs_dir_type(leaf, di)),
6107                                 &entry->type);
6108                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
6109                 put_unaligned(location.objectid, &entry->ino);
6110                 put_unaligned(found_key.offset, &entry->offset);
6111                 entries++;
6112                 addr += sizeof(struct dir_entry) + name_len;
6113                 total_len += sizeof(struct dir_entry) + name_len;
6114 next:
6115                 path->slots[0]++;
6116         }
6117         btrfs_release_path(path);
6118
6119         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6120         if (ret)
6121                 goto nopos;
6122
6123         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6124         if (ret)
6125                 goto nopos;
6126
6127         /*
6128          * Stop new entries from being returned after we return the last
6129          * entry.
6130          *
6131          * New directory entries are assigned a strictly increasing
6132          * offset.  This means that new entries created during readdir
6133          * are *guaranteed* to be seen in the future by that readdir.
6134          * This has broken buggy programs which operate on names as
6135          * they're returned by readdir.  Until we re-use freed offsets
6136          * we have this hack to stop new entries from being returned
6137          * under the assumption that they'll never reach this huge
6138          * offset.
6139          *
6140          * This is being careful not to overflow 32bit loff_t unless the
6141          * last entry requires it because doing so has broken 32bit apps
6142          * in the past.
6143          */
6144         if (ctx->pos >= INT_MAX)
6145                 ctx->pos = LLONG_MAX;
6146         else
6147                 ctx->pos = INT_MAX;
6148 nopos:
6149         ret = 0;
6150 err:
6151         if (put)
6152                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6153         btrfs_free_path(path);
6154         return ret;
6155 }
6156
6157 /*
6158  * This is somewhat expensive, updating the tree every time the
6159  * inode changes.  But, it is most likely to find the inode in cache.
6160  * FIXME, needs more benchmarking...there are no reasons other than performance
6161  * to keep or drop this code.
6162  */
6163 static int btrfs_dirty_inode(struct inode *inode)
6164 {
6165         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6166         struct btrfs_root *root = BTRFS_I(inode)->root;
6167         struct btrfs_trans_handle *trans;
6168         int ret;
6169
6170         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6171                 return 0;
6172
6173         trans = btrfs_join_transaction(root);
6174         if (IS_ERR(trans))
6175                 return PTR_ERR(trans);
6176
6177         ret = btrfs_update_inode(trans, root, inode);
6178         if (ret && ret == -ENOSPC) {
6179                 /* whoops, lets try again with the full transaction */
6180                 btrfs_end_transaction(trans);
6181                 trans = btrfs_start_transaction(root, 1);
6182                 if (IS_ERR(trans))
6183                         return PTR_ERR(trans);
6184
6185                 ret = btrfs_update_inode(trans, root, inode);
6186         }
6187         btrfs_end_transaction(trans);
6188         if (BTRFS_I(inode)->delayed_node)
6189                 btrfs_balance_delayed_items(fs_info);
6190
6191         return ret;
6192 }
6193
6194 /*
6195  * This is a copy of file_update_time.  We need this so we can return error on
6196  * ENOSPC for updating the inode in the case of file write and mmap writes.
6197  */
6198 static int btrfs_update_time(struct inode *inode, struct timespec64 *now,
6199                              int flags)
6200 {
6201         struct btrfs_root *root = BTRFS_I(inode)->root;
6202         bool dirty = flags & ~S_VERSION;
6203
6204         if (btrfs_root_readonly(root))
6205                 return -EROFS;
6206
6207         if (flags & S_VERSION)
6208                 dirty |= inode_maybe_inc_iversion(inode, dirty);
6209         if (flags & S_CTIME)
6210                 inode->i_ctime = *now;
6211         if (flags & S_MTIME)
6212                 inode->i_mtime = *now;
6213         if (flags & S_ATIME)
6214                 inode->i_atime = *now;
6215         return dirty ? btrfs_dirty_inode(inode) : 0;
6216 }
6217
6218 /*
6219  * find the highest existing sequence number in a directory
6220  * and then set the in-memory index_cnt variable to reflect
6221  * free sequence numbers
6222  */
6223 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6224 {
6225         struct btrfs_root *root = inode->root;
6226         struct btrfs_key key, found_key;
6227         struct btrfs_path *path;
6228         struct extent_buffer *leaf;
6229         int ret;
6230
6231         key.objectid = btrfs_ino(inode);
6232         key.type = BTRFS_DIR_INDEX_KEY;
6233         key.offset = (u64)-1;
6234
6235         path = btrfs_alloc_path();
6236         if (!path)
6237                 return -ENOMEM;
6238
6239         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6240         if (ret < 0)
6241                 goto out;
6242         /* FIXME: we should be able to handle this */
6243         if (ret == 0)
6244                 goto out;
6245         ret = 0;
6246
6247         /*
6248          * MAGIC NUMBER EXPLANATION:
6249          * since we search a directory based on f_pos we have to start at 2
6250          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6251          * else has to start at 2
6252          */
6253         if (path->slots[0] == 0) {
6254                 inode->index_cnt = 2;
6255                 goto out;
6256         }
6257
6258         path->slots[0]--;
6259
6260         leaf = path->nodes[0];
6261         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6262
6263         if (found_key.objectid != btrfs_ino(inode) ||
6264             found_key.type != BTRFS_DIR_INDEX_KEY) {
6265                 inode->index_cnt = 2;
6266                 goto out;
6267         }
6268
6269         inode->index_cnt = found_key.offset + 1;
6270 out:
6271         btrfs_free_path(path);
6272         return ret;
6273 }
6274
6275 /*
6276  * helper to find a free sequence number in a given directory.  This current
6277  * code is very simple, later versions will do smarter things in the btree
6278  */
6279 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6280 {
6281         int ret = 0;
6282
6283         if (dir->index_cnt == (u64)-1) {
6284                 ret = btrfs_inode_delayed_dir_index_count(dir);
6285                 if (ret) {
6286                         ret = btrfs_set_inode_index_count(dir);
6287                         if (ret)
6288                                 return ret;
6289                 }
6290         }
6291
6292         *index = dir->index_cnt;
6293         dir->index_cnt++;
6294
6295         return ret;
6296 }
6297
6298 static int btrfs_insert_inode_locked(struct inode *inode)
6299 {
6300         struct btrfs_iget_args args;
6301         args.location = &BTRFS_I(inode)->location;
6302         args.root = BTRFS_I(inode)->root;
6303
6304         return insert_inode_locked4(inode,
6305                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6306                    btrfs_find_actor, &args);
6307 }
6308
6309 /*
6310  * Inherit flags from the parent inode.
6311  *
6312  * Currently only the compression flags and the cow flags are inherited.
6313  */
6314 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6315 {
6316         unsigned int flags;
6317
6318         if (!dir)
6319                 return;
6320
6321         flags = BTRFS_I(dir)->flags;
6322
6323         if (flags & BTRFS_INODE_NOCOMPRESS) {
6324                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6325                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6326         } else if (flags & BTRFS_INODE_COMPRESS) {
6327                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6328                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6329         }
6330
6331         if (flags & BTRFS_INODE_NODATACOW) {
6332                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6333                 if (S_ISREG(inode->i_mode))
6334                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6335         }
6336
6337         btrfs_sync_inode_flags_to_i_flags(inode);
6338 }
6339
6340 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6341                                      struct btrfs_root *root,
6342                                      struct inode *dir,
6343                                      const char *name, int name_len,
6344                                      u64 ref_objectid, u64 objectid,
6345                                      umode_t mode, u64 *index)
6346 {
6347         struct btrfs_fs_info *fs_info = root->fs_info;
6348         struct inode *inode;
6349         struct btrfs_inode_item *inode_item;
6350         struct btrfs_key *location;
6351         struct btrfs_path *path;
6352         struct btrfs_inode_ref *ref;
6353         struct btrfs_key key[2];
6354         u32 sizes[2];
6355         int nitems = name ? 2 : 1;
6356         unsigned long ptr;
6357         unsigned int nofs_flag;
6358         int ret;
6359
6360         path = btrfs_alloc_path();
6361         if (!path)
6362                 return ERR_PTR(-ENOMEM);
6363
6364         nofs_flag = memalloc_nofs_save();
6365         inode = new_inode(fs_info->sb);
6366         memalloc_nofs_restore(nofs_flag);
6367         if (!inode) {
6368                 btrfs_free_path(path);
6369                 return ERR_PTR(-ENOMEM);
6370         }
6371
6372         /*
6373          * O_TMPFILE, set link count to 0, so that after this point,
6374          * we fill in an inode item with the correct link count.
6375          */
6376         if (!name)
6377                 set_nlink(inode, 0);
6378
6379         /*
6380          * we have to initialize this early, so we can reclaim the inode
6381          * number if we fail afterwards in this function.
6382          */
6383         inode->i_ino = objectid;
6384
6385         if (dir && name) {
6386                 trace_btrfs_inode_request(dir);
6387
6388                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6389                 if (ret) {
6390                         btrfs_free_path(path);
6391                         iput(inode);
6392                         return ERR_PTR(ret);
6393                 }
6394         } else if (dir) {
6395                 *index = 0;
6396         }
6397         /*
6398          * index_cnt is ignored for everything but a dir,
6399          * btrfs_set_inode_index_count has an explanation for the magic
6400          * number
6401          */
6402         BTRFS_I(inode)->index_cnt = 2;
6403         BTRFS_I(inode)->dir_index = *index;
6404         BTRFS_I(inode)->root = root;
6405         BTRFS_I(inode)->generation = trans->transid;
6406         inode->i_generation = BTRFS_I(inode)->generation;
6407
6408         /*
6409          * We could have gotten an inode number from somebody who was fsynced
6410          * and then removed in this same transaction, so let's just set full
6411          * sync since it will be a full sync anyway and this will blow away the
6412          * old info in the log.
6413          */
6414         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6415
6416         key[0].objectid = objectid;
6417         key[0].type = BTRFS_INODE_ITEM_KEY;
6418         key[0].offset = 0;
6419
6420         sizes[0] = sizeof(struct btrfs_inode_item);
6421
6422         if (name) {
6423                 /*
6424                  * Start new inodes with an inode_ref. This is slightly more
6425                  * efficient for small numbers of hard links since they will
6426                  * be packed into one item. Extended refs will kick in if we
6427                  * add more hard links than can fit in the ref item.
6428                  */
6429                 key[1].objectid = objectid;
6430                 key[1].type = BTRFS_INODE_REF_KEY;
6431                 key[1].offset = ref_objectid;
6432
6433                 sizes[1] = name_len + sizeof(*ref);
6434         }
6435
6436         location = &BTRFS_I(inode)->location;
6437         location->objectid = objectid;
6438         location->offset = 0;
6439         location->type = BTRFS_INODE_ITEM_KEY;
6440
6441         ret = btrfs_insert_inode_locked(inode);
6442         if (ret < 0) {
6443                 iput(inode);
6444                 goto fail;
6445         }
6446
6447         path->leave_spinning = 1;
6448         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6449         if (ret != 0)
6450                 goto fail_unlock;
6451
6452         inode_init_owner(inode, dir, mode);
6453         inode_set_bytes(inode, 0);
6454
6455         inode->i_mtime = current_time(inode);
6456         inode->i_atime = inode->i_mtime;
6457         inode->i_ctime = inode->i_mtime;
6458         BTRFS_I(inode)->i_otime = inode->i_mtime;
6459
6460         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6461                                   struct btrfs_inode_item);
6462         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6463                              sizeof(*inode_item));
6464         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6465
6466         if (name) {
6467                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6468                                      struct btrfs_inode_ref);
6469                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6470                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6471                 ptr = (unsigned long)(ref + 1);
6472                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6473         }
6474
6475         btrfs_mark_buffer_dirty(path->nodes[0]);
6476         btrfs_free_path(path);
6477
6478         btrfs_inherit_iflags(inode, dir);
6479
6480         if (S_ISREG(mode)) {
6481                 if (btrfs_test_opt(fs_info, NODATASUM))
6482                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6483                 if (btrfs_test_opt(fs_info, NODATACOW))
6484                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6485                                 BTRFS_INODE_NODATASUM;
6486         }
6487
6488         inode_tree_add(inode);
6489
6490         trace_btrfs_inode_new(inode);
6491         btrfs_set_inode_last_trans(trans, inode);
6492
6493         btrfs_update_root_times(trans, root);
6494
6495         ret = btrfs_inode_inherit_props(trans, inode, dir);
6496         if (ret)
6497                 btrfs_err(fs_info,
6498                           "error inheriting props for ino %llu (root %llu): %d",
6499                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6500
6501         return inode;
6502
6503 fail_unlock:
6504         discard_new_inode(inode);
6505 fail:
6506         if (dir && name)
6507                 BTRFS_I(dir)->index_cnt--;
6508         btrfs_free_path(path);
6509         return ERR_PTR(ret);
6510 }
6511
6512 /*
6513  * utility function to add 'inode' into 'parent_inode' with
6514  * a give name and a given sequence number.
6515  * if 'add_backref' is true, also insert a backref from the
6516  * inode to the parent directory.
6517  */
6518 int btrfs_add_link(struct btrfs_trans_handle *trans,
6519                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6520                    const char *name, int name_len, int add_backref, u64 index)
6521 {
6522         int ret = 0;
6523         struct btrfs_key key;
6524         struct btrfs_root *root = parent_inode->root;
6525         u64 ino = btrfs_ino(inode);
6526         u64 parent_ino = btrfs_ino(parent_inode);
6527
6528         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6529                 memcpy(&key, &inode->root->root_key, sizeof(key));
6530         } else {
6531                 key.objectid = ino;
6532                 key.type = BTRFS_INODE_ITEM_KEY;
6533                 key.offset = 0;
6534         }
6535
6536         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6537                 ret = btrfs_add_root_ref(trans, key.objectid,
6538                                          root->root_key.objectid, parent_ino,
6539                                          index, name, name_len);
6540         } else if (add_backref) {
6541                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6542                                              parent_ino, index);
6543         }
6544
6545         /* Nothing to clean up yet */
6546         if (ret)
6547                 return ret;
6548
6549         ret = btrfs_insert_dir_item(trans, name, name_len, parent_inode, &key,
6550                                     btrfs_inode_type(&inode->vfs_inode), index);
6551         if (ret == -EEXIST || ret == -EOVERFLOW)
6552                 goto fail_dir_item;
6553         else if (ret) {
6554                 btrfs_abort_transaction(trans, ret);
6555                 return ret;
6556         }
6557
6558         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6559                            name_len * 2);
6560         inode_inc_iversion(&parent_inode->vfs_inode);
6561         /*
6562          * If we are replaying a log tree, we do not want to update the mtime
6563          * and ctime of the parent directory with the current time, since the
6564          * log replay procedure is responsible for setting them to their correct
6565          * values (the ones it had when the fsync was done).
6566          */
6567         if (!test_bit(BTRFS_FS_LOG_RECOVERING, &root->fs_info->flags)) {
6568                 struct timespec64 now = current_time(&parent_inode->vfs_inode);
6569
6570                 parent_inode->vfs_inode.i_mtime = now;
6571                 parent_inode->vfs_inode.i_ctime = now;
6572         }
6573         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6574         if (ret)
6575                 btrfs_abort_transaction(trans, ret);
6576         return ret;
6577
6578 fail_dir_item:
6579         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6580                 u64 local_index;
6581                 int err;
6582                 err = btrfs_del_root_ref(trans, key.objectid,
6583                                          root->root_key.objectid, parent_ino,
6584                                          &local_index, name, name_len);
6585                 if (err)
6586                         btrfs_abort_transaction(trans, err);
6587         } else if (add_backref) {
6588                 u64 local_index;
6589                 int err;
6590
6591                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6592                                           ino, parent_ino, &local_index);
6593                 if (err)
6594                         btrfs_abort_transaction(trans, err);
6595         }
6596
6597         /* Return the original error code */
6598         return ret;
6599 }
6600
6601 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6602                             struct btrfs_inode *dir, struct dentry *dentry,
6603                             struct btrfs_inode *inode, int backref, u64 index)
6604 {
6605         int err = btrfs_add_link(trans, dir, inode,
6606                                  dentry->d_name.name, dentry->d_name.len,
6607                                  backref, index);
6608         if (err > 0)
6609                 err = -EEXIST;
6610         return err;
6611 }
6612
6613 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6614                         umode_t mode, dev_t rdev)
6615 {
6616         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6617         struct btrfs_trans_handle *trans;
6618         struct btrfs_root *root = BTRFS_I(dir)->root;
6619         struct inode *inode = NULL;
6620         int err;
6621         u64 objectid;
6622         u64 index = 0;
6623
6624         /*
6625          * 2 for inode item and ref
6626          * 2 for dir items
6627          * 1 for xattr if selinux is on
6628          */
6629         trans = btrfs_start_transaction(root, 5);
6630         if (IS_ERR(trans))
6631                 return PTR_ERR(trans);
6632
6633         err = btrfs_find_free_ino(root, &objectid);
6634         if (err)
6635                 goto out_unlock;
6636
6637         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6638                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6639                         mode, &index);
6640         if (IS_ERR(inode)) {
6641                 err = PTR_ERR(inode);
6642                 inode = NULL;
6643                 goto out_unlock;
6644         }
6645
6646         /*
6647         * If the active LSM wants to access the inode during
6648         * d_instantiate it needs these. Smack checks to see
6649         * if the filesystem supports xattrs by looking at the
6650         * ops vector.
6651         */
6652         inode->i_op = &btrfs_special_inode_operations;
6653         init_special_inode(inode, inode->i_mode, rdev);
6654
6655         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6656         if (err)
6657                 goto out_unlock;
6658
6659         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6660                         0, index);
6661         if (err)
6662                 goto out_unlock;
6663
6664         btrfs_update_inode(trans, root, inode);
6665         d_instantiate_new(dentry, inode);
6666
6667 out_unlock:
6668         btrfs_end_transaction(trans);
6669         btrfs_btree_balance_dirty(fs_info);
6670         if (err && inode) {
6671                 inode_dec_link_count(inode);
6672                 discard_new_inode(inode);
6673         }
6674         return err;
6675 }
6676
6677 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6678                         umode_t mode, bool excl)
6679 {
6680         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6681         struct btrfs_trans_handle *trans;
6682         struct btrfs_root *root = BTRFS_I(dir)->root;
6683         struct inode *inode = NULL;
6684         int err;
6685         u64 objectid;
6686         u64 index = 0;
6687
6688         /*
6689          * 2 for inode item and ref
6690          * 2 for dir items
6691          * 1 for xattr if selinux is on
6692          */
6693         trans = btrfs_start_transaction(root, 5);
6694         if (IS_ERR(trans))
6695                 return PTR_ERR(trans);
6696
6697         err = btrfs_find_free_ino(root, &objectid);
6698         if (err)
6699                 goto out_unlock;
6700
6701         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6702                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6703                         mode, &index);
6704         if (IS_ERR(inode)) {
6705                 err = PTR_ERR(inode);
6706                 inode = NULL;
6707                 goto out_unlock;
6708         }
6709         /*
6710         * If the active LSM wants to access the inode during
6711         * d_instantiate it needs these. Smack checks to see
6712         * if the filesystem supports xattrs by looking at the
6713         * ops vector.
6714         */
6715         inode->i_fop = &btrfs_file_operations;
6716         inode->i_op = &btrfs_file_inode_operations;
6717         inode->i_mapping->a_ops = &btrfs_aops;
6718
6719         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6720         if (err)
6721                 goto out_unlock;
6722
6723         err = btrfs_update_inode(trans, root, inode);
6724         if (err)
6725                 goto out_unlock;
6726
6727         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6728                         0, index);
6729         if (err)
6730                 goto out_unlock;
6731
6732         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6733         d_instantiate_new(dentry, inode);
6734
6735 out_unlock:
6736         btrfs_end_transaction(trans);
6737         if (err && inode) {
6738                 inode_dec_link_count(inode);
6739                 discard_new_inode(inode);
6740         }
6741         btrfs_btree_balance_dirty(fs_info);
6742         return err;
6743 }
6744
6745 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6746                       struct dentry *dentry)
6747 {
6748         struct btrfs_trans_handle *trans = NULL;
6749         struct btrfs_root *root = BTRFS_I(dir)->root;
6750         struct inode *inode = d_inode(old_dentry);
6751         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6752         u64 index;
6753         int err;
6754         int drop_inode = 0;
6755
6756         /* do not allow sys_link's with other subvols of the same device */
6757         if (root->root_key.objectid != BTRFS_I(inode)->root->root_key.objectid)
6758                 return -EXDEV;
6759
6760         if (inode->i_nlink >= BTRFS_LINK_MAX)
6761                 return -EMLINK;
6762
6763         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6764         if (err)
6765                 goto fail;
6766
6767         /*
6768          * 2 items for inode and inode ref
6769          * 2 items for dir items
6770          * 1 item for parent inode
6771          * 1 item for orphan item deletion if O_TMPFILE
6772          */
6773         trans = btrfs_start_transaction(root, inode->i_nlink ? 5 : 6);
6774         if (IS_ERR(trans)) {
6775                 err = PTR_ERR(trans);
6776                 trans = NULL;
6777                 goto fail;
6778         }
6779
6780         /* There are several dir indexes for this inode, clear the cache. */
6781         BTRFS_I(inode)->dir_index = 0ULL;
6782         inc_nlink(inode);
6783         inode_inc_iversion(inode);
6784         inode->i_ctime = current_time(inode);
6785         ihold(inode);
6786         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6787
6788         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6789                         1, index);
6790
6791         if (err) {
6792                 drop_inode = 1;
6793         } else {
6794                 struct dentry *parent = dentry->d_parent;
6795                 int ret;
6796
6797                 err = btrfs_update_inode(trans, root, inode);
6798                 if (err)
6799                         goto fail;
6800                 if (inode->i_nlink == 1) {
6801                         /*
6802                          * If new hard link count is 1, it's a file created
6803                          * with open(2) O_TMPFILE flag.
6804                          */
6805                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6806                         if (err)
6807                                 goto fail;
6808                 }
6809                 d_instantiate(dentry, inode);
6810                 ret = btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent,
6811                                          true, NULL);
6812                 if (ret == BTRFS_NEED_TRANS_COMMIT) {
6813                         err = btrfs_commit_transaction(trans);
6814                         trans = NULL;
6815                 }
6816         }
6817
6818 fail:
6819         if (trans)
6820                 btrfs_end_transaction(trans);
6821         if (drop_inode) {
6822                 inode_dec_link_count(inode);
6823                 iput(inode);
6824         }
6825         btrfs_btree_balance_dirty(fs_info);
6826         return err;
6827 }
6828
6829 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6830 {
6831         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6832         struct inode *inode = NULL;
6833         struct btrfs_trans_handle *trans;
6834         struct btrfs_root *root = BTRFS_I(dir)->root;
6835         int err = 0;
6836         u64 objectid = 0;
6837         u64 index = 0;
6838
6839         /*
6840          * 2 items for inode and ref
6841          * 2 items for dir items
6842          * 1 for xattr if selinux is on
6843          */
6844         trans = btrfs_start_transaction(root, 5);
6845         if (IS_ERR(trans))
6846                 return PTR_ERR(trans);
6847
6848         err = btrfs_find_free_ino(root, &objectid);
6849         if (err)
6850                 goto out_fail;
6851
6852         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6853                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6854                         S_IFDIR | mode, &index);
6855         if (IS_ERR(inode)) {
6856                 err = PTR_ERR(inode);
6857                 inode = NULL;
6858                 goto out_fail;
6859         }
6860
6861         /* these must be set before we unlock the inode */
6862         inode->i_op = &btrfs_dir_inode_operations;
6863         inode->i_fop = &btrfs_dir_file_operations;
6864
6865         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6866         if (err)
6867                 goto out_fail;
6868
6869         btrfs_i_size_write(BTRFS_I(inode), 0);
6870         err = btrfs_update_inode(trans, root, inode);
6871         if (err)
6872                 goto out_fail;
6873
6874         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6875                         dentry->d_name.name,
6876                         dentry->d_name.len, 0, index);
6877         if (err)
6878                 goto out_fail;
6879
6880         d_instantiate_new(dentry, inode);
6881
6882 out_fail:
6883         btrfs_end_transaction(trans);
6884         if (err && inode) {
6885                 inode_dec_link_count(inode);
6886                 discard_new_inode(inode);
6887         }
6888         btrfs_btree_balance_dirty(fs_info);
6889         return err;
6890 }
6891
6892 static noinline int uncompress_inline(struct btrfs_path *path,
6893                                       struct page *page,
6894                                       size_t pg_offset, u64 extent_offset,
6895                                       struct btrfs_file_extent_item *item)
6896 {
6897         int ret;
6898         struct extent_buffer *leaf = path->nodes[0];
6899         char *tmp;
6900         size_t max_size;
6901         unsigned long inline_size;
6902         unsigned long ptr;
6903         int compress_type;
6904
6905         WARN_ON(pg_offset != 0);
6906         compress_type = btrfs_file_extent_compression(leaf, item);
6907         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6908         inline_size = btrfs_file_extent_inline_item_len(leaf,
6909                                         btrfs_item_nr(path->slots[0]));
6910         tmp = kmalloc(inline_size, GFP_NOFS);
6911         if (!tmp)
6912                 return -ENOMEM;
6913         ptr = btrfs_file_extent_inline_start(item);
6914
6915         read_extent_buffer(leaf, tmp, ptr, inline_size);
6916
6917         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6918         ret = btrfs_decompress(compress_type, tmp, page,
6919                                extent_offset, inline_size, max_size);
6920
6921         /*
6922          * decompression code contains a memset to fill in any space between the end
6923          * of the uncompressed data and the end of max_size in case the decompressed
6924          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6925          * the end of an inline extent and the beginning of the next block, so we
6926          * cover that region here.
6927          */
6928
6929         if (max_size + pg_offset < PAGE_SIZE) {
6930                 char *map = kmap(page);
6931                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6932                 kunmap(page);
6933         }
6934         kfree(tmp);
6935         return ret;
6936 }
6937
6938 /*
6939  * a bit scary, this does extent mapping from logical file offset to the disk.
6940  * the ugly parts come from merging extents from the disk with the in-ram
6941  * representation.  This gets more complex because of the data=ordered code,
6942  * where the in-ram extents might be locked pending data=ordered completion.
6943  *
6944  * This also copies inline extents directly into the page.
6945  */
6946 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6947                                     struct page *page,
6948                                     size_t pg_offset, u64 start, u64 len,
6949                                     int create)
6950 {
6951         struct btrfs_fs_info *fs_info = inode->root->fs_info;
6952         int ret;
6953         int err = 0;
6954         u64 extent_start = 0;
6955         u64 extent_end = 0;
6956         u64 objectid = btrfs_ino(inode);
6957         int extent_type = -1;
6958         struct btrfs_path *path = NULL;
6959         struct btrfs_root *root = inode->root;
6960         struct btrfs_file_extent_item *item;
6961         struct extent_buffer *leaf;
6962         struct btrfs_key found_key;
6963         struct extent_map *em = NULL;
6964         struct extent_map_tree *em_tree = &inode->extent_tree;
6965         struct extent_io_tree *io_tree = &inode->io_tree;
6966         const bool new_inline = !page || create;
6967
6968         read_lock(&em_tree->lock);
6969         em = lookup_extent_mapping(em_tree, start, len);
6970         if (em)
6971                 em->bdev = fs_info->fs_devices->latest_bdev;
6972         read_unlock(&em_tree->lock);
6973
6974         if (em) {
6975                 if (em->start > start || em->start + em->len <= start)
6976                         free_extent_map(em);
6977                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6978                         free_extent_map(em);
6979                 else
6980                         goto out;
6981         }
6982         em = alloc_extent_map();
6983         if (!em) {
6984                 err = -ENOMEM;
6985                 goto out;
6986         }
6987         em->bdev = fs_info->fs_devices->latest_bdev;
6988         em->start = EXTENT_MAP_HOLE;
6989         em->orig_start = EXTENT_MAP_HOLE;
6990         em->len = (u64)-1;
6991         em->block_len = (u64)-1;
6992
6993         path = btrfs_alloc_path();
6994         if (!path) {
6995                 err = -ENOMEM;
6996                 goto out;
6997         }
6998
6999         /* Chances are we'll be called again, so go ahead and do readahead */
7000         path->reada = READA_FORWARD;
7001
7002         /*
7003          * Unless we're going to uncompress the inline extent, no sleep would
7004          * happen.
7005          */
7006         path->leave_spinning = 1;
7007
7008         ret = btrfs_lookup_file_extent(NULL, root, path, objectid, start, 0);
7009         if (ret < 0) {
7010                 err = ret;
7011                 goto out;
7012         } else if (ret > 0) {
7013                 if (path->slots[0] == 0)
7014                         goto not_found;
7015                 path->slots[0]--;
7016         }
7017
7018         leaf = path->nodes[0];
7019         item = btrfs_item_ptr(leaf, path->slots[0],
7020                               struct btrfs_file_extent_item);
7021         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7022         if (found_key.objectid != objectid ||
7023             found_key.type != BTRFS_EXTENT_DATA_KEY) {
7024                 /*
7025                  * If we backup past the first extent we want to move forward
7026                  * and see if there is an extent in front of us, otherwise we'll
7027                  * say there is a hole for our whole search range which can
7028                  * cause problems.
7029                  */
7030                 extent_end = start;
7031                 goto next;
7032         }
7033
7034         extent_type = btrfs_file_extent_type(leaf, item);
7035         extent_start = found_key.offset;
7036         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7037             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7038                 /* Only regular file could have regular/prealloc extent */
7039                 if (!S_ISREG(inode->vfs_inode.i_mode)) {
7040                         ret = -EUCLEAN;
7041                         btrfs_crit(fs_info,
7042                 "regular/prealloc extent found for non-regular inode %llu",
7043                                    btrfs_ino(inode));
7044                         goto out;
7045                 }
7046                 extent_end = extent_start +
7047                        btrfs_file_extent_num_bytes(leaf, item);
7048
7049                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7050                                                        extent_start);
7051         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7052                 size_t size;
7053
7054                 size = btrfs_file_extent_ram_bytes(leaf, item);
7055                 extent_end = ALIGN(extent_start + size,
7056                                    fs_info->sectorsize);
7057
7058                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7059                                                       path->slots[0],
7060                                                       extent_start);
7061         }
7062 next:
7063         if (start >= extent_end) {
7064                 path->slots[0]++;
7065                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7066                         ret = btrfs_next_leaf(root, path);
7067                         if (ret < 0) {
7068                                 err = ret;
7069                                 goto out;
7070                         } else if (ret > 0) {
7071                                 goto not_found;
7072                         }
7073                         leaf = path->nodes[0];
7074                 }
7075                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7076                 if (found_key.objectid != objectid ||
7077                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7078                         goto not_found;
7079                 if (start + len <= found_key.offset)
7080                         goto not_found;
7081                 if (start > found_key.offset)
7082                         goto next;
7083
7084                 /* New extent overlaps with existing one */
7085                 em->start = start;
7086                 em->orig_start = start;
7087                 em->len = found_key.offset - start;
7088                 em->block_start = EXTENT_MAP_HOLE;
7089                 goto insert;
7090         }
7091
7092         btrfs_extent_item_to_extent_map(inode, path, item,
7093                         new_inline, em);
7094
7095         if (extent_type == BTRFS_FILE_EXTENT_REG ||
7096             extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
7097                 goto insert;
7098         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
7099                 unsigned long ptr;
7100                 char *map;
7101                 size_t size;
7102                 size_t extent_offset;
7103                 size_t copy_size;
7104
7105                 if (new_inline)
7106                         goto out;
7107
7108                 size = btrfs_file_extent_ram_bytes(leaf, item);
7109                 extent_offset = page_offset(page) + pg_offset - extent_start;
7110                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7111                                   size - extent_offset);
7112                 em->start = extent_start + extent_offset;
7113                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7114                 em->orig_block_len = em->len;
7115                 em->orig_start = em->start;
7116                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7117
7118                 btrfs_set_path_blocking(path);
7119                 if (!PageUptodate(page)) {
7120                         if (btrfs_file_extent_compression(leaf, item) !=
7121                             BTRFS_COMPRESS_NONE) {
7122                                 ret = uncompress_inline(path, page, pg_offset,
7123                                                         extent_offset, item);
7124                                 if (ret) {
7125                                         err = ret;
7126                                         goto out;
7127                                 }
7128                         } else {
7129                                 map = kmap(page);
7130                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7131                                                    copy_size);
7132                                 if (pg_offset + copy_size < PAGE_SIZE) {
7133                                         memset(map + pg_offset + copy_size, 0,
7134                                                PAGE_SIZE - pg_offset -
7135                                                copy_size);
7136                                 }
7137                                 kunmap(page);
7138                         }
7139                         flush_dcache_page(page);
7140                 }
7141                 set_extent_uptodate(io_tree, em->start,
7142                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7143                 goto insert;
7144         }
7145 not_found:
7146         em->start = start;
7147         em->orig_start = start;
7148         em->len = len;
7149         em->block_start = EXTENT_MAP_HOLE;
7150 insert:
7151         btrfs_release_path(path);
7152         if (em->start > start || extent_map_end(em) <= start) {
7153                 btrfs_err(fs_info,
7154                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7155                           em->start, em->len, start, len);
7156                 err = -EIO;
7157                 goto out;
7158         }
7159
7160         err = 0;
7161         write_lock(&em_tree->lock);
7162         err = btrfs_add_extent_mapping(fs_info, em_tree, &em, start, len);
7163         write_unlock(&em_tree->lock);
7164 out:
7165         btrfs_free_path(path);
7166
7167         trace_btrfs_get_extent(root, inode, em);
7168
7169         if (err) {
7170                 free_extent_map(em);
7171                 return ERR_PTR(err);
7172         }
7173         BUG_ON(!em); /* Error is always set */
7174         return em;
7175 }
7176
7177 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7178                                            u64 start, u64 len)
7179 {
7180         struct extent_map *em;
7181         struct extent_map *hole_em = NULL;
7182         u64 delalloc_start = start;
7183         u64 end;
7184         u64 delalloc_len;
7185         u64 delalloc_end;
7186         int err = 0;
7187
7188         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7189         if (IS_ERR(em))
7190                 return em;
7191         /*
7192          * If our em maps to:
7193          * - a hole or
7194          * - a pre-alloc extent,
7195          * there might actually be delalloc bytes behind it.
7196          */
7197         if (em->block_start != EXTENT_MAP_HOLE &&
7198             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7199                 return em;
7200         else
7201                 hole_em = em;
7202
7203         /* check to see if we've wrapped (len == -1 or similar) */
7204         end = start + len;
7205         if (end < start)
7206                 end = (u64)-1;
7207         else
7208                 end -= 1;
7209
7210         em = NULL;
7211
7212         /* ok, we didn't find anything, lets look for delalloc */
7213         delalloc_len = count_range_bits(&inode->io_tree, &delalloc_start,
7214                                  end, len, EXTENT_DELALLOC, 1);
7215         delalloc_end = delalloc_start + delalloc_len;
7216         if (delalloc_end < delalloc_start)
7217                 delalloc_end = (u64)-1;
7218
7219         /*
7220          * We didn't find anything useful, return the original results from
7221          * get_extent()
7222          */
7223         if (delalloc_start > end || delalloc_end <= start) {
7224                 em = hole_em;
7225                 hole_em = NULL;
7226                 goto out;
7227         }
7228
7229         /*
7230          * Adjust the delalloc_start to make sure it doesn't go backwards from
7231          * the start they passed in
7232          */
7233         delalloc_start = max(start, delalloc_start);
7234         delalloc_len = delalloc_end - delalloc_start;
7235
7236         if (delalloc_len > 0) {
7237                 u64 hole_start;
7238                 u64 hole_len;
7239                 const u64 hole_end = extent_map_end(hole_em);
7240
7241                 em = alloc_extent_map();
7242                 if (!em) {
7243                         err = -ENOMEM;
7244                         goto out;
7245                 }
7246                 em->bdev = NULL;
7247
7248                 ASSERT(hole_em);
7249                 /*
7250                  * When btrfs_get_extent can't find anything it returns one
7251                  * huge hole
7252                  *
7253                  * Make sure what it found really fits our range, and adjust to
7254                  * make sure it is based on the start from the caller
7255                  */
7256                 if (hole_end <= start || hole_em->start > end) {
7257                        free_extent_map(hole_em);
7258                        hole_em = NULL;
7259                 } else {
7260                        hole_start = max(hole_em->start, start);
7261                        hole_len = hole_end - hole_start;
7262                 }
7263
7264                 if (hole_em && delalloc_start > hole_start) {
7265                         /*
7266                          * Our hole starts before our delalloc, so we have to
7267                          * return just the parts of the hole that go until the
7268                          * delalloc starts
7269                          */
7270                         em->len = min(hole_len, delalloc_start - hole_start);
7271                         em->start = hole_start;
7272                         em->orig_start = hole_start;
7273                         /*
7274                          * Don't adjust block start at all, it is fixed at
7275                          * EXTENT_MAP_HOLE
7276                          */
7277                         em->block_start = hole_em->block_start;
7278                         em->block_len = hole_len;
7279                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7280                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7281                 } else {
7282                         /*
7283                          * Hole is out of passed range or it starts after
7284                          * delalloc range
7285                          */
7286                         em->start = delalloc_start;
7287                         em->len = delalloc_len;
7288                         em->orig_start = delalloc_start;
7289                         em->block_start = EXTENT_MAP_DELALLOC;
7290                         em->block_len = delalloc_len;
7291                 }
7292         } else {
7293                 return hole_em;
7294         }
7295 out:
7296
7297         free_extent_map(hole_em);
7298         if (err) {
7299                 free_extent_map(em);
7300                 return ERR_PTR(err);
7301         }
7302         return em;
7303 }
7304
7305 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7306                                                   const u64 start,
7307                                                   const u64 len,
7308                                                   const u64 orig_start,
7309                                                   const u64 block_start,
7310                                                   const u64 block_len,
7311                                                   const u64 orig_block_len,
7312                                                   const u64 ram_bytes,
7313                                                   const int type)
7314 {
7315         struct extent_map *em = NULL;
7316         int ret;
7317
7318         if (type != BTRFS_ORDERED_NOCOW) {
7319                 em = create_io_em(inode, start, len, orig_start,
7320                                   block_start, block_len, orig_block_len,
7321                                   ram_bytes,
7322                                   BTRFS_COMPRESS_NONE, /* compress_type */
7323                                   type);
7324                 if (IS_ERR(em))
7325                         goto out;
7326         }
7327         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7328                                            len, block_len, type);
7329         if (ret) {
7330                 if (em) {
7331                         free_extent_map(em);
7332                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7333                                                 start + len - 1, 0);
7334                 }
7335                 em = ERR_PTR(ret);
7336         }
7337  out:
7338
7339         return em;
7340 }
7341
7342 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7343                                                   u64 start, u64 len)
7344 {
7345         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7346         struct btrfs_root *root = BTRFS_I(inode)->root;
7347         struct extent_map *em;
7348         struct btrfs_key ins;
7349         u64 alloc_hint;
7350         int ret;
7351
7352         alloc_hint = get_extent_allocation_hint(inode, start, len);
7353         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7354                                    0, alloc_hint, &ins, 1, 1);
7355         if (ret)
7356                 return ERR_PTR(ret);
7357
7358         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7359                                      ins.objectid, ins.offset, ins.offset,
7360                                      ins.offset, BTRFS_ORDERED_REGULAR);
7361         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7362         if (IS_ERR(em))
7363                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7364                                            ins.offset, 1);
7365
7366         return em;
7367 }
7368
7369 /*
7370  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7371  * block must be cow'd
7372  */
7373 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7374                               u64 *orig_start, u64 *orig_block_len,
7375                               u64 *ram_bytes)
7376 {
7377         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7378         struct btrfs_path *path;
7379         int ret;
7380         struct extent_buffer *leaf;
7381         struct btrfs_root *root = BTRFS_I(inode)->root;
7382         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7383         struct btrfs_file_extent_item *fi;
7384         struct btrfs_key key;
7385         u64 disk_bytenr;
7386         u64 backref_offset;
7387         u64 extent_end;
7388         u64 num_bytes;
7389         int slot;
7390         int found_type;
7391         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7392
7393         path = btrfs_alloc_path();
7394         if (!path)
7395                 return -ENOMEM;
7396
7397         ret = btrfs_lookup_file_extent(NULL, root, path,
7398                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7399         if (ret < 0)
7400                 goto out;
7401
7402         slot = path->slots[0];
7403         if (ret == 1) {
7404                 if (slot == 0) {
7405                         /* can't find the item, must cow */
7406                         ret = 0;
7407                         goto out;
7408                 }
7409                 slot--;
7410         }
7411         ret = 0;
7412         leaf = path->nodes[0];
7413         btrfs_item_key_to_cpu(leaf, &key, slot);
7414         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7415             key.type != BTRFS_EXTENT_DATA_KEY) {
7416                 /* not our file or wrong item type, must cow */
7417                 goto out;
7418         }
7419
7420         if (key.offset > offset) {
7421                 /* Wrong offset, must cow */
7422                 goto out;
7423         }
7424
7425         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7426         found_type = btrfs_file_extent_type(leaf, fi);
7427         if (found_type != BTRFS_FILE_EXTENT_REG &&
7428             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7429                 /* not a regular extent, must cow */
7430                 goto out;
7431         }
7432
7433         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7434                 goto out;
7435
7436         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7437         if (extent_end <= offset)
7438                 goto out;
7439
7440         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7441         if (disk_bytenr == 0)
7442                 goto out;
7443
7444         if (btrfs_file_extent_compression(leaf, fi) ||
7445             btrfs_file_extent_encryption(leaf, fi) ||
7446             btrfs_file_extent_other_encoding(leaf, fi))
7447                 goto out;
7448
7449         /*
7450          * Do the same check as in btrfs_cross_ref_exist but without the
7451          * unnecessary search.
7452          */
7453         if (btrfs_file_extent_generation(leaf, fi) <=
7454             btrfs_root_last_snapshot(&root->root_item))
7455                 goto out;
7456
7457         backref_offset = btrfs_file_extent_offset(leaf, fi);
7458
7459         if (orig_start) {
7460                 *orig_start = key.offset - backref_offset;
7461                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7462                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7463         }
7464
7465         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7466                 goto out;
7467
7468         num_bytes = min(offset + *len, extent_end) - offset;
7469         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7470                 u64 range_end;
7471
7472                 range_end = round_up(offset + num_bytes,
7473                                      root->fs_info->sectorsize) - 1;
7474                 ret = test_range_bit(io_tree, offset, range_end,
7475                                      EXTENT_DELALLOC, 0, NULL);
7476                 if (ret) {
7477                         ret = -EAGAIN;
7478                         goto out;
7479                 }
7480         }
7481
7482         btrfs_release_path(path);
7483
7484         /*
7485          * look for other files referencing this extent, if we
7486          * find any we must cow
7487          */
7488
7489         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7490                                     key.offset - backref_offset, disk_bytenr);
7491         if (ret) {
7492                 ret = 0;
7493                 goto out;
7494         }
7495
7496         /*
7497          * adjust disk_bytenr and num_bytes to cover just the bytes
7498          * in this extent we are about to write.  If there
7499          * are any csums in that range we have to cow in order
7500          * to keep the csums correct
7501          */
7502         disk_bytenr += backref_offset;
7503         disk_bytenr += offset - key.offset;
7504         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7505                 goto out;
7506         /*
7507          * all of the above have passed, it is safe to overwrite this extent
7508          * without cow
7509          */
7510         *len = num_bytes;
7511         ret = 1;
7512 out:
7513         btrfs_free_path(path);
7514         return ret;
7515 }
7516
7517 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7518                               struct extent_state **cached_state, int writing)
7519 {
7520         struct btrfs_ordered_extent *ordered;
7521         int ret = 0;
7522
7523         while (1) {
7524                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7525                                  cached_state);
7526                 /*
7527                  * We're concerned with the entire range that we're going to be
7528                  * doing DIO to, so we need to make sure there's no ordered
7529                  * extents in this range.
7530                  */
7531                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7532                                                      lockend - lockstart + 1);
7533
7534                 /*
7535                  * We need to make sure there are no buffered pages in this
7536                  * range either, we could have raced between the invalidate in
7537                  * generic_file_direct_write and locking the extent.  The
7538                  * invalidate needs to happen so that reads after a write do not
7539                  * get stale data.
7540                  */
7541                 if (!ordered &&
7542                     (!writing || !filemap_range_has_page(inode->i_mapping,
7543                                                          lockstart, lockend)))
7544                         break;
7545
7546                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7547                                      cached_state);
7548
7549                 if (ordered) {
7550                         /*
7551                          * If we are doing a DIO read and the ordered extent we
7552                          * found is for a buffered write, we can not wait for it
7553                          * to complete and retry, because if we do so we can
7554                          * deadlock with concurrent buffered writes on page
7555                          * locks. This happens only if our DIO read covers more
7556                          * than one extent map, if at this point has already
7557                          * created an ordered extent for a previous extent map
7558                          * and locked its range in the inode's io tree, and a
7559                          * concurrent write against that previous extent map's
7560                          * range and this range started (we unlock the ranges
7561                          * in the io tree only when the bios complete and
7562                          * buffered writes always lock pages before attempting
7563                          * to lock range in the io tree).
7564                          */
7565                         if (writing ||
7566                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7567                                 btrfs_start_ordered_extent(inode, ordered, 1);
7568                         else
7569                                 ret = -ENOTBLK;
7570                         btrfs_put_ordered_extent(ordered);
7571                 } else {
7572                         /*
7573                          * We could trigger writeback for this range (and wait
7574                          * for it to complete) and then invalidate the pages for
7575                          * this range (through invalidate_inode_pages2_range()),
7576                          * but that can lead us to a deadlock with a concurrent
7577                          * call to readpages() (a buffered read or a defrag call
7578                          * triggered a readahead) on a page lock due to an
7579                          * ordered dio extent we created before but did not have
7580                          * yet a corresponding bio submitted (whence it can not
7581                          * complete), which makes readpages() wait for that
7582                          * ordered extent to complete while holding a lock on
7583                          * that page.
7584                          */
7585                         ret = -ENOTBLK;
7586                 }
7587
7588                 if (ret)
7589                         break;
7590
7591                 cond_resched();
7592         }
7593
7594         return ret;
7595 }
7596
7597 /* The callers of this must take lock_extent() */
7598 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7599                                        u64 orig_start, u64 block_start,
7600                                        u64 block_len, u64 orig_block_len,
7601                                        u64 ram_bytes, int compress_type,
7602                                        int type)
7603 {
7604         struct extent_map_tree *em_tree;
7605         struct extent_map *em;
7606         struct btrfs_root *root = BTRFS_I(inode)->root;
7607         int ret;
7608
7609         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7610                type == BTRFS_ORDERED_COMPRESSED ||
7611                type == BTRFS_ORDERED_NOCOW ||
7612                type == BTRFS_ORDERED_REGULAR);
7613
7614         em_tree = &BTRFS_I(inode)->extent_tree;
7615         em = alloc_extent_map();
7616         if (!em)
7617                 return ERR_PTR(-ENOMEM);
7618
7619         em->start = start;
7620         em->orig_start = orig_start;
7621         em->len = len;
7622         em->block_len = block_len;
7623         em->block_start = block_start;
7624         em->bdev = root->fs_info->fs_devices->latest_bdev;
7625         em->orig_block_len = orig_block_len;
7626         em->ram_bytes = ram_bytes;
7627         em->generation = -1;
7628         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7629         if (type == BTRFS_ORDERED_PREALLOC) {
7630                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7631         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7632                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7633                 em->compress_type = compress_type;
7634         }
7635
7636         do {
7637                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7638                                 em->start + em->len - 1, 0);
7639                 write_lock(&em_tree->lock);
7640                 ret = add_extent_mapping(em_tree, em, 1);
7641                 write_unlock(&em_tree->lock);
7642                 /*
7643                  * The caller has taken lock_extent(), who could race with us
7644                  * to add em?
7645                  */
7646         } while (ret == -EEXIST);
7647
7648         if (ret) {
7649                 free_extent_map(em);
7650                 return ERR_PTR(ret);
7651         }
7652
7653         /* em got 2 refs now, callers needs to do free_extent_map once. */
7654         return em;
7655 }
7656
7657
7658 static int btrfs_get_blocks_direct_read(struct extent_map *em,
7659                                         struct buffer_head *bh_result,
7660                                         struct inode *inode,
7661                                         u64 start, u64 len)
7662 {
7663         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7664
7665         if (em->block_start == EXTENT_MAP_HOLE ||
7666                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7667                 return -ENOENT;
7668
7669         len = min(len, em->len - (start - em->start));
7670
7671         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7672                 inode->i_blkbits;
7673         bh_result->b_size = len;
7674         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7675         set_buffer_mapped(bh_result);
7676
7677         return 0;
7678 }
7679
7680 static int btrfs_get_blocks_direct_write(struct extent_map **map,
7681                                          struct buffer_head *bh_result,
7682                                          struct inode *inode,
7683                                          struct btrfs_dio_data *dio_data,
7684                                          u64 start, u64 len)
7685 {
7686         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7687         struct extent_map *em = *map;
7688         int ret = 0;
7689
7690         /*
7691          * We don't allocate a new extent in the following cases
7692          *
7693          * 1) The inode is marked as NODATACOW. In this case we'll just use the
7694          * existing extent.
7695          * 2) The extent is marked as PREALLOC. We're good to go here and can
7696          * just use the extent.
7697          *
7698          */
7699         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7700             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7701              em->block_start != EXTENT_MAP_HOLE)) {
7702                 int type;
7703                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7704
7705                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7706                         type = BTRFS_ORDERED_PREALLOC;
7707                 else
7708                         type = BTRFS_ORDERED_NOCOW;
7709                 len = min(len, em->len - (start - em->start));
7710                 block_start = em->block_start + (start - em->start);
7711
7712                 if (can_nocow_extent(inode, start, &len, &orig_start,
7713                                      &orig_block_len, &ram_bytes) == 1 &&
7714                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7715                         struct extent_map *em2;
7716
7717                         em2 = btrfs_create_dio_extent(inode, start, len,
7718                                                       orig_start, block_start,
7719                                                       len, orig_block_len,
7720                                                       ram_bytes, type);
7721                         btrfs_dec_nocow_writers(fs_info, block_start);
7722                         if (type == BTRFS_ORDERED_PREALLOC) {
7723                                 free_extent_map(em);
7724                                 *map = em = em2;
7725                         }
7726
7727                         if (em2 && IS_ERR(em2)) {
7728                                 ret = PTR_ERR(em2);
7729                                 goto out;
7730                         }
7731                         /*
7732                          * For inode marked NODATACOW or extent marked PREALLOC,
7733                          * use the existing or preallocated extent, so does not
7734                          * need to adjust btrfs_space_info's bytes_may_use.
7735                          */
7736                         btrfs_free_reserved_data_space_noquota(inode, start,
7737                                                                len);
7738                         goto skip_cow;
7739                 }
7740         }
7741
7742         /* this will cow the extent */
7743         len = bh_result->b_size;
7744         free_extent_map(em);
7745         *map = em = btrfs_new_extent_direct(inode, start, len);
7746         if (IS_ERR(em)) {
7747                 ret = PTR_ERR(em);
7748                 goto out;
7749         }
7750
7751         len = min(len, em->len - (start - em->start));
7752
7753 skip_cow:
7754         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7755                 inode->i_blkbits;
7756         bh_result->b_size = len;
7757         bh_result->b_bdev = fs_info->fs_devices->latest_bdev;
7758         set_buffer_mapped(bh_result);
7759
7760         if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7761                 set_buffer_new(bh_result);
7762
7763         /*
7764          * Need to update the i_size under the extent lock so buffered
7765          * readers will get the updated i_size when we unlock.
7766          */
7767         if (!dio_data->overwrite && start + len > i_size_read(inode))
7768                 i_size_write(inode, start + len);
7769
7770         WARN_ON(dio_data->reserve < len);
7771         dio_data->reserve -= len;
7772         dio_data->unsubmitted_oe_range_end = start + len;
7773         current->journal_info = dio_data;
7774 out:
7775         return ret;
7776 }
7777
7778 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7779                                    struct buffer_head *bh_result, int create)
7780 {
7781         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7782         struct extent_map *em;
7783         struct extent_state *cached_state = NULL;
7784         struct btrfs_dio_data *dio_data = NULL;
7785         u64 start = iblock << inode->i_blkbits;
7786         u64 lockstart, lockend;
7787         u64 len = bh_result->b_size;
7788         int ret = 0;
7789
7790         if (!create)
7791                 len = min_t(u64, len, fs_info->sectorsize);
7792
7793         lockstart = start;
7794         lockend = start + len - 1;
7795
7796         if (current->journal_info) {
7797                 /*
7798                  * Need to pull our outstanding extents and set journal_info to NULL so
7799                  * that anything that needs to check if there's a transaction doesn't get
7800                  * confused.
7801                  */
7802                 dio_data = current->journal_info;
7803                 current->journal_info = NULL;
7804         }
7805
7806         /*
7807          * If this errors out it's because we couldn't invalidate pagecache for
7808          * this range and we need to fallback to buffered.
7809          */
7810         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7811                                create)) {
7812                 ret = -ENOTBLK;
7813                 goto err;
7814         }
7815
7816         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7817         if (IS_ERR(em)) {
7818                 ret = PTR_ERR(em);
7819                 goto unlock_err;
7820         }
7821
7822         /*
7823          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7824          * io.  INLINE is special, and we could probably kludge it in here, but
7825          * it's still buffered so for safety lets just fall back to the generic
7826          * buffered path.
7827          *
7828          * For COMPRESSED we _have_ to read the entire extent in so we can
7829          * decompress it, so there will be buffering required no matter what we
7830          * do, so go ahead and fallback to buffered.
7831          *
7832          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7833          * to buffered IO.  Don't blame me, this is the price we pay for using
7834          * the generic code.
7835          */
7836         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7837             em->block_start == EXTENT_MAP_INLINE) {
7838                 free_extent_map(em);
7839                 ret = -ENOTBLK;
7840                 goto unlock_err;
7841         }
7842
7843         if (create) {
7844                 ret = btrfs_get_blocks_direct_write(&em, bh_result, inode,
7845                                                     dio_data, start, len);
7846                 if (ret < 0)
7847                         goto unlock_err;
7848
7849                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
7850                                      lockend, &cached_state);
7851         } else {
7852                 ret = btrfs_get_blocks_direct_read(em, bh_result, inode,
7853                                                    start, len);
7854                 /* Can be negative only if we read from a hole */
7855                 if (ret < 0) {
7856                         ret = 0;
7857                         free_extent_map(em);
7858                         goto unlock_err;
7859                 }
7860                 /*
7861                  * We need to unlock only the end area that we aren't using.
7862                  * The rest is going to be unlocked by the endio routine.
7863                  */
7864                 lockstart = start + bh_result->b_size;
7865                 if (lockstart < lockend) {
7866                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
7867                                              lockstart, lockend, &cached_state);
7868                 } else {
7869                         free_extent_state(cached_state);
7870                 }
7871         }
7872
7873         free_extent_map(em);
7874
7875         return 0;
7876
7877 unlock_err:
7878         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7879                              &cached_state);
7880 err:
7881         if (dio_data)
7882                 current->journal_info = dio_data;
7883         return ret;
7884 }
7885
7886 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7887                                                  struct bio *bio,
7888                                                  int mirror_num)
7889 {
7890         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7891         blk_status_t ret;
7892
7893         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7894
7895         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7896         if (ret)
7897                 return ret;
7898
7899         ret = btrfs_map_bio(fs_info, bio, mirror_num);
7900
7901         return ret;
7902 }
7903
7904 static int btrfs_check_dio_repairable(struct inode *inode,
7905                                       struct bio *failed_bio,
7906                                       struct io_failure_record *failrec,
7907                                       int failed_mirror)
7908 {
7909         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7910         int num_copies;
7911
7912         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7913         if (num_copies == 1) {
7914                 /*
7915                  * we only have a single copy of the data, so don't bother with
7916                  * all the retry and error correction code that follows. no
7917                  * matter what the error is, it is very likely to persist.
7918                  */
7919                 btrfs_debug(fs_info,
7920                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7921                         num_copies, failrec->this_mirror, failed_mirror);
7922                 return 0;
7923         }
7924
7925         failrec->failed_mirror = failed_mirror;
7926         failrec->this_mirror++;
7927         if (failrec->this_mirror == failed_mirror)
7928                 failrec->this_mirror++;
7929
7930         if (failrec->this_mirror > num_copies) {
7931                 btrfs_debug(fs_info,
7932                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
7933                         num_copies, failrec->this_mirror, failed_mirror);
7934                 return 0;
7935         }
7936
7937         return 1;
7938 }
7939
7940 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
7941                                    struct page *page, unsigned int pgoff,
7942                                    u64 start, u64 end, int failed_mirror,
7943                                    bio_end_io_t *repair_endio, void *repair_arg)
7944 {
7945         struct io_failure_record *failrec;
7946         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7947         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
7948         struct bio *bio;
7949         int isector;
7950         unsigned int read_mode = 0;
7951         int segs;
7952         int ret;
7953         blk_status_t status;
7954         struct bio_vec bvec;
7955
7956         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
7957
7958         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7959         if (ret)
7960                 return errno_to_blk_status(ret);
7961
7962         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7963                                          failed_mirror);
7964         if (!ret) {
7965                 free_io_failure(failure_tree, io_tree, failrec);
7966                 return BLK_STS_IOERR;
7967         }
7968
7969         segs = bio_segments(failed_bio);
7970         bio_get_first_bvec(failed_bio, &bvec);
7971         if (segs > 1 ||
7972             (bvec.bv_len > btrfs_inode_sectorsize(inode)))
7973                 read_mode |= REQ_FAILFAST_DEV;
7974
7975         isector = start - btrfs_io_bio(failed_bio)->logical;
7976         isector >>= inode->i_sb->s_blocksize_bits;
7977         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7978                                 pgoff, isector, repair_endio, repair_arg);
7979         bio->bi_opf = REQ_OP_READ | read_mode;
7980
7981         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7982                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
7983                     read_mode, failrec->this_mirror, failrec->in_validation);
7984
7985         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
7986         if (status) {
7987                 free_io_failure(failure_tree, io_tree, failrec);
7988                 bio_put(bio);
7989         }
7990
7991         return status;
7992 }
7993
7994 struct btrfs_retry_complete {
7995         struct completion done;
7996         struct inode *inode;
7997         u64 start;
7998         int uptodate;
7999 };
8000
8001 static void btrfs_retry_endio_nocsum(struct bio *bio)
8002 {
8003         struct btrfs_retry_complete *done = bio->bi_private;
8004         struct inode *inode = done->inode;
8005         struct bio_vec *bvec;
8006         struct extent_io_tree *io_tree, *failure_tree;
8007         struct bvec_iter_all iter_all;
8008
8009         if (bio->bi_status)
8010                 goto end;
8011
8012         ASSERT(bio->bi_vcnt == 1);
8013         io_tree = &BTRFS_I(inode)->io_tree;
8014         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8015         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(inode));
8016
8017         done->uptodate = 1;
8018         ASSERT(!bio_flagged(bio, BIO_CLONED));
8019         bio_for_each_segment_all(bvec, bio, iter_all)
8020                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8021                                  io_tree, done->start, bvec->bv_page,
8022                                  btrfs_ino(BTRFS_I(inode)), 0);
8023 end:
8024         complete(&done->done);
8025         bio_put(bio);
8026 }
8027
8028 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8029                                                 struct btrfs_io_bio *io_bio)
8030 {
8031         struct btrfs_fs_info *fs_info;
8032         struct bio_vec bvec;
8033         struct bvec_iter iter;
8034         struct btrfs_retry_complete done;
8035         u64 start;
8036         unsigned int pgoff;
8037         u32 sectorsize;
8038         int nr_sectors;
8039         blk_status_t ret;
8040         blk_status_t err = BLK_STS_OK;
8041
8042         fs_info = BTRFS_I(inode)->root->fs_info;
8043         sectorsize = fs_info->sectorsize;
8044
8045         start = io_bio->logical;
8046         done.inode = inode;
8047         io_bio->bio.bi_iter = io_bio->iter;
8048
8049         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8050                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8051                 pgoff = bvec.bv_offset;
8052
8053 next_block_or_try_again:
8054                 done.uptodate = 0;
8055                 done.start = start;
8056                 init_completion(&done.done);
8057
8058                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8059                                 pgoff, start, start + sectorsize - 1,
8060                                 io_bio->mirror_num,
8061                                 btrfs_retry_endio_nocsum, &done);
8062                 if (ret) {
8063                         err = ret;
8064                         goto next;
8065                 }
8066
8067                 wait_for_completion_io(&done.done);
8068
8069                 if (!done.uptodate) {
8070                         /* We might have another mirror, so try again */
8071                         goto next_block_or_try_again;
8072                 }
8073
8074 next:
8075                 start += sectorsize;
8076
8077                 nr_sectors--;
8078                 if (nr_sectors) {
8079                         pgoff += sectorsize;
8080                         ASSERT(pgoff < PAGE_SIZE);
8081                         goto next_block_or_try_again;
8082                 }
8083         }
8084
8085         return err;
8086 }
8087
8088 static void btrfs_retry_endio(struct bio *bio)
8089 {
8090         struct btrfs_retry_complete *done = bio->bi_private;
8091         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8092         struct extent_io_tree *io_tree, *failure_tree;
8093         struct inode *inode = done->inode;
8094         struct bio_vec *bvec;
8095         int uptodate;
8096         int ret;
8097         int i = 0;
8098         struct bvec_iter_all iter_all;
8099
8100         if (bio->bi_status)
8101                 goto end;
8102
8103         uptodate = 1;
8104
8105         ASSERT(bio->bi_vcnt == 1);
8106         ASSERT(bio_first_bvec_all(bio)->bv_len == btrfs_inode_sectorsize(done->inode));
8107
8108         io_tree = &BTRFS_I(inode)->io_tree;
8109         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8110
8111         ASSERT(!bio_flagged(bio, BIO_CLONED));
8112         bio_for_each_segment_all(bvec, bio, iter_all) {
8113                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8114                                              bvec->bv_offset, done->start,
8115                                              bvec->bv_len);
8116                 if (!ret)
8117                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8118                                          failure_tree, io_tree, done->start,
8119                                          bvec->bv_page,
8120                                          btrfs_ino(BTRFS_I(inode)),
8121                                          bvec->bv_offset);
8122                 else
8123                         uptodate = 0;
8124                 i++;
8125         }
8126
8127         done->uptodate = uptodate;
8128 end:
8129         complete(&done->done);
8130         bio_put(bio);
8131 }
8132
8133 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8134                 struct btrfs_io_bio *io_bio, blk_status_t err)
8135 {
8136         struct btrfs_fs_info *fs_info;
8137         struct bio_vec bvec;
8138         struct bvec_iter iter;
8139         struct btrfs_retry_complete done;
8140         u64 start;
8141         u64 offset = 0;
8142         u32 sectorsize;
8143         int nr_sectors;
8144         unsigned int pgoff;
8145         int csum_pos;
8146         bool uptodate = (err == 0);
8147         int ret;
8148         blk_status_t status;
8149
8150         fs_info = BTRFS_I(inode)->root->fs_info;
8151         sectorsize = fs_info->sectorsize;
8152
8153         err = BLK_STS_OK;
8154         start = io_bio->logical;
8155         done.inode = inode;
8156         io_bio->bio.bi_iter = io_bio->iter;
8157
8158         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8159                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8160
8161                 pgoff = bvec.bv_offset;
8162 next_block:
8163                 if (uptodate) {
8164                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8165                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8166                                         bvec.bv_page, pgoff, start, sectorsize);
8167                         if (likely(!ret))
8168                                 goto next;
8169                 }
8170 try_again:
8171                 done.uptodate = 0;
8172                 done.start = start;
8173                 init_completion(&done.done);
8174
8175                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8176                                         pgoff, start, start + sectorsize - 1,
8177                                         io_bio->mirror_num, btrfs_retry_endio,
8178                                         &done);
8179                 if (status) {
8180                         err = status;
8181                         goto next;
8182                 }
8183
8184                 wait_for_completion_io(&done.done);
8185
8186                 if (!done.uptodate) {
8187                         /* We might have another mirror, so try again */
8188                         goto try_again;
8189                 }
8190 next:
8191                 offset += sectorsize;
8192                 start += sectorsize;
8193
8194                 ASSERT(nr_sectors);
8195
8196                 nr_sectors--;
8197                 if (nr_sectors) {
8198                         pgoff += sectorsize;
8199                         ASSERT(pgoff < PAGE_SIZE);
8200                         goto next_block;
8201                 }
8202         }
8203
8204         return err;
8205 }
8206
8207 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8208                 struct btrfs_io_bio *io_bio, blk_status_t err)
8209 {
8210         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8211
8212         if (skip_csum) {
8213                 if (unlikely(err))
8214                         return __btrfs_correct_data_nocsum(inode, io_bio);
8215                 else
8216                         return BLK_STS_OK;
8217         } else {
8218                 return __btrfs_subio_endio_read(inode, io_bio, err);
8219         }
8220 }
8221
8222 static void btrfs_endio_direct_read(struct bio *bio)
8223 {
8224         struct btrfs_dio_private *dip = bio->bi_private;
8225         struct inode *inode = dip->inode;
8226         struct bio *dio_bio;
8227         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8228         blk_status_t err = bio->bi_status;
8229
8230         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8231                 err = btrfs_subio_endio_read(inode, io_bio, err);
8232
8233         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8234                       dip->logical_offset + dip->bytes - 1);
8235         dio_bio = dip->dio_bio;
8236
8237         kfree(dip);
8238
8239         dio_bio->bi_status = err;
8240         dio_end_io(dio_bio);
8241         btrfs_io_bio_free_csum(io_bio);
8242         bio_put(bio);
8243 }
8244
8245 static void __endio_write_update_ordered(struct inode *inode,
8246                                          const u64 offset, const u64 bytes,
8247                                          const bool uptodate)
8248 {
8249         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8250         struct btrfs_ordered_extent *ordered = NULL;
8251         struct btrfs_workqueue *wq;
8252         u64 ordered_offset = offset;
8253         u64 ordered_bytes = bytes;
8254         u64 last_offset;
8255
8256         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
8257                 wq = fs_info->endio_freespace_worker;
8258         else
8259                 wq = fs_info->endio_write_workers;
8260
8261         while (ordered_offset < offset + bytes) {
8262                 last_offset = ordered_offset;
8263                 if (btrfs_dec_test_first_ordered_pending(inode, &ordered,
8264                                                            &ordered_offset,
8265                                                            ordered_bytes,
8266                                                            uptodate)) {
8267                         btrfs_init_work(&ordered->work, finish_ordered_fn, NULL,
8268                                         NULL);
8269                         btrfs_queue_work(wq, &ordered->work);
8270                 }
8271                 /*
8272                  * If btrfs_dec_test_ordered_pending does not find any ordered
8273                  * extent in the range, we can exit.
8274                  */
8275                 if (ordered_offset == last_offset)
8276                         return;
8277                 /*
8278                  * Our bio might span multiple ordered extents. In this case
8279                  * we keep going until we have accounted the whole dio.
8280                  */
8281                 if (ordered_offset < offset + bytes) {
8282                         ordered_bytes = offset + bytes - ordered_offset;
8283                         ordered = NULL;
8284                 }
8285         }
8286 }
8287
8288 static void btrfs_endio_direct_write(struct bio *bio)
8289 {
8290         struct btrfs_dio_private *dip = bio->bi_private;
8291         struct bio *dio_bio = dip->dio_bio;
8292
8293         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8294                                      dip->bytes, !bio->bi_status);
8295
8296         kfree(dip);
8297
8298         dio_bio->bi_status = bio->bi_status;
8299         dio_end_io(dio_bio);
8300         bio_put(bio);
8301 }
8302
8303 static blk_status_t btrfs_submit_bio_start_direct_io(void *private_data,
8304                                     struct bio *bio, u64 offset)
8305 {
8306         struct inode *inode = private_data;
8307         blk_status_t ret;
8308         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8309         BUG_ON(ret); /* -ENOMEM */
8310         return 0;
8311 }
8312
8313 static void btrfs_end_dio_bio(struct bio *bio)
8314 {
8315         struct btrfs_dio_private *dip = bio->bi_private;
8316         blk_status_t err = bio->bi_status;
8317
8318         if (err)
8319                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8320                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8321                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8322                            bio->bi_opf,
8323                            (unsigned long long)bio->bi_iter.bi_sector,
8324                            bio->bi_iter.bi_size, err);
8325
8326         if (dip->subio_endio)
8327                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8328
8329         if (err) {
8330                 /*
8331                  * We want to perceive the errors flag being set before
8332                  * decrementing the reference count. We don't need a barrier
8333                  * since atomic operations with a return value are fully
8334                  * ordered as per atomic_t.txt
8335                  */
8336                 dip->errors = 1;
8337         }
8338
8339         /* if there are more bios still pending for this dio, just exit */
8340         if (!atomic_dec_and_test(&dip->pending_bios))
8341                 goto out;
8342
8343         if (dip->errors) {
8344                 bio_io_error(dip->orig_bio);
8345         } else {
8346                 dip->dio_bio->bi_status = BLK_STS_OK;
8347                 bio_endio(dip->orig_bio);
8348         }
8349 out:
8350         bio_put(bio);
8351 }
8352
8353 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8354                                                  struct btrfs_dio_private *dip,
8355                                                  struct bio *bio,
8356                                                  u64 file_offset)
8357 {
8358         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8359         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8360         blk_status_t ret;
8361
8362         /*
8363          * We load all the csum data we need when we submit
8364          * the first bio to reduce the csum tree search and
8365          * contention.
8366          */
8367         if (dip->logical_offset == file_offset) {
8368                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8369                                                 file_offset);
8370                 if (ret)
8371                         return ret;
8372         }
8373
8374         if (bio == dip->orig_bio)
8375                 return 0;
8376
8377         file_offset -= dip->logical_offset;
8378         file_offset >>= inode->i_sb->s_blocksize_bits;
8379         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8380
8381         return 0;
8382 }
8383
8384 static inline blk_status_t btrfs_submit_dio_bio(struct bio *bio,
8385                 struct inode *inode, u64 file_offset, int async_submit)
8386 {
8387         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8388         struct btrfs_dio_private *dip = bio->bi_private;
8389         bool write = bio_op(bio) == REQ_OP_WRITE;
8390         blk_status_t ret;
8391
8392         /* Check btrfs_submit_bio_hook() for rules about async submit. */
8393         if (async_submit)
8394                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8395
8396         if (!write) {
8397                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8398                 if (ret)
8399                         goto err;
8400         }
8401
8402         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8403                 goto map;
8404
8405         if (write && async_submit) {
8406                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8407                                           file_offset, inode,
8408                                           btrfs_submit_bio_start_direct_io);
8409                 goto err;
8410         } else if (write) {
8411                 /*
8412                  * If we aren't doing async submit, calculate the csum of the
8413                  * bio now.
8414                  */
8415                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8416                 if (ret)
8417                         goto err;
8418         } else {
8419                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8420                                                      file_offset);
8421                 if (ret)
8422                         goto err;
8423         }
8424 map:
8425         ret = btrfs_map_bio(fs_info, bio, 0);
8426 err:
8427         return ret;
8428 }
8429
8430 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8431 {
8432         struct inode *inode = dip->inode;
8433         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8434         struct bio *bio;
8435         struct bio *orig_bio = dip->orig_bio;
8436         u64 start_sector = orig_bio->bi_iter.bi_sector;
8437         u64 file_offset = dip->logical_offset;
8438         int async_submit = 0;
8439         u64 submit_len;
8440         int clone_offset = 0;
8441         int clone_len;
8442         int ret;
8443         blk_status_t status;
8444         struct btrfs_io_geometry geom;
8445
8446         submit_len = orig_bio->bi_iter.bi_size;
8447         ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8448                                     start_sector << 9, submit_len, &geom);
8449         if (ret)
8450                 return -EIO;
8451
8452         if (geom.len >= submit_len) {
8453                 bio = orig_bio;
8454                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8455                 goto submit;
8456         }
8457
8458         /* async crcs make it difficult to collect full stripe writes. */
8459         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8460                 async_submit = 0;
8461         else
8462                 async_submit = 1;
8463
8464         /* bio split */
8465         ASSERT(geom.len <= INT_MAX);
8466         atomic_inc(&dip->pending_bios);
8467         do {
8468                 clone_len = min_t(int, submit_len, geom.len);
8469
8470                 /*
8471                  * This will never fail as it's passing GPF_NOFS and
8472                  * the allocation is backed by btrfs_bioset.
8473                  */
8474                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8475                                               clone_len);
8476                 bio->bi_private = dip;
8477                 bio->bi_end_io = btrfs_end_dio_bio;
8478                 btrfs_io_bio(bio)->logical = file_offset;
8479
8480                 ASSERT(submit_len >= clone_len);
8481                 submit_len -= clone_len;
8482                 if (submit_len == 0)
8483                         break;
8484
8485                 /*
8486                  * Increase the count before we submit the bio so we know
8487                  * the end IO handler won't happen before we increase the
8488                  * count. Otherwise, the dip might get freed before we're
8489                  * done setting it up.
8490                  */
8491                 atomic_inc(&dip->pending_bios);
8492
8493                 status = btrfs_submit_dio_bio(bio, inode, file_offset,
8494                                                 async_submit);
8495                 if (status) {
8496                         bio_put(bio);
8497                         atomic_dec(&dip->pending_bios);
8498                         goto out_err;
8499                 }
8500
8501                 clone_offset += clone_len;
8502                 start_sector += clone_len >> 9;
8503                 file_offset += clone_len;
8504
8505                 ret = btrfs_get_io_geometry(fs_info, btrfs_op(orig_bio),
8506                                       start_sector << 9, submit_len, &geom);
8507                 if (ret)
8508                         goto out_err;
8509         } while (submit_len > 0);
8510
8511 submit:
8512         status = btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8513         if (!status)
8514                 return 0;
8515
8516         bio_put(bio);
8517 out_err:
8518         dip->errors = 1;
8519         /*
8520          * Before atomic variable goto zero, we must  make sure dip->errors is
8521          * perceived to be set. This ordering is ensured by the fact that an
8522          * atomic operations with a return value are fully ordered as per
8523          * atomic_t.txt
8524          */
8525         if (atomic_dec_and_test(&dip->pending_bios))
8526                 bio_io_error(dip->orig_bio);
8527
8528         /* bio_end_io() will handle error, so we needn't return it */
8529         return 0;
8530 }
8531
8532 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8533                                 loff_t file_offset)
8534 {
8535         struct btrfs_dio_private *dip = NULL;
8536         struct bio *bio = NULL;
8537         struct btrfs_io_bio *io_bio;
8538         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8539         int ret = 0;
8540
8541         bio = btrfs_bio_clone(dio_bio);
8542
8543         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8544         if (!dip) {
8545                 ret = -ENOMEM;
8546                 goto free_ordered;
8547         }
8548
8549         dip->private = dio_bio->bi_private;
8550         dip->inode = inode;
8551         dip->logical_offset = file_offset;
8552         dip->bytes = dio_bio->bi_iter.bi_size;
8553         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8554         bio->bi_private = dip;
8555         dip->orig_bio = bio;
8556         dip->dio_bio = dio_bio;
8557         atomic_set(&dip->pending_bios, 0);
8558         io_bio = btrfs_io_bio(bio);
8559         io_bio->logical = file_offset;
8560
8561         if (write) {
8562                 bio->bi_end_io = btrfs_endio_direct_write;
8563         } else {
8564                 bio->bi_end_io = btrfs_endio_direct_read;
8565                 dip->subio_endio = btrfs_subio_endio_read;
8566         }
8567
8568         /*
8569          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8570          * even if we fail to submit a bio, because in such case we do the
8571          * corresponding error handling below and it must not be done a second
8572          * time by btrfs_direct_IO().
8573          */
8574         if (write) {
8575                 struct btrfs_dio_data *dio_data = current->journal_info;
8576
8577                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8578                         dip->bytes;
8579                 dio_data->unsubmitted_oe_range_start =
8580                         dio_data->unsubmitted_oe_range_end;
8581         }
8582
8583         ret = btrfs_submit_direct_hook(dip);
8584         if (!ret)
8585                 return;
8586
8587         btrfs_io_bio_free_csum(io_bio);
8588
8589 free_ordered:
8590         /*
8591          * If we arrived here it means either we failed to submit the dip
8592          * or we either failed to clone the dio_bio or failed to allocate the
8593          * dip. If we cloned the dio_bio and allocated the dip, we can just
8594          * call bio_endio against our io_bio so that we get proper resource
8595          * cleanup if we fail to submit the dip, otherwise, we must do the
8596          * same as btrfs_endio_direct_[write|read] because we can't call these
8597          * callbacks - they require an allocated dip and a clone of dio_bio.
8598          */
8599         if (bio && dip) {
8600                 bio_io_error(bio);
8601                 /*
8602                  * The end io callbacks free our dip, do the final put on bio
8603                  * and all the cleanup and final put for dio_bio (through
8604                  * dio_end_io()).
8605                  */
8606                 dip = NULL;
8607                 bio = NULL;
8608         } else {
8609                 if (write)
8610                         __endio_write_update_ordered(inode,
8611                                                 file_offset,
8612                                                 dio_bio->bi_iter.bi_size,
8613                                                 false);
8614                 else
8615                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8616                               file_offset + dio_bio->bi_iter.bi_size - 1);
8617
8618                 dio_bio->bi_status = BLK_STS_IOERR;
8619                 /*
8620                  * Releases and cleans up our dio_bio, no need to bio_put()
8621                  * nor bio_endio()/bio_io_error() against dio_bio.
8622                  */
8623                 dio_end_io(dio_bio);
8624         }
8625         if (bio)
8626                 bio_put(bio);
8627         kfree(dip);
8628 }
8629
8630 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8631                                const struct iov_iter *iter, loff_t offset)
8632 {
8633         int seg;
8634         int i;
8635         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8636         ssize_t retval = -EINVAL;
8637
8638         if (offset & blocksize_mask)
8639                 goto out;
8640
8641         if (iov_iter_alignment(iter) & blocksize_mask)
8642                 goto out;
8643
8644         /* If this is a write we don't need to check anymore */
8645         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8646                 return 0;
8647         /*
8648          * Check to make sure we don't have duplicate iov_base's in this
8649          * iovec, if so return EINVAL, otherwise we'll get csum errors
8650          * when reading back.
8651          */
8652         for (seg = 0; seg < iter->nr_segs; seg++) {
8653                 for (i = seg + 1; i < iter->nr_segs; i++) {
8654                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8655                                 goto out;
8656                 }
8657         }
8658         retval = 0;
8659 out:
8660         return retval;
8661 }
8662
8663 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8664 {
8665         struct file *file = iocb->ki_filp;
8666         struct inode *inode = file->f_mapping->host;
8667         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8668         struct btrfs_dio_data dio_data = { 0 };
8669         struct extent_changeset *data_reserved = NULL;
8670         loff_t offset = iocb->ki_pos;
8671         size_t count = 0;
8672         int flags = 0;
8673         bool wakeup = true;
8674         bool relock = false;
8675         ssize_t ret;
8676
8677         if (check_direct_IO(fs_info, iter, offset))
8678                 return 0;
8679
8680         inode_dio_begin(inode);
8681
8682         /*
8683          * The generic stuff only does filemap_write_and_wait_range, which
8684          * isn't enough if we've written compressed pages to this area, so
8685          * we need to flush the dirty pages again to make absolutely sure
8686          * that any outstanding dirty pages are on disk.
8687          */
8688         count = iov_iter_count(iter);
8689         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8690                      &BTRFS_I(inode)->runtime_flags))
8691                 filemap_fdatawrite_range(inode->i_mapping, offset,
8692                                          offset + count - 1);
8693
8694         if (iov_iter_rw(iter) == WRITE) {
8695                 /*
8696                  * If the write DIO is beyond the EOF, we need update
8697                  * the isize, but it is protected by i_mutex. So we can
8698                  * not unlock the i_mutex at this case.
8699                  */
8700                 if (offset + count <= inode->i_size) {
8701                         dio_data.overwrite = 1;
8702                         inode_unlock(inode);
8703                         relock = true;
8704                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8705                         ret = -EAGAIN;
8706                         goto out;
8707                 }
8708                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8709                                                    offset, count);
8710                 if (ret)
8711                         goto out;
8712
8713                 /*
8714                  * We need to know how many extents we reserved so that we can
8715                  * do the accounting properly if we go over the number we
8716                  * originally calculated.  Abuse current->journal_info for this.
8717                  */
8718                 dio_data.reserve = round_up(count,
8719                                             fs_info->sectorsize);
8720                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8721                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8722                 current->journal_info = &dio_data;
8723                 down_read(&BTRFS_I(inode)->dio_sem);
8724         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8725                                      &BTRFS_I(inode)->runtime_flags)) {
8726                 inode_dio_end(inode);
8727                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8728                 wakeup = false;
8729         }
8730
8731         ret = __blockdev_direct_IO(iocb, inode,
8732                                    fs_info->fs_devices->latest_bdev,
8733                                    iter, btrfs_get_blocks_direct, NULL,
8734                                    btrfs_submit_direct, flags);
8735         if (iov_iter_rw(iter) == WRITE) {
8736                 up_read(&BTRFS_I(inode)->dio_sem);
8737                 current->journal_info = NULL;
8738                 if (ret < 0 && ret != -EIOCBQUEUED) {
8739                         if (dio_data.reserve)
8740                                 btrfs_delalloc_release_space(inode, data_reserved,
8741                                         offset, dio_data.reserve, true);
8742                         /*
8743                          * On error we might have left some ordered extents
8744                          * without submitting corresponding bios for them, so
8745                          * cleanup them up to avoid other tasks getting them
8746                          * and waiting for them to complete forever.
8747                          */
8748                         if (dio_data.unsubmitted_oe_range_start <
8749                             dio_data.unsubmitted_oe_range_end)
8750                                 __endio_write_update_ordered(inode,
8751                                         dio_data.unsubmitted_oe_range_start,
8752                                         dio_data.unsubmitted_oe_range_end -
8753                                         dio_data.unsubmitted_oe_range_start,
8754                                         false);
8755                 } else if (ret >= 0 && (size_t)ret < count)
8756                         btrfs_delalloc_release_space(inode, data_reserved,
8757                                         offset, count - (size_t)ret, true);
8758                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8759         }
8760 out:
8761         if (wakeup)
8762                 inode_dio_end(inode);
8763         if (relock)
8764                 inode_lock(inode);
8765
8766         extent_changeset_free(data_reserved);
8767         return ret;
8768 }
8769
8770 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8771
8772 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8773                 __u64 start, __u64 len)
8774 {
8775         int     ret;
8776
8777         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8778         if (ret)
8779                 return ret;
8780
8781         return extent_fiemap(inode, fieinfo, start, len);
8782 }
8783
8784 int btrfs_readpage(struct file *file, struct page *page)
8785 {
8786         struct extent_io_tree *tree;
8787         tree = &BTRFS_I(page->mapping->host)->io_tree;
8788         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8789 }
8790
8791 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8792 {
8793         struct inode *inode = page->mapping->host;
8794         int ret;
8795
8796         if (current->flags & PF_MEMALLOC) {
8797                 redirty_page_for_writepage(wbc, page);
8798                 unlock_page(page);
8799                 return 0;
8800         }
8801
8802         /*
8803          * If we are under memory pressure we will call this directly from the
8804          * VM, we need to make sure we have the inode referenced for the ordered
8805          * extent.  If not just return like we didn't do anything.
8806          */
8807         if (!igrab(inode)) {
8808                 redirty_page_for_writepage(wbc, page);
8809                 return AOP_WRITEPAGE_ACTIVATE;
8810         }
8811         ret = extent_write_full_page(page, wbc);
8812         btrfs_add_delayed_iput(inode);
8813         return ret;
8814 }
8815
8816 static int btrfs_writepages(struct address_space *mapping,
8817                             struct writeback_control *wbc)
8818 {
8819         return extent_writepages(mapping, wbc);
8820 }
8821
8822 static int
8823 btrfs_readpages(struct file *file, struct address_space *mapping,
8824                 struct list_head *pages, unsigned nr_pages)
8825 {
8826         return extent_readpages(mapping, pages, nr_pages);
8827 }
8828
8829 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8830 {
8831         int ret = try_release_extent_mapping(page, gfp_flags);
8832         if (ret == 1) {
8833                 ClearPagePrivate(page);
8834                 set_page_private(page, 0);
8835                 put_page(page);
8836         }
8837         return ret;
8838 }
8839
8840 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8841 {
8842         if (PageWriteback(page) || PageDirty(page))
8843                 return 0;
8844         return __btrfs_releasepage(page, gfp_flags);
8845 }
8846
8847 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8848                                  unsigned int length)
8849 {
8850         struct inode *inode = page->mapping->host;
8851         struct extent_io_tree *tree;
8852         struct btrfs_ordered_extent *ordered;
8853         struct extent_state *cached_state = NULL;
8854         u64 page_start = page_offset(page);
8855         u64 page_end = page_start + PAGE_SIZE - 1;
8856         u64 start;
8857         u64 end;
8858         int inode_evicting = inode->i_state & I_FREEING;
8859
8860         /*
8861          * we have the page locked, so new writeback can't start,
8862          * and the dirty bit won't be cleared while we are here.
8863          *
8864          * Wait for IO on this page so that we can safely clear
8865          * the PagePrivate2 bit and do ordered accounting
8866          */
8867         wait_on_page_writeback(page);
8868
8869         tree = &BTRFS_I(inode)->io_tree;
8870         if (offset) {
8871                 btrfs_releasepage(page, GFP_NOFS);
8872                 return;
8873         }
8874
8875         if (!inode_evicting)
8876                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8877 again:
8878         start = page_start;
8879         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8880                                         page_end - start + 1);
8881         if (ordered) {
8882                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8883                 /*
8884                  * IO on this page will never be started, so we need
8885                  * to account for any ordered extents now
8886                  */
8887                 if (!inode_evicting)
8888                         clear_extent_bit(tree, start, end,
8889                                          EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8890                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8891                                          EXTENT_DEFRAG, 1, 0, &cached_state);
8892                 /*
8893                  * whoever cleared the private bit is responsible
8894                  * for the finish_ordered_io
8895                  */
8896                 if (TestClearPagePrivate2(page)) {
8897                         struct btrfs_ordered_inode_tree *tree;
8898                         u64 new_len;
8899
8900                         tree = &BTRFS_I(inode)->ordered_tree;
8901
8902                         spin_lock_irq(&tree->lock);
8903                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8904                         new_len = start - ordered->file_offset;
8905                         if (new_len < ordered->truncated_len)
8906                                 ordered->truncated_len = new_len;
8907                         spin_unlock_irq(&tree->lock);
8908
8909                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8910                                                            start,
8911                                                            end - start + 1, 1))
8912                                 btrfs_finish_ordered_io(ordered);
8913                 }
8914                 btrfs_put_ordered_extent(ordered);
8915                 if (!inode_evicting) {
8916                         cached_state = NULL;
8917                         lock_extent_bits(tree, start, end,
8918                                          &cached_state);
8919                 }
8920
8921                 start = end + 1;
8922                 if (start < page_end)
8923                         goto again;
8924         }
8925
8926         /*
8927          * Qgroup reserved space handler
8928          * Page here will be either
8929          * 1) Already written to disk
8930          *    In this case, its reserved space is released from data rsv map
8931          *    and will be freed by delayed_ref handler finally.
8932          *    So even we call qgroup_free_data(), it won't decrease reserved
8933          *    space.
8934          * 2) Not written to disk
8935          *    This means the reserved space should be freed here. However,
8936          *    if a truncate invalidates the page (by clearing PageDirty)
8937          *    and the page is accounted for while allocating extent
8938          *    in btrfs_check_data_free_space() we let delayed_ref to
8939          *    free the entire extent.
8940          */
8941         if (PageDirty(page))
8942                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
8943         if (!inode_evicting) {
8944                 clear_extent_bit(tree, page_start, page_end, EXTENT_LOCKED |
8945                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
8946                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
8947                                  &cached_state);
8948
8949                 __btrfs_releasepage(page, GFP_NOFS);
8950         }
8951
8952         ClearPageChecked(page);
8953         if (PagePrivate(page)) {
8954                 ClearPagePrivate(page);
8955                 set_page_private(page, 0);
8956                 put_page(page);
8957         }
8958 }
8959
8960 /*
8961  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8962  * called from a page fault handler when a page is first dirtied. Hence we must
8963  * be careful to check for EOF conditions here. We set the page up correctly
8964  * for a written page which means we get ENOSPC checking when writing into
8965  * holes and correct delalloc and unwritten extent mapping on filesystems that
8966  * support these features.
8967  *
8968  * We are not allowed to take the i_mutex here so we have to play games to
8969  * protect against truncate races as the page could now be beyond EOF.  Because
8970  * truncate_setsize() writes the inode size before removing pages, once we have
8971  * the page lock we can determine safely if the page is beyond EOF. If it is not
8972  * beyond EOF, then the page is guaranteed safe against truncation until we
8973  * unlock the page.
8974  */
8975 vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
8976 {
8977         struct page *page = vmf->page;
8978         struct inode *inode = file_inode(vmf->vma->vm_file);
8979         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8980         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8981         struct btrfs_ordered_extent *ordered;
8982         struct extent_state *cached_state = NULL;
8983         struct extent_changeset *data_reserved = NULL;
8984         char *kaddr;
8985         unsigned long zero_start;
8986         loff_t size;
8987         vm_fault_t ret;
8988         int ret2;
8989         int reserved = 0;
8990         u64 reserved_space;
8991         u64 page_start;
8992         u64 page_end;
8993         u64 end;
8994
8995         reserved_space = PAGE_SIZE;
8996
8997         sb_start_pagefault(inode->i_sb);
8998         page_start = page_offset(page);
8999         page_end = page_start + PAGE_SIZE - 1;
9000         end = page_end;
9001
9002         /*
9003          * Reserving delalloc space after obtaining the page lock can lead to
9004          * deadlock. For example, if a dirty page is locked by this function
9005          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9006          * dirty page write out, then the btrfs_writepage() function could
9007          * end up waiting indefinitely to get a lock on the page currently
9008          * being processed by btrfs_page_mkwrite() function.
9009          */
9010         ret2 = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9011                                            reserved_space);
9012         if (!ret2) {
9013                 ret2 = file_update_time(vmf->vma->vm_file);
9014                 reserved = 1;
9015         }
9016         if (ret2) {
9017                 ret = vmf_error(ret2);
9018                 if (reserved)
9019                         goto out;
9020                 goto out_noreserve;
9021         }
9022
9023         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9024 again:
9025         lock_page(page);
9026         size = i_size_read(inode);
9027
9028         if ((page->mapping != inode->i_mapping) ||
9029             (page_start >= size)) {
9030                 /* page got truncated out from underneath us */
9031                 goto out_unlock;
9032         }
9033         wait_on_page_writeback(page);
9034
9035         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9036         set_page_extent_mapped(page);
9037
9038         /*
9039          * we can't set the delalloc bits if there are pending ordered
9040          * extents.  Drop our locks and wait for them to finish
9041          */
9042         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9043                         PAGE_SIZE);
9044         if (ordered) {
9045                 unlock_extent_cached(io_tree, page_start, page_end,
9046                                      &cached_state);
9047                 unlock_page(page);
9048                 btrfs_start_ordered_extent(inode, ordered, 1);
9049                 btrfs_put_ordered_extent(ordered);
9050                 goto again;
9051         }
9052
9053         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9054                 reserved_space = round_up(size - page_start,
9055                                           fs_info->sectorsize);
9056                 if (reserved_space < PAGE_SIZE) {
9057                         end = page_start + reserved_space - 1;
9058                         btrfs_delalloc_release_space(inode, data_reserved,
9059                                         page_start, PAGE_SIZE - reserved_space,
9060                                         true);
9061                 }
9062         }
9063
9064         /*
9065          * page_mkwrite gets called when the page is firstly dirtied after it's
9066          * faulted in, but write(2) could also dirty a page and set delalloc
9067          * bits, thus in this case for space account reason, we still need to
9068          * clear any delalloc bits within this page range since we have to
9069          * reserve data&meta space before lock_page() (see above comments).
9070          */
9071         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9072                           EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
9073                           EXTENT_DEFRAG, 0, 0, &cached_state);
9074
9075         ret2 = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9076                                         &cached_state);
9077         if (ret2) {
9078                 unlock_extent_cached(io_tree, page_start, page_end,
9079                                      &cached_state);
9080                 ret = VM_FAULT_SIGBUS;
9081                 goto out_unlock;
9082         }
9083         ret2 = 0;
9084
9085         /* page is wholly or partially inside EOF */
9086         if (page_start + PAGE_SIZE > size)
9087                 zero_start = offset_in_page(size);
9088         else
9089                 zero_start = PAGE_SIZE;
9090
9091         if (zero_start != PAGE_SIZE) {
9092                 kaddr = kmap(page);
9093                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9094                 flush_dcache_page(page);
9095                 kunmap(page);
9096         }
9097         ClearPageChecked(page);
9098         set_page_dirty(page);
9099         SetPageUptodate(page);
9100
9101         BTRFS_I(inode)->last_trans = fs_info->generation;
9102         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9103         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9104
9105         unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
9106
9107         if (!ret2) {
9108                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9109                 sb_end_pagefault(inode->i_sb);
9110                 extent_changeset_free(data_reserved);
9111                 return VM_FAULT_LOCKED;
9112         }
9113
9114 out_unlock:
9115         unlock_page(page);
9116 out:
9117         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9118         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9119                                      reserved_space, (ret != 0));
9120 out_noreserve:
9121         sb_end_pagefault(inode->i_sb);
9122         extent_changeset_free(data_reserved);
9123         return ret;
9124 }
9125
9126 static int btrfs_truncate(struct inode *inode, bool skip_writeback)
9127 {
9128         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9129         struct btrfs_root *root = BTRFS_I(inode)->root;
9130         struct btrfs_block_rsv *rsv;
9131         int ret;
9132         struct btrfs_trans_handle *trans;
9133         u64 mask = fs_info->sectorsize - 1;
9134         u64 min_size = btrfs_calc_metadata_size(fs_info, 1);
9135
9136         if (!skip_writeback) {
9137                 ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9138                                                (u64)-1);
9139                 if (ret)
9140                         return ret;
9141         }
9142
9143         /*
9144          * Yes ladies and gentlemen, this is indeed ugly.  We have a couple of
9145          * things going on here:
9146          *
9147          * 1) We need to reserve space to update our inode.
9148          *
9149          * 2) We need to have something to cache all the space that is going to
9150          * be free'd up by the truncate operation, but also have some slack
9151          * space reserved in case it uses space during the truncate (thank you
9152          * very much snapshotting).
9153          *
9154          * And we need these to be separate.  The fact is we can use a lot of
9155          * space doing the truncate, and we have no earthly idea how much space
9156          * we will use, so we need the truncate reservation to be separate so it
9157          * doesn't end up using space reserved for updating the inode.  We also
9158          * need to be able to stop the transaction and start a new one, which
9159          * means we need to be able to update the inode several times, and we
9160          * have no idea of knowing how many times that will be, so we can't just
9161          * reserve 1 item for the entirety of the operation, so that has to be
9162          * done separately as well.
9163          *
9164          * So that leaves us with
9165          *
9166          * 1) rsv - for the truncate reservation, which we will steal from the
9167          * transaction reservation.
9168          * 2) fs_info->trans_block_rsv - this will have 1 items worth left for
9169          * updating the inode.
9170          */
9171         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9172         if (!rsv)
9173                 return -ENOMEM;
9174         rsv->size = min_size;
9175         rsv->failfast = 1;
9176
9177         /*
9178          * 1 for the truncate slack space
9179          * 1 for updating the inode.
9180          */
9181         trans = btrfs_start_transaction(root, 2);
9182         if (IS_ERR(trans)) {
9183                 ret = PTR_ERR(trans);
9184                 goto out;
9185         }
9186
9187         /* Migrate the slack space for the truncate to our reserve */
9188         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9189                                       min_size, false);
9190         BUG_ON(ret);
9191
9192         /*
9193          * So if we truncate and then write and fsync we normally would just
9194          * write the extents that changed, which is a problem if we need to
9195          * first truncate that entire inode.  So set this flag so we write out
9196          * all of the extents in the inode to the sync log so we're completely
9197          * safe.
9198          */
9199         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9200         trans->block_rsv = rsv;
9201
9202         while (1) {
9203                 ret = btrfs_truncate_inode_items(trans, root, inode,
9204                                                  inode->i_size,
9205                                                  BTRFS_EXTENT_DATA_KEY);
9206                 trans->block_rsv = &fs_info->trans_block_rsv;
9207                 if (ret != -ENOSPC && ret != -EAGAIN)
9208                         break;
9209
9210                 ret = btrfs_update_inode(trans, root, inode);
9211                 if (ret)
9212                         break;
9213
9214                 btrfs_end_transaction(trans);
9215                 btrfs_btree_balance_dirty(fs_info);
9216
9217                 trans = btrfs_start_transaction(root, 2);
9218                 if (IS_ERR(trans)) {
9219                         ret = PTR_ERR(trans);
9220                         trans = NULL;
9221                         break;
9222                 }
9223
9224                 btrfs_block_rsv_release(fs_info, rsv, -1);
9225                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9226                                               rsv, min_size, false);
9227                 BUG_ON(ret);    /* shouldn't happen */
9228                 trans->block_rsv = rsv;
9229         }
9230
9231         /*
9232          * We can't call btrfs_truncate_block inside a trans handle as we could
9233          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9234          * we've truncated everything except the last little bit, and can do
9235          * btrfs_truncate_block and then update the disk_i_size.
9236          */
9237         if (ret == NEED_TRUNCATE_BLOCK) {
9238                 btrfs_end_transaction(trans);
9239                 btrfs_btree_balance_dirty(fs_info);
9240
9241                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9242                 if (ret)
9243                         goto out;
9244                 trans = btrfs_start_transaction(root, 1);
9245                 if (IS_ERR(trans)) {
9246                         ret = PTR_ERR(trans);
9247                         goto out;
9248                 }
9249                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9250         }
9251
9252         if (trans) {
9253                 int ret2;
9254
9255                 trans->block_rsv = &fs_info->trans_block_rsv;
9256                 ret2 = btrfs_update_inode(trans, root, inode);
9257                 if (ret2 && !ret)
9258                         ret = ret2;
9259
9260                 ret2 = btrfs_end_transaction(trans);
9261                 if (ret2 && !ret)
9262                         ret = ret2;
9263                 btrfs_btree_balance_dirty(fs_info);
9264         }
9265 out:
9266         btrfs_free_block_rsv(fs_info, rsv);
9267
9268         return ret;
9269 }
9270
9271 /*
9272  * create a new subvolume directory/inode (helper for the ioctl).
9273  */
9274 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9275                              struct btrfs_root *new_root,
9276                              struct btrfs_root *parent_root,
9277                              u64 new_dirid)
9278 {
9279         struct inode *inode;
9280         int err;
9281         u64 index = 0;
9282
9283         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9284                                 new_dirid, new_dirid,
9285                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9286                                 &index);
9287         if (IS_ERR(inode))
9288                 return PTR_ERR(inode);
9289         inode->i_op = &btrfs_dir_inode_operations;
9290         inode->i_fop = &btrfs_dir_file_operations;
9291
9292         set_nlink(inode, 1);
9293         btrfs_i_size_write(BTRFS_I(inode), 0);
9294         unlock_new_inode(inode);
9295
9296         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9297         if (err)
9298                 btrfs_err(new_root->fs_info,
9299                           "error inheriting subvolume %llu properties: %d",
9300                           new_root->root_key.objectid, err);
9301
9302         err = btrfs_update_inode(trans, new_root, inode);
9303
9304         iput(inode);
9305         return err;
9306 }
9307
9308 struct inode *btrfs_alloc_inode(struct super_block *sb)
9309 {
9310         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9311         struct btrfs_inode *ei;
9312         struct inode *inode;
9313
9314         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_KERNEL);
9315         if (!ei)
9316                 return NULL;
9317
9318         ei->root = NULL;
9319         ei->generation = 0;
9320         ei->last_trans = 0;
9321         ei->last_sub_trans = 0;
9322         ei->logged_trans = 0;
9323         ei->delalloc_bytes = 0;
9324         ei->new_delalloc_bytes = 0;
9325         ei->defrag_bytes = 0;
9326         ei->disk_i_size = 0;
9327         ei->flags = 0;
9328         ei->csum_bytes = 0;
9329         ei->index_cnt = (u64)-1;
9330         ei->dir_index = 0;
9331         ei->last_unlink_trans = 0;
9332         ei->last_log_commit = 0;
9333
9334         spin_lock_init(&ei->lock);
9335         ei->outstanding_extents = 0;
9336         if (sb->s_magic != BTRFS_TEST_MAGIC)
9337                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9338                                               BTRFS_BLOCK_RSV_DELALLOC);
9339         ei->runtime_flags = 0;
9340         ei->prop_compress = BTRFS_COMPRESS_NONE;
9341         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9342
9343         ei->delayed_node = NULL;
9344
9345         ei->i_otime.tv_sec = 0;
9346         ei->i_otime.tv_nsec = 0;
9347
9348         inode = &ei->vfs_inode;
9349         extent_map_tree_init(&ei->extent_tree);
9350         extent_io_tree_init(fs_info, &ei->io_tree, IO_TREE_INODE_IO, inode);
9351         extent_io_tree_init(fs_info, &ei->io_failure_tree,
9352                             IO_TREE_INODE_IO_FAILURE, inode);
9353         ei->io_tree.track_uptodate = true;
9354         ei->io_failure_tree.track_uptodate = true;
9355         atomic_set(&ei->sync_writers, 0);
9356         mutex_init(&ei->log_mutex);
9357         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9358         INIT_LIST_HEAD(&ei->delalloc_inodes);
9359         INIT_LIST_HEAD(&ei->delayed_iput);
9360         RB_CLEAR_NODE(&ei->rb_node);
9361         init_rwsem(&ei->dio_sem);
9362
9363         return inode;
9364 }
9365
9366 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9367 void btrfs_test_destroy_inode(struct inode *inode)
9368 {
9369         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9370         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9371 }
9372 #endif
9373
9374 void btrfs_free_inode(struct inode *inode)
9375 {
9376         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9377 }
9378
9379 void btrfs_destroy_inode(struct inode *inode)
9380 {
9381         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9382         struct btrfs_ordered_extent *ordered;
9383         struct btrfs_root *root = BTRFS_I(inode)->root;
9384
9385         WARN_ON(!hlist_empty(&inode->i_dentry));
9386         WARN_ON(inode->i_data.nrpages);
9387         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9388         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9389         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9390         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9391         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9392         WARN_ON(BTRFS_I(inode)->csum_bytes);
9393         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9394
9395         /*
9396          * This can happen where we create an inode, but somebody else also
9397          * created the same inode and we need to destroy the one we already
9398          * created.
9399          */
9400         if (!root)
9401                 return;
9402
9403         while (1) {
9404                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9405                 if (!ordered)
9406                         break;
9407                 else {
9408                         btrfs_err(fs_info,
9409                                   "found ordered extent %llu %llu on inode cleanup",
9410                                   ordered->file_offset, ordered->len);
9411                         btrfs_remove_ordered_extent(inode, ordered);
9412                         btrfs_put_ordered_extent(ordered);
9413                         btrfs_put_ordered_extent(ordered);
9414                 }
9415         }
9416         btrfs_qgroup_check_reserved_leak(inode);
9417         inode_tree_del(inode);
9418         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9419 }
9420
9421 int btrfs_drop_inode(struct inode *inode)
9422 {
9423         struct btrfs_root *root = BTRFS_I(inode)->root;
9424
9425         if (root == NULL)
9426                 return 1;
9427
9428         /* the snap/subvol tree is on deleting */
9429         if (btrfs_root_refs(&root->root_item) == 0)
9430                 return 1;
9431         else
9432                 return generic_drop_inode(inode);
9433 }
9434
9435 static void init_once(void *foo)
9436 {
9437         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9438
9439         inode_init_once(&ei->vfs_inode);
9440 }
9441
9442 void __cold btrfs_destroy_cachep(void)
9443 {
9444         /*
9445          * Make sure all delayed rcu free inodes are flushed before we
9446          * destroy cache.
9447          */
9448         rcu_barrier();
9449         kmem_cache_destroy(btrfs_inode_cachep);
9450         kmem_cache_destroy(btrfs_trans_handle_cachep);
9451         kmem_cache_destroy(btrfs_path_cachep);
9452         kmem_cache_destroy(btrfs_free_space_cachep);
9453         kmem_cache_destroy(btrfs_free_space_bitmap_cachep);
9454 }
9455
9456 int __init btrfs_init_cachep(void)
9457 {
9458         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9459                         sizeof(struct btrfs_inode), 0,
9460                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9461                         init_once);
9462         if (!btrfs_inode_cachep)
9463                 goto fail;
9464
9465         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9466                         sizeof(struct btrfs_trans_handle), 0,
9467                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9468         if (!btrfs_trans_handle_cachep)
9469                 goto fail;
9470
9471         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9472                         sizeof(struct btrfs_path), 0,
9473                         SLAB_MEM_SPREAD, NULL);
9474         if (!btrfs_path_cachep)
9475                 goto fail;
9476
9477         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9478                         sizeof(struct btrfs_free_space), 0,
9479                         SLAB_MEM_SPREAD, NULL);
9480         if (!btrfs_free_space_cachep)
9481                 goto fail;
9482
9483         btrfs_free_space_bitmap_cachep = kmem_cache_create("btrfs_free_space_bitmap",
9484                                                         PAGE_SIZE, PAGE_SIZE,
9485                                                         SLAB_RED_ZONE, NULL);
9486         if (!btrfs_free_space_bitmap_cachep)
9487                 goto fail;
9488
9489         return 0;
9490 fail:
9491         btrfs_destroy_cachep();
9492         return -ENOMEM;
9493 }
9494
9495 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9496                          u32 request_mask, unsigned int flags)
9497 {
9498         u64 delalloc_bytes;
9499         struct inode *inode = d_inode(path->dentry);
9500         u32 blocksize = inode->i_sb->s_blocksize;
9501         u32 bi_flags = BTRFS_I(inode)->flags;
9502
9503         stat->result_mask |= STATX_BTIME;
9504         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9505         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9506         if (bi_flags & BTRFS_INODE_APPEND)
9507                 stat->attributes |= STATX_ATTR_APPEND;
9508         if (bi_flags & BTRFS_INODE_COMPRESS)
9509                 stat->attributes |= STATX_ATTR_COMPRESSED;
9510         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9511                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9512         if (bi_flags & BTRFS_INODE_NODUMP)
9513                 stat->attributes |= STATX_ATTR_NODUMP;
9514
9515         stat->attributes_mask |= (STATX_ATTR_APPEND |
9516                                   STATX_ATTR_COMPRESSED |
9517                                   STATX_ATTR_IMMUTABLE |
9518                                   STATX_ATTR_NODUMP);
9519
9520         generic_fillattr(inode, stat);
9521         stat->dev = BTRFS_I(inode)->root->anon_dev;
9522
9523         spin_lock(&BTRFS_I(inode)->lock);
9524         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9525         spin_unlock(&BTRFS_I(inode)->lock);
9526         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9527                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9528         return 0;
9529 }
9530
9531 static int btrfs_rename_exchange(struct inode *old_dir,
9532                               struct dentry *old_dentry,
9533                               struct inode *new_dir,
9534                               struct dentry *new_dentry)
9535 {
9536         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9537         struct btrfs_trans_handle *trans;
9538         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9539         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9540         struct inode *new_inode = new_dentry->d_inode;
9541         struct inode *old_inode = old_dentry->d_inode;
9542         struct timespec64 ctime = current_time(old_inode);
9543         struct dentry *parent;
9544         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9545         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9546         u64 old_idx = 0;
9547         u64 new_idx = 0;
9548         u64 root_objectid;
9549         int ret;
9550         bool root_log_pinned = false;
9551         bool dest_log_pinned = false;
9552         struct btrfs_log_ctx ctx_root;
9553         struct btrfs_log_ctx ctx_dest;
9554         bool sync_log_root = false;
9555         bool sync_log_dest = false;
9556         bool commit_transaction = false;
9557
9558         /* we only allow rename subvolume link between subvolumes */
9559         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9560                 return -EXDEV;
9561
9562         btrfs_init_log_ctx(&ctx_root, old_inode);
9563         btrfs_init_log_ctx(&ctx_dest, new_inode);
9564
9565         /* close the race window with snapshot create/destroy ioctl */
9566         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9567                 down_read(&fs_info->subvol_sem);
9568         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9569                 down_read(&fs_info->subvol_sem);
9570
9571         /*
9572          * We want to reserve the absolute worst case amount of items.  So if
9573          * both inodes are subvols and we need to unlink them then that would
9574          * require 4 item modifications, but if they are both normal inodes it
9575          * would require 5 item modifications, so we'll assume their normal
9576          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9577          * should cover the worst case number of items we'll modify.
9578          */
9579         trans = btrfs_start_transaction(root, 12);
9580         if (IS_ERR(trans)) {
9581                 ret = PTR_ERR(trans);
9582                 goto out_notrans;
9583         }
9584
9585         /*
9586          * We need to find a free sequence number both in the source and
9587          * in the destination directory for the exchange.
9588          */
9589         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9590         if (ret)
9591                 goto out_fail;
9592         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9593         if (ret)
9594                 goto out_fail;
9595
9596         BTRFS_I(old_inode)->dir_index = 0ULL;
9597         BTRFS_I(new_inode)->dir_index = 0ULL;
9598
9599         /* Reference for the source. */
9600         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9601                 /* force full log commit if subvolume involved. */
9602                 btrfs_set_log_full_commit(trans);
9603         } else {
9604                 btrfs_pin_log_trans(root);
9605                 root_log_pinned = true;
9606                 ret = btrfs_insert_inode_ref(trans, dest,
9607                                              new_dentry->d_name.name,
9608                                              new_dentry->d_name.len,
9609                                              old_ino,
9610                                              btrfs_ino(BTRFS_I(new_dir)),
9611                                              old_idx);
9612                 if (ret)
9613                         goto out_fail;
9614         }
9615
9616         /* And now for the dest. */
9617         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9618                 /* force full log commit if subvolume involved. */
9619                 btrfs_set_log_full_commit(trans);
9620         } else {
9621                 btrfs_pin_log_trans(dest);
9622                 dest_log_pinned = true;
9623                 ret = btrfs_insert_inode_ref(trans, root,
9624                                              old_dentry->d_name.name,
9625                                              old_dentry->d_name.len,
9626                                              new_ino,
9627                                              btrfs_ino(BTRFS_I(old_dir)),
9628                                              new_idx);
9629                 if (ret)
9630                         goto out_fail;
9631         }
9632
9633         /* Update inode version and ctime/mtime. */
9634         inode_inc_iversion(old_dir);
9635         inode_inc_iversion(new_dir);
9636         inode_inc_iversion(old_inode);
9637         inode_inc_iversion(new_inode);
9638         old_dir->i_ctime = old_dir->i_mtime = ctime;
9639         new_dir->i_ctime = new_dir->i_mtime = ctime;
9640         old_inode->i_ctime = ctime;
9641         new_inode->i_ctime = ctime;
9642
9643         if (old_dentry->d_parent != new_dentry->d_parent) {
9644                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9645                                 BTRFS_I(old_inode), 1);
9646                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9647                                 BTRFS_I(new_inode), 1);
9648         }
9649
9650         /* src is a subvolume */
9651         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9652                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9653                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9654                                           old_dentry->d_name.name,
9655                                           old_dentry->d_name.len);
9656         } else { /* src is an inode */
9657                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9658                                            BTRFS_I(old_dentry->d_inode),
9659                                            old_dentry->d_name.name,
9660                                            old_dentry->d_name.len);
9661                 if (!ret)
9662                         ret = btrfs_update_inode(trans, root, old_inode);
9663         }
9664         if (ret) {
9665                 btrfs_abort_transaction(trans, ret);
9666                 goto out_fail;
9667         }
9668
9669         /* dest is a subvolume */
9670         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9671                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9672                 ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
9673                                           new_dentry->d_name.name,
9674                                           new_dentry->d_name.len);
9675         } else { /* dest is an inode */
9676                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9677                                            BTRFS_I(new_dentry->d_inode),
9678                                            new_dentry->d_name.name,
9679                                            new_dentry->d_name.len);
9680                 if (!ret)
9681                         ret = btrfs_update_inode(trans, dest, new_inode);
9682         }
9683         if (ret) {
9684                 btrfs_abort_transaction(trans, ret);
9685                 goto out_fail;
9686         }
9687
9688         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9689                              new_dentry->d_name.name,
9690                              new_dentry->d_name.len, 0, old_idx);
9691         if (ret) {
9692                 btrfs_abort_transaction(trans, ret);
9693                 goto out_fail;
9694         }
9695
9696         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9697                              old_dentry->d_name.name,
9698                              old_dentry->d_name.len, 0, new_idx);
9699         if (ret) {
9700                 btrfs_abort_transaction(trans, ret);
9701                 goto out_fail;
9702         }
9703
9704         if (old_inode->i_nlink == 1)
9705                 BTRFS_I(old_inode)->dir_index = old_idx;
9706         if (new_inode->i_nlink == 1)
9707                 BTRFS_I(new_inode)->dir_index = new_idx;
9708
9709         if (root_log_pinned) {
9710                 parent = new_dentry->d_parent;
9711                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
9712                                          BTRFS_I(old_dir), parent,
9713                                          false, &ctx_root);
9714                 if (ret == BTRFS_NEED_LOG_SYNC)
9715                         sync_log_root = true;
9716                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
9717                         commit_transaction = true;
9718                 ret = 0;
9719                 btrfs_end_log_trans(root);
9720                 root_log_pinned = false;
9721         }
9722         if (dest_log_pinned) {
9723                 if (!commit_transaction) {
9724                         parent = old_dentry->d_parent;
9725                         ret = btrfs_log_new_name(trans, BTRFS_I(new_inode),
9726                                                  BTRFS_I(new_dir), parent,
9727                                                  false, &ctx_dest);
9728                         if (ret == BTRFS_NEED_LOG_SYNC)
9729                                 sync_log_dest = true;
9730                         else if (ret == BTRFS_NEED_TRANS_COMMIT)
9731                                 commit_transaction = true;
9732                         ret = 0;
9733                 }
9734                 btrfs_end_log_trans(dest);
9735                 dest_log_pinned = false;
9736         }
9737 out_fail:
9738         /*
9739          * If we have pinned a log and an error happened, we unpin tasks
9740          * trying to sync the log and force them to fallback to a transaction
9741          * commit if the log currently contains any of the inodes involved in
9742          * this rename operation (to ensure we do not persist a log with an
9743          * inconsistent state for any of these inodes or leading to any
9744          * inconsistencies when replayed). If the transaction was aborted, the
9745          * abortion reason is propagated to userspace when attempting to commit
9746          * the transaction. If the log does not contain any of these inodes, we
9747          * allow the tasks to sync it.
9748          */
9749         if (ret && (root_log_pinned || dest_log_pinned)) {
9750                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9751                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9752                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9753                     (new_inode &&
9754                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9755                         btrfs_set_log_full_commit(trans);
9756
9757                 if (root_log_pinned) {
9758                         btrfs_end_log_trans(root);
9759                         root_log_pinned = false;
9760                 }
9761                 if (dest_log_pinned) {
9762                         btrfs_end_log_trans(dest);
9763                         dest_log_pinned = false;
9764                 }
9765         }
9766         if (!ret && sync_log_root && !commit_transaction) {
9767                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root,
9768                                      &ctx_root);
9769                 if (ret)
9770                         commit_transaction = true;
9771         }
9772         if (!ret && sync_log_dest && !commit_transaction) {
9773                 ret = btrfs_sync_log(trans, BTRFS_I(new_inode)->root,
9774                                      &ctx_dest);
9775                 if (ret)
9776                         commit_transaction = true;
9777         }
9778         if (commit_transaction) {
9779                 /*
9780                  * We may have set commit_transaction when logging the new name
9781                  * in the destination root, in which case we left the source
9782                  * root context in the list of log contextes. So make sure we
9783                  * remove it to avoid invalid memory accesses, since the context
9784                  * was allocated in our stack frame.
9785                  */
9786                 if (sync_log_root) {
9787                         mutex_lock(&root->log_mutex);
9788                         list_del_init(&ctx_root.list);
9789                         mutex_unlock(&root->log_mutex);
9790                 }
9791                 ret = btrfs_commit_transaction(trans);
9792         } else {
9793                 int ret2;
9794
9795                 ret2 = btrfs_end_transaction(trans);
9796                 ret = ret ? ret : ret2;
9797         }
9798 out_notrans:
9799         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9800                 up_read(&fs_info->subvol_sem);
9801         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9802                 up_read(&fs_info->subvol_sem);
9803
9804         ASSERT(list_empty(&ctx_root.list));
9805         ASSERT(list_empty(&ctx_dest.list));
9806
9807         return ret;
9808 }
9809
9810 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9811                                      struct btrfs_root *root,
9812                                      struct inode *dir,
9813                                      struct dentry *dentry)
9814 {
9815         int ret;
9816         struct inode *inode;
9817         u64 objectid;
9818         u64 index;
9819
9820         ret = btrfs_find_free_ino(root, &objectid);
9821         if (ret)
9822                 return ret;
9823
9824         inode = btrfs_new_inode(trans, root, dir,
9825                                 dentry->d_name.name,
9826                                 dentry->d_name.len,
9827                                 btrfs_ino(BTRFS_I(dir)),
9828                                 objectid,
9829                                 S_IFCHR | WHITEOUT_MODE,
9830                                 &index);
9831
9832         if (IS_ERR(inode)) {
9833                 ret = PTR_ERR(inode);
9834                 return ret;
9835         }
9836
9837         inode->i_op = &btrfs_special_inode_operations;
9838         init_special_inode(inode, inode->i_mode,
9839                 WHITEOUT_DEV);
9840
9841         ret = btrfs_init_inode_security(trans, inode, dir,
9842                                 &dentry->d_name);
9843         if (ret)
9844                 goto out;
9845
9846         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9847                                 BTRFS_I(inode), 0, index);
9848         if (ret)
9849                 goto out;
9850
9851         ret = btrfs_update_inode(trans, root, inode);
9852 out:
9853         unlock_new_inode(inode);
9854         if (ret)
9855                 inode_dec_link_count(inode);
9856         iput(inode);
9857
9858         return ret;
9859 }
9860
9861 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9862                            struct inode *new_dir, struct dentry *new_dentry,
9863                            unsigned int flags)
9864 {
9865         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9866         struct btrfs_trans_handle *trans;
9867         unsigned int trans_num_items;
9868         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9869         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9870         struct inode *new_inode = d_inode(new_dentry);
9871         struct inode *old_inode = d_inode(old_dentry);
9872         u64 index = 0;
9873         u64 root_objectid;
9874         int ret;
9875         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9876         bool log_pinned = false;
9877         struct btrfs_log_ctx ctx;
9878         bool sync_log = false;
9879         bool commit_transaction = false;
9880
9881         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9882                 return -EPERM;
9883
9884         /* we only allow rename subvolume link between subvolumes */
9885         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9886                 return -EXDEV;
9887
9888         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9889             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9890                 return -ENOTEMPTY;
9891
9892         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9893             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9894                 return -ENOTEMPTY;
9895
9896
9897         /* check for collisions, even if the  name isn't there */
9898         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9899                              new_dentry->d_name.name,
9900                              new_dentry->d_name.len);
9901
9902         if (ret) {
9903                 if (ret == -EEXIST) {
9904                         /* we shouldn't get
9905                          * eexist without a new_inode */
9906                         if (WARN_ON(!new_inode)) {
9907                                 return ret;
9908                         }
9909                 } else {
9910                         /* maybe -EOVERFLOW */
9911                         return ret;
9912                 }
9913         }
9914         ret = 0;
9915
9916         /*
9917          * we're using rename to replace one file with another.  Start IO on it
9918          * now so  we don't add too much work to the end of the transaction
9919          */
9920         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9921                 filemap_flush(old_inode->i_mapping);
9922
9923         /* close the racy window with snapshot create/destroy ioctl */
9924         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9925                 down_read(&fs_info->subvol_sem);
9926         /*
9927          * We want to reserve the absolute worst case amount of items.  So if
9928          * both inodes are subvols and we need to unlink them then that would
9929          * require 4 item modifications, but if they are both normal inodes it
9930          * would require 5 item modifications, so we'll assume they are normal
9931          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9932          * should cover the worst case number of items we'll modify.
9933          * If our rename has the whiteout flag, we need more 5 units for the
9934          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9935          * when selinux is enabled).
9936          */
9937         trans_num_items = 11;
9938         if (flags & RENAME_WHITEOUT)
9939                 trans_num_items += 5;
9940         trans = btrfs_start_transaction(root, trans_num_items);
9941         if (IS_ERR(trans)) {
9942                 ret = PTR_ERR(trans);
9943                 goto out_notrans;
9944         }
9945
9946         if (dest != root)
9947                 btrfs_record_root_in_trans(trans, dest);
9948
9949         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
9950         if (ret)
9951                 goto out_fail;
9952
9953         BTRFS_I(old_inode)->dir_index = 0ULL;
9954         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9955                 /* force full log commit if subvolume involved. */
9956                 btrfs_set_log_full_commit(trans);
9957         } else {
9958                 btrfs_pin_log_trans(root);
9959                 log_pinned = true;
9960                 ret = btrfs_insert_inode_ref(trans, dest,
9961                                              new_dentry->d_name.name,
9962                                              new_dentry->d_name.len,
9963                                              old_ino,
9964                                              btrfs_ino(BTRFS_I(new_dir)), index);
9965                 if (ret)
9966                         goto out_fail;
9967         }
9968
9969         inode_inc_iversion(old_dir);
9970         inode_inc_iversion(new_dir);
9971         inode_inc_iversion(old_inode);
9972         old_dir->i_ctime = old_dir->i_mtime =
9973         new_dir->i_ctime = new_dir->i_mtime =
9974         old_inode->i_ctime = current_time(old_dir);
9975
9976         if (old_dentry->d_parent != new_dentry->d_parent)
9977                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9978                                 BTRFS_I(old_inode), 1);
9979
9980         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9981                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9982                 ret = btrfs_unlink_subvol(trans, old_dir, root_objectid,
9983                                         old_dentry->d_name.name,
9984                                         old_dentry->d_name.len);
9985         } else {
9986                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9987                                         BTRFS_I(d_inode(old_dentry)),
9988                                         old_dentry->d_name.name,
9989                                         old_dentry->d_name.len);
9990                 if (!ret)
9991                         ret = btrfs_update_inode(trans, root, old_inode);
9992         }
9993         if (ret) {
9994                 btrfs_abort_transaction(trans, ret);
9995                 goto out_fail;
9996         }
9997
9998         if (new_inode) {
9999                 inode_inc_iversion(new_inode);
10000                 new_inode->i_ctime = current_time(new_inode);
10001                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10002                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10003                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10004                         ret = btrfs_unlink_subvol(trans, new_dir, root_objectid,
10005                                                 new_dentry->d_name.name,
10006                                                 new_dentry->d_name.len);
10007                         BUG_ON(new_inode->i_nlink == 0);
10008                 } else {
10009                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10010                                                  BTRFS_I(d_inode(new_dentry)),
10011                                                  new_dentry->d_name.name,
10012                                                  new_dentry->d_name.len);
10013                 }
10014                 if (!ret && new_inode->i_nlink == 0)
10015                         ret = btrfs_orphan_add(trans,
10016                                         BTRFS_I(d_inode(new_dentry)));
10017                 if (ret) {
10018                         btrfs_abort_transaction(trans, ret);
10019                         goto out_fail;
10020                 }
10021         }
10022
10023         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10024                              new_dentry->d_name.name,
10025                              new_dentry->d_name.len, 0, index);
10026         if (ret) {
10027                 btrfs_abort_transaction(trans, ret);
10028                 goto out_fail;
10029         }
10030
10031         if (old_inode->i_nlink == 1)
10032                 BTRFS_I(old_inode)->dir_index = index;
10033
10034         if (log_pinned) {
10035                 struct dentry *parent = new_dentry->d_parent;
10036
10037                 btrfs_init_log_ctx(&ctx, old_inode);
10038                 ret = btrfs_log_new_name(trans, BTRFS_I(old_inode),
10039                                          BTRFS_I(old_dir), parent,
10040                                          false, &ctx);
10041                 if (ret == BTRFS_NEED_LOG_SYNC)
10042                         sync_log = true;
10043                 else if (ret == BTRFS_NEED_TRANS_COMMIT)
10044                         commit_transaction = true;
10045                 ret = 0;
10046                 btrfs_end_log_trans(root);
10047                 log_pinned = false;
10048         }
10049
10050         if (flags & RENAME_WHITEOUT) {
10051                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10052                                                 old_dentry);
10053
10054                 if (ret) {
10055                         btrfs_abort_transaction(trans, ret);
10056                         goto out_fail;
10057                 }
10058         }
10059 out_fail:
10060         /*
10061          * If we have pinned the log and an error happened, we unpin tasks
10062          * trying to sync the log and force them to fallback to a transaction
10063          * commit if the log currently contains any of the inodes involved in
10064          * this rename operation (to ensure we do not persist a log with an
10065          * inconsistent state for any of these inodes or leading to any
10066          * inconsistencies when replayed). If the transaction was aborted, the
10067          * abortion reason is propagated to userspace when attempting to commit
10068          * the transaction. If the log does not contain any of these inodes, we
10069          * allow the tasks to sync it.
10070          */
10071         if (ret && log_pinned) {
10072                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10073                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10074                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10075                     (new_inode &&
10076                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10077                         btrfs_set_log_full_commit(trans);
10078
10079                 btrfs_end_log_trans(root);
10080                 log_pinned = false;
10081         }
10082         if (!ret && sync_log) {
10083                 ret = btrfs_sync_log(trans, BTRFS_I(old_inode)->root, &ctx);
10084                 if (ret)
10085                         commit_transaction = true;
10086         }
10087         if (commit_transaction) {
10088                 ret = btrfs_commit_transaction(trans);
10089         } else {
10090                 int ret2;
10091
10092                 ret2 = btrfs_end_transaction(trans);
10093                 ret = ret ? ret : ret2;
10094         }
10095 out_notrans:
10096         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10097                 up_read(&fs_info->subvol_sem);
10098
10099         return ret;
10100 }
10101
10102 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10103                          struct inode *new_dir, struct dentry *new_dentry,
10104                          unsigned int flags)
10105 {
10106         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10107                 return -EINVAL;
10108
10109         if (flags & RENAME_EXCHANGE)
10110                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10111                                           new_dentry);
10112
10113         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10114 }
10115
10116 struct btrfs_delalloc_work {
10117         struct inode *inode;
10118         struct completion completion;
10119         struct list_head list;
10120         struct btrfs_work work;
10121 };
10122
10123 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10124 {
10125         struct btrfs_delalloc_work *delalloc_work;
10126         struct inode *inode;
10127
10128         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10129                                      work);
10130         inode = delalloc_work->inode;
10131         filemap_flush(inode->i_mapping);
10132         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10133                                 &BTRFS_I(inode)->runtime_flags))
10134                 filemap_flush(inode->i_mapping);
10135
10136         iput(inode);
10137         complete(&delalloc_work->completion);
10138 }
10139
10140 static struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode)
10141 {
10142         struct btrfs_delalloc_work *work;
10143
10144         work = kmalloc(sizeof(*work), GFP_NOFS);
10145         if (!work)
10146                 return NULL;
10147
10148         init_completion(&work->completion);
10149         INIT_LIST_HEAD(&work->list);
10150         work->inode = inode;
10151         btrfs_init_work(&work->work, btrfs_run_delalloc_work, NULL, NULL);
10152
10153         return work;
10154 }
10155
10156 /*
10157  * some fairly slow code that needs optimization. This walks the list
10158  * of all the inodes with pending delalloc and forces them to disk.
10159  */
10160 static int start_delalloc_inodes(struct btrfs_root *root, int nr, bool snapshot)
10161 {
10162         struct btrfs_inode *binode;
10163         struct inode *inode;
10164         struct btrfs_delalloc_work *work, *next;
10165         struct list_head works;
10166         struct list_head splice;
10167         int ret = 0;
10168
10169         INIT_LIST_HEAD(&works);
10170         INIT_LIST_HEAD(&splice);
10171
10172         mutex_lock(&root->delalloc_mutex);
10173         spin_lock(&root->delalloc_lock);
10174         list_splice_init(&root->delalloc_inodes, &splice);
10175         while (!list_empty(&splice)) {
10176                 binode = list_entry(splice.next, struct btrfs_inode,
10177                                     delalloc_inodes);
10178
10179                 list_move_tail(&binode->delalloc_inodes,
10180                                &root->delalloc_inodes);
10181                 inode = igrab(&binode->vfs_inode);
10182                 if (!inode) {
10183                         cond_resched_lock(&root->delalloc_lock);
10184                         continue;
10185                 }
10186                 spin_unlock(&root->delalloc_lock);
10187
10188                 if (snapshot)
10189                         set_bit(BTRFS_INODE_SNAPSHOT_FLUSH,
10190                                 &binode->runtime_flags);
10191                 work = btrfs_alloc_delalloc_work(inode);
10192                 if (!work) {
10193                         iput(inode);
10194                         ret = -ENOMEM;
10195                         goto out;
10196                 }
10197                 list_add_tail(&work->list, &works);
10198                 btrfs_queue_work(root->fs_info->flush_workers,
10199                                  &work->work);
10200                 ret++;
10201                 if (nr != -1 && ret >= nr)
10202                         goto out;
10203                 cond_resched();
10204                 spin_lock(&root->delalloc_lock);
10205         }
10206         spin_unlock(&root->delalloc_lock);
10207
10208 out:
10209         list_for_each_entry_safe(work, next, &works, list) {
10210                 list_del_init(&work->list);
10211                 wait_for_completion(&work->completion);
10212                 kfree(work);
10213         }
10214
10215         if (!list_empty(&splice)) {
10216                 spin_lock(&root->delalloc_lock);
10217                 list_splice_tail(&splice, &root->delalloc_inodes);
10218                 spin_unlock(&root->delalloc_lock);
10219         }
10220         mutex_unlock(&root->delalloc_mutex);
10221         return ret;
10222 }
10223
10224 int btrfs_start_delalloc_snapshot(struct btrfs_root *root)
10225 {
10226         struct btrfs_fs_info *fs_info = root->fs_info;
10227         int ret;
10228
10229         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10230                 return -EROFS;
10231
10232         ret = start_delalloc_inodes(root, -1, true);
10233         if (ret > 0)
10234                 ret = 0;
10235         return ret;
10236 }
10237
10238 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int nr)
10239 {
10240         struct btrfs_root *root;
10241         struct list_head splice;
10242         int ret;
10243
10244         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10245                 return -EROFS;
10246
10247         INIT_LIST_HEAD(&splice);
10248
10249         mutex_lock(&fs_info->delalloc_root_mutex);
10250         spin_lock(&fs_info->delalloc_root_lock);
10251         list_splice_init(&fs_info->delalloc_roots, &splice);
10252         while (!list_empty(&splice) && nr) {
10253                 root = list_first_entry(&splice, struct btrfs_root,
10254                                         delalloc_root);
10255                 root = btrfs_grab_fs_root(root);
10256                 BUG_ON(!root);
10257                 list_move_tail(&root->delalloc_root,
10258                                &fs_info->delalloc_roots);
10259                 spin_unlock(&fs_info->delalloc_root_lock);
10260
10261                 ret = start_delalloc_inodes(root, nr, false);
10262                 btrfs_put_fs_root(root);
10263                 if (ret < 0)
10264                         goto out;
10265
10266                 if (nr != -1) {
10267                         nr -= ret;
10268                         WARN_ON(nr < 0);
10269                 }
10270                 spin_lock(&fs_info->delalloc_root_lock);
10271         }
10272         spin_unlock(&fs_info->delalloc_root_lock);
10273
10274         ret = 0;
10275 out:
10276         if (!list_empty(&splice)) {
10277                 spin_lock(&fs_info->delalloc_root_lock);
10278                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10279                 spin_unlock(&fs_info->delalloc_root_lock);
10280         }
10281         mutex_unlock(&fs_info->delalloc_root_mutex);
10282         return ret;
10283 }
10284
10285 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10286                          const char *symname)
10287 {
10288         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10289         struct btrfs_trans_handle *trans;
10290         struct btrfs_root *root = BTRFS_I(dir)->root;
10291         struct btrfs_path *path;
10292         struct btrfs_key key;
10293         struct inode *inode = NULL;
10294         int err;
10295         u64 objectid;
10296         u64 index = 0;
10297         int name_len;
10298         int datasize;
10299         unsigned long ptr;
10300         struct btrfs_file_extent_item *ei;
10301         struct extent_buffer *leaf;
10302
10303         name_len = strlen(symname);
10304         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10305                 return -ENAMETOOLONG;
10306
10307         /*
10308          * 2 items for inode item and ref
10309          * 2 items for dir items
10310          * 1 item for updating parent inode item
10311          * 1 item for the inline extent item
10312          * 1 item for xattr if selinux is on
10313          */
10314         trans = btrfs_start_transaction(root, 7);
10315         if (IS_ERR(trans))
10316                 return PTR_ERR(trans);
10317
10318         err = btrfs_find_free_ino(root, &objectid);
10319         if (err)
10320                 goto out_unlock;
10321
10322         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10323                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10324                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10325         if (IS_ERR(inode)) {
10326                 err = PTR_ERR(inode);
10327                 inode = NULL;
10328                 goto out_unlock;
10329         }
10330
10331         /*
10332         * If the active LSM wants to access the inode during
10333         * d_instantiate it needs these. Smack checks to see
10334         * if the filesystem supports xattrs by looking at the
10335         * ops vector.
10336         */
10337         inode->i_fop = &btrfs_file_operations;
10338         inode->i_op = &btrfs_file_inode_operations;
10339         inode->i_mapping->a_ops = &btrfs_aops;
10340         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10341
10342         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10343         if (err)
10344                 goto out_unlock;
10345
10346         path = btrfs_alloc_path();
10347         if (!path) {
10348                 err = -ENOMEM;
10349                 goto out_unlock;
10350         }
10351         key.objectid = btrfs_ino(BTRFS_I(inode));
10352         key.offset = 0;
10353         key.type = BTRFS_EXTENT_DATA_KEY;
10354         datasize = btrfs_file_extent_calc_inline_size(name_len);
10355         err = btrfs_insert_empty_item(trans, root, path, &key,
10356                                       datasize);
10357         if (err) {
10358                 btrfs_free_path(path);
10359                 goto out_unlock;
10360         }
10361         leaf = path->nodes[0];
10362         ei = btrfs_item_ptr(leaf, path->slots[0],
10363                             struct btrfs_file_extent_item);
10364         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10365         btrfs_set_file_extent_type(leaf, ei,
10366                                    BTRFS_FILE_EXTENT_INLINE);
10367         btrfs_set_file_extent_encryption(leaf, ei, 0);
10368         btrfs_set_file_extent_compression(leaf, ei, 0);
10369         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10370         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10371
10372         ptr = btrfs_file_extent_inline_start(ei);
10373         write_extent_buffer(leaf, symname, ptr, name_len);
10374         btrfs_mark_buffer_dirty(leaf);
10375         btrfs_free_path(path);
10376
10377         inode->i_op = &btrfs_symlink_inode_operations;
10378         inode_nohighmem(inode);
10379         inode_set_bytes(inode, name_len);
10380         btrfs_i_size_write(BTRFS_I(inode), name_len);
10381         err = btrfs_update_inode(trans, root, inode);
10382         /*
10383          * Last step, add directory indexes for our symlink inode. This is the
10384          * last step to avoid extra cleanup of these indexes if an error happens
10385          * elsewhere above.
10386          */
10387         if (!err)
10388                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10389                                 BTRFS_I(inode), 0, index);
10390         if (err)
10391                 goto out_unlock;
10392
10393         d_instantiate_new(dentry, inode);
10394
10395 out_unlock:
10396         btrfs_end_transaction(trans);
10397         if (err && inode) {
10398                 inode_dec_link_count(inode);
10399                 discard_new_inode(inode);
10400         }
10401         btrfs_btree_balance_dirty(fs_info);
10402         return err;
10403 }
10404
10405 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10406                                        u64 start, u64 num_bytes, u64 min_size,
10407                                        loff_t actual_len, u64 *alloc_hint,
10408                                        struct btrfs_trans_handle *trans)
10409 {
10410         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10411         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10412         struct extent_map *em;
10413         struct btrfs_root *root = BTRFS_I(inode)->root;
10414         struct btrfs_key ins;
10415         u64 cur_offset = start;
10416         u64 i_size;
10417         u64 cur_bytes;
10418         u64 last_alloc = (u64)-1;
10419         int ret = 0;
10420         bool own_trans = true;
10421         u64 end = start + num_bytes - 1;
10422
10423         if (trans)
10424                 own_trans = false;
10425         while (num_bytes > 0) {
10426                 if (own_trans) {
10427                         trans = btrfs_start_transaction(root, 3);
10428                         if (IS_ERR(trans)) {
10429                                 ret = PTR_ERR(trans);
10430                                 break;
10431                         }
10432                 }
10433
10434                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10435                 cur_bytes = max(cur_bytes, min_size);
10436                 /*
10437                  * If we are severely fragmented we could end up with really
10438                  * small allocations, so if the allocator is returning small
10439                  * chunks lets make its job easier by only searching for those
10440                  * sized chunks.
10441                  */
10442                 cur_bytes = min(cur_bytes, last_alloc);
10443                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10444                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10445                 if (ret) {
10446                         if (own_trans)
10447                                 btrfs_end_transaction(trans);
10448                         break;
10449                 }
10450                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10451
10452                 last_alloc = ins.offset;
10453                 ret = insert_reserved_file_extent(trans, inode,
10454                                                   cur_offset, ins.objectid,
10455                                                   ins.offset, ins.offset,
10456                                                   ins.offset, 0, 0, 0,
10457                                                   BTRFS_FILE_EXTENT_PREALLOC);
10458                 if (ret) {
10459                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10460                                                    ins.offset, 0);
10461                         btrfs_abort_transaction(trans, ret);
10462                         if (own_trans)
10463                                 btrfs_end_transaction(trans);
10464                         break;
10465                 }
10466
10467                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10468                                         cur_offset + ins.offset -1, 0);
10469
10470                 em = alloc_extent_map();
10471                 if (!em) {
10472                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10473                                 &BTRFS_I(inode)->runtime_flags);
10474                         goto next;
10475                 }
10476
10477                 em->start = cur_offset;
10478                 em->orig_start = cur_offset;
10479                 em->len = ins.offset;
10480                 em->block_start = ins.objectid;
10481                 em->block_len = ins.offset;
10482                 em->orig_block_len = ins.offset;
10483                 em->ram_bytes = ins.offset;
10484                 em->bdev = fs_info->fs_devices->latest_bdev;
10485                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10486                 em->generation = trans->transid;
10487
10488                 while (1) {
10489                         write_lock(&em_tree->lock);
10490                         ret = add_extent_mapping(em_tree, em, 1);
10491                         write_unlock(&em_tree->lock);
10492                         if (ret != -EEXIST)
10493                                 break;
10494                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10495                                                 cur_offset + ins.offset - 1,
10496                                                 0);
10497                 }
10498                 free_extent_map(em);
10499 next:
10500                 num_bytes -= ins.offset;
10501                 cur_offset += ins.offset;
10502                 *alloc_hint = ins.objectid + ins.offset;
10503
10504                 inode_inc_iversion(inode);
10505                 inode->i_ctime = current_time(inode);
10506                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10507                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10508                     (actual_len > inode->i_size) &&
10509                     (cur_offset > inode->i_size)) {
10510                         if (cur_offset > actual_len)
10511                                 i_size = actual_len;
10512                         else
10513                                 i_size = cur_offset;
10514                         i_size_write(inode, i_size);
10515                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10516                 }
10517
10518                 ret = btrfs_update_inode(trans, root, inode);
10519
10520                 if (ret) {
10521                         btrfs_abort_transaction(trans, ret);
10522                         if (own_trans)
10523                                 btrfs_end_transaction(trans);
10524                         break;
10525                 }
10526
10527                 if (own_trans)
10528                         btrfs_end_transaction(trans);
10529         }
10530         if (cur_offset < end)
10531                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10532                         end - cur_offset + 1);
10533         return ret;
10534 }
10535
10536 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10537                               u64 start, u64 num_bytes, u64 min_size,
10538                               loff_t actual_len, u64 *alloc_hint)
10539 {
10540         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10541                                            min_size, actual_len, alloc_hint,
10542                                            NULL);
10543 }
10544
10545 int btrfs_prealloc_file_range_trans(struct inode *inode,
10546                                     struct btrfs_trans_handle *trans, int mode,
10547                                     u64 start, u64 num_bytes, u64 min_size,
10548                                     loff_t actual_len, u64 *alloc_hint)
10549 {
10550         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10551                                            min_size, actual_len, alloc_hint, trans);
10552 }
10553
10554 static int btrfs_set_page_dirty(struct page *page)
10555 {
10556         return __set_page_dirty_nobuffers(page);
10557 }
10558
10559 static int btrfs_permission(struct inode *inode, int mask)
10560 {
10561         struct btrfs_root *root = BTRFS_I(inode)->root;
10562         umode_t mode = inode->i_mode;
10563
10564         if (mask & MAY_WRITE &&
10565             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10566                 if (btrfs_root_readonly(root))
10567                         return -EROFS;
10568                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10569                         return -EACCES;
10570         }
10571         return generic_permission(inode, mask);
10572 }
10573
10574 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10575 {
10576         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10577         struct btrfs_trans_handle *trans;
10578         struct btrfs_root *root = BTRFS_I(dir)->root;
10579         struct inode *inode = NULL;
10580         u64 objectid;
10581         u64 index;
10582         int ret = 0;
10583
10584         /*
10585          * 5 units required for adding orphan entry
10586          */
10587         trans = btrfs_start_transaction(root, 5);
10588         if (IS_ERR(trans))
10589                 return PTR_ERR(trans);
10590
10591         ret = btrfs_find_free_ino(root, &objectid);
10592         if (ret)
10593                 goto out;
10594
10595         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10596                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10597         if (IS_ERR(inode)) {
10598                 ret = PTR_ERR(inode);
10599                 inode = NULL;
10600                 goto out;
10601         }
10602
10603         inode->i_fop = &btrfs_file_operations;
10604         inode->i_op = &btrfs_file_inode_operations;
10605
10606         inode->i_mapping->a_ops = &btrfs_aops;
10607         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10608
10609         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10610         if (ret)
10611                 goto out;
10612
10613         ret = btrfs_update_inode(trans, root, inode);
10614         if (ret)
10615                 goto out;
10616         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10617         if (ret)
10618                 goto out;
10619
10620         /*
10621          * We set number of links to 0 in btrfs_new_inode(), and here we set
10622          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10623          * through:
10624          *
10625          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10626          */
10627         set_nlink(inode, 1);
10628         d_tmpfile(dentry, inode);
10629         unlock_new_inode(inode);
10630         mark_inode_dirty(inode);
10631 out:
10632         btrfs_end_transaction(trans);
10633         if (ret && inode)
10634                 discard_new_inode(inode);
10635         btrfs_btree_balance_dirty(fs_info);
10636         return ret;
10637 }
10638
10639 void btrfs_set_range_writeback(struct extent_io_tree *tree, u64 start, u64 end)
10640 {
10641         struct inode *inode = tree->private_data;
10642         unsigned long index = start >> PAGE_SHIFT;
10643         unsigned long end_index = end >> PAGE_SHIFT;
10644         struct page *page;
10645
10646         while (index <= end_index) {
10647                 page = find_get_page(inode->i_mapping, index);
10648                 ASSERT(page); /* Pages should be in the extent_io_tree */
10649                 set_page_writeback(page);
10650                 put_page(page);
10651                 index++;
10652         }
10653 }
10654
10655 #ifdef CONFIG_SWAP
10656 /*
10657  * Add an entry indicating a block group or device which is pinned by a
10658  * swapfile. Returns 0 on success, 1 if there is already an entry for it, or a
10659  * negative errno on failure.
10660  */
10661 static int btrfs_add_swapfile_pin(struct inode *inode, void *ptr,
10662                                   bool is_block_group)
10663 {
10664         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10665         struct btrfs_swapfile_pin *sp, *entry;
10666         struct rb_node **p;
10667         struct rb_node *parent = NULL;
10668
10669         sp = kmalloc(sizeof(*sp), GFP_NOFS);
10670         if (!sp)
10671                 return -ENOMEM;
10672         sp->ptr = ptr;
10673         sp->inode = inode;
10674         sp->is_block_group = is_block_group;
10675
10676         spin_lock(&fs_info->swapfile_pins_lock);
10677         p = &fs_info->swapfile_pins.rb_node;
10678         while (*p) {
10679                 parent = *p;
10680                 entry = rb_entry(parent, struct btrfs_swapfile_pin, node);
10681                 if (sp->ptr < entry->ptr ||
10682                     (sp->ptr == entry->ptr && sp->inode < entry->inode)) {
10683                         p = &(*p)->rb_left;
10684                 } else if (sp->ptr > entry->ptr ||
10685                            (sp->ptr == entry->ptr && sp->inode > entry->inode)) {
10686                         p = &(*p)->rb_right;
10687                 } else {
10688                         spin_unlock(&fs_info->swapfile_pins_lock);
10689                         kfree(sp);
10690                         return 1;
10691                 }
10692         }
10693         rb_link_node(&sp->node, parent, p);
10694         rb_insert_color(&sp->node, &fs_info->swapfile_pins);
10695         spin_unlock(&fs_info->swapfile_pins_lock);
10696         return 0;
10697 }
10698
10699 /* Free all of the entries pinned by this swapfile. */
10700 static void btrfs_free_swapfile_pins(struct inode *inode)
10701 {
10702         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10703         struct btrfs_swapfile_pin *sp;
10704         struct rb_node *node, *next;
10705
10706         spin_lock(&fs_info->swapfile_pins_lock);
10707         node = rb_first(&fs_info->swapfile_pins);
10708         while (node) {
10709                 next = rb_next(node);
10710                 sp = rb_entry(node, struct btrfs_swapfile_pin, node);
10711                 if (sp->inode == inode) {
10712                         rb_erase(&sp->node, &fs_info->swapfile_pins);
10713                         if (sp->is_block_group)
10714                                 btrfs_put_block_group(sp->ptr);
10715                         kfree(sp);
10716                 }
10717                 node = next;
10718         }
10719         spin_unlock(&fs_info->swapfile_pins_lock);
10720 }
10721
10722 struct btrfs_swap_info {
10723         u64 start;
10724         u64 block_start;
10725         u64 block_len;
10726         u64 lowest_ppage;
10727         u64 highest_ppage;
10728         unsigned long nr_pages;
10729         int nr_extents;
10730 };
10731
10732 static int btrfs_add_swap_extent(struct swap_info_struct *sis,
10733                                  struct btrfs_swap_info *bsi)
10734 {
10735         unsigned long nr_pages;
10736         u64 first_ppage, first_ppage_reported, next_ppage;
10737         int ret;
10738
10739         first_ppage = ALIGN(bsi->block_start, PAGE_SIZE) >> PAGE_SHIFT;
10740         next_ppage = ALIGN_DOWN(bsi->block_start + bsi->block_len,
10741                                 PAGE_SIZE) >> PAGE_SHIFT;
10742
10743         if (first_ppage >= next_ppage)
10744                 return 0;
10745         nr_pages = next_ppage - first_ppage;
10746
10747         first_ppage_reported = first_ppage;
10748         if (bsi->start == 0)
10749                 first_ppage_reported++;
10750         if (bsi->lowest_ppage > first_ppage_reported)
10751                 bsi->lowest_ppage = first_ppage_reported;
10752         if (bsi->highest_ppage < (next_ppage - 1))
10753                 bsi->highest_ppage = next_ppage - 1;
10754
10755         ret = add_swap_extent(sis, bsi->nr_pages, nr_pages, first_ppage);
10756         if (ret < 0)
10757                 return ret;
10758         bsi->nr_extents += ret;
10759         bsi->nr_pages += nr_pages;
10760         return 0;
10761 }
10762
10763 static void btrfs_swap_deactivate(struct file *file)
10764 {
10765         struct inode *inode = file_inode(file);
10766
10767         btrfs_free_swapfile_pins(inode);
10768         atomic_dec(&BTRFS_I(inode)->root->nr_swapfiles);
10769 }
10770
10771 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10772                                sector_t *span)
10773 {
10774         struct inode *inode = file_inode(file);
10775         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
10776         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
10777         struct extent_state *cached_state = NULL;
10778         struct extent_map *em = NULL;
10779         struct btrfs_device *device = NULL;
10780         struct btrfs_swap_info bsi = {
10781                 .lowest_ppage = (sector_t)-1ULL,
10782         };
10783         int ret = 0;
10784         u64 isize;
10785         u64 start;
10786
10787         /*
10788          * If the swap file was just created, make sure delalloc is done. If the
10789          * file changes again after this, the user is doing something stupid and
10790          * we don't really care.
10791          */
10792         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
10793         if (ret)
10794                 return ret;
10795
10796         /*
10797          * The inode is locked, so these flags won't change after we check them.
10798          */
10799         if (BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS) {
10800                 btrfs_warn(fs_info, "swapfile must not be compressed");
10801                 return -EINVAL;
10802         }
10803         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW)) {
10804                 btrfs_warn(fs_info, "swapfile must not be copy-on-write");
10805                 return -EINVAL;
10806         }
10807         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)) {
10808                 btrfs_warn(fs_info, "swapfile must not be checksummed");
10809                 return -EINVAL;
10810         }
10811
10812         /*
10813          * Balance or device remove/replace/resize can move stuff around from
10814          * under us. The EXCL_OP flag makes sure they aren't running/won't run
10815          * concurrently while we are mapping the swap extents, and
10816          * fs_info->swapfile_pins prevents them from running while the swap file
10817          * is active and moving the extents. Note that this also prevents a
10818          * concurrent device add which isn't actually necessary, but it's not
10819          * really worth the trouble to allow it.
10820          */
10821         if (test_and_set_bit(BTRFS_FS_EXCL_OP, &fs_info->flags)) {
10822                 btrfs_warn(fs_info,
10823            "cannot activate swapfile while exclusive operation is running");
10824                 return -EBUSY;
10825         }
10826         /*
10827          * Snapshots can create extents which require COW even if NODATACOW is
10828          * set. We use this counter to prevent snapshots. We must increment it
10829          * before walking the extents because we don't want a concurrent
10830          * snapshot to run after we've already checked the extents.
10831          */
10832         atomic_inc(&BTRFS_I(inode)->root->nr_swapfiles);
10833
10834         isize = ALIGN_DOWN(inode->i_size, fs_info->sectorsize);
10835
10836         lock_extent_bits(io_tree, 0, isize - 1, &cached_state);
10837         start = 0;
10838         while (start < isize) {
10839                 u64 logical_block_start, physical_block_start;
10840                 struct btrfs_block_group *bg;
10841                 u64 len = isize - start;
10842
10843                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
10844                 if (IS_ERR(em)) {
10845                         ret = PTR_ERR(em);
10846                         goto out;
10847                 }
10848
10849                 if (em->block_start == EXTENT_MAP_HOLE) {
10850                         btrfs_warn(fs_info, "swapfile must not have holes");
10851                         ret = -EINVAL;
10852                         goto out;
10853                 }
10854                 if (em->block_start == EXTENT_MAP_INLINE) {
10855                         /*
10856                          * It's unlikely we'll ever actually find ourselves
10857                          * here, as a file small enough to fit inline won't be
10858                          * big enough to store more than the swap header, but in
10859                          * case something changes in the future, let's catch it
10860                          * here rather than later.
10861                          */
10862                         btrfs_warn(fs_info, "swapfile must not be inline");
10863                         ret = -EINVAL;
10864                         goto out;
10865                 }
10866                 if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
10867                         btrfs_warn(fs_info, "swapfile must not be compressed");
10868                         ret = -EINVAL;
10869                         goto out;
10870                 }
10871
10872                 logical_block_start = em->block_start + (start - em->start);
10873                 len = min(len, em->len - (start - em->start));
10874                 free_extent_map(em);
10875                 em = NULL;
10876
10877                 ret = can_nocow_extent(inode, start, &len, NULL, NULL, NULL);
10878                 if (ret < 0) {
10879                         goto out;
10880                 } else if (ret) {
10881                         ret = 0;
10882                 } else {
10883                         btrfs_warn(fs_info,
10884                                    "swapfile must not be copy-on-write");
10885                         ret = -EINVAL;
10886                         goto out;
10887                 }
10888
10889                 em = btrfs_get_chunk_map(fs_info, logical_block_start, len);
10890                 if (IS_ERR(em)) {
10891                         ret = PTR_ERR(em);
10892                         goto out;
10893                 }
10894
10895                 if (em->map_lookup->type & BTRFS_BLOCK_GROUP_PROFILE_MASK) {
10896                         btrfs_warn(fs_info,
10897                                    "swapfile must have single data profile");
10898                         ret = -EINVAL;
10899                         goto out;
10900                 }
10901
10902                 if (device == NULL) {
10903                         device = em->map_lookup->stripes[0].dev;
10904                         ret = btrfs_add_swapfile_pin(inode, device, false);
10905                         if (ret == 1)
10906                                 ret = 0;
10907                         else if (ret)
10908                                 goto out;
10909                 } else if (device != em->map_lookup->stripes[0].dev) {
10910                         btrfs_warn(fs_info, "swapfile must be on one device");
10911                         ret = -EINVAL;
10912                         goto out;
10913                 }
10914
10915                 physical_block_start = (em->map_lookup->stripes[0].physical +
10916                                         (logical_block_start - em->start));
10917                 len = min(len, em->len - (logical_block_start - em->start));
10918                 free_extent_map(em);
10919                 em = NULL;
10920
10921                 bg = btrfs_lookup_block_group(fs_info, logical_block_start);
10922                 if (!bg) {
10923                         btrfs_warn(fs_info,
10924                            "could not find block group containing swapfile");
10925                         ret = -EINVAL;
10926                         goto out;
10927                 }
10928
10929                 ret = btrfs_add_swapfile_pin(inode, bg, true);
10930                 if (ret) {
10931                         btrfs_put_block_group(bg);
10932                         if (ret == 1)
10933                                 ret = 0;
10934                         else
10935                                 goto out;
10936                 }
10937
10938                 if (bsi.block_len &&
10939                     bsi.block_start + bsi.block_len == physical_block_start) {
10940                         bsi.block_len += len;
10941                 } else {
10942                         if (bsi.block_len) {
10943                                 ret = btrfs_add_swap_extent(sis, &bsi);
10944                                 if (ret)
10945                                         goto out;
10946                         }
10947                         bsi.start = start;
10948                         bsi.block_start = physical_block_start;
10949                         bsi.block_len = len;
10950                 }
10951
10952                 start += len;
10953         }
10954
10955         if (bsi.block_len)
10956                 ret = btrfs_add_swap_extent(sis, &bsi);
10957
10958 out:
10959         if (!IS_ERR_OR_NULL(em))
10960                 free_extent_map(em);
10961
10962         unlock_extent_cached(io_tree, 0, isize - 1, &cached_state);
10963
10964         if (ret)
10965                 btrfs_swap_deactivate(file);
10966
10967         clear_bit(BTRFS_FS_EXCL_OP, &fs_info->flags);
10968
10969         if (ret)
10970                 return ret;
10971
10972         if (device)
10973                 sis->bdev = device->bdev;
10974         *span = bsi.highest_ppage - bsi.lowest_ppage + 1;
10975         sis->max = bsi.nr_pages;
10976         sis->pages = bsi.nr_pages - 1;
10977         sis->highest_bit = bsi.nr_pages - 1;
10978         return bsi.nr_extents;
10979 }
10980 #else
10981 static void btrfs_swap_deactivate(struct file *file)
10982 {
10983 }
10984
10985 static int btrfs_swap_activate(struct swap_info_struct *sis, struct file *file,
10986                                sector_t *span)
10987 {
10988         return -EOPNOTSUPP;
10989 }
10990 #endif
10991
10992 static const struct inode_operations btrfs_dir_inode_operations = {
10993         .getattr        = btrfs_getattr,
10994         .lookup         = btrfs_lookup,
10995         .create         = btrfs_create,
10996         .unlink         = btrfs_unlink,
10997         .link           = btrfs_link,
10998         .mkdir          = btrfs_mkdir,
10999         .rmdir          = btrfs_rmdir,
11000         .rename         = btrfs_rename2,
11001         .symlink        = btrfs_symlink,
11002         .setattr        = btrfs_setattr,
11003         .mknod          = btrfs_mknod,
11004         .listxattr      = btrfs_listxattr,
11005         .permission     = btrfs_permission,
11006         .get_acl        = btrfs_get_acl,
11007         .set_acl        = btrfs_set_acl,
11008         .update_time    = btrfs_update_time,
11009         .tmpfile        = btrfs_tmpfile,
11010 };
11011 static const struct inode_operations btrfs_dir_ro_inode_operations = {
11012         .lookup         = btrfs_lookup,
11013         .permission     = btrfs_permission,
11014         .update_time    = btrfs_update_time,
11015 };
11016
11017 static const struct file_operations btrfs_dir_file_operations = {
11018         .llseek         = generic_file_llseek,
11019         .read           = generic_read_dir,
11020         .iterate_shared = btrfs_real_readdir,
11021         .open           = btrfs_opendir,
11022         .unlocked_ioctl = btrfs_ioctl,
11023 #ifdef CONFIG_COMPAT
11024         .compat_ioctl   = btrfs_compat_ioctl,
11025 #endif
11026         .release        = btrfs_release_file,
11027         .fsync          = btrfs_sync_file,
11028 };
11029
11030 static const struct extent_io_ops btrfs_extent_io_ops = {
11031         /* mandatory callbacks */
11032         .submit_bio_hook = btrfs_submit_bio_hook,
11033         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
11034 };
11035
11036 /*
11037  * btrfs doesn't support the bmap operation because swapfiles
11038  * use bmap to make a mapping of extents in the file.  They assume
11039  * these extents won't change over the life of the file and they
11040  * use the bmap result to do IO directly to the drive.
11041  *
11042  * the btrfs bmap call would return logical addresses that aren't
11043  * suitable for IO and they also will change frequently as COW
11044  * operations happen.  So, swapfile + btrfs == corruption.
11045  *
11046  * For now we're avoiding this by dropping bmap.
11047  */
11048 static const struct address_space_operations btrfs_aops = {
11049         .readpage       = btrfs_readpage,
11050         .writepage      = btrfs_writepage,
11051         .writepages     = btrfs_writepages,
11052         .readpages      = btrfs_readpages,
11053         .direct_IO      = btrfs_direct_IO,
11054         .invalidatepage = btrfs_invalidatepage,
11055         .releasepage    = btrfs_releasepage,
11056         .set_page_dirty = btrfs_set_page_dirty,
11057         .error_remove_page = generic_error_remove_page,
11058         .swap_activate  = btrfs_swap_activate,
11059         .swap_deactivate = btrfs_swap_deactivate,
11060 };
11061
11062 static const struct inode_operations btrfs_file_inode_operations = {
11063         .getattr        = btrfs_getattr,
11064         .setattr        = btrfs_setattr,
11065         .listxattr      = btrfs_listxattr,
11066         .permission     = btrfs_permission,
11067         .fiemap         = btrfs_fiemap,
11068         .get_acl        = btrfs_get_acl,
11069         .set_acl        = btrfs_set_acl,
11070         .update_time    = btrfs_update_time,
11071 };
11072 static const struct inode_operations btrfs_special_inode_operations = {
11073         .getattr        = btrfs_getattr,
11074         .setattr        = btrfs_setattr,
11075         .permission     = btrfs_permission,
11076         .listxattr      = btrfs_listxattr,
11077         .get_acl        = btrfs_get_acl,
11078         .set_acl        = btrfs_set_acl,
11079         .update_time    = btrfs_update_time,
11080 };
11081 static const struct inode_operations btrfs_symlink_inode_operations = {
11082         .get_link       = page_get_link,
11083         .getattr        = btrfs_getattr,
11084         .setattr        = btrfs_setattr,
11085         .permission     = btrfs_permission,
11086         .listxattr      = btrfs_listxattr,
11087         .update_time    = btrfs_update_time,
11088 };
11089
11090 const struct dentry_operations btrfs_dentry_operations = {
11091         .d_delete       = btrfs_dentry_delete,
11092 };