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