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