fs/btrfs/inode.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18
  19 #include <linux/kernel.h>
  20 #include <linux/bio.h>
  21 #include <linux/buffer_head.h>
  22 #include <linux/file.h>
  23 #include <linux/fs.h>
  24 #include <linux/pagemap.h>
  25 #include <linux/highmem.h>
  26 #include <linux/time.h>
  27 #include <linux/init.h>
  28 #include <linux/string.h>
  29 #include <linux/backing-dev.h>
  30 #include <linux/mpage.h>
  31 #include <linux/swap.h>
  32 #include <linux/writeback.h>
  33 #include <linux/compat.h>
  34 #include <linux/bit_spinlock.h>
  35 #include <linux/xattr.h>
  36 #include <linux/posix_acl.h>
  37 #include <linux/falloc.h>
  38 #include <linux/slab.h>
  39 #include <linux/ratelimit.h>
  40 #include <linux/mount.h>
  41 #include <linux/btrfs.h>
  42 #include <linux/blkdev.h>
  43 #include <linux/posix_acl_xattr.h>
  44 #include <linux/uio.h>
  45 #include <linux/magic.h>
  46 #include <linux/iversion.h>
  47 #include "ctree.h"
  48 #include "disk-io.h"
  49 #include "transaction.h"
  50 #include "btrfs_inode.h"
  51 #include "print-tree.h"
  52 #include "ordered-data.h"
  53 #include "xattr.h"
  54 #include "tree-log.h"
  55 #include "volumes.h"
  56 #include "compression.h"
  57 #include "locking.h"
  58 #include "free-space-cache.h"
  59 #include "inode-map.h"
  60 #include "backref.h"
  61 #include "hash.h"
  62 #include "props.h"
  63 #include "qgroup.h"
  64 #include "dedupe.h"
  65
  66 struct btrfs_iget_args {
  67         struct btrfs_key *location;
  68         struct btrfs_root *root;
  69 };
  70
  71 struct btrfs_dio_data {
  72         u64 reserve;
  73         u64 unsubmitted_oe_range_start;
  74         u64 unsubmitted_oe_range_end;
  75         int overwrite;
  76 };
  77
  78 static const struct inode_operations btrfs_dir_inode_operations;
  79 static const struct inode_operations btrfs_symlink_inode_operations;
  80 static const struct inode_operations btrfs_dir_ro_inode_operations;
  81 static const struct inode_operations btrfs_special_inode_operations;
  82 static const struct inode_operations btrfs_file_inode_operations;
  83 static const struct address_space_operations btrfs_aops;
  84 static const struct address_space_operations btrfs_symlink_aops;
  85 static const struct file_operations btrfs_dir_file_operations;
  86 static const struct extent_io_ops btrfs_extent_io_ops;
  87
  88 static struct kmem_cache *btrfs_inode_cachep;
  89 struct kmem_cache *btrfs_trans_handle_cachep;
  90 struct kmem_cache *btrfs_path_cachep;
  91 struct kmem_cache *btrfs_free_space_cachep;
  92
  93 #define S_SHIFT 12
  94 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  95         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  96         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  97         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  98         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  99         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
 100         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 101         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 102 };
 103
 104 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 105 static int btrfs_truncate(struct inode *inode);
 106 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 107 static noinline int cow_file_range(struct inode *inode,
 108                                    struct page *locked_page,
 109                                    u64 start, u64 end, u64 delalloc_end,
 110                                    int *page_started, unsigned long *nr_written,
 111                                    int unlock, struct btrfs_dedupe_hash *hash);
 112 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
 113                                        u64 orig_start, u64 block_start,
 114                                        u64 block_len, u64 orig_block_len,
 115                                        u64 ram_bytes, int compress_type,
 116                                        int type);
 117
 118 static void __endio_write_update_ordered(struct inode *inode,
 119                                          const u64 offset, const u64 bytes,
 120                                          const bool uptodate);
 121
 122 /*
 123  * Cleanup all submitted ordered extents in specified range to handle errors
 124  * from the fill_dellaloc() callback.
 125  *
 126  * NOTE: caller must ensure that when an error happens, it can not call
 127  * extent_clear_unlock_delalloc() to clear both the bits EXTENT_DO_ACCOUNTING
 128  * and EXTENT_DELALLOC simultaneously, because that causes the reserved metadata
 129  * to be released, which we want to happen only when finishing the ordered
 130  * extent (btrfs_finish_ordered_io()). Also note that the caller of the
 131  * fill_delalloc() callback already does proper cleanup for the first page of
 132  * the range, that is, it invokes the callback writepage_end_io_hook() for the
 133  * range of the first page.
 134  */
 135 static inline void btrfs_cleanup_ordered_extents(struct inode *inode,
 136                                                  const u64 offset,
 137                                                  const u64 bytes)
 138 {
 139         unsigned long index = offset >> PAGE_SHIFT;
 140         unsigned long end_index = (offset + bytes - 1) >> PAGE_SHIFT;
 141         struct page *page;
 142
 143         while (index <= end_index) {
 144                 page = find_get_page(inode->i_mapping, index);
 145                 index++;
 146                 if (!page)
 147                         continue;
 148                 ClearPagePrivate2(page);
 149                 put_page(page);
 150         }
 151         return __endio_write_update_ordered(inode, offset + PAGE_SIZE,
 152                                             bytes - PAGE_SIZE, false);
 153 }
 154
 155 static int btrfs_dirty_inode(struct inode *inode);
 156
 157 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 158 void btrfs_test_inode_set_ops(struct inode *inode)
 159 {
 160         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 161 }
 162 #endif
 163
 164 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 165                                      struct inode *inode,  struct inode *dir,
 166                                      const struct qstr *qstr)
 167 {
 168         int err;
 169
 170         err = btrfs_init_acl(trans, inode, dir);
 171         if (!err)
 172                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 173         return err;
 174 }
 175
 176 /*
 177  * this does all the hard work for inserting an inline extent into
 178  * the btree.  The caller should have done a btrfs_drop_extents so that
 179  * no overlapping inline items exist in the btree
 180  */
 181 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 182                                 struct btrfs_path *path, int extent_inserted,
 183                                 struct btrfs_root *root, struct inode *inode,
 184                                 u64 start, size_t size, size_t compressed_size,
 185                                 int compress_type,
 186                                 struct page **compressed_pages)
 187 {
 188         struct extent_buffer *leaf;
 189         struct page *page = NULL;
 190         char *kaddr;
 191         unsigned long ptr;
 192         struct btrfs_file_extent_item *ei;
 193         int ret;
 194         size_t cur_size = size;
 195         unsigned long offset;
 196
 197         if (compressed_size && compressed_pages)
 198                 cur_size = compressed_size;
 199
 200         inode_add_bytes(inode, size);
 201
 202         if (!extent_inserted) {
 203                 struct btrfs_key key;
 204                 size_t datasize;
 205
 206                 key.objectid = btrfs_ino(BTRFS_I(inode));
 207                 key.offset = start;
 208                 key.type = BTRFS_EXTENT_DATA_KEY;
 209
 210                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 211                 path->leave_spinning = 1;
 212                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 213                                               datasize);
 214                 if (ret)
 215                         goto fail;
 216         }
 217         leaf = path->nodes[0];
 218         ei = btrfs_item_ptr(leaf, path->slots[0],
 219                             struct btrfs_file_extent_item);
 220         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 221         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 222         btrfs_set_file_extent_encryption(leaf, ei, 0);
 223         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 224         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 225         ptr = btrfs_file_extent_inline_start(ei);
 226
 227         if (compress_type != BTRFS_COMPRESS_NONE) {
 228                 struct page *cpage;
 229                 int i = 0;
 230                 while (compressed_size > 0) {
 231                         cpage = compressed_pages[i];
 232                         cur_size = min_t(unsigned long, compressed_size,
 233                                        PAGE_SIZE);
 234
 235                         kaddr = kmap_atomic(cpage);
 236                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 237                         kunmap_atomic(kaddr);
 238
 239                         i++;
 240                         ptr += cur_size;
 241                         compressed_size -= cur_size;
 242                 }
 243                 btrfs_set_file_extent_compression(leaf, ei,
 244                                                   compress_type);
 245         } else {
 246                 page = find_get_page(inode->i_mapping,
 247                                      start >> PAGE_SHIFT);
 248                 btrfs_set_file_extent_compression(leaf, ei, 0);
 249                 kaddr = kmap_atomic(page);
 250                 offset = start & (PAGE_SIZE - 1);
 251                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 252                 kunmap_atomic(kaddr);
 253                 put_page(page);
 254         }
 255         btrfs_mark_buffer_dirty(leaf);
 256         btrfs_release_path(path);
 257
 258         /*
 259          * we're an inline extent, so nobody can
 260          * extend the file past i_size without locking
 261          * a page we already have locked.
 262          *
 263          * We must do any isize and inode updates
 264          * before we unlock the pages.  Otherwise we
 265          * could end up racing with unlink.
 266          */
 267         BTRFS_I(inode)->disk_i_size = inode->i_size;
 268         ret = btrfs_update_inode(trans, root, inode);
 269
 270 fail:
 271         return ret;
 272 }
 273
 274
 275 /*
 276  * conditionally insert an inline extent into the file.  This
 277  * does the checks required to make sure the data is small enough
 278  * to fit as an inline extent.
 279  */
 280 static noinline int cow_file_range_inline(struct btrfs_root *root,
 281                                           struct inode *inode, u64 start,
 282                                           u64 end, size_t compressed_size,
 283                                           int compress_type,
 284                                           struct page **compressed_pages)
 285 {
 286         struct btrfs_fs_info *fs_info = root->fs_info;
 287         struct btrfs_trans_handle *trans;
 288         u64 isize = i_size_read(inode);
 289         u64 actual_end = min(end + 1, isize);
 290         u64 inline_len = actual_end - start;
 291         u64 aligned_end = ALIGN(end, fs_info->sectorsize);
 292         u64 data_len = inline_len;
 293         int ret;
 294         struct btrfs_path *path;
 295         int extent_inserted = 0;
 296         u32 extent_item_size;
 297
 298         if (compressed_size)
 299                 data_len = compressed_size;
 300
 301         if (start > 0 ||
 302             actual_end > fs_info->sectorsize ||
 303             data_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info) ||
 304             (!compressed_size &&
 305             (actual_end & (fs_info->sectorsize - 1)) == 0) ||
 306             end + 1 < isize ||
 307             data_len > fs_info->max_inline) {
 308                 return 1;
 309         }
 310
 311         path = btrfs_alloc_path();
 312         if (!path)
 313                 return -ENOMEM;
 314
 315         trans = btrfs_join_transaction(root);
 316         if (IS_ERR(trans)) {
 317                 btrfs_free_path(path);
 318                 return PTR_ERR(trans);
 319         }
 320         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
 321
 322         if (compressed_size && compressed_pages)
 323                 extent_item_size = btrfs_file_extent_calc_inline_size(
 324                    compressed_size);
 325         else
 326                 extent_item_size = btrfs_file_extent_calc_inline_size(
 327                     inline_len);
 328
 329         ret = __btrfs_drop_extents(trans, root, inode, path,
 330                                    start, aligned_end, NULL,
 331                                    1, 1, extent_item_size, &extent_inserted);
 332         if (ret) {
 333                 btrfs_abort_transaction(trans, ret);
 334                 goto out;
 335         }
 336
 337         if (isize > actual_end)
 338                 inline_len = min_t(u64, isize, actual_end);
 339         ret = insert_inline_extent(trans, path, extent_inserted,
 340                                    root, inode, start,
 341                                    inline_len, compressed_size,
 342                                    compress_type, compressed_pages);
 343         if (ret && ret != -ENOSPC) {
 344                 btrfs_abort_transaction(trans, ret);
 345                 goto out;
 346         } else if (ret == -ENOSPC) {
 347                 ret = 1;
 348                 goto out;
 349         }
 350
 351         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 352         btrfs_drop_extent_cache(BTRFS_I(inode), start, aligned_end - 1, 0);
 353 out:
 354         /*
 355          * Don't forget to free the reserved space, as for inlined extent
 356          * it won't count as data extent, free them directly here.
 357          * And at reserve time, it's always aligned to page size, so
 358          * just free one page here.
 359          */
 360         btrfs_qgroup_free_data(inode, NULL, 0, PAGE_SIZE);
 361         btrfs_free_path(path);
 362         btrfs_end_transaction(trans);
 363         return ret;
 364 }
 365
 366 struct async_extent {
 367         u64 start;
 368         u64 ram_size;
 369         u64 compressed_size;
 370         struct page **pages;
 371         unsigned long nr_pages;
 372         int compress_type;
 373         struct list_head list;
 374 };
 375
 376 struct async_cow {
 377         struct inode *inode;
 378         struct btrfs_root *root;
 379         struct page *locked_page;
 380         u64 start;
 381         u64 end;
 382         unsigned int write_flags;
 383         struct list_head extents;
 384         struct btrfs_work work;
 385 };
 386
 387 static noinline int add_async_extent(struct async_cow *cow,
 388                                      u64 start, u64 ram_size,
 389                                      u64 compressed_size,
 390                                      struct page **pages,
 391                                      unsigned long nr_pages,
 392                                      int compress_type)
 393 {
 394         struct async_extent *async_extent;
 395
 396         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 397         BUG_ON(!async_extent); /* -ENOMEM */
 398         async_extent->start = start;
 399         async_extent->ram_size = ram_size;
 400         async_extent->compressed_size = compressed_size;
 401         async_extent->pages = pages;
 402         async_extent->nr_pages = nr_pages;
 403         async_extent->compress_type = compress_type;
 404         list_add_tail(&async_extent->list, &cow->extents);
 405         return 0;
 406 }
 407
 408 static inline int inode_need_compress(struct inode *inode, u64 start, u64 end)
 409 {
 410         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 411
 412         /* force compress */
 413         if (btrfs_test_opt(fs_info, FORCE_COMPRESS))
 414                 return 1;
 415         /* defrag ioctl */
 416         if (BTRFS_I(inode)->defrag_compress)
 417                 return 1;
 418         /* bad compression ratios */
 419         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 420                 return 0;
 421         if (btrfs_test_opt(fs_info, COMPRESS) ||
 422             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 423             BTRFS_I(inode)->prop_compress)
 424                 return btrfs_compress_heuristic(inode, start, end);
 425         return 0;
 426 }
 427
 428 static inline void inode_should_defrag(struct btrfs_inode *inode,
 429                 u64 start, u64 end, u64 num_bytes, u64 small_write)
 430 {
 431         /* If this is a small write inside eof, kick off a defrag */
 432         if (num_bytes < small_write &&
 433             (start > 0 || end + 1 < inode->disk_i_size))
 434                 btrfs_add_inode_defrag(NULL, inode);
 435 }
 436
 437 /*
 438  * we create compressed extents in two phases.  The first
 439  * phase compresses a range of pages that have already been
 440  * locked (both pages and state bits are locked).
 441  *
 442  * This is done inside an ordered work queue, and the compression
 443  * is spread across many cpus.  The actual IO submission is step
 444  * two, and the ordered work queue takes care of making sure that
 445  * happens in the same order things were put onto the queue by
 446  * writepages and friends.
 447  *
 448  * If this code finds it can't get good compression, it puts an
 449  * entry onto the work queue to write the uncompressed bytes.  This
 450  * makes sure that both compressed inodes and uncompressed inodes
 451  * are written in the same order that the flusher thread sent them
 452  * down.
 453  */
 454 static noinline void compress_file_range(struct inode *inode,
 455                                         struct page *locked_page,
 456                                         u64 start, u64 end,
 457                                         struct async_cow *async_cow,
 458                                         int *num_added)
 459 {
 460         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 461         struct btrfs_root *root = BTRFS_I(inode)->root;
 462         u64 blocksize = fs_info->sectorsize;
 463         u64 actual_end;
 464         u64 isize = i_size_read(inode);
 465         int ret = 0;
 466         struct page **pages = NULL;
 467         unsigned long nr_pages;
 468         unsigned long total_compressed = 0;
 469         unsigned long total_in = 0;
 470         int i;
 471         int will_compress;
 472         int compress_type = fs_info->compress_type;
 473         int redirty = 0;
 474
 475         inode_should_defrag(BTRFS_I(inode), start, end, end - start + 1,
 476                         SZ_16K);
 477
 478         actual_end = min_t(u64, isize, end + 1);
 479 again:
 480         will_compress = 0;
 481         nr_pages = (end >> PAGE_SHIFT) - (start >> PAGE_SHIFT) + 1;
 482         BUILD_BUG_ON((BTRFS_MAX_COMPRESSED % PAGE_SIZE) != 0);
 483         nr_pages = min_t(unsigned long, nr_pages,
 484                         BTRFS_MAX_COMPRESSED / PAGE_SIZE);
 485
 486         /*
 487          * we don't want to send crud past the end of i_size through
 488          * compression, that's just a waste of CPU time.  So, if the
 489          * end of the file is before the start of our current
 490          * requested range of bytes, we bail out to the uncompressed
 491          * cleanup code that can deal with all of this.
 492          *
 493          * It isn't really the fastest way to fix things, but this is a
 494          * very uncommon corner.
 495          */
 496         if (actual_end <= start)
 497                 goto cleanup_and_bail_uncompressed;
 498
 499         total_compressed = actual_end - start;
 500
 501         /*
 502          * skip compression for a small file range(<=blocksize) that
 503          * isn't an inline extent, since it doesn't save disk space at all.
 504          */
 505         if (total_compressed <= blocksize &&
 506            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 507                 goto cleanup_and_bail_uncompressed;
 508
 509         total_compressed = min_t(unsigned long, total_compressed,
 510                         BTRFS_MAX_UNCOMPRESSED);
 511         total_in = 0;
 512         ret = 0;
 513
 514         /*
 515          * we do compression for mount -o compress and when the
 516          * inode has not been flagged as nocompress.  This flag can
 517          * change at any time if we discover bad compression ratios.
 518          */
 519         if (inode_need_compress(inode, start, end)) {
 520                 WARN_ON(pages);
 521                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 522                 if (!pages) {
 523                         /* just bail out to the uncompressed code */
 524                         goto cont;
 525                 }
 526
 527                 if (BTRFS_I(inode)->defrag_compress)
 528                         compress_type = BTRFS_I(inode)->defrag_compress;
 529                 else if (BTRFS_I(inode)->prop_compress)
 530                         compress_type = BTRFS_I(inode)->prop_compress;
 531
 532                 /*
 533                  * we need to call clear_page_dirty_for_io on each
 534                  * page in the range.  Otherwise applications with the file
 535                  * mmap'd can wander in and change the page contents while
 536                  * we are compressing them.
 537                  *
 538                  * If the compression fails for any reason, we set the pages
 539                  * dirty again later on.
 540                  */
 541                 extent_range_clear_dirty_for_io(inode, start, end);
 542                 redirty = 1;
 543
 544                 /* Compression level is applied here and only here */
 545                 ret = btrfs_compress_pages(
 546                         compress_type | (fs_info->compress_level << 4),
 547                                            inode->i_mapping, start,
 548                                            pages,
 549                                            &nr_pages,
 550                                            &total_in,
 551                                            &total_compressed);
 552
 553                 if (!ret) {
 554                         unsigned long offset = total_compressed &
 555                                 (PAGE_SIZE - 1);
 556                         struct page *page = pages[nr_pages - 1];
 557                         char *kaddr;
 558
 559                         /* zero the tail end of the last page, we might be
 560                          * sending it down to disk
 561                          */
 562                         if (offset) {
 563                                 kaddr = kmap_atomic(page);
 564                                 memset(kaddr + offset, 0,
 565                                        PAGE_SIZE - offset);
 566                                 kunmap_atomic(kaddr);
 567                         }
 568                         will_compress = 1;
 569                 }
 570         }
 571 cont:
 572         if (start == 0) {
 573                 /* lets try to make an inline extent */
 574                 if (ret || total_in < actual_end) {
 575                         /* we didn't compress the entire range, try
 576                          * to make an uncompressed inline extent.
 577                          */
 578                         ret = cow_file_range_inline(root, inode, start, end,
 579                                             0, BTRFS_COMPRESS_NONE, NULL);
 580                 } else {
 581                         /* try making a compressed inline extent */
 582                         ret = cow_file_range_inline(root, inode, start, end,
 583                                                     total_compressed,
 584                                                     compress_type, pages);
 585                 }
 586                 if (ret <= 0) {
 587                         unsigned long clear_flags = EXTENT_DELALLOC |
 588                                 EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 589                                 EXTENT_DO_ACCOUNTING;
 590                         unsigned long page_error_op;
 591
 592                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 593
 594                         /*
 595                          * inline extent creation worked or returned error,
 596                          * we don't need to create any more async work items.
 597                          * Unlock and free up our temp pages.
 598                          *
 599                          * We use DO_ACCOUNTING here because we need the
 600                          * delalloc_release_metadata to be done _after_ we drop
 601                          * our outstanding extent for clearing delalloc for this
 602                          * range.
 603                          */
 604                         extent_clear_unlock_delalloc(inode, start, end, end,
 605                                                      NULL, clear_flags,
 606                                                      PAGE_UNLOCK |
 607                                                      PAGE_CLEAR_DIRTY |
 608                                                      PAGE_SET_WRITEBACK |
 609                                                      page_error_op |
 610                                                      PAGE_END_WRITEBACK);
 611                         goto free_pages_out;
 612                 }
 613         }
 614
 615         if (will_compress) {
 616                 /*
 617                  * we aren't doing an inline extent round the compressed size
 618                  * up to a block size boundary so the allocator does sane
 619                  * things
 620                  */
 621                 total_compressed = ALIGN(total_compressed, blocksize);
 622
 623                 /*
 624                  * one last check to make sure the compression is really a
 625                  * win, compare the page count read with the blocks on disk,
 626                  * compression must free at least one sector size
 627                  */
 628                 total_in = ALIGN(total_in, PAGE_SIZE);
 629                 if (total_compressed + blocksize <= total_in) {
 630                         *num_added += 1;
 631
 632                         /*
 633                          * The async work queues will take care of doing actual
 634                          * allocation on disk for these compressed pages, and
 635                          * will submit them to the elevator.
 636                          */
 637                         add_async_extent(async_cow, start, total_in,
 638                                         total_compressed, pages, nr_pages,
 639                                         compress_type);
 640
 641                         if (start + total_in < end) {
 642                                 start += total_in;
 643                                 pages = NULL;
 644                                 cond_resched();
 645                                 goto again;
 646                         }
 647                         return;
 648                 }
 649         }
 650         if (pages) {
 651                 /*
 652                  * the compression code ran but failed to make things smaller,
 653                  * free any pages it allocated and our page pointer array
 654                  */
 655                 for (i = 0; i < nr_pages; i++) {
 656                         WARN_ON(pages[i]->mapping);
 657                         put_page(pages[i]);
 658                 }
 659                 kfree(pages);
 660                 pages = NULL;
 661                 total_compressed = 0;
 662                 nr_pages = 0;
 663
 664                 /* flag the file so we don't compress in the future */
 665                 if (!btrfs_test_opt(fs_info, FORCE_COMPRESS) &&
 666                     !(BTRFS_I(inode)->prop_compress)) {
 667                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 668                 }
 669         }
 670 cleanup_and_bail_uncompressed:
 671         /*
 672          * No compression, but we still need to write the pages in the file
 673          * we've been given so far.  redirty the locked page if it corresponds
 674          * to our extent and set things up for the async work queue to run
 675          * cow_file_range to do the normal delalloc dance.
 676          */
 677         if (page_offset(locked_page) >= start &&
 678             page_offset(locked_page) <= end)
 679                 __set_page_dirty_nobuffers(locked_page);
 680                 /* unlocked later on in the async handlers */
 681
 682         if (redirty)
 683                 extent_range_redirty_for_io(inode, start, end);
 684         add_async_extent(async_cow, start, end - start + 1, 0, NULL, 0,
 685                          BTRFS_COMPRESS_NONE);
 686         *num_added += 1;
 687
 688         return;
 689
 690 free_pages_out:
 691         for (i = 0; i < nr_pages; i++) {
 692                 WARN_ON(pages[i]->mapping);
 693                 put_page(pages[i]);
 694         }
 695         kfree(pages);
 696 }
 697
 698 static void free_async_extent_pages(struct async_extent *async_extent)
 699 {
 700         int i;
 701
 702         if (!async_extent->pages)
 703                 return;
 704
 705         for (i = 0; i < async_extent->nr_pages; i++) {
 706                 WARN_ON(async_extent->pages[i]->mapping);
 707                 put_page(async_extent->pages[i]);
 708         }
 709         kfree(async_extent->pages);
 710         async_extent->nr_pages = 0;
 711         async_extent->pages = NULL;
 712 }
 713
 714 /*
 715  * phase two of compressed writeback.  This is the ordered portion
 716  * of the code, which only gets called in the order the work was
 717  * queued.  We walk all the async extents created by compress_file_range
 718  * and send them down to the disk.
 719  */
 720 static noinline void submit_compressed_extents(struct inode *inode,
 721                                               struct async_cow *async_cow)
 722 {
 723         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 724         struct async_extent *async_extent;
 725         u64 alloc_hint = 0;
 726         struct btrfs_key ins;
 727         struct extent_map *em;
 728         struct btrfs_root *root = BTRFS_I(inode)->root;
 729         struct extent_io_tree *io_tree;
 730         int ret = 0;
 731
 732 again:
 733         while (!list_empty(&async_cow->extents)) {
 734                 async_extent = list_entry(async_cow->extents.next,
 735                                           struct async_extent, list);
 736                 list_del(&async_extent->list);
 737
 738                 io_tree = &BTRFS_I(inode)->io_tree;
 739
 740 retry:
 741                 /* did the compression code fall back to uncompressed IO? */
 742                 if (!async_extent->pages) {
 743                         int page_started = 0;
 744                         unsigned long nr_written = 0;
 745
 746                         lock_extent(io_tree, async_extent->start,
 747                                          async_extent->start +
 748                                          async_extent->ram_size - 1);
 749
 750                         /* allocate blocks */
 751                         ret = cow_file_range(inode, async_cow->locked_page,
 752                                              async_extent->start,
 753                                              async_extent->start +
 754                                              async_extent->ram_size - 1,
 755                                              async_extent->start +
 756                                              async_extent->ram_size - 1,
 757                                              &page_started, &nr_written, 0,
 758                                              NULL);
 759
 760                         /* JDM XXX */
 761
 762                         /*
 763                          * if page_started, cow_file_range inserted an
 764                          * inline extent and took care of all the unlocking
 765                          * and IO for us.  Otherwise, we need to submit
 766                          * all those pages down to the drive.
 767                          */
 768                         if (!page_started && !ret)
 769                                 extent_write_locked_range(io_tree,
 770                                                   inode, async_extent->start,
 771                                                   async_extent->start +
 772                                                   async_extent->ram_size - 1,
 773                                                   btrfs_get_extent,
 774                                                   WB_SYNC_ALL);
 775                         else if (ret)
 776                                 unlock_page(async_cow->locked_page);
 777                         kfree(async_extent);
 778                         cond_resched();
 779                         continue;
 780                 }
 781
 782                 lock_extent(io_tree, async_extent->start,
 783                             async_extent->start + async_extent->ram_size - 1);
 784
 785                 ret = btrfs_reserve_extent(root, async_extent->ram_size,
 786                                            async_extent->compressed_size,
 787                                            async_extent->compressed_size,
 788                                            0, alloc_hint, &ins, 1, 1);
 789                 if (ret) {
 790                         free_async_extent_pages(async_extent);
 791
 792                         if (ret == -ENOSPC) {
 793                                 unlock_extent(io_tree, async_extent->start,
 794                                               async_extent->start +
 795                                               async_extent->ram_size - 1);
 796
 797                                 /*
 798                                  * we need to redirty the pages if we decide to
 799                                  * fallback to uncompressed IO, otherwise we
 800                                  * will not submit these pages down to lower
 801                                  * layers.
 802                                  */
 803                                 extent_range_redirty_for_io(inode,
 804                                                 async_extent->start,
 805                                                 async_extent->start +
 806                                                 async_extent->ram_size - 1);
 807
 808                                 goto retry;
 809                         }
 810                         goto out_free;
 811                 }
 812                 /*
 813                  * here we're doing allocation and writeback of the
 814                  * compressed pages
 815                  */
 816                 em = create_io_em(inode, async_extent->start,
 817                                   async_extent->ram_size, /* len */
 818                                   async_extent->start, /* orig_start */
 819                                   ins.objectid, /* block_start */
 820                                   ins.offset, /* block_len */
 821                                   ins.offset, /* orig_block_len */
 822                                   async_extent->ram_size, /* ram_bytes */
 823                                   async_extent->compress_type,
 824                                   BTRFS_ORDERED_COMPRESSED);
 825                 if (IS_ERR(em))
 826                         /* ret value is not necessary due to void function */
 827                         goto out_free_reserve;
 828                 free_extent_map(em);
 829
 830                 ret = btrfs_add_ordered_extent_compress(inode,
 831                                                 async_extent->start,
 832                                                 ins.objectid,
 833                                                 async_extent->ram_size,
 834                                                 ins.offset,
 835                                                 BTRFS_ORDERED_COMPRESSED,
 836                                                 async_extent->compress_type);
 837                 if (ret) {
 838                         btrfs_drop_extent_cache(BTRFS_I(inode),
 839                                                 async_extent->start,
 840                                                 async_extent->start +
 841                                                 async_extent->ram_size - 1, 0);
 842                         goto out_free_reserve;
 843                 }
 844                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 845
 846                 /*
 847                  * clear dirty, set writeback and unlock the pages.
 848                  */
 849                 extent_clear_unlock_delalloc(inode, async_extent->start,
 850                                 async_extent->start +
 851                                 async_extent->ram_size - 1,
 852                                 async_extent->start +
 853                                 async_extent->ram_size - 1,
 854                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 855                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 856                                 PAGE_SET_WRITEBACK);
 857                 if (btrfs_submit_compressed_write(inode,
 858                                     async_extent->start,
 859                                     async_extent->ram_size,
 860                                     ins.objectid,
 861                                     ins.offset, async_extent->pages,
 862                                     async_extent->nr_pages,
 863                                     async_cow->write_flags)) {
 864                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 865                         struct page *p = async_extent->pages[0];
 866                         const u64 start = async_extent->start;
 867                         const u64 end = start + async_extent->ram_size - 1;
 868
 869                         p->mapping = inode->i_mapping;
 870                         tree->ops->writepage_end_io_hook(p, start, end,
 871                                                          NULL, 0);
 872                         p->mapping = NULL;
 873                         extent_clear_unlock_delalloc(inode, start, end, end,
 874                                                      NULL, 0,
 875                                                      PAGE_END_WRITEBACK |
 876                                                      PAGE_SET_ERROR);
 877                         free_async_extent_pages(async_extent);
 878                 }
 879                 alloc_hint = ins.objectid + ins.offset;
 880                 kfree(async_extent);
 881                 cond_resched();
 882         }
 883         return;
 884 out_free_reserve:
 885         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
 886         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
 887 out_free:
 888         extent_clear_unlock_delalloc(inode, async_extent->start,
 889                                      async_extent->start +
 890                                      async_extent->ram_size - 1,
 891                                      async_extent->start +
 892                                      async_extent->ram_size - 1,
 893                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 894                                      EXTENT_DELALLOC_NEW |
 895                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 896                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 897                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 898                                      PAGE_SET_ERROR);
 899         free_async_extent_pages(async_extent);
 900         kfree(async_extent);
 901         goto again;
 902 }
 903
 904 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 905                                       u64 num_bytes)
 906 {
 907         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 908         struct extent_map *em;
 909         u64 alloc_hint = 0;
 910
 911         read_lock(&em_tree->lock);
 912         em = search_extent_mapping(em_tree, start, num_bytes);
 913         if (em) {
 914                 /*
 915                  * if block start isn't an actual block number then find the
 916                  * first block in this inode and use that as a hint.  If that
 917                  * block is also bogus then just don't worry about it.
 918                  */
 919                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 920                         free_extent_map(em);
 921                         em = search_extent_mapping(em_tree, 0, 0);
 922                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 923                                 alloc_hint = em->block_start;
 924                         if (em)
 925                                 free_extent_map(em);
 926                 } else {
 927                         alloc_hint = em->block_start;
 928                         free_extent_map(em);
 929                 }
 930         }
 931         read_unlock(&em_tree->lock);
 932
 933         return alloc_hint;
 934 }
 935
 936 /*
 937  * when extent_io.c finds a delayed allocation range in the file,
 938  * the call backs end up in this code.  The basic idea is to
 939  * allocate extents on disk for the range, and create ordered data structs
 940  * in ram to track those extents.
 941  *
 942  * locked_page is the page that writepage had locked already.  We use
 943  * it to make sure we don't do extra locks or unlocks.
 944  *
 945  * *page_started is set to one if we unlock locked_page and do everything
 946  * required to start IO on it.  It may be clean and already done with
 947  * IO when we return.
 948  */
 949 static noinline int cow_file_range(struct inode *inode,
 950                                    struct page *locked_page,
 951                                    u64 start, u64 end, u64 delalloc_end,
 952                                    int *page_started, unsigned long *nr_written,
 953                                    int unlock, struct btrfs_dedupe_hash *hash)
 954 {
 955         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 956         struct btrfs_root *root = BTRFS_I(inode)->root;
 957         u64 alloc_hint = 0;
 958         u64 num_bytes;
 959         unsigned long ram_size;
 960         u64 disk_num_bytes;
 961         u64 cur_alloc_size = 0;
 962         u64 blocksize = fs_info->sectorsize;
 963         struct btrfs_key ins;
 964         struct extent_map *em;
 965         unsigned clear_bits;
 966         unsigned long page_ops;
 967         bool extent_reserved = false;
 968         int ret = 0;
 969
 970         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
 971                 WARN_ON_ONCE(1);
 972                 ret = -EINVAL;
 973                 goto out_unlock;
 974         }
 975
 976         num_bytes = ALIGN(end - start + 1, blocksize);
 977         num_bytes = max(blocksize,  num_bytes);
 978         disk_num_bytes = num_bytes;
 979
 980         inode_should_defrag(BTRFS_I(inode), start, end, num_bytes, SZ_64K);
 981
 982         if (start == 0) {
 983                 /* lets try to make an inline extent */
 984                 ret = cow_file_range_inline(root, inode, start, end, 0,
 985                                         BTRFS_COMPRESS_NONE, NULL);
 986                 if (ret == 0) {
 987                         /*
 988                          * We use DO_ACCOUNTING here because we need the
 989                          * delalloc_release_metadata to be run _after_ we drop
 990                          * our outstanding extent for clearing delalloc for this
 991                          * range.
 992                          */
 993                         extent_clear_unlock_delalloc(inode, start, end,
 994                                      delalloc_end, NULL,
 995                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 996                                      EXTENT_DELALLOC_NEW | EXTENT_DEFRAG |
 997                                      EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
 998                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 999                                      PAGE_END_WRITEBACK);
1000                         *nr_written = *nr_written +
1001                              (end - start + PAGE_SIZE) / PAGE_SIZE;
1002                         *page_started = 1;
1003                         goto out;
1004                 } else if (ret < 0) {
1005                         goto out_unlock;
1006                 }
1007         }
1008
1009         BUG_ON(disk_num_bytes >
1010                btrfs_super_total_bytes(fs_info->super_copy));
1011
1012         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
1013         btrfs_drop_extent_cache(BTRFS_I(inode), start,
1014                         start + num_bytes - 1, 0);
1015
1016         while (disk_num_bytes > 0) {
1017                 cur_alloc_size = disk_num_bytes;
1018                 ret = btrfs_reserve_extent(root, cur_alloc_size, cur_alloc_size,
1019                                            fs_info->sectorsize, 0, alloc_hint,
1020                                            &ins, 1, 1);
1021                 if (ret < 0)
1022                         goto out_unlock;
1023                 cur_alloc_size = ins.offset;
1024                 extent_reserved = true;
1025
1026                 ram_size = ins.offset;
1027                 em = create_io_em(inode, start, ins.offset, /* len */
1028                                   start, /* orig_start */
1029                                   ins.objectid, /* block_start */
1030                                   ins.offset, /* block_len */
1031                                   ins.offset, /* orig_block_len */
1032                                   ram_size, /* ram_bytes */
1033                                   BTRFS_COMPRESS_NONE, /* compress_type */
1034                                   BTRFS_ORDERED_REGULAR /* type */);
1035                 if (IS_ERR(em))
1036                         goto out_reserve;
1037                 free_extent_map(em);
1038
1039                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1040                                                ram_size, cur_alloc_size, 0);
1041                 if (ret)
1042                         goto out_drop_extent_cache;
1043
1044                 if (root->root_key.objectid ==
1045                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1046                         ret = btrfs_reloc_clone_csums(inode, start,
1047                                                       cur_alloc_size);
1048                         /*
1049                          * Only drop cache here, and process as normal.
1050                          *
1051                          * We must not allow extent_clear_unlock_delalloc()
1052                          * at out_unlock label to free meta of this ordered
1053                          * extent, as its meta should be freed by
1054                          * btrfs_finish_ordered_io().
1055                          *
1056                          * So we must continue until @start is increased to
1057                          * skip current ordered extent.
1058                          */
1059                         if (ret)
1060                                 btrfs_drop_extent_cache(BTRFS_I(inode), start,
1061                                                 start + ram_size - 1, 0);
1062                 }
1063
1064                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1065
1066                 /* we're not doing compressed IO, don't unlock the first
1067                  * page (which the caller expects to stay locked), don't
1068                  * clear any dirty bits and don't set any writeback bits
1069                  *
1070                  * Do set the Private2 bit so we know this page was properly
1071                  * setup for writepage
1072                  */
1073                 page_ops = unlock ? PAGE_UNLOCK : 0;
1074                 page_ops |= PAGE_SET_PRIVATE2;
1075
1076                 extent_clear_unlock_delalloc(inode, start,
1077                                              start + ram_size - 1,
1078                                              delalloc_end, locked_page,
1079                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1080                                              page_ops);
1081                 if (disk_num_bytes < cur_alloc_size)
1082                         disk_num_bytes = 0;
1083                 else
1084                         disk_num_bytes -= cur_alloc_size;
1085                 num_bytes -= cur_alloc_size;
1086                 alloc_hint = ins.objectid + ins.offset;
1087                 start += cur_alloc_size;
1088                 extent_reserved = false;
1089
1090                 /*
1091                  * btrfs_reloc_clone_csums() error, since start is increased
1092                  * extent_clear_unlock_delalloc() at out_unlock label won't
1093                  * free metadata of current ordered extent, we're OK to exit.
1094                  */
1095                 if (ret)
1096                         goto out_unlock;
1097         }
1098 out:
1099         return ret;
1100
1101 out_drop_extent_cache:
1102         btrfs_drop_extent_cache(BTRFS_I(inode), start, start + ram_size - 1, 0);
1103 out_reserve:
1104         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
1105         btrfs_free_reserved_extent(fs_info, ins.objectid, ins.offset, 1);
1106 out_unlock:
1107         clear_bits = EXTENT_LOCKED | EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
1108                 EXTENT_DEFRAG | EXTENT_CLEAR_META_RESV;
1109         page_ops = PAGE_UNLOCK | PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
1110                 PAGE_END_WRITEBACK;
1111         /*
1112          * If we reserved an extent for our delalloc range (or a subrange) and
1113          * failed to create the respective ordered extent, then it means that
1114          * when we reserved the extent we decremented the extent's size from
1115          * the data space_info's bytes_may_use counter and incremented the
1116          * space_info's bytes_reserved counter by the same amount. We must make
1117          * sure extent_clear_unlock_delalloc() does not try to decrement again
1118          * the data space_info's bytes_may_use counter, therefore we do not pass
1119          * it the flag EXTENT_CLEAR_DATA_RESV.
1120          */
1121         if (extent_reserved) {
1122                 extent_clear_unlock_delalloc(inode, start,
1123                                              start + cur_alloc_size,
1124                                              start + cur_alloc_size,
1125                                              locked_page,
1126                                              clear_bits,
1127                                              page_ops);
1128                 start += cur_alloc_size;
1129                 if (start >= end)
1130                         goto out;
1131         }
1132         extent_clear_unlock_delalloc(inode, start, end, delalloc_end,
1133                                      locked_page,
1134                                      clear_bits | EXTENT_CLEAR_DATA_RESV,
1135                                      page_ops);
1136         goto out;
1137 }
1138
1139 /*
1140  * work queue call back to started compression on a file and pages
1141  */
1142 static noinline void async_cow_start(struct btrfs_work *work)
1143 {
1144         struct async_cow *async_cow;
1145         int num_added = 0;
1146         async_cow = container_of(work, struct async_cow, work);
1147
1148         compress_file_range(async_cow->inode, async_cow->locked_page,
1149                             async_cow->start, async_cow->end, async_cow,
1150                             &num_added);
1151         if (num_added == 0) {
1152                 btrfs_add_delayed_iput(async_cow->inode);
1153                 async_cow->inode = NULL;
1154         }
1155 }
1156
1157 /*
1158  * work queue call back to submit previously compressed pages
1159  */
1160 static noinline void async_cow_submit(struct btrfs_work *work)
1161 {
1162         struct btrfs_fs_info *fs_info;
1163         struct async_cow *async_cow;
1164         struct btrfs_root *root;
1165         unsigned long nr_pages;
1166
1167         async_cow = container_of(work, struct async_cow, work);
1168
1169         root = async_cow->root;
1170         fs_info = root->fs_info;
1171         nr_pages = (async_cow->end - async_cow->start + PAGE_SIZE) >>
1172                 PAGE_SHIFT;
1173
1174         /*
1175          * atomic_sub_return implies a barrier for waitqueue_active
1176          */
1177         if (atomic_sub_return(nr_pages, &fs_info->async_delalloc_pages) <
1178             5 * SZ_1M &&
1179             waitqueue_active(&fs_info->async_submit_wait))
1180                 wake_up(&fs_info->async_submit_wait);
1181
1182         if (async_cow->inode)
1183                 submit_compressed_extents(async_cow->inode, async_cow);
1184 }
1185
1186 static noinline void async_cow_free(struct btrfs_work *work)
1187 {
1188         struct async_cow *async_cow;
1189         async_cow = container_of(work, struct async_cow, work);
1190         if (async_cow->inode)
1191                 btrfs_add_delayed_iput(async_cow->inode);
1192         kfree(async_cow);
1193 }
1194
1195 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1196                                 u64 start, u64 end, int *page_started,
1197                                 unsigned long *nr_written,
1198                                 unsigned int write_flags)
1199 {
1200         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1201         struct async_cow *async_cow;
1202         struct btrfs_root *root = BTRFS_I(inode)->root;
1203         unsigned long nr_pages;
1204         u64 cur_end;
1205
1206         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1207                          1, 0, NULL, GFP_NOFS);
1208         while (start < end) {
1209                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1210                 BUG_ON(!async_cow); /* -ENOMEM */
1211                 async_cow->inode = igrab(inode);
1212                 async_cow->root = root;
1213                 async_cow->locked_page = locked_page;
1214                 async_cow->start = start;
1215                 async_cow->write_flags = write_flags;
1216
1217                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1218                     !btrfs_test_opt(fs_info, FORCE_COMPRESS))
1219                         cur_end = end;
1220                 else
1221                         cur_end = min(end, start + SZ_512K - 1);
1222
1223                 async_cow->end = cur_end;
1224                 INIT_LIST_HEAD(&async_cow->extents);
1225
1226                 btrfs_init_work(&async_cow->work,
1227                                 btrfs_delalloc_helper,
1228                                 async_cow_start, async_cow_submit,
1229                                 async_cow_free);
1230
1231                 nr_pages = (cur_end - start + PAGE_SIZE) >>
1232                         PAGE_SHIFT;
1233                 atomic_add(nr_pages, &fs_info->async_delalloc_pages);
1234
1235                 btrfs_queue_work(fs_info->delalloc_workers, &async_cow->work);
1236
1237                 *nr_written += nr_pages;
1238                 start = cur_end + 1;
1239         }
1240         *page_started = 1;
1241         return 0;
1242 }
1243
1244 static noinline int csum_exist_in_range(struct btrfs_fs_info *fs_info,
1245                                         u64 bytenr, u64 num_bytes)
1246 {
1247         int ret;
1248         struct btrfs_ordered_sum *sums;
1249         LIST_HEAD(list);
1250
1251         ret = btrfs_lookup_csums_range(fs_info->csum_root, bytenr,
1252                                        bytenr + num_bytes - 1, &list, 0);
1253         if (ret == 0 && list_empty(&list))
1254                 return 0;
1255
1256         while (!list_empty(&list)) {
1257                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1258                 list_del(&sums->list);
1259                 kfree(sums);
1260         }
1261         return 1;
1262 }
1263
1264 /*
1265  * when nowcow writeback call back.  This checks for snapshots or COW copies
1266  * of the extents that exist in the file, and COWs the file as required.
1267  *
1268  * If no cow copies or snapshots exist, we write directly to the existing
1269  * blocks on disk
1270  */
1271 static noinline int run_delalloc_nocow(struct inode *inode,
1272                                        struct page *locked_page,
1273                               u64 start, u64 end, int *page_started, int force,
1274                               unsigned long *nr_written)
1275 {
1276         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1277         struct btrfs_root *root = BTRFS_I(inode)->root;
1278         struct extent_buffer *leaf;
1279         struct btrfs_path *path;
1280         struct btrfs_file_extent_item *fi;
1281         struct btrfs_key found_key;
1282         struct extent_map *em;
1283         u64 cow_start;
1284         u64 cur_offset;
1285         u64 extent_end;
1286         u64 extent_offset;
1287         u64 disk_bytenr;
1288         u64 num_bytes;
1289         u64 disk_num_bytes;
1290         u64 ram_bytes;
1291         int extent_type;
1292         int ret, err;
1293         int type;
1294         int nocow;
1295         int check_prev = 1;
1296         bool nolock;
1297         u64 ino = btrfs_ino(BTRFS_I(inode));
1298
1299         path = btrfs_alloc_path();
1300         if (!path) {
1301                 extent_clear_unlock_delalloc(inode, start, end, end,
1302                                              locked_page,
1303                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1304                                              EXTENT_DO_ACCOUNTING |
1305                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1306                                              PAGE_CLEAR_DIRTY |
1307                                              PAGE_SET_WRITEBACK |
1308                                              PAGE_END_WRITEBACK);
1309                 return -ENOMEM;
1310         }
1311
1312         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
1313
1314         cow_start = (u64)-1;
1315         cur_offset = start;
1316         while (1) {
1317                 ret = btrfs_lookup_file_extent(NULL, root, path, ino,
1318                                                cur_offset, 0);
1319                 if (ret < 0)
1320                         goto error;
1321                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1322                         leaf = path->nodes[0];
1323                         btrfs_item_key_to_cpu(leaf, &found_key,
1324                                               path->slots[0] - 1);
1325                         if (found_key.objectid == ino &&
1326                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1327                                 path->slots[0]--;
1328                 }
1329                 check_prev = 0;
1330 next_slot:
1331                 leaf = path->nodes[0];
1332                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1333                         ret = btrfs_next_leaf(root, path);
1334                         if (ret < 0)
1335                                 goto error;
1336                         if (ret > 0)
1337                                 break;
1338                         leaf = path->nodes[0];
1339                 }
1340
1341                 nocow = 0;
1342                 disk_bytenr = 0;
1343                 num_bytes = 0;
1344                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1345
1346                 if (found_key.objectid > ino)
1347                         break;
1348                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1349                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1350                         path->slots[0]++;
1351                         goto next_slot;
1352                 }
1353                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1354                     found_key.offset > end)
1355                         break;
1356
1357                 if (found_key.offset > cur_offset) {
1358                         extent_end = found_key.offset;
1359                         extent_type = 0;
1360                         goto out_check;
1361                 }
1362
1363                 fi = btrfs_item_ptr(leaf, path->slots[0],
1364                                     struct btrfs_file_extent_item);
1365                 extent_type = btrfs_file_extent_type(leaf, fi);
1366
1367                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1368                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1369                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1370                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1371                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1372                         extent_end = found_key.offset +
1373                                 btrfs_file_extent_num_bytes(leaf, fi);
1374                         disk_num_bytes =
1375                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1376                         if (extent_end <= start) {
1377                                 path->slots[0]++;
1378                                 goto next_slot;
1379                         }
1380                         if (disk_bytenr == 0)
1381                                 goto out_check;
1382                         if (btrfs_file_extent_compression(leaf, fi) ||
1383                             btrfs_file_extent_encryption(leaf, fi) ||
1384                             btrfs_file_extent_other_encoding(leaf, fi))
1385                                 goto out_check;
1386                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1387                                 goto out_check;
1388                         if (btrfs_extent_readonly(fs_info, disk_bytenr))
1389                                 goto out_check;
1390                         if (btrfs_cross_ref_exist(root, ino,
1391                                                   found_key.offset -
1392                                                   extent_offset, disk_bytenr))
1393                                 goto out_check;
1394                         disk_bytenr += extent_offset;
1395                         disk_bytenr += cur_offset - found_key.offset;
1396                         num_bytes = min(end + 1, extent_end) - cur_offset;
1397                         /*
1398                          * if there are pending snapshots for this root,
1399                          * we fall into common COW way.
1400                          */
1401                         if (!nolock) {
1402                                 err = btrfs_start_write_no_snapshotting(root);
1403                                 if (!err)
1404                                         goto out_check;
1405                         }
1406                         /*
1407                          * force cow if csum exists in the range.
1408                          * this ensure that csum for a given extent are
1409                          * either valid or do not exist.
1410                          */
1411                         if (csum_exist_in_range(fs_info, disk_bytenr,
1412                                                 num_bytes)) {
1413                                 if (!nolock)
1414                                         btrfs_end_write_no_snapshotting(root);
1415                                 goto out_check;
1416                         }
1417                         if (!btrfs_inc_nocow_writers(fs_info, disk_bytenr)) {
1418                                 if (!nolock)
1419                                         btrfs_end_write_no_snapshotting(root);
1420                                 goto out_check;
1421                         }
1422                         nocow = 1;
1423                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1424                         extent_end = found_key.offset +
1425                                 btrfs_file_extent_inline_len(leaf,
1426                                                      path->slots[0], fi);
1427                         extent_end = ALIGN(extent_end,
1428                                            fs_info->sectorsize);
1429                 } else {
1430                         BUG_ON(1);
1431                 }
1432 out_check:
1433                 if (extent_end <= start) {
1434                         path->slots[0]++;
1435                         if (!nolock && nocow)
1436                                 btrfs_end_write_no_snapshotting(root);
1437                         if (nocow)
1438                                 btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1439                         goto next_slot;
1440                 }
1441                 if (!nocow) {
1442                         if (cow_start == (u64)-1)
1443                                 cow_start = cur_offset;
1444                         cur_offset = extent_end;
1445                         if (cur_offset > end)
1446                                 break;
1447                         path->slots[0]++;
1448                         goto next_slot;
1449                 }
1450
1451                 btrfs_release_path(path);
1452                 if (cow_start != (u64)-1) {
1453                         ret = cow_file_range(inode, locked_page,
1454                                              cow_start, found_key.offset - 1,
1455                                              end, page_started, nr_written, 1,
1456                                              NULL);
1457                         if (ret) {
1458                                 if (!nolock && nocow)
1459                                         btrfs_end_write_no_snapshotting(root);
1460                                 if (nocow)
1461                                         btrfs_dec_nocow_writers(fs_info,
1462                                                                 disk_bytenr);
1463                                 goto error;
1464                         }
1465                         cow_start = (u64)-1;
1466                 }
1467
1468                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1469                         u64 orig_start = found_key.offset - extent_offset;
1470
1471                         em = create_io_em(inode, cur_offset, num_bytes,
1472                                           orig_start,
1473                                           disk_bytenr, /* block_start */
1474                                           num_bytes, /* block_len */
1475                                           disk_num_bytes, /* orig_block_len */
1476                                           ram_bytes, BTRFS_COMPRESS_NONE,
1477                                           BTRFS_ORDERED_PREALLOC);
1478                         if (IS_ERR(em)) {
1479                                 if (!nolock && nocow)
1480                                         btrfs_end_write_no_snapshotting(root);
1481                                 if (nocow)
1482                                         btrfs_dec_nocow_writers(fs_info,
1483                                                                 disk_bytenr);
1484                                 ret = PTR_ERR(em);
1485                                 goto error;
1486                         }
1487                         free_extent_map(em);
1488                 }
1489
1490                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1491                         type = BTRFS_ORDERED_PREALLOC;
1492                 } else {
1493                         type = BTRFS_ORDERED_NOCOW;
1494                 }
1495
1496                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1497                                                num_bytes, num_bytes, type);
1498                 if (nocow)
1499                         btrfs_dec_nocow_writers(fs_info, disk_bytenr);
1500                 BUG_ON(ret); /* -ENOMEM */
1501
1502                 if (root->root_key.objectid ==
1503                     BTRFS_DATA_RELOC_TREE_OBJECTID)
1504                         /*
1505                          * Error handled later, as we must prevent
1506                          * extent_clear_unlock_delalloc() in error handler
1507                          * from freeing metadata of created ordered extent.
1508                          */
1509                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1510                                                       num_bytes);
1511
1512                 extent_clear_unlock_delalloc(inode, cur_offset,
1513                                              cur_offset + num_bytes - 1, end,
1514                                              locked_page, EXTENT_LOCKED |
1515                                              EXTENT_DELALLOC |
1516                                              EXTENT_CLEAR_DATA_RESV,
1517                                              PAGE_UNLOCK | PAGE_SET_PRIVATE2);
1518
1519                 if (!nolock && nocow)
1520                         btrfs_end_write_no_snapshotting(root);
1521                 cur_offset = extent_end;
1522
1523                 /*
1524                  * btrfs_reloc_clone_csums() error, now we're OK to call error
1525                  * handler, as metadata for created ordered extent will only
1526                  * be freed by btrfs_finish_ordered_io().
1527                  */
1528                 if (ret)
1529                         goto error;
1530                 if (cur_offset > end)
1531                         break;
1532         }
1533         btrfs_release_path(path);
1534
1535         if (cur_offset <= end && cow_start == (u64)-1) {
1536                 cow_start = cur_offset;
1537                 cur_offset = end;
1538         }
1539
1540         if (cow_start != (u64)-1) {
1541                 ret = cow_file_range(inode, locked_page, cow_start, end, end,
1542                                      page_started, nr_written, 1, NULL);
1543                 if (ret)
1544                         goto error;
1545         }
1546
1547 error:
1548         if (ret && cur_offset < end)
1549                 extent_clear_unlock_delalloc(inode, cur_offset, end, end,
1550                                              locked_page, EXTENT_LOCKED |
1551                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1552                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1553                                              PAGE_CLEAR_DIRTY |
1554                                              PAGE_SET_WRITEBACK |
1555                                              PAGE_END_WRITEBACK);
1556         btrfs_free_path(path);
1557         return ret;
1558 }
1559
1560 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1561 {
1562
1563         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1564             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1565                 return 0;
1566
1567         /*
1568          * @defrag_bytes is a hint value, no spinlock held here,
1569          * if is not zero, it means the file is defragging.
1570          * Force cow if given extent needs to be defragged.
1571          */
1572         if (BTRFS_I(inode)->defrag_bytes &&
1573             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1574                            EXTENT_DEFRAG, 0, NULL))
1575                 return 1;
1576
1577         return 0;
1578 }
1579
1580 /*
1581  * extent_io.c call back to do delayed allocation processing
1582  */
1583 static int run_delalloc_range(void *private_data, struct page *locked_page,
1584                               u64 start, u64 end, int *page_started,
1585                               unsigned long *nr_written,
1586                               struct writeback_control *wbc)
1587 {
1588         struct inode *inode = private_data;
1589         int ret;
1590         int force_cow = need_force_cow(inode, start, end);
1591         unsigned int write_flags = wbc_to_write_flags(wbc);
1592
1593         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1594                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1595                                          page_started, 1, nr_written);
1596         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1597                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1598                                          page_started, 0, nr_written);
1599         } else if (!inode_need_compress(inode, start, end)) {
1600                 ret = cow_file_range(inode, locked_page, start, end, end,
1601                                       page_started, nr_written, 1, NULL);
1602         } else {
1603                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1604                         &BTRFS_I(inode)->runtime_flags);
1605                 ret = cow_file_range_async(inode, locked_page, start, end,
1606                                            page_started, nr_written,
1607                                            write_flags);
1608         }
1609         if (ret)
1610                 btrfs_cleanup_ordered_extents(inode, start, end - start + 1);
1611         return ret;
1612 }
1613
1614 static void btrfs_split_extent_hook(void *private_data,
1615                                     struct extent_state *orig, u64 split)
1616 {
1617         struct inode *inode = private_data;
1618         u64 size;
1619
1620         /* not delalloc, ignore it */
1621         if (!(orig->state & EXTENT_DELALLOC))
1622                 return;
1623
1624         size = orig->end - orig->start + 1;
1625         if (size > BTRFS_MAX_EXTENT_SIZE) {
1626                 u32 num_extents;
1627                 u64 new_size;
1628
1629                 /*
1630                  * See the explanation in btrfs_merge_extent_hook, the same
1631                  * applies here, just in reverse.
1632                  */
1633                 new_size = orig->end - split + 1;
1634                 num_extents = count_max_extents(new_size);
1635                 new_size = split - orig->start;
1636                 num_extents += count_max_extents(new_size);
1637                 if (count_max_extents(size) >= num_extents)
1638                         return;
1639         }
1640
1641         spin_lock(&BTRFS_I(inode)->lock);
1642         btrfs_mod_outstanding_extents(BTRFS_I(inode), 1);
1643         spin_unlock(&BTRFS_I(inode)->lock);
1644 }
1645
1646 /*
1647  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1648  * extents so we can keep track of new extents that are just merged onto old
1649  * extents, such as when we are doing sequential writes, so we can properly
1650  * account for the metadata space we'll need.
1651  */
1652 static void btrfs_merge_extent_hook(void *private_data,
1653                                     struct extent_state *new,
1654                                     struct extent_state *other)
1655 {
1656         struct inode *inode = private_data;
1657         u64 new_size, old_size;
1658         u32 num_extents;
1659
1660         /* not delalloc, ignore it */
1661         if (!(other->state & EXTENT_DELALLOC))
1662                 return;
1663
1664         if (new->start > other->start)
1665                 new_size = new->end - other->start + 1;
1666         else
1667                 new_size = other->end - new->start + 1;
1668
1669         /* we're not bigger than the max, unreserve the space and go */
1670         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1671                 spin_lock(&BTRFS_I(inode)->lock);
1672                 btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1673                 spin_unlock(&BTRFS_I(inode)->lock);
1674                 return;
1675         }
1676
1677         /*
1678          * We have to add up either side to figure out how many extents were
1679          * accounted for before we merged into one big extent.  If the number of
1680          * extents we accounted for is <= the amount we need for the new range
1681          * then we can return, otherwise drop.  Think of it like this
1682          *
1683          * [ 4k][MAX_SIZE]
1684          *
1685          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1686          * need 2 outstanding extents, on one side we have 1 and the other side
1687          * we have 1 so they are == and we can return.  But in this case
1688          *
1689          * [MAX_SIZE+4k][MAX_SIZE+4k]
1690          *
1691          * Each range on their own accounts for 2 extents, but merged together
1692          * they are only 3 extents worth of accounting, so we need to drop in
1693          * this case.
1694          */
1695         old_size = other->end - other->start + 1;
1696         num_extents = count_max_extents(old_size);
1697         old_size = new->end - new->start + 1;
1698         num_extents += count_max_extents(old_size);
1699         if (count_max_extents(new_size) >= num_extents)
1700                 return;
1701
1702         spin_lock(&BTRFS_I(inode)->lock);
1703         btrfs_mod_outstanding_extents(BTRFS_I(inode), -1);
1704         spin_unlock(&BTRFS_I(inode)->lock);
1705 }
1706
1707 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1708                                       struct inode *inode)
1709 {
1710         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1711
1712         spin_lock(&root->delalloc_lock);
1713         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1714                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1715                               &root->delalloc_inodes);
1716                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1717                         &BTRFS_I(inode)->runtime_flags);
1718                 root->nr_delalloc_inodes++;
1719                 if (root->nr_delalloc_inodes == 1) {
1720                         spin_lock(&fs_info->delalloc_root_lock);
1721                         BUG_ON(!list_empty(&root->delalloc_root));
1722                         list_add_tail(&root->delalloc_root,
1723                                       &fs_info->delalloc_roots);
1724                         spin_unlock(&fs_info->delalloc_root_lock);
1725                 }
1726         }
1727         spin_unlock(&root->delalloc_lock);
1728 }
1729
1730 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1731                                      struct btrfs_inode *inode)
1732 {
1733         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1734
1735         spin_lock(&root->delalloc_lock);
1736         if (!list_empty(&inode->delalloc_inodes)) {
1737                 list_del_init(&inode->delalloc_inodes);
1738                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1739                           &inode->runtime_flags);
1740                 root->nr_delalloc_inodes--;
1741                 if (!root->nr_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         spin_unlock(&root->delalloc_lock);
1749 }
1750
1751 /*
1752  * extent_io.c set_bit_hook, used to track delayed allocation
1753  * bytes in this file, and to maintain the list of inodes that
1754  * have pending delalloc work to be done.
1755  */
1756 static void btrfs_set_bit_hook(void *private_data,
1757                                struct extent_state *state, unsigned *bits)
1758 {
1759         struct inode *inode = private_data;
1760
1761         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1762
1763         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1764                 WARN_ON(1);
1765         /*
1766          * set_bit and clear bit hooks normally require _irqsave/restore
1767          * but in this case, we are only testing for the DELALLOC
1768          * bit, which is only set or cleared with irqs on
1769          */
1770         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1771                 struct btrfs_root *root = BTRFS_I(inode)->root;
1772                 u64 len = state->end + 1 - state->start;
1773                 u32 num_extents = count_max_extents(len);
1774                 bool do_list = !btrfs_is_free_space_inode(BTRFS_I(inode));
1775
1776                 spin_lock(&BTRFS_I(inode)->lock);
1777                 btrfs_mod_outstanding_extents(BTRFS_I(inode), num_extents);
1778                 spin_unlock(&BTRFS_I(inode)->lock);
1779
1780                 /* For sanity tests */
1781                 if (btrfs_is_testing(fs_info))
1782                         return;
1783
1784                 percpu_counter_add_batch(&fs_info->delalloc_bytes, len,
1785                                          fs_info->delalloc_batch);
1786                 spin_lock(&BTRFS_I(inode)->lock);
1787                 BTRFS_I(inode)->delalloc_bytes += len;
1788                 if (*bits & EXTENT_DEFRAG)
1789                         BTRFS_I(inode)->defrag_bytes += len;
1790                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1791                                          &BTRFS_I(inode)->runtime_flags))
1792                         btrfs_add_delalloc_inodes(root, inode);
1793                 spin_unlock(&BTRFS_I(inode)->lock);
1794         }
1795
1796         if (!(state->state & EXTENT_DELALLOC_NEW) &&
1797             (*bits & EXTENT_DELALLOC_NEW)) {
1798                 spin_lock(&BTRFS_I(inode)->lock);
1799                 BTRFS_I(inode)->new_delalloc_bytes += state->end + 1 -
1800                         state->start;
1801                 spin_unlock(&BTRFS_I(inode)->lock);
1802         }
1803 }
1804
1805 /*
1806  * extent_io.c clear_bit_hook, see set_bit_hook for why
1807  */
1808 static void btrfs_clear_bit_hook(void *private_data,
1809                                  struct extent_state *state,
1810                                  unsigned *bits)
1811 {
1812         struct btrfs_inode *inode = BTRFS_I((struct inode *)private_data);
1813         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1814         u64 len = state->end + 1 - state->start;
1815         u32 num_extents = count_max_extents(len);
1816
1817         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG)) {
1818                 spin_lock(&inode->lock);
1819                 inode->defrag_bytes -= len;
1820                 spin_unlock(&inode->lock);
1821         }
1822
1823         /*
1824          * set_bit and clear bit hooks normally require _irqsave/restore
1825          * but in this case, we are only testing for the DELALLOC
1826          * bit, which is only set or cleared with irqs on
1827          */
1828         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1829                 struct btrfs_root *root = inode->root;
1830                 bool do_list = !btrfs_is_free_space_inode(inode);
1831
1832                 spin_lock(&inode->lock);
1833                 btrfs_mod_outstanding_extents(inode, -num_extents);
1834                 spin_unlock(&inode->lock);
1835
1836                 /*
1837                  * We don't reserve metadata space for space cache inodes so we
1838                  * don't need to call dellalloc_release_metadata if there is an
1839                  * error.
1840                  */
1841                 if (*bits & EXTENT_CLEAR_META_RESV &&
1842                     root != fs_info->tree_root)
1843                         btrfs_delalloc_release_metadata(inode, len);
1844
1845                 /* For sanity tests. */
1846                 if (btrfs_is_testing(fs_info))
1847                         return;
1848
1849                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID &&
1850                     do_list && !(state->state & EXTENT_NORESERVE) &&
1851                     (*bits & EXTENT_CLEAR_DATA_RESV))
1852                         btrfs_free_reserved_data_space_noquota(
1853                                         &inode->vfs_inode,
1854                                         state->start, len);
1855
1856                 percpu_counter_add_batch(&fs_info->delalloc_bytes, -len,
1857                                          fs_info->delalloc_batch);
1858                 spin_lock(&inode->lock);
1859                 inode->delalloc_bytes -= len;
1860                 if (do_list && inode->delalloc_bytes == 0 &&
1861                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1862                                         &inode->runtime_flags))
1863                         btrfs_del_delalloc_inode(root, inode);
1864                 spin_unlock(&inode->lock);
1865         }
1866
1867         if ((state->state & EXTENT_DELALLOC_NEW) &&
1868             (*bits & EXTENT_DELALLOC_NEW)) {
1869                 spin_lock(&inode->lock);
1870                 ASSERT(inode->new_delalloc_bytes >= len);
1871                 inode->new_delalloc_bytes -= len;
1872                 spin_unlock(&inode->lock);
1873         }
1874 }
1875
1876 /*
1877  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1878  * we don't create bios that span stripes or chunks
1879  *
1880  * return 1 if page cannot be merged to bio
1881  * return 0 if page can be merged to bio
1882  * return error otherwise
1883  */
1884 int btrfs_merge_bio_hook(struct page *page, unsigned long offset,
1885                          size_t size, struct bio *bio,
1886                          unsigned long bio_flags)
1887 {
1888         struct inode *inode = page->mapping->host;
1889         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1890         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1891         u64 length = 0;
1892         u64 map_length;
1893         int ret;
1894
1895         if (bio_flags & EXTENT_BIO_COMPRESSED)
1896                 return 0;
1897
1898         length = bio->bi_iter.bi_size;
1899         map_length = length;
1900         ret = btrfs_map_block(fs_info, btrfs_op(bio), logical, &map_length,
1901                               NULL, 0);
1902         if (ret < 0)
1903                 return ret;
1904         if (map_length < length + size)
1905                 return 1;
1906         return 0;
1907 }
1908
1909 /*
1910  * in order to insert checksums into the metadata in large chunks,
1911  * we wait until bio submission time.   All the pages in the bio are
1912  * checksummed and sums are attached onto the ordered extent record.
1913  *
1914  * At IO completion time the cums attached on the ordered extent record
1915  * are inserted into the btree
1916  */
1917 static blk_status_t __btrfs_submit_bio_start(void *private_data, struct bio *bio,
1918                                     int mirror_num, unsigned long bio_flags,
1919                                     u64 bio_offset)
1920 {
1921         struct inode *inode = private_data;
1922         blk_status_t ret = 0;
1923
1924         ret = btrfs_csum_one_bio(inode, bio, 0, 0);
1925         BUG_ON(ret); /* -ENOMEM */
1926         return 0;
1927 }
1928
1929 /*
1930  * in order to insert checksums into the metadata in large chunks,
1931  * we wait until bio submission time.   All the pages in the bio are
1932  * checksummed and sums are attached onto the ordered extent record.
1933  *
1934  * At IO completion time the cums attached on the ordered extent record
1935  * are inserted into the btree
1936  */
1937 static blk_status_t __btrfs_submit_bio_done(void *private_data, struct bio *bio,
1938                           int mirror_num, unsigned long bio_flags,
1939                           u64 bio_offset)
1940 {
1941         struct inode *inode = private_data;
1942         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1943         blk_status_t ret;
1944
1945         ret = btrfs_map_bio(fs_info, bio, mirror_num, 1);
1946         if (ret) {
1947                 bio->bi_status = ret;
1948                 bio_endio(bio);
1949         }
1950         return ret;
1951 }
1952
1953 /*
1954  * extent_io.c submission hook. This does the right thing for csum calculation
1955  * on write, or reading the csums from the tree before a read
1956  */
1957 static blk_status_t btrfs_submit_bio_hook(void *private_data, struct bio *bio,
1958                                  int mirror_num, unsigned long bio_flags,
1959                                  u64 bio_offset)
1960 {
1961         struct inode *inode = private_data;
1962         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1963         struct btrfs_root *root = BTRFS_I(inode)->root;
1964         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1965         blk_status_t ret = 0;
1966         int skip_sum;
1967         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1968
1969         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1970
1971         if (btrfs_is_free_space_inode(BTRFS_I(inode)))
1972                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1973
1974         if (bio_op(bio) != REQ_OP_WRITE) {
1975                 ret = btrfs_bio_wq_end_io(fs_info, bio, metadata);
1976                 if (ret)
1977                         goto out;
1978
1979                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1980                         ret = btrfs_submit_compressed_read(inode, bio,
1981                                                            mirror_num,
1982                                                            bio_flags);
1983                         goto out;
1984                 } else if (!skip_sum) {
1985                         ret = btrfs_lookup_bio_sums(inode, bio, NULL);
1986                         if (ret)
1987                                 goto out;
1988                 }
1989                 goto mapit;
1990         } else if (async && !skip_sum) {
1991                 /* csum items have already been cloned */
1992                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1993                         goto mapit;
1994                 /* we're doing a write, do the async checksumming */
1995                 ret = btrfs_wq_submit_bio(fs_info, bio, mirror_num, bio_flags,
1996                                           bio_offset, inode,
1997                                           __btrfs_submit_bio_start,
1998                                           __btrfs_submit_bio_done);
1999                 goto out;
2000         } else if (!skip_sum) {
2001                 ret = btrfs_csum_one_bio(inode, bio, 0, 0);
2002                 if (ret)
2003                         goto out;
2004         }
2005
2006 mapit:
2007         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
2008
2009 out:
2010         if (ret) {
2011                 bio->bi_status = ret;
2012                 bio_endio(bio);
2013         }
2014         return ret;
2015 }
2016
2017 /*
2018  * given a list of ordered sums record them in the inode.  This happens
2019  * at IO completion time based on sums calculated at bio submission time.
2020  */
2021 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
2022                              struct inode *inode, struct list_head *list)
2023 {
2024         struct btrfs_ordered_sum *sum;
2025
2026         list_for_each_entry(sum, list, list) {
2027                 trans->adding_csums = 1;
2028                 btrfs_csum_file_blocks(trans,
2029                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
2030                 trans->adding_csums = 0;
2031         }
2032         return 0;
2033 }
2034
2035 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
2036                               unsigned int extra_bits,
2037                               struct extent_state **cached_state, int dedupe)
2038 {
2039         WARN_ON((end & (PAGE_SIZE - 1)) == 0);
2040         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
2041                                    extra_bits, cached_state);
2042 }
2043
2044 /* see btrfs_writepage_start_hook for details on why this is required */
2045 struct btrfs_writepage_fixup {
2046         struct page *page;
2047         struct btrfs_work work;
2048 };
2049
2050 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
2051 {
2052         struct btrfs_writepage_fixup *fixup;
2053         struct btrfs_ordered_extent *ordered;
2054         struct extent_state *cached_state = NULL;
2055         struct extent_changeset *data_reserved = NULL;
2056         struct page *page;
2057         struct inode *inode;
2058         u64 page_start;
2059         u64 page_end;
2060         int ret;
2061
2062         fixup = container_of(work, struct btrfs_writepage_fixup, work);
2063         page = fixup->page;
2064 again:
2065         lock_page(page);
2066         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
2067                 ClearPageChecked(page);
2068                 goto out_page;
2069         }
2070
2071         inode = page->mapping->host;
2072         page_start = page_offset(page);
2073         page_end = page_offset(page) + PAGE_SIZE - 1;
2074
2075         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
2076                          &cached_state);
2077
2078         /* already ordered? We're done */
2079         if (PagePrivate2(page))
2080                 goto out;
2081
2082         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
2083                                         PAGE_SIZE);
2084         if (ordered) {
2085                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2086                                      page_end, &cached_state, GFP_NOFS);
2087                 unlock_page(page);
2088                 btrfs_start_ordered_extent(inode, ordered, 1);
2089                 btrfs_put_ordered_extent(ordered);
2090                 goto again;
2091         }
2092
2093         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
2094                                            PAGE_SIZE);
2095         if (ret) {
2096                 mapping_set_error(page->mapping, ret);
2097                 end_extent_writepage(page, ret, page_start, page_end);
2098                 ClearPageChecked(page);
2099                 goto out;
2100          }
2101
2102         btrfs_set_extent_delalloc(inode, page_start, page_end, 0, &cached_state,
2103                                   0);
2104         ClearPageChecked(page);
2105         set_page_dirty(page);
2106         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
2107 out:
2108         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2109                              &cached_state, GFP_NOFS);
2110 out_page:
2111         unlock_page(page);
2112         put_page(page);
2113         kfree(fixup);
2114         extent_changeset_free(data_reserved);
2115 }
2116
2117 /*
2118  * There are a few paths in the higher layers of the kernel that directly
2119  * set the page dirty bit without asking the filesystem if it is a
2120  * good idea.  This causes problems because we want to make sure COW
2121  * properly happens and the data=ordered rules are followed.
2122  *
2123  * In our case any range that doesn't have the ORDERED bit set
2124  * hasn't been properly setup for IO.  We kick off an async process
2125  * to fix it up.  The async helper will wait for ordered extents, set
2126  * the delalloc bit and make it safe to write the page.
2127  */
2128 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2129 {
2130         struct inode *inode = page->mapping->host;
2131         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2132         struct btrfs_writepage_fixup *fixup;
2133
2134         /* this page is properly in the ordered list */
2135         if (TestClearPagePrivate2(page))
2136                 return 0;
2137
2138         if (PageChecked(page))
2139                 return -EAGAIN;
2140
2141         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2142         if (!fixup)
2143                 return -EAGAIN;
2144
2145         SetPageChecked(page);
2146         get_page(page);
2147         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2148                         btrfs_writepage_fixup_worker, NULL, NULL);
2149         fixup->page = page;
2150         btrfs_queue_work(fs_info->fixup_workers, &fixup->work);
2151         return -EBUSY;
2152 }
2153
2154 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2155                                        struct inode *inode, u64 file_pos,
2156                                        u64 disk_bytenr, u64 disk_num_bytes,
2157                                        u64 num_bytes, u64 ram_bytes,
2158                                        u8 compression, u8 encryption,
2159                                        u16 other_encoding, int extent_type)
2160 {
2161         struct btrfs_root *root = BTRFS_I(inode)->root;
2162         struct btrfs_file_extent_item *fi;
2163         struct btrfs_path *path;
2164         struct extent_buffer *leaf;
2165         struct btrfs_key ins;
2166         u64 qg_released;
2167         int extent_inserted = 0;
2168         int ret;
2169
2170         path = btrfs_alloc_path();
2171         if (!path)
2172                 return -ENOMEM;
2173
2174         /*
2175          * we may be replacing one extent in the tree with another.
2176          * The new extent is pinned in the extent map, and we don't want
2177          * to drop it from the cache until it is completely in the btree.
2178          *
2179          * So, tell btrfs_drop_extents to leave this extent in the cache.
2180          * the caller is expected to unpin it and allow it to be merged
2181          * with the others.
2182          */
2183         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2184                                    file_pos + num_bytes, NULL, 0,
2185                                    1, sizeof(*fi), &extent_inserted);
2186         if (ret)
2187                 goto out;
2188
2189         if (!extent_inserted) {
2190                 ins.objectid = btrfs_ino(BTRFS_I(inode));
2191                 ins.offset = file_pos;
2192                 ins.type = BTRFS_EXTENT_DATA_KEY;
2193
2194                 path->leave_spinning = 1;
2195                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2196                                               sizeof(*fi));
2197                 if (ret)
2198                         goto out;
2199         }
2200         leaf = path->nodes[0];
2201         fi = btrfs_item_ptr(leaf, path->slots[0],
2202                             struct btrfs_file_extent_item);
2203         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2204         btrfs_set_file_extent_type(leaf, fi, extent_type);
2205         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2206         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2207         btrfs_set_file_extent_offset(leaf, fi, 0);
2208         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2209         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2210         btrfs_set_file_extent_compression(leaf, fi, compression);
2211         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2212         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2213
2214         btrfs_mark_buffer_dirty(leaf);
2215         btrfs_release_path(path);
2216
2217         inode_add_bytes(inode, num_bytes);
2218
2219         ins.objectid = disk_bytenr;
2220         ins.offset = disk_num_bytes;
2221         ins.type = BTRFS_EXTENT_ITEM_KEY;
2222
2223         /*
2224          * Release the reserved range from inode dirty range map, as it is
2225          * already moved into delayed_ref_head
2226          */
2227         ret = btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2228         if (ret < 0)
2229                 goto out;
2230         qg_released = ret;
2231         ret = btrfs_alloc_reserved_file_extent(trans, root,
2232                                                btrfs_ino(BTRFS_I(inode)),
2233                                                file_pos, qg_released, &ins);
2234 out:
2235         btrfs_free_path(path);
2236
2237         return ret;
2238 }
2239
2240 /* snapshot-aware defrag */
2241 struct sa_defrag_extent_backref {
2242         struct rb_node node;
2243         struct old_sa_defrag_extent *old;
2244         u64 root_id;
2245         u64 inum;
2246         u64 file_pos;
2247         u64 extent_offset;
2248         u64 num_bytes;
2249         u64 generation;
2250 };
2251
2252 struct old_sa_defrag_extent {
2253         struct list_head list;
2254         struct new_sa_defrag_extent *new;
2255
2256         u64 extent_offset;
2257         u64 bytenr;
2258         u64 offset;
2259         u64 len;
2260         int count;
2261 };
2262
2263 struct new_sa_defrag_extent {
2264         struct rb_root root;
2265         struct list_head head;
2266         struct btrfs_path *path;
2267         struct inode *inode;
2268         u64 file_pos;
2269         u64 len;
2270         u64 bytenr;
2271         u64 disk_len;
2272         u8 compress_type;
2273 };
2274
2275 static int backref_comp(struct sa_defrag_extent_backref *b1,
2276                         struct sa_defrag_extent_backref *b2)
2277 {
2278         if (b1->root_id < b2->root_id)
2279                 return -1;
2280         else if (b1->root_id > b2->root_id)
2281                 return 1;
2282
2283         if (b1->inum < b2->inum)
2284                 return -1;
2285         else if (b1->inum > b2->inum)
2286                 return 1;
2287
2288         if (b1->file_pos < b2->file_pos)
2289                 return -1;
2290         else if (b1->file_pos > b2->file_pos)
2291                 return 1;
2292
2293         /*
2294          * [------------------------------] ===> (a range of space)
2295          *     |<--->|   |<---->| =============> (fs/file tree A)
2296          * |<---------------------------->| ===> (fs/file tree B)
2297          *
2298          * A range of space can refer to two file extents in one tree while
2299          * refer to only one file extent in another tree.
2300          *
2301          * So we may process a disk offset more than one time(two extents in A)
2302          * and locate at the same extent(one extent in B), then insert two same
2303          * backrefs(both refer to the extent in B).
2304          */
2305         return 0;
2306 }
2307
2308 static void backref_insert(struct rb_root *root,
2309                            struct sa_defrag_extent_backref *backref)
2310 {
2311         struct rb_node **p = &root->rb_node;
2312         struct rb_node *parent = NULL;
2313         struct sa_defrag_extent_backref *entry;
2314         int ret;
2315
2316         while (*p) {
2317                 parent = *p;
2318                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2319
2320                 ret = backref_comp(backref, entry);
2321                 if (ret < 0)
2322                         p = &(*p)->rb_left;
2323                 else
2324                         p = &(*p)->rb_right;
2325         }
2326
2327         rb_link_node(&backref->node, parent, p);
2328         rb_insert_color(&backref->node, root);
2329 }
2330
2331 /*
2332  * Note the backref might has changed, and in this case we just return 0.
2333  */
2334 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2335                                        void *ctx)
2336 {
2337         struct btrfs_file_extent_item *extent;
2338         struct old_sa_defrag_extent *old = ctx;
2339         struct new_sa_defrag_extent *new = old->new;
2340         struct btrfs_path *path = new->path;
2341         struct btrfs_key key;
2342         struct btrfs_root *root;
2343         struct sa_defrag_extent_backref *backref;
2344         struct extent_buffer *leaf;
2345         struct inode *inode = new->inode;
2346         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2347         int slot;
2348         int ret;
2349         u64 extent_offset;
2350         u64 num_bytes;
2351
2352         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2353             inum == btrfs_ino(BTRFS_I(inode)))
2354                 return 0;
2355
2356         key.objectid = root_id;
2357         key.type = BTRFS_ROOT_ITEM_KEY;
2358         key.offset = (u64)-1;
2359
2360         root = btrfs_read_fs_root_no_name(fs_info, &key);
2361         if (IS_ERR(root)) {
2362                 if (PTR_ERR(root) == -ENOENT)
2363                         return 0;
2364                 WARN_ON(1);
2365                 btrfs_debug(fs_info, "inum=%llu, offset=%llu, root_id=%llu",
2366                          inum, offset, root_id);
2367                 return PTR_ERR(root);
2368         }
2369
2370         key.objectid = inum;
2371         key.type = BTRFS_EXTENT_DATA_KEY;
2372         if (offset > (u64)-1 << 32)
2373                 key.offset = 0;
2374         else
2375                 key.offset = offset;
2376
2377         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2378         if (WARN_ON(ret < 0))
2379                 return ret;
2380         ret = 0;
2381
2382         while (1) {
2383                 cond_resched();
2384
2385                 leaf = path->nodes[0];
2386                 slot = path->slots[0];
2387
2388                 if (slot >= btrfs_header_nritems(leaf)) {
2389                         ret = btrfs_next_leaf(root, path);
2390                         if (ret < 0) {
2391                                 goto out;
2392                         } else if (ret > 0) {
2393                                 ret = 0;
2394                                 goto out;
2395                         }
2396                         continue;
2397                 }
2398
2399                 path->slots[0]++;
2400
2401                 btrfs_item_key_to_cpu(leaf, &key, slot);
2402
2403                 if (key.objectid > inum)
2404                         goto out;
2405
2406                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2407                         continue;
2408
2409                 extent = btrfs_item_ptr(leaf, slot,
2410                                         struct btrfs_file_extent_item);
2411
2412                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2413                         continue;
2414
2415                 /*
2416                  * 'offset' refers to the exact key.offset,
2417                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2418                  * (key.offset - extent_offset).
2419                  */
2420                 if (key.offset != offset)
2421                         continue;
2422
2423                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2424                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2425
2426                 if (extent_offset >= old->extent_offset + old->offset +
2427                     old->len || extent_offset + num_bytes <=
2428                     old->extent_offset + old->offset)
2429                         continue;
2430                 break;
2431         }
2432
2433         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2434         if (!backref) {
2435                 ret = -ENOENT;
2436                 goto out;
2437         }
2438
2439         backref->root_id = root_id;
2440         backref->inum = inum;
2441         backref->file_pos = offset;
2442         backref->num_bytes = num_bytes;
2443         backref->extent_offset = extent_offset;
2444         backref->generation = btrfs_file_extent_generation(leaf, extent);
2445         backref->old = old;
2446         backref_insert(&new->root, backref);
2447         old->count++;
2448 out:
2449         btrfs_release_path(path);
2450         WARN_ON(ret);
2451         return ret;
2452 }
2453
2454 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2455                                    struct new_sa_defrag_extent *new)
2456 {
2457         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2458         struct old_sa_defrag_extent *old, *tmp;
2459         int ret;
2460
2461         new->path = path;
2462
2463         list_for_each_entry_safe(old, tmp, &new->head, list) {
2464                 ret = iterate_inodes_from_logical(old->bytenr +
2465                                                   old->extent_offset, fs_info,
2466                                                   path, record_one_backref,
2467                                                   old, false);
2468                 if (ret < 0 && ret != -ENOENT)
2469                         return false;
2470
2471                 /* no backref to be processed for this extent */
2472                 if (!old->count) {
2473                         list_del(&old->list);
2474                         kfree(old);
2475                 }
2476         }
2477
2478         if (list_empty(&new->head))
2479                 return false;
2480
2481         return true;
2482 }
2483
2484 static int relink_is_mergable(struct extent_buffer *leaf,
2485                               struct btrfs_file_extent_item *fi,
2486                               struct new_sa_defrag_extent *new)
2487 {
2488         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2489                 return 0;
2490
2491         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2492                 return 0;
2493
2494         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2495                 return 0;
2496
2497         if (btrfs_file_extent_encryption(leaf, fi) ||
2498             btrfs_file_extent_other_encoding(leaf, fi))
2499                 return 0;
2500
2501         return 1;
2502 }
2503
2504 /*
2505  * Note the backref might has changed, and in this case we just return 0.
2506  */
2507 static noinline int relink_extent_backref(struct btrfs_path *path,
2508                                  struct sa_defrag_extent_backref *prev,
2509                                  struct sa_defrag_extent_backref *backref)
2510 {
2511         struct btrfs_file_extent_item *extent;
2512         struct btrfs_file_extent_item *item;
2513         struct btrfs_ordered_extent *ordered;
2514         struct btrfs_trans_handle *trans;
2515         struct btrfs_root *root;
2516         struct btrfs_key key;
2517         struct extent_buffer *leaf;
2518         struct old_sa_defrag_extent *old = backref->old;
2519         struct new_sa_defrag_extent *new = old->new;
2520         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2521         struct inode *inode;
2522         struct extent_state *cached = NULL;
2523         int ret = 0;
2524         u64 start;
2525         u64 len;
2526         u64 lock_start;
2527         u64 lock_end;
2528         bool merge = false;
2529         int index;
2530
2531         if (prev && prev->root_id == backref->root_id &&
2532             prev->inum == backref->inum &&
2533             prev->file_pos + prev->num_bytes == backref->file_pos)
2534                 merge = true;
2535
2536         /* step 1: get root */
2537         key.objectid = backref->root_id;
2538         key.type = BTRFS_ROOT_ITEM_KEY;
2539         key.offset = (u64)-1;
2540
2541         index = srcu_read_lock(&fs_info->subvol_srcu);
2542
2543         root = btrfs_read_fs_root_no_name(fs_info, &key);
2544         if (IS_ERR(root)) {
2545                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2546                 if (PTR_ERR(root) == -ENOENT)
2547                         return 0;
2548                 return PTR_ERR(root);
2549         }
2550
2551         if (btrfs_root_readonly(root)) {
2552                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2553                 return 0;
2554         }
2555
2556         /* step 2: get inode */
2557         key.objectid = backref->inum;
2558         key.type = BTRFS_INODE_ITEM_KEY;
2559         key.offset = 0;
2560
2561         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2562         if (IS_ERR(inode)) {
2563                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2564                 return 0;
2565         }
2566
2567         srcu_read_unlock(&fs_info->subvol_srcu, index);
2568
2569         /* step 3: relink backref */
2570         lock_start = backref->file_pos;
2571         lock_end = backref->file_pos + backref->num_bytes - 1;
2572         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2573                          &cached);
2574
2575         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2576         if (ordered) {
2577                 btrfs_put_ordered_extent(ordered);
2578                 goto out_unlock;
2579         }
2580
2581         trans = btrfs_join_transaction(root);
2582         if (IS_ERR(trans)) {
2583                 ret = PTR_ERR(trans);
2584                 goto out_unlock;
2585         }
2586
2587         key.objectid = backref->inum;
2588         key.type = BTRFS_EXTENT_DATA_KEY;
2589         key.offset = backref->file_pos;
2590
2591         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2592         if (ret < 0) {
2593                 goto out_free_path;
2594         } else if (ret > 0) {
2595                 ret = 0;
2596                 goto out_free_path;
2597         }
2598
2599         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2600                                 struct btrfs_file_extent_item);
2601
2602         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2603             backref->generation)
2604                 goto out_free_path;
2605
2606         btrfs_release_path(path);
2607
2608         start = backref->file_pos;
2609         if (backref->extent_offset < old->extent_offset + old->offset)
2610                 start += old->extent_offset + old->offset -
2611                          backref->extent_offset;
2612
2613         len = min(backref->extent_offset + backref->num_bytes,
2614                   old->extent_offset + old->offset + old->len);
2615         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2616
2617         ret = btrfs_drop_extents(trans, root, inode, start,
2618                                  start + len, 1);
2619         if (ret)
2620                 goto out_free_path;
2621 again:
2622         key.objectid = btrfs_ino(BTRFS_I(inode));
2623         key.type = BTRFS_EXTENT_DATA_KEY;
2624         key.offset = start;
2625
2626         path->leave_spinning = 1;
2627         if (merge) {
2628                 struct btrfs_file_extent_item *fi;
2629                 u64 extent_len;
2630                 struct btrfs_key found_key;
2631
2632                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2633                 if (ret < 0)
2634                         goto out_free_path;
2635
2636                 path->slots[0]--;
2637                 leaf = path->nodes[0];
2638                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2639
2640                 fi = btrfs_item_ptr(leaf, path->slots[0],
2641                                     struct btrfs_file_extent_item);
2642                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2643
2644                 if (extent_len + found_key.offset == start &&
2645                     relink_is_mergable(leaf, fi, new)) {
2646                         btrfs_set_file_extent_num_bytes(leaf, fi,
2647                                                         extent_len + len);
2648                         btrfs_mark_buffer_dirty(leaf);
2649                         inode_add_bytes(inode, len);
2650
2651                         ret = 1;
2652                         goto out_free_path;
2653                 } else {
2654                         merge = false;
2655                         btrfs_release_path(path);
2656                         goto again;
2657                 }
2658         }
2659
2660         ret = btrfs_insert_empty_item(trans, root, path, &key,
2661                                         sizeof(*extent));
2662         if (ret) {
2663                 btrfs_abort_transaction(trans, ret);
2664                 goto out_free_path;
2665         }
2666
2667         leaf = path->nodes[0];
2668         item = btrfs_item_ptr(leaf, path->slots[0],
2669                                 struct btrfs_file_extent_item);
2670         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2671         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2672         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2673         btrfs_set_file_extent_num_bytes(leaf, item, len);
2674         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2675         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2676         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2677         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2678         btrfs_set_file_extent_encryption(leaf, item, 0);
2679         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2680
2681         btrfs_mark_buffer_dirty(leaf);
2682         inode_add_bytes(inode, len);
2683         btrfs_release_path(path);
2684
2685         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2686                         new->disk_len, 0,
2687                         backref->root_id, backref->inum,
2688                         new->file_pos); /* start - extent_offset */
2689         if (ret) {
2690                 btrfs_abort_transaction(trans, ret);
2691                 goto out_free_path;
2692         }
2693
2694         ret = 1;
2695 out_free_path:
2696         btrfs_release_path(path);
2697         path->leave_spinning = 0;
2698         btrfs_end_transaction(trans);
2699 out_unlock:
2700         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2701                              &cached, GFP_NOFS);
2702         iput(inode);
2703         return ret;
2704 }
2705
2706 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2707 {
2708         struct old_sa_defrag_extent *old, *tmp;
2709
2710         if (!new)
2711                 return;
2712
2713         list_for_each_entry_safe(old, tmp, &new->head, list) {
2714                 kfree(old);
2715         }
2716         kfree(new);
2717 }
2718
2719 static void relink_file_extents(struct new_sa_defrag_extent *new)
2720 {
2721         struct btrfs_fs_info *fs_info = btrfs_sb(new->inode->i_sb);
2722         struct btrfs_path *path;
2723         struct sa_defrag_extent_backref *backref;
2724         struct sa_defrag_extent_backref *prev = NULL;
2725         struct inode *inode;
2726         struct btrfs_root *root;
2727         struct rb_node *node;
2728         int ret;
2729
2730         inode = new->inode;
2731         root = BTRFS_I(inode)->root;
2732
2733         path = btrfs_alloc_path();
2734         if (!path)
2735                 return;
2736
2737         if (!record_extent_backrefs(path, new)) {
2738                 btrfs_free_path(path);
2739                 goto out;
2740         }
2741         btrfs_release_path(path);
2742
2743         while (1) {
2744                 node = rb_first(&new->root);
2745                 if (!node)
2746                         break;
2747                 rb_erase(node, &new->root);
2748
2749                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2750
2751                 ret = relink_extent_backref(path, prev, backref);
2752                 WARN_ON(ret < 0);
2753
2754                 kfree(prev);
2755
2756                 if (ret == 1)
2757                         prev = backref;
2758                 else
2759                         prev = NULL;
2760                 cond_resched();
2761         }
2762         kfree(prev);
2763
2764         btrfs_free_path(path);
2765 out:
2766         free_sa_defrag_extent(new);
2767
2768         atomic_dec(&fs_info->defrag_running);
2769         wake_up(&fs_info->transaction_wait);
2770 }
2771
2772 static struct new_sa_defrag_extent *
2773 record_old_file_extents(struct inode *inode,
2774                         struct btrfs_ordered_extent *ordered)
2775 {
2776         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2777         struct btrfs_root *root = BTRFS_I(inode)->root;
2778         struct btrfs_path *path;
2779         struct btrfs_key key;
2780         struct old_sa_defrag_extent *old;
2781         struct new_sa_defrag_extent *new;
2782         int ret;
2783
2784         new = kmalloc(sizeof(*new), GFP_NOFS);
2785         if (!new)
2786                 return NULL;
2787
2788         new->inode = inode;
2789         new->file_pos = ordered->file_offset;
2790         new->len = ordered->len;
2791         new->bytenr = ordered->start;
2792         new->disk_len = ordered->disk_len;
2793         new->compress_type = ordered->compress_type;
2794         new->root = RB_ROOT;
2795         INIT_LIST_HEAD(&new->head);
2796
2797         path = btrfs_alloc_path();
2798         if (!path)
2799                 goto out_kfree;
2800
2801         key.objectid = btrfs_ino(BTRFS_I(inode));
2802         key.type = BTRFS_EXTENT_DATA_KEY;
2803         key.offset = new->file_pos;
2804
2805         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2806         if (ret < 0)
2807                 goto out_free_path;
2808         if (ret > 0 && path->slots[0] > 0)
2809                 path->slots[0]--;
2810
2811         /* find out all the old extents for the file range */
2812         while (1) {
2813                 struct btrfs_file_extent_item *extent;
2814                 struct extent_buffer *l;
2815                 int slot;
2816                 u64 num_bytes;
2817                 u64 offset;
2818                 u64 end;
2819                 u64 disk_bytenr;
2820                 u64 extent_offset;
2821
2822                 l = path->nodes[0];
2823                 slot = path->slots[0];
2824
2825                 if (slot >= btrfs_header_nritems(l)) {
2826                         ret = btrfs_next_leaf(root, path);
2827                         if (ret < 0)
2828                                 goto out_free_path;
2829                         else if (ret > 0)
2830                                 break;
2831                         continue;
2832                 }
2833
2834                 btrfs_item_key_to_cpu(l, &key, slot);
2835
2836                 if (key.objectid != btrfs_ino(BTRFS_I(inode)))
2837                         break;
2838                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2839                         break;
2840                 if (key.offset >= new->file_pos + new->len)
2841                         break;
2842
2843                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2844
2845                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2846                 if (key.offset + num_bytes < new->file_pos)
2847                         goto next;
2848
2849                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2850                 if (!disk_bytenr)
2851                         goto next;
2852
2853                 extent_offset = btrfs_file_extent_offset(l, extent);
2854
2855                 old = kmalloc(sizeof(*old), GFP_NOFS);
2856                 if (!old)
2857                         goto out_free_path;
2858
2859                 offset = max(new->file_pos, key.offset);
2860                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2861
2862                 old->bytenr = disk_bytenr;
2863                 old->extent_offset = extent_offset;
2864                 old->offset = offset - key.offset;
2865                 old->len = end - offset;
2866                 old->new = new;
2867                 old->count = 0;
2868                 list_add_tail(&old->list, &new->head);
2869 next:
2870                 path->slots[0]++;
2871                 cond_resched();
2872         }
2873
2874         btrfs_free_path(path);
2875         atomic_inc(&fs_info->defrag_running);
2876
2877         return new;
2878
2879 out_free_path:
2880         btrfs_free_path(path);
2881 out_kfree:
2882         free_sa_defrag_extent(new);
2883         return NULL;
2884 }
2885
2886 static void btrfs_release_delalloc_bytes(struct btrfs_fs_info *fs_info,
2887                                          u64 start, u64 len)
2888 {
2889         struct btrfs_block_group_cache *cache;
2890
2891         cache = btrfs_lookup_block_group(fs_info, start);
2892         ASSERT(cache);
2893
2894         spin_lock(&cache->lock);
2895         cache->delalloc_bytes -= len;
2896         spin_unlock(&cache->lock);
2897
2898         btrfs_put_block_group(cache);
2899 }
2900
2901 /* as ordered data IO finishes, this gets called so we can finish
2902  * an ordered extent if the range of bytes in the file it covers are
2903  * fully written.
2904  */
2905 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2906 {
2907         struct inode *inode = ordered_extent->inode;
2908         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2909         struct btrfs_root *root = BTRFS_I(inode)->root;
2910         struct btrfs_trans_handle *trans = NULL;
2911         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2912         struct extent_state *cached_state = NULL;
2913         struct new_sa_defrag_extent *new = NULL;
2914         int compress_type = 0;
2915         int ret = 0;
2916         u64 logical_len = ordered_extent->len;
2917         bool nolock;
2918         bool truncated = false;
2919         bool range_locked = false;
2920         bool clear_new_delalloc_bytes = false;
2921
2922         if (!test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2923             !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags) &&
2924             !test_bit(BTRFS_ORDERED_DIRECT, &ordered_extent->flags))
2925                 clear_new_delalloc_bytes = true;
2926
2927         nolock = btrfs_is_free_space_inode(BTRFS_I(inode));
2928
2929         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2930                 ret = -EIO;
2931                 goto out;
2932         }
2933
2934         btrfs_free_io_failure_record(BTRFS_I(inode),
2935                         ordered_extent->file_offset,
2936                         ordered_extent->file_offset +
2937                         ordered_extent->len - 1);
2938
2939         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2940                 truncated = true;
2941                 logical_len = ordered_extent->truncated_len;
2942                 /* Truncated the entire extent, don't bother adding */
2943                 if (!logical_len)
2944                         goto out;
2945         }
2946
2947         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2948                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2949
2950                 /*
2951                  * For mwrite(mmap + memset to write) case, we still reserve
2952                  * space for NOCOW range.
2953                  * As NOCOW won't cause a new delayed ref, just free the space
2954                  */
2955                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
2956                                        ordered_extent->len);
2957                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2958                 if (nolock)
2959                         trans = btrfs_join_transaction_nolock(root);
2960                 else
2961                         trans = btrfs_join_transaction(root);
2962                 if (IS_ERR(trans)) {
2963                         ret = PTR_ERR(trans);
2964                         trans = NULL;
2965                         goto out;
2966                 }
2967                 trans->block_rsv = &BTRFS_I(inode)->block_rsv;
2968                 ret = btrfs_update_inode_fallback(trans, root, inode);
2969                 if (ret) /* -ENOMEM or corruption */
2970                         btrfs_abort_transaction(trans, ret);
2971                 goto out;
2972         }
2973
2974         range_locked = true;
2975         lock_extent_bits(io_tree, ordered_extent->file_offset,
2976                          ordered_extent->file_offset + ordered_extent->len - 1,
2977                          &cached_state);
2978
2979         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2980                         ordered_extent->file_offset + ordered_extent->len - 1,
2981                         EXTENT_DEFRAG, 0, cached_state);
2982         if (ret) {
2983                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2984                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2985                         /* the inode is shared */
2986                         new = record_old_file_extents(inode, ordered_extent);
2987
2988                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2989                         ordered_extent->file_offset + ordered_extent->len - 1,
2990                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2991         }
2992
2993         if (nolock)
2994                 trans = btrfs_join_transaction_nolock(root);
2995         else
2996                 trans = btrfs_join_transaction(root);
2997         if (IS_ERR(trans)) {
2998                 ret = PTR_ERR(trans);
2999                 trans = NULL;
3000                 goto out;
3001         }
3002
3003         trans->block_rsv = &BTRFS_I(inode)->block_rsv;
3004
3005         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
3006                 compress_type = ordered_extent->compress_type;
3007         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
3008                 BUG_ON(compress_type);
3009                 btrfs_qgroup_free_data(inode, NULL, ordered_extent->file_offset,
3010                                        ordered_extent->len);
3011                 ret = btrfs_mark_extent_written(trans, BTRFS_I(inode),
3012                                                 ordered_extent->file_offset,
3013                                                 ordered_extent->file_offset +
3014                                                 logical_len);
3015         } else {
3016                 BUG_ON(root == fs_info->tree_root);
3017                 ret = insert_reserved_file_extent(trans, inode,
3018                                                 ordered_extent->file_offset,
3019                                                 ordered_extent->start,
3020                                                 ordered_extent->disk_len,
3021                                                 logical_len, logical_len,
3022                                                 compress_type, 0, 0,
3023                                                 BTRFS_FILE_EXTENT_REG);
3024                 if (!ret)
3025                         btrfs_release_delalloc_bytes(fs_info,
3026                                                      ordered_extent->start,
3027                                                      ordered_extent->disk_len);
3028         }
3029         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
3030                            ordered_extent->file_offset, ordered_extent->len,
3031                            trans->transid);
3032         if (ret < 0) {
3033                 btrfs_abort_transaction(trans, ret);
3034                 goto out;
3035         }
3036
3037         add_pending_csums(trans, inode, &ordered_extent->list);
3038
3039         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
3040         ret = btrfs_update_inode_fallback(trans, root, inode);
3041         if (ret) { /* -ENOMEM or corruption */
3042                 btrfs_abort_transaction(trans, ret);
3043                 goto out;
3044         }
3045         ret = 0;
3046 out:
3047         if (range_locked || clear_new_delalloc_bytes) {
3048                 unsigned int clear_bits = 0;
3049
3050                 if (range_locked)
3051                         clear_bits |= EXTENT_LOCKED;
3052                 if (clear_new_delalloc_bytes)
3053                         clear_bits |= EXTENT_DELALLOC_NEW;
3054                 clear_extent_bit(&BTRFS_I(inode)->io_tree,
3055                                  ordered_extent->file_offset,
3056                                  ordered_extent->file_offset +
3057                                  ordered_extent->len - 1,
3058                                  clear_bits,
3059                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
3060                                  0, &cached_state, GFP_NOFS);
3061         }
3062
3063         if (trans)
3064                 btrfs_end_transaction(trans);
3065
3066         if (ret || truncated) {
3067                 u64 start, end;
3068
3069                 if (truncated)
3070                         start = ordered_extent->file_offset + logical_len;
3071                 else
3072                         start = ordered_extent->file_offset;
3073                 end = ordered_extent->file_offset + ordered_extent->len - 1;
3074                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
3075
3076                 /* Drop the cache for the part of the extent we didn't write. */
3077                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end, 0);
3078
3079                 /*
3080                  * If the ordered extent had an IOERR or something else went
3081                  * wrong we need to return the space for this ordered extent
3082                  * back to the allocator.  We only free the extent in the
3083                  * truncated case if we didn't write out the extent at all.
3084                  */
3085                 if ((ret || !logical_len) &&
3086                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
3087                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
3088                         btrfs_free_reserved_extent(fs_info,
3089                                                    ordered_extent->start,
3090                                                    ordered_extent->disk_len, 1);
3091         }
3092
3093
3094         /*
3095          * This needs to be done to make sure anybody waiting knows we are done
3096          * updating everything for this ordered extent.
3097          */
3098         btrfs_remove_ordered_extent(inode, ordered_extent);
3099
3100         /* for snapshot-aware defrag */
3101         if (new) {
3102                 if (ret) {
3103                         free_sa_defrag_extent(new);
3104                         atomic_dec(&fs_info->defrag_running);
3105                 } else {
3106                         relink_file_extents(new);
3107                 }
3108         }
3109
3110         /* once for us */
3111         btrfs_put_ordered_extent(ordered_extent);
3112         /* once for the tree */
3113         btrfs_put_ordered_extent(ordered_extent);
3114
3115         return ret;
3116 }
3117
3118 static void finish_ordered_fn(struct btrfs_work *work)
3119 {
3120         struct btrfs_ordered_extent *ordered_extent;
3121         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3122         btrfs_finish_ordered_io(ordered_extent);
3123 }
3124
3125 static void btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3126                                 struct extent_state *state, int uptodate)
3127 {
3128         struct inode *inode = page->mapping->host;
3129         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3130         struct btrfs_ordered_extent *ordered_extent = NULL;
3131         struct btrfs_workqueue *wq;
3132         btrfs_work_func_t func;
3133
3134         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3135
3136         ClearPagePrivate2(page);
3137         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3138                                             end - start + 1, uptodate))
3139                 return;
3140
3141         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
3142                 wq = fs_info->endio_freespace_worker;
3143                 func = btrfs_freespace_write_helper;
3144         } else {
3145                 wq = fs_info->endio_write_workers;
3146                 func = btrfs_endio_write_helper;
3147         }
3148
3149         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3150                         NULL);
3151         btrfs_queue_work(wq, &ordered_extent->work);
3152 }
3153
3154 static int __readpage_endio_check(struct inode *inode,
3155                                   struct btrfs_io_bio *io_bio,
3156                                   int icsum, struct page *page,
3157                                   int pgoff, u64 start, size_t len)
3158 {
3159         char *kaddr;
3160         u32 csum_expected;
3161         u32 csum = ~(u32)0;
3162
3163         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3164
3165         kaddr = kmap_atomic(page);
3166         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3167         btrfs_csum_final(csum, (u8 *)&csum);
3168         if (csum != csum_expected)
3169                 goto zeroit;
3170
3171         kunmap_atomic(kaddr);
3172         return 0;
3173 zeroit:
3174         btrfs_print_data_csum_error(BTRFS_I(inode), start, csum, csum_expected,
3175                                     io_bio->mirror_num);
3176         memset(kaddr + pgoff, 1, len);
3177         flush_dcache_page(page);
3178         kunmap_atomic(kaddr);
3179         return -EIO;
3180 }
3181
3182 /*
3183  * when reads are done, we need to check csums to verify the data is correct
3184  * if there's a match, we allow the bio to finish.  If not, the code in
3185  * extent_io.c will try to find good copies for us.
3186  */
3187 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3188                                       u64 phy_offset, struct page *page,
3189                                       u64 start, u64 end, int mirror)
3190 {
3191         size_t offset = start - page_offset(page);
3192         struct inode *inode = page->mapping->host;
3193         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3194         struct btrfs_root *root = BTRFS_I(inode)->root;
3195
3196         if (PageChecked(page)) {
3197                 ClearPageChecked(page);
3198                 return 0;
3199         }
3200
3201         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3202                 return 0;
3203
3204         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3205             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3206                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM);
3207                 return 0;
3208         }
3209
3210         phy_offset >>= inode->i_sb->s_blocksize_bits;
3211         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3212                                       start, (size_t)(end - start + 1));
3213 }
3214
3215 void btrfs_add_delayed_iput(struct inode *inode)
3216 {
3217         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3218         struct btrfs_inode *binode = BTRFS_I(inode);
3219
3220         if (atomic_add_unless(&inode->i_count, -1, 1))
3221                 return;
3222
3223         spin_lock(&fs_info->delayed_iput_lock);
3224         if (binode->delayed_iput_count == 0) {
3225                 ASSERT(list_empty(&binode->delayed_iput));
3226                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3227         } else {
3228                 binode->delayed_iput_count++;
3229         }
3230         spin_unlock(&fs_info->delayed_iput_lock);
3231 }
3232
3233 void btrfs_run_delayed_iputs(struct btrfs_fs_info *fs_info)
3234 {
3235
3236         spin_lock(&fs_info->delayed_iput_lock);
3237         while (!list_empty(&fs_info->delayed_iputs)) {
3238                 struct btrfs_inode *inode;
3239
3240                 inode = list_first_entry(&fs_info->delayed_iputs,
3241                                 struct btrfs_inode, delayed_iput);
3242                 if (inode->delayed_iput_count) {
3243                         inode->delayed_iput_count--;
3244                         list_move_tail(&inode->delayed_iput,
3245                                         &fs_info->delayed_iputs);
3246                 } else {
3247                         list_del_init(&inode->delayed_iput);
3248                 }
3249                 spin_unlock(&fs_info->delayed_iput_lock);
3250                 iput(&inode->vfs_inode);
3251                 spin_lock(&fs_info->delayed_iput_lock);
3252         }
3253         spin_unlock(&fs_info->delayed_iput_lock);
3254 }
3255
3256 /*
3257  * This is called in transaction commit time. If there are no orphan
3258  * files in the subvolume, it removes orphan item and frees block_rsv
3259  * structure.
3260  */
3261 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3262                               struct btrfs_root *root)
3263 {
3264         struct btrfs_fs_info *fs_info = root->fs_info;
3265         struct btrfs_block_rsv *block_rsv;
3266         int ret;
3267
3268         if (atomic_read(&root->orphan_inodes) ||
3269             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3270                 return;
3271
3272         spin_lock(&root->orphan_lock);
3273         if (atomic_read(&root->orphan_inodes)) {
3274                 spin_unlock(&root->orphan_lock);
3275                 return;
3276         }
3277
3278         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3279                 spin_unlock(&root->orphan_lock);
3280                 return;
3281         }
3282
3283         block_rsv = root->orphan_block_rsv;
3284         root->orphan_block_rsv = NULL;
3285         spin_unlock(&root->orphan_lock);
3286
3287         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3288             btrfs_root_refs(&root->root_item) > 0) {
3289                 ret = btrfs_del_orphan_item(trans, fs_info->tree_root,
3290                                             root->root_key.objectid);
3291                 if (ret)
3292                         btrfs_abort_transaction(trans, ret);
3293                 else
3294                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3295                                   &root->state);
3296         }
3297
3298         if (block_rsv) {
3299                 WARN_ON(block_rsv->size > 0);
3300                 btrfs_free_block_rsv(fs_info, block_rsv);
3301         }
3302 }
3303
3304 /*
3305  * This creates an orphan entry for the given inode in case something goes
3306  * wrong in the middle of an unlink/truncate.
3307  *
3308  * NOTE: caller of this function should reserve 5 units of metadata for
3309  *       this function.
3310  */
3311 int btrfs_orphan_add(struct btrfs_trans_handle *trans,
3312                 struct btrfs_inode *inode)
3313 {
3314         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
3315         struct btrfs_root *root = inode->root;
3316         struct btrfs_block_rsv *block_rsv = NULL;
3317         int reserve = 0;
3318         int insert = 0;
3319         int ret;
3320
3321         if (!root->orphan_block_rsv) {
3322                 block_rsv = btrfs_alloc_block_rsv(fs_info,
3323                                                   BTRFS_BLOCK_RSV_TEMP);
3324                 if (!block_rsv)
3325                         return -ENOMEM;
3326         }
3327
3328         spin_lock(&root->orphan_lock);
3329         if (!root->orphan_block_rsv) {
3330                 root->orphan_block_rsv = block_rsv;
3331         } else if (block_rsv) {
3332                 btrfs_free_block_rsv(fs_info, block_rsv);
3333                 block_rsv = NULL;
3334         }
3335
3336         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3337                               &inode->runtime_flags)) {
3338 #if 0
3339                 /*
3340                  * For proper ENOSPC handling, we should do orphan
3341                  * cleanup when mounting. But this introduces backward
3342                  * compatibility issue.
3343                  */
3344                 if (!xchg(&root->orphan_item_inserted, 1))
3345                         insert = 2;
3346                 else
3347                         insert = 1;
3348 #endif
3349                 insert = 1;
3350                 atomic_inc(&root->orphan_inodes);
3351         }
3352
3353         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3354                               &inode->runtime_flags))
3355                 reserve = 1;
3356         spin_unlock(&root->orphan_lock);
3357
3358         /* grab metadata reservation from transaction handle */
3359         if (reserve) {
3360                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3361                 ASSERT(!ret);
3362                 if (ret) {
3363                         atomic_dec(&root->orphan_inodes);
3364                         clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3365                                   &inode->runtime_flags);
3366                         if (insert)
3367                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3368                                           &inode->runtime_flags);
3369                         return ret;
3370                 }
3371         }
3372
3373         /* insert an orphan item to track this unlinked/truncated file */
3374         if (insert >= 1) {
3375                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3376                 if (ret) {
3377                         atomic_dec(&root->orphan_inodes);
3378                         if (reserve) {
3379                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3380                                           &inode->runtime_flags);
3381                                 btrfs_orphan_release_metadata(inode);
3382                         }
3383                         if (ret != -EEXIST) {
3384                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3385                                           &inode->runtime_flags);
3386                                 btrfs_abort_transaction(trans, ret);
3387                                 return ret;
3388                         }
3389                 }
3390                 ret = 0;
3391         }
3392
3393         /* insert an orphan item to track subvolume contains orphan files */
3394         if (insert >= 2) {
3395                 ret = btrfs_insert_orphan_item(trans, fs_info->tree_root,
3396                                                root->root_key.objectid);
3397                 if (ret && ret != -EEXIST) {
3398                         btrfs_abort_transaction(trans, ret);
3399                         return ret;
3400                 }
3401         }
3402         return 0;
3403 }
3404
3405 /*
3406  * We have done the truncate/delete so we can go ahead and remove the orphan
3407  * item for this particular inode.
3408  */
3409 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3410                             struct btrfs_inode *inode)
3411 {
3412         struct btrfs_root *root = inode->root;
3413         int delete_item = 0;
3414         int release_rsv = 0;
3415         int ret = 0;
3416
3417         spin_lock(&root->orphan_lock);
3418         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3419                                &inode->runtime_flags))
3420                 delete_item = 1;
3421
3422         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3423                                &inode->runtime_flags))
3424                 release_rsv = 1;
3425         spin_unlock(&root->orphan_lock);
3426
3427         if (delete_item) {
3428                 atomic_dec(&root->orphan_inodes);
3429                 if (trans)
3430                         ret = btrfs_del_orphan_item(trans, root,
3431                                                     btrfs_ino(inode));
3432         }
3433
3434         if (release_rsv)
3435                 btrfs_orphan_release_metadata(inode);
3436
3437         return ret;
3438 }
3439
3440 /*
3441  * this cleans up any orphans that may be left on the list from the last use
3442  * of this root.
3443  */
3444 int btrfs_orphan_cleanup(struct btrfs_root *root)
3445 {
3446         struct btrfs_fs_info *fs_info = root->fs_info;
3447         struct btrfs_path *path;
3448         struct extent_buffer *leaf;
3449         struct btrfs_key key, found_key;
3450         struct btrfs_trans_handle *trans;
3451         struct inode *inode;
3452         u64 last_objectid = 0;
3453         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3454
3455         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3456                 return 0;
3457
3458         path = btrfs_alloc_path();
3459         if (!path) {
3460                 ret = -ENOMEM;
3461                 goto out;
3462         }
3463         path->reada = READA_BACK;
3464
3465         key.objectid = BTRFS_ORPHAN_OBJECTID;
3466         key.type = BTRFS_ORPHAN_ITEM_KEY;
3467         key.offset = (u64)-1;
3468
3469         while (1) {
3470                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3471                 if (ret < 0)
3472                         goto out;
3473
3474                 /*
3475                  * if ret == 0 means we found what we were searching for, which
3476                  * is weird, but possible, so only screw with path if we didn't
3477                  * find the key and see if we have stuff that matches
3478                  */
3479                 if (ret > 0) {
3480                         ret = 0;
3481                         if (path->slots[0] == 0)
3482                                 break;
3483                         path->slots[0]--;
3484                 }
3485
3486                 /* pull out the item */
3487                 leaf = path->nodes[0];
3488                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3489
3490                 /* make sure the item matches what we want */
3491                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3492                         break;
3493                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3494                         break;
3495
3496                 /* release the path since we're done with it */
3497                 btrfs_release_path(path);
3498
3499                 /*
3500                  * this is where we are basically btrfs_lookup, without the
3501                  * crossing root thing.  we store the inode number in the
3502                  * offset of the orphan item.
3503                  */
3504
3505                 if (found_key.offset == last_objectid) {
3506                         btrfs_err(fs_info,
3507                                   "Error removing orphan entry, stopping orphan cleanup");
3508                         ret = -EINVAL;
3509                         goto out;
3510                 }
3511
3512                 last_objectid = found_key.offset;
3513
3514                 found_key.objectid = found_key.offset;
3515                 found_key.type = BTRFS_INODE_ITEM_KEY;
3516                 found_key.offset = 0;
3517                 inode = btrfs_iget(fs_info->sb, &found_key, root, NULL);
3518                 ret = PTR_ERR_OR_ZERO(inode);
3519                 if (ret && ret != -ENOENT)
3520                         goto out;
3521
3522                 if (ret == -ENOENT && root == fs_info->tree_root) {
3523                         struct btrfs_root *dead_root;
3524                         struct btrfs_fs_info *fs_info = root->fs_info;
3525                         int is_dead_root = 0;
3526
3527                         /*
3528                          * this is an orphan in the tree root. Currently these
3529                          * could come from 2 sources:
3530                          *  a) a snapshot deletion in progress
3531                          *  b) a free space cache inode
3532                          * We need to distinguish those two, as the snapshot
3533                          * orphan must not get deleted.
3534                          * find_dead_roots already ran before us, so if this
3535                          * is a snapshot deletion, we should find the root
3536                          * in the dead_roots list
3537                          */
3538                         spin_lock(&fs_info->trans_lock);
3539                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3540                                             root_list) {
3541                                 if (dead_root->root_key.objectid ==
3542                                     found_key.objectid) {
3543                                         is_dead_root = 1;
3544                                         break;
3545                                 }
3546                         }
3547                         spin_unlock(&fs_info->trans_lock);
3548                         if (is_dead_root) {
3549                                 /* prevent this orphan from being found again */
3550                                 key.offset = found_key.objectid - 1;
3551                                 continue;
3552                         }
3553                 }
3554                 /*
3555                  * Inode is already gone but the orphan item is still there,
3556                  * kill the orphan item.
3557                  */
3558                 if (ret == -ENOENT) {
3559                         trans = btrfs_start_transaction(root, 1);
3560                         if (IS_ERR(trans)) {
3561                                 ret = PTR_ERR(trans);
3562                                 goto out;
3563                         }
3564                         btrfs_debug(fs_info, "auto deleting %Lu",
3565                                     found_key.objectid);
3566                         ret = btrfs_del_orphan_item(trans, root,
3567                                                     found_key.objectid);
3568                         btrfs_end_transaction(trans);
3569                         if (ret)
3570                                 goto out;
3571                         continue;
3572                 }
3573
3574                 /*
3575                  * add this inode to the orphan list so btrfs_orphan_del does
3576                  * the proper thing when we hit it
3577                  */
3578                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3579                         &BTRFS_I(inode)->runtime_flags);
3580                 atomic_inc(&root->orphan_inodes);
3581
3582                 /* if we have links, this was a truncate, lets do that */
3583                 if (inode->i_nlink) {
3584                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3585                                 iput(inode);
3586                                 continue;
3587                         }
3588                         nr_truncate++;
3589
3590                         /* 1 for the orphan item deletion. */
3591                         trans = btrfs_start_transaction(root, 1);
3592                         if (IS_ERR(trans)) {
3593                                 iput(inode);
3594                                 ret = PTR_ERR(trans);
3595                                 goto out;
3596                         }
3597                         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
3598                         btrfs_end_transaction(trans);
3599                         if (ret) {
3600                                 iput(inode);
3601                                 goto out;
3602                         }
3603
3604                         ret = btrfs_truncate(inode);
3605                         if (ret)
3606                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
3607                 } else {
3608                         nr_unlink++;
3609                 }
3610
3611                 /* this will do delete_inode and everything for us */
3612                 iput(inode);
3613                 if (ret)
3614                         goto out;
3615         }
3616         /* release the path since we're done with it */
3617         btrfs_release_path(path);
3618
3619         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3620
3621         if (root->orphan_block_rsv)
3622                 btrfs_block_rsv_release(fs_info, root->orphan_block_rsv,
3623                                         (u64)-1);
3624
3625         if (root->orphan_block_rsv ||
3626             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3627                 trans = btrfs_join_transaction(root);
3628                 if (!IS_ERR(trans))
3629                         btrfs_end_transaction(trans);
3630         }
3631
3632         if (nr_unlink)
3633                 btrfs_debug(fs_info, "unlinked %d orphans", nr_unlink);
3634         if (nr_truncate)
3635                 btrfs_debug(fs_info, "truncated %d orphans", nr_truncate);
3636
3637 out:
3638         if (ret)
3639                 btrfs_err(fs_info, "could not do orphan cleanup %d", ret);
3640         btrfs_free_path(path);
3641         return ret;
3642 }
3643
3644 /*
3645  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3646  * don't find any xattrs, we know there can't be any acls.
3647  *
3648  * slot is the slot the inode is in, objectid is the objectid of the inode
3649  */
3650 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3651                                           int slot, u64 objectid,
3652                                           int *first_xattr_slot)
3653 {
3654         u32 nritems = btrfs_header_nritems(leaf);
3655         struct btrfs_key found_key;
3656         static u64 xattr_access = 0;
3657         static u64 xattr_default = 0;
3658         int scanned = 0;
3659
3660         if (!xattr_access) {
3661                 xattr_access = btrfs_name_hash(XATTR_NAME_POSIX_ACL_ACCESS,
3662                                         strlen(XATTR_NAME_POSIX_ACL_ACCESS));
3663                 xattr_default = btrfs_name_hash(XATTR_NAME_POSIX_ACL_DEFAULT,
3664                                         strlen(XATTR_NAME_POSIX_ACL_DEFAULT));
3665         }
3666
3667         slot++;
3668         *first_xattr_slot = -1;
3669         while (slot < nritems) {
3670                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3671
3672                 /* we found a different objectid, there must not be acls */
3673                 if (found_key.objectid != objectid)
3674                         return 0;
3675
3676                 /* we found an xattr, assume we've got an acl */
3677                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3678                         if (*first_xattr_slot == -1)
3679                                 *first_xattr_slot = slot;
3680                         if (found_key.offset == xattr_access ||
3681                             found_key.offset == xattr_default)
3682                                 return 1;
3683                 }
3684
3685                 /*
3686                  * we found a key greater than an xattr key, there can't
3687                  * be any acls later on
3688                  */
3689                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3690                         return 0;
3691
3692                 slot++;
3693                 scanned++;
3694
3695                 /*
3696                  * it goes inode, inode backrefs, xattrs, extents,
3697                  * so if there are a ton of hard links to an inode there can
3698                  * be a lot of backrefs.  Don't waste time searching too hard,
3699                  * this is just an optimization
3700                  */
3701                 if (scanned >= 8)
3702                         break;
3703         }
3704         /* we hit the end of the leaf before we found an xattr or
3705          * something larger than an xattr.  We have to assume the inode
3706          * has acls
3707          */
3708         if (*first_xattr_slot == -1)
3709                 *first_xattr_slot = slot;
3710         return 1;
3711 }
3712
3713 /*
3714  * read an inode from the btree into the in-memory inode
3715  */
3716 static int btrfs_read_locked_inode(struct inode *inode)
3717 {
3718         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
3719         struct btrfs_path *path;
3720         struct extent_buffer *leaf;
3721         struct btrfs_inode_item *inode_item;
3722         struct btrfs_root *root = BTRFS_I(inode)->root;
3723         struct btrfs_key location;
3724         unsigned long ptr;
3725         int maybe_acls;
3726         u32 rdev;
3727         int ret;
3728         bool filled = false;
3729         int first_xattr_slot;
3730
3731         ret = btrfs_fill_inode(inode, &rdev);
3732         if (!ret)
3733                 filled = true;
3734
3735         path = btrfs_alloc_path();
3736         if (!path) {
3737                 ret = -ENOMEM;
3738                 goto make_bad;
3739         }
3740
3741         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3742
3743         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3744         if (ret) {
3745                 if (ret > 0)
3746                         ret = -ENOENT;
3747                 goto make_bad;
3748         }
3749
3750         leaf = path->nodes[0];
3751
3752         if (filled)
3753                 goto cache_index;
3754
3755         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3756                                     struct btrfs_inode_item);
3757         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3758         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3759         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3760         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3761         btrfs_i_size_write(BTRFS_I(inode), btrfs_inode_size(leaf, inode_item));
3762
3763         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3764         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3765
3766         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3767         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3768
3769         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3770         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3771
3772         BTRFS_I(inode)->i_otime.tv_sec =
3773                 btrfs_timespec_sec(leaf, &inode_item->otime);
3774         BTRFS_I(inode)->i_otime.tv_nsec =
3775                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3776
3777         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3778         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3779         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3780
3781         inode_set_iversion_queried(inode,
3782                                    btrfs_inode_sequence(leaf, inode_item));
3783         inode->i_generation = BTRFS_I(inode)->generation;
3784         inode->i_rdev = 0;
3785         rdev = btrfs_inode_rdev(leaf, inode_item);
3786
3787         BTRFS_I(inode)->index_cnt = (u64)-1;
3788         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3789
3790 cache_index:
3791         /*
3792          * If we were modified in the current generation and evicted from memory
3793          * and then re-read we need to do a full sync since we don't have any
3794          * idea about which extents were modified before we were evicted from
3795          * cache.
3796          *
3797          * This is required for both inode re-read from disk and delayed inode
3798          * in delayed_nodes_tree.
3799          */
3800         if (BTRFS_I(inode)->last_trans == fs_info->generation)
3801                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3802                         &BTRFS_I(inode)->runtime_flags);
3803
3804         /*
3805          * We don't persist the id of the transaction where an unlink operation
3806          * against the inode was last made. So here we assume the inode might
3807          * have been evicted, and therefore the exact value of last_unlink_trans
3808          * lost, and set it to last_trans to avoid metadata inconsistencies
3809          * between the inode and its parent if the inode is fsync'ed and the log
3810          * replayed. For example, in the scenario:
3811          *
3812          * touch mydir/foo
3813          * ln mydir/foo mydir/bar
3814          * sync
3815          * unlink mydir/bar
3816          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3817          * xfs_io -c fsync mydir/foo
3818          * <power failure>
3819          * mount fs, triggers fsync log replay
3820          *
3821          * We must make sure that when we fsync our inode foo we also log its
3822          * parent inode, otherwise after log replay the parent still has the
3823          * dentry with the "bar" name but our inode foo has a link count of 1
3824          * and doesn't have an inode ref with the name "bar" anymore.
3825          *
3826          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3827          * but it guarantees correctness at the expense of occasional full
3828          * transaction commits on fsync if our inode is a directory, or if our
3829          * inode is not a directory, logging its parent unnecessarily.
3830          */
3831         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3832
3833         path->slots[0]++;
3834         if (inode->i_nlink != 1 ||
3835             path->slots[0] >= btrfs_header_nritems(leaf))
3836                 goto cache_acl;
3837
3838         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3839         if (location.objectid != btrfs_ino(BTRFS_I(inode)))
3840                 goto cache_acl;
3841
3842         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3843         if (location.type == BTRFS_INODE_REF_KEY) {
3844                 struct btrfs_inode_ref *ref;
3845
3846                 ref = (struct btrfs_inode_ref *)ptr;
3847                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3848         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3849                 struct btrfs_inode_extref *extref;
3850
3851                 extref = (struct btrfs_inode_extref *)ptr;
3852                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3853                                                                      extref);
3854         }
3855 cache_acl:
3856         /*
3857          * try to precache a NULL acl entry for files that don't have
3858          * any xattrs or acls
3859          */
3860         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3861                         btrfs_ino(BTRFS_I(inode)), &first_xattr_slot);
3862         if (first_xattr_slot != -1) {
3863                 path->slots[0] = first_xattr_slot;
3864                 ret = btrfs_load_inode_props(inode, path);
3865                 if (ret)
3866                         btrfs_err(fs_info,
3867                                   "error loading props for ino %llu (root %llu): %d",
3868                                   btrfs_ino(BTRFS_I(inode)),
3869                                   root->root_key.objectid, ret);
3870         }
3871         btrfs_free_path(path);
3872
3873         if (!maybe_acls)
3874                 cache_no_acl(inode);
3875
3876         switch (inode->i_mode & S_IFMT) {
3877         case S_IFREG:
3878                 inode->i_mapping->a_ops = &btrfs_aops;
3879                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3880                 inode->i_fop = &btrfs_file_operations;
3881                 inode->i_op = &btrfs_file_inode_operations;
3882                 break;
3883         case S_IFDIR:
3884                 inode->i_fop = &btrfs_dir_file_operations;
3885                 inode->i_op = &btrfs_dir_inode_operations;
3886                 break;
3887         case S_IFLNK:
3888                 inode->i_op = &btrfs_symlink_inode_operations;
3889                 inode_nohighmem(inode);
3890                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3891                 break;
3892         default:
3893                 inode->i_op = &btrfs_special_inode_operations;
3894                 init_special_inode(inode, inode->i_mode, rdev);
3895                 break;
3896         }
3897
3898         btrfs_update_iflags(inode);
3899         return 0;
3900
3901 make_bad:
3902         btrfs_free_path(path);
3903         make_bad_inode(inode);
3904         return ret;
3905 }
3906
3907 /*
3908  * given a leaf and an inode, copy the inode fields into the leaf
3909  */
3910 static void fill_inode_item(struct btrfs_trans_handle *trans,
3911                             struct extent_buffer *leaf,
3912                             struct btrfs_inode_item *item,
3913                             struct inode *inode)
3914 {
3915         struct btrfs_map_token token;
3916
3917         btrfs_init_map_token(&token);
3918
3919         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3920         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3921         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3922                                    &token);
3923         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3924         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3925
3926         btrfs_set_token_timespec_sec(leaf, &item->atime,
3927                                      inode->i_atime.tv_sec, &token);
3928         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3929                                       inode->i_atime.tv_nsec, &token);
3930
3931         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3932                                      inode->i_mtime.tv_sec, &token);
3933         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3934                                       inode->i_mtime.tv_nsec, &token);
3935
3936         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3937                                      inode->i_ctime.tv_sec, &token);
3938         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3939                                       inode->i_ctime.tv_nsec, &token);
3940
3941         btrfs_set_token_timespec_sec(leaf, &item->otime,
3942                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3943         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3944                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3945
3946         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3947                                      &token);
3948         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3949                                          &token);
3950         btrfs_set_token_inode_sequence(leaf, item, inode_peek_iversion(inode),
3951                                        &token);
3952         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3953         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3954         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3955         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3956 }
3957
3958 /*
3959  * copy everything in the in-memory inode into the btree.
3960  */
3961 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3962                                 struct btrfs_root *root, struct inode *inode)
3963 {
3964         struct btrfs_inode_item *inode_item;
3965         struct btrfs_path *path;
3966         struct extent_buffer *leaf;
3967         int ret;
3968
3969         path = btrfs_alloc_path();
3970         if (!path)
3971                 return -ENOMEM;
3972
3973         path->leave_spinning = 1;
3974         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3975                                  1);
3976         if (ret) {
3977                 if (ret > 0)
3978                         ret = -ENOENT;
3979                 goto failed;
3980         }
3981
3982         leaf = path->nodes[0];
3983         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3984                                     struct btrfs_inode_item);
3985
3986         fill_inode_item(trans, leaf, inode_item, inode);
3987         btrfs_mark_buffer_dirty(leaf);
3988         btrfs_set_inode_last_trans(trans, inode);
3989         ret = 0;
3990 failed:
3991         btrfs_free_path(path);
3992         return ret;
3993 }
3994
3995 /*
3996  * copy everything in the in-memory inode into the btree.
3997  */
3998 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3999                                 struct btrfs_root *root, struct inode *inode)
4000 {
4001         struct btrfs_fs_info *fs_info = root->fs_info;
4002         int ret;
4003
4004         /*
4005          * If the inode is a free space inode, we can deadlock during commit
4006          * if we put it into the delayed code.
4007          *
4008          * The data relocation inode should also be directly updated
4009          * without delay
4010          */
4011         if (!btrfs_is_free_space_inode(BTRFS_I(inode))
4012             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
4013             && !test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
4014                 btrfs_update_root_times(trans, root);
4015
4016                 ret = btrfs_delayed_update_inode(trans, root, inode);
4017                 if (!ret)
4018                         btrfs_set_inode_last_trans(trans, inode);
4019                 return ret;
4020         }
4021
4022         return btrfs_update_inode_item(trans, root, inode);
4023 }
4024
4025 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
4026                                          struct btrfs_root *root,
4027                                          struct inode *inode)
4028 {
4029         int ret;
4030
4031         ret = btrfs_update_inode(trans, root, inode);
4032         if (ret == -ENOSPC)
4033                 return btrfs_update_inode_item(trans, root, inode);
4034         return ret;
4035 }
4036
4037 /*
4038  * unlink helper that gets used here in inode.c and in the tree logging
4039  * recovery code.  It remove a link in a directory with a given name, and
4040  * also drops the back refs in the inode to the directory
4041  */
4042 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4043                                 struct btrfs_root *root,
4044                                 struct btrfs_inode *dir,
4045                                 struct btrfs_inode *inode,
4046                                 const char *name, int name_len)
4047 {
4048         struct btrfs_fs_info *fs_info = root->fs_info;
4049         struct btrfs_path *path;
4050         int ret = 0;
4051         struct extent_buffer *leaf;
4052         struct btrfs_dir_item *di;
4053         struct btrfs_key key;
4054         u64 index;
4055         u64 ino = btrfs_ino(inode);
4056         u64 dir_ino = btrfs_ino(dir);
4057
4058         path = btrfs_alloc_path();
4059         if (!path) {
4060                 ret = -ENOMEM;
4061                 goto out;
4062         }
4063
4064         path->leave_spinning = 1;
4065         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4066                                     name, name_len, -1);
4067         if (IS_ERR(di)) {
4068                 ret = PTR_ERR(di);
4069                 goto err;
4070         }
4071         if (!di) {
4072                 ret = -ENOENT;
4073                 goto err;
4074         }
4075         leaf = path->nodes[0];
4076         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4077         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4078         if (ret)
4079                 goto err;
4080         btrfs_release_path(path);
4081
4082         /*
4083          * If we don't have dir index, we have to get it by looking up
4084          * the inode ref, since we get the inode ref, remove it directly,
4085          * it is unnecessary to do delayed deletion.
4086          *
4087          * But if we have dir index, needn't search inode ref to get it.
4088          * Since the inode ref is close to the inode item, it is better
4089          * that we delay to delete it, and just do this deletion when
4090          * we update the inode item.
4091          */
4092         if (inode->dir_index) {
4093                 ret = btrfs_delayed_delete_inode_ref(inode);
4094                 if (!ret) {
4095                         index = inode->dir_index;
4096                         goto skip_backref;
4097                 }
4098         }
4099
4100         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
4101                                   dir_ino, &index);
4102         if (ret) {
4103                 btrfs_info(fs_info,
4104                         "failed to delete reference to %.*s, inode %llu parent %llu",
4105                         name_len, name, ino, dir_ino);
4106                 btrfs_abort_transaction(trans, ret);
4107                 goto err;
4108         }
4109 skip_backref:
4110         ret = btrfs_delete_delayed_dir_index(trans, fs_info, dir, index);
4111         if (ret) {
4112                 btrfs_abort_transaction(trans, ret);
4113                 goto err;
4114         }
4115
4116         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len, inode,
4117                         dir_ino);
4118         if (ret != 0 && ret != -ENOENT) {
4119                 btrfs_abort_transaction(trans, ret);
4120                 goto err;
4121         }
4122
4123         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len, dir,
4124                         index);
4125         if (ret == -ENOENT)
4126                 ret = 0;
4127         else if (ret)
4128                 btrfs_abort_transaction(trans, ret);
4129 err:
4130         btrfs_free_path(path);
4131         if (ret)
4132                 goto out;
4133
4134         btrfs_i_size_write(dir, dir->vfs_inode.i_size - name_len * 2);
4135         inode_inc_iversion(&inode->vfs_inode);
4136         inode_inc_iversion(&dir->vfs_inode);
4137         inode->vfs_inode.i_ctime = dir->vfs_inode.i_mtime =
4138                 dir->vfs_inode.i_ctime = current_time(&inode->vfs_inode);
4139         ret = btrfs_update_inode(trans, root, &dir->vfs_inode);
4140 out:
4141         return ret;
4142 }
4143
4144 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4145                        struct btrfs_root *root,
4146                        struct btrfs_inode *dir, struct btrfs_inode *inode,
4147                        const char *name, int name_len)
4148 {
4149         int ret;
4150         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4151         if (!ret) {
4152                 drop_nlink(&inode->vfs_inode);
4153                 ret = btrfs_update_inode(trans, root, &inode->vfs_inode);
4154         }
4155         return ret;
4156 }
4157
4158 /*
4159  * helper to start transaction for unlink and rmdir.
4160  *
4161  * unlink and rmdir are special in btrfs, they do not always free space, so
4162  * if we cannot make our reservations the normal way try and see if there is
4163  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4164  * allow the unlink to occur.
4165  */
4166 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4167 {
4168         struct btrfs_root *root = BTRFS_I(dir)->root;
4169
4170         /*
4171          * 1 for the possible orphan item
4172          * 1 for the dir item
4173          * 1 for the dir index
4174          * 1 for the inode ref
4175          * 1 for the inode
4176          */
4177         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4178 }
4179
4180 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4181 {
4182         struct btrfs_root *root = BTRFS_I(dir)->root;
4183         struct btrfs_trans_handle *trans;
4184         struct inode *inode = d_inode(dentry);
4185         int ret;
4186
4187         trans = __unlink_start_trans(dir);
4188         if (IS_ERR(trans))
4189                 return PTR_ERR(trans);
4190
4191         btrfs_record_unlink_dir(trans, BTRFS_I(dir), BTRFS_I(d_inode(dentry)),
4192                         0);
4193
4194         ret = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4195                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4196                         dentry->d_name.len);
4197         if (ret)
4198                 goto out;
4199
4200         if (inode->i_nlink == 0) {
4201                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
4202                 if (ret)
4203                         goto out;
4204         }
4205
4206 out:
4207         btrfs_end_transaction(trans);
4208         btrfs_btree_balance_dirty(root->fs_info);
4209         return ret;
4210 }
4211
4212 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4213                         struct btrfs_root *root,
4214                         struct inode *dir, u64 objectid,
4215                         const char *name, int name_len)
4216 {
4217         struct btrfs_fs_info *fs_info = root->fs_info;
4218         struct btrfs_path *path;
4219         struct extent_buffer *leaf;
4220         struct btrfs_dir_item *di;
4221         struct btrfs_key key;
4222         u64 index;
4223         int ret;
4224         u64 dir_ino = btrfs_ino(BTRFS_I(dir));
4225
4226         path = btrfs_alloc_path();
4227         if (!path)
4228                 return -ENOMEM;
4229
4230         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4231                                    name, name_len, -1);
4232         if (IS_ERR_OR_NULL(di)) {
4233                 if (!di)
4234                         ret = -ENOENT;
4235                 else
4236                         ret = PTR_ERR(di);
4237                 goto out;
4238         }
4239
4240         leaf = path->nodes[0];
4241         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4242         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4243         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4244         if (ret) {
4245                 btrfs_abort_transaction(trans, ret);
4246                 goto out;
4247         }
4248         btrfs_release_path(path);
4249
4250         ret = btrfs_del_root_ref(trans, fs_info, objectid,
4251                                  root->root_key.objectid, dir_ino,
4252                                  &index, name, name_len);
4253         if (ret < 0) {
4254                 if (ret != -ENOENT) {
4255                         btrfs_abort_transaction(trans, ret);
4256                         goto out;
4257                 }
4258                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4259                                                  name, name_len);
4260                 if (IS_ERR_OR_NULL(di)) {
4261                         if (!di)
4262                                 ret = -ENOENT;
4263                         else
4264                                 ret = PTR_ERR(di);
4265                         btrfs_abort_transaction(trans, ret);
4266                         goto out;
4267                 }
4268
4269                 leaf = path->nodes[0];
4270                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4271                 btrfs_release_path(path);
4272                 index = key.offset;
4273         }
4274         btrfs_release_path(path);
4275
4276         ret = btrfs_delete_delayed_dir_index(trans, fs_info, BTRFS_I(dir), index);
4277         if (ret) {
4278                 btrfs_abort_transaction(trans, ret);
4279                 goto out;
4280         }
4281
4282         btrfs_i_size_write(BTRFS_I(dir), dir->i_size - name_len * 2);
4283         inode_inc_iversion(dir);
4284         dir->i_mtime = dir->i_ctime = current_time(dir);
4285         ret = btrfs_update_inode_fallback(trans, root, dir);
4286         if (ret)
4287                 btrfs_abort_transaction(trans, ret);
4288 out:
4289         btrfs_free_path(path);
4290         return ret;
4291 }
4292
4293 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4294 {
4295         struct inode *inode = d_inode(dentry);
4296         int err = 0;
4297         struct btrfs_root *root = BTRFS_I(dir)->root;
4298         struct btrfs_trans_handle *trans;
4299         u64 last_unlink_trans;
4300
4301         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4302                 return -ENOTEMPTY;
4303         if (btrfs_ino(BTRFS_I(inode)) == BTRFS_FIRST_FREE_OBJECTID)
4304                 return -EPERM;
4305
4306         trans = __unlink_start_trans(dir);
4307         if (IS_ERR(trans))
4308                 return PTR_ERR(trans);
4309
4310         if (unlikely(btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4311                 err = btrfs_unlink_subvol(trans, root, dir,
4312                                           BTRFS_I(inode)->location.objectid,
4313                                           dentry->d_name.name,
4314                                           dentry->d_name.len);
4315                 goto out;
4316         }
4317
4318         err = btrfs_orphan_add(trans, BTRFS_I(inode));
4319         if (err)
4320                 goto out;
4321
4322         last_unlink_trans = BTRFS_I(inode)->last_unlink_trans;
4323
4324         /* now the directory is empty */
4325         err = btrfs_unlink_inode(trans, root, BTRFS_I(dir),
4326                         BTRFS_I(d_inode(dentry)), dentry->d_name.name,
4327                         dentry->d_name.len);
4328         if (!err) {
4329                 btrfs_i_size_write(BTRFS_I(inode), 0);
4330                 /*
4331                  * Propagate the last_unlink_trans value of the deleted dir to
4332                  * its parent directory. This is to prevent an unrecoverable
4333                  * log tree in the case we do something like this:
4334                  * 1) create dir foo
4335                  * 2) create snapshot under dir foo
4336                  * 3) delete the snapshot
4337                  * 4) rmdir foo
4338                  * 5) mkdir foo
4339                  * 6) fsync foo or some file inside foo
4340                  */
4341                 if (last_unlink_trans >= trans->transid)
4342                         BTRFS_I(dir)->last_unlink_trans = last_unlink_trans;
4343         }
4344 out:
4345         btrfs_end_transaction(trans);
4346         btrfs_btree_balance_dirty(root->fs_info);
4347
4348         return err;
4349 }
4350
4351 static int truncate_space_check(struct btrfs_trans_handle *trans,
4352                                 struct btrfs_root *root,
4353                                 u64 bytes_deleted)
4354 {
4355         struct btrfs_fs_info *fs_info = root->fs_info;
4356         int ret;
4357
4358         /*
4359          * This is only used to apply pressure to the enospc system, we don't
4360          * intend to use this reservation at all.
4361          */
4362         bytes_deleted = btrfs_csum_bytes_to_leaves(fs_info, bytes_deleted);
4363         bytes_deleted *= fs_info->nodesize;
4364         ret = btrfs_block_rsv_add(root, &fs_info->trans_block_rsv,
4365                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4366         if (!ret) {
4367                 trace_btrfs_space_reservation(fs_info, "transaction",
4368                                               trans->transid,
4369                                               bytes_deleted, 1);
4370                 trans->bytes_reserved += bytes_deleted;
4371         }
4372         return ret;
4373
4374 }
4375
4376 /*
4377  * Return this if we need to call truncate_block for the last bit of the
4378  * truncate.
4379  */
4380 #define NEED_TRUNCATE_BLOCK 1
4381
4382 /*
4383  * this can truncate away extent items, csum items and directory items.
4384  * It starts at a high offset and removes keys until it can't find
4385  * any higher than new_size
4386  *
4387  * csum items that cross the new i_size are truncated to the new size
4388  * as well.
4389  *
4390  * min_type is the minimum key type to truncate down to.  If set to 0, this
4391  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4392  */
4393 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4394                                struct btrfs_root *root,
4395                                struct inode *inode,
4396                                u64 new_size, u32 min_type)
4397 {
4398         struct btrfs_fs_info *fs_info = root->fs_info;
4399         struct btrfs_path *path;
4400         struct extent_buffer *leaf;
4401         struct btrfs_file_extent_item *fi;
4402         struct btrfs_key key;
4403         struct btrfs_key found_key;
4404         u64 extent_start = 0;
4405         u64 extent_num_bytes = 0;
4406         u64 extent_offset = 0;
4407         u64 item_end = 0;
4408         u64 last_size = new_size;
4409         u32 found_type = (u8)-1;
4410         int found_extent;
4411         int del_item;
4412         int pending_del_nr = 0;
4413         int pending_del_slot = 0;
4414         int extent_type = -1;
4415         int ret;
4416         int err = 0;
4417         u64 ino = btrfs_ino(BTRFS_I(inode));
4418         u64 bytes_deleted = 0;
4419         bool be_nice = false;
4420         bool should_throttle = false;
4421         bool should_end = false;
4422
4423         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4424
4425         /*
4426          * for non-free space inodes and ref cows, we want to back off from
4427          * time to time
4428          */
4429         if (!btrfs_is_free_space_inode(BTRFS_I(inode)) &&
4430             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4431                 be_nice = true;
4432
4433         path = btrfs_alloc_path();
4434         if (!path)
4435                 return -ENOMEM;
4436         path->reada = READA_BACK;
4437
4438         /*
4439          * We want to drop from the next block forward in case this new size is
4440          * not block aligned since we will be keeping the last block of the
4441          * extent just the way it is.
4442          */
4443         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4444             root == fs_info->tree_root)
4445                 btrfs_drop_extent_cache(BTRFS_I(inode), ALIGN(new_size,
4446                                         fs_info->sectorsize),
4447                                         (u64)-1, 0);
4448
4449         /*
4450          * This function is also used to drop the items in the log tree before
4451          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4452          * it is used to drop the loged items. So we shouldn't kill the delayed
4453          * items.
4454          */
4455         if (min_type == 0 && root == BTRFS_I(inode)->root)
4456                 btrfs_kill_delayed_inode_items(BTRFS_I(inode));
4457
4458         key.objectid = ino;
4459         key.offset = (u64)-1;
4460         key.type = (u8)-1;
4461
4462 search_again:
4463         /*
4464          * with a 16K leaf size and 128MB extents, you can actually queue
4465          * up a huge file in a single leaf.  Most of the time that
4466          * bytes_deleted is > 0, it will be huge by the time we get here
4467          */
4468         if (be_nice && bytes_deleted > SZ_32M) {
4469                 if (btrfs_should_end_transaction(trans)) {
4470                         err = -EAGAIN;
4471                         goto error;
4472                 }
4473         }
4474
4475
4476         path->leave_spinning = 1;
4477         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4478         if (ret < 0) {
4479                 err = ret;
4480                 goto out;
4481         }
4482
4483         if (ret > 0) {
4484                 /* there are no items in the tree for us to truncate, we're
4485                  * done
4486                  */
4487                 if (path->slots[0] == 0)
4488                         goto out;
4489                 path->slots[0]--;
4490         }
4491
4492         while (1) {
4493                 fi = NULL;
4494                 leaf = path->nodes[0];
4495                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4496                 found_type = found_key.type;
4497
4498                 if (found_key.objectid != ino)
4499                         break;
4500
4501                 if (found_type < min_type)
4502                         break;
4503
4504                 item_end = found_key.offset;
4505                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4506                         fi = btrfs_item_ptr(leaf, path->slots[0],
4507                                             struct btrfs_file_extent_item);
4508                         extent_type = btrfs_file_extent_type(leaf, fi);
4509                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4510                                 item_end +=
4511                                     btrfs_file_extent_num_bytes(leaf, fi);
4512
4513                                 trace_btrfs_truncate_show_fi_regular(
4514                                         BTRFS_I(inode), leaf, fi,
4515                                         found_key.offset);
4516                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4517                                 item_end += btrfs_file_extent_inline_len(leaf,
4518                                                          path->slots[0], fi);
4519
4520                                 trace_btrfs_truncate_show_fi_inline(
4521                                         BTRFS_I(inode), leaf, fi, path->slots[0],
4522                                         found_key.offset);
4523                         }
4524                         item_end--;
4525                 }
4526                 if (found_type > min_type) {
4527                         del_item = 1;
4528                 } else {
4529                         if (item_end < new_size)
4530                                 break;
4531                         if (found_key.offset >= new_size)
4532                                 del_item = 1;
4533                         else
4534                                 del_item = 0;
4535                 }
4536                 found_extent = 0;
4537                 /* FIXME, shrink the extent if the ref count is only 1 */
4538                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4539                         goto delete;
4540
4541                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4542                         u64 num_dec;
4543                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4544                         if (!del_item) {
4545                                 u64 orig_num_bytes =
4546                                         btrfs_file_extent_num_bytes(leaf, fi);
4547                                 extent_num_bytes = ALIGN(new_size -
4548                                                 found_key.offset,
4549                                                 fs_info->sectorsize);
4550                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4551                                                          extent_num_bytes);
4552                                 num_dec = (orig_num_bytes -
4553                                            extent_num_bytes);
4554                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4555                                              &root->state) &&
4556                                     extent_start != 0)
4557                                         inode_sub_bytes(inode, num_dec);
4558                                 btrfs_mark_buffer_dirty(leaf);
4559                         } else {
4560                                 extent_num_bytes =
4561                                         btrfs_file_extent_disk_num_bytes(leaf,
4562                                                                          fi);
4563                                 extent_offset = found_key.offset -
4564                                         btrfs_file_extent_offset(leaf, fi);
4565
4566                                 /* FIXME blocksize != 4096 */
4567                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4568                                 if (extent_start != 0) {
4569                                         found_extent = 1;
4570                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4571                                                      &root->state))
4572                                                 inode_sub_bytes(inode, num_dec);
4573                                 }
4574                         }
4575                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4576                         /*
4577                          * we can't truncate inline items that have had
4578                          * special encodings
4579                          */
4580                         if (!del_item &&
4581                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4582                             btrfs_file_extent_other_encoding(leaf, fi) == 0 &&
4583                             btrfs_file_extent_compression(leaf, fi) == 0) {
4584                                 u32 size = (u32)(new_size - found_key.offset);
4585
4586                                 btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4587                                 size = btrfs_file_extent_calc_inline_size(size);
4588                                 btrfs_truncate_item(root->fs_info, path, size, 1);
4589                         } else if (!del_item) {
4590                                 /*
4591                                  * We have to bail so the last_size is set to
4592                                  * just before this extent.
4593                                  */
4594                                 err = NEED_TRUNCATE_BLOCK;
4595                                 break;
4596                         }
4597
4598                         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4599                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4600                 }
4601 delete:
4602                 if (del_item)
4603                         last_size = found_key.offset;
4604                 else
4605                         last_size = new_size;
4606                 if (del_item) {
4607                         if (!pending_del_nr) {
4608                                 /* no pending yet, add ourselves */
4609                                 pending_del_slot = path->slots[0];
4610                                 pending_del_nr = 1;
4611                         } else if (pending_del_nr &&
4612                                    path->slots[0] + 1 == pending_del_slot) {
4613                                 /* hop on the pending chunk */
4614                                 pending_del_nr++;
4615                                 pending_del_slot = path->slots[0];
4616                         } else {
4617                                 BUG();
4618                         }
4619                 } else {
4620                         break;
4621                 }
4622                 should_throttle = false;
4623
4624                 if (found_extent &&
4625                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4626                      root == fs_info->tree_root)) {
4627                         btrfs_set_path_blocking(path);
4628                         bytes_deleted += extent_num_bytes;
4629                         ret = btrfs_free_extent(trans, root, extent_start,
4630                                                 extent_num_bytes, 0,
4631                                                 btrfs_header_owner(leaf),
4632                                                 ino, extent_offset);
4633                         BUG_ON(ret);
4634                         if (btrfs_should_throttle_delayed_refs(trans, fs_info))
4635                                 btrfs_async_run_delayed_refs(fs_info,
4636                                         trans->delayed_ref_updates * 2,
4637                                         trans->transid, 0);
4638                         if (be_nice) {
4639                                 if (truncate_space_check(trans, root,
4640                                                          extent_num_bytes)) {
4641                                         should_end = true;
4642                                 }
4643                                 if (btrfs_should_throttle_delayed_refs(trans,
4644                                                                        fs_info))
4645                                         should_throttle = true;
4646                         }
4647                 }
4648
4649                 if (found_type == BTRFS_INODE_ITEM_KEY)
4650                         break;
4651
4652                 if (path->slots[0] == 0 ||
4653                     path->slots[0] != pending_del_slot ||
4654                     should_throttle || should_end) {
4655                         if (pending_del_nr) {
4656                                 ret = btrfs_del_items(trans, root, path,
4657                                                 pending_del_slot,
4658                                                 pending_del_nr);
4659                                 if (ret) {
4660                                         btrfs_abort_transaction(trans, ret);
4661                                         goto error;
4662                                 }
4663                                 pending_del_nr = 0;
4664                         }
4665                         btrfs_release_path(path);
4666                         if (should_throttle) {
4667                                 unsigned long updates = trans->delayed_ref_updates;
4668                                 if (updates) {
4669                                         trans->delayed_ref_updates = 0;
4670                                         ret = btrfs_run_delayed_refs(trans,
4671                                                                    fs_info,
4672                                                                    updates * 2);
4673                                         if (ret && !err)
4674                                                 err = ret;
4675                                 }
4676                         }
4677                         /*
4678                          * if we failed to refill our space rsv, bail out
4679                          * and let the transaction restart
4680                          */
4681                         if (should_end) {
4682                                 err = -EAGAIN;
4683                                 goto error;
4684                         }
4685                         goto search_again;
4686                 } else {
4687                         path->slots[0]--;
4688                 }
4689         }
4690 out:
4691         if (pending_del_nr) {
4692                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4693                                       pending_del_nr);
4694                 if (ret)
4695                         btrfs_abort_transaction(trans, ret);
4696         }
4697 error:
4698         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4699                 ASSERT(last_size >= new_size);
4700                 if (!err && last_size > new_size)
4701                         last_size = new_size;
4702                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4703         }
4704
4705         btrfs_free_path(path);
4706
4707         if (be_nice && bytes_deleted > SZ_32M) {
4708                 unsigned long updates = trans->delayed_ref_updates;
4709                 if (updates) {
4710                         trans->delayed_ref_updates = 0;
4711                         ret = btrfs_run_delayed_refs(trans, fs_info,
4712                                                      updates * 2);
4713                         if (ret && !err)
4714                                 err = ret;
4715                 }
4716         }
4717         return err;
4718 }
4719
4720 /*
4721  * btrfs_truncate_block - read, zero a chunk and write a block
4722  * @inode - inode that we're zeroing
4723  * @from - the offset to start zeroing
4724  * @len - the length to zero, 0 to zero the entire range respective to the
4725  *      offset
4726  * @front - zero up to the offset instead of from the offset on
4727  *
4728  * This will find the block for the "from" offset and cow the block and zero the
4729  * part we want to zero.  This is used with truncate and hole punching.
4730  */
4731 int btrfs_truncate_block(struct inode *inode, loff_t from, loff_t len,
4732                         int front)
4733 {
4734         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4735         struct address_space *mapping = inode->i_mapping;
4736         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4737         struct btrfs_ordered_extent *ordered;
4738         struct extent_state *cached_state = NULL;
4739         struct extent_changeset *data_reserved = NULL;
4740         char *kaddr;
4741         u32 blocksize = fs_info->sectorsize;
4742         pgoff_t index = from >> PAGE_SHIFT;
4743         unsigned offset = from & (blocksize - 1);
4744         struct page *page;
4745         gfp_t mask = btrfs_alloc_write_mask(mapping);
4746         int ret = 0;
4747         u64 block_start;
4748         u64 block_end;
4749
4750         if ((offset & (blocksize - 1)) == 0 &&
4751             (!len || ((len & (blocksize - 1)) == 0)))
4752                 goto out;
4753
4754         block_start = round_down(from, blocksize);
4755         block_end = block_start + blocksize - 1;
4756
4757         ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
4758                                            block_start, blocksize);
4759         if (ret)
4760                 goto out;
4761
4762 again:
4763         page = find_or_create_page(mapping, index, mask);
4764         if (!page) {
4765                 btrfs_delalloc_release_space(inode, data_reserved,
4766                                              block_start, blocksize);
4767                 btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4768                 ret = -ENOMEM;
4769                 goto out;
4770         }
4771
4772         if (!PageUptodate(page)) {
4773                 ret = btrfs_readpage(NULL, page);
4774                 lock_page(page);
4775                 if (page->mapping != mapping) {
4776                         unlock_page(page);
4777                         put_page(page);
4778                         goto again;
4779                 }
4780                 if (!PageUptodate(page)) {
4781                         ret = -EIO;
4782                         goto out_unlock;
4783                 }
4784         }
4785         wait_on_page_writeback(page);
4786
4787         lock_extent_bits(io_tree, block_start, block_end, &cached_state);
4788         set_page_extent_mapped(page);
4789
4790         ordered = btrfs_lookup_ordered_extent(inode, block_start);
4791         if (ordered) {
4792                 unlock_extent_cached(io_tree, block_start, block_end,
4793                                      &cached_state, GFP_NOFS);
4794                 unlock_page(page);
4795                 put_page(page);
4796                 btrfs_start_ordered_extent(inode, ordered, 1);
4797                 btrfs_put_ordered_extent(ordered);
4798                 goto again;
4799         }
4800
4801         clear_extent_bit(&BTRFS_I(inode)->io_tree, block_start, block_end,
4802                           EXTENT_DIRTY | EXTENT_DELALLOC |
4803                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4804                           0, 0, &cached_state, GFP_NOFS);
4805
4806         ret = btrfs_set_extent_delalloc(inode, block_start, block_end, 0,
4807                                         &cached_state, 0);
4808         if (ret) {
4809                 unlock_extent_cached(io_tree, block_start, block_end,
4810                                      &cached_state, GFP_NOFS);
4811                 goto out_unlock;
4812         }
4813
4814         if (offset != blocksize) {
4815                 if (!len)
4816                         len = blocksize - offset;
4817                 kaddr = kmap(page);
4818                 if (front)
4819                         memset(kaddr + (block_start - page_offset(page)),
4820                                 0, offset);
4821                 else
4822                         memset(kaddr + (block_start - page_offset(page)) +  offset,
4823                                 0, len);
4824                 flush_dcache_page(page);
4825                 kunmap(page);
4826         }
4827         ClearPageChecked(page);
4828         set_page_dirty(page);
4829         unlock_extent_cached(io_tree, block_start, block_end, &cached_state,
4830                              GFP_NOFS);
4831
4832 out_unlock:
4833         if (ret)
4834                 btrfs_delalloc_release_space(inode, data_reserved, block_start,
4835                                              blocksize);
4836         btrfs_delalloc_release_extents(BTRFS_I(inode), blocksize);
4837         unlock_page(page);
4838         put_page(page);
4839 out:
4840         extent_changeset_free(data_reserved);
4841         return ret;
4842 }
4843
4844 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4845                              u64 offset, u64 len)
4846 {
4847         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4848         struct btrfs_trans_handle *trans;
4849         int ret;
4850
4851         /*
4852          * Still need to make sure the inode looks like it's been updated so
4853          * that any holes get logged if we fsync.
4854          */
4855         if (btrfs_fs_incompat(fs_info, NO_HOLES)) {
4856                 BTRFS_I(inode)->last_trans = fs_info->generation;
4857                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4858                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4859                 return 0;
4860         }
4861
4862         /*
4863          * 1 - for the one we're dropping
4864          * 1 - for the one we're adding
4865          * 1 - for updating the inode.
4866          */
4867         trans = btrfs_start_transaction(root, 3);
4868         if (IS_ERR(trans))
4869                 return PTR_ERR(trans);
4870
4871         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4872         if (ret) {
4873                 btrfs_abort_transaction(trans, ret);
4874                 btrfs_end_transaction(trans);
4875                 return ret;
4876         }
4877
4878         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(BTRFS_I(inode)),
4879                         offset, 0, 0, len, 0, len, 0, 0, 0);
4880         if (ret)
4881                 btrfs_abort_transaction(trans, ret);
4882         else
4883                 btrfs_update_inode(trans, root, inode);
4884         btrfs_end_transaction(trans);
4885         return ret;
4886 }
4887
4888 /*
4889  * This function puts in dummy file extents for the area we're creating a hole
4890  * for.  So if we are truncating this file to a larger size we need to insert
4891  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4892  * the range between oldsize and size
4893  */
4894 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4895 {
4896         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
4897         struct btrfs_root *root = BTRFS_I(inode)->root;
4898         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4899         struct extent_map *em = NULL;
4900         struct extent_state *cached_state = NULL;
4901         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4902         u64 hole_start = ALIGN(oldsize, fs_info->sectorsize);
4903         u64 block_end = ALIGN(size, fs_info->sectorsize);
4904         u64 last_byte;
4905         u64 cur_offset;
4906         u64 hole_size;
4907         int err = 0;
4908
4909         /*
4910          * If our size started in the middle of a block we need to zero out the
4911          * rest of the block before we expand the i_size, otherwise we could
4912          * expose stale data.
4913          */
4914         err = btrfs_truncate_block(inode, oldsize, 0, 0);
4915         if (err)
4916                 return err;
4917
4918         if (size <= hole_start)
4919                 return 0;
4920
4921         while (1) {
4922                 struct btrfs_ordered_extent *ordered;
4923
4924                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4925                                  &cached_state);
4926                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), hole_start,
4927                                                      block_end - hole_start);
4928                 if (!ordered)
4929                         break;
4930                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4931                                      &cached_state, GFP_NOFS);
4932                 btrfs_start_ordered_extent(inode, ordered, 1);
4933                 btrfs_put_ordered_extent(ordered);
4934         }
4935
4936         cur_offset = hole_start;
4937         while (1) {
4938                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
4939                                 block_end - cur_offset, 0);
4940                 if (IS_ERR(em)) {
4941                         err = PTR_ERR(em);
4942                         em = NULL;
4943                         break;
4944                 }
4945                 last_byte = min(extent_map_end(em), block_end);
4946                 last_byte = ALIGN(last_byte, fs_info->sectorsize);
4947                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4948                         struct extent_map *hole_em;
4949                         hole_size = last_byte - cur_offset;
4950
4951                         err = maybe_insert_hole(root, inode, cur_offset,
4952                                                 hole_size);
4953                         if (err)
4954                                 break;
4955                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
4956                                                 cur_offset + hole_size - 1, 0);
4957                         hole_em = alloc_extent_map();
4958                         if (!hole_em) {
4959                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4960                                         &BTRFS_I(inode)->runtime_flags);
4961                                 goto next;
4962                         }
4963                         hole_em->start = cur_offset;
4964                         hole_em->len = hole_size;
4965                         hole_em->orig_start = cur_offset;
4966
4967                         hole_em->block_start = EXTENT_MAP_HOLE;
4968                         hole_em->block_len = 0;
4969                         hole_em->orig_block_len = 0;
4970                         hole_em->ram_bytes = hole_size;
4971                         hole_em->bdev = fs_info->fs_devices->latest_bdev;
4972                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4973                         hole_em->generation = fs_info->generation;
4974
4975                         while (1) {
4976                                 write_lock(&em_tree->lock);
4977                                 err = add_extent_mapping(em_tree, hole_em, 1);
4978                                 write_unlock(&em_tree->lock);
4979                                 if (err != -EEXIST)
4980                                         break;
4981                                 btrfs_drop_extent_cache(BTRFS_I(inode),
4982                                                         cur_offset,
4983                                                         cur_offset +
4984                                                         hole_size - 1, 0);
4985                         }
4986                         free_extent_map(hole_em);
4987                 }
4988 next:
4989                 free_extent_map(em);
4990                 em = NULL;
4991                 cur_offset = last_byte;
4992                 if (cur_offset >= block_end)
4993                         break;
4994         }
4995         free_extent_map(em);
4996         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4997                              GFP_NOFS);
4998         return err;
4999 }
5000
5001 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
5002 {
5003         struct btrfs_root *root = BTRFS_I(inode)->root;
5004         struct btrfs_trans_handle *trans;
5005         loff_t oldsize = i_size_read(inode);
5006         loff_t newsize = attr->ia_size;
5007         int mask = attr->ia_valid;
5008         int ret;
5009
5010         /*
5011          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
5012          * special case where we need to update the times despite not having
5013          * these flags set.  For all other operations the VFS set these flags
5014          * explicitly if it wants a timestamp update.
5015          */
5016         if (newsize != oldsize) {
5017                 inode_inc_iversion(inode);
5018                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
5019                         inode->i_ctime = inode->i_mtime =
5020                                 current_time(inode);
5021         }
5022
5023         if (newsize > oldsize) {
5024                 /*
5025                  * Don't do an expanding truncate while snapshotting is ongoing.
5026                  * This is to ensure the snapshot captures a fully consistent
5027                  * state of this file - if the snapshot captures this expanding
5028                  * truncation, it must capture all writes that happened before
5029                  * this truncation.
5030                  */
5031                 btrfs_wait_for_snapshot_creation(root);
5032                 ret = btrfs_cont_expand(inode, oldsize, newsize);
5033                 if (ret) {
5034                         btrfs_end_write_no_snapshotting(root);
5035                         return ret;
5036                 }
5037
5038                 trans = btrfs_start_transaction(root, 1);
5039                 if (IS_ERR(trans)) {
5040                         btrfs_end_write_no_snapshotting(root);
5041                         return PTR_ERR(trans);
5042                 }
5043
5044                 i_size_write(inode, newsize);
5045                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
5046                 pagecache_isize_extended(inode, oldsize, newsize);
5047                 ret = btrfs_update_inode(trans, root, inode);
5048                 btrfs_end_write_no_snapshotting(root);
5049                 btrfs_end_transaction(trans);
5050         } else {
5051
5052                 /*
5053                  * We're truncating a file that used to have good data down to
5054                  * zero. Make sure it gets into the ordered flush list so that
5055                  * any new writes get down to disk quickly.
5056                  */
5057                 if (newsize == 0)
5058                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
5059                                 &BTRFS_I(inode)->runtime_flags);
5060
5061                 /*
5062                  * 1 for the orphan item we're going to add
5063                  * 1 for the orphan item deletion.
5064                  */
5065                 trans = btrfs_start_transaction(root, 2);
5066                 if (IS_ERR(trans))
5067                         return PTR_ERR(trans);
5068
5069                 /*
5070                  * We need to do this in case we fail at _any_ point during the
5071                  * actual truncate.  Once we do the truncate_setsize we could
5072                  * invalidate pages which forces any outstanding ordered io to
5073                  * be instantly completed which will give us extents that need
5074                  * to be truncated.  If we fail to get an orphan inode down we
5075                  * could have left over extents that were never meant to live,
5076                  * so we need to guarantee from this point on that everything
5077                  * will be consistent.
5078                  */
5079                 ret = btrfs_orphan_add(trans, BTRFS_I(inode));
5080                 btrfs_end_transaction(trans);
5081                 if (ret)
5082                         return ret;
5083
5084                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
5085                 truncate_setsize(inode, newsize);
5086
5087                 /* Disable nonlocked read DIO to avoid the end less truncate */
5088                 btrfs_inode_block_unlocked_dio(BTRFS_I(inode));
5089                 inode_dio_wait(inode);
5090                 btrfs_inode_resume_unlocked_dio(BTRFS_I(inode));
5091
5092                 ret = btrfs_truncate(inode);
5093                 if (ret && inode->i_nlink) {
5094                         int err;
5095
5096                         /* To get a stable disk_i_size */
5097                         err = btrfs_wait_ordered_range(inode, 0, (u64)-1);
5098                         if (err) {
5099                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5100                                 return err;
5101                         }
5102
5103                         /*
5104                          * failed to truncate, disk_i_size is only adjusted down
5105                          * as we remove extents, so it should represent the true
5106                          * size of the inode, so reset the in memory size and
5107                          * delete our orphan entry.
5108                          */
5109                         trans = btrfs_join_transaction(root);
5110                         if (IS_ERR(trans)) {
5111                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5112                                 return ret;
5113                         }
5114                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5115                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
5116                         if (err)
5117                                 btrfs_abort_transaction(trans, err);
5118                         btrfs_end_transaction(trans);
5119                 }
5120         }
5121
5122         return ret;
5123 }
5124
5125 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5126 {
5127         struct inode *inode = d_inode(dentry);
5128         struct btrfs_root *root = BTRFS_I(inode)->root;
5129         int err;
5130
5131         if (btrfs_root_readonly(root))
5132                 return -EROFS;
5133
5134         err = setattr_prepare(dentry, attr);
5135         if (err)
5136                 return err;
5137
5138         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5139                 err = btrfs_setsize(inode, attr);
5140                 if (err)
5141                         return err;
5142         }
5143
5144         if (attr->ia_valid) {
5145                 setattr_copy(inode, attr);
5146                 inode_inc_iversion(inode);
5147                 err = btrfs_dirty_inode(inode);
5148
5149                 if (!err && attr->ia_valid & ATTR_MODE)
5150                         err = posix_acl_chmod(inode, inode->i_mode);
5151         }
5152
5153         return err;
5154 }
5155
5156 /*
5157  * While truncating the inode pages during eviction, we get the VFS calling
5158  * btrfs_invalidatepage() against each page of the inode. This is slow because
5159  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5160  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5161  * extent_state structures over and over, wasting lots of time.
5162  *
5163  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5164  * those expensive operations on a per page basis and do only the ordered io
5165  * finishing, while we release here the extent_map and extent_state structures,
5166  * without the excessive merging and splitting.
5167  */
5168 static void evict_inode_truncate_pages(struct inode *inode)
5169 {
5170         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5171         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5172         struct rb_node *node;
5173
5174         ASSERT(inode->i_state & I_FREEING);
5175         truncate_inode_pages_final(&inode->i_data);
5176
5177         write_lock(&map_tree->lock);
5178         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5179                 struct extent_map *em;
5180
5181                 node = rb_first(&map_tree->map);
5182                 em = rb_entry(node, struct extent_map, rb_node);
5183                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5184                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5185                 remove_extent_mapping(map_tree, em);
5186                 free_extent_map(em);
5187                 if (need_resched()) {
5188                         write_unlock(&map_tree->lock);
5189                         cond_resched();
5190                         write_lock(&map_tree->lock);
5191                 }
5192         }
5193         write_unlock(&map_tree->lock);
5194
5195         /*
5196          * Keep looping until we have no more ranges in the io tree.
5197          * We can have ongoing bios started by readpages (called from readahead)
5198          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5199          * still in progress (unlocked the pages in the bio but did not yet
5200          * unlocked the ranges in the io tree). Therefore this means some
5201          * ranges can still be locked and eviction started because before
5202          * submitting those bios, which are executed by a separate task (work
5203          * queue kthread), inode references (inode->i_count) were not taken
5204          * (which would be dropped in the end io callback of each bio).
5205          * Therefore here we effectively end up waiting for those bios and
5206          * anyone else holding locked ranges without having bumped the inode's
5207          * reference count - if we don't do it, when they access the inode's
5208          * io_tree to unlock a range it may be too late, leading to an
5209          * use-after-free issue.
5210          */
5211         spin_lock(&io_tree->lock);
5212         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5213                 struct extent_state *state;
5214                 struct extent_state *cached_state = NULL;
5215                 u64 start;
5216                 u64 end;
5217
5218                 node = rb_first(&io_tree->state);
5219                 state = rb_entry(node, struct extent_state, rb_node);
5220                 start = state->start;
5221                 end = state->end;
5222                 spin_unlock(&io_tree->lock);
5223
5224                 lock_extent_bits(io_tree, start, end, &cached_state);
5225
5226                 /*
5227                  * If still has DELALLOC flag, the extent didn't reach disk,
5228                  * and its reserved space won't be freed by delayed_ref.
5229                  * So we need to free its reserved space here.
5230                  * (Refer to comment in btrfs_invalidatepage, case 2)
5231                  *
5232                  * Note, end is the bytenr of last byte, so we need + 1 here.
5233                  */
5234                 if (state->state & EXTENT_DELALLOC)
5235                         btrfs_qgroup_free_data(inode, NULL, start, end - start + 1);
5236
5237                 clear_extent_bit(io_tree, start, end,
5238                                  EXTENT_LOCKED | EXTENT_DIRTY |
5239                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5240                                  EXTENT_DEFRAG, 1, 1,
5241                                  &cached_state, GFP_NOFS);
5242
5243                 cond_resched();
5244                 spin_lock(&io_tree->lock);
5245         }
5246         spin_unlock(&io_tree->lock);
5247 }
5248
5249 void btrfs_evict_inode(struct inode *inode)
5250 {
5251         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5252         struct btrfs_trans_handle *trans;
5253         struct btrfs_root *root = BTRFS_I(inode)->root;
5254         struct btrfs_block_rsv *rsv, *global_rsv;
5255         int steal_from_global = 0;
5256         u64 min_size;
5257         int ret;
5258
5259         trace_btrfs_inode_evict(inode);
5260
5261         if (!root) {
5262                 kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
5263                 return;
5264         }
5265
5266         min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
5267
5268         evict_inode_truncate_pages(inode);
5269
5270         if (inode->i_nlink &&
5271             ((btrfs_root_refs(&root->root_item) != 0 &&
5272               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5273              btrfs_is_free_space_inode(BTRFS_I(inode))))
5274                 goto no_delete;
5275
5276         if (is_bad_inode(inode)) {
5277                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5278                 goto no_delete;
5279         }
5280         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5281         if (!special_file(inode->i_mode))
5282                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5283
5284         btrfs_free_io_failure_record(BTRFS_I(inode), 0, (u64)-1);
5285
5286         if (test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags)) {
5287                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5288                                  &BTRFS_I(inode)->runtime_flags));
5289                 goto no_delete;
5290         }
5291
5292         if (inode->i_nlink > 0) {
5293                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5294                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5295                 goto no_delete;
5296         }
5297
5298         ret = btrfs_commit_inode_delayed_inode(BTRFS_I(inode));
5299         if (ret) {
5300                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5301                 goto no_delete;
5302         }
5303
5304         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
5305         if (!rsv) {
5306                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5307                 goto no_delete;
5308         }
5309         rsv->size = min_size;
5310         rsv->failfast = 1;
5311         global_rsv = &fs_info->global_block_rsv;
5312
5313         btrfs_i_size_write(BTRFS_I(inode), 0);
5314
5315         /*
5316          * This is a bit simpler than btrfs_truncate since we've already
5317          * reserved our space for our orphan item in the unlink, so we just
5318          * need to reserve some slack space in case we add bytes and update
5319          * inode item when doing the truncate.
5320          */
5321         while (1) {
5322                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5323                                              BTRFS_RESERVE_FLUSH_LIMIT);
5324
5325                 /*
5326                  * Try and steal from the global reserve since we will
5327                  * likely not use this space anyway, we want to try as
5328                  * hard as possible to get this to work.
5329                  */
5330                 if (ret)
5331                         steal_from_global++;
5332                 else
5333                         steal_from_global = 0;
5334                 ret = 0;
5335
5336                 /*
5337                  * steal_from_global == 0: we reserved stuff, hooray!
5338                  * steal_from_global == 1: we didn't reserve stuff, boo!
5339                  * steal_from_global == 2: we've committed, still not a lot of
5340                  * room but maybe we'll have room in the global reserve this
5341                  * time.
5342                  * steal_from_global == 3: abandon all hope!
5343                  */
5344                 if (steal_from_global > 2) {
5345                         btrfs_warn(fs_info,
5346                                    "Could not get space for a delete, will truncate on mount %d",
5347                                    ret);
5348                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5349                         btrfs_free_block_rsv(fs_info, rsv);
5350                         goto no_delete;
5351                 }
5352
5353                 trans = btrfs_join_transaction(root);
5354                 if (IS_ERR(trans)) {
5355                         btrfs_orphan_del(NULL, BTRFS_I(inode));
5356                         btrfs_free_block_rsv(fs_info, rsv);
5357                         goto no_delete;
5358                 }
5359
5360                 /*
5361                  * We can't just steal from the global reserve, we need to make
5362                  * sure there is room to do it, if not we need to commit and try
5363                  * again.
5364                  */
5365                 if (steal_from_global) {
5366                         if (!btrfs_check_space_for_delayed_refs(trans, fs_info))
5367                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5368                                                               min_size, 0);
5369                         else
5370                                 ret = -ENOSPC;
5371                 }
5372
5373                 /*
5374                  * Couldn't steal from the global reserve, we have too much
5375                  * pending stuff built up, commit the transaction and try it
5376                  * again.
5377                  */
5378                 if (ret) {
5379                         ret = btrfs_commit_transaction(trans);
5380                         if (ret) {
5381                                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5382                                 btrfs_free_block_rsv(fs_info, rsv);
5383                                 goto no_delete;
5384                         }
5385                         continue;
5386                 } else {
5387                         steal_from_global = 0;
5388                 }
5389
5390                 trans->block_rsv = rsv;
5391
5392                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5393                 if (ret != -ENOSPC && ret != -EAGAIN)
5394                         break;
5395
5396                 trans->block_rsv = &fs_info->trans_block_rsv;
5397                 btrfs_end_transaction(trans);
5398                 trans = NULL;
5399                 btrfs_btree_balance_dirty(fs_info);
5400         }
5401
5402         btrfs_free_block_rsv(fs_info, rsv);
5403
5404         /*
5405          * Errors here aren't a big deal, it just means we leave orphan items
5406          * in the tree.  They will be cleaned up on the next mount.
5407          */
5408         if (ret == 0) {
5409                 trans->block_rsv = root->orphan_block_rsv;
5410                 btrfs_orphan_del(trans, BTRFS_I(inode));
5411         } else {
5412                 btrfs_orphan_del(NULL, BTRFS_I(inode));
5413         }
5414
5415         trans->block_rsv = &fs_info->trans_block_rsv;
5416         if (!(root == fs_info->tree_root ||
5417               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5418                 btrfs_return_ino(root, btrfs_ino(BTRFS_I(inode)));
5419
5420         btrfs_end_transaction(trans);
5421         btrfs_btree_balance_dirty(fs_info);
5422 no_delete:
5423         btrfs_remove_delayed_node(BTRFS_I(inode));
5424         clear_inode(inode);
5425 }
5426
5427 /*
5428  * this returns the key found in the dir entry in the location pointer.
5429  * If no dir entries were found, location->objectid is 0.
5430  */
5431 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5432                                struct btrfs_key *location)
5433 {
5434         const char *name = dentry->d_name.name;
5435         int namelen = dentry->d_name.len;
5436         struct btrfs_dir_item *di;
5437         struct btrfs_path *path;
5438         struct btrfs_root *root = BTRFS_I(dir)->root;
5439         int ret = 0;
5440
5441         path = btrfs_alloc_path();
5442         if (!path)
5443                 return -ENOMEM;
5444
5445         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(BTRFS_I(dir)),
5446                         name, namelen, 0);
5447         if (IS_ERR(di))
5448                 ret = PTR_ERR(di);
5449
5450         if (IS_ERR_OR_NULL(di))
5451                 goto out_err;
5452
5453         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5454         if (location->type != BTRFS_INODE_ITEM_KEY &&
5455             location->type != BTRFS_ROOT_ITEM_KEY) {
5456                 btrfs_warn(root->fs_info,
5457 "%s gets something invalid in DIR_ITEM (name %s, directory ino %llu, location(%llu %u %llu))",
5458                            __func__, name, btrfs_ino(BTRFS_I(dir)),
5459                            location->objectid, location->type, location->offset);
5460                 goto out_err;
5461         }
5462 out:
5463         btrfs_free_path(path);
5464         return ret;
5465 out_err:
5466         location->objectid = 0;
5467         goto out;
5468 }
5469
5470 /*
5471  * when we hit a tree root in a directory, the btrfs part of the inode
5472  * needs to be changed to reflect the root directory of the tree root.  This
5473  * is kind of like crossing a mount point.
5474  */
5475 static int fixup_tree_root_location(struct btrfs_fs_info *fs_info,
5476                                     struct inode *dir,
5477                                     struct dentry *dentry,
5478                                     struct btrfs_key *location,
5479                                     struct btrfs_root **sub_root)
5480 {
5481         struct btrfs_path *path;
5482         struct btrfs_root *new_root;
5483         struct btrfs_root_ref *ref;
5484         struct extent_buffer *leaf;
5485         struct btrfs_key key;
5486         int ret;
5487         int err = 0;
5488
5489         path = btrfs_alloc_path();
5490         if (!path) {
5491                 err = -ENOMEM;
5492                 goto out;
5493         }
5494
5495         err = -ENOENT;
5496         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5497         key.type = BTRFS_ROOT_REF_KEY;
5498         key.offset = location->objectid;
5499
5500         ret = btrfs_search_slot(NULL, fs_info->tree_root, &key, path, 0, 0);
5501         if (ret) {
5502                 if (ret < 0)
5503                         err = ret;
5504                 goto out;
5505         }
5506
5507         leaf = path->nodes[0];
5508         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5509         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(BTRFS_I(dir)) ||
5510             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5511                 goto out;
5512
5513         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5514                                    (unsigned long)(ref + 1),
5515                                    dentry->d_name.len);
5516         if (ret)
5517                 goto out;
5518
5519         btrfs_release_path(path);
5520
5521         new_root = btrfs_read_fs_root_no_name(fs_info, location);
5522         if (IS_ERR(new_root)) {
5523                 err = PTR_ERR(new_root);
5524                 goto out;
5525         }
5526
5527         *sub_root = new_root;
5528         location->objectid = btrfs_root_dirid(&new_root->root_item);
5529         location->type = BTRFS_INODE_ITEM_KEY;
5530         location->offset = 0;
5531         err = 0;
5532 out:
5533         btrfs_free_path(path);
5534         return err;
5535 }
5536
5537 static void inode_tree_add(struct inode *inode)
5538 {
5539         struct btrfs_root *root = BTRFS_I(inode)->root;
5540         struct btrfs_inode *entry;
5541         struct rb_node **p;
5542         struct rb_node *parent;
5543         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5544         u64 ino = btrfs_ino(BTRFS_I(inode));
5545
5546         if (inode_unhashed(inode))
5547                 return;
5548         parent = NULL;
5549         spin_lock(&root->inode_lock);
5550         p = &root->inode_tree.rb_node;
5551         while (*p) {
5552                 parent = *p;
5553                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5554
5555                 if (ino < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5556                         p = &parent->rb_left;
5557                 else if (ino > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5558                         p = &parent->rb_right;
5559                 else {
5560                         WARN_ON(!(entry->vfs_inode.i_state &
5561                                   (I_WILL_FREE | I_FREEING)));
5562                         rb_replace_node(parent, new, &root->inode_tree);
5563                         RB_CLEAR_NODE(parent);
5564                         spin_unlock(&root->inode_lock);
5565                         return;
5566                 }
5567         }
5568         rb_link_node(new, parent, p);
5569         rb_insert_color(new, &root->inode_tree);
5570         spin_unlock(&root->inode_lock);
5571 }
5572
5573 static void inode_tree_del(struct inode *inode)
5574 {
5575         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5576         struct btrfs_root *root = BTRFS_I(inode)->root;
5577         int empty = 0;
5578
5579         spin_lock(&root->inode_lock);
5580         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5581                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5582                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5583                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5584         }
5585         spin_unlock(&root->inode_lock);
5586
5587         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5588                 synchronize_srcu(&fs_info->subvol_srcu);
5589                 spin_lock(&root->inode_lock);
5590                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5591                 spin_unlock(&root->inode_lock);
5592                 if (empty)
5593                         btrfs_add_dead_root(root);
5594         }
5595 }
5596
5597 void btrfs_invalidate_inodes(struct btrfs_root *root)
5598 {
5599         struct btrfs_fs_info *fs_info = root->fs_info;
5600         struct rb_node *node;
5601         struct rb_node *prev;
5602         struct btrfs_inode *entry;
5603         struct inode *inode;
5604         u64 objectid = 0;
5605
5606         if (!test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
5607                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5608
5609         spin_lock(&root->inode_lock);
5610 again:
5611         node = root->inode_tree.rb_node;
5612         prev = NULL;
5613         while (node) {
5614                 prev = node;
5615                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5616
5617                 if (objectid < btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5618                         node = node->rb_left;
5619                 else if (objectid > btrfs_ino(BTRFS_I(&entry->vfs_inode)))
5620                         node = node->rb_right;
5621                 else
5622                         break;
5623         }
5624         if (!node) {
5625                 while (prev) {
5626                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5627                         if (objectid <= btrfs_ino(BTRFS_I(&entry->vfs_inode))) {
5628                                 node = prev;
5629                                 break;
5630                         }
5631                         prev = rb_next(prev);
5632                 }
5633         }
5634         while (node) {
5635                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5636                 objectid = btrfs_ino(BTRFS_I(&entry->vfs_inode)) + 1;
5637                 inode = igrab(&entry->vfs_inode);
5638                 if (inode) {
5639                         spin_unlock(&root->inode_lock);
5640                         if (atomic_read(&inode->i_count) > 1)
5641                                 d_prune_aliases(inode);
5642                         /*
5643                          * btrfs_drop_inode will have it removed from
5644                          * the inode cache when its usage count
5645                          * hits zero.
5646                          */
5647                         iput(inode);
5648                         cond_resched();
5649                         spin_lock(&root->inode_lock);
5650                         goto again;
5651                 }
5652
5653                 if (cond_resched_lock(&root->inode_lock))
5654                         goto again;
5655
5656                 node = rb_next(node);
5657         }
5658         spin_unlock(&root->inode_lock);
5659 }
5660
5661 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5662 {
5663         struct btrfs_iget_args *args = p;
5664         inode->i_ino = args->location->objectid;
5665         memcpy(&BTRFS_I(inode)->location, args->location,
5666                sizeof(*args->location));
5667         BTRFS_I(inode)->root = args->root;
5668         return 0;
5669 }
5670
5671 static int btrfs_find_actor(struct inode *inode, void *opaque)
5672 {
5673         struct btrfs_iget_args *args = opaque;
5674         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5675                 args->root == BTRFS_I(inode)->root;
5676 }
5677
5678 static struct inode *btrfs_iget_locked(struct super_block *s,
5679                                        struct btrfs_key *location,
5680                                        struct btrfs_root *root)
5681 {
5682         struct inode *inode;
5683         struct btrfs_iget_args args;
5684         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5685
5686         args.location = location;
5687         args.root = root;
5688
5689         inode = iget5_locked(s, hashval, btrfs_find_actor,
5690                              btrfs_init_locked_inode,
5691                              (void *)&args);
5692         return inode;
5693 }
5694
5695 /* Get an inode object given its location and corresponding root.
5696  * Returns in *is_new if the inode was read from disk
5697  */
5698 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5699                          struct btrfs_root *root, int *new)
5700 {
5701         struct inode *inode;
5702
5703         inode = btrfs_iget_locked(s, location, root);
5704         if (!inode)
5705                 return ERR_PTR(-ENOMEM);
5706
5707         if (inode->i_state & I_NEW) {
5708                 int ret;
5709
5710                 ret = btrfs_read_locked_inode(inode);
5711                 if (!is_bad_inode(inode)) {
5712                         inode_tree_add(inode);
5713                         unlock_new_inode(inode);
5714                         if (new)
5715                                 *new = 1;
5716                 } else {
5717                         unlock_new_inode(inode);
5718                         iput(inode);
5719                         ASSERT(ret < 0);
5720                         inode = ERR_PTR(ret < 0 ? ret : -ESTALE);
5721                 }
5722         }
5723
5724         return inode;
5725 }
5726
5727 static struct inode *new_simple_dir(struct super_block *s,
5728                                     struct btrfs_key *key,
5729                                     struct btrfs_root *root)
5730 {
5731         struct inode *inode = new_inode(s);
5732
5733         if (!inode)
5734                 return ERR_PTR(-ENOMEM);
5735
5736         BTRFS_I(inode)->root = root;
5737         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5738         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5739
5740         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5741         inode->i_op = &btrfs_dir_ro_inode_operations;
5742         inode->i_opflags &= ~IOP_XATTR;
5743         inode->i_fop = &simple_dir_operations;
5744         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5745         inode->i_mtime = current_time(inode);
5746         inode->i_atime = inode->i_mtime;
5747         inode->i_ctime = inode->i_mtime;
5748         BTRFS_I(inode)->i_otime = inode->i_mtime;
5749
5750         return inode;
5751 }
5752
5753 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5754 {
5755         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
5756         struct inode *inode;
5757         struct btrfs_root *root = BTRFS_I(dir)->root;
5758         struct btrfs_root *sub_root = root;
5759         struct btrfs_key location;
5760         int index;
5761         int ret = 0;
5762
5763         if (dentry->d_name.len > BTRFS_NAME_LEN)
5764                 return ERR_PTR(-ENAMETOOLONG);
5765
5766         ret = btrfs_inode_by_name(dir, dentry, &location);
5767         if (ret < 0)
5768                 return ERR_PTR(ret);
5769
5770         if (location.objectid == 0)
5771                 return ERR_PTR(-ENOENT);
5772
5773         if (location.type == BTRFS_INODE_ITEM_KEY) {
5774                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5775                 return inode;
5776         }
5777
5778         index = srcu_read_lock(&fs_info->subvol_srcu);
5779         ret = fixup_tree_root_location(fs_info, dir, dentry,
5780                                        &location, &sub_root);
5781         if (ret < 0) {
5782                 if (ret != -ENOENT)
5783                         inode = ERR_PTR(ret);
5784                 else
5785                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5786         } else {
5787                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5788         }
5789         srcu_read_unlock(&fs_info->subvol_srcu, index);
5790
5791         if (!IS_ERR(inode) && root != sub_root) {
5792                 down_read(&fs_info->cleanup_work_sem);
5793                 if (!sb_rdonly(inode->i_sb))
5794                         ret = btrfs_orphan_cleanup(sub_root);
5795                 up_read(&fs_info->cleanup_work_sem);
5796                 if (ret) {
5797                         iput(inode);
5798                         inode = ERR_PTR(ret);
5799                 }
5800         }
5801
5802         return inode;
5803 }
5804
5805 static int btrfs_dentry_delete(const struct dentry *dentry)
5806 {
5807         struct btrfs_root *root;
5808         struct inode *inode = d_inode(dentry);
5809
5810         if (!inode && !IS_ROOT(dentry))
5811                 inode = d_inode(dentry->d_parent);
5812
5813         if (inode) {
5814                 root = BTRFS_I(inode)->root;
5815                 if (btrfs_root_refs(&root->root_item) == 0)
5816                         return 1;
5817
5818                 if (btrfs_ino(BTRFS_I(inode)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5819                         return 1;
5820         }
5821         return 0;
5822 }
5823
5824 static void btrfs_dentry_release(struct dentry *dentry)
5825 {
5826         kfree(dentry->d_fsdata);
5827 }
5828
5829 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5830                                    unsigned int flags)
5831 {
5832         struct inode *inode;
5833
5834         inode = btrfs_lookup_dentry(dir, dentry);
5835         if (IS_ERR(inode)) {
5836                 if (PTR_ERR(inode) == -ENOENT)
5837                         inode = NULL;
5838                 else
5839                         return ERR_CAST(inode);
5840         }
5841
5842         return d_splice_alias(inode, dentry);
5843 }
5844
5845 unsigned char btrfs_filetype_table[] = {
5846         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5847 };
5848
5849 /*
5850  * All this infrastructure exists because dir_emit can fault, and we are holding
5851  * the tree lock when doing readdir.  For now just allocate a buffer and copy
5852  * our information into that, and then dir_emit from the buffer.  This is
5853  * similar to what NFS does, only we don't keep the buffer around in pagecache
5854  * because I'm afraid I'll mess that up.  Long term we need to make filldir do
5855  * copy_to_user_inatomic so we don't have to worry about page faulting under the
5856  * tree lock.
5857  */
5858 static int btrfs_opendir(struct inode *inode, struct file *file)
5859 {
5860         struct btrfs_file_private *private;
5861
5862         private = kzalloc(sizeof(struct btrfs_file_private), GFP_KERNEL);
5863         if (!private)
5864                 return -ENOMEM;
5865         private->filldir_buf = kzalloc(PAGE_SIZE, GFP_KERNEL);
5866         if (!private->filldir_buf) {
5867                 kfree(private);
5868                 return -ENOMEM;
5869         }
5870         file->private_data = private;
5871         return 0;
5872 }
5873
5874 struct dir_entry {
5875         u64 ino;
5876         u64 offset;
5877         unsigned type;
5878         int name_len;
5879 };
5880
5881 static int btrfs_filldir(void *addr, int entries, struct dir_context *ctx)
5882 {
5883         while (entries--) {
5884                 struct dir_entry *entry = addr;
5885                 char *name = (char *)(entry + 1);
5886
5887                 ctx->pos = entry->offset;
5888                 if (!dir_emit(ctx, name, entry->name_len, entry->ino,
5889                               entry->type))
5890                         return 1;
5891                 addr += sizeof(struct dir_entry) + entry->name_len;
5892                 ctx->pos++;
5893         }
5894         return 0;
5895 }
5896
5897 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5898 {
5899         struct inode *inode = file_inode(file);
5900         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
5901         struct btrfs_root *root = BTRFS_I(inode)->root;
5902         struct btrfs_file_private *private = file->private_data;
5903         struct btrfs_dir_item *di;
5904         struct btrfs_key key;
5905         struct btrfs_key found_key;
5906         struct btrfs_path *path;
5907         void *addr;
5908         struct list_head ins_list;
5909         struct list_head del_list;
5910         int ret;
5911         struct extent_buffer *leaf;
5912         int slot;
5913         char *name_ptr;
5914         int name_len;
5915         int entries = 0;
5916         int total_len = 0;
5917         bool put = false;
5918         struct btrfs_key location;
5919
5920         if (!dir_emit_dots(file, ctx))
5921                 return 0;
5922
5923         path = btrfs_alloc_path();
5924         if (!path)
5925                 return -ENOMEM;
5926
5927         addr = private->filldir_buf;
5928         path->reada = READA_FORWARD;
5929
5930         INIT_LIST_HEAD(&ins_list);
5931         INIT_LIST_HEAD(&del_list);
5932         put = btrfs_readdir_get_delayed_items(inode, &ins_list, &del_list);
5933
5934 again:
5935         key.type = BTRFS_DIR_INDEX_KEY;
5936         key.offset = ctx->pos;
5937         key.objectid = btrfs_ino(BTRFS_I(inode));
5938
5939         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5940         if (ret < 0)
5941                 goto err;
5942
5943         while (1) {
5944                 struct dir_entry *entry;
5945
5946                 leaf = path->nodes[0];
5947                 slot = path->slots[0];
5948                 if (slot >= btrfs_header_nritems(leaf)) {
5949                         ret = btrfs_next_leaf(root, path);
5950                         if (ret < 0)
5951                                 goto err;
5952                         else if (ret > 0)
5953                                 break;
5954                         continue;
5955                 }
5956
5957                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5958
5959                 if (found_key.objectid != key.objectid)
5960                         break;
5961                 if (found_key.type != BTRFS_DIR_INDEX_KEY)
5962                         break;
5963                 if (found_key.offset < ctx->pos)
5964                         goto next;
5965                 if (btrfs_should_delete_dir_index(&del_list, found_key.offset))
5966                         goto next;
5967                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5968                 if (verify_dir_item(fs_info, leaf, slot, di))
5969                         goto next;
5970
5971                 name_len = btrfs_dir_name_len(leaf, di);
5972                 if ((total_len + sizeof(struct dir_entry) + name_len) >=
5973                     PAGE_SIZE) {
5974                         btrfs_release_path(path);
5975                         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
5976                         if (ret)
5977                                 goto nopos;
5978                         addr = private->filldir_buf;
5979                         entries = 0;
5980                         total_len = 0;
5981                         goto again;
5982                 }
5983
5984                 entry = addr;
5985                 entry->name_len = name_len;
5986                 name_ptr = (char *)(entry + 1);
5987                 read_extent_buffer(leaf, name_ptr, (unsigned long)(di + 1),
5988                                    name_len);
5989                 entry->type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5990                 btrfs_dir_item_key_to_cpu(leaf, di, &location);
5991                 entry->ino = location.objectid;
5992                 entry->offset = found_key.offset;
5993                 entries++;
5994                 addr += sizeof(struct dir_entry) + name_len;
5995                 total_len += sizeof(struct dir_entry) + name_len;
5996 next:
5997                 path->slots[0]++;
5998         }
5999         btrfs_release_path(path);
6000
6001         ret = btrfs_filldir(private->filldir_buf, entries, ctx);
6002         if (ret)
6003                 goto nopos;
6004
6005         ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
6006         if (ret)
6007                 goto nopos;
6008
6009         /*
6010          * Stop new entries from being returned after we return the last
6011          * entry.
6012          *
6013          * New directory entries are assigned a strictly increasing
6014          * offset.  This means that new entries created during readdir
6015          * are *guaranteed* to be seen in the future by that readdir.
6016          * This has broken buggy programs which operate on names as
6017          * they're returned by readdir.  Until we re-use freed offsets
6018          * we have this hack to stop new entries from being returned
6019          * under the assumption that they'll never reach this huge
6020          * offset.
6021          *
6022          * This is being careful not to overflow 32bit loff_t unless the
6023          * last entry requires it because doing so has broken 32bit apps
6024          * in the past.
6025          */
6026         if (ctx->pos >= INT_MAX)
6027                 ctx->pos = LLONG_MAX;
6028         else
6029                 ctx->pos = INT_MAX;
6030 nopos:
6031         ret = 0;
6032 err:
6033         if (put)
6034                 btrfs_readdir_put_delayed_items(inode, &ins_list, &del_list);
6035         btrfs_free_path(path);
6036         return ret;
6037 }
6038
6039 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
6040 {
6041         struct btrfs_root *root = BTRFS_I(inode)->root;
6042         struct btrfs_trans_handle *trans;
6043         int ret = 0;
6044         bool nolock = false;
6045
6046         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6047                 return 0;
6048
6049         if (btrfs_fs_closing(root->fs_info) &&
6050                         btrfs_is_free_space_inode(BTRFS_I(inode)))
6051                 nolock = true;
6052
6053         if (wbc->sync_mode == WB_SYNC_ALL) {
6054                 if (nolock)
6055                         trans = btrfs_join_transaction_nolock(root);
6056                 else
6057                         trans = btrfs_join_transaction(root);
6058                 if (IS_ERR(trans))
6059                         return PTR_ERR(trans);
6060                 ret = btrfs_commit_transaction(trans);
6061         }
6062         return ret;
6063 }
6064
6065 /*
6066  * This is somewhat expensive, updating the tree every time the
6067  * inode changes.  But, it is most likely to find the inode in cache.
6068  * FIXME, needs more benchmarking...there are no reasons other than performance
6069  * to keep or drop this code.
6070  */
6071 static int btrfs_dirty_inode(struct inode *inode)
6072 {
6073         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6074         struct btrfs_root *root = BTRFS_I(inode)->root;
6075         struct btrfs_trans_handle *trans;
6076         int ret;
6077
6078         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
6079                 return 0;
6080
6081         trans = btrfs_join_transaction(root);
6082         if (IS_ERR(trans))
6083                 return PTR_ERR(trans);
6084
6085         ret = btrfs_update_inode(trans, root, inode);
6086         if (ret && ret == -ENOSPC) {
6087                 /* whoops, lets try again with the full transaction */
6088                 btrfs_end_transaction(trans);
6089                 trans = btrfs_start_transaction(root, 1);
6090                 if (IS_ERR(trans))
6091                         return PTR_ERR(trans);
6092
6093                 ret = btrfs_update_inode(trans, root, inode);
6094         }
6095         btrfs_end_transaction(trans);
6096         if (BTRFS_I(inode)->delayed_node)
6097                 btrfs_balance_delayed_items(fs_info);
6098
6099         return ret;
6100 }
6101
6102 /*
6103  * This is a copy of file_update_time.  We need this so we can return error on
6104  * ENOSPC for updating the inode in the case of file write and mmap writes.
6105  */
6106 static int btrfs_update_time(struct inode *inode, struct timespec *now,
6107                              int flags)
6108 {
6109         struct btrfs_root *root = BTRFS_I(inode)->root;
6110
6111         if (btrfs_root_readonly(root))
6112                 return -EROFS;
6113
6114         if (flags & S_VERSION)
6115                 inode_inc_iversion(inode);
6116         if (flags & S_CTIME)
6117                 inode->i_ctime = *now;
6118         if (flags & S_MTIME)
6119                 inode->i_mtime = *now;
6120         if (flags & S_ATIME)
6121                 inode->i_atime = *now;
6122         return btrfs_dirty_inode(inode);
6123 }
6124
6125 /*
6126  * find the highest existing sequence number in a directory
6127  * and then set the in-memory index_cnt variable to reflect
6128  * free sequence numbers
6129  */
6130 static int btrfs_set_inode_index_count(struct btrfs_inode *inode)
6131 {
6132         struct btrfs_root *root = inode->root;
6133         struct btrfs_key key, found_key;
6134         struct btrfs_path *path;
6135         struct extent_buffer *leaf;
6136         int ret;
6137
6138         key.objectid = btrfs_ino(inode);
6139         key.type = BTRFS_DIR_INDEX_KEY;
6140         key.offset = (u64)-1;
6141
6142         path = btrfs_alloc_path();
6143         if (!path)
6144                 return -ENOMEM;
6145
6146         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6147         if (ret < 0)
6148                 goto out;
6149         /* FIXME: we should be able to handle this */
6150         if (ret == 0)
6151                 goto out;
6152         ret = 0;
6153
6154         /*
6155          * MAGIC NUMBER EXPLANATION:
6156          * since we search a directory based on f_pos we have to start at 2
6157          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6158          * else has to start at 2
6159          */
6160         if (path->slots[0] == 0) {
6161                 inode->index_cnt = 2;
6162                 goto out;
6163         }
6164
6165         path->slots[0]--;
6166
6167         leaf = path->nodes[0];
6168         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6169
6170         if (found_key.objectid != btrfs_ino(inode) ||
6171             found_key.type != BTRFS_DIR_INDEX_KEY) {
6172                 inode->index_cnt = 2;
6173                 goto out;
6174         }
6175
6176         inode->index_cnt = found_key.offset + 1;
6177 out:
6178         btrfs_free_path(path);
6179         return ret;
6180 }
6181
6182 /*
6183  * helper to find a free sequence number in a given directory.  This current
6184  * code is very simple, later versions will do smarter things in the btree
6185  */
6186 int btrfs_set_inode_index(struct btrfs_inode *dir, u64 *index)
6187 {
6188         int ret = 0;
6189
6190         if (dir->index_cnt == (u64)-1) {
6191                 ret = btrfs_inode_delayed_dir_index_count(dir);
6192                 if (ret) {
6193                         ret = btrfs_set_inode_index_count(dir);
6194                         if (ret)
6195                                 return ret;
6196                 }
6197         }
6198
6199         *index = dir->index_cnt;
6200         dir->index_cnt++;
6201
6202         return ret;
6203 }
6204
6205 static int btrfs_insert_inode_locked(struct inode *inode)
6206 {
6207         struct btrfs_iget_args args;
6208         args.location = &BTRFS_I(inode)->location;
6209         args.root = BTRFS_I(inode)->root;
6210
6211         return insert_inode_locked4(inode,
6212                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6213                    btrfs_find_actor, &args);
6214 }
6215
6216 /*
6217  * Inherit flags from the parent inode.
6218  *
6219  * Currently only the compression flags and the cow flags are inherited.
6220  */
6221 static void btrfs_inherit_iflags(struct inode *inode, struct inode *dir)
6222 {
6223         unsigned int flags;
6224
6225         if (!dir)
6226                 return;
6227
6228         flags = BTRFS_I(dir)->flags;
6229
6230         if (flags & BTRFS_INODE_NOCOMPRESS) {
6231                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_COMPRESS;
6232                 BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
6233         } else if (flags & BTRFS_INODE_COMPRESS) {
6234                 BTRFS_I(inode)->flags &= ~BTRFS_INODE_NOCOMPRESS;
6235                 BTRFS_I(inode)->flags |= BTRFS_INODE_COMPRESS;
6236         }
6237
6238         if (flags & BTRFS_INODE_NODATACOW) {
6239                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW;
6240                 if (S_ISREG(inode->i_mode))
6241                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6242         }
6243
6244         btrfs_update_iflags(inode);
6245 }
6246
6247 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6248                                      struct btrfs_root *root,
6249                                      struct inode *dir,
6250                                      const char *name, int name_len,
6251                                      u64 ref_objectid, u64 objectid,
6252                                      umode_t mode, u64 *index)
6253 {
6254         struct btrfs_fs_info *fs_info = root->fs_info;
6255         struct inode *inode;
6256         struct btrfs_inode_item *inode_item;
6257         struct btrfs_key *location;
6258         struct btrfs_path *path;
6259         struct btrfs_inode_ref *ref;
6260         struct btrfs_key key[2];
6261         u32 sizes[2];
6262         int nitems = name ? 2 : 1;
6263         unsigned long ptr;
6264         int ret;
6265
6266         path = btrfs_alloc_path();
6267         if (!path)
6268                 return ERR_PTR(-ENOMEM);
6269
6270         inode = new_inode(fs_info->sb);
6271         if (!inode) {
6272                 btrfs_free_path(path);
6273                 return ERR_PTR(-ENOMEM);
6274         }
6275
6276         /*
6277          * O_TMPFILE, set link count to 0, so that after this point,
6278          * we fill in an inode item with the correct link count.
6279          */
6280         if (!name)
6281                 set_nlink(inode, 0);
6282
6283         /*
6284          * we have to initialize this early, so we can reclaim the inode
6285          * number if we fail afterwards in this function.
6286          */
6287         inode->i_ino = objectid;
6288
6289         if (dir && name) {
6290                 trace_btrfs_inode_request(dir);
6291
6292                 ret = btrfs_set_inode_index(BTRFS_I(dir), index);
6293                 if (ret) {
6294                         btrfs_free_path(path);
6295                         iput(inode);
6296                         return ERR_PTR(ret);
6297                 }
6298         } else if (dir) {
6299                 *index = 0;
6300         }
6301         /*
6302          * index_cnt is ignored for everything but a dir,
6303          * btrfs_get_inode_index_count has an explanation for the magic
6304          * number
6305          */
6306         BTRFS_I(inode)->index_cnt = 2;
6307         BTRFS_I(inode)->dir_index = *index;
6308         BTRFS_I(inode)->root = root;
6309         BTRFS_I(inode)->generation = trans->transid;
6310         inode->i_generation = BTRFS_I(inode)->generation;
6311
6312         /*
6313          * We could have gotten an inode number from somebody who was fsynced
6314          * and then removed in this same transaction, so let's just set full
6315          * sync since it will be a full sync anyway and this will blow away the
6316          * old info in the log.
6317          */
6318         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6319
6320         key[0].objectid = objectid;
6321         key[0].type = BTRFS_INODE_ITEM_KEY;
6322         key[0].offset = 0;
6323
6324         sizes[0] = sizeof(struct btrfs_inode_item);
6325
6326         if (name) {
6327                 /*
6328                  * Start new inodes with an inode_ref. This is slightly more
6329                  * efficient for small numbers of hard links since they will
6330                  * be packed into one item. Extended refs will kick in if we
6331                  * add more hard links than can fit in the ref item.
6332                  */
6333                 key[1].objectid = objectid;
6334                 key[1].type = BTRFS_INODE_REF_KEY;
6335                 key[1].offset = ref_objectid;
6336
6337                 sizes[1] = name_len + sizeof(*ref);
6338         }
6339
6340         location = &BTRFS_I(inode)->location;
6341         location->objectid = objectid;
6342         location->offset = 0;
6343         location->type = BTRFS_INODE_ITEM_KEY;
6344
6345         ret = btrfs_insert_inode_locked(inode);
6346         if (ret < 0)
6347                 goto fail;
6348
6349         path->leave_spinning = 1;
6350         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6351         if (ret != 0)
6352                 goto fail_unlock;
6353
6354         inode_init_owner(inode, dir, mode);
6355         inode_set_bytes(inode, 0);
6356
6357         inode->i_mtime = current_time(inode);
6358         inode->i_atime = inode->i_mtime;
6359         inode->i_ctime = inode->i_mtime;
6360         BTRFS_I(inode)->i_otime = inode->i_mtime;
6361
6362         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6363                                   struct btrfs_inode_item);
6364         memzero_extent_buffer(path->nodes[0], (unsigned long)inode_item,
6365                              sizeof(*inode_item));
6366         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6367
6368         if (name) {
6369                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6370                                      struct btrfs_inode_ref);
6371                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6372                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6373                 ptr = (unsigned long)(ref + 1);
6374                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6375         }
6376
6377         btrfs_mark_buffer_dirty(path->nodes[0]);
6378         btrfs_free_path(path);
6379
6380         btrfs_inherit_iflags(inode, dir);
6381
6382         if (S_ISREG(mode)) {
6383                 if (btrfs_test_opt(fs_info, NODATASUM))
6384                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6385                 if (btrfs_test_opt(fs_info, NODATACOW))
6386                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6387                                 BTRFS_INODE_NODATASUM;
6388         }
6389
6390         inode_tree_add(inode);
6391
6392         trace_btrfs_inode_new(inode);
6393         btrfs_set_inode_last_trans(trans, inode);
6394
6395         btrfs_update_root_times(trans, root);
6396
6397         ret = btrfs_inode_inherit_props(trans, inode, dir);
6398         if (ret)
6399                 btrfs_err(fs_info,
6400                           "error inheriting props for ino %llu (root %llu): %d",
6401                         btrfs_ino(BTRFS_I(inode)), root->root_key.objectid, ret);
6402
6403         return inode;
6404
6405 fail_unlock:
6406         unlock_new_inode(inode);
6407 fail:
6408         if (dir && name)
6409                 BTRFS_I(dir)->index_cnt--;
6410         btrfs_free_path(path);
6411         iput(inode);
6412         return ERR_PTR(ret);
6413 }
6414
6415 static inline u8 btrfs_inode_type(struct inode *inode)
6416 {
6417         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6418 }
6419
6420 /*
6421  * utility function to add 'inode' into 'parent_inode' with
6422  * a give name and a given sequence number.
6423  * if 'add_backref' is true, also insert a backref from the
6424  * inode to the parent directory.
6425  */
6426 int btrfs_add_link(struct btrfs_trans_handle *trans,
6427                    struct btrfs_inode *parent_inode, struct btrfs_inode *inode,
6428                    const char *name, int name_len, int add_backref, u64 index)
6429 {
6430         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6431         int ret = 0;
6432         struct btrfs_key key;
6433         struct btrfs_root *root = parent_inode->root;
6434         u64 ino = btrfs_ino(inode);
6435         u64 parent_ino = btrfs_ino(parent_inode);
6436
6437         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6438                 memcpy(&key, &inode->root->root_key, sizeof(key));
6439         } else {
6440                 key.objectid = ino;
6441                 key.type = BTRFS_INODE_ITEM_KEY;
6442                 key.offset = 0;
6443         }
6444
6445         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6446                 ret = btrfs_add_root_ref(trans, fs_info, key.objectid,
6447                                          root->root_key.objectid, parent_ino,
6448                                          index, name, name_len);
6449         } else if (add_backref) {
6450                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6451                                              parent_ino, index);
6452         }
6453
6454         /* Nothing to clean up yet */
6455         if (ret)
6456                 return ret;
6457
6458         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6459                                     parent_inode, &key,
6460                                     btrfs_inode_type(&inode->vfs_inode), index);
6461         if (ret == -EEXIST || ret == -EOVERFLOW)
6462                 goto fail_dir_item;
6463         else if (ret) {
6464                 btrfs_abort_transaction(trans, ret);
6465                 return ret;
6466         }
6467
6468         btrfs_i_size_write(parent_inode, parent_inode->vfs_inode.i_size +
6469                            name_len * 2);
6470         inode_inc_iversion(&parent_inode->vfs_inode);
6471         parent_inode->vfs_inode.i_mtime = parent_inode->vfs_inode.i_ctime =
6472                 current_time(&parent_inode->vfs_inode);
6473         ret = btrfs_update_inode(trans, root, &parent_inode->vfs_inode);
6474         if (ret)
6475                 btrfs_abort_transaction(trans, ret);
6476         return ret;
6477
6478 fail_dir_item:
6479         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6480                 u64 local_index;
6481                 int err;
6482                 err = btrfs_del_root_ref(trans, fs_info, key.objectid,
6483                                          root->root_key.objectid, parent_ino,
6484                                          &local_index, name, name_len);
6485
6486         } else if (add_backref) {
6487                 u64 local_index;
6488                 int err;
6489
6490                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6491                                           ino, parent_ino, &local_index);
6492         }
6493         return ret;
6494 }
6495
6496 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6497                             struct btrfs_inode *dir, struct dentry *dentry,
6498                             struct btrfs_inode *inode, int backref, u64 index)
6499 {
6500         int err = btrfs_add_link(trans, dir, inode,
6501                                  dentry->d_name.name, dentry->d_name.len,
6502                                  backref, index);
6503         if (err > 0)
6504                 err = -EEXIST;
6505         return err;
6506 }
6507
6508 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6509                         umode_t mode, dev_t rdev)
6510 {
6511         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6512         struct btrfs_trans_handle *trans;
6513         struct btrfs_root *root = BTRFS_I(dir)->root;
6514         struct inode *inode = NULL;
6515         int err;
6516         int drop_inode = 0;
6517         u64 objectid;
6518         u64 index = 0;
6519
6520         /*
6521          * 2 for inode item and ref
6522          * 2 for dir items
6523          * 1 for xattr if selinux is on
6524          */
6525         trans = btrfs_start_transaction(root, 5);
6526         if (IS_ERR(trans))
6527                 return PTR_ERR(trans);
6528
6529         err = btrfs_find_free_ino(root, &objectid);
6530         if (err)
6531                 goto out_unlock;
6532
6533         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6534                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6535                         mode, &index);
6536         if (IS_ERR(inode)) {
6537                 err = PTR_ERR(inode);
6538                 goto out_unlock;
6539         }
6540
6541         /*
6542         * If the active LSM wants to access the inode during
6543         * d_instantiate it needs these. Smack checks to see
6544         * if the filesystem supports xattrs by looking at the
6545         * ops vector.
6546         */
6547         inode->i_op = &btrfs_special_inode_operations;
6548         init_special_inode(inode, inode->i_mode, rdev);
6549
6550         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6551         if (err)
6552                 goto out_unlock_inode;
6553
6554         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6555                         0, index);
6556         if (err) {
6557                 goto out_unlock_inode;
6558         } else {
6559                 btrfs_update_inode(trans, root, inode);
6560                 unlock_new_inode(inode);
6561                 d_instantiate(dentry, inode);
6562         }
6563
6564 out_unlock:
6565         btrfs_end_transaction(trans);
6566         btrfs_balance_delayed_items(fs_info);
6567         btrfs_btree_balance_dirty(fs_info);
6568         if (drop_inode) {
6569                 inode_dec_link_count(inode);
6570                 iput(inode);
6571         }
6572         return err;
6573
6574 out_unlock_inode:
6575         drop_inode = 1;
6576         unlock_new_inode(inode);
6577         goto out_unlock;
6578
6579 }
6580
6581 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6582                         umode_t mode, bool excl)
6583 {
6584         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6585         struct btrfs_trans_handle *trans;
6586         struct btrfs_root *root = BTRFS_I(dir)->root;
6587         struct inode *inode = NULL;
6588         int drop_inode_on_err = 0;
6589         int err;
6590         u64 objectid;
6591         u64 index = 0;
6592
6593         /*
6594          * 2 for inode item and ref
6595          * 2 for dir items
6596          * 1 for xattr if selinux is on
6597          */
6598         trans = btrfs_start_transaction(root, 5);
6599         if (IS_ERR(trans))
6600                 return PTR_ERR(trans);
6601
6602         err = btrfs_find_free_ino(root, &objectid);
6603         if (err)
6604                 goto out_unlock;
6605
6606         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6607                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6608                         mode, &index);
6609         if (IS_ERR(inode)) {
6610                 err = PTR_ERR(inode);
6611                 goto out_unlock;
6612         }
6613         drop_inode_on_err = 1;
6614         /*
6615         * If the active LSM wants to access the inode during
6616         * d_instantiate it needs these. Smack checks to see
6617         * if the filesystem supports xattrs by looking at the
6618         * ops vector.
6619         */
6620         inode->i_fop = &btrfs_file_operations;
6621         inode->i_op = &btrfs_file_inode_operations;
6622         inode->i_mapping->a_ops = &btrfs_aops;
6623
6624         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6625         if (err)
6626                 goto out_unlock_inode;
6627
6628         err = btrfs_update_inode(trans, root, inode);
6629         if (err)
6630                 goto out_unlock_inode;
6631
6632         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6633                         0, index);
6634         if (err)
6635                 goto out_unlock_inode;
6636
6637         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6638         unlock_new_inode(inode);
6639         d_instantiate(dentry, inode);
6640
6641 out_unlock:
6642         btrfs_end_transaction(trans);
6643         if (err && drop_inode_on_err) {
6644                 inode_dec_link_count(inode);
6645                 iput(inode);
6646         }
6647         btrfs_balance_delayed_items(fs_info);
6648         btrfs_btree_balance_dirty(fs_info);
6649         return err;
6650
6651 out_unlock_inode:
6652         unlock_new_inode(inode);
6653         goto out_unlock;
6654
6655 }
6656
6657 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6658                       struct dentry *dentry)
6659 {
6660         struct btrfs_trans_handle *trans = NULL;
6661         struct btrfs_root *root = BTRFS_I(dir)->root;
6662         struct inode *inode = d_inode(old_dentry);
6663         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
6664         u64 index;
6665         int err;
6666         int drop_inode = 0;
6667
6668         /* do not allow sys_link's with other subvols of the same device */
6669         if (root->objectid != BTRFS_I(inode)->root->objectid)
6670                 return -EXDEV;
6671
6672         if (inode->i_nlink >= BTRFS_LINK_MAX)
6673                 return -EMLINK;
6674
6675         err = btrfs_set_inode_index(BTRFS_I(dir), &index);
6676         if (err)
6677                 goto fail;
6678
6679         /*
6680          * 2 items for inode and inode ref
6681          * 2 items for dir items
6682          * 1 item for parent inode
6683          */
6684         trans = btrfs_start_transaction(root, 5);
6685         if (IS_ERR(trans)) {
6686                 err = PTR_ERR(trans);
6687                 trans = NULL;
6688                 goto fail;
6689         }
6690
6691         /* There are several dir indexes for this inode, clear the cache. */
6692         BTRFS_I(inode)->dir_index = 0ULL;
6693         inc_nlink(inode);
6694         inode_inc_iversion(inode);
6695         inode->i_ctime = current_time(inode);
6696         ihold(inode);
6697         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6698
6699         err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry, BTRFS_I(inode),
6700                         1, index);
6701
6702         if (err) {
6703                 drop_inode = 1;
6704         } else {
6705                 struct dentry *parent = dentry->d_parent;
6706                 err = btrfs_update_inode(trans, root, inode);
6707                 if (err)
6708                         goto fail;
6709                 if (inode->i_nlink == 1) {
6710                         /*
6711                          * If new hard link count is 1, it's a file created
6712                          * with open(2) O_TMPFILE flag.
6713                          */
6714                         err = btrfs_orphan_del(trans, BTRFS_I(inode));
6715                         if (err)
6716                                 goto fail;
6717                 }
6718                 d_instantiate(dentry, inode);
6719                 btrfs_log_new_name(trans, BTRFS_I(inode), NULL, parent);
6720         }
6721
6722         btrfs_balance_delayed_items(fs_info);
6723 fail:
6724         if (trans)
6725                 btrfs_end_transaction(trans);
6726         if (drop_inode) {
6727                 inode_dec_link_count(inode);
6728                 iput(inode);
6729         }
6730         btrfs_btree_balance_dirty(fs_info);
6731         return err;
6732 }
6733
6734 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6735 {
6736         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
6737         struct inode *inode = NULL;
6738         struct btrfs_trans_handle *trans;
6739         struct btrfs_root *root = BTRFS_I(dir)->root;
6740         int err = 0;
6741         int drop_on_err = 0;
6742         u64 objectid = 0;
6743         u64 index = 0;
6744
6745         /*
6746          * 2 items for inode and ref
6747          * 2 items for dir items
6748          * 1 for xattr if selinux is on
6749          */
6750         trans = btrfs_start_transaction(root, 5);
6751         if (IS_ERR(trans))
6752                 return PTR_ERR(trans);
6753
6754         err = btrfs_find_free_ino(root, &objectid);
6755         if (err)
6756                 goto out_fail;
6757
6758         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6759                         dentry->d_name.len, btrfs_ino(BTRFS_I(dir)), objectid,
6760                         S_IFDIR | mode, &index);
6761         if (IS_ERR(inode)) {
6762                 err = PTR_ERR(inode);
6763                 goto out_fail;
6764         }
6765
6766         drop_on_err = 1;
6767         /* these must be set before we unlock the inode */
6768         inode->i_op = &btrfs_dir_inode_operations;
6769         inode->i_fop = &btrfs_dir_file_operations;
6770
6771         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6772         if (err)
6773                 goto out_fail_inode;
6774
6775         btrfs_i_size_write(BTRFS_I(inode), 0);
6776         err = btrfs_update_inode(trans, root, inode);
6777         if (err)
6778                 goto out_fail_inode;
6779
6780         err = btrfs_add_link(trans, BTRFS_I(dir), BTRFS_I(inode),
6781                         dentry->d_name.name,
6782                         dentry->d_name.len, 0, index);
6783         if (err)
6784                 goto out_fail_inode;
6785
6786         d_instantiate(dentry, inode);
6787         /*
6788          * mkdir is special.  We're unlocking after we call d_instantiate
6789          * to avoid a race with nfsd calling d_instantiate.
6790          */
6791         unlock_new_inode(inode);
6792         drop_on_err = 0;
6793
6794 out_fail:
6795         btrfs_end_transaction(trans);
6796         if (drop_on_err) {
6797                 inode_dec_link_count(inode);
6798                 iput(inode);
6799         }
6800         btrfs_balance_delayed_items(fs_info);
6801         btrfs_btree_balance_dirty(fs_info);
6802         return err;
6803
6804 out_fail_inode:
6805         unlock_new_inode(inode);
6806         goto out_fail;
6807 }
6808
6809 /* Find next extent map of a given extent map, caller needs to ensure locks */
6810 static struct extent_map *next_extent_map(struct extent_map *em)
6811 {
6812         struct rb_node *next;
6813
6814         next = rb_next(&em->rb_node);
6815         if (!next)
6816                 return NULL;
6817         return container_of(next, struct extent_map, rb_node);
6818 }
6819
6820 static struct extent_map *prev_extent_map(struct extent_map *em)
6821 {
6822         struct rb_node *prev;
6823
6824         prev = rb_prev(&em->rb_node);
6825         if (!prev)
6826                 return NULL;
6827         return container_of(prev, struct extent_map, rb_node);
6828 }
6829
6830 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6831  * the existing extent is the nearest extent to map_start,
6832  * and an extent that you want to insert, deal with overlap and insert
6833  * the best fitted new extent into the tree.
6834  */
6835 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6836                                 struct extent_map *existing,
6837                                 struct extent_map *em,
6838                                 u64 map_start)
6839 {
6840         struct extent_map *prev;
6841         struct extent_map *next;
6842         u64 start;
6843         u64 end;
6844         u64 start_diff;
6845
6846         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6847
6848         if (existing->start > map_start) {
6849                 next = existing;
6850                 prev = prev_extent_map(next);
6851         } else {
6852                 prev = existing;
6853                 next = next_extent_map(prev);
6854         }
6855
6856         start = prev ? extent_map_end(prev) : em->start;
6857         start = max_t(u64, start, em->start);
6858         end = next ? next->start : extent_map_end(em);
6859         end = min_t(u64, end, extent_map_end(em));
6860         start_diff = start - em->start;
6861         em->start = start;
6862         em->len = end - start;
6863         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6864             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6865                 em->block_start += start_diff;
6866                 em->block_len -= start_diff;
6867         }
6868         return add_extent_mapping(em_tree, em, 0);
6869 }
6870
6871 static noinline int uncompress_inline(struct btrfs_path *path,
6872                                       struct page *page,
6873                                       size_t pg_offset, u64 extent_offset,
6874                                       struct btrfs_file_extent_item *item)
6875 {
6876         int ret;
6877         struct extent_buffer *leaf = path->nodes[0];
6878         char *tmp;
6879         size_t max_size;
6880         unsigned long inline_size;
6881         unsigned long ptr;
6882         int compress_type;
6883
6884         WARN_ON(pg_offset != 0);
6885         compress_type = btrfs_file_extent_compression(leaf, item);
6886         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6887         inline_size = btrfs_file_extent_inline_item_len(leaf,
6888                                         btrfs_item_nr(path->slots[0]));
6889         tmp = kmalloc(inline_size, GFP_NOFS);
6890         if (!tmp)
6891                 return -ENOMEM;
6892         ptr = btrfs_file_extent_inline_start(item);
6893
6894         read_extent_buffer(leaf, tmp, ptr, inline_size);
6895
6896         max_size = min_t(unsigned long, PAGE_SIZE, max_size);
6897         ret = btrfs_decompress(compress_type, tmp, page,
6898                                extent_offset, inline_size, max_size);
6899
6900         /*
6901          * decompression code contains a memset to fill in any space between the end
6902          * of the uncompressed data and the end of max_size in case the decompressed
6903          * data ends up shorter than ram_bytes.  That doesn't cover the hole between
6904          * the end of an inline extent and the beginning of the next block, so we
6905          * cover that region here.
6906          */
6907
6908         if (max_size + pg_offset < PAGE_SIZE) {
6909                 char *map = kmap(page);
6910                 memset(map + pg_offset + max_size, 0, PAGE_SIZE - max_size - pg_offset);
6911                 kunmap(page);
6912         }
6913         kfree(tmp);
6914         return ret;
6915 }
6916
6917 /*
6918  * a bit scary, this does extent mapping from logical file offset to the disk.
6919  * the ugly parts come from merging extents from the disk with the in-ram
6920  * representation.  This gets more complex because of the data=ordered code,
6921  * where the in-ram extents might be locked pending data=ordered completion.
6922  *
6923  * This also copies inline extents directly into the page.
6924  */
6925 struct extent_map *btrfs_get_extent(struct btrfs_inode *inode,
6926                 struct page *page,
6927             size_t pg_offset, u64 start, u64 len,
6928                 int create)
6929 {
6930         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
6931         int ret;
6932         int err = 0;
6933         u64 extent_start = 0;
6934         u64 extent_end = 0;
6935         u64 objectid = btrfs_ino(inode);
6936         u32 found_type;
6937         struct btrfs_path *path = NULL;
6938         struct btrfs_root *root = inode->root;
6939         struct btrfs_file_extent_item *item;
6940         struct extent_buffer *leaf;
6941         struct btrfs_key found_key;
6942         struct extent_map *em = NULL;
6943         struct extent_map_tree *em_tree = &inode->extent_tree;
6944         struct extent_io_tree *io_tree = &inode->io_tree;
6945         struct btrfs_trans_handle *trans = NULL;
6946         const bool new_inline = !page || create;
6947
6948 again:
6949         read_lock(&em_tree->lock);
6950         em = lookup_extent_mapping(em_tree, start, len);
6951         if (em)
6952                 em->bdev = fs_info->fs_devices->latest_bdev;
6953         read_unlock(&em_tree->lock);
6954
6955         if (em) {
6956                 if (em->start > start || em->start + em->len <= start)
6957                         free_extent_map(em);
6958                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6959                         free_extent_map(em);
6960                 else
6961                         goto out;
6962         }
6963         em = alloc_extent_map();
6964         if (!em) {
6965                 err = -ENOMEM;
6966                 goto out;
6967         }
6968         em->bdev = fs_info->fs_devices->latest_bdev;
6969         em->start = EXTENT_MAP_HOLE;
6970         em->orig_start = EXTENT_MAP_HOLE;
6971         em->len = (u64)-1;
6972         em->block_len = (u64)-1;
6973
6974         if (!path) {
6975                 path = btrfs_alloc_path();
6976                 if (!path) {
6977                         err = -ENOMEM;
6978                         goto out;
6979                 }
6980                 /*
6981                  * Chances are we'll be called again, so go ahead and do
6982                  * readahead
6983                  */
6984                 path->reada = READA_FORWARD;
6985         }
6986
6987         ret = btrfs_lookup_file_extent(trans, root, path,
6988                                        objectid, start, trans != NULL);
6989         if (ret < 0) {
6990                 err = ret;
6991                 goto out;
6992         }
6993
6994         if (ret != 0) {
6995                 if (path->slots[0] == 0)
6996                         goto not_found;
6997                 path->slots[0]--;
6998         }
6999
7000         leaf = path->nodes[0];
7001         item = btrfs_item_ptr(leaf, path->slots[0],
7002                               struct btrfs_file_extent_item);
7003         /* are we inside the extent that was found? */
7004         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7005         found_type = found_key.type;
7006         if (found_key.objectid != objectid ||
7007             found_type != BTRFS_EXTENT_DATA_KEY) {
7008                 /*
7009                  * If we backup past the first extent we want to move forward
7010                  * and see if there is an extent in front of us, otherwise we'll
7011                  * say there is a hole for our whole search range which can
7012                  * cause problems.
7013                  */
7014                 extent_end = start;
7015                 goto next;
7016         }
7017
7018         found_type = btrfs_file_extent_type(leaf, item);
7019         extent_start = found_key.offset;
7020         if (found_type == BTRFS_FILE_EXTENT_REG ||
7021             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7022                 extent_end = extent_start +
7023                        btrfs_file_extent_num_bytes(leaf, item);
7024
7025                 trace_btrfs_get_extent_show_fi_regular(inode, leaf, item,
7026                                                        extent_start);
7027         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7028                 size_t size;
7029                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7030                 extent_end = ALIGN(extent_start + size,
7031                                    fs_info->sectorsize);
7032
7033                 trace_btrfs_get_extent_show_fi_inline(inode, leaf, item,
7034                                                       path->slots[0],
7035                                                       extent_start);
7036         }
7037 next:
7038         if (start >= extent_end) {
7039                 path->slots[0]++;
7040                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
7041                         ret = btrfs_next_leaf(root, path);
7042                         if (ret < 0) {
7043                                 err = ret;
7044                                 goto out;
7045                         }
7046                         if (ret > 0)
7047                                 goto not_found;
7048                         leaf = path->nodes[0];
7049                 }
7050                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7051                 if (found_key.objectid != objectid ||
7052                     found_key.type != BTRFS_EXTENT_DATA_KEY)
7053                         goto not_found;
7054                 if (start + len <= found_key.offset)
7055                         goto not_found;
7056                 if (start > found_key.offset)
7057                         goto next;
7058                 em->start = start;
7059                 em->orig_start = start;
7060                 em->len = found_key.offset - start;
7061                 goto not_found_em;
7062         }
7063
7064         btrfs_extent_item_to_extent_map(inode, path, item,
7065                         new_inline, em);
7066
7067         if (found_type == BTRFS_FILE_EXTENT_REG ||
7068             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7069                 goto insert;
7070         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
7071                 unsigned long ptr;
7072                 char *map;
7073                 size_t size;
7074                 size_t extent_offset;
7075                 size_t copy_size;
7076
7077                 if (new_inline)
7078                         goto out;
7079
7080                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
7081                 extent_offset = page_offset(page) + pg_offset - extent_start;
7082                 copy_size = min_t(u64, PAGE_SIZE - pg_offset,
7083                                   size - extent_offset);
7084                 em->start = extent_start + extent_offset;
7085                 em->len = ALIGN(copy_size, fs_info->sectorsize);
7086                 em->orig_block_len = em->len;
7087                 em->orig_start = em->start;
7088                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
7089                 if (create == 0 && !PageUptodate(page)) {
7090                         if (btrfs_file_extent_compression(leaf, item) !=
7091                             BTRFS_COMPRESS_NONE) {
7092                                 ret = uncompress_inline(path, page, pg_offset,
7093                                                         extent_offset, item);
7094                                 if (ret) {
7095                                         err = ret;
7096                                         goto out;
7097                                 }
7098                         } else {
7099                                 map = kmap(page);
7100                                 read_extent_buffer(leaf, map + pg_offset, ptr,
7101                                                    copy_size);
7102                                 if (pg_offset + copy_size < PAGE_SIZE) {
7103                                         memset(map + pg_offset + copy_size, 0,
7104                                                PAGE_SIZE - pg_offset -
7105                                                copy_size);
7106                                 }
7107                                 kunmap(page);
7108                         }
7109                         flush_dcache_page(page);
7110                 } else if (create && PageUptodate(page)) {
7111                         BUG();
7112                         if (!trans) {
7113                                 kunmap(page);
7114                                 free_extent_map(em);
7115                                 em = NULL;
7116
7117                                 btrfs_release_path(path);
7118                                 trans = btrfs_join_transaction(root);
7119
7120                                 if (IS_ERR(trans))
7121                                         return ERR_CAST(trans);
7122                                 goto again;
7123                         }
7124                         map = kmap(page);
7125                         write_extent_buffer(leaf, map + pg_offset, ptr,
7126                                             copy_size);
7127                         kunmap(page);
7128                         btrfs_mark_buffer_dirty(leaf);
7129                 }
7130                 set_extent_uptodate(io_tree, em->start,
7131                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
7132                 goto insert;
7133         }
7134 not_found:
7135         em->start = start;
7136         em->orig_start = start;
7137         em->len = len;
7138 not_found_em:
7139         em->block_start = EXTENT_MAP_HOLE;
7140         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
7141 insert:
7142         btrfs_release_path(path);
7143         if (em->start > start || extent_map_end(em) <= start) {
7144                 btrfs_err(fs_info,
7145                           "bad extent! em: [%llu %llu] passed [%llu %llu]",
7146                           em->start, em->len, start, len);
7147                 err = -EIO;
7148                 goto out;
7149         }
7150
7151         err = 0;
7152         write_lock(&em_tree->lock);
7153         ret = add_extent_mapping(em_tree, em, 0);
7154         /* it is possible that someone inserted the extent into the tree
7155          * while we had the lock dropped.  It is also possible that
7156          * an overlapping map exists in the tree
7157          */
7158         if (ret == -EEXIST) {
7159                 struct extent_map *existing;
7160
7161                 ret = 0;
7162
7163                 existing = search_extent_mapping(em_tree, start, len);
7164                 /*
7165                  * existing will always be non-NULL, since there must be
7166                  * extent causing the -EEXIST.
7167                  */
7168                 if (existing->start == em->start &&
7169                     extent_map_end(existing) >= extent_map_end(em) &&
7170                     em->block_start == existing->block_start) {
7171                         /*
7172                          * The existing extent map already encompasses the
7173                          * entire extent map we tried to add.
7174                          */
7175                         free_extent_map(em);
7176                         em = existing;
7177                         err = 0;
7178
7179                 } else if (start >= extent_map_end(existing) ||
7180                     start <= existing->start) {
7181                         /*
7182                          * The existing extent map is the one nearest to
7183                          * the [start, start + len) range which overlaps
7184                          */
7185                         err = merge_extent_mapping(em_tree, existing,
7186                                                    em, start);
7187                         free_extent_map(existing);
7188                         if (err) {
7189                                 free_extent_map(em);
7190                                 em = NULL;
7191                         }
7192                 } else {
7193                         free_extent_map(em);
7194                         em = existing;
7195                         err = 0;
7196                 }
7197         }
7198         write_unlock(&em_tree->lock);
7199 out:
7200
7201         trace_btrfs_get_extent(root, inode, em);
7202
7203         btrfs_free_path(path);
7204         if (trans) {
7205                 ret = btrfs_end_transaction(trans);
7206                 if (!err)
7207                         err = ret;
7208         }
7209         if (err) {
7210                 free_extent_map(em);
7211                 return ERR_PTR(err);
7212         }
7213         BUG_ON(!em); /* Error is always set */
7214         return em;
7215 }
7216
7217 struct extent_map *btrfs_get_extent_fiemap(struct btrfs_inode *inode,
7218                 struct page *page,
7219                 size_t pg_offset, u64 start, u64 len,
7220                 int create)
7221 {
7222         struct extent_map *em;
7223         struct extent_map *hole_em = NULL;
7224         u64 range_start = start;
7225         u64 end;
7226         u64 found;
7227         u64 found_end;
7228         int err = 0;
7229
7230         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7231         if (IS_ERR(em))
7232                 return em;
7233         /*
7234          * If our em maps to:
7235          * - a hole or
7236          * - a pre-alloc extent,
7237          * there might actually be delalloc bytes behind it.
7238          */
7239         if (em->block_start != EXTENT_MAP_HOLE &&
7240             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7241                 return em;
7242         else
7243                 hole_em = em;
7244
7245         /* check to see if we've wrapped (len == -1 or similar) */
7246         end = start + len;
7247         if (end < start)
7248                 end = (u64)-1;
7249         else
7250                 end -= 1;
7251
7252         em = NULL;
7253
7254         /* ok, we didn't find anything, lets look for delalloc */
7255         found = count_range_bits(&inode->io_tree, &range_start,
7256                                  end, len, EXTENT_DELALLOC, 1);
7257         found_end = range_start + found;
7258         if (found_end < range_start)
7259                 found_end = (u64)-1;
7260
7261         /*
7262          * we didn't find anything useful, return
7263          * the original results from get_extent()
7264          */
7265         if (range_start > end || found_end <= start) {
7266                 em = hole_em;
7267                 hole_em = NULL;
7268                 goto out;
7269         }
7270
7271         /* adjust the range_start to make sure it doesn't
7272          * go backwards from the start they passed in
7273          */
7274         range_start = max(start, range_start);
7275         found = found_end - range_start;
7276
7277         if (found > 0) {
7278                 u64 hole_start = start;
7279                 u64 hole_len = len;
7280
7281                 em = alloc_extent_map();
7282                 if (!em) {
7283                         err = -ENOMEM;
7284                         goto out;
7285                 }
7286                 /*
7287                  * when btrfs_get_extent can't find anything it
7288                  * returns one huge hole
7289                  *
7290                  * make sure what it found really fits our range, and
7291                  * adjust to make sure it is based on the start from
7292                  * the caller
7293                  */
7294                 if (hole_em) {
7295                         u64 calc_end = extent_map_end(hole_em);
7296
7297                         if (calc_end <= start || (hole_em->start > end)) {
7298                                 free_extent_map(hole_em);
7299                                 hole_em = NULL;
7300                         } else {
7301                                 hole_start = max(hole_em->start, start);
7302                                 hole_len = calc_end - hole_start;
7303                         }
7304                 }
7305                 em->bdev = NULL;
7306                 if (hole_em && range_start > hole_start) {
7307                         /* our hole starts before our delalloc, so we
7308                          * have to return just the parts of the hole
7309                          * that go until  the delalloc starts
7310                          */
7311                         em->len = min(hole_len,
7312                                       range_start - hole_start);
7313                         em->start = hole_start;
7314                         em->orig_start = hole_start;
7315                         /*
7316                          * don't adjust block start at all,
7317                          * it is fixed at EXTENT_MAP_HOLE
7318                          */
7319                         em->block_start = hole_em->block_start;
7320                         em->block_len = hole_len;
7321                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7322                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7323                 } else {
7324                         em->start = range_start;
7325                         em->len = found;
7326                         em->orig_start = range_start;
7327                         em->block_start = EXTENT_MAP_DELALLOC;
7328                         em->block_len = found;
7329                 }
7330         } else if (hole_em) {
7331                 return hole_em;
7332         }
7333 out:
7334
7335         free_extent_map(hole_em);
7336         if (err) {
7337                 free_extent_map(em);
7338                 return ERR_PTR(err);
7339         }
7340         return em;
7341 }
7342
7343 static struct extent_map *btrfs_create_dio_extent(struct inode *inode,
7344                                                   const u64 start,
7345                                                   const u64 len,
7346                                                   const u64 orig_start,
7347                                                   const u64 block_start,
7348                                                   const u64 block_len,
7349                                                   const u64 orig_block_len,
7350                                                   const u64 ram_bytes,
7351                                                   const int type)
7352 {
7353         struct extent_map *em = NULL;
7354         int ret;
7355
7356         if (type != BTRFS_ORDERED_NOCOW) {
7357                 em = create_io_em(inode, start, len, orig_start,
7358                                   block_start, block_len, orig_block_len,
7359                                   ram_bytes,
7360                                   BTRFS_COMPRESS_NONE, /* compress_type */
7361                                   type);
7362                 if (IS_ERR(em))
7363                         goto out;
7364         }
7365         ret = btrfs_add_ordered_extent_dio(inode, start, block_start,
7366                                            len, block_len, type);
7367         if (ret) {
7368                 if (em) {
7369                         free_extent_map(em);
7370                         btrfs_drop_extent_cache(BTRFS_I(inode), start,
7371                                                 start + len - 1, 0);
7372                 }
7373                 em = ERR_PTR(ret);
7374         }
7375  out:
7376
7377         return em;
7378 }
7379
7380 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7381                                                   u64 start, u64 len)
7382 {
7383         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7384         struct btrfs_root *root = BTRFS_I(inode)->root;
7385         struct extent_map *em;
7386         struct btrfs_key ins;
7387         u64 alloc_hint;
7388         int ret;
7389
7390         alloc_hint = get_extent_allocation_hint(inode, start, len);
7391         ret = btrfs_reserve_extent(root, len, len, fs_info->sectorsize,
7392                                    0, alloc_hint, &ins, 1, 1);
7393         if (ret)
7394                 return ERR_PTR(ret);
7395
7396         em = btrfs_create_dio_extent(inode, start, ins.offset, start,
7397                                      ins.objectid, ins.offset, ins.offset,
7398                                      ins.offset, BTRFS_ORDERED_REGULAR);
7399         btrfs_dec_block_group_reservations(fs_info, ins.objectid);
7400         if (IS_ERR(em))
7401                 btrfs_free_reserved_extent(fs_info, ins.objectid,
7402                                            ins.offset, 1);
7403
7404         return em;
7405 }
7406
7407 /*
7408  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7409  * block must be cow'd
7410  */
7411 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7412                               u64 *orig_start, u64 *orig_block_len,
7413                               u64 *ram_bytes)
7414 {
7415         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7416         struct btrfs_path *path;
7417         int ret;
7418         struct extent_buffer *leaf;
7419         struct btrfs_root *root = BTRFS_I(inode)->root;
7420         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7421         struct btrfs_file_extent_item *fi;
7422         struct btrfs_key key;
7423         u64 disk_bytenr;
7424         u64 backref_offset;
7425         u64 extent_end;
7426         u64 num_bytes;
7427         int slot;
7428         int found_type;
7429         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7430
7431         path = btrfs_alloc_path();
7432         if (!path)
7433                 return -ENOMEM;
7434
7435         ret = btrfs_lookup_file_extent(NULL, root, path,
7436                         btrfs_ino(BTRFS_I(inode)), offset, 0);
7437         if (ret < 0)
7438                 goto out;
7439
7440         slot = path->slots[0];
7441         if (ret == 1) {
7442                 if (slot == 0) {
7443                         /* can't find the item, must cow */
7444                         ret = 0;
7445                         goto out;
7446                 }
7447                 slot--;
7448         }
7449         ret = 0;
7450         leaf = path->nodes[0];
7451         btrfs_item_key_to_cpu(leaf, &key, slot);
7452         if (key.objectid != btrfs_ino(BTRFS_I(inode)) ||
7453             key.type != BTRFS_EXTENT_DATA_KEY) {
7454                 /* not our file or wrong item type, must cow */
7455                 goto out;
7456         }
7457
7458         if (key.offset > offset) {
7459                 /* Wrong offset, must cow */
7460                 goto out;
7461         }
7462
7463         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7464         found_type = btrfs_file_extent_type(leaf, fi);
7465         if (found_type != BTRFS_FILE_EXTENT_REG &&
7466             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7467                 /* not a regular extent, must cow */
7468                 goto out;
7469         }
7470
7471         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7472                 goto out;
7473
7474         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7475         if (extent_end <= offset)
7476                 goto out;
7477
7478         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7479         if (disk_bytenr == 0)
7480                 goto out;
7481
7482         if (btrfs_file_extent_compression(leaf, fi) ||
7483             btrfs_file_extent_encryption(leaf, fi) ||
7484             btrfs_file_extent_other_encoding(leaf, fi))
7485                 goto out;
7486
7487         backref_offset = btrfs_file_extent_offset(leaf, fi);
7488
7489         if (orig_start) {
7490                 *orig_start = key.offset - backref_offset;
7491                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7492                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7493         }
7494
7495         if (btrfs_extent_readonly(fs_info, disk_bytenr))
7496                 goto out;
7497
7498         num_bytes = min(offset + *len, extent_end) - offset;
7499         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7500                 u64 range_end;
7501
7502                 range_end = round_up(offset + num_bytes,
7503                                      root->fs_info->sectorsize) - 1;
7504                 ret = test_range_bit(io_tree, offset, range_end,
7505                                      EXTENT_DELALLOC, 0, NULL);
7506                 if (ret) {
7507                         ret = -EAGAIN;
7508                         goto out;
7509                 }
7510         }
7511
7512         btrfs_release_path(path);
7513
7514         /*
7515          * look for other files referencing this extent, if we
7516          * find any we must cow
7517          */
7518
7519         ret = btrfs_cross_ref_exist(root, btrfs_ino(BTRFS_I(inode)),
7520                                     key.offset - backref_offset, disk_bytenr);
7521         if (ret) {
7522                 ret = 0;
7523                 goto out;
7524         }
7525
7526         /*
7527          * adjust disk_bytenr and num_bytes to cover just the bytes
7528          * in this extent we are about to write.  If there
7529          * are any csums in that range we have to cow in order
7530          * to keep the csums correct
7531          */
7532         disk_bytenr += backref_offset;
7533         disk_bytenr += offset - key.offset;
7534         if (csum_exist_in_range(fs_info, disk_bytenr, num_bytes))
7535                 goto out;
7536         /*
7537          * all of the above have passed, it is safe to overwrite this extent
7538          * without cow
7539          */
7540         *len = num_bytes;
7541         ret = 1;
7542 out:
7543         btrfs_free_path(path);
7544         return ret;
7545 }
7546
7547 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7548 {
7549         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7550         bool found = false;
7551         void **pagep = NULL;
7552         struct page *page = NULL;
7553         unsigned long start_idx;
7554         unsigned long end_idx;
7555
7556         start_idx = start >> PAGE_SHIFT;
7557
7558         /*
7559          * end is the last byte in the last page.  end == start is legal
7560          */
7561         end_idx = end >> PAGE_SHIFT;
7562
7563         rcu_read_lock();
7564
7565         /* Most of the code in this while loop is lifted from
7566          * find_get_page.  It's been modified to begin searching from a
7567          * page and return just the first page found in that range.  If the
7568          * found idx is less than or equal to the end idx then we know that
7569          * a page exists.  If no pages are found or if those pages are
7570          * outside of the range then we're fine (yay!) */
7571         while (page == NULL &&
7572                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7573                 page = radix_tree_deref_slot(pagep);
7574                 if (unlikely(!page))
7575                         break;
7576
7577                 if (radix_tree_exception(page)) {
7578                         if (radix_tree_deref_retry(page)) {
7579                                 page = NULL;
7580                                 continue;
7581                         }
7582                         /*
7583                          * Otherwise, shmem/tmpfs must be storing a swap entry
7584                          * here as an exceptional entry: so return it without
7585                          * attempting to raise page count.
7586                          */
7587                         page = NULL;
7588                         break; /* TODO: Is this relevant for this use case? */
7589                 }
7590
7591                 if (!page_cache_get_speculative(page)) {
7592                         page = NULL;
7593                         continue;
7594                 }
7595
7596                 /*
7597                  * Has the page moved?
7598                  * This is part of the lockless pagecache protocol. See
7599                  * include/linux/pagemap.h for details.
7600                  */
7601                 if (unlikely(page != *pagep)) {
7602                         put_page(page);
7603                         page = NULL;
7604                 }
7605         }
7606
7607         if (page) {
7608                 if (page->index <= end_idx)
7609                         found = true;
7610                 put_page(page);
7611         }
7612
7613         rcu_read_unlock();
7614         return found;
7615 }
7616
7617 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7618                               struct extent_state **cached_state, int writing)
7619 {
7620         struct btrfs_ordered_extent *ordered;
7621         int ret = 0;
7622
7623         while (1) {
7624                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7625                                  cached_state);
7626                 /*
7627                  * We're concerned with the entire range that we're going to be
7628                  * doing DIO to, so we need to make sure there's no ordered
7629                  * extents in this range.
7630                  */
7631                 ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), lockstart,
7632                                                      lockend - lockstart + 1);
7633
7634                 /*
7635                  * We need to make sure there are no buffered pages in this
7636                  * range either, we could have raced between the invalidate in
7637                  * generic_file_direct_write and locking the extent.  The
7638                  * invalidate needs to happen so that reads after a write do not
7639                  * get stale data.
7640                  */
7641                 if (!ordered &&
7642                     (!writing ||
7643                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7644                         break;
7645
7646                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7647                                      cached_state, GFP_NOFS);
7648
7649                 if (ordered) {
7650                         /*
7651                          * If we are doing a DIO read and the ordered extent we
7652                          * found is for a buffered write, we can not wait for it
7653                          * to complete and retry, because if we do so we can
7654                          * deadlock with concurrent buffered writes on page
7655                          * locks. This happens only if our DIO read covers more
7656                          * than one extent map, if at this point has already
7657                          * created an ordered extent for a previous extent map
7658                          * and locked its range in the inode's io tree, and a
7659                          * concurrent write against that previous extent map's
7660                          * range and this range started (we unlock the ranges
7661                          * in the io tree only when the bios complete and
7662                          * buffered writes always lock pages before attempting
7663                          * to lock range in the io tree).
7664                          */
7665                         if (writing ||
7666                             test_bit(BTRFS_ORDERED_DIRECT, &ordered->flags))
7667                                 btrfs_start_ordered_extent(inode, ordered, 1);
7668                         else
7669                                 ret = -ENOTBLK;
7670                         btrfs_put_ordered_extent(ordered);
7671                 } else {
7672                         /*
7673                          * We could trigger writeback for this range (and wait
7674                          * for it to complete) and then invalidate the pages for
7675                          * this range (through invalidate_inode_pages2_range()),
7676                          * but that can lead us to a deadlock with a concurrent
7677                          * call to readpages() (a buffered read or a defrag call
7678                          * triggered a readahead) on a page lock due to an
7679                          * ordered dio extent we created before but did not have
7680                          * yet a corresponding bio submitted (whence it can not
7681                          * complete), which makes readpages() wait for that
7682                          * ordered extent to complete while holding a lock on
7683                          * that page.
7684                          */
7685                         ret = -ENOTBLK;
7686                 }
7687
7688                 if (ret)
7689                         break;
7690
7691                 cond_resched();
7692         }
7693
7694         return ret;
7695 }
7696
7697 /* The callers of this must take lock_extent() */
7698 static struct extent_map *create_io_em(struct inode *inode, u64 start, u64 len,
7699                                        u64 orig_start, u64 block_start,
7700                                        u64 block_len, u64 orig_block_len,
7701                                        u64 ram_bytes, int compress_type,
7702                                        int type)
7703 {
7704         struct extent_map_tree *em_tree;
7705         struct extent_map *em;
7706         struct btrfs_root *root = BTRFS_I(inode)->root;
7707         int ret;
7708
7709         ASSERT(type == BTRFS_ORDERED_PREALLOC ||
7710                type == BTRFS_ORDERED_COMPRESSED ||
7711                type == BTRFS_ORDERED_NOCOW ||
7712                type == BTRFS_ORDERED_REGULAR);
7713
7714         em_tree = &BTRFS_I(inode)->extent_tree;
7715         em = alloc_extent_map();
7716         if (!em)
7717                 return ERR_PTR(-ENOMEM);
7718
7719         em->start = start;
7720         em->orig_start = orig_start;
7721         em->len = len;
7722         em->block_len = block_len;
7723         em->block_start = block_start;
7724         em->bdev = root->fs_info->fs_devices->latest_bdev;
7725         em->orig_block_len = orig_block_len;
7726         em->ram_bytes = ram_bytes;
7727         em->generation = -1;
7728         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7729         if (type == BTRFS_ORDERED_PREALLOC) {
7730                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7731         } else if (type == BTRFS_ORDERED_COMPRESSED) {
7732                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
7733                 em->compress_type = compress_type;
7734         }
7735
7736         do {
7737                 btrfs_drop_extent_cache(BTRFS_I(inode), em->start,
7738                                 em->start + em->len - 1, 0);
7739                 write_lock(&em_tree->lock);
7740                 ret = add_extent_mapping(em_tree, em, 1);
7741                 write_unlock(&em_tree->lock);
7742                 /*
7743                  * The caller has taken lock_extent(), who could race with us
7744                  * to add em?
7745                  */
7746         } while (ret == -EEXIST);
7747
7748         if (ret) {
7749                 free_extent_map(em);
7750                 return ERR_PTR(ret);
7751         }
7752
7753         /* em got 2 refs now, callers needs to do free_extent_map once. */
7754         return em;
7755 }
7756
7757 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7758                                    struct buffer_head *bh_result, int create)
7759 {
7760         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7761         struct extent_map *em;
7762         struct extent_state *cached_state = NULL;
7763         struct btrfs_dio_data *dio_data = NULL;
7764         u64 start = iblock << inode->i_blkbits;
7765         u64 lockstart, lockend;
7766         u64 len = bh_result->b_size;
7767         int unlock_bits = EXTENT_LOCKED;
7768         int ret = 0;
7769
7770         if (create)
7771                 unlock_bits |= EXTENT_DIRTY;
7772         else
7773                 len = min_t(u64, len, fs_info->sectorsize);
7774
7775         lockstart = start;
7776         lockend = start + len - 1;
7777
7778         if (current->journal_info) {
7779                 /*
7780                  * Need to pull our outstanding extents and set journal_info to NULL so
7781                  * that anything that needs to check if there's a transaction doesn't get
7782                  * confused.
7783                  */
7784                 dio_data = current->journal_info;
7785                 current->journal_info = NULL;
7786         }
7787
7788         /*
7789          * If this errors out it's because we couldn't invalidate pagecache for
7790          * this range and we need to fallback to buffered.
7791          */
7792         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7793                                create)) {
7794                 ret = -ENOTBLK;
7795                 goto err;
7796         }
7797
7798         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, start, len, 0);
7799         if (IS_ERR(em)) {
7800                 ret = PTR_ERR(em);
7801                 goto unlock_err;
7802         }
7803
7804         /*
7805          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7806          * io.  INLINE is special, and we could probably kludge it in here, but
7807          * it's still buffered so for safety lets just fall back to the generic
7808          * buffered path.
7809          *
7810          * For COMPRESSED we _have_ to read the entire extent in so we can
7811          * decompress it, so there will be buffering required no matter what we
7812          * do, so go ahead and fallback to buffered.
7813          *
7814          * We return -ENOTBLK because that's what makes DIO go ahead and go back
7815          * to buffered IO.  Don't blame me, this is the price we pay for using
7816          * the generic code.
7817          */
7818         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7819             em->block_start == EXTENT_MAP_INLINE) {
7820                 free_extent_map(em);
7821                 ret = -ENOTBLK;
7822                 goto unlock_err;
7823         }
7824
7825         /* Just a good old fashioned hole, return */
7826         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7827                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7828                 free_extent_map(em);
7829                 goto unlock_err;
7830         }
7831
7832         /*
7833          * We don't allocate a new extent in the following cases
7834          *
7835          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7836          * existing extent.
7837          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7838          * just use the extent.
7839          *
7840          */
7841         if (!create) {
7842                 len = min(len, em->len - (start - em->start));
7843                 lockstart = start + len;
7844                 goto unlock;
7845         }
7846
7847         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7848             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7849              em->block_start != EXTENT_MAP_HOLE)) {
7850                 int type;
7851                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7852
7853                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7854                         type = BTRFS_ORDERED_PREALLOC;
7855                 else
7856                         type = BTRFS_ORDERED_NOCOW;
7857                 len = min(len, em->len - (start - em->start));
7858                 block_start = em->block_start + (start - em->start);
7859
7860                 if (can_nocow_extent(inode, start, &len, &orig_start,
7861                                      &orig_block_len, &ram_bytes) == 1 &&
7862                     btrfs_inc_nocow_writers(fs_info, block_start)) {
7863                         struct extent_map *em2;
7864
7865                         em2 = btrfs_create_dio_extent(inode, start, len,
7866                                                       orig_start, block_start,
7867                                                       len, orig_block_len,
7868                                                       ram_bytes, type);
7869                         btrfs_dec_nocow_writers(fs_info, block_start);
7870                         if (type == BTRFS_ORDERED_PREALLOC) {
7871                                 free_extent_map(em);
7872                                 em = em2;
7873                         }
7874                         if (em2 && IS_ERR(em2)) {
7875                                 ret = PTR_ERR(em2);
7876                                 goto unlock_err;
7877                         }
7878                         /*
7879                          * For inode marked NODATACOW or extent marked PREALLOC,
7880                          * use the existing or preallocated extent, so does not
7881                          * need to adjust btrfs_space_info's bytes_may_use.
7882                          */
7883                         btrfs_free_reserved_data_space_noquota(inode,
7884                                         start, len);
7885                         goto unlock;
7886                 }
7887         }
7888
7889         /*
7890          * this will cow the extent, reset the len in case we changed
7891          * it above
7892          */
7893         len = bh_result->b_size;
7894         free_extent_map(em);
7895         em = btrfs_new_extent_direct(inode, start, len);
7896         if (IS_ERR(em)) {
7897                 ret = PTR_ERR(em);
7898                 goto unlock_err;
7899         }
7900         len = min(len, em->len - (start - em->start));
7901 unlock:
7902         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7903                 inode->i_blkbits;
7904         bh_result->b_size = len;
7905         bh_result->b_bdev = em->bdev;
7906         set_buffer_mapped(bh_result);
7907         if (create) {
7908                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7909                         set_buffer_new(bh_result);
7910
7911                 /*
7912                  * Need to update the i_size under the extent lock so buffered
7913                  * readers will get the updated i_size when we unlock.
7914                  */
7915                 if (!dio_data->overwrite && start + len > i_size_read(inode))
7916                         i_size_write(inode, start + len);
7917
7918                 WARN_ON(dio_data->reserve < len);
7919                 dio_data->reserve -= len;
7920                 dio_data->unsubmitted_oe_range_end = start + len;
7921                 current->journal_info = dio_data;
7922         }
7923
7924         /*
7925          * In the case of write we need to clear and unlock the entire range,
7926          * in the case of read we need to unlock only the end area that we
7927          * aren't using if there is any left over space.
7928          */
7929         if (lockstart < lockend) {
7930                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7931                                  lockend, unlock_bits, 1, 0,
7932                                  &cached_state, GFP_NOFS);
7933         } else {
7934                 free_extent_state(cached_state);
7935         }
7936
7937         free_extent_map(em);
7938
7939         return 0;
7940
7941 unlock_err:
7942         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7943                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7944 err:
7945         if (dio_data)
7946                 current->journal_info = dio_data;
7947         return ret;
7948 }
7949
7950 static inline blk_status_t submit_dio_repair_bio(struct inode *inode,
7951                                                  struct bio *bio,
7952                                                  int mirror_num)
7953 {
7954         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7955         blk_status_t ret;
7956
7957         BUG_ON(bio_op(bio) == REQ_OP_WRITE);
7958
7959         bio_get(bio);
7960
7961         ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DIO_REPAIR);
7962         if (ret)
7963                 goto err;
7964
7965         ret = btrfs_map_bio(fs_info, bio, mirror_num, 0);
7966 err:
7967         bio_put(bio);
7968         return ret;
7969 }
7970
7971 static int btrfs_check_dio_repairable(struct inode *inode,
7972                                       struct bio *failed_bio,
7973                                       struct io_failure_record *failrec,
7974                                       int failed_mirror)
7975 {
7976         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
7977         int num_copies;
7978
7979         num_copies = btrfs_num_copies(fs_info, failrec->logical, failrec->len);
7980         if (num_copies == 1) {
7981                 /*
7982                  * we only have a single copy of the data, so don't bother with
7983                  * all the retry and error correction code that follows. no
7984                  * matter what the error is, it is very likely to persist.
7985                  */
7986                 btrfs_debug(fs_info,
7987                         "Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d",
7988                         num_copies, failrec->this_mirror, failed_mirror);
7989                 return 0;
7990         }
7991
7992         failrec->failed_mirror = failed_mirror;
7993         failrec->this_mirror++;
7994         if (failrec->this_mirror == failed_mirror)
7995                 failrec->this_mirror++;
7996
7997         if (failrec->this_mirror > num_copies) {
7998                 btrfs_debug(fs_info,
7999                         "Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d",
8000                         num_copies, failrec->this_mirror, failed_mirror);
8001                 return 0;
8002         }
8003
8004         return 1;
8005 }
8006
8007 static blk_status_t dio_read_error(struct inode *inode, struct bio *failed_bio,
8008                                    struct page *page, unsigned int pgoff,
8009                                    u64 start, u64 end, int failed_mirror,
8010                                    bio_end_io_t *repair_endio, void *repair_arg)
8011 {
8012         struct io_failure_record *failrec;
8013         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8014         struct extent_io_tree *failure_tree = &BTRFS_I(inode)->io_failure_tree;
8015         struct bio *bio;
8016         int isector;
8017         unsigned int read_mode = 0;
8018         int segs;
8019         int ret;
8020         blk_status_t status;
8021
8022         BUG_ON(bio_op(failed_bio) == REQ_OP_WRITE);
8023
8024         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
8025         if (ret)
8026                 return errno_to_blk_status(ret);
8027
8028         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
8029                                          failed_mirror);
8030         if (!ret) {
8031                 free_io_failure(failure_tree, io_tree, failrec);
8032                 return BLK_STS_IOERR;
8033         }
8034
8035         segs = bio_segments(failed_bio);
8036         if (segs > 1 ||
8037             (failed_bio->bi_io_vec->bv_len > btrfs_inode_sectorsize(inode)))
8038                 read_mode |= REQ_FAILFAST_DEV;
8039
8040         isector = start - btrfs_io_bio(failed_bio)->logical;
8041         isector >>= inode->i_sb->s_blocksize_bits;
8042         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
8043                                 pgoff, isector, repair_endio, repair_arg);
8044         bio_set_op_attrs(bio, REQ_OP_READ, read_mode);
8045
8046         btrfs_debug(BTRFS_I(inode)->root->fs_info,
8047                     "repair DIO read error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d",
8048                     read_mode, failrec->this_mirror, failrec->in_validation);
8049
8050         status = submit_dio_repair_bio(inode, bio, failrec->this_mirror);
8051         if (status) {
8052                 free_io_failure(failure_tree, io_tree, failrec);
8053                 bio_put(bio);
8054         }
8055
8056         return status;
8057 }
8058
8059 struct btrfs_retry_complete {
8060         struct completion done;
8061         struct inode *inode;
8062         u64 start;
8063         int uptodate;
8064 };
8065
8066 static void btrfs_retry_endio_nocsum(struct bio *bio)
8067 {
8068         struct btrfs_retry_complete *done = bio->bi_private;
8069         struct inode *inode = done->inode;
8070         struct bio_vec *bvec;
8071         struct extent_io_tree *io_tree, *failure_tree;
8072         int i;
8073
8074         if (bio->bi_status)
8075                 goto end;
8076
8077         ASSERT(bio->bi_vcnt == 1);
8078         io_tree = &BTRFS_I(inode)->io_tree;
8079         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8080         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(inode));
8081
8082         done->uptodate = 1;
8083         ASSERT(!bio_flagged(bio, BIO_CLONED));
8084         bio_for_each_segment_all(bvec, bio, i)
8085                 clean_io_failure(BTRFS_I(inode)->root->fs_info, failure_tree,
8086                                  io_tree, done->start, bvec->bv_page,
8087                                  btrfs_ino(BTRFS_I(inode)), 0);
8088 end:
8089         complete(&done->done);
8090         bio_put(bio);
8091 }
8092
8093 static blk_status_t __btrfs_correct_data_nocsum(struct inode *inode,
8094                                                 struct btrfs_io_bio *io_bio)
8095 {
8096         struct btrfs_fs_info *fs_info;
8097         struct bio_vec bvec;
8098         struct bvec_iter iter;
8099         struct btrfs_retry_complete done;
8100         u64 start;
8101         unsigned int pgoff;
8102         u32 sectorsize;
8103         int nr_sectors;
8104         blk_status_t ret;
8105         blk_status_t err = BLK_STS_OK;
8106
8107         fs_info = BTRFS_I(inode)->root->fs_info;
8108         sectorsize = fs_info->sectorsize;
8109
8110         start = io_bio->logical;
8111         done.inode = inode;
8112         io_bio->bio.bi_iter = io_bio->iter;
8113
8114         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8115                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8116                 pgoff = bvec.bv_offset;
8117
8118 next_block_or_try_again:
8119                 done.uptodate = 0;
8120                 done.start = start;
8121                 init_completion(&done.done);
8122
8123                 ret = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8124                                 pgoff, start, start + sectorsize - 1,
8125                                 io_bio->mirror_num,
8126                                 btrfs_retry_endio_nocsum, &done);
8127                 if (ret) {
8128                         err = ret;
8129                         goto next;
8130                 }
8131
8132                 wait_for_completion_io(&done.done);
8133
8134                 if (!done.uptodate) {
8135                         /* We might have another mirror, so try again */
8136                         goto next_block_or_try_again;
8137                 }
8138
8139 next:
8140                 start += sectorsize;
8141
8142                 nr_sectors--;
8143                 if (nr_sectors) {
8144                         pgoff += sectorsize;
8145                         ASSERT(pgoff < PAGE_SIZE);
8146                         goto next_block_or_try_again;
8147                 }
8148         }
8149
8150         return err;
8151 }
8152
8153 static void btrfs_retry_endio(struct bio *bio)
8154 {
8155         struct btrfs_retry_complete *done = bio->bi_private;
8156         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8157         struct extent_io_tree *io_tree, *failure_tree;
8158         struct inode *inode = done->inode;
8159         struct bio_vec *bvec;
8160         int uptodate;
8161         int ret;
8162         int i;
8163
8164         if (bio->bi_status)
8165                 goto end;
8166
8167         uptodate = 1;
8168
8169         ASSERT(bio->bi_vcnt == 1);
8170         ASSERT(bio->bi_io_vec->bv_len == btrfs_inode_sectorsize(done->inode));
8171
8172         io_tree = &BTRFS_I(inode)->io_tree;
8173         failure_tree = &BTRFS_I(inode)->io_failure_tree;
8174
8175         ASSERT(!bio_flagged(bio, BIO_CLONED));
8176         bio_for_each_segment_all(bvec, bio, i) {
8177                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
8178                                              bvec->bv_offset, done->start,
8179                                              bvec->bv_len);
8180                 if (!ret)
8181                         clean_io_failure(BTRFS_I(inode)->root->fs_info,
8182                                          failure_tree, io_tree, done->start,
8183                                          bvec->bv_page,
8184                                          btrfs_ino(BTRFS_I(inode)),
8185                                          bvec->bv_offset);
8186                 else
8187                         uptodate = 0;
8188         }
8189
8190         done->uptodate = uptodate;
8191 end:
8192         complete(&done->done);
8193         bio_put(bio);
8194 }
8195
8196 static blk_status_t __btrfs_subio_endio_read(struct inode *inode,
8197                 struct btrfs_io_bio *io_bio, blk_status_t err)
8198 {
8199         struct btrfs_fs_info *fs_info;
8200         struct bio_vec bvec;
8201         struct bvec_iter iter;
8202         struct btrfs_retry_complete done;
8203         u64 start;
8204         u64 offset = 0;
8205         u32 sectorsize;
8206         int nr_sectors;
8207         unsigned int pgoff;
8208         int csum_pos;
8209         bool uptodate = (err == 0);
8210         int ret;
8211         blk_status_t status;
8212
8213         fs_info = BTRFS_I(inode)->root->fs_info;
8214         sectorsize = fs_info->sectorsize;
8215
8216         err = BLK_STS_OK;
8217         start = io_bio->logical;
8218         done.inode = inode;
8219         io_bio->bio.bi_iter = io_bio->iter;
8220
8221         bio_for_each_segment(bvec, &io_bio->bio, iter) {
8222                 nr_sectors = BTRFS_BYTES_TO_BLKS(fs_info, bvec.bv_len);
8223
8224                 pgoff = bvec.bv_offset;
8225 next_block:
8226                 if (uptodate) {
8227                         csum_pos = BTRFS_BYTES_TO_BLKS(fs_info, offset);
8228                         ret = __readpage_endio_check(inode, io_bio, csum_pos,
8229                                         bvec.bv_page, pgoff, start, sectorsize);
8230                         if (likely(!ret))
8231                                 goto next;
8232                 }
8233 try_again:
8234                 done.uptodate = 0;
8235                 done.start = start;
8236                 init_completion(&done.done);
8237
8238                 status = dio_read_error(inode, &io_bio->bio, bvec.bv_page,
8239                                         pgoff, start, start + sectorsize - 1,
8240                                         io_bio->mirror_num, btrfs_retry_endio,
8241                                         &done);
8242                 if (status) {
8243                         err = status;
8244                         goto next;
8245                 }
8246
8247                 wait_for_completion_io(&done.done);
8248
8249                 if (!done.uptodate) {
8250                         /* We might have another mirror, so try again */
8251                         goto try_again;
8252                 }
8253 next:
8254                 offset += sectorsize;
8255                 start += sectorsize;
8256
8257                 ASSERT(nr_sectors);
8258
8259                 nr_sectors--;
8260                 if (nr_sectors) {
8261                         pgoff += sectorsize;
8262                         ASSERT(pgoff < PAGE_SIZE);
8263                         goto next_block;
8264                 }
8265         }
8266
8267         return err;
8268 }
8269
8270 static blk_status_t btrfs_subio_endio_read(struct inode *inode,
8271                 struct btrfs_io_bio *io_bio, blk_status_t err)
8272 {
8273         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8274
8275         if (skip_csum) {
8276                 if (unlikely(err))
8277                         return __btrfs_correct_data_nocsum(inode, io_bio);
8278                 else
8279                         return BLK_STS_OK;
8280         } else {
8281                 return __btrfs_subio_endio_read(inode, io_bio, err);
8282         }
8283 }
8284
8285 static void btrfs_endio_direct_read(struct bio *bio)
8286 {
8287         struct btrfs_dio_private *dip = bio->bi_private;
8288         struct inode *inode = dip->inode;
8289         struct bio *dio_bio;
8290         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8291         blk_status_t err = bio->bi_status;
8292
8293         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
8294                 err = btrfs_subio_endio_read(inode, io_bio, err);
8295
8296         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
8297                       dip->logical_offset + dip->bytes - 1);
8298         dio_bio = dip->dio_bio;
8299
8300         kfree(dip);
8301
8302         dio_bio->bi_status = err;
8303         dio_end_io(dio_bio);
8304
8305         if (io_bio->end_io)
8306                 io_bio->end_io(io_bio, blk_status_to_errno(err));
8307         bio_put(bio);
8308 }
8309
8310 static void __endio_write_update_ordered(struct inode *inode,
8311                                          const u64 offset, const u64 bytes,
8312                                          const bool uptodate)
8313 {
8314         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8315         struct btrfs_ordered_extent *ordered = NULL;
8316         struct btrfs_workqueue *wq;
8317         btrfs_work_func_t func;
8318         u64 ordered_offset = offset;
8319         u64 ordered_bytes = bytes;
8320         u64 last_offset;
8321         int ret;
8322
8323         if (btrfs_is_free_space_inode(BTRFS_I(inode))) {
8324                 wq = fs_info->endio_freespace_worker;
8325                 func = btrfs_freespace_write_helper;
8326         } else {
8327                 wq = fs_info->endio_write_workers;
8328                 func = btrfs_endio_write_helper;
8329         }
8330
8331 again:
8332         last_offset = ordered_offset;
8333         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
8334                                                    &ordered_offset,
8335                                                    ordered_bytes,
8336                                                    uptodate);
8337         if (!ret)
8338                 goto out_test;
8339
8340         btrfs_init_work(&ordered->work, func, finish_ordered_fn, NULL, NULL);
8341         btrfs_queue_work(wq, &ordered->work);
8342 out_test:
8343         /*
8344          * If btrfs_dec_test_ordered_pending does not find any ordered extent
8345          * in the range, we can exit.
8346          */
8347         if (ordered_offset == last_offset)
8348                 return;
8349         /*
8350          * our bio might span multiple ordered extents.  If we haven't
8351          * completed the accounting for the whole dio, go back and try again
8352          */
8353         if (ordered_offset < offset + bytes) {
8354                 ordered_bytes = offset + bytes - ordered_offset;
8355                 ordered = NULL;
8356                 goto again;
8357         }
8358 }
8359
8360 static void btrfs_endio_direct_write(struct bio *bio)
8361 {
8362         struct btrfs_dio_private *dip = bio->bi_private;
8363         struct bio *dio_bio = dip->dio_bio;
8364
8365         __endio_write_update_ordered(dip->inode, dip->logical_offset,
8366                                      dip->bytes, !bio->bi_status);
8367
8368         kfree(dip);
8369
8370         dio_bio->bi_status = bio->bi_status;
8371         dio_end_io(dio_bio);
8372         bio_put(bio);
8373 }
8374
8375 static blk_status_t __btrfs_submit_bio_start_direct_io(void *private_data,
8376                                     struct bio *bio, int mirror_num,
8377                                     unsigned long bio_flags, u64 offset)
8378 {
8379         struct inode *inode = private_data;
8380         blk_status_t ret;
8381         ret = btrfs_csum_one_bio(inode, bio, offset, 1);
8382         BUG_ON(ret); /* -ENOMEM */
8383         return 0;
8384 }
8385
8386 static void btrfs_end_dio_bio(struct bio *bio)
8387 {
8388         struct btrfs_dio_private *dip = bio->bi_private;
8389         blk_status_t err = bio->bi_status;
8390
8391         if (err)
8392                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8393                            "direct IO failed ino %llu rw %d,%u sector %#Lx len %u err no %d",
8394                            btrfs_ino(BTRFS_I(dip->inode)), bio_op(bio),
8395                            bio->bi_opf,
8396                            (unsigned long long)bio->bi_iter.bi_sector,
8397                            bio->bi_iter.bi_size, err);
8398
8399         if (dip->subio_endio)
8400                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8401
8402         if (err) {
8403                 dip->errors = 1;
8404
8405                 /*
8406                  * before atomic variable goto zero, we must make sure
8407                  * dip->errors is perceived to be set.
8408                  */
8409                 smp_mb__before_atomic();
8410         }
8411
8412         /* if there are more bios still pending for this dio, just exit */
8413         if (!atomic_dec_and_test(&dip->pending_bios))
8414                 goto out;
8415
8416         if (dip->errors) {
8417                 bio_io_error(dip->orig_bio);
8418         } else {
8419                 dip->dio_bio->bi_status = BLK_STS_OK;
8420                 bio_endio(dip->orig_bio);
8421         }
8422 out:
8423         bio_put(bio);
8424 }
8425
8426 static inline blk_status_t btrfs_lookup_and_bind_dio_csum(struct inode *inode,
8427                                                  struct btrfs_dio_private *dip,
8428                                                  struct bio *bio,
8429                                                  u64 file_offset)
8430 {
8431         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8432         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8433         blk_status_t ret;
8434
8435         /*
8436          * We load all the csum data we need when we submit
8437          * the first bio to reduce the csum tree search and
8438          * contention.
8439          */
8440         if (dip->logical_offset == file_offset) {
8441                 ret = btrfs_lookup_bio_sums_dio(inode, dip->orig_bio,
8442                                                 file_offset);
8443                 if (ret)
8444                         return ret;
8445         }
8446
8447         if (bio == dip->orig_bio)
8448                 return 0;
8449
8450         file_offset -= dip->logical_offset;
8451         file_offset >>= inode->i_sb->s_blocksize_bits;
8452         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8453
8454         return 0;
8455 }
8456
8457 static inline blk_status_t
8458 __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode, u64 file_offset,
8459                        int async_submit)
8460 {
8461         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8462         struct btrfs_dio_private *dip = bio->bi_private;
8463         bool write = bio_op(bio) == REQ_OP_WRITE;
8464         blk_status_t ret;
8465
8466         if (async_submit)
8467                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8468
8469         bio_get(bio);
8470
8471         if (!write) {
8472                 ret = btrfs_bio_wq_end_io(fs_info, bio, BTRFS_WQ_ENDIO_DATA);
8473                 if (ret)
8474                         goto err;
8475         }
8476
8477         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
8478                 goto map;
8479
8480         if (write && async_submit) {
8481                 ret = btrfs_wq_submit_bio(fs_info, bio, 0, 0,
8482                                           file_offset, inode,
8483                                           __btrfs_submit_bio_start_direct_io,
8484                                           __btrfs_submit_bio_done);
8485                 goto err;
8486         } else if (write) {
8487                 /*
8488                  * If we aren't doing async submit, calculate the csum of the
8489                  * bio now.
8490                  */
8491                 ret = btrfs_csum_one_bio(inode, bio, file_offset, 1);
8492                 if (ret)
8493                         goto err;
8494         } else {
8495                 ret = btrfs_lookup_and_bind_dio_csum(inode, dip, bio,
8496                                                      file_offset);
8497                 if (ret)
8498                         goto err;
8499         }
8500 map:
8501         ret = btrfs_map_bio(fs_info, bio, 0, 0);
8502 err:
8503         bio_put(bio);
8504         return ret;
8505 }
8506
8507 static int btrfs_submit_direct_hook(struct btrfs_dio_private *dip)
8508 {
8509         struct inode *inode = dip->inode;
8510         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8511         struct bio *bio;
8512         struct bio *orig_bio = dip->orig_bio;
8513         u64 start_sector = orig_bio->bi_iter.bi_sector;
8514         u64 file_offset = dip->logical_offset;
8515         u64 map_length;
8516         int async_submit = 0;
8517         u64 submit_len;
8518         int clone_offset = 0;
8519         int clone_len;
8520         int ret;
8521         blk_status_t status;
8522
8523         map_length = orig_bio->bi_iter.bi_size;
8524         submit_len = map_length;
8525         ret = btrfs_map_block(fs_info, btrfs_op(orig_bio), start_sector << 9,
8526                               &map_length, NULL, 0);
8527         if (ret)
8528                 return -EIO;
8529
8530         if (map_length >= submit_len) {
8531                 bio = orig_bio;
8532                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8533                 goto submit;
8534         }
8535
8536         /* async crcs make it difficult to collect full stripe writes. */
8537         if (btrfs_data_alloc_profile(fs_info) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8538                 async_submit = 0;
8539         else
8540                 async_submit = 1;
8541
8542         /* bio split */
8543         ASSERT(map_length <= INT_MAX);
8544         atomic_inc(&dip->pending_bios);
8545         do {
8546                 clone_len = min_t(int, submit_len, map_length);
8547
8548                 /*
8549                  * This will never fail as it's passing GPF_NOFS and
8550                  * the allocation is backed by btrfs_bioset.
8551                  */
8552                 bio = btrfs_bio_clone_partial(orig_bio, clone_offset,
8553                                               clone_len);
8554                 bio->bi_private = dip;
8555                 bio->bi_end_io = btrfs_end_dio_bio;
8556                 btrfs_io_bio(bio)->logical = file_offset;
8557
8558                 ASSERT(submit_len >= clone_len);
8559                 submit_len -= clone_len;
8560                 if (submit_len == 0)
8561                         break;
8562
8563                 /*
8564                  * Increase the count before we submit the bio so we know
8565                  * the end IO handler won't happen before we increase the
8566                  * count. Otherwise, the dip might get freed before we're
8567                  * done setting it up.
8568                  */
8569                 atomic_inc(&dip->pending_bios);
8570
8571                 status = __btrfs_submit_dio_bio(bio, inode, file_offset,
8572                                                 async_submit);
8573                 if (status) {
8574                         bio_put(bio);
8575                         atomic_dec(&dip->pending_bios);
8576                         goto out_err;
8577                 }
8578
8579                 clone_offset += clone_len;
8580                 start_sector += clone_len >> 9;
8581                 file_offset += clone_len;
8582
8583                 map_length = submit_len;
8584                 ret = btrfs_map_block(fs_info, btrfs_op(orig_bio),
8585                                       start_sector << 9, &map_length, NULL, 0);
8586                 if (ret)
8587                         goto out_err;
8588         } while (submit_len > 0);
8589
8590 submit:
8591         status = __btrfs_submit_dio_bio(bio, inode, file_offset, async_submit);
8592         if (!status)
8593                 return 0;
8594
8595         bio_put(bio);
8596 out_err:
8597         dip->errors = 1;
8598         /*
8599          * before atomic variable goto zero, we must
8600          * make sure dip->errors is perceived to be set.
8601          */
8602         smp_mb__before_atomic();
8603         if (atomic_dec_and_test(&dip->pending_bios))
8604                 bio_io_error(dip->orig_bio);
8605
8606         /* bio_end_io() will handle error, so we needn't return it */
8607         return 0;
8608 }
8609
8610 static void btrfs_submit_direct(struct bio *dio_bio, struct inode *inode,
8611                                 loff_t file_offset)
8612 {
8613         struct btrfs_dio_private *dip = NULL;
8614         struct bio *bio = NULL;
8615         struct btrfs_io_bio *io_bio;
8616         bool write = (bio_op(dio_bio) == REQ_OP_WRITE);
8617         int ret = 0;
8618
8619         bio = btrfs_bio_clone(dio_bio);
8620
8621         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8622         if (!dip) {
8623                 ret = -ENOMEM;
8624                 goto free_ordered;
8625         }
8626
8627         dip->private = dio_bio->bi_private;
8628         dip->inode = inode;
8629         dip->logical_offset = file_offset;
8630         dip->bytes = dio_bio->bi_iter.bi_size;
8631         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8632         bio->bi_private = dip;
8633         dip->orig_bio = bio;
8634         dip->dio_bio = dio_bio;
8635         atomic_set(&dip->pending_bios, 0);
8636         io_bio = btrfs_io_bio(bio);
8637         io_bio->logical = file_offset;
8638
8639         if (write) {
8640                 bio->bi_end_io = btrfs_endio_direct_write;
8641         } else {
8642                 bio->bi_end_io = btrfs_endio_direct_read;
8643                 dip->subio_endio = btrfs_subio_endio_read;
8644         }
8645
8646         /*
8647          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8648          * even if we fail to submit a bio, because in such case we do the
8649          * corresponding error handling below and it must not be done a second
8650          * time by btrfs_direct_IO().
8651          */
8652         if (write) {
8653                 struct btrfs_dio_data *dio_data = current->journal_info;
8654
8655                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8656                         dip->bytes;
8657                 dio_data->unsubmitted_oe_range_start =
8658                         dio_data->unsubmitted_oe_range_end;
8659         }
8660
8661         ret = btrfs_submit_direct_hook(dip);
8662         if (!ret)
8663                 return;
8664
8665         if (io_bio->end_io)
8666                 io_bio->end_io(io_bio, ret);
8667
8668 free_ordered:
8669         /*
8670          * If we arrived here it means either we failed to submit the dip
8671          * or we either failed to clone the dio_bio or failed to allocate the
8672          * dip. If we cloned the dio_bio and allocated the dip, we can just
8673          * call bio_endio against our io_bio so that we get proper resource
8674          * cleanup if we fail to submit the dip, otherwise, we must do the
8675          * same as btrfs_endio_direct_[write|read] because we can't call these
8676          * callbacks - they require an allocated dip and a clone of dio_bio.
8677          */
8678         if (bio && dip) {
8679                 bio_io_error(bio);
8680                 /*
8681                  * The end io callbacks free our dip, do the final put on bio
8682                  * and all the cleanup and final put for dio_bio (through
8683                  * dio_end_io()).
8684                  */
8685                 dip = NULL;
8686                 bio = NULL;
8687         } else {
8688                 if (write)
8689                         __endio_write_update_ordered(inode,
8690                                                 file_offset,
8691                                                 dio_bio->bi_iter.bi_size,
8692                                                 false);
8693                 else
8694                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8695                               file_offset + dio_bio->bi_iter.bi_size - 1);
8696
8697                 dio_bio->bi_status = BLK_STS_IOERR;
8698                 /*
8699                  * Releases and cleans up our dio_bio, no need to bio_put()
8700                  * nor bio_endio()/bio_io_error() against dio_bio.
8701                  */
8702                 dio_end_io(dio_bio);
8703         }
8704         if (bio)
8705                 bio_put(bio);
8706         kfree(dip);
8707 }
8708
8709 static ssize_t check_direct_IO(struct btrfs_fs_info *fs_info,
8710                                const struct iov_iter *iter, loff_t offset)
8711 {
8712         int seg;
8713         int i;
8714         unsigned int blocksize_mask = fs_info->sectorsize - 1;
8715         ssize_t retval = -EINVAL;
8716
8717         if (offset & blocksize_mask)
8718                 goto out;
8719
8720         if (iov_iter_alignment(iter) & blocksize_mask)
8721                 goto out;
8722
8723         /* If this is a write we don't need to check anymore */
8724         if (iov_iter_rw(iter) != READ || !iter_is_iovec(iter))
8725                 return 0;
8726         /*
8727          * Check to make sure we don't have duplicate iov_base's in this
8728          * iovec, if so return EINVAL, otherwise we'll get csum errors
8729          * when reading back.
8730          */
8731         for (seg = 0; seg < iter->nr_segs; seg++) {
8732                 for (i = seg + 1; i < iter->nr_segs; i++) {
8733                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8734                                 goto out;
8735                 }
8736         }
8737         retval = 0;
8738 out:
8739         return retval;
8740 }
8741
8742 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
8743 {
8744         struct file *file = iocb->ki_filp;
8745         struct inode *inode = file->f_mapping->host;
8746         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
8747         struct btrfs_dio_data dio_data = { 0 };
8748         struct extent_changeset *data_reserved = NULL;
8749         loff_t offset = iocb->ki_pos;
8750         size_t count = 0;
8751         int flags = 0;
8752         bool wakeup = true;
8753         bool relock = false;
8754         ssize_t ret;
8755
8756         if (check_direct_IO(fs_info, iter, offset))
8757                 return 0;
8758
8759         inode_dio_begin(inode);
8760
8761         /*
8762          * The generic stuff only does filemap_write_and_wait_range, which
8763          * isn't enough if we've written compressed pages to this area, so
8764          * we need to flush the dirty pages again to make absolutely sure
8765          * that any outstanding dirty pages are on disk.
8766          */
8767         count = iov_iter_count(iter);
8768         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8769                      &BTRFS_I(inode)->runtime_flags))
8770                 filemap_fdatawrite_range(inode->i_mapping, offset,
8771                                          offset + count - 1);
8772
8773         if (iov_iter_rw(iter) == WRITE) {
8774                 /*
8775                  * If the write DIO is beyond the EOF, we need update
8776                  * the isize, but it is protected by i_mutex. So we can
8777                  * not unlock the i_mutex at this case.
8778                  */
8779                 if (offset + count <= inode->i_size) {
8780                         dio_data.overwrite = 1;
8781                         inode_unlock(inode);
8782                         relock = true;
8783                 } else if (iocb->ki_flags & IOCB_NOWAIT) {
8784                         ret = -EAGAIN;
8785                         goto out;
8786                 }
8787                 ret = btrfs_delalloc_reserve_space(inode, &data_reserved,
8788                                                    offset, count);
8789                 if (ret)
8790                         goto out;
8791
8792                 /*
8793                  * We need to know how many extents we reserved so that we can
8794                  * do the accounting properly if we go over the number we
8795                  * originally calculated.  Abuse current->journal_info for this.
8796                  */
8797                 dio_data.reserve = round_up(count,
8798                                             fs_info->sectorsize);
8799                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8800                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8801                 current->journal_info = &dio_data;
8802                 down_read(&BTRFS_I(inode)->dio_sem);
8803         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8804                                      &BTRFS_I(inode)->runtime_flags)) {
8805                 inode_dio_end(inode);
8806                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8807                 wakeup = false;
8808         }
8809
8810         ret = __blockdev_direct_IO(iocb, inode,
8811                                    fs_info->fs_devices->latest_bdev,
8812                                    iter, btrfs_get_blocks_direct, NULL,
8813                                    btrfs_submit_direct, flags);
8814         if (iov_iter_rw(iter) == WRITE) {
8815                 up_read(&BTRFS_I(inode)->dio_sem);
8816                 current->journal_info = NULL;
8817                 if (ret < 0 && ret != -EIOCBQUEUED) {
8818                         if (dio_data.reserve)
8819                                 btrfs_delalloc_release_space(inode, data_reserved,
8820                                         offset, dio_data.reserve);
8821                         /*
8822                          * On error we might have left some ordered extents
8823                          * without submitting corresponding bios for them, so
8824                          * cleanup them up to avoid other tasks getting them
8825                          * and waiting for them to complete forever.
8826                          */
8827                         if (dio_data.unsubmitted_oe_range_start <
8828                             dio_data.unsubmitted_oe_range_end)
8829                                 __endio_write_update_ordered(inode,
8830                                         dio_data.unsubmitted_oe_range_start,
8831                                         dio_data.unsubmitted_oe_range_end -
8832                                         dio_data.unsubmitted_oe_range_start,
8833                                         false);
8834                 } else if (ret >= 0 && (size_t)ret < count)
8835                         btrfs_delalloc_release_space(inode, data_reserved,
8836                                         offset, count - (size_t)ret);
8837                 btrfs_delalloc_release_extents(BTRFS_I(inode), count);
8838         }
8839 out:
8840         if (wakeup)
8841                 inode_dio_end(inode);
8842         if (relock)
8843                 inode_lock(inode);
8844
8845         extent_changeset_free(data_reserved);
8846         return ret;
8847 }
8848
8849 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8850
8851 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8852                 __u64 start, __u64 len)
8853 {
8854         int     ret;
8855
8856         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8857         if (ret)
8858                 return ret;
8859
8860         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8861 }
8862
8863 int btrfs_readpage(struct file *file, struct page *page)
8864 {
8865         struct extent_io_tree *tree;
8866         tree = &BTRFS_I(page->mapping->host)->io_tree;
8867         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8868 }
8869
8870 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8871 {
8872         struct extent_io_tree *tree;
8873         struct inode *inode = page->mapping->host;
8874         int ret;
8875
8876         if (current->flags & PF_MEMALLOC) {
8877                 redirty_page_for_writepage(wbc, page);
8878                 unlock_page(page);
8879                 return 0;
8880         }
8881
8882         /*
8883          * If we are under memory pressure we will call this directly from the
8884          * VM, we need to make sure we have the inode referenced for the ordered
8885          * extent.  If not just return like we didn't do anything.
8886          */
8887         if (!igrab(inode)) {
8888                 redirty_page_for_writepage(wbc, page);
8889                 return AOP_WRITEPAGE_ACTIVATE;
8890         }
8891         tree = &BTRFS_I(page->mapping->host)->io_tree;
8892         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8893         btrfs_add_delayed_iput(inode);
8894         return ret;
8895 }
8896
8897 static int btrfs_writepages(struct address_space *mapping,
8898                             struct writeback_control *wbc)
8899 {
8900         struct extent_io_tree *tree;
8901
8902         tree = &BTRFS_I(mapping->host)->io_tree;
8903         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8904 }
8905
8906 static int
8907 btrfs_readpages(struct file *file, struct address_space *mapping,
8908                 struct list_head *pages, unsigned nr_pages)
8909 {
8910         struct extent_io_tree *tree;
8911         tree = &BTRFS_I(mapping->host)->io_tree;
8912         return extent_readpages(tree, mapping, pages, nr_pages,
8913                                 btrfs_get_extent);
8914 }
8915 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8916 {
8917         struct extent_io_tree *tree;
8918         struct extent_map_tree *map;
8919         int ret;
8920
8921         tree = &BTRFS_I(page->mapping->host)->io_tree;
8922         map = &BTRFS_I(page->mapping->host)->extent_tree;
8923         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8924         if (ret == 1) {
8925                 ClearPagePrivate(page);
8926                 set_page_private(page, 0);
8927                 put_page(page);
8928         }
8929         return ret;
8930 }
8931
8932 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8933 {
8934         if (PageWriteback(page) || PageDirty(page))
8935                 return 0;
8936         return __btrfs_releasepage(page, gfp_flags);
8937 }
8938
8939 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8940                                  unsigned int length)
8941 {
8942         struct inode *inode = page->mapping->host;
8943         struct extent_io_tree *tree;
8944         struct btrfs_ordered_extent *ordered;
8945         struct extent_state *cached_state = NULL;
8946         u64 page_start = page_offset(page);
8947         u64 page_end = page_start + PAGE_SIZE - 1;
8948         u64 start;
8949         u64 end;
8950         int inode_evicting = inode->i_state & I_FREEING;
8951
8952         /*
8953          * we have the page locked, so new writeback can't start,
8954          * and the dirty bit won't be cleared while we are here.
8955          *
8956          * Wait for IO on this page so that we can safely clear
8957          * the PagePrivate2 bit and do ordered accounting
8958          */
8959         wait_on_page_writeback(page);
8960
8961         tree = &BTRFS_I(inode)->io_tree;
8962         if (offset) {
8963                 btrfs_releasepage(page, GFP_NOFS);
8964                 return;
8965         }
8966
8967         if (!inode_evicting)
8968                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8969 again:
8970         start = page_start;
8971         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), start,
8972                                         page_end - start + 1);
8973         if (ordered) {
8974                 end = min(page_end, ordered->file_offset + ordered->len - 1);
8975                 /*
8976                  * IO on this page will never be started, so we need
8977                  * to account for any ordered extents now
8978                  */
8979                 if (!inode_evicting)
8980                         clear_extent_bit(tree, start, end,
8981                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8982                                          EXTENT_DELALLOC_NEW |
8983                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8984                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8985                                          GFP_NOFS);
8986                 /*
8987                  * whoever cleared the private bit is responsible
8988                  * for the finish_ordered_io
8989                  */
8990                 if (TestClearPagePrivate2(page)) {
8991                         struct btrfs_ordered_inode_tree *tree;
8992                         u64 new_len;
8993
8994                         tree = &BTRFS_I(inode)->ordered_tree;
8995
8996                         spin_lock_irq(&tree->lock);
8997                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8998                         new_len = start - ordered->file_offset;
8999                         if (new_len < ordered->truncated_len)
9000                                 ordered->truncated_len = new_len;
9001                         spin_unlock_irq(&tree->lock);
9002
9003                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
9004                                                            start,
9005                                                            end - start + 1, 1))
9006                                 btrfs_finish_ordered_io(ordered);
9007                 }
9008                 btrfs_put_ordered_extent(ordered);
9009                 if (!inode_evicting) {
9010                         cached_state = NULL;
9011                         lock_extent_bits(tree, start, end,
9012                                          &cached_state);
9013                 }
9014
9015                 start = end + 1;
9016                 if (start < page_end)
9017                         goto again;
9018         }
9019
9020         /*
9021          * Qgroup reserved space handler
9022          * Page here will be either
9023          * 1) Already written to disk
9024          *    In this case, its reserved space is released from data rsv map
9025          *    and will be freed by delayed_ref handler finally.
9026          *    So even we call qgroup_free_data(), it won't decrease reserved
9027          *    space.
9028          * 2) Not written to disk
9029          *    This means the reserved space should be freed here. However,
9030          *    if a truncate invalidates the page (by clearing PageDirty)
9031          *    and the page is accounted for while allocating extent
9032          *    in btrfs_check_data_free_space() we let delayed_ref to
9033          *    free the entire extent.
9034          */
9035         if (PageDirty(page))
9036                 btrfs_qgroup_free_data(inode, NULL, page_start, PAGE_SIZE);
9037         if (!inode_evicting) {
9038                 clear_extent_bit(tree, page_start, page_end,
9039                                  EXTENT_LOCKED | EXTENT_DIRTY |
9040                                  EXTENT_DELALLOC | EXTENT_DELALLOC_NEW |
9041                                  EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG, 1, 1,
9042                                  &cached_state, GFP_NOFS);
9043
9044                 __btrfs_releasepage(page, GFP_NOFS);
9045         }
9046
9047         ClearPageChecked(page);
9048         if (PagePrivate(page)) {
9049                 ClearPagePrivate(page);
9050                 set_page_private(page, 0);
9051                 put_page(page);
9052         }
9053 }
9054
9055 /*
9056  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
9057  * called from a page fault handler when a page is first dirtied. Hence we must
9058  * be careful to check for EOF conditions here. We set the page up correctly
9059  * for a written page which means we get ENOSPC checking when writing into
9060  * holes and correct delalloc and unwritten extent mapping on filesystems that
9061  * support these features.
9062  *
9063  * We are not allowed to take the i_mutex here so we have to play games to
9064  * protect against truncate races as the page could now be beyond EOF.  Because
9065  * vmtruncate() writes the inode size before removing pages, once we have the
9066  * page lock we can determine safely if the page is beyond EOF. If it is not
9067  * beyond EOF, then the page is guaranteed safe against truncation until we
9068  * unlock the page.
9069  */
9070 int btrfs_page_mkwrite(struct vm_fault *vmf)
9071 {
9072         struct page *page = vmf->page;
9073         struct inode *inode = file_inode(vmf->vma->vm_file);
9074         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9075         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
9076         struct btrfs_ordered_extent *ordered;
9077         struct extent_state *cached_state = NULL;
9078         struct extent_changeset *data_reserved = NULL;
9079         char *kaddr;
9080         unsigned long zero_start;
9081         loff_t size;
9082         int ret;
9083         int reserved = 0;
9084         u64 reserved_space;
9085         u64 page_start;
9086         u64 page_end;
9087         u64 end;
9088
9089         reserved_space = PAGE_SIZE;
9090
9091         sb_start_pagefault(inode->i_sb);
9092         page_start = page_offset(page);
9093         page_end = page_start + PAGE_SIZE - 1;
9094         end = page_end;
9095
9096         /*
9097          * Reserving delalloc space after obtaining the page lock can lead to
9098          * deadlock. For example, if a dirty page is locked by this function
9099          * and the call to btrfs_delalloc_reserve_space() ends up triggering
9100          * dirty page write out, then the btrfs_writepage() function could
9101          * end up waiting indefinitely to get a lock on the page currently
9102          * being processed by btrfs_page_mkwrite() function.
9103          */
9104         ret = btrfs_delalloc_reserve_space(inode, &data_reserved, page_start,
9105                                            reserved_space);
9106         if (!ret) {
9107                 ret = file_update_time(vmf->vma->vm_file);
9108                 reserved = 1;
9109         }
9110         if (ret) {
9111                 if (ret == -ENOMEM)
9112                         ret = VM_FAULT_OOM;
9113                 else /* -ENOSPC, -EIO, etc */
9114                         ret = VM_FAULT_SIGBUS;
9115                 if (reserved)
9116                         goto out;
9117                 goto out_noreserve;
9118         }
9119
9120         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
9121 again:
9122         lock_page(page);
9123         size = i_size_read(inode);
9124
9125         if ((page->mapping != inode->i_mapping) ||
9126             (page_start >= size)) {
9127                 /* page got truncated out from underneath us */
9128                 goto out_unlock;
9129         }
9130         wait_on_page_writeback(page);
9131
9132         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
9133         set_page_extent_mapped(page);
9134
9135         /*
9136          * we can't set the delalloc bits if there are pending ordered
9137          * extents.  Drop our locks and wait for them to finish
9138          */
9139         ordered = btrfs_lookup_ordered_range(BTRFS_I(inode), page_start,
9140                         PAGE_SIZE);
9141         if (ordered) {
9142                 unlock_extent_cached(io_tree, page_start, page_end,
9143                                      &cached_state, GFP_NOFS);
9144                 unlock_page(page);
9145                 btrfs_start_ordered_extent(inode, ordered, 1);
9146                 btrfs_put_ordered_extent(ordered);
9147                 goto again;
9148         }
9149
9150         if (page->index == ((size - 1) >> PAGE_SHIFT)) {
9151                 reserved_space = round_up(size - page_start,
9152                                           fs_info->sectorsize);
9153                 if (reserved_space < PAGE_SIZE) {
9154                         end = page_start + reserved_space - 1;
9155                         btrfs_delalloc_release_space(inode, data_reserved,
9156                                         page_start, PAGE_SIZE - reserved_space);
9157                 }
9158         }
9159
9160         /*
9161          * page_mkwrite gets called when the page is firstly dirtied after it's
9162          * faulted in, but write(2) could also dirty a page and set delalloc
9163          * bits, thus in this case for space account reason, we still need to
9164          * clear any delalloc bits within this page range since we have to
9165          * reserve data&meta space before lock_page() (see above comments).
9166          */
9167         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, end,
9168                           EXTENT_DIRTY | EXTENT_DELALLOC |
9169                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
9170                           0, 0, &cached_state, GFP_NOFS);
9171
9172         ret = btrfs_set_extent_delalloc(inode, page_start, end, 0,
9173                                         &cached_state, 0);
9174         if (ret) {
9175                 unlock_extent_cached(io_tree, page_start, page_end,
9176                                      &cached_state, GFP_NOFS);
9177                 ret = VM_FAULT_SIGBUS;
9178                 goto out_unlock;
9179         }
9180         ret = 0;
9181
9182         /* page is wholly or partially inside EOF */
9183         if (page_start + PAGE_SIZE > size)
9184                 zero_start = size & ~PAGE_MASK;
9185         else
9186                 zero_start = PAGE_SIZE;
9187
9188         if (zero_start != PAGE_SIZE) {
9189                 kaddr = kmap(page);
9190                 memset(kaddr + zero_start, 0, PAGE_SIZE - zero_start);
9191                 flush_dcache_page(page);
9192                 kunmap(page);
9193         }
9194         ClearPageChecked(page);
9195         set_page_dirty(page);
9196         SetPageUptodate(page);
9197
9198         BTRFS_I(inode)->last_trans = fs_info->generation;
9199         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
9200         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
9201
9202         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
9203
9204 out_unlock:
9205         if (!ret) {
9206                 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9207                 sb_end_pagefault(inode->i_sb);
9208                 extent_changeset_free(data_reserved);
9209                 return VM_FAULT_LOCKED;
9210         }
9211         unlock_page(page);
9212 out:
9213         btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE);
9214         btrfs_delalloc_release_space(inode, data_reserved, page_start,
9215                                      reserved_space);
9216 out_noreserve:
9217         sb_end_pagefault(inode->i_sb);
9218         extent_changeset_free(data_reserved);
9219         return ret;
9220 }
9221
9222 static int btrfs_truncate(struct inode *inode)
9223 {
9224         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9225         struct btrfs_root *root = BTRFS_I(inode)->root;
9226         struct btrfs_block_rsv *rsv;
9227         int ret = 0;
9228         int err = 0;
9229         struct btrfs_trans_handle *trans;
9230         u64 mask = fs_info->sectorsize - 1;
9231         u64 min_size = btrfs_calc_trunc_metadata_size(fs_info, 1);
9232
9233         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
9234                                        (u64)-1);
9235         if (ret)
9236                 return ret;
9237
9238         /*
9239          * Yes ladies and gentlemen, this is indeed ugly.  The fact is we have
9240          * 3 things going on here
9241          *
9242          * 1) We need to reserve space for our orphan item and the space to
9243          * delete our orphan item.  Lord knows we don't want to have a dangling
9244          * orphan item because we didn't reserve space to remove it.
9245          *
9246          * 2) We need to reserve space to update our inode.
9247          *
9248          * 3) We need to have something to cache all the space that is going to
9249          * be free'd up by the truncate operation, but also have some slack
9250          * space reserved in case it uses space during the truncate (thank you
9251          * very much snapshotting).
9252          *
9253          * And we need these to all be separate.  The fact is we can use a lot of
9254          * space doing the truncate, and we have no earthly idea how much space
9255          * we will use, so we need the truncate reservation to be separate so it
9256          * doesn't end up using space reserved for updating the inode or
9257          * removing the orphan item.  We also need to be able to stop the
9258          * transaction and start a new one, which means we need to be able to
9259          * update the inode several times, and we have no idea of knowing how
9260          * many times that will be, so we can't just reserve 1 item for the
9261          * entirety of the operation, so that has to be done separately as well.
9262          * Then there is the orphan item, which does indeed need to be held on
9263          * to for the whole operation, and we need nobody to touch this reserved
9264          * space except the orphan code.
9265          *
9266          * So that leaves us with
9267          *
9268          * 1) root->orphan_block_rsv - for the orphan deletion.
9269          * 2) rsv - for the truncate reservation, which we will steal from the
9270          * transaction reservation.
9271          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
9272          * updating the inode.
9273          */
9274         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
9275         if (!rsv)
9276                 return -ENOMEM;
9277         rsv->size = min_size;
9278         rsv->failfast = 1;
9279
9280         /*
9281          * 1 for the truncate slack space
9282          * 1 for updating the inode.
9283          */
9284         trans = btrfs_start_transaction(root, 2);
9285         if (IS_ERR(trans)) {
9286                 err = PTR_ERR(trans);
9287                 goto out;
9288         }
9289
9290         /* Migrate the slack space for the truncate to our reserve */
9291         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
9292                                       min_size, 0);
9293         BUG_ON(ret);
9294
9295         /*
9296          * So if we truncate and then write and fsync we normally would just
9297          * write the extents that changed, which is a problem if we need to
9298          * first truncate that entire inode.  So set this flag so we write out
9299          * all of the extents in the inode to the sync log so we're completely
9300          * safe.
9301          */
9302         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
9303         trans->block_rsv = rsv;
9304
9305         while (1) {
9306                 ret = btrfs_truncate_inode_items(trans, root, inode,
9307                                                  inode->i_size,
9308                                                  BTRFS_EXTENT_DATA_KEY);
9309                 trans->block_rsv = &fs_info->trans_block_rsv;
9310                 if (ret != -ENOSPC && ret != -EAGAIN) {
9311                         err = ret;
9312                         break;
9313                 }
9314
9315                 ret = btrfs_update_inode(trans, root, inode);
9316                 if (ret) {
9317                         err = ret;
9318                         break;
9319                 }
9320
9321                 btrfs_end_transaction(trans);
9322                 btrfs_btree_balance_dirty(fs_info);
9323
9324                 trans = btrfs_start_transaction(root, 2);
9325                 if (IS_ERR(trans)) {
9326                         ret = err = PTR_ERR(trans);
9327                         trans = NULL;
9328                         break;
9329                 }
9330
9331                 btrfs_block_rsv_release(fs_info, rsv, -1);
9332                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
9333                                               rsv, min_size, 0);
9334                 BUG_ON(ret);    /* shouldn't happen */
9335                 trans->block_rsv = rsv;
9336         }
9337
9338         /*
9339          * We can't call btrfs_truncate_block inside a trans handle as we could
9340          * deadlock with freeze, if we got NEED_TRUNCATE_BLOCK then we know
9341          * we've truncated everything except the last little bit, and can do
9342          * btrfs_truncate_block and then update the disk_i_size.
9343          */
9344         if (ret == NEED_TRUNCATE_BLOCK) {
9345                 btrfs_end_transaction(trans);
9346                 btrfs_btree_balance_dirty(fs_info);
9347
9348                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
9349                 if (ret)
9350                         goto out;
9351                 trans = btrfs_start_transaction(root, 1);
9352                 if (IS_ERR(trans)) {
9353                         ret = PTR_ERR(trans);
9354                         goto out;
9355                 }
9356                 btrfs_ordered_update_i_size(inode, inode->i_size, NULL);
9357         }
9358
9359         if (ret == 0 && inode->i_nlink > 0) {
9360                 trans->block_rsv = root->orphan_block_rsv;
9361                 ret = btrfs_orphan_del(trans, BTRFS_I(inode));
9362                 if (ret)
9363                         err = ret;
9364         }
9365
9366         if (trans) {
9367                 trans->block_rsv = &fs_info->trans_block_rsv;
9368                 ret = btrfs_update_inode(trans, root, inode);
9369                 if (ret && !err)
9370                         err = ret;
9371
9372                 ret = btrfs_end_transaction(trans);
9373                 btrfs_btree_balance_dirty(fs_info);
9374         }
9375 out:
9376         btrfs_free_block_rsv(fs_info, rsv);
9377
9378         if (ret && !err)
9379                 err = ret;
9380
9381         return err;
9382 }
9383
9384 /*
9385  * create a new subvolume directory/inode (helper for the ioctl).
9386  */
9387 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9388                              struct btrfs_root *new_root,
9389                              struct btrfs_root *parent_root,
9390                              u64 new_dirid)
9391 {
9392         struct inode *inode;
9393         int err;
9394         u64 index = 0;
9395
9396         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9397                                 new_dirid, new_dirid,
9398                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9399                                 &index);
9400         if (IS_ERR(inode))
9401                 return PTR_ERR(inode);
9402         inode->i_op = &btrfs_dir_inode_operations;
9403         inode->i_fop = &btrfs_dir_file_operations;
9404
9405         set_nlink(inode, 1);
9406         btrfs_i_size_write(BTRFS_I(inode), 0);
9407         unlock_new_inode(inode);
9408
9409         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9410         if (err)
9411                 btrfs_err(new_root->fs_info,
9412                           "error inheriting subvolume %llu properties: %d",
9413                           new_root->root_key.objectid, err);
9414
9415         err = btrfs_update_inode(trans, new_root, inode);
9416
9417         iput(inode);
9418         return err;
9419 }
9420
9421 struct inode *btrfs_alloc_inode(struct super_block *sb)
9422 {
9423         struct btrfs_fs_info *fs_info = btrfs_sb(sb);
9424         struct btrfs_inode *ei;
9425         struct inode *inode;
9426
9427         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9428         if (!ei)
9429                 return NULL;
9430
9431         ei->root = NULL;
9432         ei->generation = 0;
9433         ei->last_trans = 0;
9434         ei->last_sub_trans = 0;
9435         ei->logged_trans = 0;
9436         ei->delalloc_bytes = 0;
9437         ei->new_delalloc_bytes = 0;
9438         ei->defrag_bytes = 0;
9439         ei->disk_i_size = 0;
9440         ei->flags = 0;
9441         ei->csum_bytes = 0;
9442         ei->index_cnt = (u64)-1;
9443         ei->dir_index = 0;
9444         ei->last_unlink_trans = 0;
9445         ei->last_log_commit = 0;
9446         ei->delayed_iput_count = 0;
9447
9448         spin_lock_init(&ei->lock);
9449         ei->outstanding_extents = 0;
9450         if (sb->s_magic != BTRFS_TEST_MAGIC)
9451                 btrfs_init_metadata_block_rsv(fs_info, &ei->block_rsv,
9452                                               BTRFS_BLOCK_RSV_DELALLOC);
9453         ei->runtime_flags = 0;
9454         ei->prop_compress = BTRFS_COMPRESS_NONE;
9455         ei->defrag_compress = BTRFS_COMPRESS_NONE;
9456
9457         ei->delayed_node = NULL;
9458
9459         ei->i_otime.tv_sec = 0;
9460         ei->i_otime.tv_nsec = 0;
9461
9462         inode = &ei->vfs_inode;
9463         extent_map_tree_init(&ei->extent_tree);
9464         extent_io_tree_init(&ei->io_tree, inode);
9465         extent_io_tree_init(&ei->io_failure_tree, inode);
9466         ei->io_tree.track_uptodate = 1;
9467         ei->io_failure_tree.track_uptodate = 1;
9468         atomic_set(&ei->sync_writers, 0);
9469         mutex_init(&ei->log_mutex);
9470         mutex_init(&ei->delalloc_mutex);
9471         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9472         INIT_LIST_HEAD(&ei->delalloc_inodes);
9473         INIT_LIST_HEAD(&ei->delayed_iput);
9474         RB_CLEAR_NODE(&ei->rb_node);
9475         init_rwsem(&ei->dio_sem);
9476
9477         return inode;
9478 }
9479
9480 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9481 void btrfs_test_destroy_inode(struct inode *inode)
9482 {
9483         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9484         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9485 }
9486 #endif
9487
9488 static void btrfs_i_callback(struct rcu_head *head)
9489 {
9490         struct inode *inode = container_of(head, struct inode, i_rcu);
9491         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9492 }
9493
9494 void btrfs_destroy_inode(struct inode *inode)
9495 {
9496         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
9497         struct btrfs_ordered_extent *ordered;
9498         struct btrfs_root *root = BTRFS_I(inode)->root;
9499
9500         WARN_ON(!hlist_empty(&inode->i_dentry));
9501         WARN_ON(inode->i_data.nrpages);
9502         WARN_ON(BTRFS_I(inode)->block_rsv.reserved);
9503         WARN_ON(BTRFS_I(inode)->block_rsv.size);
9504         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9505         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9506         WARN_ON(BTRFS_I(inode)->new_delalloc_bytes);
9507         WARN_ON(BTRFS_I(inode)->csum_bytes);
9508         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9509
9510         /*
9511          * This can happen where we create an inode, but somebody else also
9512          * created the same inode and we need to destroy the one we already
9513          * created.
9514          */
9515         if (!root)
9516                 goto free;
9517
9518         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9519                      &BTRFS_I(inode)->runtime_flags)) {
9520                 btrfs_info(fs_info, "inode %llu still on the orphan list",
9521                            btrfs_ino(BTRFS_I(inode)));
9522                 atomic_dec(&root->orphan_inodes);
9523         }
9524
9525         while (1) {
9526                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9527                 if (!ordered)
9528                         break;
9529                 else {
9530                         btrfs_err(fs_info,
9531                                   "found ordered extent %llu %llu on inode cleanup",
9532                                   ordered->file_offset, ordered->len);
9533                         btrfs_remove_ordered_extent(inode, ordered);
9534                         btrfs_put_ordered_extent(ordered);
9535                         btrfs_put_ordered_extent(ordered);
9536                 }
9537         }
9538         btrfs_qgroup_check_reserved_leak(inode);
9539         inode_tree_del(inode);
9540         btrfs_drop_extent_cache(BTRFS_I(inode), 0, (u64)-1, 0);
9541 free:
9542         call_rcu(&inode->i_rcu, btrfs_i_callback);
9543 }
9544
9545 int btrfs_drop_inode(struct inode *inode)
9546 {
9547         struct btrfs_root *root = BTRFS_I(inode)->root;
9548
9549         if (root == NULL)
9550                 return 1;
9551
9552         /* the snap/subvol tree is on deleting */
9553         if (btrfs_root_refs(&root->root_item) == 0)
9554                 return 1;
9555         else
9556                 return generic_drop_inode(inode);
9557 }
9558
9559 static void init_once(void *foo)
9560 {
9561         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9562
9563         inode_init_once(&ei->vfs_inode);
9564 }
9565
9566 void btrfs_destroy_cachep(void)
9567 {
9568         /*
9569          * Make sure all delayed rcu free inodes are flushed before we
9570          * destroy cache.
9571          */
9572         rcu_barrier();
9573         kmem_cache_destroy(btrfs_inode_cachep);
9574         kmem_cache_destroy(btrfs_trans_handle_cachep);
9575         kmem_cache_destroy(btrfs_path_cachep);
9576         kmem_cache_destroy(btrfs_free_space_cachep);
9577 }
9578
9579 int btrfs_init_cachep(void)
9580 {
9581         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9582                         sizeof(struct btrfs_inode), 0,
9583                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD | SLAB_ACCOUNT,
9584                         init_once);
9585         if (!btrfs_inode_cachep)
9586                 goto fail;
9587
9588         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9589                         sizeof(struct btrfs_trans_handle), 0,
9590                         SLAB_TEMPORARY | SLAB_MEM_SPREAD, NULL);
9591         if (!btrfs_trans_handle_cachep)
9592                 goto fail;
9593
9594         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9595                         sizeof(struct btrfs_path), 0,
9596                         SLAB_MEM_SPREAD, NULL);
9597         if (!btrfs_path_cachep)
9598                 goto fail;
9599
9600         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9601                         sizeof(struct btrfs_free_space), 0,
9602                         SLAB_MEM_SPREAD, NULL);
9603         if (!btrfs_free_space_cachep)
9604                 goto fail;
9605
9606         return 0;
9607 fail:
9608         btrfs_destroy_cachep();
9609         return -ENOMEM;
9610 }
9611
9612 static int btrfs_getattr(const struct path *path, struct kstat *stat,
9613                          u32 request_mask, unsigned int flags)
9614 {
9615         u64 delalloc_bytes;
9616         struct inode *inode = d_inode(path->dentry);
9617         u32 blocksize = inode->i_sb->s_blocksize;
9618         u32 bi_flags = BTRFS_I(inode)->flags;
9619
9620         stat->result_mask |= STATX_BTIME;
9621         stat->btime.tv_sec = BTRFS_I(inode)->i_otime.tv_sec;
9622         stat->btime.tv_nsec = BTRFS_I(inode)->i_otime.tv_nsec;
9623         if (bi_flags & BTRFS_INODE_APPEND)
9624                 stat->attributes |= STATX_ATTR_APPEND;
9625         if (bi_flags & BTRFS_INODE_COMPRESS)
9626                 stat->attributes |= STATX_ATTR_COMPRESSED;
9627         if (bi_flags & BTRFS_INODE_IMMUTABLE)
9628                 stat->attributes |= STATX_ATTR_IMMUTABLE;
9629         if (bi_flags & BTRFS_INODE_NODUMP)
9630                 stat->attributes |= STATX_ATTR_NODUMP;
9631
9632         stat->attributes_mask |= (STATX_ATTR_APPEND |
9633                                   STATX_ATTR_COMPRESSED |
9634                                   STATX_ATTR_IMMUTABLE |
9635                                   STATX_ATTR_NODUMP);
9636
9637         generic_fillattr(inode, stat);
9638         stat->dev = BTRFS_I(inode)->root->anon_dev;
9639
9640         spin_lock(&BTRFS_I(inode)->lock);
9641         delalloc_bytes = BTRFS_I(inode)->new_delalloc_bytes;
9642         spin_unlock(&BTRFS_I(inode)->lock);
9643         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9644                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9645         return 0;
9646 }
9647
9648 static int btrfs_rename_exchange(struct inode *old_dir,
9649                               struct dentry *old_dentry,
9650                               struct inode *new_dir,
9651                               struct dentry *new_dentry)
9652 {
9653         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9654         struct btrfs_trans_handle *trans;
9655         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9656         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9657         struct inode *new_inode = new_dentry->d_inode;
9658         struct inode *old_inode = old_dentry->d_inode;
9659         struct timespec ctime = current_time(old_inode);
9660         struct dentry *parent;
9661         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9662         u64 new_ino = btrfs_ino(BTRFS_I(new_inode));
9663         u64 old_idx = 0;
9664         u64 new_idx = 0;
9665         u64 root_objectid;
9666         int ret;
9667         bool root_log_pinned = false;
9668         bool dest_log_pinned = false;
9669
9670         /* we only allow rename subvolume link between subvolumes */
9671         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9672                 return -EXDEV;
9673
9674         /* close the race window with snapshot create/destroy ioctl */
9675         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9676                 down_read(&fs_info->subvol_sem);
9677         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9678                 down_read(&fs_info->subvol_sem);
9679
9680         /*
9681          * We want to reserve the absolute worst case amount of items.  So if
9682          * both inodes are subvols and we need to unlink them then that would
9683          * require 4 item modifications, but if they are both normal inodes it
9684          * would require 5 item modifications, so we'll assume their normal
9685          * inodes.  So 5 * 2 is 10, plus 2 for the new links, so 12 total items
9686          * should cover the worst case number of items we'll modify.
9687          */
9688         trans = btrfs_start_transaction(root, 12);
9689         if (IS_ERR(trans)) {
9690                 ret = PTR_ERR(trans);
9691                 goto out_notrans;
9692         }
9693
9694         /*
9695          * We need to find a free sequence number both in the source and
9696          * in the destination directory for the exchange.
9697          */
9698         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &old_idx);
9699         if (ret)
9700                 goto out_fail;
9701         ret = btrfs_set_inode_index(BTRFS_I(old_dir), &new_idx);
9702         if (ret)
9703                 goto out_fail;
9704
9705         BTRFS_I(old_inode)->dir_index = 0ULL;
9706         BTRFS_I(new_inode)->dir_index = 0ULL;
9707
9708         /* Reference for the source. */
9709         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9710                 /* force full log commit if subvolume involved. */
9711                 btrfs_set_log_full_commit(fs_info, trans);
9712         } else {
9713                 btrfs_pin_log_trans(root);
9714                 root_log_pinned = true;
9715                 ret = btrfs_insert_inode_ref(trans, dest,
9716                                              new_dentry->d_name.name,
9717                                              new_dentry->d_name.len,
9718                                              old_ino,
9719                                              btrfs_ino(BTRFS_I(new_dir)),
9720                                              old_idx);
9721                 if (ret)
9722                         goto out_fail;
9723         }
9724
9725         /* And now for the dest. */
9726         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9727                 /* force full log commit if subvolume involved. */
9728                 btrfs_set_log_full_commit(fs_info, trans);
9729         } else {
9730                 btrfs_pin_log_trans(dest);
9731                 dest_log_pinned = true;
9732                 ret = btrfs_insert_inode_ref(trans, root,
9733                                              old_dentry->d_name.name,
9734                                              old_dentry->d_name.len,
9735                                              new_ino,
9736                                              btrfs_ino(BTRFS_I(old_dir)),
9737                                              new_idx);
9738                 if (ret)
9739                         goto out_fail;
9740         }
9741
9742         /* Update inode version and ctime/mtime. */
9743         inode_inc_iversion(old_dir);
9744         inode_inc_iversion(new_dir);
9745         inode_inc_iversion(old_inode);
9746         inode_inc_iversion(new_inode);
9747         old_dir->i_ctime = old_dir->i_mtime = ctime;
9748         new_dir->i_ctime = new_dir->i_mtime = ctime;
9749         old_inode->i_ctime = ctime;
9750         new_inode->i_ctime = ctime;
9751
9752         if (old_dentry->d_parent != new_dentry->d_parent) {
9753                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
9754                                 BTRFS_I(old_inode), 1);
9755                 btrfs_record_unlink_dir(trans, BTRFS_I(new_dir),
9756                                 BTRFS_I(new_inode), 1);
9757         }
9758
9759         /* src is a subvolume */
9760         if (old_ino == BTRFS_FIRST_FREE_OBJECTID) {
9761                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9762                 ret = btrfs_unlink_subvol(trans, root, old_dir,
9763                                           root_objectid,
9764                                           old_dentry->d_name.name,
9765                                           old_dentry->d_name.len);
9766         } else { /* src is an inode */
9767                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
9768                                            BTRFS_I(old_dentry->d_inode),
9769                                            old_dentry->d_name.name,
9770                                            old_dentry->d_name.len);
9771                 if (!ret)
9772                         ret = btrfs_update_inode(trans, root, old_inode);
9773         }
9774         if (ret) {
9775                 btrfs_abort_transaction(trans, ret);
9776                 goto out_fail;
9777         }
9778
9779         /* dest is a subvolume */
9780         if (new_ino == BTRFS_FIRST_FREE_OBJECTID) {
9781                 root_objectid = BTRFS_I(new_inode)->root->root_key.objectid;
9782                 ret = btrfs_unlink_subvol(trans, dest, new_dir,
9783                                           root_objectid,
9784                                           new_dentry->d_name.name,
9785                                           new_dentry->d_name.len);
9786         } else { /* dest is an inode */
9787                 ret = __btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
9788                                            BTRFS_I(new_dentry->d_inode),
9789                                            new_dentry->d_name.name,
9790                                            new_dentry->d_name.len);
9791                 if (!ret)
9792                         ret = btrfs_update_inode(trans, dest, new_inode);
9793         }
9794         if (ret) {
9795                 btrfs_abort_transaction(trans, ret);
9796                 goto out_fail;
9797         }
9798
9799         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
9800                              new_dentry->d_name.name,
9801                              new_dentry->d_name.len, 0, old_idx);
9802         if (ret) {
9803                 btrfs_abort_transaction(trans, ret);
9804                 goto out_fail;
9805         }
9806
9807         ret = btrfs_add_link(trans, BTRFS_I(old_dir), BTRFS_I(new_inode),
9808                              old_dentry->d_name.name,
9809                              old_dentry->d_name.len, 0, new_idx);
9810         if (ret) {
9811                 btrfs_abort_transaction(trans, ret);
9812                 goto out_fail;
9813         }
9814
9815         if (old_inode->i_nlink == 1)
9816                 BTRFS_I(old_inode)->dir_index = old_idx;
9817         if (new_inode->i_nlink == 1)
9818                 BTRFS_I(new_inode)->dir_index = new_idx;
9819
9820         if (root_log_pinned) {
9821                 parent = new_dentry->d_parent;
9822                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
9823                                 parent);
9824                 btrfs_end_log_trans(root);
9825                 root_log_pinned = false;
9826         }
9827         if (dest_log_pinned) {
9828                 parent = old_dentry->d_parent;
9829                 btrfs_log_new_name(trans, BTRFS_I(new_inode), BTRFS_I(new_dir),
9830                                 parent);
9831                 btrfs_end_log_trans(dest);
9832                 dest_log_pinned = false;
9833         }
9834 out_fail:
9835         /*
9836          * If we have pinned a log and an error happened, we unpin tasks
9837          * trying to sync the log and force them to fallback to a transaction
9838          * commit if the log currently contains any of the inodes involved in
9839          * this rename operation (to ensure we do not persist a log with an
9840          * inconsistent state for any of these inodes or leading to any
9841          * inconsistencies when replayed). If the transaction was aborted, the
9842          * abortion reason is propagated to userspace when attempting to commit
9843          * the transaction. If the log does not contain any of these inodes, we
9844          * allow the tasks to sync it.
9845          */
9846         if (ret && (root_log_pinned || dest_log_pinned)) {
9847                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
9848                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
9849                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
9850                     (new_inode &&
9851                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
9852                         btrfs_set_log_full_commit(fs_info, trans);
9853
9854                 if (root_log_pinned) {
9855                         btrfs_end_log_trans(root);
9856                         root_log_pinned = false;
9857                 }
9858                 if (dest_log_pinned) {
9859                         btrfs_end_log_trans(dest);
9860                         dest_log_pinned = false;
9861                 }
9862         }
9863         ret = btrfs_end_transaction(trans);
9864 out_notrans:
9865         if (new_ino == BTRFS_FIRST_FREE_OBJECTID)
9866                 up_read(&fs_info->subvol_sem);
9867         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9868                 up_read(&fs_info->subvol_sem);
9869
9870         return ret;
9871 }
9872
9873 static int btrfs_whiteout_for_rename(struct btrfs_trans_handle *trans,
9874                                      struct btrfs_root *root,
9875                                      struct inode *dir,
9876                                      struct dentry *dentry)
9877 {
9878         int ret;
9879         struct inode *inode;
9880         u64 objectid;
9881         u64 index;
9882
9883         ret = btrfs_find_free_ino(root, &objectid);
9884         if (ret)
9885                 return ret;
9886
9887         inode = btrfs_new_inode(trans, root, dir,
9888                                 dentry->d_name.name,
9889                                 dentry->d_name.len,
9890                                 btrfs_ino(BTRFS_I(dir)),
9891                                 objectid,
9892                                 S_IFCHR | WHITEOUT_MODE,
9893                                 &index);
9894
9895         if (IS_ERR(inode)) {
9896                 ret = PTR_ERR(inode);
9897                 return ret;
9898         }
9899
9900         inode->i_op = &btrfs_special_inode_operations;
9901         init_special_inode(inode, inode->i_mode,
9902                 WHITEOUT_DEV);
9903
9904         ret = btrfs_init_inode_security(trans, inode, dir,
9905                                 &dentry->d_name);
9906         if (ret)
9907                 goto out;
9908
9909         ret = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
9910                                 BTRFS_I(inode), 0, index);
9911         if (ret)
9912                 goto out;
9913
9914         ret = btrfs_update_inode(trans, root, inode);
9915 out:
9916         unlock_new_inode(inode);
9917         if (ret)
9918                 inode_dec_link_count(inode);
9919         iput(inode);
9920
9921         return ret;
9922 }
9923
9924 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9925                            struct inode *new_dir, struct dentry *new_dentry,
9926                            unsigned int flags)
9927 {
9928         struct btrfs_fs_info *fs_info = btrfs_sb(old_dir->i_sb);
9929         struct btrfs_trans_handle *trans;
9930         unsigned int trans_num_items;
9931         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9932         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9933         struct inode *new_inode = d_inode(new_dentry);
9934         struct inode *old_inode = d_inode(old_dentry);
9935         u64 index = 0;
9936         u64 root_objectid;
9937         int ret;
9938         u64 old_ino = btrfs_ino(BTRFS_I(old_inode));
9939         bool log_pinned = false;
9940
9941         if (btrfs_ino(BTRFS_I(new_dir)) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9942                 return -EPERM;
9943
9944         /* we only allow rename subvolume link between subvolumes */
9945         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9946                 return -EXDEV;
9947
9948         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9949             (new_inode && btrfs_ino(BTRFS_I(new_inode)) == BTRFS_FIRST_FREE_OBJECTID))
9950                 return -ENOTEMPTY;
9951
9952         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9953             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9954                 return -ENOTEMPTY;
9955
9956
9957         /* check for collisions, even if the  name isn't there */
9958         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9959                              new_dentry->d_name.name,
9960                              new_dentry->d_name.len);
9961
9962         if (ret) {
9963                 if (ret == -EEXIST) {
9964                         /* we shouldn't get
9965                          * eexist without a new_inode */
9966                         if (WARN_ON(!new_inode)) {
9967                                 return ret;
9968                         }
9969                 } else {
9970                         /* maybe -EOVERFLOW */
9971                         return ret;
9972                 }
9973         }
9974         ret = 0;
9975
9976         /*
9977          * we're using rename to replace one file with another.  Start IO on it
9978          * now so  we don't add too much work to the end of the transaction
9979          */
9980         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9981                 filemap_flush(old_inode->i_mapping);
9982
9983         /* close the racy window with snapshot create/destroy ioctl */
9984         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9985                 down_read(&fs_info->subvol_sem);
9986         /*
9987          * We want to reserve the absolute worst case amount of items.  So if
9988          * both inodes are subvols and we need to unlink them then that would
9989          * require 4 item modifications, but if they are both normal inodes it
9990          * would require 5 item modifications, so we'll assume they are normal
9991          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9992          * should cover the worst case number of items we'll modify.
9993          * If our rename has the whiteout flag, we need more 5 units for the
9994          * new inode (1 inode item, 1 inode ref, 2 dir items and 1 xattr item
9995          * when selinux is enabled).
9996          */
9997         trans_num_items = 11;
9998         if (flags & RENAME_WHITEOUT)
9999                 trans_num_items += 5;
10000         trans = btrfs_start_transaction(root, trans_num_items);
10001         if (IS_ERR(trans)) {
10002                 ret = PTR_ERR(trans);
10003                 goto out_notrans;
10004         }
10005
10006         if (dest != root)
10007                 btrfs_record_root_in_trans(trans, dest);
10008
10009         ret = btrfs_set_inode_index(BTRFS_I(new_dir), &index);
10010         if (ret)
10011                 goto out_fail;
10012
10013         BTRFS_I(old_inode)->dir_index = 0ULL;
10014         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10015                 /* force full log commit if subvolume involved. */
10016                 btrfs_set_log_full_commit(fs_info, trans);
10017         } else {
10018                 btrfs_pin_log_trans(root);
10019                 log_pinned = true;
10020                 ret = btrfs_insert_inode_ref(trans, dest,
10021                                              new_dentry->d_name.name,
10022                                              new_dentry->d_name.len,
10023                                              old_ino,
10024                                              btrfs_ino(BTRFS_I(new_dir)), index);
10025                 if (ret)
10026                         goto out_fail;
10027         }
10028
10029         inode_inc_iversion(old_dir);
10030         inode_inc_iversion(new_dir);
10031         inode_inc_iversion(old_inode);
10032         old_dir->i_ctime = old_dir->i_mtime =
10033         new_dir->i_ctime = new_dir->i_mtime =
10034         old_inode->i_ctime = current_time(old_dir);
10035
10036         if (old_dentry->d_parent != new_dentry->d_parent)
10037                 btrfs_record_unlink_dir(trans, BTRFS_I(old_dir),
10038                                 BTRFS_I(old_inode), 1);
10039
10040         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
10041                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
10042                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
10043                                         old_dentry->d_name.name,
10044                                         old_dentry->d_name.len);
10045         } else {
10046                 ret = __btrfs_unlink_inode(trans, root, BTRFS_I(old_dir),
10047                                         BTRFS_I(d_inode(old_dentry)),
10048                                         old_dentry->d_name.name,
10049                                         old_dentry->d_name.len);
10050                 if (!ret)
10051                         ret = btrfs_update_inode(trans, root, old_inode);
10052         }
10053         if (ret) {
10054                 btrfs_abort_transaction(trans, ret);
10055                 goto out_fail;
10056         }
10057
10058         if (new_inode) {
10059                 inode_inc_iversion(new_inode);
10060                 new_inode->i_ctime = current_time(new_inode);
10061                 if (unlikely(btrfs_ino(BTRFS_I(new_inode)) ==
10062                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
10063                         root_objectid = BTRFS_I(new_inode)->location.objectid;
10064                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
10065                                                 root_objectid,
10066                                                 new_dentry->d_name.name,
10067                                                 new_dentry->d_name.len);
10068                         BUG_ON(new_inode->i_nlink == 0);
10069                 } else {
10070                         ret = btrfs_unlink_inode(trans, dest, BTRFS_I(new_dir),
10071                                                  BTRFS_I(d_inode(new_dentry)),
10072                                                  new_dentry->d_name.name,
10073                                                  new_dentry->d_name.len);
10074                 }
10075                 if (!ret && new_inode->i_nlink == 0)
10076                         ret = btrfs_orphan_add(trans,
10077                                         BTRFS_I(d_inode(new_dentry)));
10078                 if (ret) {
10079                         btrfs_abort_transaction(trans, ret);
10080                         goto out_fail;
10081                 }
10082         }
10083
10084         ret = btrfs_add_link(trans, BTRFS_I(new_dir), BTRFS_I(old_inode),
10085                              new_dentry->d_name.name,
10086                              new_dentry->d_name.len, 0, index);
10087         if (ret) {
10088                 btrfs_abort_transaction(trans, ret);
10089                 goto out_fail;
10090         }
10091
10092         if (old_inode->i_nlink == 1)
10093                 BTRFS_I(old_inode)->dir_index = index;
10094
10095         if (log_pinned) {
10096                 struct dentry *parent = new_dentry->d_parent;
10097
10098                 btrfs_log_new_name(trans, BTRFS_I(old_inode), BTRFS_I(old_dir),
10099                                 parent);
10100                 btrfs_end_log_trans(root);
10101                 log_pinned = false;
10102         }
10103
10104         if (flags & RENAME_WHITEOUT) {
10105                 ret = btrfs_whiteout_for_rename(trans, root, old_dir,
10106                                                 old_dentry);
10107
10108                 if (ret) {
10109                         btrfs_abort_transaction(trans, ret);
10110                         goto out_fail;
10111                 }
10112         }
10113 out_fail:
10114         /*
10115          * If we have pinned the log and an error happened, we unpin tasks
10116          * trying to sync the log and force them to fallback to a transaction
10117          * commit if the log currently contains any of the inodes involved in
10118          * this rename operation (to ensure we do not persist a log with an
10119          * inconsistent state for any of these inodes or leading to any
10120          * inconsistencies when replayed). If the transaction was aborted, the
10121          * abortion reason is propagated to userspace when attempting to commit
10122          * the transaction. If the log does not contain any of these inodes, we
10123          * allow the tasks to sync it.
10124          */
10125         if (ret && log_pinned) {
10126                 if (btrfs_inode_in_log(BTRFS_I(old_dir), fs_info->generation) ||
10127                     btrfs_inode_in_log(BTRFS_I(new_dir), fs_info->generation) ||
10128                     btrfs_inode_in_log(BTRFS_I(old_inode), fs_info->generation) ||
10129                     (new_inode &&
10130                      btrfs_inode_in_log(BTRFS_I(new_inode), fs_info->generation)))
10131                         btrfs_set_log_full_commit(fs_info, trans);
10132
10133                 btrfs_end_log_trans(root);
10134                 log_pinned = false;
10135         }
10136         btrfs_end_transaction(trans);
10137 out_notrans:
10138         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
10139                 up_read(&fs_info->subvol_sem);
10140
10141         return ret;
10142 }
10143
10144 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
10145                          struct inode *new_dir, struct dentry *new_dentry,
10146                          unsigned int flags)
10147 {
10148         if (flags & ~(RENAME_NOREPLACE | RENAME_EXCHANGE | RENAME_WHITEOUT))
10149                 return -EINVAL;
10150
10151         if (flags & RENAME_EXCHANGE)
10152                 return btrfs_rename_exchange(old_dir, old_dentry, new_dir,
10153                                           new_dentry);
10154
10155         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry, flags);
10156 }
10157
10158 static void btrfs_run_delalloc_work(struct btrfs_work *work)
10159 {
10160         struct btrfs_delalloc_work *delalloc_work;
10161         struct inode *inode;
10162
10163         delalloc_work = container_of(work, struct btrfs_delalloc_work,
10164                                      work);
10165         inode = delalloc_work->inode;
10166         filemap_flush(inode->i_mapping);
10167         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
10168                                 &BTRFS_I(inode)->runtime_flags))
10169                 filemap_flush(inode->i_mapping);
10170
10171         if (delalloc_work->delay_iput)
10172                 btrfs_add_delayed_iput(inode);
10173         else
10174                 iput(inode);
10175         complete(&delalloc_work->completion);
10176 }
10177
10178 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
10179                                                     int delay_iput)
10180 {
10181         struct btrfs_delalloc_work *work;
10182
10183         work = kmalloc(sizeof(*work), GFP_NOFS);
10184         if (!work)
10185                 return NULL;
10186
10187         init_completion(&work->completion);
10188         INIT_LIST_HEAD(&work->list);
10189         work->inode = inode;
10190         work->delay_iput = delay_iput;
10191         WARN_ON_ONCE(!inode);
10192         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
10193                         btrfs_run_delalloc_work, NULL, NULL);
10194
10195         return work;
10196 }
10197
10198 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
10199 {
10200         wait_for_completion(&work->completion);
10201         kfree(work);
10202 }
10203
10204 /*
10205  * some fairly slow code that needs optimization. This walks the list
10206  * of all the inodes with pending delalloc and forces them to disk.
10207  */
10208 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
10209                                    int nr)
10210 {
10211         struct btrfs_inode *binode;
10212         struct inode *inode;
10213         struct btrfs_delalloc_work *work, *next;
10214         struct list_head works;
10215         struct list_head splice;
10216         int ret = 0;
10217
10218         INIT_LIST_HEAD(&works);
10219         INIT_LIST_HEAD(&splice);
10220
10221         mutex_lock(&root->delalloc_mutex);
10222         spin_lock(&root->delalloc_lock);
10223         list_splice_init(&root->delalloc_inodes, &splice);
10224         while (!list_empty(&splice)) {
10225                 binode = list_entry(splice.next, struct btrfs_inode,
10226                                     delalloc_inodes);
10227
10228                 list_move_tail(&binode->delalloc_inodes,
10229                                &root->delalloc_inodes);
10230                 inode = igrab(&binode->vfs_inode);
10231                 if (!inode) {
10232                         cond_resched_lock(&root->delalloc_lock);
10233                         continue;
10234                 }
10235                 spin_unlock(&root->delalloc_lock);
10236
10237                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
10238                 if (!work) {
10239                         if (delay_iput)
10240                                 btrfs_add_delayed_iput(inode);
10241                         else
10242                                 iput(inode);
10243                         ret = -ENOMEM;
10244                         goto out;
10245                 }
10246                 list_add_tail(&work->list, &works);
10247                 btrfs_queue_work(root->fs_info->flush_workers,
10248                                  &work->work);
10249                 ret++;
10250                 if (nr != -1 && ret >= nr)
10251                         goto out;
10252                 cond_resched();
10253                 spin_lock(&root->delalloc_lock);
10254         }
10255         spin_unlock(&root->delalloc_lock);
10256
10257 out:
10258         list_for_each_entry_safe(work, next, &works, list) {
10259                 list_del_init(&work->list);
10260                 btrfs_wait_and_free_delalloc_work(work);
10261         }
10262
10263         if (!list_empty_careful(&splice)) {
10264                 spin_lock(&root->delalloc_lock);
10265                 list_splice_tail(&splice, &root->delalloc_inodes);
10266                 spin_unlock(&root->delalloc_lock);
10267         }
10268         mutex_unlock(&root->delalloc_mutex);
10269         return ret;
10270 }
10271
10272 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
10273 {
10274         struct btrfs_fs_info *fs_info = root->fs_info;
10275         int ret;
10276
10277         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10278                 return -EROFS;
10279
10280         ret = __start_delalloc_inodes(root, delay_iput, -1);
10281         if (ret > 0)
10282                 ret = 0;
10283         return ret;
10284 }
10285
10286 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
10287                                int nr)
10288 {
10289         struct btrfs_root *root;
10290         struct list_head splice;
10291         int ret;
10292
10293         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
10294                 return -EROFS;
10295
10296         INIT_LIST_HEAD(&splice);
10297
10298         mutex_lock(&fs_info->delalloc_root_mutex);
10299         spin_lock(&fs_info->delalloc_root_lock);
10300         list_splice_init(&fs_info->delalloc_roots, &splice);
10301         while (!list_empty(&splice) && nr) {
10302                 root = list_first_entry(&splice, struct btrfs_root,
10303                                         delalloc_root);
10304                 root = btrfs_grab_fs_root(root);
10305                 BUG_ON(!root);
10306                 list_move_tail(&root->delalloc_root,
10307                                &fs_info->delalloc_roots);
10308                 spin_unlock(&fs_info->delalloc_root_lock);
10309
10310                 ret = __start_delalloc_inodes(root, delay_iput, nr);
10311                 btrfs_put_fs_root(root);
10312                 if (ret < 0)
10313                         goto out;
10314
10315                 if (nr != -1) {
10316                         nr -= ret;
10317                         WARN_ON(nr < 0);
10318                 }
10319                 spin_lock(&fs_info->delalloc_root_lock);
10320         }
10321         spin_unlock(&fs_info->delalloc_root_lock);
10322
10323         ret = 0;
10324 out:
10325         if (!list_empty_careful(&splice)) {
10326                 spin_lock(&fs_info->delalloc_root_lock);
10327                 list_splice_tail(&splice, &fs_info->delalloc_roots);
10328                 spin_unlock(&fs_info->delalloc_root_lock);
10329         }
10330         mutex_unlock(&fs_info->delalloc_root_mutex);
10331         return ret;
10332 }
10333
10334 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
10335                          const char *symname)
10336 {
10337         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10338         struct btrfs_trans_handle *trans;
10339         struct btrfs_root *root = BTRFS_I(dir)->root;
10340         struct btrfs_path *path;
10341         struct btrfs_key key;
10342         struct inode *inode = NULL;
10343         int err;
10344         int drop_inode = 0;
10345         u64 objectid;
10346         u64 index = 0;
10347         int name_len;
10348         int datasize;
10349         unsigned long ptr;
10350         struct btrfs_file_extent_item *ei;
10351         struct extent_buffer *leaf;
10352
10353         name_len = strlen(symname);
10354         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(fs_info))
10355                 return -ENAMETOOLONG;
10356
10357         /*
10358          * 2 items for inode item and ref
10359          * 2 items for dir items
10360          * 1 item for updating parent inode item
10361          * 1 item for the inline extent item
10362          * 1 item for xattr if selinux is on
10363          */
10364         trans = btrfs_start_transaction(root, 7);
10365         if (IS_ERR(trans))
10366                 return PTR_ERR(trans);
10367
10368         err = btrfs_find_free_ino(root, &objectid);
10369         if (err)
10370                 goto out_unlock;
10371
10372         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
10373                                 dentry->d_name.len, btrfs_ino(BTRFS_I(dir)),
10374                                 objectid, S_IFLNK|S_IRWXUGO, &index);
10375         if (IS_ERR(inode)) {
10376                 err = PTR_ERR(inode);
10377                 goto out_unlock;
10378         }
10379
10380         /*
10381         * If the active LSM wants to access the inode during
10382         * d_instantiate it needs these. Smack checks to see
10383         * if the filesystem supports xattrs by looking at the
10384         * ops vector.
10385         */
10386         inode->i_fop = &btrfs_file_operations;
10387         inode->i_op = &btrfs_file_inode_operations;
10388         inode->i_mapping->a_ops = &btrfs_aops;
10389         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10390
10391         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
10392         if (err)
10393                 goto out_unlock_inode;
10394
10395         path = btrfs_alloc_path();
10396         if (!path) {
10397                 err = -ENOMEM;
10398                 goto out_unlock_inode;
10399         }
10400         key.objectid = btrfs_ino(BTRFS_I(inode));
10401         key.offset = 0;
10402         key.type = BTRFS_EXTENT_DATA_KEY;
10403         datasize = btrfs_file_extent_calc_inline_size(name_len);
10404         err = btrfs_insert_empty_item(trans, root, path, &key,
10405                                       datasize);
10406         if (err) {
10407                 btrfs_free_path(path);
10408                 goto out_unlock_inode;
10409         }
10410         leaf = path->nodes[0];
10411         ei = btrfs_item_ptr(leaf, path->slots[0],
10412                             struct btrfs_file_extent_item);
10413         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
10414         btrfs_set_file_extent_type(leaf, ei,
10415                                    BTRFS_FILE_EXTENT_INLINE);
10416         btrfs_set_file_extent_encryption(leaf, ei, 0);
10417         btrfs_set_file_extent_compression(leaf, ei, 0);
10418         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
10419         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
10420
10421         ptr = btrfs_file_extent_inline_start(ei);
10422         write_extent_buffer(leaf, symname, ptr, name_len);
10423         btrfs_mark_buffer_dirty(leaf);
10424         btrfs_free_path(path);
10425
10426         inode->i_op = &btrfs_symlink_inode_operations;
10427         inode_nohighmem(inode);
10428         inode->i_mapping->a_ops = &btrfs_symlink_aops;
10429         inode_set_bytes(inode, name_len);
10430         btrfs_i_size_write(BTRFS_I(inode), name_len);
10431         err = btrfs_update_inode(trans, root, inode);
10432         /*
10433          * Last step, add directory indexes for our symlink inode. This is the
10434          * last step to avoid extra cleanup of these indexes if an error happens
10435          * elsewhere above.
10436          */
10437         if (!err)
10438                 err = btrfs_add_nondir(trans, BTRFS_I(dir), dentry,
10439                                 BTRFS_I(inode), 0, index);
10440         if (err) {
10441                 drop_inode = 1;
10442                 goto out_unlock_inode;
10443         }
10444
10445         unlock_new_inode(inode);
10446         d_instantiate(dentry, inode);
10447
10448 out_unlock:
10449         btrfs_end_transaction(trans);
10450         if (drop_inode) {
10451                 inode_dec_link_count(inode);
10452                 iput(inode);
10453         }
10454         btrfs_btree_balance_dirty(fs_info);
10455         return err;
10456
10457 out_unlock_inode:
10458         drop_inode = 1;
10459         unlock_new_inode(inode);
10460         goto out_unlock;
10461 }
10462
10463 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
10464                                        u64 start, u64 num_bytes, u64 min_size,
10465                                        loff_t actual_len, u64 *alloc_hint,
10466                                        struct btrfs_trans_handle *trans)
10467 {
10468         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
10469         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
10470         struct extent_map *em;
10471         struct btrfs_root *root = BTRFS_I(inode)->root;
10472         struct btrfs_key ins;
10473         u64 cur_offset = start;
10474         u64 i_size;
10475         u64 cur_bytes;
10476         u64 last_alloc = (u64)-1;
10477         int ret = 0;
10478         bool own_trans = true;
10479         u64 end = start + num_bytes - 1;
10480
10481         if (trans)
10482                 own_trans = false;
10483         while (num_bytes > 0) {
10484                 if (own_trans) {
10485                         trans = btrfs_start_transaction(root, 3);
10486                         if (IS_ERR(trans)) {
10487                                 ret = PTR_ERR(trans);
10488                                 break;
10489                         }
10490                 }
10491
10492                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
10493                 cur_bytes = max(cur_bytes, min_size);
10494                 /*
10495                  * If we are severely fragmented we could end up with really
10496                  * small allocations, so if the allocator is returning small
10497                  * chunks lets make its job easier by only searching for those
10498                  * sized chunks.
10499                  */
10500                 cur_bytes = min(cur_bytes, last_alloc);
10501                 ret = btrfs_reserve_extent(root, cur_bytes, cur_bytes,
10502                                 min_size, 0, *alloc_hint, &ins, 1, 0);
10503                 if (ret) {
10504                         if (own_trans)
10505                                 btrfs_end_transaction(trans);
10506                         break;
10507                 }
10508                 btrfs_dec_block_group_reservations(fs_info, ins.objectid);
10509
10510                 last_alloc = ins.offset;
10511                 ret = insert_reserved_file_extent(trans, inode,
10512                                                   cur_offset, ins.objectid,
10513                                                   ins.offset, ins.offset,
10514                                                   ins.offset, 0, 0, 0,
10515                                                   BTRFS_FILE_EXTENT_PREALLOC);
10516                 if (ret) {
10517                         btrfs_free_reserved_extent(fs_info, ins.objectid,
10518                                                    ins.offset, 0);
10519                         btrfs_abort_transaction(trans, ret);
10520                         if (own_trans)
10521                                 btrfs_end_transaction(trans);
10522                         break;
10523                 }
10524
10525                 btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10526                                         cur_offset + ins.offset -1, 0);
10527
10528                 em = alloc_extent_map();
10529                 if (!em) {
10530                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
10531                                 &BTRFS_I(inode)->runtime_flags);
10532                         goto next;
10533                 }
10534
10535                 em->start = cur_offset;
10536                 em->orig_start = cur_offset;
10537                 em->len = ins.offset;
10538                 em->block_start = ins.objectid;
10539                 em->block_len = ins.offset;
10540                 em->orig_block_len = ins.offset;
10541                 em->ram_bytes = ins.offset;
10542                 em->bdev = fs_info->fs_devices->latest_bdev;
10543                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
10544                 em->generation = trans->transid;
10545
10546                 while (1) {
10547                         write_lock(&em_tree->lock);
10548                         ret = add_extent_mapping(em_tree, em, 1);
10549                         write_unlock(&em_tree->lock);
10550                         if (ret != -EEXIST)
10551                                 break;
10552                         btrfs_drop_extent_cache(BTRFS_I(inode), cur_offset,
10553                                                 cur_offset + ins.offset - 1,
10554                                                 0);
10555                 }
10556                 free_extent_map(em);
10557 next:
10558                 num_bytes -= ins.offset;
10559                 cur_offset += ins.offset;
10560                 *alloc_hint = ins.objectid + ins.offset;
10561
10562                 inode_inc_iversion(inode);
10563                 inode->i_ctime = current_time(inode);
10564                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
10565                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
10566                     (actual_len > inode->i_size) &&
10567                     (cur_offset > inode->i_size)) {
10568                         if (cur_offset > actual_len)
10569                                 i_size = actual_len;
10570                         else
10571                                 i_size = cur_offset;
10572                         i_size_write(inode, i_size);
10573                         btrfs_ordered_update_i_size(inode, i_size, NULL);
10574                 }
10575
10576                 ret = btrfs_update_inode(trans, root, inode);
10577
10578                 if (ret) {
10579                         btrfs_abort_transaction(trans, ret);
10580                         if (own_trans)
10581                                 btrfs_end_transaction(trans);
10582                         break;
10583                 }
10584
10585                 if (own_trans)
10586                         btrfs_end_transaction(trans);
10587         }
10588         if (cur_offset < end)
10589                 btrfs_free_reserved_data_space(inode, NULL, cur_offset,
10590                         end - cur_offset + 1);
10591         return ret;
10592 }
10593
10594 int btrfs_prealloc_file_range(struct inode *inode, int mode,
10595                               u64 start, u64 num_bytes, u64 min_size,
10596                               loff_t actual_len, u64 *alloc_hint)
10597 {
10598         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10599                                            min_size, actual_len, alloc_hint,
10600                                            NULL);
10601 }
10602
10603 int btrfs_prealloc_file_range_trans(struct inode *inode,
10604                                     struct btrfs_trans_handle *trans, int mode,
10605                                     u64 start, u64 num_bytes, u64 min_size,
10606                                     loff_t actual_len, u64 *alloc_hint)
10607 {
10608         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
10609                                            min_size, actual_len, alloc_hint, trans);
10610 }
10611
10612 static int btrfs_set_page_dirty(struct page *page)
10613 {
10614         return __set_page_dirty_nobuffers(page);
10615 }
10616
10617 static int btrfs_permission(struct inode *inode, int mask)
10618 {
10619         struct btrfs_root *root = BTRFS_I(inode)->root;
10620         umode_t mode = inode->i_mode;
10621
10622         if (mask & MAY_WRITE &&
10623             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
10624                 if (btrfs_root_readonly(root))
10625                         return -EROFS;
10626                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
10627                         return -EACCES;
10628         }
10629         return generic_permission(inode, mask);
10630 }
10631
10632 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
10633 {
10634         struct btrfs_fs_info *fs_info = btrfs_sb(dir->i_sb);
10635         struct btrfs_trans_handle *trans;
10636         struct btrfs_root *root = BTRFS_I(dir)->root;
10637         struct inode *inode = NULL;
10638         u64 objectid;
10639         u64 index;
10640         int ret = 0;
10641
10642         /*
10643          * 5 units required for adding orphan entry
10644          */
10645         trans = btrfs_start_transaction(root, 5);
10646         if (IS_ERR(trans))
10647                 return PTR_ERR(trans);
10648
10649         ret = btrfs_find_free_ino(root, &objectid);
10650         if (ret)
10651                 goto out;
10652
10653         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
10654                         btrfs_ino(BTRFS_I(dir)), objectid, mode, &index);
10655         if (IS_ERR(inode)) {
10656                 ret = PTR_ERR(inode);
10657                 inode = NULL;
10658                 goto out;
10659         }
10660
10661         inode->i_fop = &btrfs_file_operations;
10662         inode->i_op = &btrfs_file_inode_operations;
10663
10664         inode->i_mapping->a_ops = &btrfs_aops;
10665         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
10666
10667         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
10668         if (ret)
10669                 goto out_inode;
10670
10671         ret = btrfs_update_inode(trans, root, inode);
10672         if (ret)
10673                 goto out_inode;
10674         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
10675         if (ret)
10676                 goto out_inode;
10677
10678         /*
10679          * We set number of links to 0 in btrfs_new_inode(), and here we set
10680          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
10681          * through:
10682          *
10683          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
10684          */
10685         set_nlink(inode, 1);
10686         unlock_new_inode(inode);
10687         d_tmpfile(dentry, inode);
10688         mark_inode_dirty(inode);
10689
10690 out:
10691         btrfs_end_transaction(trans);
10692         if (ret)
10693                 iput(inode);
10694         btrfs_balance_delayed_items(fs_info);
10695         btrfs_btree_balance_dirty(fs_info);
10696         return ret;
10697
10698 out_inode:
10699         unlock_new_inode(inode);
10700         goto out;
10701
10702 }
10703
10704 __attribute__((const))
10705 static int btrfs_readpage_io_failed_hook(struct page *page, int failed_mirror)
10706 {
10707         return -EAGAIN;
10708 }
10709
10710 static struct btrfs_fs_info *iotree_fs_info(void *private_data)
10711 {
10712         struct inode *inode = private_data;
10713         return btrfs_sb(inode->i_sb);
10714 }
10715
10716 static void btrfs_check_extent_io_range(void *private_data, const char *caller,
10717                                         u64 start, u64 end)
10718 {
10719         struct inode *inode = private_data;
10720         u64 isize;
10721
10722         isize = i_size_read(inode);
10723         if (end >= PAGE_SIZE && (end % 2) == 0 && end != isize - 1) {
10724                 btrfs_debug_rl(BTRFS_I(inode)->root->fs_info,
10725                     "%s: ino %llu isize %llu odd range [%llu,%llu]",
10726                         caller, btrfs_ino(BTRFS_I(inode)), isize, start, end);
10727         }
10728 }
10729
10730 void btrfs_set_range_writeback(void *private_data, u64 start, u64 end)
10731 {
10732         struct inode *inode = private_data;
10733         unsigned long index = start >> PAGE_SHIFT;
10734         unsigned long end_index = end >> PAGE_SHIFT;
10735         struct page *page;
10736
10737         while (index <= end_index) {
10738                 page = find_get_page(inode->i_mapping, index);
10739                 ASSERT(page); /* Pages should be in the extent_io_tree */
10740                 set_page_writeback(page);
10741                 put_page(page);
10742                 index++;
10743         }
10744 }
10745
10746 static const struct inode_operations btrfs_dir_inode_operations = {
10747         .getattr        = btrfs_getattr,
10748         .lookup         = btrfs_lookup,
10749         .create         = btrfs_create,
10750         .unlink         = btrfs_unlink,
10751         .link           = btrfs_link,
10752         .mkdir          = btrfs_mkdir,
10753         .rmdir          = btrfs_rmdir,
10754         .rename         = btrfs_rename2,
10755         .symlink        = btrfs_symlink,
10756         .setattr        = btrfs_setattr,
10757         .mknod          = btrfs_mknod,
10758         .listxattr      = btrfs_listxattr,
10759         .permission     = btrfs_permission,
10760         .get_acl        = btrfs_get_acl,
10761         .set_acl        = btrfs_set_acl,
10762         .update_time    = btrfs_update_time,
10763         .tmpfile        = btrfs_tmpfile,
10764 };
10765 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10766         .lookup         = btrfs_lookup,
10767         .permission     = btrfs_permission,
10768         .update_time    = btrfs_update_time,
10769 };
10770
10771 static const struct file_operations btrfs_dir_file_operations = {
10772         .llseek         = generic_file_llseek,
10773         .read           = generic_read_dir,
10774         .iterate_shared = btrfs_real_readdir,
10775         .open           = btrfs_opendir,
10776         .unlocked_ioctl = btrfs_ioctl,
10777 #ifdef CONFIG_COMPAT
10778         .compat_ioctl   = btrfs_compat_ioctl,
10779 #endif
10780         .release        = btrfs_release_file,
10781         .fsync          = btrfs_sync_file,
10782 };
10783
10784 static const struct extent_io_ops btrfs_extent_io_ops = {
10785         /* mandatory callbacks */
10786         .submit_bio_hook = btrfs_submit_bio_hook,
10787         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10788         .merge_bio_hook = btrfs_merge_bio_hook,
10789         .readpage_io_failed_hook = btrfs_readpage_io_failed_hook,
10790         .tree_fs_info = iotree_fs_info,
10791         .set_range_writeback = btrfs_set_range_writeback,
10792
10793         /* optional callbacks */
10794         .fill_delalloc = run_delalloc_range,
10795         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10796         .writepage_start_hook = btrfs_writepage_start_hook,
10797         .set_bit_hook = btrfs_set_bit_hook,
10798         .clear_bit_hook = btrfs_clear_bit_hook,
10799         .merge_extent_hook = btrfs_merge_extent_hook,
10800         .split_extent_hook = btrfs_split_extent_hook,
10801         .check_extent_io_range = btrfs_check_extent_io_range,
10802 };
10803
10804 /*
10805  * btrfs doesn't support the bmap operation because swapfiles
10806  * use bmap to make a mapping of extents in the file.  They assume
10807  * these extents won't change over the life of the file and they
10808  * use the bmap result to do IO directly to the drive.
10809  *
10810  * the btrfs bmap call would return logical addresses that aren't
10811  * suitable for IO and they also will change frequently as COW
10812  * operations happen.  So, swapfile + btrfs == corruption.
10813  *
10814  * For now we're avoiding this by dropping bmap.
10815  */
10816 static const struct address_space_operations btrfs_aops = {
10817         .readpage       = btrfs_readpage,
10818         .writepage      = btrfs_writepage,
10819         .writepages     = btrfs_writepages,
10820         .readpages      = btrfs_readpages,
10821         .direct_IO      = btrfs_direct_IO,
10822         .invalidatepage = btrfs_invalidatepage,
10823         .releasepage    = btrfs_releasepage,
10824         .set_page_dirty = btrfs_set_page_dirty,
10825         .error_remove_page = generic_error_remove_page,
10826 };
10827
10828 static const struct address_space_operations btrfs_symlink_aops = {
10829         .readpage       = btrfs_readpage,
10830         .writepage      = btrfs_writepage,
10831         .invalidatepage = btrfs_invalidatepage,
10832         .releasepage    = btrfs_releasepage,
10833 };
10834
10835 static const struct inode_operations btrfs_file_inode_operations = {
10836         .getattr        = btrfs_getattr,
10837         .setattr        = btrfs_setattr,
10838         .listxattr      = btrfs_listxattr,
10839         .permission     = btrfs_permission,
10840         .fiemap         = btrfs_fiemap,
10841         .get_acl        = btrfs_get_acl,
10842         .set_acl        = btrfs_set_acl,
10843         .update_time    = btrfs_update_time,
10844 };
10845 static const struct inode_operations btrfs_special_inode_operations = {
10846         .getattr        = btrfs_getattr,
10847         .setattr        = btrfs_setattr,
10848         .permission     = btrfs_permission,
10849         .listxattr      = btrfs_listxattr,
10850         .get_acl        = btrfs_get_acl,
10851         .set_acl        = btrfs_set_acl,
10852         .update_time    = btrfs_update_time,
10853 };
10854 static const struct inode_operations btrfs_symlink_inode_operations = {
10855         .get_link       = page_get_link,
10856         .getattr        = btrfs_getattr,
10857         .setattr        = btrfs_setattr,
10858         .permission     = btrfs_permission,
10859         .listxattr      = btrfs_listxattr,
10860         .update_time    = btrfs_update_time,
10861 };
10862
10863 const struct dentry_operations btrfs_dentry_operations = {
10864         .d_delete       = btrfs_dentry_delete,
10865         .d_release      = btrfs_dentry_release,
10866 };