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