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/statfs.h>
  34 #include <linux/compat.h>
  35 #include <linux/bit_spinlock.h>
  36 #include <linux/xattr.h>
  37 #include <linux/posix_acl.h>
  38 #include <linux/falloc.h>
  39 #include <linux/slab.h>
  40 #include <linux/ratelimit.h>
  41 #include <linux/mount.h>
  42 #include <linux/btrfs.h>
  43 #include <linux/blkdev.h>
  44 #include <linux/posix_acl_xattr.h>
  45 #include <linux/uio.h>
  46 #include "ctree.h"
  47 #include "disk-io.h"
  48 #include "transaction.h"
  49 #include "btrfs_inode.h"
  50 #include "print-tree.h"
  51 #include "ordered-data.h"
  52 #include "xattr.h"
  53 #include "tree-log.h"
  54 #include "volumes.h"
  55 #include "compression.h"
  56 #include "locking.h"
  57 #include "free-space-cache.h"
  58 #include "inode-map.h"
  59 #include "backref.h"
  60 #include "hash.h"
  61 #include "props.h"
  62 #include "qgroup.h"
  63
  64 struct btrfs_iget_args {
  65         struct btrfs_key *location;
  66         struct btrfs_root *root;
  67 };
  68
  69 struct btrfs_dio_data {
  70         u64 outstanding_extents;
  71         u64 reserve;
  72         u64 unsubmitted_oe_range_start;
  73         u64 unsubmitted_oe_range_end;
  74 };
  75
  76 static const struct inode_operations btrfs_dir_inode_operations;
  77 static const struct inode_operations btrfs_symlink_inode_operations;
  78 static const struct inode_operations btrfs_dir_ro_inode_operations;
  79 static const struct inode_operations btrfs_special_inode_operations;
  80 static const struct inode_operations btrfs_file_inode_operations;
  81 static const struct address_space_operations btrfs_aops;
  82 static const struct address_space_operations btrfs_symlink_aops;
  83 static const struct file_operations btrfs_dir_file_operations;
  84 static const struct extent_io_ops btrfs_extent_io_ops;
  85
  86 static struct kmem_cache *btrfs_inode_cachep;
  87 struct kmem_cache *btrfs_trans_handle_cachep;
  88 struct kmem_cache *btrfs_transaction_cachep;
  89 struct kmem_cache *btrfs_path_cachep;
  90 struct kmem_cache *btrfs_free_space_cachep;
  91
  92 #define S_SHIFT 12
  93 static const unsigned char btrfs_type_by_mode[S_IFMT >> S_SHIFT] = {
  94         [S_IFREG >> S_SHIFT]    = BTRFS_FT_REG_FILE,
  95         [S_IFDIR >> S_SHIFT]    = BTRFS_FT_DIR,
  96         [S_IFCHR >> S_SHIFT]    = BTRFS_FT_CHRDEV,
  97         [S_IFBLK >> S_SHIFT]    = BTRFS_FT_BLKDEV,
  98         [S_IFIFO >> S_SHIFT]    = BTRFS_FT_FIFO,
  99         [S_IFSOCK >> S_SHIFT]   = BTRFS_FT_SOCK,
 100         [S_IFLNK >> S_SHIFT]    = BTRFS_FT_SYMLINK,
 101 };
 102
 103 static int btrfs_setsize(struct inode *inode, struct iattr *attr);
 104 static int btrfs_truncate(struct inode *inode);
 105 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent);
 106 static noinline int cow_file_range(struct inode *inode,
 107                                    struct page *locked_page,
 108                                    u64 start, u64 end, int *page_started,
 109                                    unsigned long *nr_written, int unlock);
 110 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
 111                                            u64 len, u64 orig_start,
 112                                            u64 block_start, u64 block_len,
 113                                            u64 orig_block_len, u64 ram_bytes,
 114                                            int type);
 115
 116 static int btrfs_dirty_inode(struct inode *inode);
 117
 118 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
 119 void btrfs_test_inode_set_ops(struct inode *inode)
 120 {
 121         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
 122 }
 123 #endif
 124
 125 static int btrfs_init_inode_security(struct btrfs_trans_handle *trans,
 126                                      struct inode *inode,  struct inode *dir,
 127                                      const struct qstr *qstr)
 128 {
 129         int err;
 130
 131         err = btrfs_init_acl(trans, inode, dir);
 132         if (!err)
 133                 err = btrfs_xattr_security_init(trans, inode, dir, qstr);
 134         return err;
 135 }
 136
 137 /*
 138  * this does all the hard work for inserting an inline extent into
 139  * the btree.  The caller should have done a btrfs_drop_extents so that
 140  * no overlapping inline items exist in the btree
 141  */
 142 static int insert_inline_extent(struct btrfs_trans_handle *trans,
 143                                 struct btrfs_path *path, int extent_inserted,
 144                                 struct btrfs_root *root, struct inode *inode,
 145                                 u64 start, size_t size, size_t compressed_size,
 146                                 int compress_type,
 147                                 struct page **compressed_pages)
 148 {
 149         struct extent_buffer *leaf;
 150         struct page *page = NULL;
 151         char *kaddr;
 152         unsigned long ptr;
 153         struct btrfs_file_extent_item *ei;
 154         int err = 0;
 155         int ret;
 156         size_t cur_size = size;
 157         unsigned long offset;
 158
 159         if (compressed_size && compressed_pages)
 160                 cur_size = compressed_size;
 161
 162         inode_add_bytes(inode, size);
 163
 164         if (!extent_inserted) {
 165                 struct btrfs_key key;
 166                 size_t datasize;
 167
 168                 key.objectid = btrfs_ino(inode);
 169                 key.offset = start;
 170                 key.type = BTRFS_EXTENT_DATA_KEY;
 171
 172                 datasize = btrfs_file_extent_calc_inline_size(cur_size);
 173                 path->leave_spinning = 1;
 174                 ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                               datasize);
 176                 if (ret) {
 177                         err = ret;
 178                         goto fail;
 179                 }
 180         }
 181         leaf = path->nodes[0];
 182         ei = btrfs_item_ptr(leaf, path->slots[0],
 183                             struct btrfs_file_extent_item);
 184         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
 185         btrfs_set_file_extent_type(leaf, ei, BTRFS_FILE_EXTENT_INLINE);
 186         btrfs_set_file_extent_encryption(leaf, ei, 0);
 187         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
 188         btrfs_set_file_extent_ram_bytes(leaf, ei, size);
 189         ptr = btrfs_file_extent_inline_start(ei);
 190
 191         if (compress_type != BTRFS_COMPRESS_NONE) {
 192                 struct page *cpage;
 193                 int i = 0;
 194                 while (compressed_size > 0) {
 195                         cpage = compressed_pages[i];
 196                         cur_size = min_t(unsigned long, compressed_size,
 197                                        PAGE_CACHE_SIZE);
 198
 199                         kaddr = kmap_atomic(cpage);
 200                         write_extent_buffer(leaf, kaddr, ptr, cur_size);
 201                         kunmap_atomic(kaddr);
 202
 203                         i++;
 204                         ptr += cur_size;
 205                         compressed_size -= cur_size;
 206                 }
 207                 btrfs_set_file_extent_compression(leaf, ei,
 208                                                   compress_type);
 209         } else {
 210                 page = find_get_page(inode->i_mapping,
 211                                      start >> PAGE_CACHE_SHIFT);
 212                 btrfs_set_file_extent_compression(leaf, ei, 0);
 213                 kaddr = kmap_atomic(page);
 214                 offset = start & (PAGE_CACHE_SIZE - 1);
 215                 write_extent_buffer(leaf, kaddr + offset, ptr, size);
 216                 kunmap_atomic(kaddr);
 217                 page_cache_release(page);
 218         }
 219         btrfs_mark_buffer_dirty(leaf);
 220         btrfs_release_path(path);
 221
 222         /*
 223          * we're an inline extent, so nobody can
 224          * extend the file past i_size without locking
 225          * a page we already have locked.
 226          *
 227          * We must do any isize and inode updates
 228          * before we unlock the pages.  Otherwise we
 229          * could end up racing with unlink.
 230          */
 231         BTRFS_I(inode)->disk_i_size = inode->i_size;
 232         ret = btrfs_update_inode(trans, root, inode);
 233
 234         return ret;
 235 fail:
 236         return err;
 237 }
 238
 239
 240 /*
 241  * conditionally insert an inline extent into the file.  This
 242  * does the checks required to make sure the data is small enough
 243  * to fit as an inline extent.
 244  */
 245 static noinline int cow_file_range_inline(struct btrfs_root *root,
 246                                           struct inode *inode, u64 start,
 247                                           u64 end, size_t compressed_size,
 248                                           int compress_type,
 249                                           struct page **compressed_pages)
 250 {
 251         struct btrfs_trans_handle *trans;
 252         u64 isize = i_size_read(inode);
 253         u64 actual_end = min(end + 1, isize);
 254         u64 inline_len = actual_end - start;
 255         u64 aligned_end = ALIGN(end, root->sectorsize);
 256         u64 data_len = inline_len;
 257         int ret;
 258         struct btrfs_path *path;
 259         int extent_inserted = 0;
 260         u32 extent_item_size;
 261
 262         if (compressed_size)
 263                 data_len = compressed_size;
 264
 265         if (start > 0 ||
 266             actual_end > PAGE_CACHE_SIZE ||
 267             data_len > BTRFS_MAX_INLINE_DATA_SIZE(root) ||
 268             (!compressed_size &&
 269             (actual_end & (root->sectorsize - 1)) == 0) ||
 270             end + 1 < isize ||
 271             data_len > root->fs_info->max_inline) {
 272                 return 1;
 273         }
 274
 275         path = btrfs_alloc_path();
 276         if (!path)
 277                 return -ENOMEM;
 278
 279         trans = btrfs_join_transaction(root);
 280         if (IS_ERR(trans)) {
 281                 btrfs_free_path(path);
 282                 return PTR_ERR(trans);
 283         }
 284         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
 285
 286         if (compressed_size && compressed_pages)
 287                 extent_item_size = btrfs_file_extent_calc_inline_size(
 288                    compressed_size);
 289         else
 290                 extent_item_size = btrfs_file_extent_calc_inline_size(
 291                     inline_len);
 292
 293         ret = __btrfs_drop_extents(trans, root, inode, path,
 294                                    start, aligned_end, NULL,
 295                                    1, 1, extent_item_size, &extent_inserted);
 296         if (ret) {
 297                 btrfs_abort_transaction(trans, root, ret);
 298                 goto out;
 299         }
 300
 301         if (isize > actual_end)
 302                 inline_len = min_t(u64, isize, actual_end);
 303         ret = insert_inline_extent(trans, path, extent_inserted,
 304                                    root, inode, start,
 305                                    inline_len, compressed_size,
 306                                    compress_type, compressed_pages);
 307         if (ret && ret != -ENOSPC) {
 308                 btrfs_abort_transaction(trans, root, ret);
 309                 goto out;
 310         } else if (ret == -ENOSPC) {
 311                 ret = 1;
 312                 goto out;
 313         }
 314
 315         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
 316         btrfs_delalloc_release_metadata(inode, end + 1 - start);
 317         btrfs_drop_extent_cache(inode, start, aligned_end - 1, 0);
 318 out:
 319         /*
 320          * Don't forget to free the reserved space, as for inlined extent
 321          * it won't count as data extent, free them directly here.
 322          * And at reserve time, it's always aligned to page size, so
 323          * just free one page here.
 324          */
 325         btrfs_qgroup_free_data(inode, 0, PAGE_CACHE_SIZE);
 326         btrfs_free_path(path);
 327         btrfs_end_transaction(trans, root);
 328         return ret;
 329 }
 330
 331 struct async_extent {
 332         u64 start;
 333         u64 ram_size;
 334         u64 compressed_size;
 335         struct page **pages;
 336         unsigned long nr_pages;
 337         int compress_type;
 338         struct list_head list;
 339 };
 340
 341 struct async_cow {
 342         struct inode *inode;
 343         struct btrfs_root *root;
 344         struct page *locked_page;
 345         u64 start;
 346         u64 end;
 347         struct list_head extents;
 348         struct btrfs_work work;
 349 };
 350
 351 static noinline int add_async_extent(struct async_cow *cow,
 352                                      u64 start, u64 ram_size,
 353                                      u64 compressed_size,
 354                                      struct page **pages,
 355                                      unsigned long nr_pages,
 356                                      int compress_type)
 357 {
 358         struct async_extent *async_extent;
 359
 360         async_extent = kmalloc(sizeof(*async_extent), GFP_NOFS);
 361         BUG_ON(!async_extent); /* -ENOMEM */
 362         async_extent->start = start;
 363         async_extent->ram_size = ram_size;
 364         async_extent->compressed_size = compressed_size;
 365         async_extent->pages = pages;
 366         async_extent->nr_pages = nr_pages;
 367         async_extent->compress_type = compress_type;
 368         list_add_tail(&async_extent->list, &cow->extents);
 369         return 0;
 370 }
 371
 372 static inline int inode_need_compress(struct inode *inode)
 373 {
 374         struct btrfs_root *root = BTRFS_I(inode)->root;
 375
 376         /* force compress */
 377         if (btrfs_test_opt(root, FORCE_COMPRESS))
 378                 return 1;
 379         /* bad compression ratios */
 380         if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS)
 381                 return 0;
 382         if (btrfs_test_opt(root, COMPRESS) ||
 383             BTRFS_I(inode)->flags & BTRFS_INODE_COMPRESS ||
 384             BTRFS_I(inode)->force_compress)
 385                 return 1;
 386         return 0;
 387 }
 388
 389 /*
 390  * we create compressed extents in two phases.  The first
 391  * phase compresses a range of pages that have already been
 392  * locked (both pages and state bits are locked).
 393  *
 394  * This is done inside an ordered work queue, and the compression
 395  * is spread across many cpus.  The actual IO submission is step
 396  * two, and the ordered work queue takes care of making sure that
 397  * happens in the same order things were put onto the queue by
 398  * writepages and friends.
 399  *
 400  * If this code finds it can't get good compression, it puts an
 401  * entry onto the work queue to write the uncompressed bytes.  This
 402  * makes sure that both compressed inodes and uncompressed inodes
 403  * are written in the same order that the flusher thread sent them
 404  * down.
 405  */
 406 static noinline void compress_file_range(struct inode *inode,
 407                                         struct page *locked_page,
 408                                         u64 start, u64 end,
 409                                         struct async_cow *async_cow,
 410                                         int *num_added)
 411 {
 412         struct btrfs_root *root = BTRFS_I(inode)->root;
 413         u64 num_bytes;
 414         u64 blocksize = root->sectorsize;
 415         u64 actual_end;
 416         u64 isize = i_size_read(inode);
 417         int ret = 0;
 418         struct page **pages = NULL;
 419         unsigned long nr_pages;
 420         unsigned long nr_pages_ret = 0;
 421         unsigned long total_compressed = 0;
 422         unsigned long total_in = 0;
 423         unsigned long max_compressed = SZ_128K;
 424         unsigned long max_uncompressed = SZ_128K;
 425         int i;
 426         int will_compress;
 427         int compress_type = root->fs_info->compress_type;
 428         int redirty = 0;
 429
 430         /* if this is a small write inside eof, kick off a defrag */
 431         if ((end - start + 1) < SZ_16K &&
 432             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 433                 btrfs_add_inode_defrag(NULL, inode);
 434
 435         actual_end = min_t(u64, isize, end + 1);
 436 again:
 437         will_compress = 0;
 438         nr_pages = (end >> PAGE_CACHE_SHIFT) - (start >> PAGE_CACHE_SHIFT) + 1;
 439         nr_pages = min_t(unsigned long, nr_pages, SZ_128K / PAGE_CACHE_SIZE);
 440
 441         /*
 442          * we don't want to send crud past the end of i_size through
 443          * compression, that's just a waste of CPU time.  So, if the
 444          * end of the file is before the start of our current
 445          * requested range of bytes, we bail out to the uncompressed
 446          * cleanup code that can deal with all of this.
 447          *
 448          * It isn't really the fastest way to fix things, but this is a
 449          * very uncommon corner.
 450          */
 451         if (actual_end <= start)
 452                 goto cleanup_and_bail_uncompressed;
 453
 454         total_compressed = actual_end - start;
 455
 456         /*
 457          * skip compression for a small file range(<=blocksize) that
 458          * isn't an inline extent, since it dosen't save disk space at all.
 459          */
 460         if (total_compressed <= blocksize &&
 461            (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 462                 goto cleanup_and_bail_uncompressed;
 463
 464         /* we want to make sure that amount of ram required to uncompress
 465          * an extent is reasonable, so we limit the total size in ram
 466          * of a compressed extent to 128k.  This is a crucial number
 467          * because it also controls how easily we can spread reads across
 468          * cpus for decompression.
 469          *
 470          * We also want to make sure the amount of IO required to do
 471          * a random read is reasonably small, so we limit the size of
 472          * a compressed extent to 128k.
 473          */
 474         total_compressed = min(total_compressed, max_uncompressed);
 475         num_bytes = ALIGN(end - start + 1, blocksize);
 476         num_bytes = max(blocksize,  num_bytes);
 477         total_in = 0;
 478         ret = 0;
 479
 480         /*
 481          * we do compression for mount -o compress and when the
 482          * inode has not been flagged as nocompress.  This flag can
 483          * change at any time if we discover bad compression ratios.
 484          */
 485         if (inode_need_compress(inode)) {
 486                 WARN_ON(pages);
 487                 pages = kcalloc(nr_pages, sizeof(struct page *), GFP_NOFS);
 488                 if (!pages) {
 489                         /* just bail out to the uncompressed code */
 490                         goto cont;
 491                 }
 492
 493                 if (BTRFS_I(inode)->force_compress)
 494                         compress_type = BTRFS_I(inode)->force_compress;
 495
 496                 /*
 497                  * we need to call clear_page_dirty_for_io on each
 498                  * page in the range.  Otherwise applications with the file
 499                  * mmap'd can wander in and change the page contents while
 500                  * we are compressing them.
 501                  *
 502                  * If the compression fails for any reason, we set the pages
 503                  * dirty again later on.
 504                  */
 505                 extent_range_clear_dirty_for_io(inode, start, end);
 506                 redirty = 1;
 507                 ret = btrfs_compress_pages(compress_type,
 508                                            inode->i_mapping, start,
 509                                            total_compressed, pages,
 510                                            nr_pages, &nr_pages_ret,
 511                                            &total_in,
 512                                            &total_compressed,
 513                                            max_compressed);
 514
 515                 if (!ret) {
 516                         unsigned long offset = total_compressed &
 517                                 (PAGE_CACHE_SIZE - 1);
 518                         struct page *page = pages[nr_pages_ret - 1];
 519                         char *kaddr;
 520
 521                         /* zero the tail end of the last page, we might be
 522                          * sending it down to disk
 523                          */
 524                         if (offset) {
 525                                 kaddr = kmap_atomic(page);
 526                                 memset(kaddr + offset, 0,
 527                                        PAGE_CACHE_SIZE - offset);
 528                                 kunmap_atomic(kaddr);
 529                         }
 530                         will_compress = 1;
 531                 }
 532         }
 533 cont:
 534         if (start == 0) {
 535                 /* lets try to make an inline extent */
 536                 if (ret || total_in < (actual_end - start)) {
 537                         /* we didn't compress the entire range, try
 538                          * to make an uncompressed inline extent.
 539                          */
 540                         ret = cow_file_range_inline(root, inode, start, end,
 541                                                     0, 0, NULL);
 542                 } else {
 543                         /* try making a compressed inline extent */
 544                         ret = cow_file_range_inline(root, inode, start, end,
 545                                                     total_compressed,
 546                                                     compress_type, pages);
 547                 }
 548                 if (ret <= 0) {
 549                         unsigned long clear_flags = EXTENT_DELALLOC |
 550                                 EXTENT_DEFRAG;
 551                         unsigned long page_error_op;
 552
 553                         clear_flags |= (ret < 0) ? EXTENT_DO_ACCOUNTING : 0;
 554                         page_error_op = ret < 0 ? PAGE_SET_ERROR : 0;
 555
 556                         /*
 557                          * inline extent creation worked or returned error,
 558                          * we don't need to create any more async work items.
 559                          * Unlock and free up our temp pages.
 560                          */
 561                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 562                                                      clear_flags, PAGE_UNLOCK |
 563                                                      PAGE_CLEAR_DIRTY |
 564                                                      PAGE_SET_WRITEBACK |
 565                                                      page_error_op |
 566                                                      PAGE_END_WRITEBACK);
 567                         goto free_pages_out;
 568                 }
 569         }
 570
 571         if (will_compress) {
 572                 /*
 573                  * we aren't doing an inline extent round the compressed size
 574                  * up to a block size boundary so the allocator does sane
 575                  * things
 576                  */
 577                 total_compressed = ALIGN(total_compressed, blocksize);
 578
 579                 /*
 580                  * one last check to make sure the compression is really a
 581                  * win, compare the page count read with the blocks on disk
 582                  */
 583                 total_in = ALIGN(total_in, PAGE_CACHE_SIZE);
 584                 if (total_compressed >= total_in) {
 585                         will_compress = 0;
 586                 } else {
 587                         num_bytes = total_in;
 588                 }
 589         }
 590         if (!will_compress && pages) {
 591                 /*
 592                  * the compression code ran but failed to make things smaller,
 593                  * free any pages it allocated and our page pointer array
 594                  */
 595                 for (i = 0; i < nr_pages_ret; i++) {
 596                         WARN_ON(pages[i]->mapping);
 597                         page_cache_release(pages[i]);
 598                 }
 599                 kfree(pages);
 600                 pages = NULL;
 601                 total_compressed = 0;
 602                 nr_pages_ret = 0;
 603
 604                 /* flag the file so we don't compress in the future */
 605                 if (!btrfs_test_opt(root, FORCE_COMPRESS) &&
 606                     !(BTRFS_I(inode)->force_compress)) {
 607                         BTRFS_I(inode)->flags |= BTRFS_INODE_NOCOMPRESS;
 608                 }
 609         }
 610         if (will_compress) {
 611                 *num_added += 1;
 612
 613                 /* the async work queues will take care of doing actual
 614                  * allocation on disk for these compressed pages,
 615                  * and will submit them to the elevator.
 616                  */
 617                 add_async_extent(async_cow, start, num_bytes,
 618                                  total_compressed, pages, nr_pages_ret,
 619                                  compress_type);
 620
 621                 if (start + num_bytes < end) {
 622                         start += num_bytes;
 623                         pages = NULL;
 624                         cond_resched();
 625                         goto again;
 626                 }
 627         } else {
 628 cleanup_and_bail_uncompressed:
 629                 /*
 630                  * No compression, but we still need to write the pages in
 631                  * the file we've been given so far.  redirty the locked
 632                  * page if it corresponds to our extent and set things up
 633                  * for the async work queue to run cow_file_range to do
 634                  * the normal delalloc dance
 635                  */
 636                 if (page_offset(locked_page) >= start &&
 637                     page_offset(locked_page) <= end) {
 638                         __set_page_dirty_nobuffers(locked_page);
 639                         /* unlocked later on in the async handlers */
 640                 }
 641                 if (redirty)
 642                         extent_range_redirty_for_io(inode, start, end);
 643                 add_async_extent(async_cow, start, end - start + 1,
 644                                  0, NULL, 0, BTRFS_COMPRESS_NONE);
 645                 *num_added += 1;
 646         }
 647
 648         return;
 649
 650 free_pages_out:
 651         for (i = 0; i < nr_pages_ret; i++) {
 652                 WARN_ON(pages[i]->mapping);
 653                 page_cache_release(pages[i]);
 654         }
 655         kfree(pages);
 656 }
 657
 658 static void free_async_extent_pages(struct async_extent *async_extent)
 659 {
 660         int i;
 661
 662         if (!async_extent->pages)
 663                 return;
 664
 665         for (i = 0; i < async_extent->nr_pages; i++) {
 666                 WARN_ON(async_extent->pages[i]->mapping);
 667                 page_cache_release(async_extent->pages[i]);
 668         }
 669         kfree(async_extent->pages);
 670         async_extent->nr_pages = 0;
 671         async_extent->pages = NULL;
 672 }
 673
 674 /*
 675  * phase two of compressed writeback.  This is the ordered portion
 676  * of the code, which only gets called in the order the work was
 677  * queued.  We walk all the async extents created by compress_file_range
 678  * and send them down to the disk.
 679  */
 680 static noinline void submit_compressed_extents(struct inode *inode,
 681                                               struct async_cow *async_cow)
 682 {
 683         struct async_extent *async_extent;
 684         u64 alloc_hint = 0;
 685         struct btrfs_key ins;
 686         struct extent_map *em;
 687         struct btrfs_root *root = BTRFS_I(inode)->root;
 688         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 689         struct extent_io_tree *io_tree;
 690         int ret = 0;
 691
 692 again:
 693         while (!list_empty(&async_cow->extents)) {
 694                 async_extent = list_entry(async_cow->extents.next,
 695                                           struct async_extent, list);
 696                 list_del(&async_extent->list);
 697
 698                 io_tree = &BTRFS_I(inode)->io_tree;
 699
 700 retry:
 701                 /* did the compression code fall back to uncompressed IO? */
 702                 if (!async_extent->pages) {
 703                         int page_started = 0;
 704                         unsigned long nr_written = 0;
 705
 706                         lock_extent(io_tree, async_extent->start,
 707                                          async_extent->start +
 708                                          async_extent->ram_size - 1);
 709
 710                         /* allocate blocks */
 711                         ret = cow_file_range(inode, async_cow->locked_page,
 712                                              async_extent->start,
 713                                              async_extent->start +
 714                                              async_extent->ram_size - 1,
 715                                              &page_started, &nr_written, 0);
 716
 717                         /* JDM XXX */
 718
 719                         /*
 720                          * if page_started, cow_file_range inserted an
 721                          * inline extent and took care of all the unlocking
 722                          * and IO for us.  Otherwise, we need to submit
 723                          * all those pages down to the drive.
 724                          */
 725                         if (!page_started && !ret)
 726                                 extent_write_locked_range(io_tree,
 727                                                   inode, async_extent->start,
 728                                                   async_extent->start +
 729                                                   async_extent->ram_size - 1,
 730                                                   btrfs_get_extent,
 731                                                   WB_SYNC_ALL);
 732                         else if (ret)
 733                                 unlock_page(async_cow->locked_page);
 734                         kfree(async_extent);
 735                         cond_resched();
 736                         continue;
 737                 }
 738
 739                 lock_extent(io_tree, async_extent->start,
 740                             async_extent->start + async_extent->ram_size - 1);
 741
 742                 ret = btrfs_reserve_extent(root,
 743                                            async_extent->compressed_size,
 744                                            async_extent->compressed_size,
 745                                            0, alloc_hint, &ins, 1, 1);
 746                 if (ret) {
 747                         free_async_extent_pages(async_extent);
 748
 749                         if (ret == -ENOSPC) {
 750                                 unlock_extent(io_tree, async_extent->start,
 751                                               async_extent->start +
 752                                               async_extent->ram_size - 1);
 753
 754                                 /*
 755                                  * we need to redirty the pages if we decide to
 756                                  * fallback to uncompressed IO, otherwise we
 757                                  * will not submit these pages down to lower
 758                                  * layers.
 759                                  */
 760                                 extent_range_redirty_for_io(inode,
 761                                                 async_extent->start,
 762                                                 async_extent->start +
 763                                                 async_extent->ram_size - 1);
 764
 765                                 goto retry;
 766                         }
 767                         goto out_free;
 768                 }
 769                 /*
 770                  * here we're doing allocation and writeback of the
 771                  * compressed pages
 772                  */
 773                 btrfs_drop_extent_cache(inode, async_extent->start,
 774                                         async_extent->start +
 775                                         async_extent->ram_size - 1, 0);
 776
 777                 em = alloc_extent_map();
 778                 if (!em) {
 779                         ret = -ENOMEM;
 780                         goto out_free_reserve;
 781                 }
 782                 em->start = async_extent->start;
 783                 em->len = async_extent->ram_size;
 784                 em->orig_start = em->start;
 785                 em->mod_start = em->start;
 786                 em->mod_len = em->len;
 787
 788                 em->block_start = ins.objectid;
 789                 em->block_len = ins.offset;
 790                 em->orig_block_len = ins.offset;
 791                 em->ram_bytes = async_extent->ram_size;
 792                 em->bdev = root->fs_info->fs_devices->latest_bdev;
 793                 em->compress_type = async_extent->compress_type;
 794                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
 795                 set_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 796                 em->generation = -1;
 797
 798                 while (1) {
 799                         write_lock(&em_tree->lock);
 800                         ret = add_extent_mapping(em_tree, em, 1);
 801                         write_unlock(&em_tree->lock);
 802                         if (ret != -EEXIST) {
 803                                 free_extent_map(em);
 804                                 break;
 805                         }
 806                         btrfs_drop_extent_cache(inode, async_extent->start,
 807                                                 async_extent->start +
 808                                                 async_extent->ram_size - 1, 0);
 809                 }
 810
 811                 if (ret)
 812                         goto out_free_reserve;
 813
 814                 ret = btrfs_add_ordered_extent_compress(inode,
 815                                                 async_extent->start,
 816                                                 ins.objectid,
 817                                                 async_extent->ram_size,
 818                                                 ins.offset,
 819                                                 BTRFS_ORDERED_COMPRESSED,
 820                                                 async_extent->compress_type);
 821                 if (ret) {
 822                         btrfs_drop_extent_cache(inode, async_extent->start,
 823                                                 async_extent->start +
 824                                                 async_extent->ram_size - 1, 0);
 825                         goto out_free_reserve;
 826                 }
 827
 828                 /*
 829                  * clear dirty, set writeback and unlock the pages.
 830                  */
 831                 extent_clear_unlock_delalloc(inode, async_extent->start,
 832                                 async_extent->start +
 833                                 async_extent->ram_size - 1,
 834                                 NULL, EXTENT_LOCKED | EXTENT_DELALLOC,
 835                                 PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 836                                 PAGE_SET_WRITEBACK);
 837                 ret = btrfs_submit_compressed_write(inode,
 838                                     async_extent->start,
 839                                     async_extent->ram_size,
 840                                     ins.objectid,
 841                                     ins.offset, async_extent->pages,
 842                                     async_extent->nr_pages);
 843                 if (ret) {
 844                         struct extent_io_tree *tree = &BTRFS_I(inode)->io_tree;
 845                         struct page *p = async_extent->pages[0];
 846                         const u64 start = async_extent->start;
 847                         const u64 end = start + async_extent->ram_size - 1;
 848
 849                         p->mapping = inode->i_mapping;
 850                         tree->ops->writepage_end_io_hook(p, start, end,
 851                                                          NULL, 0);
 852                         p->mapping = NULL;
 853                         extent_clear_unlock_delalloc(inode, start, end, NULL, 0,
 854                                                      PAGE_END_WRITEBACK |
 855                                                      PAGE_SET_ERROR);
 856                         free_async_extent_pages(async_extent);
 857                 }
 858                 alloc_hint = ins.objectid + ins.offset;
 859                 kfree(async_extent);
 860                 cond_resched();
 861         }
 862         return;
 863 out_free_reserve:
 864         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
 865 out_free:
 866         extent_clear_unlock_delalloc(inode, async_extent->start,
 867                                      async_extent->start +
 868                                      async_extent->ram_size - 1,
 869                                      NULL, EXTENT_LOCKED | EXTENT_DELALLOC |
 870                                      EXTENT_DEFRAG | EXTENT_DO_ACCOUNTING,
 871                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
 872                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK |
 873                                      PAGE_SET_ERROR);
 874         free_async_extent_pages(async_extent);
 875         kfree(async_extent);
 876         goto again;
 877 }
 878
 879 static u64 get_extent_allocation_hint(struct inode *inode, u64 start,
 880                                       u64 num_bytes)
 881 {
 882         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 883         struct extent_map *em;
 884         u64 alloc_hint = 0;
 885
 886         read_lock(&em_tree->lock);
 887         em = search_extent_mapping(em_tree, start, num_bytes);
 888         if (em) {
 889                 /*
 890                  * if block start isn't an actual block number then find the
 891                  * first block in this inode and use that as a hint.  If that
 892                  * block is also bogus then just don't worry about it.
 893                  */
 894                 if (em->block_start >= EXTENT_MAP_LAST_BYTE) {
 895                         free_extent_map(em);
 896                         em = search_extent_mapping(em_tree, 0, 0);
 897                         if (em && em->block_start < EXTENT_MAP_LAST_BYTE)
 898                                 alloc_hint = em->block_start;
 899                         if (em)
 900                                 free_extent_map(em);
 901                 } else {
 902                         alloc_hint = em->block_start;
 903                         free_extent_map(em);
 904                 }
 905         }
 906         read_unlock(&em_tree->lock);
 907
 908         return alloc_hint;
 909 }
 910
 911 /*
 912  * when extent_io.c finds a delayed allocation range in the file,
 913  * the call backs end up in this code.  The basic idea is to
 914  * allocate extents on disk for the range, and create ordered data structs
 915  * in ram to track those extents.
 916  *
 917  * locked_page is the page that writepage had locked already.  We use
 918  * it to make sure we don't do extra locks or unlocks.
 919  *
 920  * *page_started is set to one if we unlock locked_page and do everything
 921  * required to start IO on it.  It may be clean and already done with
 922  * IO when we return.
 923  */
 924 static noinline int cow_file_range(struct inode *inode,
 925                                    struct page *locked_page,
 926                                    u64 start, u64 end, int *page_started,
 927                                    unsigned long *nr_written,
 928                                    int unlock)
 929 {
 930         struct btrfs_root *root = BTRFS_I(inode)->root;
 931         u64 alloc_hint = 0;
 932         u64 num_bytes;
 933         unsigned long ram_size;
 934         u64 disk_num_bytes;
 935         u64 cur_alloc_size;
 936         u64 blocksize = root->sectorsize;
 937         struct btrfs_key ins;
 938         struct extent_map *em;
 939         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 940         int ret = 0;
 941
 942         if (btrfs_is_free_space_inode(inode)) {
 943                 WARN_ON_ONCE(1);
 944                 ret = -EINVAL;
 945                 goto out_unlock;
 946         }
 947
 948         num_bytes = ALIGN(end - start + 1, blocksize);
 949         num_bytes = max(blocksize,  num_bytes);
 950         disk_num_bytes = num_bytes;
 951
 952         /* if this is a small write inside eof, kick off defrag */
 953         if (num_bytes < SZ_64K &&
 954             (start > 0 || end + 1 < BTRFS_I(inode)->disk_i_size))
 955                 btrfs_add_inode_defrag(NULL, inode);
 956
 957         if (start == 0) {
 958                 /* lets try to make an inline extent */
 959                 ret = cow_file_range_inline(root, inode, start, end, 0, 0,
 960                                             NULL);
 961                 if (ret == 0) {
 962                         extent_clear_unlock_delalloc(inode, start, end, NULL,
 963                                      EXTENT_LOCKED | EXTENT_DELALLOC |
 964                                      EXTENT_DEFRAG, PAGE_UNLOCK |
 965                                      PAGE_CLEAR_DIRTY | PAGE_SET_WRITEBACK |
 966                                      PAGE_END_WRITEBACK);
 967
 968                         *nr_written = *nr_written +
 969                              (end - start + PAGE_CACHE_SIZE) / PAGE_CACHE_SIZE;
 970                         *page_started = 1;
 971                         goto out;
 972                 } else if (ret < 0) {
 973                         goto out_unlock;
 974                 }
 975         }
 976
 977         BUG_ON(disk_num_bytes >
 978                btrfs_super_total_bytes(root->fs_info->super_copy));
 979
 980         alloc_hint = get_extent_allocation_hint(inode, start, num_bytes);
 981         btrfs_drop_extent_cache(inode, start, start + num_bytes - 1, 0);
 982
 983         while (disk_num_bytes > 0) {
 984                 unsigned long op;
 985
 986                 cur_alloc_size = disk_num_bytes;
 987                 ret = btrfs_reserve_extent(root, cur_alloc_size,
 988                                            root->sectorsize, 0, alloc_hint,
 989                                            &ins, 1, 1);
 990                 if (ret < 0)
 991                         goto out_unlock;
 992
 993                 em = alloc_extent_map();
 994                 if (!em) {
 995                         ret = -ENOMEM;
 996                         goto out_reserve;
 997                 }
 998                 em->start = start;
 999                 em->orig_start = em->start;
1000                 ram_size = ins.offset;
1001                 em->len = ins.offset;
1002                 em->mod_start = em->start;
1003                 em->mod_len = em->len;
1004
1005                 em->block_start = ins.objectid;
1006                 em->block_len = ins.offset;
1007                 em->orig_block_len = ins.offset;
1008                 em->ram_bytes = ram_size;
1009                 em->bdev = root->fs_info->fs_devices->latest_bdev;
1010                 set_bit(EXTENT_FLAG_PINNED, &em->flags);
1011                 em->generation = -1;
1012
1013                 while (1) {
1014                         write_lock(&em_tree->lock);
1015                         ret = add_extent_mapping(em_tree, em, 1);
1016                         write_unlock(&em_tree->lock);
1017                         if (ret != -EEXIST) {
1018                                 free_extent_map(em);
1019                                 break;
1020                         }
1021                         btrfs_drop_extent_cache(inode, start,
1022                                                 start + ram_size - 1, 0);
1023                 }
1024                 if (ret)
1025                         goto out_reserve;
1026
1027                 cur_alloc_size = ins.offset;
1028                 ret = btrfs_add_ordered_extent(inode, start, ins.objectid,
1029                                                ram_size, cur_alloc_size, 0);
1030                 if (ret)
1031                         goto out_drop_extent_cache;
1032
1033                 if (root->root_key.objectid ==
1034                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1035                         ret = btrfs_reloc_clone_csums(inode, start,
1036                                                       cur_alloc_size);
1037                         if (ret)
1038                                 goto out_drop_extent_cache;
1039                 }
1040
1041                 if (disk_num_bytes < cur_alloc_size)
1042                         break;
1043
1044                 /* we're not doing compressed IO, don't unlock the first
1045                  * page (which the caller expects to stay locked), don't
1046                  * clear any dirty bits and don't set any writeback bits
1047                  *
1048                  * Do set the Private2 bit so we know this page was properly
1049                  * setup for writepage
1050                  */
1051                 op = unlock ? PAGE_UNLOCK : 0;
1052                 op |= PAGE_SET_PRIVATE2;
1053
1054                 extent_clear_unlock_delalloc(inode, start,
1055                                              start + ram_size - 1, locked_page,
1056                                              EXTENT_LOCKED | EXTENT_DELALLOC,
1057                                              op);
1058                 disk_num_bytes -= cur_alloc_size;
1059                 num_bytes -= cur_alloc_size;
1060                 alloc_hint = ins.objectid + ins.offset;
1061                 start += cur_alloc_size;
1062         }
1063 out:
1064         return ret;
1065
1066 out_drop_extent_cache:
1067         btrfs_drop_extent_cache(inode, start, start + ram_size - 1, 0);
1068 out_reserve:
1069         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
1070 out_unlock:
1071         extent_clear_unlock_delalloc(inode, start, end, locked_page,
1072                                      EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
1073                                      EXTENT_DELALLOC | EXTENT_DEFRAG,
1074                                      PAGE_UNLOCK | PAGE_CLEAR_DIRTY |
1075                                      PAGE_SET_WRITEBACK | PAGE_END_WRITEBACK);
1076         goto out;
1077 }
1078
1079 /*
1080  * work queue call back to started compression on a file and pages
1081  */
1082 static noinline void async_cow_start(struct btrfs_work *work)
1083 {
1084         struct async_cow *async_cow;
1085         int num_added = 0;
1086         async_cow = container_of(work, struct async_cow, work);
1087
1088         compress_file_range(async_cow->inode, async_cow->locked_page,
1089                             async_cow->start, async_cow->end, async_cow,
1090                             &num_added);
1091         if (num_added == 0) {
1092                 btrfs_add_delayed_iput(async_cow->inode);
1093                 async_cow->inode = NULL;
1094         }
1095 }
1096
1097 /*
1098  * work queue call back to submit previously compressed pages
1099  */
1100 static noinline void async_cow_submit(struct btrfs_work *work)
1101 {
1102         struct async_cow *async_cow;
1103         struct btrfs_root *root;
1104         unsigned long nr_pages;
1105
1106         async_cow = container_of(work, struct async_cow, work);
1107
1108         root = async_cow->root;
1109         nr_pages = (async_cow->end - async_cow->start + PAGE_CACHE_SIZE) >>
1110                 PAGE_CACHE_SHIFT;
1111
1112         /*
1113          * atomic_sub_return implies a barrier for waitqueue_active
1114          */
1115         if (atomic_sub_return(nr_pages, &root->fs_info->async_delalloc_pages) <
1116             5 * SZ_1M &&
1117             waitqueue_active(&root->fs_info->async_submit_wait))
1118                 wake_up(&root->fs_info->async_submit_wait);
1119
1120         if (async_cow->inode)
1121                 submit_compressed_extents(async_cow->inode, async_cow);
1122 }
1123
1124 static noinline void async_cow_free(struct btrfs_work *work)
1125 {
1126         struct async_cow *async_cow;
1127         async_cow = container_of(work, struct async_cow, work);
1128         if (async_cow->inode)
1129                 btrfs_add_delayed_iput(async_cow->inode);
1130         kfree(async_cow);
1131 }
1132
1133 static int cow_file_range_async(struct inode *inode, struct page *locked_page,
1134                                 u64 start, u64 end, int *page_started,
1135                                 unsigned long *nr_written)
1136 {
1137         struct async_cow *async_cow;
1138         struct btrfs_root *root = BTRFS_I(inode)->root;
1139         unsigned long nr_pages;
1140         u64 cur_end;
1141         int limit = 10 * SZ_1M;
1142
1143         clear_extent_bit(&BTRFS_I(inode)->io_tree, start, end, EXTENT_LOCKED,
1144                          1, 0, NULL, GFP_NOFS);
1145         while (start < end) {
1146                 async_cow = kmalloc(sizeof(*async_cow), GFP_NOFS);
1147                 BUG_ON(!async_cow); /* -ENOMEM */
1148                 async_cow->inode = igrab(inode);
1149                 async_cow->root = root;
1150                 async_cow->locked_page = locked_page;
1151                 async_cow->start = start;
1152
1153                 if (BTRFS_I(inode)->flags & BTRFS_INODE_NOCOMPRESS &&
1154                     !btrfs_test_opt(root, FORCE_COMPRESS))
1155                         cur_end = end;
1156                 else
1157                         cur_end = min(end, start + SZ_512K - 1);
1158
1159                 async_cow->end = cur_end;
1160                 INIT_LIST_HEAD(&async_cow->extents);
1161
1162                 btrfs_init_work(&async_cow->work,
1163                                 btrfs_delalloc_helper,
1164                                 async_cow_start, async_cow_submit,
1165                                 async_cow_free);
1166
1167                 nr_pages = (cur_end - start + PAGE_CACHE_SIZE) >>
1168                         PAGE_CACHE_SHIFT;
1169                 atomic_add(nr_pages, &root->fs_info->async_delalloc_pages);
1170
1171                 btrfs_queue_work(root->fs_info->delalloc_workers,
1172                                  &async_cow->work);
1173
1174                 if (atomic_read(&root->fs_info->async_delalloc_pages) > limit) {
1175                         wait_event(root->fs_info->async_submit_wait,
1176                            (atomic_read(&root->fs_info->async_delalloc_pages) <
1177                             limit));
1178                 }
1179
1180                 while (atomic_read(&root->fs_info->async_submit_draining) &&
1181                       atomic_read(&root->fs_info->async_delalloc_pages)) {
1182                         wait_event(root->fs_info->async_submit_wait,
1183                           (atomic_read(&root->fs_info->async_delalloc_pages) ==
1184                            0));
1185                 }
1186
1187                 *nr_written += nr_pages;
1188                 start = cur_end + 1;
1189         }
1190         *page_started = 1;
1191         return 0;
1192 }
1193
1194 static noinline int csum_exist_in_range(struct btrfs_root *root,
1195                                         u64 bytenr, u64 num_bytes)
1196 {
1197         int ret;
1198         struct btrfs_ordered_sum *sums;
1199         LIST_HEAD(list);
1200
1201         ret = btrfs_lookup_csums_range(root->fs_info->csum_root, bytenr,
1202                                        bytenr + num_bytes - 1, &list, 0);
1203         if (ret == 0 && list_empty(&list))
1204                 return 0;
1205
1206         while (!list_empty(&list)) {
1207                 sums = list_entry(list.next, struct btrfs_ordered_sum, list);
1208                 list_del(&sums->list);
1209                 kfree(sums);
1210         }
1211         return 1;
1212 }
1213
1214 /*
1215  * when nowcow writeback call back.  This checks for snapshots or COW copies
1216  * of the extents that exist in the file, and COWs the file as required.
1217  *
1218  * If no cow copies or snapshots exist, we write directly to the existing
1219  * blocks on disk
1220  */
1221 static noinline int run_delalloc_nocow(struct inode *inode,
1222                                        struct page *locked_page,
1223                               u64 start, u64 end, int *page_started, int force,
1224                               unsigned long *nr_written)
1225 {
1226         struct btrfs_root *root = BTRFS_I(inode)->root;
1227         struct btrfs_trans_handle *trans;
1228         struct extent_buffer *leaf;
1229         struct btrfs_path *path;
1230         struct btrfs_file_extent_item *fi;
1231         struct btrfs_key found_key;
1232         u64 cow_start;
1233         u64 cur_offset;
1234         u64 extent_end;
1235         u64 extent_offset;
1236         u64 disk_bytenr;
1237         u64 num_bytes;
1238         u64 disk_num_bytes;
1239         u64 ram_bytes;
1240         int extent_type;
1241         int ret, err;
1242         int type;
1243         int nocow;
1244         int check_prev = 1;
1245         bool nolock;
1246         u64 ino = btrfs_ino(inode);
1247
1248         path = btrfs_alloc_path();
1249         if (!path) {
1250                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1251                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1252                                              EXTENT_DO_ACCOUNTING |
1253                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1254                                              PAGE_CLEAR_DIRTY |
1255                                              PAGE_SET_WRITEBACK |
1256                                              PAGE_END_WRITEBACK);
1257                 return -ENOMEM;
1258         }
1259
1260         nolock = btrfs_is_free_space_inode(inode);
1261
1262         if (nolock)
1263                 trans = btrfs_join_transaction_nolock(root);
1264         else
1265                 trans = btrfs_join_transaction(root);
1266
1267         if (IS_ERR(trans)) {
1268                 extent_clear_unlock_delalloc(inode, start, end, locked_page,
1269                                              EXTENT_LOCKED | EXTENT_DELALLOC |
1270                                              EXTENT_DO_ACCOUNTING |
1271                                              EXTENT_DEFRAG, PAGE_UNLOCK |
1272                                              PAGE_CLEAR_DIRTY |
1273                                              PAGE_SET_WRITEBACK |
1274                                              PAGE_END_WRITEBACK);
1275                 btrfs_free_path(path);
1276                 return PTR_ERR(trans);
1277         }
1278
1279         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
1280
1281         cow_start = (u64)-1;
1282         cur_offset = start;
1283         while (1) {
1284                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
1285                                                cur_offset, 0);
1286                 if (ret < 0)
1287                         goto error;
1288                 if (ret > 0 && path->slots[0] > 0 && check_prev) {
1289                         leaf = path->nodes[0];
1290                         btrfs_item_key_to_cpu(leaf, &found_key,
1291                                               path->slots[0] - 1);
1292                         if (found_key.objectid == ino &&
1293                             found_key.type == BTRFS_EXTENT_DATA_KEY)
1294                                 path->slots[0]--;
1295                 }
1296                 check_prev = 0;
1297 next_slot:
1298                 leaf = path->nodes[0];
1299                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1300                         ret = btrfs_next_leaf(root, path);
1301                         if (ret < 0)
1302                                 goto error;
1303                         if (ret > 0)
1304                                 break;
1305                         leaf = path->nodes[0];
1306                 }
1307
1308                 nocow = 0;
1309                 disk_bytenr = 0;
1310                 num_bytes = 0;
1311                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1312
1313                 if (found_key.objectid > ino)
1314                         break;
1315                 if (WARN_ON_ONCE(found_key.objectid < ino) ||
1316                     found_key.type < BTRFS_EXTENT_DATA_KEY) {
1317                         path->slots[0]++;
1318                         goto next_slot;
1319                 }
1320                 if (found_key.type > BTRFS_EXTENT_DATA_KEY ||
1321                     found_key.offset > end)
1322                         break;
1323
1324                 if (found_key.offset > cur_offset) {
1325                         extent_end = found_key.offset;
1326                         extent_type = 0;
1327                         goto out_check;
1328                 }
1329
1330                 fi = btrfs_item_ptr(leaf, path->slots[0],
1331                                     struct btrfs_file_extent_item);
1332                 extent_type = btrfs_file_extent_type(leaf, fi);
1333
1334                 ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
1335                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
1336                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1337                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1338                         extent_offset = btrfs_file_extent_offset(leaf, fi);
1339                         extent_end = found_key.offset +
1340                                 btrfs_file_extent_num_bytes(leaf, fi);
1341                         disk_num_bytes =
1342                                 btrfs_file_extent_disk_num_bytes(leaf, fi);
1343                         if (extent_end <= start) {
1344                                 path->slots[0]++;
1345                                 goto next_slot;
1346                         }
1347                         if (disk_bytenr == 0)
1348                                 goto out_check;
1349                         if (btrfs_file_extent_compression(leaf, fi) ||
1350                             btrfs_file_extent_encryption(leaf, fi) ||
1351                             btrfs_file_extent_other_encoding(leaf, fi))
1352                                 goto out_check;
1353                         if (extent_type == BTRFS_FILE_EXTENT_REG && !force)
1354                                 goto out_check;
1355                         if (btrfs_extent_readonly(root, disk_bytenr))
1356                                 goto out_check;
1357                         if (btrfs_cross_ref_exist(trans, root, ino,
1358                                                   found_key.offset -
1359                                                   extent_offset, disk_bytenr))
1360                                 goto out_check;
1361                         disk_bytenr += extent_offset;
1362                         disk_bytenr += cur_offset - found_key.offset;
1363                         num_bytes = min(end + 1, extent_end) - cur_offset;
1364                         /*
1365                          * if there are pending snapshots for this root,
1366                          * we fall into common COW way.
1367                          */
1368                         if (!nolock) {
1369                                 err = btrfs_start_write_no_snapshoting(root);
1370                                 if (!err)
1371                                         goto out_check;
1372                         }
1373                         /*
1374                          * force cow if csum exists in the range.
1375                          * this ensure that csum for a given extent are
1376                          * either valid or do not exist.
1377                          */
1378                         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
1379                                 goto out_check;
1380                         nocow = 1;
1381                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
1382                         extent_end = found_key.offset +
1383                                 btrfs_file_extent_inline_len(leaf,
1384                                                      path->slots[0], fi);
1385                         extent_end = ALIGN(extent_end, root->sectorsize);
1386                 } else {
1387                         BUG_ON(1);
1388                 }
1389 out_check:
1390                 if (extent_end <= start) {
1391                         path->slots[0]++;
1392                         if (!nolock && nocow)
1393                                 btrfs_end_write_no_snapshoting(root);
1394                         goto next_slot;
1395                 }
1396                 if (!nocow) {
1397                         if (cow_start == (u64)-1)
1398                                 cow_start = cur_offset;
1399                         cur_offset = extent_end;
1400                         if (cur_offset > end)
1401                                 break;
1402                         path->slots[0]++;
1403                         goto next_slot;
1404                 }
1405
1406                 btrfs_release_path(path);
1407                 if (cow_start != (u64)-1) {
1408                         ret = cow_file_range(inode, locked_page,
1409                                              cow_start, found_key.offset - 1,
1410                                              page_started, nr_written, 1);
1411                         if (ret) {
1412                                 if (!nolock && nocow)
1413                                         btrfs_end_write_no_snapshoting(root);
1414                                 goto error;
1415                         }
1416                         cow_start = (u64)-1;
1417                 }
1418
1419                 if (extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
1420                         struct extent_map *em;
1421                         struct extent_map_tree *em_tree;
1422                         em_tree = &BTRFS_I(inode)->extent_tree;
1423                         em = alloc_extent_map();
1424                         BUG_ON(!em); /* -ENOMEM */
1425                         em->start = cur_offset;
1426                         em->orig_start = found_key.offset - extent_offset;
1427                         em->len = num_bytes;
1428                         em->block_len = num_bytes;
1429                         em->block_start = disk_bytenr;
1430                         em->orig_block_len = disk_num_bytes;
1431                         em->ram_bytes = ram_bytes;
1432                         em->bdev = root->fs_info->fs_devices->latest_bdev;
1433                         em->mod_start = em->start;
1434                         em->mod_len = em->len;
1435                         set_bit(EXTENT_FLAG_PINNED, &em->flags);
1436                         set_bit(EXTENT_FLAG_FILLING, &em->flags);
1437                         em->generation = -1;
1438                         while (1) {
1439                                 write_lock(&em_tree->lock);
1440                                 ret = add_extent_mapping(em_tree, em, 1);
1441                                 write_unlock(&em_tree->lock);
1442                                 if (ret != -EEXIST) {
1443                                         free_extent_map(em);
1444                                         break;
1445                                 }
1446                                 btrfs_drop_extent_cache(inode, em->start,
1447                                                 em->start + em->len - 1, 0);
1448                         }
1449                         type = BTRFS_ORDERED_PREALLOC;
1450                 } else {
1451                         type = BTRFS_ORDERED_NOCOW;
1452                 }
1453
1454                 ret = btrfs_add_ordered_extent(inode, cur_offset, disk_bytenr,
1455                                                num_bytes, num_bytes, type);
1456                 BUG_ON(ret); /* -ENOMEM */
1457
1458                 if (root->root_key.objectid ==
1459                     BTRFS_DATA_RELOC_TREE_OBJECTID) {
1460                         ret = btrfs_reloc_clone_csums(inode, cur_offset,
1461                                                       num_bytes);
1462                         if (ret) {
1463                                 if (!nolock && nocow)
1464                                         btrfs_end_write_no_snapshoting(root);
1465                                 goto error;
1466                         }
1467                 }
1468
1469                 extent_clear_unlock_delalloc(inode, cur_offset,
1470                                              cur_offset + num_bytes - 1,
1471                                              locked_page, EXTENT_LOCKED |
1472                                              EXTENT_DELALLOC, PAGE_UNLOCK |
1473                                              PAGE_SET_PRIVATE2);
1474                 if (!nolock && nocow)
1475                         btrfs_end_write_no_snapshoting(root);
1476                 cur_offset = extent_end;
1477                 if (cur_offset > end)
1478                         break;
1479         }
1480         btrfs_release_path(path);
1481
1482         if (cur_offset <= end && cow_start == (u64)-1) {
1483                 cow_start = cur_offset;
1484                 cur_offset = end;
1485         }
1486
1487         if (cow_start != (u64)-1) {
1488                 ret = cow_file_range(inode, locked_page, cow_start, end,
1489                                      page_started, nr_written, 1);
1490                 if (ret)
1491                         goto error;
1492         }
1493
1494 error:
1495         err = btrfs_end_transaction(trans, root);
1496         if (!ret)
1497                 ret = err;
1498
1499         if (ret && cur_offset < end)
1500                 extent_clear_unlock_delalloc(inode, cur_offset, end,
1501                                              locked_page, EXTENT_LOCKED |
1502                                              EXTENT_DELALLOC | EXTENT_DEFRAG |
1503                                              EXTENT_DO_ACCOUNTING, PAGE_UNLOCK |
1504                                              PAGE_CLEAR_DIRTY |
1505                                              PAGE_SET_WRITEBACK |
1506                                              PAGE_END_WRITEBACK);
1507         btrfs_free_path(path);
1508         return ret;
1509 }
1510
1511 static inline int need_force_cow(struct inode *inode, u64 start, u64 end)
1512 {
1513
1514         if (!(BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
1515             !(BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC))
1516                 return 0;
1517
1518         /*
1519          * @defrag_bytes is a hint value, no spinlock held here,
1520          * if is not zero, it means the file is defragging.
1521          * Force cow if given extent needs to be defragged.
1522          */
1523         if (BTRFS_I(inode)->defrag_bytes &&
1524             test_range_bit(&BTRFS_I(inode)->io_tree, start, end,
1525                            EXTENT_DEFRAG, 0, NULL))
1526                 return 1;
1527
1528         return 0;
1529 }
1530
1531 /*
1532  * extent_io.c call back to do delayed allocation processing
1533  */
1534 static int run_delalloc_range(struct inode *inode, struct page *locked_page,
1535                               u64 start, u64 end, int *page_started,
1536                               unsigned long *nr_written)
1537 {
1538         int ret;
1539         int force_cow = need_force_cow(inode, start, end);
1540
1541         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW && !force_cow) {
1542                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1543                                          page_started, 1, nr_written);
1544         } else if (BTRFS_I(inode)->flags & BTRFS_INODE_PREALLOC && !force_cow) {
1545                 ret = run_delalloc_nocow(inode, locked_page, start, end,
1546                                          page_started, 0, nr_written);
1547         } else if (!inode_need_compress(inode)) {
1548                 ret = cow_file_range(inode, locked_page, start, end,
1549                                       page_started, nr_written, 1);
1550         } else {
1551                 set_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1552                         &BTRFS_I(inode)->runtime_flags);
1553                 ret = cow_file_range_async(inode, locked_page, start, end,
1554                                            page_started, nr_written);
1555         }
1556         return ret;
1557 }
1558
1559 static void btrfs_split_extent_hook(struct inode *inode,
1560                                     struct extent_state *orig, u64 split)
1561 {
1562         u64 size;
1563
1564         /* not delalloc, ignore it */
1565         if (!(orig->state & EXTENT_DELALLOC))
1566                 return;
1567
1568         size = orig->end - orig->start + 1;
1569         if (size > BTRFS_MAX_EXTENT_SIZE) {
1570                 u64 num_extents;
1571                 u64 new_size;
1572
1573                 /*
1574                  * See the explanation in btrfs_merge_extent_hook, the same
1575                  * applies here, just in reverse.
1576                  */
1577                 new_size = orig->end - split + 1;
1578                 num_extents = div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1579                                         BTRFS_MAX_EXTENT_SIZE);
1580                 new_size = split - orig->start;
1581                 num_extents += div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1582                                         BTRFS_MAX_EXTENT_SIZE);
1583                 if (div64_u64(size + BTRFS_MAX_EXTENT_SIZE - 1,
1584                               BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1585                         return;
1586         }
1587
1588         spin_lock(&BTRFS_I(inode)->lock);
1589         BTRFS_I(inode)->outstanding_extents++;
1590         spin_unlock(&BTRFS_I(inode)->lock);
1591 }
1592
1593 /*
1594  * extent_io.c merge_extent_hook, used to track merged delayed allocation
1595  * extents so we can keep track of new extents that are just merged onto old
1596  * extents, such as when we are doing sequential writes, so we can properly
1597  * account for the metadata space we'll need.
1598  */
1599 static void btrfs_merge_extent_hook(struct inode *inode,
1600                                     struct extent_state *new,
1601                                     struct extent_state *other)
1602 {
1603         u64 new_size, old_size;
1604         u64 num_extents;
1605
1606         /* not delalloc, ignore it */
1607         if (!(other->state & EXTENT_DELALLOC))
1608                 return;
1609
1610         if (new->start > other->start)
1611                 new_size = new->end - other->start + 1;
1612         else
1613                 new_size = other->end - new->start + 1;
1614
1615         /* we're not bigger than the max, unreserve the space and go */
1616         if (new_size <= BTRFS_MAX_EXTENT_SIZE) {
1617                 spin_lock(&BTRFS_I(inode)->lock);
1618                 BTRFS_I(inode)->outstanding_extents--;
1619                 spin_unlock(&BTRFS_I(inode)->lock);
1620                 return;
1621         }
1622
1623         /*
1624          * We have to add up either side to figure out how many extents were
1625          * accounted for before we merged into one big extent.  If the number of
1626          * extents we accounted for is <= the amount we need for the new range
1627          * then we can return, otherwise drop.  Think of it like this
1628          *
1629          * [ 4k][MAX_SIZE]
1630          *
1631          * So we've grown the extent by a MAX_SIZE extent, this would mean we
1632          * need 2 outstanding extents, on one side we have 1 and the other side
1633          * we have 1 so they are == and we can return.  But in this case
1634          *
1635          * [MAX_SIZE+4k][MAX_SIZE+4k]
1636          *
1637          * Each range on their own accounts for 2 extents, but merged together
1638          * they are only 3 extents worth of accounting, so we need to drop in
1639          * this case.
1640          */
1641         old_size = other->end - other->start + 1;
1642         num_extents = div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1643                                 BTRFS_MAX_EXTENT_SIZE);
1644         old_size = new->end - new->start + 1;
1645         num_extents += div64_u64(old_size + BTRFS_MAX_EXTENT_SIZE - 1,
1646                                  BTRFS_MAX_EXTENT_SIZE);
1647
1648         if (div64_u64(new_size + BTRFS_MAX_EXTENT_SIZE - 1,
1649                       BTRFS_MAX_EXTENT_SIZE) >= num_extents)
1650                 return;
1651
1652         spin_lock(&BTRFS_I(inode)->lock);
1653         BTRFS_I(inode)->outstanding_extents--;
1654         spin_unlock(&BTRFS_I(inode)->lock);
1655 }
1656
1657 static void btrfs_add_delalloc_inodes(struct btrfs_root *root,
1658                                       struct inode *inode)
1659 {
1660         spin_lock(&root->delalloc_lock);
1661         if (list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1662                 list_add_tail(&BTRFS_I(inode)->delalloc_inodes,
1663                               &root->delalloc_inodes);
1664                 set_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1665                         &BTRFS_I(inode)->runtime_flags);
1666                 root->nr_delalloc_inodes++;
1667                 if (root->nr_delalloc_inodes == 1) {
1668                         spin_lock(&root->fs_info->delalloc_root_lock);
1669                         BUG_ON(!list_empty(&root->delalloc_root));
1670                         list_add_tail(&root->delalloc_root,
1671                                       &root->fs_info->delalloc_roots);
1672                         spin_unlock(&root->fs_info->delalloc_root_lock);
1673                 }
1674         }
1675         spin_unlock(&root->delalloc_lock);
1676 }
1677
1678 static void btrfs_del_delalloc_inode(struct btrfs_root *root,
1679                                      struct inode *inode)
1680 {
1681         spin_lock(&root->delalloc_lock);
1682         if (!list_empty(&BTRFS_I(inode)->delalloc_inodes)) {
1683                 list_del_init(&BTRFS_I(inode)->delalloc_inodes);
1684                 clear_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1685                           &BTRFS_I(inode)->runtime_flags);
1686                 root->nr_delalloc_inodes--;
1687                 if (!root->nr_delalloc_inodes) {
1688                         spin_lock(&root->fs_info->delalloc_root_lock);
1689                         BUG_ON(list_empty(&root->delalloc_root));
1690                         list_del_init(&root->delalloc_root);
1691                         spin_unlock(&root->fs_info->delalloc_root_lock);
1692                 }
1693         }
1694         spin_unlock(&root->delalloc_lock);
1695 }
1696
1697 /*
1698  * extent_io.c set_bit_hook, used to track delayed allocation
1699  * bytes in this file, and to maintain the list of inodes that
1700  * have pending delalloc work to be done.
1701  */
1702 static void btrfs_set_bit_hook(struct inode *inode,
1703                                struct extent_state *state, unsigned *bits)
1704 {
1705
1706         if ((*bits & EXTENT_DEFRAG) && !(*bits & EXTENT_DELALLOC))
1707                 WARN_ON(1);
1708         /*
1709          * set_bit and clear bit hooks normally require _irqsave/restore
1710          * but in this case, we are only testing for the DELALLOC
1711          * bit, which is only set or cleared with irqs on
1712          */
1713         if (!(state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1714                 struct btrfs_root *root = BTRFS_I(inode)->root;
1715                 u64 len = state->end + 1 - state->start;
1716                 bool do_list = !btrfs_is_free_space_inode(inode);
1717
1718                 if (*bits & EXTENT_FIRST_DELALLOC) {
1719                         *bits &= ~EXTENT_FIRST_DELALLOC;
1720                 } else {
1721                         spin_lock(&BTRFS_I(inode)->lock);
1722                         BTRFS_I(inode)->outstanding_extents++;
1723                         spin_unlock(&BTRFS_I(inode)->lock);
1724                 }
1725
1726                 /* For sanity tests */
1727                 if (btrfs_test_is_dummy_root(root))
1728                         return;
1729
1730                 __percpu_counter_add(&root->fs_info->delalloc_bytes, len,
1731                                      root->fs_info->delalloc_batch);
1732                 spin_lock(&BTRFS_I(inode)->lock);
1733                 BTRFS_I(inode)->delalloc_bytes += len;
1734                 if (*bits & EXTENT_DEFRAG)
1735                         BTRFS_I(inode)->defrag_bytes += len;
1736                 if (do_list && !test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1737                                          &BTRFS_I(inode)->runtime_flags))
1738                         btrfs_add_delalloc_inodes(root, inode);
1739                 spin_unlock(&BTRFS_I(inode)->lock);
1740         }
1741 }
1742
1743 /*
1744  * extent_io.c clear_bit_hook, see set_bit_hook for why
1745  */
1746 static void btrfs_clear_bit_hook(struct inode *inode,
1747                                  struct extent_state *state,
1748                                  unsigned *bits)
1749 {
1750         u64 len = state->end + 1 - state->start;
1751         u64 num_extents = div64_u64(len + BTRFS_MAX_EXTENT_SIZE -1,
1752                                     BTRFS_MAX_EXTENT_SIZE);
1753
1754         spin_lock(&BTRFS_I(inode)->lock);
1755         if ((state->state & EXTENT_DEFRAG) && (*bits & EXTENT_DEFRAG))
1756                 BTRFS_I(inode)->defrag_bytes -= len;
1757         spin_unlock(&BTRFS_I(inode)->lock);
1758
1759         /*
1760          * set_bit and clear bit hooks normally require _irqsave/restore
1761          * but in this case, we are only testing for the DELALLOC
1762          * bit, which is only set or cleared with irqs on
1763          */
1764         if ((state->state & EXTENT_DELALLOC) && (*bits & EXTENT_DELALLOC)) {
1765                 struct btrfs_root *root = BTRFS_I(inode)->root;
1766                 bool do_list = !btrfs_is_free_space_inode(inode);
1767
1768                 if (*bits & EXTENT_FIRST_DELALLOC) {
1769                         *bits &= ~EXTENT_FIRST_DELALLOC;
1770                 } else if (!(*bits & EXTENT_DO_ACCOUNTING)) {
1771                         spin_lock(&BTRFS_I(inode)->lock);
1772                         BTRFS_I(inode)->outstanding_extents -= num_extents;
1773                         spin_unlock(&BTRFS_I(inode)->lock);
1774                 }
1775
1776                 /*
1777                  * We don't reserve metadata space for space cache inodes so we
1778                  * don't need to call dellalloc_release_metadata if there is an
1779                  * error.
1780                  */
1781                 if (*bits & EXTENT_DO_ACCOUNTING &&
1782                     root != root->fs_info->tree_root)
1783                         btrfs_delalloc_release_metadata(inode, len);
1784
1785                 /* For sanity tests. */
1786                 if (btrfs_test_is_dummy_root(root))
1787                         return;
1788
1789                 if (root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
1790                     && do_list && !(state->state & EXTENT_NORESERVE))
1791                         btrfs_free_reserved_data_space_noquota(inode,
1792                                         state->start, len);
1793
1794                 __percpu_counter_add(&root->fs_info->delalloc_bytes, -len,
1795                                      root->fs_info->delalloc_batch);
1796                 spin_lock(&BTRFS_I(inode)->lock);
1797                 BTRFS_I(inode)->delalloc_bytes -= len;
1798                 if (do_list && BTRFS_I(inode)->delalloc_bytes == 0 &&
1799                     test_bit(BTRFS_INODE_IN_DELALLOC_LIST,
1800                              &BTRFS_I(inode)->runtime_flags))
1801                         btrfs_del_delalloc_inode(root, inode);
1802                 spin_unlock(&BTRFS_I(inode)->lock);
1803         }
1804 }
1805
1806 /*
1807  * extent_io.c merge_bio_hook, this must check the chunk tree to make sure
1808  * we don't create bios that span stripes or chunks
1809  */
1810 int btrfs_merge_bio_hook(int rw, struct page *page, unsigned long offset,
1811                          size_t size, struct bio *bio,
1812                          unsigned long bio_flags)
1813 {
1814         struct btrfs_root *root = BTRFS_I(page->mapping->host)->root;
1815         u64 logical = (u64)bio->bi_iter.bi_sector << 9;
1816         u64 length = 0;
1817         u64 map_length;
1818         int ret;
1819
1820         if (bio_flags & EXTENT_BIO_COMPRESSED)
1821                 return 0;
1822
1823         length = bio->bi_iter.bi_size;
1824         map_length = length;
1825         ret = btrfs_map_block(root->fs_info, rw, logical,
1826                               &map_length, NULL, 0);
1827         /* Will always return 0 with map_multi == NULL */
1828         BUG_ON(ret < 0);
1829         if (map_length < length + size)
1830                 return 1;
1831         return 0;
1832 }
1833
1834 /*
1835  * in order to insert checksums into the metadata in large chunks,
1836  * we wait until bio submission time.   All the pages in the bio are
1837  * checksummed and sums are attached onto the ordered extent record.
1838  *
1839  * At IO completion time the cums attached on the ordered extent record
1840  * are inserted into the btree
1841  */
1842 static int __btrfs_submit_bio_start(struct inode *inode, int rw,
1843                                     struct bio *bio, int mirror_num,
1844                                     unsigned long bio_flags,
1845                                     u64 bio_offset)
1846 {
1847         struct btrfs_root *root = BTRFS_I(inode)->root;
1848         int ret = 0;
1849
1850         ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1851         BUG_ON(ret); /* -ENOMEM */
1852         return 0;
1853 }
1854
1855 /*
1856  * in order to insert checksums into the metadata in large chunks,
1857  * we wait until bio submission time.   All the pages in the bio are
1858  * checksummed and sums are attached onto the ordered extent record.
1859  *
1860  * At IO completion time the cums attached on the ordered extent record
1861  * are inserted into the btree
1862  */
1863 static int __btrfs_submit_bio_done(struct inode *inode, int rw, struct bio *bio,
1864                           int mirror_num, unsigned long bio_flags,
1865                           u64 bio_offset)
1866 {
1867         struct btrfs_root *root = BTRFS_I(inode)->root;
1868         int ret;
1869
1870         ret = btrfs_map_bio(root, rw, bio, mirror_num, 1);
1871         if (ret) {
1872                 bio->bi_error = ret;
1873                 bio_endio(bio);
1874         }
1875         return ret;
1876 }
1877
1878 /*
1879  * extent_io.c submission hook. This does the right thing for csum calculation
1880  * on write, or reading the csums from the tree before a read
1881  */
1882 static int btrfs_submit_bio_hook(struct inode *inode, int rw, struct bio *bio,
1883                           int mirror_num, unsigned long bio_flags,
1884                           u64 bio_offset)
1885 {
1886         struct btrfs_root *root = BTRFS_I(inode)->root;
1887         enum btrfs_wq_endio_type metadata = BTRFS_WQ_ENDIO_DATA;
1888         int ret = 0;
1889         int skip_sum;
1890         int async = !atomic_read(&BTRFS_I(inode)->sync_writers);
1891
1892         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
1893
1894         if (btrfs_is_free_space_inode(inode))
1895                 metadata = BTRFS_WQ_ENDIO_FREE_SPACE;
1896
1897         if (!(rw & REQ_WRITE)) {
1898                 ret = btrfs_bio_wq_end_io(root->fs_info, bio, metadata);
1899                 if (ret)
1900                         goto out;
1901
1902                 if (bio_flags & EXTENT_BIO_COMPRESSED) {
1903                         ret = btrfs_submit_compressed_read(inode, bio,
1904                                                            mirror_num,
1905                                                            bio_flags);
1906                         goto out;
1907                 } else if (!skip_sum) {
1908                         ret = btrfs_lookup_bio_sums(root, inode, bio, NULL);
1909                         if (ret)
1910                                 goto out;
1911                 }
1912                 goto mapit;
1913         } else if (async && !skip_sum) {
1914                 /* csum items have already been cloned */
1915                 if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
1916                         goto mapit;
1917                 /* we're doing a write, do the async checksumming */
1918                 ret = btrfs_wq_submit_bio(BTRFS_I(inode)->root->fs_info,
1919                                    inode, rw, bio, mirror_num,
1920                                    bio_flags, bio_offset,
1921                                    __btrfs_submit_bio_start,
1922                                    __btrfs_submit_bio_done);
1923                 goto out;
1924         } else if (!skip_sum) {
1925                 ret = btrfs_csum_one_bio(root, inode, bio, 0, 0);
1926                 if (ret)
1927                         goto out;
1928         }
1929
1930 mapit:
1931         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
1932
1933 out:
1934         if (ret < 0) {
1935                 bio->bi_error = ret;
1936                 bio_endio(bio);
1937         }
1938         return ret;
1939 }
1940
1941 /*
1942  * given a list of ordered sums record them in the inode.  This happens
1943  * at IO completion time based on sums calculated at bio submission time.
1944  */
1945 static noinline int add_pending_csums(struct btrfs_trans_handle *trans,
1946                              struct inode *inode, u64 file_offset,
1947                              struct list_head *list)
1948 {
1949         struct btrfs_ordered_sum *sum;
1950
1951         list_for_each_entry(sum, list, list) {
1952                 trans->adding_csums = 1;
1953                 btrfs_csum_file_blocks(trans,
1954                        BTRFS_I(inode)->root->fs_info->csum_root, sum);
1955                 trans->adding_csums = 0;
1956         }
1957         return 0;
1958 }
1959
1960 int btrfs_set_extent_delalloc(struct inode *inode, u64 start, u64 end,
1961                               struct extent_state **cached_state)
1962 {
1963         WARN_ON((end & (PAGE_CACHE_SIZE - 1)) == 0);
1964         return set_extent_delalloc(&BTRFS_I(inode)->io_tree, start, end,
1965                                    cached_state, GFP_NOFS);
1966 }
1967
1968 /* see btrfs_writepage_start_hook for details on why this is required */
1969 struct btrfs_writepage_fixup {
1970         struct page *page;
1971         struct btrfs_work work;
1972 };
1973
1974 static void btrfs_writepage_fixup_worker(struct btrfs_work *work)
1975 {
1976         struct btrfs_writepage_fixup *fixup;
1977         struct btrfs_ordered_extent *ordered;
1978         struct extent_state *cached_state = NULL;
1979         struct page *page;
1980         struct inode *inode;
1981         u64 page_start;
1982         u64 page_end;
1983         int ret;
1984
1985         fixup = container_of(work, struct btrfs_writepage_fixup, work);
1986         page = fixup->page;
1987 again:
1988         lock_page(page);
1989         if (!page->mapping || !PageDirty(page) || !PageChecked(page)) {
1990                 ClearPageChecked(page);
1991                 goto out_page;
1992         }
1993
1994         inode = page->mapping->host;
1995         page_start = page_offset(page);
1996         page_end = page_offset(page) + PAGE_CACHE_SIZE - 1;
1997
1998         lock_extent_bits(&BTRFS_I(inode)->io_tree, page_start, page_end,
1999                          &cached_state);
2000
2001         /* already ordered? We're done */
2002         if (PagePrivate2(page))
2003                 goto out;
2004
2005         ordered = btrfs_lookup_ordered_extent(inode, page_start);
2006         if (ordered) {
2007                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start,
2008                                      page_end, &cached_state, GFP_NOFS);
2009                 unlock_page(page);
2010                 btrfs_start_ordered_extent(inode, ordered, 1);
2011                 btrfs_put_ordered_extent(ordered);
2012                 goto again;
2013         }
2014
2015         ret = btrfs_delalloc_reserve_space(inode, page_start,
2016                                            PAGE_CACHE_SIZE);
2017         if (ret) {
2018                 mapping_set_error(page->mapping, ret);
2019                 end_extent_writepage(page, ret, page_start, page_end);
2020                 ClearPageChecked(page);
2021                 goto out;
2022          }
2023
2024         btrfs_set_extent_delalloc(inode, page_start, page_end, &cached_state);
2025         ClearPageChecked(page);
2026         set_page_dirty(page);
2027 out:
2028         unlock_extent_cached(&BTRFS_I(inode)->io_tree, page_start, page_end,
2029                              &cached_state, GFP_NOFS);
2030 out_page:
2031         unlock_page(page);
2032         page_cache_release(page);
2033         kfree(fixup);
2034 }
2035
2036 /*
2037  * There are a few paths in the higher layers of the kernel that directly
2038  * set the page dirty bit without asking the filesystem if it is a
2039  * good idea.  This causes problems because we want to make sure COW
2040  * properly happens and the data=ordered rules are followed.
2041  *
2042  * In our case any range that doesn't have the ORDERED bit set
2043  * hasn't been properly setup for IO.  We kick off an async process
2044  * to fix it up.  The async helper will wait for ordered extents, set
2045  * the delalloc bit and make it safe to write the page.
2046  */
2047 static int btrfs_writepage_start_hook(struct page *page, u64 start, u64 end)
2048 {
2049         struct inode *inode = page->mapping->host;
2050         struct btrfs_writepage_fixup *fixup;
2051         struct btrfs_root *root = BTRFS_I(inode)->root;
2052
2053         /* this page is properly in the ordered list */
2054         if (TestClearPagePrivate2(page))
2055                 return 0;
2056
2057         if (PageChecked(page))
2058                 return -EAGAIN;
2059
2060         fixup = kzalloc(sizeof(*fixup), GFP_NOFS);
2061         if (!fixup)
2062                 return -EAGAIN;
2063
2064         SetPageChecked(page);
2065         page_cache_get(page);
2066         btrfs_init_work(&fixup->work, btrfs_fixup_helper,
2067                         btrfs_writepage_fixup_worker, NULL, NULL);
2068         fixup->page = page;
2069         btrfs_queue_work(root->fs_info->fixup_workers, &fixup->work);
2070         return -EBUSY;
2071 }
2072
2073 static int insert_reserved_file_extent(struct btrfs_trans_handle *trans,
2074                                        struct inode *inode, u64 file_pos,
2075                                        u64 disk_bytenr, u64 disk_num_bytes,
2076                                        u64 num_bytes, u64 ram_bytes,
2077                                        u8 compression, u8 encryption,
2078                                        u16 other_encoding, int extent_type)
2079 {
2080         struct btrfs_root *root = BTRFS_I(inode)->root;
2081         struct btrfs_file_extent_item *fi;
2082         struct btrfs_path *path;
2083         struct extent_buffer *leaf;
2084         struct btrfs_key ins;
2085         int extent_inserted = 0;
2086         int ret;
2087
2088         path = btrfs_alloc_path();
2089         if (!path)
2090                 return -ENOMEM;
2091
2092         /*
2093          * we may be replacing one extent in the tree with another.
2094          * The new extent is pinned in the extent map, and we don't want
2095          * to drop it from the cache until it is completely in the btree.
2096          *
2097          * So, tell btrfs_drop_extents to leave this extent in the cache.
2098          * the caller is expected to unpin it and allow it to be merged
2099          * with the others.
2100          */
2101         ret = __btrfs_drop_extents(trans, root, inode, path, file_pos,
2102                                    file_pos + num_bytes, NULL, 0,
2103                                    1, sizeof(*fi), &extent_inserted);
2104         if (ret)
2105                 goto out;
2106
2107         if (!extent_inserted) {
2108                 ins.objectid = btrfs_ino(inode);
2109                 ins.offset = file_pos;
2110                 ins.type = BTRFS_EXTENT_DATA_KEY;
2111
2112                 path->leave_spinning = 1;
2113                 ret = btrfs_insert_empty_item(trans, root, path, &ins,
2114                                               sizeof(*fi));
2115                 if (ret)
2116                         goto out;
2117         }
2118         leaf = path->nodes[0];
2119         fi = btrfs_item_ptr(leaf, path->slots[0],
2120                             struct btrfs_file_extent_item);
2121         btrfs_set_file_extent_generation(leaf, fi, trans->transid);
2122         btrfs_set_file_extent_type(leaf, fi, extent_type);
2123         btrfs_set_file_extent_disk_bytenr(leaf, fi, disk_bytenr);
2124         btrfs_set_file_extent_disk_num_bytes(leaf, fi, disk_num_bytes);
2125         btrfs_set_file_extent_offset(leaf, fi, 0);
2126         btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2127         btrfs_set_file_extent_ram_bytes(leaf, fi, ram_bytes);
2128         btrfs_set_file_extent_compression(leaf, fi, compression);
2129         btrfs_set_file_extent_encryption(leaf, fi, encryption);
2130         btrfs_set_file_extent_other_encoding(leaf, fi, other_encoding);
2131
2132         btrfs_mark_buffer_dirty(leaf);
2133         btrfs_release_path(path);
2134
2135         inode_add_bytes(inode, num_bytes);
2136
2137         ins.objectid = disk_bytenr;
2138         ins.offset = disk_num_bytes;
2139         ins.type = BTRFS_EXTENT_ITEM_KEY;
2140         ret = btrfs_alloc_reserved_file_extent(trans, root,
2141                                         root->root_key.objectid,
2142                                         btrfs_ino(inode), file_pos,
2143                                         ram_bytes, &ins);
2144         /*
2145          * Release the reserved range from inode dirty range map, as it is
2146          * already moved into delayed_ref_head
2147          */
2148         btrfs_qgroup_release_data(inode, file_pos, ram_bytes);
2149 out:
2150         btrfs_free_path(path);
2151
2152         return ret;
2153 }
2154
2155 /* snapshot-aware defrag */
2156 struct sa_defrag_extent_backref {
2157         struct rb_node node;
2158         struct old_sa_defrag_extent *old;
2159         u64 root_id;
2160         u64 inum;
2161         u64 file_pos;
2162         u64 extent_offset;
2163         u64 num_bytes;
2164         u64 generation;
2165 };
2166
2167 struct old_sa_defrag_extent {
2168         struct list_head list;
2169         struct new_sa_defrag_extent *new;
2170
2171         u64 extent_offset;
2172         u64 bytenr;
2173         u64 offset;
2174         u64 len;
2175         int count;
2176 };
2177
2178 struct new_sa_defrag_extent {
2179         struct rb_root root;
2180         struct list_head head;
2181         struct btrfs_path *path;
2182         struct inode *inode;
2183         u64 file_pos;
2184         u64 len;
2185         u64 bytenr;
2186         u64 disk_len;
2187         u8 compress_type;
2188 };
2189
2190 static int backref_comp(struct sa_defrag_extent_backref *b1,
2191                         struct sa_defrag_extent_backref *b2)
2192 {
2193         if (b1->root_id < b2->root_id)
2194                 return -1;
2195         else if (b1->root_id > b2->root_id)
2196                 return 1;
2197
2198         if (b1->inum < b2->inum)
2199                 return -1;
2200         else if (b1->inum > b2->inum)
2201                 return 1;
2202
2203         if (b1->file_pos < b2->file_pos)
2204                 return -1;
2205         else if (b1->file_pos > b2->file_pos)
2206                 return 1;
2207
2208         /*
2209          * [------------------------------] ===> (a range of space)
2210          *     |<--->|   |<---->| =============> (fs/file tree A)
2211          * |<---------------------------->| ===> (fs/file tree B)
2212          *
2213          * A range of space can refer to two file extents in one tree while
2214          * refer to only one file extent in another tree.
2215          *
2216          * So we may process a disk offset more than one time(two extents in A)
2217          * and locate at the same extent(one extent in B), then insert two same
2218          * backrefs(both refer to the extent in B).
2219          */
2220         return 0;
2221 }
2222
2223 static void backref_insert(struct rb_root *root,
2224                            struct sa_defrag_extent_backref *backref)
2225 {
2226         struct rb_node **p = &root->rb_node;
2227         struct rb_node *parent = NULL;
2228         struct sa_defrag_extent_backref *entry;
2229         int ret;
2230
2231         while (*p) {
2232                 parent = *p;
2233                 entry = rb_entry(parent, struct sa_defrag_extent_backref, node);
2234
2235                 ret = backref_comp(backref, entry);
2236                 if (ret < 0)
2237                         p = &(*p)->rb_left;
2238                 else
2239                         p = &(*p)->rb_right;
2240         }
2241
2242         rb_link_node(&backref->node, parent, p);
2243         rb_insert_color(&backref->node, root);
2244 }
2245
2246 /*
2247  * Note the backref might has changed, and in this case we just return 0.
2248  */
2249 static noinline int record_one_backref(u64 inum, u64 offset, u64 root_id,
2250                                        void *ctx)
2251 {
2252         struct btrfs_file_extent_item *extent;
2253         struct btrfs_fs_info *fs_info;
2254         struct old_sa_defrag_extent *old = ctx;
2255         struct new_sa_defrag_extent *new = old->new;
2256         struct btrfs_path *path = new->path;
2257         struct btrfs_key key;
2258         struct btrfs_root *root;
2259         struct sa_defrag_extent_backref *backref;
2260         struct extent_buffer *leaf;
2261         struct inode *inode = new->inode;
2262         int slot;
2263         int ret;
2264         u64 extent_offset;
2265         u64 num_bytes;
2266
2267         if (BTRFS_I(inode)->root->root_key.objectid == root_id &&
2268             inum == btrfs_ino(inode))
2269                 return 0;
2270
2271         key.objectid = root_id;
2272         key.type = BTRFS_ROOT_ITEM_KEY;
2273         key.offset = (u64)-1;
2274
2275         fs_info = BTRFS_I(inode)->root->fs_info;
2276         root = btrfs_read_fs_root_no_name(fs_info, &key);
2277         if (IS_ERR(root)) {
2278                 if (PTR_ERR(root) == -ENOENT)
2279                         return 0;
2280                 WARN_ON(1);
2281                 pr_debug("inum=%llu, offset=%llu, root_id=%llu\n",
2282                          inum, offset, root_id);
2283                 return PTR_ERR(root);
2284         }
2285
2286         key.objectid = inum;
2287         key.type = BTRFS_EXTENT_DATA_KEY;
2288         if (offset > (u64)-1 << 32)
2289                 key.offset = 0;
2290         else
2291                 key.offset = offset;
2292
2293         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2294         if (WARN_ON(ret < 0))
2295                 return ret;
2296         ret = 0;
2297
2298         while (1) {
2299                 cond_resched();
2300
2301                 leaf = path->nodes[0];
2302                 slot = path->slots[0];
2303
2304                 if (slot >= btrfs_header_nritems(leaf)) {
2305                         ret = btrfs_next_leaf(root, path);
2306                         if (ret < 0) {
2307                                 goto out;
2308                         } else if (ret > 0) {
2309                                 ret = 0;
2310                                 goto out;
2311                         }
2312                         continue;
2313                 }
2314
2315                 path->slots[0]++;
2316
2317                 btrfs_item_key_to_cpu(leaf, &key, slot);
2318
2319                 if (key.objectid > inum)
2320                         goto out;
2321
2322                 if (key.objectid < inum || key.type != BTRFS_EXTENT_DATA_KEY)
2323                         continue;
2324
2325                 extent = btrfs_item_ptr(leaf, slot,
2326                                         struct btrfs_file_extent_item);
2327
2328                 if (btrfs_file_extent_disk_bytenr(leaf, extent) != old->bytenr)
2329                         continue;
2330
2331                 /*
2332                  * 'offset' refers to the exact key.offset,
2333                  * NOT the 'offset' field in btrfs_extent_data_ref, ie.
2334                  * (key.offset - extent_offset).
2335                  */
2336                 if (key.offset != offset)
2337                         continue;
2338
2339                 extent_offset = btrfs_file_extent_offset(leaf, extent);
2340                 num_bytes = btrfs_file_extent_num_bytes(leaf, extent);
2341
2342                 if (extent_offset >= old->extent_offset + old->offset +
2343                     old->len || extent_offset + num_bytes <=
2344                     old->extent_offset + old->offset)
2345                         continue;
2346                 break;
2347         }
2348
2349         backref = kmalloc(sizeof(*backref), GFP_NOFS);
2350         if (!backref) {
2351                 ret = -ENOENT;
2352                 goto out;
2353         }
2354
2355         backref->root_id = root_id;
2356         backref->inum = inum;
2357         backref->file_pos = offset;
2358         backref->num_bytes = num_bytes;
2359         backref->extent_offset = extent_offset;
2360         backref->generation = btrfs_file_extent_generation(leaf, extent);
2361         backref->old = old;
2362         backref_insert(&new->root, backref);
2363         old->count++;
2364 out:
2365         btrfs_release_path(path);
2366         WARN_ON(ret);
2367         return ret;
2368 }
2369
2370 static noinline bool record_extent_backrefs(struct btrfs_path *path,
2371                                    struct new_sa_defrag_extent *new)
2372 {
2373         struct btrfs_fs_info *fs_info = BTRFS_I(new->inode)->root->fs_info;
2374         struct old_sa_defrag_extent *old, *tmp;
2375         int ret;
2376
2377         new->path = path;
2378
2379         list_for_each_entry_safe(old, tmp, &new->head, list) {
2380                 ret = iterate_inodes_from_logical(old->bytenr +
2381                                                   old->extent_offset, fs_info,
2382                                                   path, record_one_backref,
2383                                                   old);
2384                 if (ret < 0 && ret != -ENOENT)
2385                         return false;
2386
2387                 /* no backref to be processed for this extent */
2388                 if (!old->count) {
2389                         list_del(&old->list);
2390                         kfree(old);
2391                 }
2392         }
2393
2394         if (list_empty(&new->head))
2395                 return false;
2396
2397         return true;
2398 }
2399
2400 static int relink_is_mergable(struct extent_buffer *leaf,
2401                               struct btrfs_file_extent_item *fi,
2402                               struct new_sa_defrag_extent *new)
2403 {
2404         if (btrfs_file_extent_disk_bytenr(leaf, fi) != new->bytenr)
2405                 return 0;
2406
2407         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2408                 return 0;
2409
2410         if (btrfs_file_extent_compression(leaf, fi) != new->compress_type)
2411                 return 0;
2412
2413         if (btrfs_file_extent_encryption(leaf, fi) ||
2414             btrfs_file_extent_other_encoding(leaf, fi))
2415                 return 0;
2416
2417         return 1;
2418 }
2419
2420 /*
2421  * Note the backref might has changed, and in this case we just return 0.
2422  */
2423 static noinline int relink_extent_backref(struct btrfs_path *path,
2424                                  struct sa_defrag_extent_backref *prev,
2425                                  struct sa_defrag_extent_backref *backref)
2426 {
2427         struct btrfs_file_extent_item *extent;
2428         struct btrfs_file_extent_item *item;
2429         struct btrfs_ordered_extent *ordered;
2430         struct btrfs_trans_handle *trans;
2431         struct btrfs_fs_info *fs_info;
2432         struct btrfs_root *root;
2433         struct btrfs_key key;
2434         struct extent_buffer *leaf;
2435         struct old_sa_defrag_extent *old = backref->old;
2436         struct new_sa_defrag_extent *new = old->new;
2437         struct inode *src_inode = new->inode;
2438         struct inode *inode;
2439         struct extent_state *cached = NULL;
2440         int ret = 0;
2441         u64 start;
2442         u64 len;
2443         u64 lock_start;
2444         u64 lock_end;
2445         bool merge = false;
2446         int index;
2447
2448         if (prev && prev->root_id == backref->root_id &&
2449             prev->inum == backref->inum &&
2450             prev->file_pos + prev->num_bytes == backref->file_pos)
2451                 merge = true;
2452
2453         /* step 1: get root */
2454         key.objectid = backref->root_id;
2455         key.type = BTRFS_ROOT_ITEM_KEY;
2456         key.offset = (u64)-1;
2457
2458         fs_info = BTRFS_I(src_inode)->root->fs_info;
2459         index = srcu_read_lock(&fs_info->subvol_srcu);
2460
2461         root = btrfs_read_fs_root_no_name(fs_info, &key);
2462         if (IS_ERR(root)) {
2463                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2464                 if (PTR_ERR(root) == -ENOENT)
2465                         return 0;
2466                 return PTR_ERR(root);
2467         }
2468
2469         if (btrfs_root_readonly(root)) {
2470                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2471                 return 0;
2472         }
2473
2474         /* step 2: get inode */
2475         key.objectid = backref->inum;
2476         key.type = BTRFS_INODE_ITEM_KEY;
2477         key.offset = 0;
2478
2479         inode = btrfs_iget(fs_info->sb, &key, root, NULL);
2480         if (IS_ERR(inode)) {
2481                 srcu_read_unlock(&fs_info->subvol_srcu, index);
2482                 return 0;
2483         }
2484
2485         srcu_read_unlock(&fs_info->subvol_srcu, index);
2486
2487         /* step 3: relink backref */
2488         lock_start = backref->file_pos;
2489         lock_end = backref->file_pos + backref->num_bytes - 1;
2490         lock_extent_bits(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2491                          &cached);
2492
2493         ordered = btrfs_lookup_first_ordered_extent(inode, lock_end);
2494         if (ordered) {
2495                 btrfs_put_ordered_extent(ordered);
2496                 goto out_unlock;
2497         }
2498
2499         trans = btrfs_join_transaction(root);
2500         if (IS_ERR(trans)) {
2501                 ret = PTR_ERR(trans);
2502                 goto out_unlock;
2503         }
2504
2505         key.objectid = backref->inum;
2506         key.type = BTRFS_EXTENT_DATA_KEY;
2507         key.offset = backref->file_pos;
2508
2509         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2510         if (ret < 0) {
2511                 goto out_free_path;
2512         } else if (ret > 0) {
2513                 ret = 0;
2514                 goto out_free_path;
2515         }
2516
2517         extent = btrfs_item_ptr(path->nodes[0], path->slots[0],
2518                                 struct btrfs_file_extent_item);
2519
2520         if (btrfs_file_extent_generation(path->nodes[0], extent) !=
2521             backref->generation)
2522                 goto out_free_path;
2523
2524         btrfs_release_path(path);
2525
2526         start = backref->file_pos;
2527         if (backref->extent_offset < old->extent_offset + old->offset)
2528                 start += old->extent_offset + old->offset -
2529                          backref->extent_offset;
2530
2531         len = min(backref->extent_offset + backref->num_bytes,
2532                   old->extent_offset + old->offset + old->len);
2533         len -= max(backref->extent_offset, old->extent_offset + old->offset);
2534
2535         ret = btrfs_drop_extents(trans, root, inode, start,
2536                                  start + len, 1);
2537         if (ret)
2538                 goto out_free_path;
2539 again:
2540         key.objectid = btrfs_ino(inode);
2541         key.type = BTRFS_EXTENT_DATA_KEY;
2542         key.offset = start;
2543
2544         path->leave_spinning = 1;
2545         if (merge) {
2546                 struct btrfs_file_extent_item *fi;
2547                 u64 extent_len;
2548                 struct btrfs_key found_key;
2549
2550                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2551                 if (ret < 0)
2552                         goto out_free_path;
2553
2554                 path->slots[0]--;
2555                 leaf = path->nodes[0];
2556                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
2557
2558                 fi = btrfs_item_ptr(leaf, path->slots[0],
2559                                     struct btrfs_file_extent_item);
2560                 extent_len = btrfs_file_extent_num_bytes(leaf, fi);
2561
2562                 if (extent_len + found_key.offset == start &&
2563                     relink_is_mergable(leaf, fi, new)) {
2564                         btrfs_set_file_extent_num_bytes(leaf, fi,
2565                                                         extent_len + len);
2566                         btrfs_mark_buffer_dirty(leaf);
2567                         inode_add_bytes(inode, len);
2568
2569                         ret = 1;
2570                         goto out_free_path;
2571                 } else {
2572                         merge = false;
2573                         btrfs_release_path(path);
2574                         goto again;
2575                 }
2576         }
2577
2578         ret = btrfs_insert_empty_item(trans, root, path, &key,
2579                                         sizeof(*extent));
2580         if (ret) {
2581                 btrfs_abort_transaction(trans, root, ret);
2582                 goto out_free_path;
2583         }
2584
2585         leaf = path->nodes[0];
2586         item = btrfs_item_ptr(leaf, path->slots[0],
2587                                 struct btrfs_file_extent_item);
2588         btrfs_set_file_extent_disk_bytenr(leaf, item, new->bytenr);
2589         btrfs_set_file_extent_disk_num_bytes(leaf, item, new->disk_len);
2590         btrfs_set_file_extent_offset(leaf, item, start - new->file_pos);
2591         btrfs_set_file_extent_num_bytes(leaf, item, len);
2592         btrfs_set_file_extent_ram_bytes(leaf, item, new->len);
2593         btrfs_set_file_extent_generation(leaf, item, trans->transid);
2594         btrfs_set_file_extent_type(leaf, item, BTRFS_FILE_EXTENT_REG);
2595         btrfs_set_file_extent_compression(leaf, item, new->compress_type);
2596         btrfs_set_file_extent_encryption(leaf, item, 0);
2597         btrfs_set_file_extent_other_encoding(leaf, item, 0);
2598
2599         btrfs_mark_buffer_dirty(leaf);
2600         inode_add_bytes(inode, len);
2601         btrfs_release_path(path);
2602
2603         ret = btrfs_inc_extent_ref(trans, root, new->bytenr,
2604                         new->disk_len, 0,
2605                         backref->root_id, backref->inum,
2606                         new->file_pos); /* start - extent_offset */
2607         if (ret) {
2608                 btrfs_abort_transaction(trans, root, ret);
2609                 goto out_free_path;
2610         }
2611
2612         ret = 1;
2613 out_free_path:
2614         btrfs_release_path(path);
2615         path->leave_spinning = 0;
2616         btrfs_end_transaction(trans, root);
2617 out_unlock:
2618         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lock_start, lock_end,
2619                              &cached, GFP_NOFS);
2620         iput(inode);
2621         return ret;
2622 }
2623
2624 static void free_sa_defrag_extent(struct new_sa_defrag_extent *new)
2625 {
2626         struct old_sa_defrag_extent *old, *tmp;
2627
2628         if (!new)
2629                 return;
2630
2631         list_for_each_entry_safe(old, tmp, &new->head, list) {
2632                 kfree(old);
2633         }
2634         kfree(new);
2635 }
2636
2637 static void relink_file_extents(struct new_sa_defrag_extent *new)
2638 {
2639         struct btrfs_path *path;
2640         struct sa_defrag_extent_backref *backref;
2641         struct sa_defrag_extent_backref *prev = NULL;
2642         struct inode *inode;
2643         struct btrfs_root *root;
2644         struct rb_node *node;
2645         int ret;
2646
2647         inode = new->inode;
2648         root = BTRFS_I(inode)->root;
2649
2650         path = btrfs_alloc_path();
2651         if (!path)
2652                 return;
2653
2654         if (!record_extent_backrefs(path, new)) {
2655                 btrfs_free_path(path);
2656                 goto out;
2657         }
2658         btrfs_release_path(path);
2659
2660         while (1) {
2661                 node = rb_first(&new->root);
2662                 if (!node)
2663                         break;
2664                 rb_erase(node, &new->root);
2665
2666                 backref = rb_entry(node, struct sa_defrag_extent_backref, node);
2667
2668                 ret = relink_extent_backref(path, prev, backref);
2669                 WARN_ON(ret < 0);
2670
2671                 kfree(prev);
2672
2673                 if (ret == 1)
2674                         prev = backref;
2675                 else
2676                         prev = NULL;
2677                 cond_resched();
2678         }
2679         kfree(prev);
2680
2681         btrfs_free_path(path);
2682 out:
2683         free_sa_defrag_extent(new);
2684
2685         atomic_dec(&root->fs_info->defrag_running);
2686         wake_up(&root->fs_info->transaction_wait);
2687 }
2688
2689 static struct new_sa_defrag_extent *
2690 record_old_file_extents(struct inode *inode,
2691                         struct btrfs_ordered_extent *ordered)
2692 {
2693         struct btrfs_root *root = BTRFS_I(inode)->root;
2694         struct btrfs_path *path;
2695         struct btrfs_key key;
2696         struct old_sa_defrag_extent *old;
2697         struct new_sa_defrag_extent *new;
2698         int ret;
2699
2700         new = kmalloc(sizeof(*new), GFP_NOFS);
2701         if (!new)
2702                 return NULL;
2703
2704         new->inode = inode;
2705         new->file_pos = ordered->file_offset;
2706         new->len = ordered->len;
2707         new->bytenr = ordered->start;
2708         new->disk_len = ordered->disk_len;
2709         new->compress_type = ordered->compress_type;
2710         new->root = RB_ROOT;
2711         INIT_LIST_HEAD(&new->head);
2712
2713         path = btrfs_alloc_path();
2714         if (!path)
2715                 goto out_kfree;
2716
2717         key.objectid = btrfs_ino(inode);
2718         key.type = BTRFS_EXTENT_DATA_KEY;
2719         key.offset = new->file_pos;
2720
2721         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
2722         if (ret < 0)
2723                 goto out_free_path;
2724         if (ret > 0 && path->slots[0] > 0)
2725                 path->slots[0]--;
2726
2727         /* find out all the old extents for the file range */
2728         while (1) {
2729                 struct btrfs_file_extent_item *extent;
2730                 struct extent_buffer *l;
2731                 int slot;
2732                 u64 num_bytes;
2733                 u64 offset;
2734                 u64 end;
2735                 u64 disk_bytenr;
2736                 u64 extent_offset;
2737
2738                 l = path->nodes[0];
2739                 slot = path->slots[0];
2740
2741                 if (slot >= btrfs_header_nritems(l)) {
2742                         ret = btrfs_next_leaf(root, path);
2743                         if (ret < 0)
2744                                 goto out_free_path;
2745                         else if (ret > 0)
2746                                 break;
2747                         continue;
2748                 }
2749
2750                 btrfs_item_key_to_cpu(l, &key, slot);
2751
2752                 if (key.objectid != btrfs_ino(inode))
2753                         break;
2754                 if (key.type != BTRFS_EXTENT_DATA_KEY)
2755                         break;
2756                 if (key.offset >= new->file_pos + new->len)
2757                         break;
2758
2759                 extent = btrfs_item_ptr(l, slot, struct btrfs_file_extent_item);
2760
2761                 num_bytes = btrfs_file_extent_num_bytes(l, extent);
2762                 if (key.offset + num_bytes < new->file_pos)
2763                         goto next;
2764
2765                 disk_bytenr = btrfs_file_extent_disk_bytenr(l, extent);
2766                 if (!disk_bytenr)
2767                         goto next;
2768
2769                 extent_offset = btrfs_file_extent_offset(l, extent);
2770
2771                 old = kmalloc(sizeof(*old), GFP_NOFS);
2772                 if (!old)
2773                         goto out_free_path;
2774
2775                 offset = max(new->file_pos, key.offset);
2776                 end = min(new->file_pos + new->len, key.offset + num_bytes);
2777
2778                 old->bytenr = disk_bytenr;
2779                 old->extent_offset = extent_offset;
2780                 old->offset = offset - key.offset;
2781                 old->len = end - offset;
2782                 old->new = new;
2783                 old->count = 0;
2784                 list_add_tail(&old->list, &new->head);
2785 next:
2786                 path->slots[0]++;
2787                 cond_resched();
2788         }
2789
2790         btrfs_free_path(path);
2791         atomic_inc(&root->fs_info->defrag_running);
2792
2793         return new;
2794
2795 out_free_path:
2796         btrfs_free_path(path);
2797 out_kfree:
2798         free_sa_defrag_extent(new);
2799         return NULL;
2800 }
2801
2802 static void btrfs_release_delalloc_bytes(struct btrfs_root *root,
2803                                          u64 start, u64 len)
2804 {
2805         struct btrfs_block_group_cache *cache;
2806
2807         cache = btrfs_lookup_block_group(root->fs_info, start);
2808         ASSERT(cache);
2809
2810         spin_lock(&cache->lock);
2811         cache->delalloc_bytes -= len;
2812         spin_unlock(&cache->lock);
2813
2814         btrfs_put_block_group(cache);
2815 }
2816
2817 /* as ordered data IO finishes, this gets called so we can finish
2818  * an ordered extent if the range of bytes in the file it covers are
2819  * fully written.
2820  */
2821 static int btrfs_finish_ordered_io(struct btrfs_ordered_extent *ordered_extent)
2822 {
2823         struct inode *inode = ordered_extent->inode;
2824         struct btrfs_root *root = BTRFS_I(inode)->root;
2825         struct btrfs_trans_handle *trans = NULL;
2826         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
2827         struct extent_state *cached_state = NULL;
2828         struct new_sa_defrag_extent *new = NULL;
2829         int compress_type = 0;
2830         int ret = 0;
2831         u64 logical_len = ordered_extent->len;
2832         bool nolock;
2833         bool truncated = false;
2834
2835         nolock = btrfs_is_free_space_inode(inode);
2836
2837         if (test_bit(BTRFS_ORDERED_IOERR, &ordered_extent->flags)) {
2838                 ret = -EIO;
2839                 goto out;
2840         }
2841
2842         btrfs_free_io_failure_record(inode, ordered_extent->file_offset,
2843                                      ordered_extent->file_offset +
2844                                      ordered_extent->len - 1);
2845
2846         if (test_bit(BTRFS_ORDERED_TRUNCATED, &ordered_extent->flags)) {
2847                 truncated = true;
2848                 logical_len = ordered_extent->truncated_len;
2849                 /* Truncated the entire extent, don't bother adding */
2850                 if (!logical_len)
2851                         goto out;
2852         }
2853
2854         if (test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags)) {
2855                 BUG_ON(!list_empty(&ordered_extent->list)); /* Logic error */
2856
2857                 /*
2858                  * For mwrite(mmap + memset to write) case, we still reserve
2859                  * space for NOCOW range.
2860                  * As NOCOW won't cause a new delayed ref, just free the space
2861                  */
2862                 btrfs_qgroup_free_data(inode, ordered_extent->file_offset,
2863                                        ordered_extent->len);
2864                 btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2865                 if (nolock)
2866                         trans = btrfs_join_transaction_nolock(root);
2867                 else
2868                         trans = btrfs_join_transaction(root);
2869                 if (IS_ERR(trans)) {
2870                         ret = PTR_ERR(trans);
2871                         trans = NULL;
2872                         goto out;
2873                 }
2874                 trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2875                 ret = btrfs_update_inode_fallback(trans, root, inode);
2876                 if (ret) /* -ENOMEM or corruption */
2877                         btrfs_abort_transaction(trans, root, ret);
2878                 goto out;
2879         }
2880
2881         lock_extent_bits(io_tree, ordered_extent->file_offset,
2882                          ordered_extent->file_offset + ordered_extent->len - 1,
2883                          &cached_state);
2884
2885         ret = test_range_bit(io_tree, ordered_extent->file_offset,
2886                         ordered_extent->file_offset + ordered_extent->len - 1,
2887                         EXTENT_DEFRAG, 1, cached_state);
2888         if (ret) {
2889                 u64 last_snapshot = btrfs_root_last_snapshot(&root->root_item);
2890                 if (0 && last_snapshot >= BTRFS_I(inode)->generation)
2891                         /* the inode is shared */
2892                         new = record_old_file_extents(inode, ordered_extent);
2893
2894                 clear_extent_bit(io_tree, ordered_extent->file_offset,
2895                         ordered_extent->file_offset + ordered_extent->len - 1,
2896                         EXTENT_DEFRAG, 0, 0, &cached_state, GFP_NOFS);
2897         }
2898
2899         if (nolock)
2900                 trans = btrfs_join_transaction_nolock(root);
2901         else
2902                 trans = btrfs_join_transaction(root);
2903         if (IS_ERR(trans)) {
2904                 ret = PTR_ERR(trans);
2905                 trans = NULL;
2906                 goto out_unlock;
2907         }
2908
2909         trans->block_rsv = &root->fs_info->delalloc_block_rsv;
2910
2911         if (test_bit(BTRFS_ORDERED_COMPRESSED, &ordered_extent->flags))
2912                 compress_type = ordered_extent->compress_type;
2913         if (test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags)) {
2914                 BUG_ON(compress_type);
2915                 ret = btrfs_mark_extent_written(trans, inode,
2916                                                 ordered_extent->file_offset,
2917                                                 ordered_extent->file_offset +
2918                                                 logical_len);
2919         } else {
2920                 BUG_ON(root == root->fs_info->tree_root);
2921                 ret = insert_reserved_file_extent(trans, inode,
2922                                                 ordered_extent->file_offset,
2923                                                 ordered_extent->start,
2924                                                 ordered_extent->disk_len,
2925                                                 logical_len, logical_len,
2926                                                 compress_type, 0, 0,
2927                                                 BTRFS_FILE_EXTENT_REG);
2928                 if (!ret)
2929                         btrfs_release_delalloc_bytes(root,
2930                                                      ordered_extent->start,
2931                                                      ordered_extent->disk_len);
2932         }
2933         unpin_extent_cache(&BTRFS_I(inode)->extent_tree,
2934                            ordered_extent->file_offset, ordered_extent->len,
2935                            trans->transid);
2936         if (ret < 0) {
2937                 btrfs_abort_transaction(trans, root, ret);
2938                 goto out_unlock;
2939         }
2940
2941         add_pending_csums(trans, inode, ordered_extent->file_offset,
2942                           &ordered_extent->list);
2943
2944         btrfs_ordered_update_i_size(inode, 0, ordered_extent);
2945         ret = btrfs_update_inode_fallback(trans, root, inode);
2946         if (ret) { /* -ENOMEM or corruption */
2947                 btrfs_abort_transaction(trans, root, ret);
2948                 goto out_unlock;
2949         }
2950         ret = 0;
2951 out_unlock:
2952         unlock_extent_cached(io_tree, ordered_extent->file_offset,
2953                              ordered_extent->file_offset +
2954                              ordered_extent->len - 1, &cached_state, GFP_NOFS);
2955 out:
2956         if (root != root->fs_info->tree_root)
2957                 btrfs_delalloc_release_metadata(inode, ordered_extent->len);
2958         if (trans)
2959                 btrfs_end_transaction(trans, root);
2960
2961         if (ret || truncated) {
2962                 u64 start, end;
2963
2964                 if (truncated)
2965                         start = ordered_extent->file_offset + logical_len;
2966                 else
2967                         start = ordered_extent->file_offset;
2968                 end = ordered_extent->file_offset + ordered_extent->len - 1;
2969                 clear_extent_uptodate(io_tree, start, end, NULL, GFP_NOFS);
2970
2971                 /* Drop the cache for the part of the extent we didn't write. */
2972                 btrfs_drop_extent_cache(inode, start, end, 0);
2973
2974                 /*
2975                  * If the ordered extent had an IOERR or something else went
2976                  * wrong we need to return the space for this ordered extent
2977                  * back to the allocator.  We only free the extent in the
2978                  * truncated case if we didn't write out the extent at all.
2979                  */
2980                 if ((ret || !logical_len) &&
2981                     !test_bit(BTRFS_ORDERED_NOCOW, &ordered_extent->flags) &&
2982                     !test_bit(BTRFS_ORDERED_PREALLOC, &ordered_extent->flags))
2983                         btrfs_free_reserved_extent(root, ordered_extent->start,
2984                                                    ordered_extent->disk_len, 1);
2985         }
2986
2987
2988         /*
2989          * This needs to be done to make sure anybody waiting knows we are done
2990          * updating everything for this ordered extent.
2991          */
2992         btrfs_remove_ordered_extent(inode, ordered_extent);
2993
2994         /* for snapshot-aware defrag */
2995         if (new) {
2996                 if (ret) {
2997                         free_sa_defrag_extent(new);
2998                         atomic_dec(&root->fs_info->defrag_running);
2999                 } else {
3000                         relink_file_extents(new);
3001                 }
3002         }
3003
3004         /* once for us */
3005         btrfs_put_ordered_extent(ordered_extent);
3006         /* once for the tree */
3007         btrfs_put_ordered_extent(ordered_extent);
3008
3009         return ret;
3010 }
3011
3012 static void finish_ordered_fn(struct btrfs_work *work)
3013 {
3014         struct btrfs_ordered_extent *ordered_extent;
3015         ordered_extent = container_of(work, struct btrfs_ordered_extent, work);
3016         btrfs_finish_ordered_io(ordered_extent);
3017 }
3018
3019 static int btrfs_writepage_end_io_hook(struct page *page, u64 start, u64 end,
3020                                 struct extent_state *state, int uptodate)
3021 {
3022         struct inode *inode = page->mapping->host;
3023         struct btrfs_root *root = BTRFS_I(inode)->root;
3024         struct btrfs_ordered_extent *ordered_extent = NULL;
3025         struct btrfs_workqueue *wq;
3026         btrfs_work_func_t func;
3027
3028         trace_btrfs_writepage_end_io_hook(page, start, end, uptodate);
3029
3030         ClearPagePrivate2(page);
3031         if (!btrfs_dec_test_ordered_pending(inode, &ordered_extent, start,
3032                                             end - start + 1, uptodate))
3033                 return 0;
3034
3035         if (btrfs_is_free_space_inode(inode)) {
3036                 wq = root->fs_info->endio_freespace_worker;
3037                 func = btrfs_freespace_write_helper;
3038         } else {
3039                 wq = root->fs_info->endio_write_workers;
3040                 func = btrfs_endio_write_helper;
3041         }
3042
3043         btrfs_init_work(&ordered_extent->work, func, finish_ordered_fn, NULL,
3044                         NULL);
3045         btrfs_queue_work(wq, &ordered_extent->work);
3046
3047         return 0;
3048 }
3049
3050 static int __readpage_endio_check(struct inode *inode,
3051                                   struct btrfs_io_bio *io_bio,
3052                                   int icsum, struct page *page,
3053                                   int pgoff, u64 start, size_t len)
3054 {
3055         char *kaddr;
3056         u32 csum_expected;
3057         u32 csum = ~(u32)0;
3058
3059         csum_expected = *(((u32 *)io_bio->csum) + icsum);
3060
3061         kaddr = kmap_atomic(page);
3062         csum = btrfs_csum_data(kaddr + pgoff, csum,  len);
3063         btrfs_csum_final(csum, (char *)&csum);
3064         if (csum != csum_expected)
3065                 goto zeroit;
3066
3067         kunmap_atomic(kaddr);
3068         return 0;
3069 zeroit:
3070         btrfs_warn_rl(BTRFS_I(inode)->root->fs_info,
3071                 "csum failed ino %llu off %llu csum %u expected csum %u",
3072                            btrfs_ino(inode), start, csum, csum_expected);
3073         memset(kaddr + pgoff, 1, len);
3074         flush_dcache_page(page);
3075         kunmap_atomic(kaddr);
3076         if (csum_expected == 0)
3077                 return 0;
3078         return -EIO;
3079 }
3080
3081 /*
3082  * when reads are done, we need to check csums to verify the data is correct
3083  * if there's a match, we allow the bio to finish.  If not, the code in
3084  * extent_io.c will try to find good copies for us.
3085  */
3086 static int btrfs_readpage_end_io_hook(struct btrfs_io_bio *io_bio,
3087                                       u64 phy_offset, struct page *page,
3088                                       u64 start, u64 end, int mirror)
3089 {
3090         size_t offset = start - page_offset(page);
3091         struct inode *inode = page->mapping->host;
3092         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
3093         struct btrfs_root *root = BTRFS_I(inode)->root;
3094
3095         if (PageChecked(page)) {
3096                 ClearPageChecked(page);
3097                 return 0;
3098         }
3099
3100         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM)
3101                 return 0;
3102
3103         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID &&
3104             test_range_bit(io_tree, start, end, EXTENT_NODATASUM, 1, NULL)) {
3105                 clear_extent_bits(io_tree, start, end, EXTENT_NODATASUM,
3106                                   GFP_NOFS);
3107                 return 0;
3108         }
3109
3110         phy_offset >>= inode->i_sb->s_blocksize_bits;
3111         return __readpage_endio_check(inode, io_bio, phy_offset, page, offset,
3112                                       start, (size_t)(end - start + 1));
3113 }
3114
3115 void btrfs_add_delayed_iput(struct inode *inode)
3116 {
3117         struct btrfs_fs_info *fs_info = BTRFS_I(inode)->root->fs_info;
3118         struct btrfs_inode *binode = BTRFS_I(inode);
3119
3120         if (atomic_add_unless(&inode->i_count, -1, 1))
3121                 return;
3122
3123         spin_lock(&fs_info->delayed_iput_lock);
3124         if (binode->delayed_iput_count == 0) {
3125                 ASSERT(list_empty(&binode->delayed_iput));
3126                 list_add_tail(&binode->delayed_iput, &fs_info->delayed_iputs);
3127         } else {
3128                 binode->delayed_iput_count++;
3129         }
3130         spin_unlock(&fs_info->delayed_iput_lock);
3131 }
3132
3133 void btrfs_run_delayed_iputs(struct btrfs_root *root)
3134 {
3135         struct btrfs_fs_info *fs_info = root->fs_info;
3136
3137         spin_lock(&fs_info->delayed_iput_lock);
3138         while (!list_empty(&fs_info->delayed_iputs)) {
3139                 struct btrfs_inode *inode;
3140
3141                 inode = list_first_entry(&fs_info->delayed_iputs,
3142                                 struct btrfs_inode, delayed_iput);
3143                 if (inode->delayed_iput_count) {
3144                         inode->delayed_iput_count--;
3145                         list_move_tail(&inode->delayed_iput,
3146                                         &fs_info->delayed_iputs);
3147                 } else {
3148                         list_del_init(&inode->delayed_iput);
3149                 }
3150                 spin_unlock(&fs_info->delayed_iput_lock);
3151                 iput(&inode->vfs_inode);
3152                 spin_lock(&fs_info->delayed_iput_lock);
3153         }
3154         spin_unlock(&fs_info->delayed_iput_lock);
3155 }
3156
3157 /*
3158  * This is called in transaction commit time. If there are no orphan
3159  * files in the subvolume, it removes orphan item and frees block_rsv
3160  * structure.
3161  */
3162 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3163                               struct btrfs_root *root)
3164 {
3165         struct btrfs_block_rsv *block_rsv;
3166         int ret;
3167
3168         if (atomic_read(&root->orphan_inodes) ||
3169             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3170                 return;
3171
3172         spin_lock(&root->orphan_lock);
3173         if (atomic_read(&root->orphan_inodes)) {
3174                 spin_unlock(&root->orphan_lock);
3175                 return;
3176         }
3177
3178         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3179                 spin_unlock(&root->orphan_lock);
3180                 return;
3181         }
3182
3183         block_rsv = root->orphan_block_rsv;
3184         root->orphan_block_rsv = NULL;
3185         spin_unlock(&root->orphan_lock);
3186
3187         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3188             btrfs_root_refs(&root->root_item) > 0) {
3189                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3190                                             root->root_key.objectid);
3191                 if (ret)
3192                         btrfs_abort_transaction(trans, root, ret);
3193                 else
3194                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3195                                   &root->state);
3196         }
3197
3198         if (block_rsv) {
3199                 WARN_ON(block_rsv->size > 0);
3200                 btrfs_free_block_rsv(root, block_rsv);
3201         }
3202 }
3203
3204 /*
3205  * This creates an orphan entry for the given inode in case something goes
3206  * wrong in the middle of an unlink/truncate.
3207  *
3208  * NOTE: caller of this function should reserve 5 units of metadata for
3209  *       this function.
3210  */
3211 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3212 {
3213         struct btrfs_root *root = BTRFS_I(inode)->root;
3214         struct btrfs_block_rsv *block_rsv = NULL;
3215         int reserve = 0;
3216         int insert = 0;
3217         int ret;
3218
3219         if (!root->orphan_block_rsv) {
3220                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3221                 if (!block_rsv)
3222                         return -ENOMEM;
3223         }
3224
3225         spin_lock(&root->orphan_lock);
3226         if (!root->orphan_block_rsv) {
3227                 root->orphan_block_rsv = block_rsv;
3228         } else if (block_rsv) {
3229                 btrfs_free_block_rsv(root, block_rsv);
3230                 block_rsv = NULL;
3231         }
3232
3233         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3234                               &BTRFS_I(inode)->runtime_flags)) {
3235 #if 0
3236                 /*
3237                  * For proper ENOSPC handling, we should do orphan
3238                  * cleanup when mounting. But this introduces backward
3239                  * compatibility issue.
3240                  */
3241                 if (!xchg(&root->orphan_item_inserted, 1))
3242                         insert = 2;
3243                 else
3244                         insert = 1;
3245 #endif
3246                 insert = 1;
3247                 atomic_inc(&root->orphan_inodes);
3248         }
3249
3250         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3251                               &BTRFS_I(inode)->runtime_flags))
3252                 reserve = 1;
3253         spin_unlock(&root->orphan_lock);
3254
3255         /* grab metadata reservation from transaction handle */
3256         if (reserve) {
3257                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3258                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3259         }
3260
3261         /* insert an orphan item to track this unlinked/truncated file */
3262         if (insert >= 1) {
3263                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3264                 if (ret) {
3265                         atomic_dec(&root->orphan_inodes);
3266                         if (reserve) {
3267                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3268                                           &BTRFS_I(inode)->runtime_flags);
3269                                 btrfs_orphan_release_metadata(inode);
3270                         }
3271                         if (ret != -EEXIST) {
3272                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3273                                           &BTRFS_I(inode)->runtime_flags);
3274                                 btrfs_abort_transaction(trans, root, ret);
3275                                 return ret;
3276                         }
3277                 }
3278                 ret = 0;
3279         }
3280
3281         /* insert an orphan item to track subvolume contains orphan files */
3282         if (insert >= 2) {
3283                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3284                                                root->root_key.objectid);
3285                 if (ret && ret != -EEXIST) {
3286                         btrfs_abort_transaction(trans, root, ret);
3287                         return ret;
3288                 }
3289         }
3290         return 0;
3291 }
3292
3293 /*
3294  * We have done the truncate/delete so we can go ahead and remove the orphan
3295  * item for this particular inode.
3296  */
3297 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3298                             struct inode *inode)
3299 {
3300         struct btrfs_root *root = BTRFS_I(inode)->root;
3301         int delete_item = 0;
3302         int release_rsv = 0;
3303         int ret = 0;
3304
3305         spin_lock(&root->orphan_lock);
3306         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3307                                &BTRFS_I(inode)->runtime_flags))
3308                 delete_item = 1;
3309
3310         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3311                                &BTRFS_I(inode)->runtime_flags))
3312                 release_rsv = 1;
3313         spin_unlock(&root->orphan_lock);
3314
3315         if (delete_item) {
3316                 atomic_dec(&root->orphan_inodes);
3317                 if (trans)
3318                         ret = btrfs_del_orphan_item(trans, root,
3319                                                     btrfs_ino(inode));
3320         }
3321
3322         if (release_rsv)
3323                 btrfs_orphan_release_metadata(inode);
3324
3325         return ret;
3326 }
3327
3328 /*
3329  * this cleans up any orphans that may be left on the list from the last use
3330  * of this root.
3331  */
3332 int btrfs_orphan_cleanup(struct btrfs_root *root)
3333 {
3334         struct btrfs_path *path;
3335         struct extent_buffer *leaf;
3336         struct btrfs_key key, found_key;
3337         struct btrfs_trans_handle *trans;
3338         struct inode *inode;
3339         u64 last_objectid = 0;
3340         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3341
3342         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3343                 return 0;
3344
3345         path = btrfs_alloc_path();
3346         if (!path) {
3347                 ret = -ENOMEM;
3348                 goto out;
3349         }
3350         path->reada = READA_BACK;
3351
3352         key.objectid = BTRFS_ORPHAN_OBJECTID;
3353         key.type = BTRFS_ORPHAN_ITEM_KEY;
3354         key.offset = (u64)-1;
3355
3356         while (1) {
3357                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3358                 if (ret < 0)
3359                         goto out;
3360
3361                 /*
3362                  * if ret == 0 means we found what we were searching for, which
3363                  * is weird, but possible, so only screw with path if we didn't
3364                  * find the key and see if we have stuff that matches
3365                  */
3366                 if (ret > 0) {
3367                         ret = 0;
3368                         if (path->slots[0] == 0)
3369                                 break;
3370                         path->slots[0]--;
3371                 }
3372
3373                 /* pull out the item */
3374                 leaf = path->nodes[0];
3375                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3376
3377                 /* make sure the item matches what we want */
3378                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3379                         break;
3380                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3381                         break;
3382
3383                 /* release the path since we're done with it */
3384                 btrfs_release_path(path);
3385
3386                 /*
3387                  * this is where we are basically btrfs_lookup, without the
3388                  * crossing root thing.  we store the inode number in the
3389                  * offset of the orphan item.
3390                  */
3391
3392                 if (found_key.offset == last_objectid) {
3393                         btrfs_err(root->fs_info,
3394                                 "Error removing orphan entry, stopping orphan cleanup");
3395                         ret = -EINVAL;
3396                         goto out;
3397                 }
3398
3399                 last_objectid = found_key.offset;
3400
3401                 found_key.objectid = found_key.offset;
3402                 found_key.type = BTRFS_INODE_ITEM_KEY;
3403                 found_key.offset = 0;
3404                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3405                 ret = PTR_ERR_OR_ZERO(inode);
3406                 if (ret && ret != -ESTALE)
3407                         goto out;
3408
3409                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3410                         struct btrfs_root *dead_root;
3411                         struct btrfs_fs_info *fs_info = root->fs_info;
3412                         int is_dead_root = 0;
3413
3414                         /*
3415                          * this is an orphan in the tree root. Currently these
3416                          * could come from 2 sources:
3417                          *  a) a snapshot deletion in progress
3418                          *  b) a free space cache inode
3419                          * We need to distinguish those two, as the snapshot
3420                          * orphan must not get deleted.
3421                          * find_dead_roots already ran before us, so if this
3422                          * is a snapshot deletion, we should find the root
3423                          * in the dead_roots list
3424                          */
3425                         spin_lock(&fs_info->trans_lock);
3426                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3427                                             root_list) {
3428                                 if (dead_root->root_key.objectid ==
3429                                     found_key.objectid) {
3430                                         is_dead_root = 1;
3431                                         break;
3432                                 }
3433                         }
3434                         spin_unlock(&fs_info->trans_lock);
3435                         if (is_dead_root) {
3436                                 /* prevent this orphan from being found again */
3437                                 key.offset = found_key.objectid - 1;
3438                                 continue;
3439                         }
3440                 }
3441                 /*
3442                  * Inode is already gone but the orphan item is still there,
3443                  * kill the orphan item.
3444                  */
3445                 if (ret == -ESTALE) {
3446                         trans = btrfs_start_transaction(root, 1);
3447                         if (IS_ERR(trans)) {
3448                                 ret = PTR_ERR(trans);
3449                                 goto out;
3450                         }
3451                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3452                                 found_key.objectid);
3453                         ret = btrfs_del_orphan_item(trans, root,
3454                                                     found_key.objectid);
3455                         btrfs_end_transaction(trans, root);
3456                         if (ret)
3457                                 goto out;
3458                         continue;
3459                 }
3460
3461                 /*
3462                  * add this inode to the orphan list so btrfs_orphan_del does
3463                  * the proper thing when we hit it
3464                  */
3465                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3466                         &BTRFS_I(inode)->runtime_flags);
3467                 atomic_inc(&root->orphan_inodes);
3468
3469                 /* if we have links, this was a truncate, lets do that */
3470                 if (inode->i_nlink) {
3471                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3472                                 iput(inode);
3473                                 continue;
3474                         }
3475                         nr_truncate++;
3476
3477                         /* 1 for the orphan item deletion. */
3478                         trans = btrfs_start_transaction(root, 1);
3479                         if (IS_ERR(trans)) {
3480                                 iput(inode);
3481                                 ret = PTR_ERR(trans);
3482                                 goto out;
3483                         }
3484                         ret = btrfs_orphan_add(trans, inode);
3485                         btrfs_end_transaction(trans, root);
3486                         if (ret) {
3487                                 iput(inode);
3488                                 goto out;
3489                         }
3490
3491                         ret = btrfs_truncate(inode);
3492                         if (ret)
3493                                 btrfs_orphan_del(NULL, inode);
3494                 } else {
3495                         nr_unlink++;
3496                 }
3497
3498                 /* this will do delete_inode and everything for us */
3499                 iput(inode);
3500                 if (ret)
3501                         goto out;
3502         }
3503         /* release the path since we're done with it */
3504         btrfs_release_path(path);
3505
3506         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3507
3508         if (root->orphan_block_rsv)
3509                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3510                                         (u64)-1);
3511
3512         if (root->orphan_block_rsv ||
3513             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3514                 trans = btrfs_join_transaction(root);
3515                 if (!IS_ERR(trans))
3516                         btrfs_end_transaction(trans, root);
3517         }
3518
3519         if (nr_unlink)
3520                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3521         if (nr_truncate)
3522                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3523
3524 out:
3525         if (ret)
3526                 btrfs_err(root->fs_info,
3527                         "could not do orphan cleanup %d", ret);
3528         btrfs_free_path(path);
3529         return ret;
3530 }
3531
3532 /*
3533  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3534  * don't find any xattrs, we know there can't be any acls.
3535  *
3536  * slot is the slot the inode is in, objectid is the objectid of the inode
3537  */
3538 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3539                                           int slot, u64 objectid,
3540                                           int *first_xattr_slot)
3541 {
3542         u32 nritems = btrfs_header_nritems(leaf);
3543         struct btrfs_key found_key;
3544         static u64 xattr_access = 0;
3545         static u64 xattr_default = 0;
3546         int scanned = 0;
3547
3548         if (!xattr_access) {
3549                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3550                                         strlen(POSIX_ACL_XATTR_ACCESS));
3551                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3552                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3553         }
3554
3555         slot++;
3556         *first_xattr_slot = -1;
3557         while (slot < nritems) {
3558                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3559
3560                 /* we found a different objectid, there must not be acls */
3561                 if (found_key.objectid != objectid)
3562                         return 0;
3563
3564                 /* we found an xattr, assume we've got an acl */
3565                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3566                         if (*first_xattr_slot == -1)
3567                                 *first_xattr_slot = slot;
3568                         if (found_key.offset == xattr_access ||
3569                             found_key.offset == xattr_default)
3570                                 return 1;
3571                 }
3572
3573                 /*
3574                  * we found a key greater than an xattr key, there can't
3575                  * be any acls later on
3576                  */
3577                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3578                         return 0;
3579
3580                 slot++;
3581                 scanned++;
3582
3583                 /*
3584                  * it goes inode, inode backrefs, xattrs, extents,
3585                  * so if there are a ton of hard links to an inode there can
3586                  * be a lot of backrefs.  Don't waste time searching too hard,
3587                  * this is just an optimization
3588                  */
3589                 if (scanned >= 8)
3590                         break;
3591         }
3592         /* we hit the end of the leaf before we found an xattr or
3593          * something larger than an xattr.  We have to assume the inode
3594          * has acls
3595          */
3596         if (*first_xattr_slot == -1)
3597                 *first_xattr_slot = slot;
3598         return 1;
3599 }
3600
3601 /*
3602  * read an inode from the btree into the in-memory inode
3603  */
3604 static void btrfs_read_locked_inode(struct inode *inode)
3605 {
3606         struct btrfs_path *path;
3607         struct extent_buffer *leaf;
3608         struct btrfs_inode_item *inode_item;
3609         struct btrfs_root *root = BTRFS_I(inode)->root;
3610         struct btrfs_key location;
3611         unsigned long ptr;
3612         int maybe_acls;
3613         u32 rdev;
3614         int ret;
3615         bool filled = false;
3616         int first_xattr_slot;
3617
3618         ret = btrfs_fill_inode(inode, &rdev);
3619         if (!ret)
3620                 filled = true;
3621
3622         path = btrfs_alloc_path();
3623         if (!path)
3624                 goto make_bad;
3625
3626         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3627
3628         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3629         if (ret)
3630                 goto make_bad;
3631
3632         leaf = path->nodes[0];
3633
3634         if (filled)
3635                 goto cache_index;
3636
3637         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3638                                     struct btrfs_inode_item);
3639         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3640         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3641         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3642         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3643         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3644
3645         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3646         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3647
3648         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3649         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3650
3651         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3652         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3653
3654         BTRFS_I(inode)->i_otime.tv_sec =
3655                 btrfs_timespec_sec(leaf, &inode_item->otime);
3656         BTRFS_I(inode)->i_otime.tv_nsec =
3657                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3658
3659         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3660         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3661         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3662
3663         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3664         inode->i_generation = BTRFS_I(inode)->generation;
3665         inode->i_rdev = 0;
3666         rdev = btrfs_inode_rdev(leaf, inode_item);
3667
3668         BTRFS_I(inode)->index_cnt = (u64)-1;
3669         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3670
3671 cache_index:
3672         /*
3673          * If we were modified in the current generation and evicted from memory
3674          * and then re-read we need to do a full sync since we don't have any
3675          * idea about which extents were modified before we were evicted from
3676          * cache.
3677          *
3678          * This is required for both inode re-read from disk and delayed inode
3679          * in delayed_nodes_tree.
3680          */
3681         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3682                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3683                         &BTRFS_I(inode)->runtime_flags);
3684
3685         /*
3686          * We don't persist the id of the transaction where an unlink operation
3687          * against the inode was last made. So here we assume the inode might
3688          * have been evicted, and therefore the exact value of last_unlink_trans
3689          * lost, and set it to last_trans to avoid metadata inconsistencies
3690          * between the inode and its parent if the inode is fsync'ed and the log
3691          * replayed. For example, in the scenario:
3692          *
3693          * touch mydir/foo
3694          * ln mydir/foo mydir/bar
3695          * sync
3696          * unlink mydir/bar
3697          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3698          * xfs_io -c fsync mydir/foo
3699          * <power failure>
3700          * mount fs, triggers fsync log replay
3701          *
3702          * We must make sure that when we fsync our inode foo we also log its
3703          * parent inode, otherwise after log replay the parent still has the
3704          * dentry with the "bar" name but our inode foo has a link count of 1
3705          * and doesn't have an inode ref with the name "bar" anymore.
3706          *
3707          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3708          * but it guarantees correctness at the expense of ocassional full
3709          * transaction commits on fsync if our inode is a directory, or if our
3710          * inode is not a directory, logging its parent unnecessarily.
3711          */
3712         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3713
3714         path->slots[0]++;
3715         if (inode->i_nlink != 1 ||
3716             path->slots[0] >= btrfs_header_nritems(leaf))
3717                 goto cache_acl;
3718
3719         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3720         if (location.objectid != btrfs_ino(inode))
3721                 goto cache_acl;
3722
3723         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3724         if (location.type == BTRFS_INODE_REF_KEY) {
3725                 struct btrfs_inode_ref *ref;
3726
3727                 ref = (struct btrfs_inode_ref *)ptr;
3728                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3729         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3730                 struct btrfs_inode_extref *extref;
3731
3732                 extref = (struct btrfs_inode_extref *)ptr;
3733                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3734                                                                      extref);
3735         }
3736 cache_acl:
3737         /*
3738          * try to precache a NULL acl entry for files that don't have
3739          * any xattrs or acls
3740          */
3741         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3742                                            btrfs_ino(inode), &first_xattr_slot);
3743         if (first_xattr_slot != -1) {
3744                 path->slots[0] = first_xattr_slot;
3745                 ret = btrfs_load_inode_props(inode, path);
3746                 if (ret)
3747                         btrfs_err(root->fs_info,
3748                                   "error loading props for ino %llu (root %llu): %d",
3749                                   btrfs_ino(inode),
3750                                   root->root_key.objectid, ret);
3751         }
3752         btrfs_free_path(path);
3753
3754         if (!maybe_acls)
3755                 cache_no_acl(inode);
3756
3757         switch (inode->i_mode & S_IFMT) {
3758         case S_IFREG:
3759                 inode->i_mapping->a_ops = &btrfs_aops;
3760                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3761                 inode->i_fop = &btrfs_file_operations;
3762                 inode->i_op = &btrfs_file_inode_operations;
3763                 break;
3764         case S_IFDIR:
3765                 inode->i_fop = &btrfs_dir_file_operations;
3766                 if (root == root->fs_info->tree_root)
3767                         inode->i_op = &btrfs_dir_ro_inode_operations;
3768                 else
3769                         inode->i_op = &btrfs_dir_inode_operations;
3770                 break;
3771         case S_IFLNK:
3772                 inode->i_op = &btrfs_symlink_inode_operations;
3773                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3774                 break;
3775         default:
3776                 inode->i_op = &btrfs_special_inode_operations;
3777                 init_special_inode(inode, inode->i_mode, rdev);
3778                 break;
3779         }
3780
3781         btrfs_update_iflags(inode);
3782         return;
3783
3784 make_bad:
3785         btrfs_free_path(path);
3786         make_bad_inode(inode);
3787 }
3788
3789 /*
3790  * given a leaf and an inode, copy the inode fields into the leaf
3791  */
3792 static void fill_inode_item(struct btrfs_trans_handle *trans,
3793                             struct extent_buffer *leaf,
3794                             struct btrfs_inode_item *item,
3795                             struct inode *inode)
3796 {
3797         struct btrfs_map_token token;
3798
3799         btrfs_init_map_token(&token);
3800
3801         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3802         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3803         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3804                                    &token);
3805         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3806         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3807
3808         btrfs_set_token_timespec_sec(leaf, &item->atime,
3809                                      inode->i_atime.tv_sec, &token);
3810         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3811                                       inode->i_atime.tv_nsec, &token);
3812
3813         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3814                                      inode->i_mtime.tv_sec, &token);
3815         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3816                                       inode->i_mtime.tv_nsec, &token);
3817
3818         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3819                                      inode->i_ctime.tv_sec, &token);
3820         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3821                                       inode->i_ctime.tv_nsec, &token);
3822
3823         btrfs_set_token_timespec_sec(leaf, &item->otime,
3824                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3825         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3826                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3827
3828         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3829                                      &token);
3830         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3831                                          &token);
3832         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3833         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3834         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3835         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3836         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3837 }
3838
3839 /*
3840  * copy everything in the in-memory inode into the btree.
3841  */
3842 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3843                                 struct btrfs_root *root, struct inode *inode)
3844 {
3845         struct btrfs_inode_item *inode_item;
3846         struct btrfs_path *path;
3847         struct extent_buffer *leaf;
3848         int ret;
3849
3850         path = btrfs_alloc_path();
3851         if (!path)
3852                 return -ENOMEM;
3853
3854         path->leave_spinning = 1;
3855         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3856                                  1);
3857         if (ret) {
3858                 if (ret > 0)
3859                         ret = -ENOENT;
3860                 goto failed;
3861         }
3862
3863         leaf = path->nodes[0];
3864         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3865                                     struct btrfs_inode_item);
3866
3867         fill_inode_item(trans, leaf, inode_item, inode);
3868         btrfs_mark_buffer_dirty(leaf);
3869         btrfs_set_inode_last_trans(trans, inode);
3870         ret = 0;
3871 failed:
3872         btrfs_free_path(path);
3873         return ret;
3874 }
3875
3876 /*
3877  * copy everything in the in-memory inode into the btree.
3878  */
3879 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3880                                 struct btrfs_root *root, struct inode *inode)
3881 {
3882         int ret;
3883
3884         /*
3885          * If the inode is a free space inode, we can deadlock during commit
3886          * if we put it into the delayed code.
3887          *
3888          * The data relocation inode should also be directly updated
3889          * without delay
3890          */
3891         if (!btrfs_is_free_space_inode(inode)
3892             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3893             && !root->fs_info->log_root_recovering) {
3894                 btrfs_update_root_times(trans, root);
3895
3896                 ret = btrfs_delayed_update_inode(trans, root, inode);
3897                 if (!ret)
3898                         btrfs_set_inode_last_trans(trans, inode);
3899                 return ret;
3900         }
3901
3902         return btrfs_update_inode_item(trans, root, inode);
3903 }
3904
3905 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3906                                          struct btrfs_root *root,
3907                                          struct inode *inode)
3908 {
3909         int ret;
3910
3911         ret = btrfs_update_inode(trans, root, inode);
3912         if (ret == -ENOSPC)
3913                 return btrfs_update_inode_item(trans, root, inode);
3914         return ret;
3915 }
3916
3917 /*
3918  * unlink helper that gets used here in inode.c and in the tree logging
3919  * recovery code.  It remove a link in a directory with a given name, and
3920  * also drops the back refs in the inode to the directory
3921  */
3922 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3923                                 struct btrfs_root *root,
3924                                 struct inode *dir, struct inode *inode,
3925                                 const char *name, int name_len)
3926 {
3927         struct btrfs_path *path;
3928         int ret = 0;
3929         struct extent_buffer *leaf;
3930         struct btrfs_dir_item *di;
3931         struct btrfs_key key;
3932         u64 index;
3933         u64 ino = btrfs_ino(inode);
3934         u64 dir_ino = btrfs_ino(dir);
3935
3936         path = btrfs_alloc_path();
3937         if (!path) {
3938                 ret = -ENOMEM;
3939                 goto out;
3940         }
3941
3942         path->leave_spinning = 1;
3943         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3944                                     name, name_len, -1);
3945         if (IS_ERR(di)) {
3946                 ret = PTR_ERR(di);
3947                 goto err;
3948         }
3949         if (!di) {
3950                 ret = -ENOENT;
3951                 goto err;
3952         }
3953         leaf = path->nodes[0];
3954         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3955         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3956         if (ret)
3957                 goto err;
3958         btrfs_release_path(path);
3959
3960         /*
3961          * If we don't have dir index, we have to get it by looking up
3962          * the inode ref, since we get the inode ref, remove it directly,
3963          * it is unnecessary to do delayed deletion.
3964          *
3965          * But if we have dir index, needn't search inode ref to get it.
3966          * Since the inode ref is close to the inode item, it is better
3967          * that we delay to delete it, and just do this deletion when
3968          * we update the inode item.
3969          */
3970         if (BTRFS_I(inode)->dir_index) {
3971                 ret = btrfs_delayed_delete_inode_ref(inode);
3972                 if (!ret) {
3973                         index = BTRFS_I(inode)->dir_index;
3974                         goto skip_backref;
3975                 }
3976         }
3977
3978         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3979                                   dir_ino, &index);
3980         if (ret) {
3981                 btrfs_info(root->fs_info,
3982                         "failed to delete reference to %.*s, inode %llu parent %llu",
3983                         name_len, name, ino, dir_ino);
3984                 btrfs_abort_transaction(trans, root, ret);
3985                 goto err;
3986         }
3987 skip_backref:
3988         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3989         if (ret) {
3990                 btrfs_abort_transaction(trans, root, ret);
3991                 goto err;
3992         }
3993
3994         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3995                                          inode, dir_ino);
3996         if (ret != 0 && ret != -ENOENT) {
3997                 btrfs_abort_transaction(trans, root, ret);
3998                 goto err;
3999         }
4000
4001         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4002                                            dir, index);
4003         if (ret == -ENOENT)
4004                 ret = 0;
4005         else if (ret)
4006                 btrfs_abort_transaction(trans, root, ret);
4007 err:
4008         btrfs_free_path(path);
4009         if (ret)
4010                 goto out;
4011
4012         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4013         inode_inc_iversion(inode);
4014         inode_inc_iversion(dir);
4015         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4016         ret = btrfs_update_inode(trans, root, dir);
4017 out:
4018         return ret;
4019 }
4020
4021 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4022                        struct btrfs_root *root,
4023                        struct inode *dir, struct inode *inode,
4024                        const char *name, int name_len)
4025 {
4026         int ret;
4027         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4028         if (!ret) {
4029                 drop_nlink(inode);
4030                 ret = btrfs_update_inode(trans, root, inode);
4031         }
4032         return ret;
4033 }
4034
4035 /*
4036  * helper to start transaction for unlink and rmdir.
4037  *
4038  * unlink and rmdir are special in btrfs, they do not always free space, so
4039  * if we cannot make our reservations the normal way try and see if there is
4040  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4041  * allow the unlink to occur.
4042  */
4043 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4044 {
4045         struct btrfs_root *root = BTRFS_I(dir)->root;
4046
4047         /*
4048          * 1 for the possible orphan item
4049          * 1 for the dir item
4050          * 1 for the dir index
4051          * 1 for the inode ref
4052          * 1 for the inode
4053          */
4054         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4055 }
4056
4057 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4058 {
4059         struct btrfs_root *root = BTRFS_I(dir)->root;
4060         struct btrfs_trans_handle *trans;
4061         struct inode *inode = d_inode(dentry);
4062         int ret;
4063
4064         trans = __unlink_start_trans(dir);
4065         if (IS_ERR(trans))
4066                 return PTR_ERR(trans);
4067
4068         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4069
4070         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4071                                  dentry->d_name.name, dentry->d_name.len);
4072         if (ret)
4073                 goto out;
4074
4075         if (inode->i_nlink == 0) {
4076                 ret = btrfs_orphan_add(trans, inode);
4077                 if (ret)
4078                         goto out;
4079         }
4080
4081 out:
4082         btrfs_end_transaction(trans, root);
4083         btrfs_btree_balance_dirty(root);
4084         return ret;
4085 }
4086
4087 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4088                         struct btrfs_root *root,
4089                         struct inode *dir, u64 objectid,
4090                         const char *name, int name_len)
4091 {
4092         struct btrfs_path *path;
4093         struct extent_buffer *leaf;
4094         struct btrfs_dir_item *di;
4095         struct btrfs_key key;
4096         u64 index;
4097         int ret;
4098         u64 dir_ino = btrfs_ino(dir);
4099
4100         path = btrfs_alloc_path();
4101         if (!path)
4102                 return -ENOMEM;
4103
4104         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4105                                    name, name_len, -1);
4106         if (IS_ERR_OR_NULL(di)) {
4107                 if (!di)
4108                         ret = -ENOENT;
4109                 else
4110                         ret = PTR_ERR(di);
4111                 goto out;
4112         }
4113
4114         leaf = path->nodes[0];
4115         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4116         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4117         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4118         if (ret) {
4119                 btrfs_abort_transaction(trans, root, ret);
4120                 goto out;
4121         }
4122         btrfs_release_path(path);
4123
4124         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4125                                  objectid, root->root_key.objectid,
4126                                  dir_ino, &index, name, name_len);
4127         if (ret < 0) {
4128                 if (ret != -ENOENT) {
4129                         btrfs_abort_transaction(trans, root, ret);
4130                         goto out;
4131                 }
4132                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4133                                                  name, name_len);
4134                 if (IS_ERR_OR_NULL(di)) {
4135                         if (!di)
4136                                 ret = -ENOENT;
4137                         else
4138                                 ret = PTR_ERR(di);
4139                         btrfs_abort_transaction(trans, root, ret);
4140                         goto out;
4141                 }
4142
4143                 leaf = path->nodes[0];
4144                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4145                 btrfs_release_path(path);
4146                 index = key.offset;
4147         }
4148         btrfs_release_path(path);
4149
4150         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4151         if (ret) {
4152                 btrfs_abort_transaction(trans, root, ret);
4153                 goto out;
4154         }
4155
4156         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4157         inode_inc_iversion(dir);
4158         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4159         ret = btrfs_update_inode_fallback(trans, root, dir);
4160         if (ret)
4161                 btrfs_abort_transaction(trans, root, ret);
4162 out:
4163         btrfs_free_path(path);
4164         return ret;
4165 }
4166
4167 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4168 {
4169         struct inode *inode = d_inode(dentry);
4170         int err = 0;
4171         struct btrfs_root *root = BTRFS_I(dir)->root;
4172         struct btrfs_trans_handle *trans;
4173
4174         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4175                 return -ENOTEMPTY;
4176         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4177                 return -EPERM;
4178
4179         trans = __unlink_start_trans(dir);
4180         if (IS_ERR(trans))
4181                 return PTR_ERR(trans);
4182
4183         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4184                 err = btrfs_unlink_subvol(trans, root, dir,
4185                                           BTRFS_I(inode)->location.objectid,
4186                                           dentry->d_name.name,
4187                                           dentry->d_name.len);
4188                 goto out;
4189         }
4190
4191         err = btrfs_orphan_add(trans, inode);
4192         if (err)
4193                 goto out;
4194
4195         /* now the directory is empty */
4196         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4197                                  dentry->d_name.name, dentry->d_name.len);
4198         if (!err)
4199                 btrfs_i_size_write(inode, 0);
4200 out:
4201         btrfs_end_transaction(trans, root);
4202         btrfs_btree_balance_dirty(root);
4203
4204         return err;
4205 }
4206
4207 static int truncate_space_check(struct btrfs_trans_handle *trans,
4208                                 struct btrfs_root *root,
4209                                 u64 bytes_deleted)
4210 {
4211         int ret;
4212
4213         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4214         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4215                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4216         if (!ret)
4217                 trans->bytes_reserved += bytes_deleted;
4218         return ret;
4219
4220 }
4221
4222 static int truncate_inline_extent(struct inode *inode,
4223                                   struct btrfs_path *path,
4224                                   struct btrfs_key *found_key,
4225                                   const u64 item_end,
4226                                   const u64 new_size)
4227 {
4228         struct extent_buffer *leaf = path->nodes[0];
4229         int slot = path->slots[0];
4230         struct btrfs_file_extent_item *fi;
4231         u32 size = (u32)(new_size - found_key->offset);
4232         struct btrfs_root *root = BTRFS_I(inode)->root;
4233
4234         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4235
4236         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4237                 loff_t offset = new_size;
4238                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4239
4240                 /*
4241                  * Zero out the remaining of the last page of our inline extent,
4242                  * instead of directly truncating our inline extent here - that
4243                  * would be much more complex (decompressing all the data, then
4244                  * compressing the truncated data, which might be bigger than
4245                  * the size of the inline extent, resize the extent, etc).
4246                  * We release the path because to get the page we might need to
4247                  * read the extent item from disk (data not in the page cache).
4248                  */
4249                 btrfs_release_path(path);
4250                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4251         }
4252
4253         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4254         size = btrfs_file_extent_calc_inline_size(size);
4255         btrfs_truncate_item(root, path, size, 1);
4256
4257         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4258                 inode_sub_bytes(inode, item_end + 1 - new_size);
4259
4260         return 0;
4261 }
4262
4263 /*
4264  * this can truncate away extent items, csum items and directory items.
4265  * It starts at a high offset and removes keys until it can't find
4266  * any higher than new_size
4267  *
4268  * csum items that cross the new i_size are truncated to the new size
4269  * as well.
4270  *
4271  * min_type is the minimum key type to truncate down to.  If set to 0, this
4272  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4273  */
4274 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4275                                struct btrfs_root *root,
4276                                struct inode *inode,
4277                                u64 new_size, u32 min_type)
4278 {
4279         struct btrfs_path *path;
4280         struct extent_buffer *leaf;
4281         struct btrfs_file_extent_item *fi;
4282         struct btrfs_key key;
4283         struct btrfs_key found_key;
4284         u64 extent_start = 0;
4285         u64 extent_num_bytes = 0;
4286         u64 extent_offset = 0;
4287         u64 item_end = 0;
4288         u64 last_size = new_size;
4289         u32 found_type = (u8)-1;
4290         int found_extent;
4291         int del_item;
4292         int pending_del_nr = 0;
4293         int pending_del_slot = 0;
4294         int extent_type = -1;
4295         int ret;
4296         int err = 0;
4297         u64 ino = btrfs_ino(inode);
4298         u64 bytes_deleted = 0;
4299         bool be_nice = 0;
4300         bool should_throttle = 0;
4301         bool should_end = 0;
4302
4303         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4304
4305         /*
4306          * for non-free space inodes and ref cows, we want to back off from
4307          * time to time
4308          */
4309         if (!btrfs_is_free_space_inode(inode) &&
4310             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4311                 be_nice = 1;
4312
4313         path = btrfs_alloc_path();
4314         if (!path)
4315                 return -ENOMEM;
4316         path->reada = READA_BACK;
4317
4318         /*
4319          * We want to drop from the next block forward in case this new size is
4320          * not block aligned since we will be keeping the last block of the
4321          * extent just the way it is.
4322          */
4323         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4324             root == root->fs_info->tree_root)
4325                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4326                                         root->sectorsize), (u64)-1, 0);
4327
4328         /*
4329          * This function is also used to drop the items in the log tree before
4330          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4331          * it is used to drop the loged items. So we shouldn't kill the delayed
4332          * items.
4333          */
4334         if (min_type == 0 && root == BTRFS_I(inode)->root)
4335                 btrfs_kill_delayed_inode_items(inode);
4336
4337         key.objectid = ino;
4338         key.offset = (u64)-1;
4339         key.type = (u8)-1;
4340
4341 search_again:
4342         /*
4343          * with a 16K leaf size and 128MB extents, you can actually queue
4344          * up a huge file in a single leaf.  Most of the time that
4345          * bytes_deleted is > 0, it will be huge by the time we get here
4346          */
4347         if (be_nice && bytes_deleted > SZ_32M) {
4348                 if (btrfs_should_end_transaction(trans, root)) {
4349                         err = -EAGAIN;
4350                         goto error;
4351                 }
4352         }
4353
4354
4355         path->leave_spinning = 1;
4356         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4357         if (ret < 0) {
4358                 err = ret;
4359                 goto out;
4360         }
4361
4362         if (ret > 0) {
4363                 /* there are no items in the tree for us to truncate, we're
4364                  * done
4365                  */
4366                 if (path->slots[0] == 0)
4367                         goto out;
4368                 path->slots[0]--;
4369         }
4370
4371         while (1) {
4372                 fi = NULL;
4373                 leaf = path->nodes[0];
4374                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4375                 found_type = found_key.type;
4376
4377                 if (found_key.objectid != ino)
4378                         break;
4379
4380                 if (found_type < min_type)
4381                         break;
4382
4383                 item_end = found_key.offset;
4384                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4385                         fi = btrfs_item_ptr(leaf, path->slots[0],
4386                                             struct btrfs_file_extent_item);
4387                         extent_type = btrfs_file_extent_type(leaf, fi);
4388                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4389                                 item_end +=
4390                                     btrfs_file_extent_num_bytes(leaf, fi);
4391                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4392                                 item_end += btrfs_file_extent_inline_len(leaf,
4393                                                          path->slots[0], fi);
4394                         }
4395                         item_end--;
4396                 }
4397                 if (found_type > min_type) {
4398                         del_item = 1;
4399                 } else {
4400                         if (item_end < new_size)
4401                                 break;
4402                         if (found_key.offset >= new_size)
4403                                 del_item = 1;
4404                         else
4405                                 del_item = 0;
4406                 }
4407                 found_extent = 0;
4408                 /* FIXME, shrink the extent if the ref count is only 1 */
4409                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4410                         goto delete;
4411
4412                 if (del_item)
4413                         last_size = found_key.offset;
4414                 else
4415                         last_size = new_size;
4416
4417                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4418                         u64 num_dec;
4419                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4420                         if (!del_item) {
4421                                 u64 orig_num_bytes =
4422                                         btrfs_file_extent_num_bytes(leaf, fi);
4423                                 extent_num_bytes = ALIGN(new_size -
4424                                                 found_key.offset,
4425                                                 root->sectorsize);
4426                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4427                                                          extent_num_bytes);
4428                                 num_dec = (orig_num_bytes -
4429                                            extent_num_bytes);
4430                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4431                                              &root->state) &&
4432                                     extent_start != 0)
4433                                         inode_sub_bytes(inode, num_dec);
4434                                 btrfs_mark_buffer_dirty(leaf);
4435                         } else {
4436                                 extent_num_bytes =
4437                                         btrfs_file_extent_disk_num_bytes(leaf,
4438                                                                          fi);
4439                                 extent_offset = found_key.offset -
4440                                         btrfs_file_extent_offset(leaf, fi);
4441
4442                                 /* FIXME blocksize != 4096 */
4443                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4444                                 if (extent_start != 0) {
4445                                         found_extent = 1;
4446                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4447                                                      &root->state))
4448                                                 inode_sub_bytes(inode, num_dec);
4449                                 }
4450                         }
4451                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4452                         /*
4453                          * we can't truncate inline items that have had
4454                          * special encodings
4455                          */
4456                         if (!del_item &&
4457                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4458                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4459
4460                                 /*
4461                                  * Need to release path in order to truncate a
4462                                  * compressed extent. So delete any accumulated
4463                                  * extent items so far.
4464                                  */
4465                                 if (btrfs_file_extent_compression(leaf, fi) !=
4466                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4467                                         err = btrfs_del_items(trans, root, path,
4468                                                               pending_del_slot,
4469                                                               pending_del_nr);
4470                                         if (err) {
4471                                                 btrfs_abort_transaction(trans,
4472                                                                         root,
4473                                                                         err);
4474                                                 goto error;
4475                                         }
4476                                         pending_del_nr = 0;
4477                                 }
4478
4479                                 err = truncate_inline_extent(inode, path,
4480                                                              &found_key,
4481                                                              item_end,
4482                                                              new_size);
4483                                 if (err) {
4484                                         btrfs_abort_transaction(trans,
4485                                                                 root, err);
4486                                         goto error;
4487                                 }
4488                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4489                                             &root->state)) {
4490                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4491                         }
4492                 }
4493 delete:
4494                 if (del_item) {
4495                         if (!pending_del_nr) {
4496                                 /* no pending yet, add ourselves */
4497                                 pending_del_slot = path->slots[0];
4498                                 pending_del_nr = 1;
4499                         } else if (pending_del_nr &&
4500                                    path->slots[0] + 1 == pending_del_slot) {
4501                                 /* hop on the pending chunk */
4502                                 pending_del_nr++;
4503                                 pending_del_slot = path->slots[0];
4504                         } else {
4505                                 BUG();
4506                         }
4507                 } else {
4508                         break;
4509                 }
4510                 should_throttle = 0;
4511
4512                 if (found_extent &&
4513                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4514                      root == root->fs_info->tree_root)) {
4515                         btrfs_set_path_blocking(path);
4516                         bytes_deleted += extent_num_bytes;
4517                         ret = btrfs_free_extent(trans, root, extent_start,
4518                                                 extent_num_bytes, 0,
4519                                                 btrfs_header_owner(leaf),
4520                                                 ino, extent_offset);
4521                         BUG_ON(ret);
4522                         if (btrfs_should_throttle_delayed_refs(trans, root))
4523                                 btrfs_async_run_delayed_refs(root,
4524                                         trans->delayed_ref_updates * 2, 0);
4525                         if (be_nice) {
4526                                 if (truncate_space_check(trans, root,
4527                                                          extent_num_bytes)) {
4528                                         should_end = 1;
4529                                 }
4530                                 if (btrfs_should_throttle_delayed_refs(trans,
4531                                                                        root)) {
4532                                         should_throttle = 1;
4533                                 }
4534                         }
4535                 }
4536
4537                 if (found_type == BTRFS_INODE_ITEM_KEY)
4538                         break;
4539
4540                 if (path->slots[0] == 0 ||
4541                     path->slots[0] != pending_del_slot ||
4542                     should_throttle || should_end) {
4543                         if (pending_del_nr) {
4544                                 ret = btrfs_del_items(trans, root, path,
4545                                                 pending_del_slot,
4546                                                 pending_del_nr);
4547                                 if (ret) {
4548                                         btrfs_abort_transaction(trans,
4549                                                                 root, ret);
4550                                         goto error;
4551                                 }
4552                                 pending_del_nr = 0;
4553                         }
4554                         btrfs_release_path(path);
4555                         if (should_throttle) {
4556                                 unsigned long updates = trans->delayed_ref_updates;
4557                                 if (updates) {
4558                                         trans->delayed_ref_updates = 0;
4559                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4560                                         if (ret && !err)
4561                                                 err = ret;
4562                                 }
4563                         }
4564                         /*
4565                          * if we failed to refill our space rsv, bail out
4566                          * and let the transaction restart
4567                          */
4568                         if (should_end) {
4569                                 err = -EAGAIN;
4570                                 goto error;
4571                         }
4572                         goto search_again;
4573                 } else {
4574                         path->slots[0]--;
4575                 }
4576         }
4577 out:
4578         if (pending_del_nr) {
4579                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4580                                       pending_del_nr);
4581                 if (ret)
4582                         btrfs_abort_transaction(trans, root, ret);
4583         }
4584 error:
4585         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4586                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4587
4588         btrfs_free_path(path);
4589
4590         if (be_nice && bytes_deleted > SZ_32M) {
4591                 unsigned long updates = trans->delayed_ref_updates;
4592                 if (updates) {
4593                         trans->delayed_ref_updates = 0;
4594                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4595                         if (ret && !err)
4596                                 err = ret;
4597                 }
4598         }
4599         return err;
4600 }
4601
4602 /*
4603  * btrfs_truncate_page - read, zero a chunk and write a page
4604  * @inode - inode that we're zeroing
4605  * @from - the offset to start zeroing
4606  * @len - the length to zero, 0 to zero the entire range respective to the
4607  *      offset
4608  * @front - zero up to the offset instead of from the offset on
4609  *
4610  * This will find the page for the "from" offset and cow the page and zero the
4611  * part we want to zero.  This is used with truncate and hole punching.
4612  */
4613 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4614                         int front)
4615 {
4616         struct address_space *mapping = inode->i_mapping;
4617         struct btrfs_root *root = BTRFS_I(inode)->root;
4618         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4619         struct btrfs_ordered_extent *ordered;
4620         struct extent_state *cached_state = NULL;
4621         char *kaddr;
4622         u32 blocksize = root->sectorsize;
4623         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4624         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4625         struct page *page;
4626         gfp_t mask = btrfs_alloc_write_mask(mapping);
4627         int ret = 0;
4628         u64 page_start;
4629         u64 page_end;
4630
4631         if ((offset & (blocksize - 1)) == 0 &&
4632             (!len || ((len & (blocksize - 1)) == 0)))
4633                 goto out;
4634         ret = btrfs_delalloc_reserve_space(inode,
4635                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4636         if (ret)
4637                 goto out;
4638
4639 again:
4640         page = find_or_create_page(mapping, index, mask);
4641         if (!page) {
4642                 btrfs_delalloc_release_space(inode,
4643                                 round_down(from, PAGE_CACHE_SIZE),
4644                                 PAGE_CACHE_SIZE);
4645                 ret = -ENOMEM;
4646                 goto out;
4647         }
4648
4649         page_start = page_offset(page);
4650         page_end = page_start + PAGE_CACHE_SIZE - 1;
4651
4652         if (!PageUptodate(page)) {
4653                 ret = btrfs_readpage(NULL, page);
4654                 lock_page(page);
4655                 if (page->mapping != mapping) {
4656                         unlock_page(page);
4657                         page_cache_release(page);
4658                         goto again;
4659                 }
4660                 if (!PageUptodate(page)) {
4661                         ret = -EIO;
4662                         goto out_unlock;
4663                 }
4664         }
4665         wait_on_page_writeback(page);
4666
4667         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
4668         set_page_extent_mapped(page);
4669
4670         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4671         if (ordered) {
4672                 unlock_extent_cached(io_tree, page_start, page_end,
4673                                      &cached_state, GFP_NOFS);
4674                 unlock_page(page);
4675                 page_cache_release(page);
4676                 btrfs_start_ordered_extent(inode, ordered, 1);
4677                 btrfs_put_ordered_extent(ordered);
4678                 goto again;
4679         }
4680
4681         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4682                           EXTENT_DIRTY | EXTENT_DELALLOC |
4683                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4684                           0, 0, &cached_state, GFP_NOFS);
4685
4686         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4687                                         &cached_state);
4688         if (ret) {
4689                 unlock_extent_cached(io_tree, page_start, page_end,
4690                                      &cached_state, GFP_NOFS);
4691                 goto out_unlock;
4692         }
4693
4694         if (offset != PAGE_CACHE_SIZE) {
4695                 if (!len)
4696                         len = PAGE_CACHE_SIZE - offset;
4697                 kaddr = kmap(page);
4698                 if (front)
4699                         memset(kaddr, 0, offset);
4700                 else
4701                         memset(kaddr + offset, 0, len);
4702                 flush_dcache_page(page);
4703                 kunmap(page);
4704         }
4705         ClearPageChecked(page);
4706         set_page_dirty(page);
4707         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4708                              GFP_NOFS);
4709
4710 out_unlock:
4711         if (ret)
4712                 btrfs_delalloc_release_space(inode, page_start,
4713                                              PAGE_CACHE_SIZE);
4714         unlock_page(page);
4715         page_cache_release(page);
4716 out:
4717         return ret;
4718 }
4719
4720 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4721                              u64 offset, u64 len)
4722 {
4723         struct btrfs_trans_handle *trans;
4724         int ret;
4725
4726         /*
4727          * Still need to make sure the inode looks like it's been updated so
4728          * that any holes get logged if we fsync.
4729          */
4730         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4731                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4732                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4733                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4734                 return 0;
4735         }
4736
4737         /*
4738          * 1 - for the one we're dropping
4739          * 1 - for the one we're adding
4740          * 1 - for updating the inode.
4741          */
4742         trans = btrfs_start_transaction(root, 3);
4743         if (IS_ERR(trans))
4744                 return PTR_ERR(trans);
4745
4746         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4747         if (ret) {
4748                 btrfs_abort_transaction(trans, root, ret);
4749                 btrfs_end_transaction(trans, root);
4750                 return ret;
4751         }
4752
4753         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4754                                        0, 0, len, 0, len, 0, 0, 0);
4755         if (ret)
4756                 btrfs_abort_transaction(trans, root, ret);
4757         else
4758                 btrfs_update_inode(trans, root, inode);
4759         btrfs_end_transaction(trans, root);
4760         return ret;
4761 }
4762
4763 /*
4764  * This function puts in dummy file extents for the area we're creating a hole
4765  * for.  So if we are truncating this file to a larger size we need to insert
4766  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4767  * the range between oldsize and size
4768  */
4769 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4770 {
4771         struct btrfs_root *root = BTRFS_I(inode)->root;
4772         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4773         struct extent_map *em = NULL;
4774         struct extent_state *cached_state = NULL;
4775         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4776         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4777         u64 block_end = ALIGN(size, root->sectorsize);
4778         u64 last_byte;
4779         u64 cur_offset;
4780         u64 hole_size;
4781         int err = 0;
4782
4783         /*
4784          * If our size started in the middle of a page we need to zero out the
4785          * rest of the page before we expand the i_size, otherwise we could
4786          * expose stale data.
4787          */
4788         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4789         if (err)
4790                 return err;
4791
4792         if (size <= hole_start)
4793                 return 0;
4794
4795         while (1) {
4796                 struct btrfs_ordered_extent *ordered;
4797
4798                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4799                                  &cached_state);
4800                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4801                                                      block_end - hole_start);
4802                 if (!ordered)
4803                         break;
4804                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4805                                      &cached_state, GFP_NOFS);
4806                 btrfs_start_ordered_extent(inode, ordered, 1);
4807                 btrfs_put_ordered_extent(ordered);
4808         }
4809
4810         cur_offset = hole_start;
4811         while (1) {
4812                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4813                                 block_end - cur_offset, 0);
4814                 if (IS_ERR(em)) {
4815                         err = PTR_ERR(em);
4816                         em = NULL;
4817                         break;
4818                 }
4819                 last_byte = min(extent_map_end(em), block_end);
4820                 last_byte = ALIGN(last_byte , root->sectorsize);
4821                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4822                         struct extent_map *hole_em;
4823                         hole_size = last_byte - cur_offset;
4824
4825                         err = maybe_insert_hole(root, inode, cur_offset,
4826                                                 hole_size);
4827                         if (err)
4828                                 break;
4829                         btrfs_drop_extent_cache(inode, cur_offset,
4830                                                 cur_offset + hole_size - 1, 0);
4831                         hole_em = alloc_extent_map();
4832                         if (!hole_em) {
4833                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4834                                         &BTRFS_I(inode)->runtime_flags);
4835                                 goto next;
4836                         }
4837                         hole_em->start = cur_offset;
4838                         hole_em->len = hole_size;
4839                         hole_em->orig_start = cur_offset;
4840
4841                         hole_em->block_start = EXTENT_MAP_HOLE;
4842                         hole_em->block_len = 0;
4843                         hole_em->orig_block_len = 0;
4844                         hole_em->ram_bytes = hole_size;
4845                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4846                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4847                         hole_em->generation = root->fs_info->generation;
4848
4849                         while (1) {
4850                                 write_lock(&em_tree->lock);
4851                                 err = add_extent_mapping(em_tree, hole_em, 1);
4852                                 write_unlock(&em_tree->lock);
4853                                 if (err != -EEXIST)
4854                                         break;
4855                                 btrfs_drop_extent_cache(inode, cur_offset,
4856                                                         cur_offset +
4857                                                         hole_size - 1, 0);
4858                         }
4859                         free_extent_map(hole_em);
4860                 }
4861 next:
4862                 free_extent_map(em);
4863                 em = NULL;
4864                 cur_offset = last_byte;
4865                 if (cur_offset >= block_end)
4866                         break;
4867         }
4868         free_extent_map(em);
4869         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4870                              GFP_NOFS);
4871         return err;
4872 }
4873
4874 static int wait_snapshoting_atomic_t(atomic_t *a)
4875 {
4876         schedule();
4877         return 0;
4878 }
4879
4880 static void wait_for_snapshot_creation(struct btrfs_root *root)
4881 {
4882         while (true) {
4883                 int ret;
4884
4885                 ret = btrfs_start_write_no_snapshoting(root);
4886                 if (ret)
4887                         break;
4888                 wait_on_atomic_t(&root->will_be_snapshoted,
4889                                  wait_snapshoting_atomic_t,
4890                                  TASK_UNINTERRUPTIBLE);
4891         }
4892 }
4893
4894 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4895 {
4896         struct btrfs_root *root = BTRFS_I(inode)->root;
4897         struct btrfs_trans_handle *trans;
4898         loff_t oldsize = i_size_read(inode);
4899         loff_t newsize = attr->ia_size;
4900         int mask = attr->ia_valid;
4901         int ret;
4902
4903         /*
4904          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4905          * special case where we need to update the times despite not having
4906          * these flags set.  For all other operations the VFS set these flags
4907          * explicitly if it wants a timestamp update.
4908          */
4909         if (newsize != oldsize) {
4910                 inode_inc_iversion(inode);
4911                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4912                         inode->i_ctime = inode->i_mtime =
4913                                 current_fs_time(inode->i_sb);
4914         }
4915
4916         if (newsize > oldsize) {
4917                 truncate_pagecache(inode, newsize);
4918                 /*
4919                  * Don't do an expanding truncate while snapshoting is ongoing.
4920                  * This is to ensure the snapshot captures a fully consistent
4921                  * state of this file - if the snapshot captures this expanding
4922                  * truncation, it must capture all writes that happened before
4923                  * this truncation.
4924                  */
4925                 wait_for_snapshot_creation(root);
4926                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4927                 if (ret) {
4928                         btrfs_end_write_no_snapshoting(root);
4929                         return ret;
4930                 }
4931
4932                 trans = btrfs_start_transaction(root, 1);
4933                 if (IS_ERR(trans)) {
4934                         btrfs_end_write_no_snapshoting(root);
4935                         return PTR_ERR(trans);
4936                 }
4937
4938                 i_size_write(inode, newsize);
4939                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4940                 ret = btrfs_update_inode(trans, root, inode);
4941                 btrfs_end_write_no_snapshoting(root);
4942                 btrfs_end_transaction(trans, root);
4943         } else {
4944
4945                 /*
4946                  * We're truncating a file that used to have good data down to
4947                  * zero. Make sure it gets into the ordered flush list so that
4948                  * any new writes get down to disk quickly.
4949                  */
4950                 if (newsize == 0)
4951                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4952                                 &BTRFS_I(inode)->runtime_flags);
4953
4954                 /*
4955                  * 1 for the orphan item we're going to add
4956                  * 1 for the orphan item deletion.
4957                  */
4958                 trans = btrfs_start_transaction(root, 2);
4959                 if (IS_ERR(trans))
4960                         return PTR_ERR(trans);
4961
4962                 /*
4963                  * We need to do this in case we fail at _any_ point during the
4964                  * actual truncate.  Once we do the truncate_setsize we could
4965                  * invalidate pages which forces any outstanding ordered io to
4966                  * be instantly completed which will give us extents that need
4967                  * to be truncated.  If we fail to get an orphan inode down we
4968                  * could have left over extents that were never meant to live,
4969                  * so we need to garuntee from this point on that everything
4970                  * will be consistent.
4971                  */
4972                 ret = btrfs_orphan_add(trans, inode);
4973                 btrfs_end_transaction(trans, root);
4974                 if (ret)
4975                         return ret;
4976
4977                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4978                 truncate_setsize(inode, newsize);
4979
4980                 /* Disable nonlocked read DIO to avoid the end less truncate */
4981                 btrfs_inode_block_unlocked_dio(inode);
4982                 inode_dio_wait(inode);
4983                 btrfs_inode_resume_unlocked_dio(inode);
4984
4985                 ret = btrfs_truncate(inode);
4986                 if (ret && inode->i_nlink) {
4987                         int err;
4988
4989                         /*
4990                          * failed to truncate, disk_i_size is only adjusted down
4991                          * as we remove extents, so it should represent the true
4992                          * size of the inode, so reset the in memory size and
4993                          * delete our orphan entry.
4994                          */
4995                         trans = btrfs_join_transaction(root);
4996                         if (IS_ERR(trans)) {
4997                                 btrfs_orphan_del(NULL, inode);
4998                                 return ret;
4999                         }
5000                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5001                         err = btrfs_orphan_del(trans, inode);
5002                         if (err)
5003                                 btrfs_abort_transaction(trans, root, err);
5004                         btrfs_end_transaction(trans, root);
5005                 }
5006         }
5007
5008         return ret;
5009 }
5010
5011 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5012 {
5013         struct inode *inode = d_inode(dentry);
5014         struct btrfs_root *root = BTRFS_I(inode)->root;
5015         int err;
5016
5017         if (btrfs_root_readonly(root))
5018                 return -EROFS;
5019
5020         err = inode_change_ok(inode, attr);
5021         if (err)
5022                 return err;
5023
5024         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5025                 err = btrfs_setsize(inode, attr);
5026                 if (err)
5027                         return err;
5028         }
5029
5030         if (attr->ia_valid) {
5031                 setattr_copy(inode, attr);
5032                 inode_inc_iversion(inode);
5033                 err = btrfs_dirty_inode(inode);
5034
5035                 if (!err && attr->ia_valid & ATTR_MODE)
5036                         err = posix_acl_chmod(inode, inode->i_mode);
5037         }
5038
5039         return err;
5040 }
5041
5042 /*
5043  * While truncating the inode pages during eviction, we get the VFS calling
5044  * btrfs_invalidatepage() against each page of the inode. This is slow because
5045  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5046  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5047  * extent_state structures over and over, wasting lots of time.
5048  *
5049  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5050  * those expensive operations on a per page basis and do only the ordered io
5051  * finishing, while we release here the extent_map and extent_state structures,
5052  * without the excessive merging and splitting.
5053  */
5054 static void evict_inode_truncate_pages(struct inode *inode)
5055 {
5056         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5057         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5058         struct rb_node *node;
5059
5060         ASSERT(inode->i_state & I_FREEING);
5061         truncate_inode_pages_final(&inode->i_data);
5062
5063         write_lock(&map_tree->lock);
5064         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5065                 struct extent_map *em;
5066
5067                 node = rb_first(&map_tree->map);
5068                 em = rb_entry(node, struct extent_map, rb_node);
5069                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5070                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5071                 remove_extent_mapping(map_tree, em);
5072                 free_extent_map(em);
5073                 if (need_resched()) {
5074                         write_unlock(&map_tree->lock);
5075                         cond_resched();
5076                         write_lock(&map_tree->lock);
5077                 }
5078         }
5079         write_unlock(&map_tree->lock);
5080
5081         /*
5082          * Keep looping until we have no more ranges in the io tree.
5083          * We can have ongoing bios started by readpages (called from readahead)
5084          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5085          * still in progress (unlocked the pages in the bio but did not yet
5086          * unlocked the ranges in the io tree). Therefore this means some
5087          * ranges can still be locked and eviction started because before
5088          * submitting those bios, which are executed by a separate task (work
5089          * queue kthread), inode references (inode->i_count) were not taken
5090          * (which would be dropped in the end io callback of each bio).
5091          * Therefore here we effectively end up waiting for those bios and
5092          * anyone else holding locked ranges without having bumped the inode's
5093          * reference count - if we don't do it, when they access the inode's
5094          * io_tree to unlock a range it may be too late, leading to an
5095          * use-after-free issue.
5096          */
5097         spin_lock(&io_tree->lock);
5098         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5099                 struct extent_state *state;
5100                 struct extent_state *cached_state = NULL;
5101                 u64 start;
5102                 u64 end;
5103
5104                 node = rb_first(&io_tree->state);
5105                 state = rb_entry(node, struct extent_state, rb_node);
5106                 start = state->start;
5107                 end = state->end;
5108                 spin_unlock(&io_tree->lock);
5109
5110                 lock_extent_bits(io_tree, start, end, &cached_state);
5111
5112                 /*
5113                  * If still has DELALLOC flag, the extent didn't reach disk,
5114                  * and its reserved space won't be freed by delayed_ref.
5115                  * So we need to free its reserved space here.
5116                  * (Refer to comment in btrfs_invalidatepage, case 2)
5117                  *
5118                  * Note, end is the bytenr of last byte, so we need + 1 here.
5119                  */
5120                 if (state->state & EXTENT_DELALLOC)
5121                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5122
5123                 clear_extent_bit(io_tree, start, end,
5124                                  EXTENT_LOCKED | EXTENT_DIRTY |
5125                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5126                                  EXTENT_DEFRAG, 1, 1,
5127                                  &cached_state, GFP_NOFS);
5128
5129                 cond_resched();
5130                 spin_lock(&io_tree->lock);
5131         }
5132         spin_unlock(&io_tree->lock);
5133 }
5134
5135 void btrfs_evict_inode(struct inode *inode)
5136 {
5137         struct btrfs_trans_handle *trans;
5138         struct btrfs_root *root = BTRFS_I(inode)->root;
5139         struct btrfs_block_rsv *rsv, *global_rsv;
5140         int steal_from_global = 0;
5141         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5142         int ret;
5143
5144         trace_btrfs_inode_evict(inode);
5145
5146         evict_inode_truncate_pages(inode);
5147
5148         if (inode->i_nlink &&
5149             ((btrfs_root_refs(&root->root_item) != 0 &&
5150               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5151              btrfs_is_free_space_inode(inode)))
5152                 goto no_delete;
5153
5154         if (is_bad_inode(inode)) {
5155                 btrfs_orphan_del(NULL, inode);
5156                 goto no_delete;
5157         }
5158         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5159         if (!special_file(inode->i_mode))
5160                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5161
5162         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5163
5164         if (root->fs_info->log_root_recovering) {
5165                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5166                                  &BTRFS_I(inode)->runtime_flags));
5167                 goto no_delete;
5168         }
5169
5170         if (inode->i_nlink > 0) {
5171                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5172                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5173                 goto no_delete;
5174         }
5175
5176         ret = btrfs_commit_inode_delayed_inode(inode);
5177         if (ret) {
5178                 btrfs_orphan_del(NULL, inode);
5179                 goto no_delete;
5180         }
5181
5182         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5183         if (!rsv) {
5184                 btrfs_orphan_del(NULL, inode);
5185                 goto no_delete;
5186         }
5187         rsv->size = min_size;
5188         rsv->failfast = 1;
5189         global_rsv = &root->fs_info->global_block_rsv;
5190
5191         btrfs_i_size_write(inode, 0);
5192
5193         /*
5194          * This is a bit simpler than btrfs_truncate since we've already
5195          * reserved our space for our orphan item in the unlink, so we just
5196          * need to reserve some slack space in case we add bytes and update
5197          * inode item when doing the truncate.
5198          */
5199         while (1) {
5200                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5201                                              BTRFS_RESERVE_FLUSH_LIMIT);
5202
5203                 /*
5204                  * Try and steal from the global reserve since we will
5205                  * likely not use this space anyway, we want to try as
5206                  * hard as possible to get this to work.
5207                  */
5208                 if (ret)
5209                         steal_from_global++;
5210                 else
5211                         steal_from_global = 0;
5212                 ret = 0;
5213
5214                 /*
5215                  * steal_from_global == 0: we reserved stuff, hooray!
5216                  * steal_from_global == 1: we didn't reserve stuff, boo!
5217                  * steal_from_global == 2: we've committed, still not a lot of
5218                  * room but maybe we'll have room in the global reserve this
5219                  * time.
5220                  * steal_from_global == 3: abandon all hope!
5221                  */
5222                 if (steal_from_global > 2) {
5223                         btrfs_warn(root->fs_info,
5224                                 "Could not get space for a delete, will truncate on mount %d",
5225                                 ret);
5226                         btrfs_orphan_del(NULL, inode);
5227                         btrfs_free_block_rsv(root, rsv);
5228                         goto no_delete;
5229                 }
5230
5231                 trans = btrfs_join_transaction(root);
5232                 if (IS_ERR(trans)) {
5233                         btrfs_orphan_del(NULL, inode);
5234                         btrfs_free_block_rsv(root, rsv);
5235                         goto no_delete;
5236                 }
5237
5238                 /*
5239                  * We can't just steal from the global reserve, we need tomake
5240                  * sure there is room to do it, if not we need to commit and try
5241                  * again.
5242                  */
5243                 if (steal_from_global) {
5244                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5245                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5246                                                               min_size);
5247                         else
5248                                 ret = -ENOSPC;
5249                 }
5250
5251                 /*
5252                  * Couldn't steal from the global reserve, we have too much
5253                  * pending stuff built up, commit the transaction and try it
5254                  * again.
5255                  */
5256                 if (ret) {
5257                         ret = btrfs_commit_transaction(trans, root);
5258                         if (ret) {
5259                                 btrfs_orphan_del(NULL, inode);
5260                                 btrfs_free_block_rsv(root, rsv);
5261                                 goto no_delete;
5262                         }
5263                         continue;
5264                 } else {
5265                         steal_from_global = 0;
5266                 }
5267
5268                 trans->block_rsv = rsv;
5269
5270                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5271                 if (ret != -ENOSPC && ret != -EAGAIN)
5272                         break;
5273
5274                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5275                 btrfs_end_transaction(trans, root);
5276                 trans = NULL;
5277                 btrfs_btree_balance_dirty(root);
5278         }
5279
5280         btrfs_free_block_rsv(root, rsv);
5281
5282         /*
5283          * Errors here aren't a big deal, it just means we leave orphan items
5284          * in the tree.  They will be cleaned up on the next mount.
5285          */
5286         if (ret == 0) {
5287                 trans->block_rsv = root->orphan_block_rsv;
5288                 btrfs_orphan_del(trans, inode);
5289         } else {
5290                 btrfs_orphan_del(NULL, inode);
5291         }
5292
5293         trans->block_rsv = &root->fs_info->trans_block_rsv;
5294         if (!(root == root->fs_info->tree_root ||
5295               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5296                 btrfs_return_ino(root, btrfs_ino(inode));
5297
5298         btrfs_end_transaction(trans, root);
5299         btrfs_btree_balance_dirty(root);
5300 no_delete:
5301         btrfs_remove_delayed_node(inode);
5302         clear_inode(inode);
5303 }
5304
5305 /*
5306  * this returns the key found in the dir entry in the location pointer.
5307  * If no dir entries were found, location->objectid is 0.
5308  */
5309 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5310                                struct btrfs_key *location)
5311 {
5312         const char *name = dentry->d_name.name;
5313         int namelen = dentry->d_name.len;
5314         struct btrfs_dir_item *di;
5315         struct btrfs_path *path;
5316         struct btrfs_root *root = BTRFS_I(dir)->root;
5317         int ret = 0;
5318
5319         path = btrfs_alloc_path();
5320         if (!path)
5321                 return -ENOMEM;
5322
5323         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5324                                     namelen, 0);
5325         if (IS_ERR(di))
5326                 ret = PTR_ERR(di);
5327
5328         if (IS_ERR_OR_NULL(di))
5329                 goto out_err;
5330
5331         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5332 out:
5333         btrfs_free_path(path);
5334         return ret;
5335 out_err:
5336         location->objectid = 0;
5337         goto out;
5338 }
5339
5340 /*
5341  * when we hit a tree root in a directory, the btrfs part of the inode
5342  * needs to be changed to reflect the root directory of the tree root.  This
5343  * is kind of like crossing a mount point.
5344  */
5345 static int fixup_tree_root_location(struct btrfs_root *root,
5346                                     struct inode *dir,
5347                                     struct dentry *dentry,
5348                                     struct btrfs_key *location,
5349                                     struct btrfs_root **sub_root)
5350 {
5351         struct btrfs_path *path;
5352         struct btrfs_root *new_root;
5353         struct btrfs_root_ref *ref;
5354         struct extent_buffer *leaf;
5355         struct btrfs_key key;
5356         int ret;
5357         int err = 0;
5358
5359         path = btrfs_alloc_path();
5360         if (!path) {
5361                 err = -ENOMEM;
5362                 goto out;
5363         }
5364
5365         err = -ENOENT;
5366         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5367         key.type = BTRFS_ROOT_REF_KEY;
5368         key.offset = location->objectid;
5369
5370         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5371                                 0, 0);
5372         if (ret) {
5373                 if (ret < 0)
5374                         err = ret;
5375                 goto out;
5376         }
5377
5378         leaf = path->nodes[0];
5379         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5380         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5381             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5382                 goto out;
5383
5384         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5385                                    (unsigned long)(ref + 1),
5386                                    dentry->d_name.len);
5387         if (ret)
5388                 goto out;
5389
5390         btrfs_release_path(path);
5391
5392         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5393         if (IS_ERR(new_root)) {
5394                 err = PTR_ERR(new_root);
5395                 goto out;
5396         }
5397
5398         *sub_root = new_root;
5399         location->objectid = btrfs_root_dirid(&new_root->root_item);
5400         location->type = BTRFS_INODE_ITEM_KEY;
5401         location->offset = 0;
5402         err = 0;
5403 out:
5404         btrfs_free_path(path);
5405         return err;
5406 }
5407
5408 static void inode_tree_add(struct inode *inode)
5409 {
5410         struct btrfs_root *root = BTRFS_I(inode)->root;
5411         struct btrfs_inode *entry;
5412         struct rb_node **p;
5413         struct rb_node *parent;
5414         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5415         u64 ino = btrfs_ino(inode);
5416
5417         if (inode_unhashed(inode))
5418                 return;
5419         parent = NULL;
5420         spin_lock(&root->inode_lock);
5421         p = &root->inode_tree.rb_node;
5422         while (*p) {
5423                 parent = *p;
5424                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5425
5426                 if (ino < btrfs_ino(&entry->vfs_inode))
5427                         p = &parent->rb_left;
5428                 else if (ino > btrfs_ino(&entry->vfs_inode))
5429                         p = &parent->rb_right;
5430                 else {
5431                         WARN_ON(!(entry->vfs_inode.i_state &
5432                                   (I_WILL_FREE | I_FREEING)));
5433                         rb_replace_node(parent, new, &root->inode_tree);
5434                         RB_CLEAR_NODE(parent);
5435                         spin_unlock(&root->inode_lock);
5436                         return;
5437                 }
5438         }
5439         rb_link_node(new, parent, p);
5440         rb_insert_color(new, &root->inode_tree);
5441         spin_unlock(&root->inode_lock);
5442 }
5443
5444 static void inode_tree_del(struct inode *inode)
5445 {
5446         struct btrfs_root *root = BTRFS_I(inode)->root;
5447         int empty = 0;
5448
5449         spin_lock(&root->inode_lock);
5450         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5451                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5452                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5453                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5454         }
5455         spin_unlock(&root->inode_lock);
5456
5457         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5458                 synchronize_srcu(&root->fs_info->subvol_srcu);
5459                 spin_lock(&root->inode_lock);
5460                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5461                 spin_unlock(&root->inode_lock);
5462                 if (empty)
5463                         btrfs_add_dead_root(root);
5464         }
5465 }
5466
5467 void btrfs_invalidate_inodes(struct btrfs_root *root)
5468 {
5469         struct rb_node *node;
5470         struct rb_node *prev;
5471         struct btrfs_inode *entry;
5472         struct inode *inode;
5473         u64 objectid = 0;
5474
5475         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5476                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5477
5478         spin_lock(&root->inode_lock);
5479 again:
5480         node = root->inode_tree.rb_node;
5481         prev = NULL;
5482         while (node) {
5483                 prev = node;
5484                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5485
5486                 if (objectid < btrfs_ino(&entry->vfs_inode))
5487                         node = node->rb_left;
5488                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5489                         node = node->rb_right;
5490                 else
5491                         break;
5492         }
5493         if (!node) {
5494                 while (prev) {
5495                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5496                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5497                                 node = prev;
5498                                 break;
5499                         }
5500                         prev = rb_next(prev);
5501                 }
5502         }
5503         while (node) {
5504                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5505                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5506                 inode = igrab(&entry->vfs_inode);
5507                 if (inode) {
5508                         spin_unlock(&root->inode_lock);
5509                         if (atomic_read(&inode->i_count) > 1)
5510                                 d_prune_aliases(inode);
5511                         /*
5512                          * btrfs_drop_inode will have it removed from
5513                          * the inode cache when its usage count
5514                          * hits zero.
5515                          */
5516                         iput(inode);
5517                         cond_resched();
5518                         spin_lock(&root->inode_lock);
5519                         goto again;
5520                 }
5521
5522                 if (cond_resched_lock(&root->inode_lock))
5523                         goto again;
5524
5525                 node = rb_next(node);
5526         }
5527         spin_unlock(&root->inode_lock);
5528 }
5529
5530 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5531 {
5532         struct btrfs_iget_args *args = p;
5533         inode->i_ino = args->location->objectid;
5534         memcpy(&BTRFS_I(inode)->location, args->location,
5535                sizeof(*args->location));
5536         BTRFS_I(inode)->root = args->root;
5537         return 0;
5538 }
5539
5540 static int btrfs_find_actor(struct inode *inode, void *opaque)
5541 {
5542         struct btrfs_iget_args *args = opaque;
5543         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5544                 args->root == BTRFS_I(inode)->root;
5545 }
5546
5547 static struct inode *btrfs_iget_locked(struct super_block *s,
5548                                        struct btrfs_key *location,
5549                                        struct btrfs_root *root)
5550 {
5551         struct inode *inode;
5552         struct btrfs_iget_args args;
5553         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5554
5555         args.location = location;
5556         args.root = root;
5557
5558         inode = iget5_locked(s, hashval, btrfs_find_actor,
5559                              btrfs_init_locked_inode,
5560                              (void *)&args);
5561         return inode;
5562 }
5563
5564 /* Get an inode object given its location and corresponding root.
5565  * Returns in *is_new if the inode was read from disk
5566  */
5567 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5568                          struct btrfs_root *root, int *new)
5569 {
5570         struct inode *inode;
5571
5572         inode = btrfs_iget_locked(s, location, root);
5573         if (!inode)
5574                 return ERR_PTR(-ENOMEM);
5575
5576         if (inode->i_state & I_NEW) {
5577                 btrfs_read_locked_inode(inode);
5578                 if (!is_bad_inode(inode)) {
5579                         inode_tree_add(inode);
5580                         unlock_new_inode(inode);
5581                         if (new)
5582                                 *new = 1;
5583                 } else {
5584                         unlock_new_inode(inode);
5585                         iput(inode);
5586                         inode = ERR_PTR(-ESTALE);
5587                 }
5588         }
5589
5590         return inode;
5591 }
5592
5593 static struct inode *new_simple_dir(struct super_block *s,
5594                                     struct btrfs_key *key,
5595                                     struct btrfs_root *root)
5596 {
5597         struct inode *inode = new_inode(s);
5598
5599         if (!inode)
5600                 return ERR_PTR(-ENOMEM);
5601
5602         BTRFS_I(inode)->root = root;
5603         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5604         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5605
5606         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5607         inode->i_op = &btrfs_dir_ro_inode_operations;
5608         inode->i_fop = &simple_dir_operations;
5609         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5610         inode->i_mtime = CURRENT_TIME;
5611         inode->i_atime = inode->i_mtime;
5612         inode->i_ctime = inode->i_mtime;
5613         BTRFS_I(inode)->i_otime = inode->i_mtime;
5614
5615         return inode;
5616 }
5617
5618 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5619 {
5620         struct inode *inode;
5621         struct btrfs_root *root = BTRFS_I(dir)->root;
5622         struct btrfs_root *sub_root = root;
5623         struct btrfs_key location;
5624         int index;
5625         int ret = 0;
5626
5627         if (dentry->d_name.len > BTRFS_NAME_LEN)
5628                 return ERR_PTR(-ENAMETOOLONG);
5629
5630         ret = btrfs_inode_by_name(dir, dentry, &location);
5631         if (ret < 0)
5632                 return ERR_PTR(ret);
5633
5634         if (location.objectid == 0)
5635                 return ERR_PTR(-ENOENT);
5636
5637         if (location.type == BTRFS_INODE_ITEM_KEY) {
5638                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5639                 return inode;
5640         }
5641
5642         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5643
5644         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5645         ret = fixup_tree_root_location(root, dir, dentry,
5646                                        &location, &sub_root);
5647         if (ret < 0) {
5648                 if (ret != -ENOENT)
5649                         inode = ERR_PTR(ret);
5650                 else
5651                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5652         } else {
5653                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5654         }
5655         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5656
5657         if (!IS_ERR(inode) && root != sub_root) {
5658                 down_read(&root->fs_info->cleanup_work_sem);
5659                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5660                         ret = btrfs_orphan_cleanup(sub_root);
5661                 up_read(&root->fs_info->cleanup_work_sem);
5662                 if (ret) {
5663                         iput(inode);
5664                         inode = ERR_PTR(ret);
5665                 }
5666         }
5667
5668         return inode;
5669 }
5670
5671 static int btrfs_dentry_delete(const struct dentry *dentry)
5672 {
5673         struct btrfs_root *root;
5674         struct inode *inode = d_inode(dentry);
5675
5676         if (!inode && !IS_ROOT(dentry))
5677                 inode = d_inode(dentry->d_parent);
5678
5679         if (inode) {
5680                 root = BTRFS_I(inode)->root;
5681                 if (btrfs_root_refs(&root->root_item) == 0)
5682                         return 1;
5683
5684                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5685                         return 1;
5686         }
5687         return 0;
5688 }
5689
5690 static void btrfs_dentry_release(struct dentry *dentry)
5691 {
5692         kfree(dentry->d_fsdata);
5693 }
5694
5695 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5696                                    unsigned int flags)
5697 {
5698         struct inode *inode;
5699
5700         inode = btrfs_lookup_dentry(dir, dentry);
5701         if (IS_ERR(inode)) {
5702                 if (PTR_ERR(inode) == -ENOENT)
5703                         inode = NULL;
5704                 else
5705                         return ERR_CAST(inode);
5706         }
5707
5708         return d_splice_alias(inode, dentry);
5709 }
5710
5711 unsigned char btrfs_filetype_table[] = {
5712         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5713 };
5714
5715 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5716 {
5717         struct inode *inode = file_inode(file);
5718         struct btrfs_root *root = BTRFS_I(inode)->root;
5719         struct btrfs_item *item;
5720         struct btrfs_dir_item *di;
5721         struct btrfs_key key;
5722         struct btrfs_key found_key;
5723         struct btrfs_path *path;
5724         struct list_head ins_list;
5725         struct list_head del_list;
5726         int ret;
5727         struct extent_buffer *leaf;
5728         int slot;
5729         unsigned char d_type;
5730         int over = 0;
5731         u32 di_cur;
5732         u32 di_total;
5733         u32 di_len;
5734         int key_type = BTRFS_DIR_INDEX_KEY;
5735         char tmp_name[32];
5736         char *name_ptr;
5737         int name_len;
5738         int is_curr = 0;        /* ctx->pos points to the current index? */
5739
5740         /* FIXME, use a real flag for deciding about the key type */
5741         if (root->fs_info->tree_root == root)
5742                 key_type = BTRFS_DIR_ITEM_KEY;
5743
5744         if (!dir_emit_dots(file, ctx))
5745                 return 0;
5746
5747         path = btrfs_alloc_path();
5748         if (!path)
5749                 return -ENOMEM;
5750
5751         path->reada = READA_FORWARD;
5752
5753         if (key_type == BTRFS_DIR_INDEX_KEY) {
5754                 INIT_LIST_HEAD(&ins_list);
5755                 INIT_LIST_HEAD(&del_list);
5756                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5757         }
5758
5759         key.type = key_type;
5760         key.offset = ctx->pos;
5761         key.objectid = btrfs_ino(inode);
5762
5763         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5764         if (ret < 0)
5765                 goto err;
5766
5767         while (1) {
5768                 leaf = path->nodes[0];
5769                 slot = path->slots[0];
5770                 if (slot >= btrfs_header_nritems(leaf)) {
5771                         ret = btrfs_next_leaf(root, path);
5772                         if (ret < 0)
5773                                 goto err;
5774                         else if (ret > 0)
5775                                 break;
5776                         continue;
5777                 }
5778
5779                 item = btrfs_item_nr(slot);
5780                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5781
5782                 if (found_key.objectid != key.objectid)
5783                         break;
5784                 if (found_key.type != key_type)
5785                         break;
5786                 if (found_key.offset < ctx->pos)
5787                         goto next;
5788                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5789                     btrfs_should_delete_dir_index(&del_list,
5790                                                   found_key.offset))
5791                         goto next;
5792
5793                 ctx->pos = found_key.offset;
5794                 is_curr = 1;
5795
5796                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5797                 di_cur = 0;
5798                 di_total = btrfs_item_size(leaf, item);
5799
5800                 while (di_cur < di_total) {
5801                         struct btrfs_key location;
5802
5803                         if (verify_dir_item(root, leaf, di))
5804                                 break;
5805
5806                         name_len = btrfs_dir_name_len(leaf, di);
5807                         if (name_len <= sizeof(tmp_name)) {
5808                                 name_ptr = tmp_name;
5809                         } else {
5810                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5811                                 if (!name_ptr) {
5812                                         ret = -ENOMEM;
5813                                         goto err;
5814                                 }
5815                         }
5816                         read_extent_buffer(leaf, name_ptr,
5817                                            (unsigned long)(di + 1), name_len);
5818
5819                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5820                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5821
5822
5823                         /* is this a reference to our own snapshot? If so
5824                          * skip it.
5825                          *
5826                          * In contrast to old kernels, we insert the snapshot's
5827                          * dir item and dir index after it has been created, so
5828                          * we won't find a reference to our own snapshot. We
5829                          * still keep the following code for backward
5830                          * compatibility.
5831                          */
5832                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5833                             location.objectid == root->root_key.objectid) {
5834                                 over = 0;
5835                                 goto skip;
5836                         }
5837                         over = !dir_emit(ctx, name_ptr, name_len,
5838                                        location.objectid, d_type);
5839
5840 skip:
5841                         if (name_ptr != tmp_name)
5842                                 kfree(name_ptr);
5843
5844                         if (over)
5845                                 goto nopos;
5846                         di_len = btrfs_dir_name_len(leaf, di) +
5847                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5848                         di_cur += di_len;
5849                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5850                 }
5851 next:
5852                 path->slots[0]++;
5853         }
5854
5855         if (key_type == BTRFS_DIR_INDEX_KEY) {
5856                 if (is_curr)
5857                         ctx->pos++;
5858                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5859                 if (ret)
5860                         goto nopos;
5861         }
5862
5863         /* Reached end of directory/root. Bump pos past the last item. */
5864         ctx->pos++;
5865
5866         /*
5867          * Stop new entries from being returned after we return the last
5868          * entry.
5869          *
5870          * New directory entries are assigned a strictly increasing
5871          * offset.  This means that new entries created during readdir
5872          * are *guaranteed* to be seen in the future by that readdir.
5873          * This has broken buggy programs which operate on names as
5874          * they're returned by readdir.  Until we re-use freed offsets
5875          * we have this hack to stop new entries from being returned
5876          * under the assumption that they'll never reach this huge
5877          * offset.
5878          *
5879          * This is being careful not to overflow 32bit loff_t unless the
5880          * last entry requires it because doing so has broken 32bit apps
5881          * in the past.
5882          */
5883         if (key_type == BTRFS_DIR_INDEX_KEY) {
5884                 if (ctx->pos >= INT_MAX)
5885                         ctx->pos = LLONG_MAX;
5886                 else
5887                         ctx->pos = INT_MAX;
5888         }
5889 nopos:
5890         ret = 0;
5891 err:
5892         if (key_type == BTRFS_DIR_INDEX_KEY)
5893                 btrfs_put_delayed_items(&ins_list, &del_list);
5894         btrfs_free_path(path);
5895         return ret;
5896 }
5897
5898 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5899 {
5900         struct btrfs_root *root = BTRFS_I(inode)->root;
5901         struct btrfs_trans_handle *trans;
5902         int ret = 0;
5903         bool nolock = false;
5904
5905         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5906                 return 0;
5907
5908         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5909                 nolock = true;
5910
5911         if (wbc->sync_mode == WB_SYNC_ALL) {
5912                 if (nolock)
5913                         trans = btrfs_join_transaction_nolock(root);
5914                 else
5915                         trans = btrfs_join_transaction(root);
5916                 if (IS_ERR(trans))
5917                         return PTR_ERR(trans);
5918                 ret = btrfs_commit_transaction(trans, root);
5919         }
5920         return ret;
5921 }
5922
5923 /*
5924  * This is somewhat expensive, updating the tree every time the
5925  * inode changes.  But, it is most likely to find the inode in cache.
5926  * FIXME, needs more benchmarking...there are no reasons other than performance
5927  * to keep or drop this code.
5928  */
5929 static int btrfs_dirty_inode(struct inode *inode)
5930 {
5931         struct btrfs_root *root = BTRFS_I(inode)->root;
5932         struct btrfs_trans_handle *trans;
5933         int ret;
5934
5935         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5936                 return 0;
5937
5938         trans = btrfs_join_transaction(root);
5939         if (IS_ERR(trans))
5940                 return PTR_ERR(trans);
5941
5942         ret = btrfs_update_inode(trans, root, inode);
5943         if (ret && ret == -ENOSPC) {
5944                 /* whoops, lets try again with the full transaction */
5945                 btrfs_end_transaction(trans, root);
5946                 trans = btrfs_start_transaction(root, 1);
5947                 if (IS_ERR(trans))
5948                         return PTR_ERR(trans);
5949
5950                 ret = btrfs_update_inode(trans, root, inode);
5951         }
5952         btrfs_end_transaction(trans, root);
5953         if (BTRFS_I(inode)->delayed_node)
5954                 btrfs_balance_delayed_items(root);
5955
5956         return ret;
5957 }
5958
5959 /*
5960  * This is a copy of file_update_time.  We need this so we can return error on
5961  * ENOSPC for updating the inode in the case of file write and mmap writes.
5962  */
5963 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5964                              int flags)
5965 {
5966         struct btrfs_root *root = BTRFS_I(inode)->root;
5967
5968         if (btrfs_root_readonly(root))
5969                 return -EROFS;
5970
5971         if (flags & S_VERSION)
5972                 inode_inc_iversion(inode);
5973         if (flags & S_CTIME)
5974                 inode->i_ctime = *now;
5975         if (flags & S_MTIME)
5976                 inode->i_mtime = *now;
5977         if (flags & S_ATIME)
5978                 inode->i_atime = *now;
5979         return btrfs_dirty_inode(inode);
5980 }
5981
5982 /*
5983  * find the highest existing sequence number in a directory
5984  * and then set the in-memory index_cnt variable to reflect
5985  * free sequence numbers
5986  */
5987 static int btrfs_set_inode_index_count(struct inode *inode)
5988 {
5989         struct btrfs_root *root = BTRFS_I(inode)->root;
5990         struct btrfs_key key, found_key;
5991         struct btrfs_path *path;
5992         struct extent_buffer *leaf;
5993         int ret;
5994
5995         key.objectid = btrfs_ino(inode);
5996         key.type = BTRFS_DIR_INDEX_KEY;
5997         key.offset = (u64)-1;
5998
5999         path = btrfs_alloc_path();
6000         if (!path)
6001                 return -ENOMEM;
6002
6003         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6004         if (ret < 0)
6005                 goto out;
6006         /* FIXME: we should be able to handle this */
6007         if (ret == 0)
6008                 goto out;
6009         ret = 0;
6010
6011         /*
6012          * MAGIC NUMBER EXPLANATION:
6013          * since we search a directory based on f_pos we have to start at 2
6014          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6015          * else has to start at 2
6016          */
6017         if (path->slots[0] == 0) {
6018                 BTRFS_I(inode)->index_cnt = 2;
6019                 goto out;
6020         }
6021
6022         path->slots[0]--;
6023
6024         leaf = path->nodes[0];
6025         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6026
6027         if (found_key.objectid != btrfs_ino(inode) ||
6028             found_key.type != BTRFS_DIR_INDEX_KEY) {
6029                 BTRFS_I(inode)->index_cnt = 2;
6030                 goto out;
6031         }
6032
6033         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6034 out:
6035         btrfs_free_path(path);
6036         return ret;
6037 }
6038
6039 /*
6040  * helper to find a free sequence number in a given directory.  This current
6041  * code is very simple, later versions will do smarter things in the btree
6042  */
6043 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6044 {
6045         int ret = 0;
6046
6047         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6048                 ret = btrfs_inode_delayed_dir_index_count(dir);
6049                 if (ret) {
6050                         ret = btrfs_set_inode_index_count(dir);
6051                         if (ret)
6052                                 return ret;
6053                 }
6054         }
6055
6056         *index = BTRFS_I(dir)->index_cnt;
6057         BTRFS_I(dir)->index_cnt++;
6058
6059         return ret;
6060 }
6061
6062 static int btrfs_insert_inode_locked(struct inode *inode)
6063 {
6064         struct btrfs_iget_args args;
6065         args.location = &BTRFS_I(inode)->location;
6066         args.root = BTRFS_I(inode)->root;
6067
6068         return insert_inode_locked4(inode,
6069                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6070                    btrfs_find_actor, &args);
6071 }
6072
6073 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6074                                      struct btrfs_root *root,
6075                                      struct inode *dir,
6076                                      const char *name, int name_len,
6077                                      u64 ref_objectid, u64 objectid,
6078                                      umode_t mode, u64 *index)
6079 {
6080         struct inode *inode;
6081         struct btrfs_inode_item *inode_item;
6082         struct btrfs_key *location;
6083         struct btrfs_path *path;
6084         struct btrfs_inode_ref *ref;
6085         struct btrfs_key key[2];
6086         u32 sizes[2];
6087         int nitems = name ? 2 : 1;
6088         unsigned long ptr;
6089         int ret;
6090
6091         path = btrfs_alloc_path();
6092         if (!path)
6093                 return ERR_PTR(-ENOMEM);
6094
6095         inode = new_inode(root->fs_info->sb);
6096         if (!inode) {
6097                 btrfs_free_path(path);
6098                 return ERR_PTR(-ENOMEM);
6099         }
6100
6101         /*
6102          * O_TMPFILE, set link count to 0, so that after this point,
6103          * we fill in an inode item with the correct link count.
6104          */
6105         if (!name)
6106                 set_nlink(inode, 0);
6107
6108         /*
6109          * we have to initialize this early, so we can reclaim the inode
6110          * number if we fail afterwards in this function.
6111          */
6112         inode->i_ino = objectid;
6113
6114         if (dir && name) {
6115                 trace_btrfs_inode_request(dir);
6116
6117                 ret = btrfs_set_inode_index(dir, index);
6118                 if (ret) {
6119                         btrfs_free_path(path);
6120                         iput(inode);
6121                         return ERR_PTR(ret);
6122                 }
6123         } else if (dir) {
6124                 *index = 0;
6125         }
6126         /*
6127          * index_cnt is ignored for everything but a dir,
6128          * btrfs_get_inode_index_count has an explanation for the magic
6129          * number
6130          */
6131         BTRFS_I(inode)->index_cnt = 2;
6132         BTRFS_I(inode)->dir_index = *index;
6133         BTRFS_I(inode)->root = root;
6134         BTRFS_I(inode)->generation = trans->transid;
6135         inode->i_generation = BTRFS_I(inode)->generation;
6136
6137         /*
6138          * We could have gotten an inode number from somebody who was fsynced
6139          * and then removed in this same transaction, so let's just set full
6140          * sync since it will be a full sync anyway and this will blow away the
6141          * old info in the log.
6142          */
6143         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6144
6145         key[0].objectid = objectid;
6146         key[0].type = BTRFS_INODE_ITEM_KEY;
6147         key[0].offset = 0;
6148
6149         sizes[0] = sizeof(struct btrfs_inode_item);
6150
6151         if (name) {
6152                 /*
6153                  * Start new inodes with an inode_ref. This is slightly more
6154                  * efficient for small numbers of hard links since they will
6155                  * be packed into one item. Extended refs will kick in if we
6156                  * add more hard links than can fit in the ref item.
6157                  */
6158                 key[1].objectid = objectid;
6159                 key[1].type = BTRFS_INODE_REF_KEY;
6160                 key[1].offset = ref_objectid;
6161
6162                 sizes[1] = name_len + sizeof(*ref);
6163         }
6164
6165         location = &BTRFS_I(inode)->location;
6166         location->objectid = objectid;
6167         location->offset = 0;
6168         location->type = BTRFS_INODE_ITEM_KEY;
6169
6170         ret = btrfs_insert_inode_locked(inode);
6171         if (ret < 0)
6172                 goto fail;
6173
6174         path->leave_spinning = 1;
6175         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6176         if (ret != 0)
6177                 goto fail_unlock;
6178
6179         inode_init_owner(inode, dir, mode);
6180         inode_set_bytes(inode, 0);
6181
6182         inode->i_mtime = CURRENT_TIME;
6183         inode->i_atime = inode->i_mtime;
6184         inode->i_ctime = inode->i_mtime;
6185         BTRFS_I(inode)->i_otime = inode->i_mtime;
6186
6187         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6188                                   struct btrfs_inode_item);
6189         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6190                              sizeof(*inode_item));
6191         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6192
6193         if (name) {
6194                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6195                                      struct btrfs_inode_ref);
6196                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6197                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6198                 ptr = (unsigned long)(ref + 1);
6199                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6200         }
6201
6202         btrfs_mark_buffer_dirty(path->nodes[0]);
6203         btrfs_free_path(path);
6204
6205         btrfs_inherit_iflags(inode, dir);
6206
6207         if (S_ISREG(mode)) {
6208                 if (btrfs_test_opt(root, NODATASUM))
6209                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6210                 if (btrfs_test_opt(root, NODATACOW))
6211                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6212                                 BTRFS_INODE_NODATASUM;
6213         }
6214
6215         inode_tree_add(inode);
6216
6217         trace_btrfs_inode_new(inode);
6218         btrfs_set_inode_last_trans(trans, inode);
6219
6220         btrfs_update_root_times(trans, root);
6221
6222         ret = btrfs_inode_inherit_props(trans, inode, dir);
6223         if (ret)
6224                 btrfs_err(root->fs_info,
6225                           "error inheriting props for ino %llu (root %llu): %d",
6226                           btrfs_ino(inode), root->root_key.objectid, ret);
6227
6228         return inode;
6229
6230 fail_unlock:
6231         unlock_new_inode(inode);
6232 fail:
6233         if (dir && name)
6234                 BTRFS_I(dir)->index_cnt--;
6235         btrfs_free_path(path);
6236         iput(inode);
6237         return ERR_PTR(ret);
6238 }
6239
6240 static inline u8 btrfs_inode_type(struct inode *inode)
6241 {
6242         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6243 }
6244
6245 /*
6246  * utility function to add 'inode' into 'parent_inode' with
6247  * a give name and a given sequence number.
6248  * if 'add_backref' is true, also insert a backref from the
6249  * inode to the parent directory.
6250  */
6251 int btrfs_add_link(struct btrfs_trans_handle *trans,
6252                    struct inode *parent_inode, struct inode *inode,
6253                    const char *name, int name_len, int add_backref, u64 index)
6254 {
6255         int ret = 0;
6256         struct btrfs_key key;
6257         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6258         u64 ino = btrfs_ino(inode);
6259         u64 parent_ino = btrfs_ino(parent_inode);
6260
6261         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6262                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6263         } else {
6264                 key.objectid = ino;
6265                 key.type = BTRFS_INODE_ITEM_KEY;
6266                 key.offset = 0;
6267         }
6268
6269         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6270                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6271                                          key.objectid, root->root_key.objectid,
6272                                          parent_ino, index, name, name_len);
6273         } else if (add_backref) {
6274                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6275                                              parent_ino, index);
6276         }
6277
6278         /* Nothing to clean up yet */
6279         if (ret)
6280                 return ret;
6281
6282         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6283                                     parent_inode, &key,
6284                                     btrfs_inode_type(inode), index);
6285         if (ret == -EEXIST || ret == -EOVERFLOW)
6286                 goto fail_dir_item;
6287         else if (ret) {
6288                 btrfs_abort_transaction(trans, root, ret);
6289                 return ret;
6290         }
6291
6292         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6293                            name_len * 2);
6294         inode_inc_iversion(parent_inode);
6295         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6296         ret = btrfs_update_inode(trans, root, parent_inode);
6297         if (ret)
6298                 btrfs_abort_transaction(trans, root, ret);
6299         return ret;
6300
6301 fail_dir_item:
6302         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6303                 u64 local_index;
6304                 int err;
6305                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6306                                  key.objectid, root->root_key.objectid,
6307                                  parent_ino, &local_index, name, name_len);
6308
6309         } else if (add_backref) {
6310                 u64 local_index;
6311                 int err;
6312
6313                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6314                                           ino, parent_ino, &local_index);
6315         }
6316         return ret;
6317 }
6318
6319 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6320                             struct inode *dir, struct dentry *dentry,
6321                             struct inode *inode, int backref, u64 index)
6322 {
6323         int err = btrfs_add_link(trans, dir, inode,
6324                                  dentry->d_name.name, dentry->d_name.len,
6325                                  backref, index);
6326         if (err > 0)
6327                 err = -EEXIST;
6328         return err;
6329 }
6330
6331 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6332                         umode_t mode, dev_t rdev)
6333 {
6334         struct btrfs_trans_handle *trans;
6335         struct btrfs_root *root = BTRFS_I(dir)->root;
6336         struct inode *inode = NULL;
6337         int err;
6338         int drop_inode = 0;
6339         u64 objectid;
6340         u64 index = 0;
6341
6342         /*
6343          * 2 for inode item and ref
6344          * 2 for dir items
6345          * 1 for xattr if selinux is on
6346          */
6347         trans = btrfs_start_transaction(root, 5);
6348         if (IS_ERR(trans))
6349                 return PTR_ERR(trans);
6350
6351         err = btrfs_find_free_ino(root, &objectid);
6352         if (err)
6353                 goto out_unlock;
6354
6355         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6356                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6357                                 mode, &index);
6358         if (IS_ERR(inode)) {
6359                 err = PTR_ERR(inode);
6360                 goto out_unlock;
6361         }
6362
6363         /*
6364         * If the active LSM wants to access the inode during
6365         * d_instantiate it needs these. Smack checks to see
6366         * if the filesystem supports xattrs by looking at the
6367         * ops vector.
6368         */
6369         inode->i_op = &btrfs_special_inode_operations;
6370         init_special_inode(inode, inode->i_mode, rdev);
6371
6372         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6373         if (err)
6374                 goto out_unlock_inode;
6375
6376         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6377         if (err) {
6378                 goto out_unlock_inode;
6379         } else {
6380                 btrfs_update_inode(trans, root, inode);
6381                 unlock_new_inode(inode);
6382                 d_instantiate(dentry, inode);
6383         }
6384
6385 out_unlock:
6386         btrfs_end_transaction(trans, root);
6387         btrfs_balance_delayed_items(root);
6388         btrfs_btree_balance_dirty(root);
6389         if (drop_inode) {
6390                 inode_dec_link_count(inode);
6391                 iput(inode);
6392         }
6393         return err;
6394
6395 out_unlock_inode:
6396         drop_inode = 1;
6397         unlock_new_inode(inode);
6398         goto out_unlock;
6399
6400 }
6401
6402 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6403                         umode_t mode, bool excl)
6404 {
6405         struct btrfs_trans_handle *trans;
6406         struct btrfs_root *root = BTRFS_I(dir)->root;
6407         struct inode *inode = NULL;
6408         int drop_inode_on_err = 0;
6409         int err;
6410         u64 objectid;
6411         u64 index = 0;
6412
6413         /*
6414          * 2 for inode item and ref
6415          * 2 for dir items
6416          * 1 for xattr if selinux is on
6417          */
6418         trans = btrfs_start_transaction(root, 5);
6419         if (IS_ERR(trans))
6420                 return PTR_ERR(trans);
6421
6422         err = btrfs_find_free_ino(root, &objectid);
6423         if (err)
6424                 goto out_unlock;
6425
6426         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6427                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6428                                 mode, &index);
6429         if (IS_ERR(inode)) {
6430                 err = PTR_ERR(inode);
6431                 goto out_unlock;
6432         }
6433         drop_inode_on_err = 1;
6434         /*
6435         * If the active LSM wants to access the inode during
6436         * d_instantiate it needs these. Smack checks to see
6437         * if the filesystem supports xattrs by looking at the
6438         * ops vector.
6439         */
6440         inode->i_fop = &btrfs_file_operations;
6441         inode->i_op = &btrfs_file_inode_operations;
6442         inode->i_mapping->a_ops = &btrfs_aops;
6443
6444         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6445         if (err)
6446                 goto out_unlock_inode;
6447
6448         err = btrfs_update_inode(trans, root, inode);
6449         if (err)
6450                 goto out_unlock_inode;
6451
6452         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6453         if (err)
6454                 goto out_unlock_inode;
6455
6456         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6457         unlock_new_inode(inode);
6458         d_instantiate(dentry, inode);
6459
6460 out_unlock:
6461         btrfs_end_transaction(trans, root);
6462         if (err && drop_inode_on_err) {
6463                 inode_dec_link_count(inode);
6464                 iput(inode);
6465         }
6466         btrfs_balance_delayed_items(root);
6467         btrfs_btree_balance_dirty(root);
6468         return err;
6469
6470 out_unlock_inode:
6471         unlock_new_inode(inode);
6472         goto out_unlock;
6473
6474 }
6475
6476 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6477                       struct dentry *dentry)
6478 {
6479         struct btrfs_trans_handle *trans = NULL;
6480         struct btrfs_root *root = BTRFS_I(dir)->root;
6481         struct inode *inode = d_inode(old_dentry);
6482         u64 index;
6483         int err;
6484         int drop_inode = 0;
6485
6486         /* do not allow sys_link's with other subvols of the same device */
6487         if (root->objectid != BTRFS_I(inode)->root->objectid)
6488                 return -EXDEV;
6489
6490         if (inode->i_nlink >= BTRFS_LINK_MAX)
6491                 return -EMLINK;
6492
6493         err = btrfs_set_inode_index(dir, &index);
6494         if (err)
6495                 goto fail;
6496
6497         /*
6498          * 2 items for inode and inode ref
6499          * 2 items for dir items
6500          * 1 item for parent inode
6501          */
6502         trans = btrfs_start_transaction(root, 5);
6503         if (IS_ERR(trans)) {
6504                 err = PTR_ERR(trans);
6505                 trans = NULL;
6506                 goto fail;
6507         }
6508
6509         /* There are several dir indexes for this inode, clear the cache. */
6510         BTRFS_I(inode)->dir_index = 0ULL;
6511         inc_nlink(inode);
6512         inode_inc_iversion(inode);
6513         inode->i_ctime = CURRENT_TIME;
6514         ihold(inode);
6515         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6516
6517         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6518
6519         if (err) {
6520                 drop_inode = 1;
6521         } else {
6522                 struct dentry *parent = dentry->d_parent;
6523                 err = btrfs_update_inode(trans, root, inode);
6524                 if (err)
6525                         goto fail;
6526                 if (inode->i_nlink == 1) {
6527                         /*
6528                          * If new hard link count is 1, it's a file created
6529                          * with open(2) O_TMPFILE flag.
6530                          */
6531                         err = btrfs_orphan_del(trans, inode);
6532                         if (err)
6533                                 goto fail;
6534                 }
6535                 d_instantiate(dentry, inode);
6536                 btrfs_log_new_name(trans, inode, NULL, parent);
6537         }
6538
6539         btrfs_balance_delayed_items(root);
6540 fail:
6541         if (trans)
6542                 btrfs_end_transaction(trans, root);
6543         if (drop_inode) {
6544                 inode_dec_link_count(inode);
6545                 iput(inode);
6546         }
6547         btrfs_btree_balance_dirty(root);
6548         return err;
6549 }
6550
6551 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6552 {
6553         struct inode *inode = NULL;
6554         struct btrfs_trans_handle *trans;
6555         struct btrfs_root *root = BTRFS_I(dir)->root;
6556         int err = 0;
6557         int drop_on_err = 0;
6558         u64 objectid = 0;
6559         u64 index = 0;
6560
6561         /*
6562          * 2 items for inode and ref
6563          * 2 items for dir items
6564          * 1 for xattr if selinux is on
6565          */
6566         trans = btrfs_start_transaction(root, 5);
6567         if (IS_ERR(trans))
6568                 return PTR_ERR(trans);
6569
6570         err = btrfs_find_free_ino(root, &objectid);
6571         if (err)
6572                 goto out_fail;
6573
6574         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6575                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6576                                 S_IFDIR | mode, &index);
6577         if (IS_ERR(inode)) {
6578                 err = PTR_ERR(inode);
6579                 goto out_fail;
6580         }
6581
6582         drop_on_err = 1;
6583         /* these must be set before we unlock the inode */
6584         inode->i_op = &btrfs_dir_inode_operations;
6585         inode->i_fop = &btrfs_dir_file_operations;
6586
6587         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6588         if (err)
6589                 goto out_fail_inode;
6590
6591         btrfs_i_size_write(inode, 0);
6592         err = btrfs_update_inode(trans, root, inode);
6593         if (err)
6594                 goto out_fail_inode;
6595
6596         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6597                              dentry->d_name.len, 0, index);
6598         if (err)
6599                 goto out_fail_inode;
6600
6601         d_instantiate(dentry, inode);
6602         /*
6603          * mkdir is special.  We're unlocking after we call d_instantiate
6604          * to avoid a race with nfsd calling d_instantiate.
6605          */
6606         unlock_new_inode(inode);
6607         drop_on_err = 0;
6608
6609 out_fail:
6610         btrfs_end_transaction(trans, root);
6611         if (drop_on_err) {
6612                 inode_dec_link_count(inode);
6613                 iput(inode);
6614         }
6615         btrfs_balance_delayed_items(root);
6616         btrfs_btree_balance_dirty(root);
6617         return err;
6618
6619 out_fail_inode:
6620         unlock_new_inode(inode);
6621         goto out_fail;
6622 }
6623
6624 /* Find next extent map of a given extent map, caller needs to ensure locks */
6625 static struct extent_map *next_extent_map(struct extent_map *em)
6626 {
6627         struct rb_node *next;
6628
6629         next = rb_next(&em->rb_node);
6630         if (!next)
6631                 return NULL;
6632         return container_of(next, struct extent_map, rb_node);
6633 }
6634
6635 static struct extent_map *prev_extent_map(struct extent_map *em)
6636 {
6637         struct rb_node *prev;
6638
6639         prev = rb_prev(&em->rb_node);
6640         if (!prev)
6641                 return NULL;
6642         return container_of(prev, struct extent_map, rb_node);
6643 }
6644
6645 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6646  * the existing extent is the nearest extent to map_start,
6647  * and an extent that you want to insert, deal with overlap and insert
6648  * the best fitted new extent into the tree.
6649  */
6650 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6651                                 struct extent_map *existing,
6652                                 struct extent_map *em,
6653                                 u64 map_start)
6654 {
6655         struct extent_map *prev;
6656         struct extent_map *next;
6657         u64 start;
6658         u64 end;
6659         u64 start_diff;
6660
6661         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6662
6663         if (existing->start > map_start) {
6664                 next = existing;
6665                 prev = prev_extent_map(next);
6666         } else {
6667                 prev = existing;
6668                 next = next_extent_map(prev);
6669         }
6670
6671         start = prev ? extent_map_end(prev) : em->start;
6672         start = max_t(u64, start, em->start);
6673         end = next ? next->start : extent_map_end(em);
6674         end = min_t(u64, end, extent_map_end(em));
6675         start_diff = start - em->start;
6676         em->start = start;
6677         em->len = end - start;
6678         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6679             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6680                 em->block_start += start_diff;
6681                 em->block_len -= start_diff;
6682         }
6683         return add_extent_mapping(em_tree, em, 0);
6684 }
6685
6686 static noinline int uncompress_inline(struct btrfs_path *path,
6687                                       struct page *page,
6688                                       size_t pg_offset, u64 extent_offset,
6689                                       struct btrfs_file_extent_item *item)
6690 {
6691         int ret;
6692         struct extent_buffer *leaf = path->nodes[0];
6693         char *tmp;
6694         size_t max_size;
6695         unsigned long inline_size;
6696         unsigned long ptr;
6697         int compress_type;
6698
6699         WARN_ON(pg_offset != 0);
6700         compress_type = btrfs_file_extent_compression(leaf, item);
6701         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6702         inline_size = btrfs_file_extent_inline_item_len(leaf,
6703                                         btrfs_item_nr(path->slots[0]));
6704         tmp = kmalloc(inline_size, GFP_NOFS);
6705         if (!tmp)
6706                 return -ENOMEM;
6707         ptr = btrfs_file_extent_inline_start(item);
6708
6709         read_extent_buffer(leaf, tmp, ptr, inline_size);
6710
6711         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6712         ret = btrfs_decompress(compress_type, tmp, page,
6713                                extent_offset, inline_size, max_size);
6714         kfree(tmp);
6715         return ret;
6716 }
6717
6718 /*
6719  * a bit scary, this does extent mapping from logical file offset to the disk.
6720  * the ugly parts come from merging extents from the disk with the in-ram
6721  * representation.  This gets more complex because of the data=ordered code,
6722  * where the in-ram extents might be locked pending data=ordered completion.
6723  *
6724  * This also copies inline extents directly into the page.
6725  */
6726
6727 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6728                                     size_t pg_offset, u64 start, u64 len,
6729                                     int create)
6730 {
6731         int ret;
6732         int err = 0;
6733         u64 extent_start = 0;
6734         u64 extent_end = 0;
6735         u64 objectid = btrfs_ino(inode);
6736         u32 found_type;
6737         struct btrfs_path *path = NULL;
6738         struct btrfs_root *root = BTRFS_I(inode)->root;
6739         struct btrfs_file_extent_item *item;
6740         struct extent_buffer *leaf;
6741         struct btrfs_key found_key;
6742         struct extent_map *em = NULL;
6743         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6744         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6745         struct btrfs_trans_handle *trans = NULL;
6746         const bool new_inline = !page || create;
6747
6748 again:
6749         read_lock(&em_tree->lock);
6750         em = lookup_extent_mapping(em_tree, start, len);
6751         if (em)
6752                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6753         read_unlock(&em_tree->lock);
6754
6755         if (em) {
6756                 if (em->start > start || em->start + em->len <= start)
6757                         free_extent_map(em);
6758                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6759                         free_extent_map(em);
6760                 else
6761                         goto out;
6762         }
6763         em = alloc_extent_map();
6764         if (!em) {
6765                 err = -ENOMEM;
6766                 goto out;
6767         }
6768         em->bdev = root->fs_info->fs_devices->latest_bdev;
6769         em->start = EXTENT_MAP_HOLE;
6770         em->orig_start = EXTENT_MAP_HOLE;
6771         em->len = (u64)-1;
6772         em->block_len = (u64)-1;
6773
6774         if (!path) {
6775                 path = btrfs_alloc_path();
6776                 if (!path) {
6777                         err = -ENOMEM;
6778                         goto out;
6779                 }
6780                 /*
6781                  * Chances are we'll be called again, so go ahead and do
6782                  * readahead
6783                  */
6784                 path->reada = READA_FORWARD;
6785         }
6786
6787         ret = btrfs_lookup_file_extent(trans, root, path,
6788                                        objectid, start, trans != NULL);
6789         if (ret < 0) {
6790                 err = ret;
6791                 goto out;
6792         }
6793
6794         if (ret != 0) {
6795                 if (path->slots[0] == 0)
6796                         goto not_found;
6797                 path->slots[0]--;
6798         }
6799
6800         leaf = path->nodes[0];
6801         item = btrfs_item_ptr(leaf, path->slots[0],
6802                               struct btrfs_file_extent_item);
6803         /* are we inside the extent that was found? */
6804         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6805         found_type = found_key.type;
6806         if (found_key.objectid != objectid ||
6807             found_type != BTRFS_EXTENT_DATA_KEY) {
6808                 /*
6809                  * If we backup past the first extent we want to move forward
6810                  * and see if there is an extent in front of us, otherwise we'll
6811                  * say there is a hole for our whole search range which can
6812                  * cause problems.
6813                  */
6814                 extent_end = start;
6815                 goto next;
6816         }
6817
6818         found_type = btrfs_file_extent_type(leaf, item);
6819         extent_start = found_key.offset;
6820         if (found_type == BTRFS_FILE_EXTENT_REG ||
6821             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6822                 extent_end = extent_start +
6823                        btrfs_file_extent_num_bytes(leaf, item);
6824         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6825                 size_t size;
6826                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6827                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6828         }
6829 next:
6830         if (start >= extent_end) {
6831                 path->slots[0]++;
6832                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6833                         ret = btrfs_next_leaf(root, path);
6834                         if (ret < 0) {
6835                                 err = ret;
6836                                 goto out;
6837                         }
6838                         if (ret > 0)
6839                                 goto not_found;
6840                         leaf = path->nodes[0];
6841                 }
6842                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6843                 if (found_key.objectid != objectid ||
6844                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6845                         goto not_found;
6846                 if (start + len <= found_key.offset)
6847                         goto not_found;
6848                 if (start > found_key.offset)
6849                         goto next;
6850                 em->start = start;
6851                 em->orig_start = start;
6852                 em->len = found_key.offset - start;
6853                 goto not_found_em;
6854         }
6855
6856         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6857
6858         if (found_type == BTRFS_FILE_EXTENT_REG ||
6859             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6860                 goto insert;
6861         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6862                 unsigned long ptr;
6863                 char *map;
6864                 size_t size;
6865                 size_t extent_offset;
6866                 size_t copy_size;
6867
6868                 if (new_inline)
6869                         goto out;
6870
6871                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6872                 extent_offset = page_offset(page) + pg_offset - extent_start;
6873                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6874                                 size - extent_offset);
6875                 em->start = extent_start + extent_offset;
6876                 em->len = ALIGN(copy_size, root->sectorsize);
6877                 em->orig_block_len = em->len;
6878                 em->orig_start = em->start;
6879                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6880                 if (create == 0 && !PageUptodate(page)) {
6881                         if (btrfs_file_extent_compression(leaf, item) !=
6882                             BTRFS_COMPRESS_NONE) {
6883                                 ret = uncompress_inline(path, page, pg_offset,
6884                                                         extent_offset, item);
6885                                 if (ret) {
6886                                         err = ret;
6887                                         goto out;
6888                                 }
6889                         } else {
6890                                 map = kmap(page);
6891                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6892                                                    copy_size);
6893                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6894                                         memset(map + pg_offset + copy_size, 0,
6895                                                PAGE_CACHE_SIZE - pg_offset -
6896                                                copy_size);
6897                                 }
6898                                 kunmap(page);
6899                         }
6900                         flush_dcache_page(page);
6901                 } else if (create && PageUptodate(page)) {
6902                         BUG();
6903                         if (!trans) {
6904                                 kunmap(page);
6905                                 free_extent_map(em);
6906                                 em = NULL;
6907
6908                                 btrfs_release_path(path);
6909                                 trans = btrfs_join_transaction(root);
6910
6911                                 if (IS_ERR(trans))
6912                                         return ERR_CAST(trans);
6913                                 goto again;
6914                         }
6915                         map = kmap(page);
6916                         write_extent_buffer(leaf, map + pg_offset, ptr,
6917                                             copy_size);
6918                         kunmap(page);
6919                         btrfs_mark_buffer_dirty(leaf);
6920                 }
6921                 set_extent_uptodate(io_tree, em->start,
6922                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6923                 goto insert;
6924         }
6925 not_found:
6926         em->start = start;
6927         em->orig_start = start;
6928         em->len = len;
6929 not_found_em:
6930         em->block_start = EXTENT_MAP_HOLE;
6931         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6932 insert:
6933         btrfs_release_path(path);
6934         if (em->start > start || extent_map_end(em) <= start) {
6935                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6936                         em->start, em->len, start, len);
6937                 err = -EIO;
6938                 goto out;
6939         }
6940
6941         err = 0;
6942         write_lock(&em_tree->lock);
6943         ret = add_extent_mapping(em_tree, em, 0);
6944         /* it is possible that someone inserted the extent into the tree
6945          * while we had the lock dropped.  It is also possible that
6946          * an overlapping map exists in the tree
6947          */
6948         if (ret == -EEXIST) {
6949                 struct extent_map *existing;
6950
6951                 ret = 0;
6952
6953                 existing = search_extent_mapping(em_tree, start, len);
6954                 /*
6955                  * existing will always be non-NULL, since there must be
6956                  * extent causing the -EEXIST.
6957                  */
6958                 if (start >= extent_map_end(existing) ||
6959                     start <= existing->start) {
6960                         /*
6961                          * The existing extent map is the one nearest to
6962                          * the [start, start + len) range which overlaps
6963                          */
6964                         err = merge_extent_mapping(em_tree, existing,
6965                                                    em, start);
6966                         free_extent_map(existing);
6967                         if (err) {
6968                                 free_extent_map(em);
6969                                 em = NULL;
6970                         }
6971                 } else {
6972                         free_extent_map(em);
6973                         em = existing;
6974                         err = 0;
6975                 }
6976         }
6977         write_unlock(&em_tree->lock);
6978 out:
6979
6980         trace_btrfs_get_extent(root, em);
6981
6982         btrfs_free_path(path);
6983         if (trans) {
6984                 ret = btrfs_end_transaction(trans, root);
6985                 if (!err)
6986                         err = ret;
6987         }
6988         if (err) {
6989                 free_extent_map(em);
6990                 return ERR_PTR(err);
6991         }
6992         BUG_ON(!em); /* Error is always set */
6993         return em;
6994 }
6995
6996 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
6997                                            size_t pg_offset, u64 start, u64 len,
6998                                            int create)
6999 {
7000         struct extent_map *em;
7001         struct extent_map *hole_em = NULL;
7002         u64 range_start = start;
7003         u64 end;
7004         u64 found;
7005         u64 found_end;
7006         int err = 0;
7007
7008         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7009         if (IS_ERR(em))
7010                 return em;
7011         if (em) {
7012                 /*
7013                  * if our em maps to
7014                  * -  a hole or
7015                  * -  a pre-alloc extent,
7016                  * there might actually be delalloc bytes behind it.
7017                  */
7018                 if (em->block_start != EXTENT_MAP_HOLE &&
7019                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7020                         return em;
7021                 else
7022                         hole_em = em;
7023         }
7024
7025         /* check to see if we've wrapped (len == -1 or similar) */
7026         end = start + len;
7027         if (end < start)
7028                 end = (u64)-1;
7029         else
7030                 end -= 1;
7031
7032         em = NULL;
7033
7034         /* ok, we didn't find anything, lets look for delalloc */
7035         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7036                                  end, len, EXTENT_DELALLOC, 1);
7037         found_end = range_start + found;
7038         if (found_end < range_start)
7039                 found_end = (u64)-1;
7040
7041         /*
7042          * we didn't find anything useful, return
7043          * the original results from get_extent()
7044          */
7045         if (range_start > end || found_end <= start) {
7046                 em = hole_em;
7047                 hole_em = NULL;
7048                 goto out;
7049         }
7050
7051         /* adjust the range_start to make sure it doesn't
7052          * go backwards from the start they passed in
7053          */
7054         range_start = max(start, range_start);
7055         found = found_end - range_start;
7056
7057         if (found > 0) {
7058                 u64 hole_start = start;
7059                 u64 hole_len = len;
7060
7061                 em = alloc_extent_map();
7062                 if (!em) {
7063                         err = -ENOMEM;
7064                         goto out;
7065                 }
7066                 /*
7067                  * when btrfs_get_extent can't find anything it
7068                  * returns one huge hole
7069                  *
7070                  * make sure what it found really fits our range, and
7071                  * adjust to make sure it is based on the start from
7072                  * the caller
7073                  */
7074                 if (hole_em) {
7075                         u64 calc_end = extent_map_end(hole_em);
7076
7077                         if (calc_end <= start || (hole_em->start > end)) {
7078                                 free_extent_map(hole_em);
7079                                 hole_em = NULL;
7080                         } else {
7081                                 hole_start = max(hole_em->start, start);
7082                                 hole_len = calc_end - hole_start;
7083                         }
7084                 }
7085                 em->bdev = NULL;
7086                 if (hole_em && range_start > hole_start) {
7087                         /* our hole starts before our delalloc, so we
7088                          * have to return just the parts of the hole
7089                          * that go until  the delalloc starts
7090                          */
7091                         em->len = min(hole_len,
7092                                       range_start - hole_start);
7093                         em->start = hole_start;
7094                         em->orig_start = hole_start;
7095                         /*
7096                          * don't adjust block start at all,
7097                          * it is fixed at EXTENT_MAP_HOLE
7098                          */
7099                         em->block_start = hole_em->block_start;
7100                         em->block_len = hole_len;
7101                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7102                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7103                 } else {
7104                         em->start = range_start;
7105                         em->len = found;
7106                         em->orig_start = range_start;
7107                         em->block_start = EXTENT_MAP_DELALLOC;
7108                         em->block_len = found;
7109                 }
7110         } else if (hole_em) {
7111                 return hole_em;
7112         }
7113 out:
7114
7115         free_extent_map(hole_em);
7116         if (err) {
7117                 free_extent_map(em);
7118                 return ERR_PTR(err);
7119         }
7120         return em;
7121 }
7122
7123 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7124                                                   u64 start, u64 len)
7125 {
7126         struct btrfs_root *root = BTRFS_I(inode)->root;
7127         struct extent_map *em;
7128         struct btrfs_key ins;
7129         u64 alloc_hint;
7130         int ret;
7131
7132         alloc_hint = get_extent_allocation_hint(inode, start, len);
7133         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7134                                    alloc_hint, &ins, 1, 1);
7135         if (ret)
7136                 return ERR_PTR(ret);
7137
7138         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7139                               ins.offset, ins.offset, ins.offset, 0);
7140         if (IS_ERR(em)) {
7141                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7142                 return em;
7143         }
7144
7145         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7146                                            ins.offset, ins.offset, 0);
7147         if (ret) {
7148                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7149                 free_extent_map(em);
7150                 return ERR_PTR(ret);
7151         }
7152
7153         return em;
7154 }
7155
7156 /*
7157  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7158  * block must be cow'd
7159  */
7160 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7161                               u64 *orig_start, u64 *orig_block_len,
7162                               u64 *ram_bytes)
7163 {
7164         struct btrfs_trans_handle *trans;
7165         struct btrfs_path *path;
7166         int ret;
7167         struct extent_buffer *leaf;
7168         struct btrfs_root *root = BTRFS_I(inode)->root;
7169         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7170         struct btrfs_file_extent_item *fi;
7171         struct btrfs_key key;
7172         u64 disk_bytenr;
7173         u64 backref_offset;
7174         u64 extent_end;
7175         u64 num_bytes;
7176         int slot;
7177         int found_type;
7178         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7179
7180         path = btrfs_alloc_path();
7181         if (!path)
7182                 return -ENOMEM;
7183
7184         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7185                                        offset, 0);
7186         if (ret < 0)
7187                 goto out;
7188
7189         slot = path->slots[0];
7190         if (ret == 1) {
7191                 if (slot == 0) {
7192                         /* can't find the item, must cow */
7193                         ret = 0;
7194                         goto out;
7195                 }
7196                 slot--;
7197         }
7198         ret = 0;
7199         leaf = path->nodes[0];
7200         btrfs_item_key_to_cpu(leaf, &key, slot);
7201         if (key.objectid != btrfs_ino(inode) ||
7202             key.type != BTRFS_EXTENT_DATA_KEY) {
7203                 /* not our file or wrong item type, must cow */
7204                 goto out;
7205         }
7206
7207         if (key.offset > offset) {
7208                 /* Wrong offset, must cow */
7209                 goto out;
7210         }
7211
7212         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7213         found_type = btrfs_file_extent_type(leaf, fi);
7214         if (found_type != BTRFS_FILE_EXTENT_REG &&
7215             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7216                 /* not a regular extent, must cow */
7217                 goto out;
7218         }
7219
7220         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7221                 goto out;
7222
7223         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7224         if (extent_end <= offset)
7225                 goto out;
7226
7227         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7228         if (disk_bytenr == 0)
7229                 goto out;
7230
7231         if (btrfs_file_extent_compression(leaf, fi) ||
7232             btrfs_file_extent_encryption(leaf, fi) ||
7233             btrfs_file_extent_other_encoding(leaf, fi))
7234                 goto out;
7235
7236         backref_offset = btrfs_file_extent_offset(leaf, fi);
7237
7238         if (orig_start) {
7239                 *orig_start = key.offset - backref_offset;
7240                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7241                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7242         }
7243
7244         if (btrfs_extent_readonly(root, disk_bytenr))
7245                 goto out;
7246
7247         num_bytes = min(offset + *len, extent_end) - offset;
7248         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7249                 u64 range_end;
7250
7251                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7252                 ret = test_range_bit(io_tree, offset, range_end,
7253                                      EXTENT_DELALLOC, 0, NULL);
7254                 if (ret) {
7255                         ret = -EAGAIN;
7256                         goto out;
7257                 }
7258         }
7259
7260         btrfs_release_path(path);
7261
7262         /*
7263          * look for other files referencing this extent, if we
7264          * find any we must cow
7265          */
7266         trans = btrfs_join_transaction(root);
7267         if (IS_ERR(trans)) {
7268                 ret = 0;
7269                 goto out;
7270         }
7271
7272         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7273                                     key.offset - backref_offset, disk_bytenr);
7274         btrfs_end_transaction(trans, root);
7275         if (ret) {
7276                 ret = 0;
7277                 goto out;
7278         }
7279
7280         /*
7281          * adjust disk_bytenr and num_bytes to cover just the bytes
7282          * in this extent we are about to write.  If there
7283          * are any csums in that range we have to cow in order
7284          * to keep the csums correct
7285          */
7286         disk_bytenr += backref_offset;
7287         disk_bytenr += offset - key.offset;
7288         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7289                                 goto out;
7290         /*
7291          * all of the above have passed, it is safe to overwrite this extent
7292          * without cow
7293          */
7294         *len = num_bytes;
7295         ret = 1;
7296 out:
7297         btrfs_free_path(path);
7298         return ret;
7299 }
7300
7301 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7302 {
7303         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7304         int found = false;
7305         void **pagep = NULL;
7306         struct page *page = NULL;
7307         int start_idx;
7308         int end_idx;
7309
7310         start_idx = start >> PAGE_CACHE_SHIFT;
7311
7312         /*
7313          * end is the last byte in the last page.  end == start is legal
7314          */
7315         end_idx = end >> PAGE_CACHE_SHIFT;
7316
7317         rcu_read_lock();
7318
7319         /* Most of the code in this while loop is lifted from
7320          * find_get_page.  It's been modified to begin searching from a
7321          * page and return just the first page found in that range.  If the
7322          * found idx is less than or equal to the end idx then we know that
7323          * a page exists.  If no pages are found or if those pages are
7324          * outside of the range then we're fine (yay!) */
7325         while (page == NULL &&
7326                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7327                 page = radix_tree_deref_slot(pagep);
7328                 if (unlikely(!page))
7329                         break;
7330
7331                 if (radix_tree_exception(page)) {
7332                         if (radix_tree_deref_retry(page)) {
7333                                 page = NULL;
7334                                 continue;
7335                         }
7336                         /*
7337                          * Otherwise, shmem/tmpfs must be storing a swap entry
7338                          * here as an exceptional entry: so return it without
7339                          * attempting to raise page count.
7340                          */
7341                         page = NULL;
7342                         break; /* TODO: Is this relevant for this use case? */
7343                 }
7344
7345                 if (!page_cache_get_speculative(page)) {
7346                         page = NULL;
7347                         continue;
7348                 }
7349
7350                 /*
7351                  * Has the page moved?
7352                  * This is part of the lockless pagecache protocol. See
7353                  * include/linux/pagemap.h for details.
7354                  */
7355                 if (unlikely(page != *pagep)) {
7356                         page_cache_release(page);
7357                         page = NULL;
7358                 }
7359         }
7360
7361         if (page) {
7362                 if (page->index <= end_idx)
7363                         found = true;
7364                 page_cache_release(page);
7365         }
7366
7367         rcu_read_unlock();
7368         return found;
7369 }
7370
7371 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7372                               struct extent_state **cached_state, int writing)
7373 {
7374         struct btrfs_ordered_extent *ordered;
7375         int ret = 0;
7376
7377         while (1) {
7378                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7379                                  cached_state);
7380                 /*
7381                  * We're concerned with the entire range that we're going to be
7382                  * doing DIO to, so we need to make sure theres no ordered
7383                  * extents in this range.
7384                  */
7385                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7386                                                      lockend - lockstart + 1);
7387
7388                 /*
7389                  * We need to make sure there are no buffered pages in this
7390                  * range either, we could have raced between the invalidate in
7391                  * generic_file_direct_write and locking the extent.  The
7392                  * invalidate needs to happen so that reads after a write do not
7393                  * get stale data.
7394                  */
7395                 if (!ordered &&
7396                     (!writing ||
7397                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7398                         break;
7399
7400                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7401                                      cached_state, GFP_NOFS);
7402
7403                 if (ordered) {
7404                         btrfs_start_ordered_extent(inode, ordered, 1);
7405                         btrfs_put_ordered_extent(ordered);
7406                 } else {
7407                         /*
7408                          * We could trigger writeback for this range (and wait
7409                          * for it to complete) and then invalidate the pages for
7410                          * this range (through invalidate_inode_pages2_range()),
7411                          * but that can lead us to a deadlock with a concurrent
7412                          * call to readpages() (a buffered read or a defrag call
7413                          * triggered a readahead) on a page lock due to an
7414                          * ordered dio extent we created before but did not have
7415                          * yet a corresponding bio submitted (whence it can not
7416                          * complete), which makes readpages() wait for that
7417                          * ordered extent to complete while holding a lock on
7418                          * that page.
7419                          */
7420                         ret = -ENOTBLK;
7421                         break;
7422                 }
7423
7424                 cond_resched();
7425         }
7426
7427         return ret;
7428 }
7429
7430 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7431                                            u64 len, u64 orig_start,
7432                                            u64 block_start, u64 block_len,
7433                                            u64 orig_block_len, u64 ram_bytes,
7434                                            int type)
7435 {
7436         struct extent_map_tree *em_tree;
7437         struct extent_map *em;
7438         struct btrfs_root *root = BTRFS_I(inode)->root;
7439         int ret;
7440
7441         em_tree = &BTRFS_I(inode)->extent_tree;
7442         em = alloc_extent_map();
7443         if (!em)
7444                 return ERR_PTR(-ENOMEM);
7445
7446         em->start = start;
7447         em->orig_start = orig_start;
7448         em->mod_start = start;
7449         em->mod_len = len;
7450         em->len = len;
7451         em->block_len = block_len;
7452         em->block_start = block_start;
7453         em->bdev = root->fs_info->fs_devices->latest_bdev;
7454         em->orig_block_len = orig_block_len;
7455         em->ram_bytes = ram_bytes;
7456         em->generation = -1;
7457         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7458         if (type == BTRFS_ORDERED_PREALLOC)
7459                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7460
7461         do {
7462                 btrfs_drop_extent_cache(inode, em->start,
7463                                 em->start + em->len - 1, 0);
7464                 write_lock(&em_tree->lock);
7465                 ret = add_extent_mapping(em_tree, em, 1);
7466                 write_unlock(&em_tree->lock);
7467         } while (ret == -EEXIST);
7468
7469         if (ret) {
7470                 free_extent_map(em);
7471                 return ERR_PTR(ret);
7472         }
7473
7474         return em;
7475 }
7476
7477 static void adjust_dio_outstanding_extents(struct inode *inode,
7478                                            struct btrfs_dio_data *dio_data,
7479                                            const u64 len)
7480 {
7481         unsigned num_extents;
7482
7483         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7484                                            BTRFS_MAX_EXTENT_SIZE);
7485         /*
7486          * If we have an outstanding_extents count still set then we're
7487          * within our reservation, otherwise we need to adjust our inode
7488          * counter appropriately.
7489          */
7490         if (dio_data->outstanding_extents) {
7491                 dio_data->outstanding_extents -= num_extents;
7492         } else {
7493                 spin_lock(&BTRFS_I(inode)->lock);
7494                 BTRFS_I(inode)->outstanding_extents += num_extents;
7495                 spin_unlock(&BTRFS_I(inode)->lock);
7496         }
7497 }
7498
7499 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7500                                    struct buffer_head *bh_result, int create)
7501 {
7502         struct extent_map *em;
7503         struct btrfs_root *root = BTRFS_I(inode)->root;
7504         struct extent_state *cached_state = NULL;
7505         struct btrfs_dio_data *dio_data = NULL;
7506         u64 start = iblock << inode->i_blkbits;
7507         u64 lockstart, lockend;
7508         u64 len = bh_result->b_size;
7509         int unlock_bits = EXTENT_LOCKED;
7510         int ret = 0;
7511
7512         if (create)
7513                 unlock_bits |= EXTENT_DIRTY;
7514         else
7515                 len = min_t(u64, len, root->sectorsize);
7516
7517         lockstart = start;
7518         lockend = start + len - 1;
7519
7520         if (current->journal_info) {
7521                 /*
7522                  * Need to pull our outstanding extents and set journal_info to NULL so
7523                  * that anything that needs to check if there's a transction doesn't get
7524                  * confused.
7525                  */
7526                 dio_data = current->journal_info;
7527                 current->journal_info = NULL;
7528         }
7529
7530         /*
7531          * If this errors out it's because we couldn't invalidate pagecache for
7532          * this range and we need to fallback to buffered.
7533          */
7534         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7535                                create)) {
7536                 ret = -ENOTBLK;
7537                 goto err;
7538         }
7539
7540         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7541         if (IS_ERR(em)) {
7542                 ret = PTR_ERR(em);
7543                 goto unlock_err;
7544         }
7545
7546         /*
7547          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7548          * io.  INLINE is special, and we could probably kludge it in here, but
7549          * it's still buffered so for safety lets just fall back to the generic
7550          * buffered path.
7551          *
7552          * For COMPRESSED we _have_ to read the entire extent in so we can
7553          * decompress it, so there will be buffering required no matter what we
7554          * do, so go ahead and fallback to buffered.
7555          *
7556          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7557          * to buffered IO.  Don't blame me, this is the price we pay for using
7558          * the generic code.
7559          */
7560         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7561             em->block_start == EXTENT_MAP_INLINE) {
7562                 free_extent_map(em);
7563                 ret = -ENOTBLK;
7564                 goto unlock_err;
7565         }
7566
7567         /* Just a good old fashioned hole, return */
7568         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7569                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7570                 free_extent_map(em);
7571                 goto unlock_err;
7572         }
7573
7574         /*
7575          * We don't allocate a new extent in the following cases
7576          *
7577          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7578          * existing extent.
7579          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7580          * just use the extent.
7581          *
7582          */
7583         if (!create) {
7584                 len = min(len, em->len - (start - em->start));
7585                 lockstart = start + len;
7586                 goto unlock;
7587         }
7588
7589         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7590             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7591              em->block_start != EXTENT_MAP_HOLE)) {
7592                 int type;
7593                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7594
7595                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7596                         type = BTRFS_ORDERED_PREALLOC;
7597                 else
7598                         type = BTRFS_ORDERED_NOCOW;
7599                 len = min(len, em->len - (start - em->start));
7600                 block_start = em->block_start + (start - em->start);
7601
7602                 if (can_nocow_extent(inode, start, &len, &orig_start,
7603                                      &orig_block_len, &ram_bytes) == 1) {
7604                         if (type == BTRFS_ORDERED_PREALLOC) {
7605                                 free_extent_map(em);
7606                                 em = create_pinned_em(inode, start, len,
7607                                                        orig_start,
7608                                                        block_start, len,
7609                                                        orig_block_len,
7610                                                        ram_bytes, type);
7611                                 if (IS_ERR(em)) {
7612                                         ret = PTR_ERR(em);
7613                                         goto unlock_err;
7614                                 }
7615                         }
7616
7617                         ret = btrfs_add_ordered_extent_dio(inode, start,
7618                                            block_start, len, len, type);
7619                         if (ret) {
7620                                 free_extent_map(em);
7621                                 goto unlock_err;
7622                         }
7623                         goto unlock;
7624                 }
7625         }
7626
7627         /*
7628          * this will cow the extent, reset the len in case we changed
7629          * it above
7630          */
7631         len = bh_result->b_size;
7632         free_extent_map(em);
7633         em = btrfs_new_extent_direct(inode, start, len);
7634         if (IS_ERR(em)) {
7635                 ret = PTR_ERR(em);
7636                 goto unlock_err;
7637         }
7638         len = min(len, em->len - (start - em->start));
7639 unlock:
7640         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7641                 inode->i_blkbits;
7642         bh_result->b_size = len;
7643         bh_result->b_bdev = em->bdev;
7644         set_buffer_mapped(bh_result);
7645         if (create) {
7646                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7647                         set_buffer_new(bh_result);
7648
7649                 /*
7650                  * Need to update the i_size under the extent lock so buffered
7651                  * readers will get the updated i_size when we unlock.
7652                  */
7653                 if (start + len > i_size_read(inode))
7654                         i_size_write(inode, start + len);
7655
7656                 adjust_dio_outstanding_extents(inode, dio_data, len);
7657                 btrfs_free_reserved_data_space(inode, start, len);
7658                 WARN_ON(dio_data->reserve < len);
7659                 dio_data->reserve -= len;
7660                 dio_data->unsubmitted_oe_range_end = start + len;
7661                 current->journal_info = dio_data;
7662         }
7663
7664         /*
7665          * In the case of write we need to clear and unlock the entire range,
7666          * in the case of read we need to unlock only the end area that we
7667          * aren't using if there is any left over space.
7668          */
7669         if (lockstart < lockend) {
7670                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7671                                  lockend, unlock_bits, 1, 0,
7672                                  &cached_state, GFP_NOFS);
7673         } else {
7674                 free_extent_state(cached_state);
7675         }
7676
7677         free_extent_map(em);
7678
7679         return 0;
7680
7681 unlock_err:
7682         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7683                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7684 err:
7685         if (dio_data)
7686                 current->journal_info = dio_data;
7687         /*
7688          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7689          * write less data then expected, so that we don't underflow our inode's
7690          * outstanding extents counter.
7691          */
7692         if (create && dio_data)
7693                 adjust_dio_outstanding_extents(inode, dio_data, len);
7694
7695         return ret;
7696 }
7697
7698 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7699                                         int rw, int mirror_num)
7700 {
7701         struct btrfs_root *root = BTRFS_I(inode)->root;
7702         int ret;
7703
7704         BUG_ON(rw & REQ_WRITE);
7705
7706         bio_get(bio);
7707
7708         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7709                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7710         if (ret)
7711                 goto err;
7712
7713         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7714 err:
7715         bio_put(bio);
7716         return ret;
7717 }
7718
7719 static int btrfs_check_dio_repairable(struct inode *inode,
7720                                       struct bio *failed_bio,
7721                                       struct io_failure_record *failrec,
7722                                       int failed_mirror)
7723 {
7724         int num_copies;
7725
7726         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7727                                       failrec->logical, failrec->len);
7728         if (num_copies == 1) {
7729                 /*
7730                  * we only have a single copy of the data, so don't bother with
7731                  * all the retry and error correction code that follows. no
7732                  * matter what the error is, it is very likely to persist.
7733                  */
7734                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7735                          num_copies, failrec->this_mirror, failed_mirror);
7736                 return 0;
7737         }
7738
7739         failrec->failed_mirror = failed_mirror;
7740         failrec->this_mirror++;
7741         if (failrec->this_mirror == failed_mirror)
7742                 failrec->this_mirror++;
7743
7744         if (failrec->this_mirror > num_copies) {
7745                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7746                          num_copies, failrec->this_mirror, failed_mirror);
7747                 return 0;
7748         }
7749
7750         return 1;
7751 }
7752
7753 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7754                           struct page *page, u64 start, u64 end,
7755                           int failed_mirror, bio_end_io_t *repair_endio,
7756                           void *repair_arg)
7757 {
7758         struct io_failure_record *failrec;
7759         struct bio *bio;
7760         int isector;
7761         int read_mode;
7762         int ret;
7763
7764         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7765
7766         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7767         if (ret)
7768                 return ret;
7769
7770         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7771                                          failed_mirror);
7772         if (!ret) {
7773                 free_io_failure(inode, failrec);
7774                 return -EIO;
7775         }
7776
7777         if (failed_bio->bi_vcnt > 1)
7778                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7779         else
7780                 read_mode = READ_SYNC;
7781
7782         isector = start - btrfs_io_bio(failed_bio)->logical;
7783         isector >>= inode->i_sb->s_blocksize_bits;
7784         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7785                                       0, isector, repair_endio, repair_arg);
7786         if (!bio) {
7787                 free_io_failure(inode, failrec);
7788                 return -EIO;
7789         }
7790
7791         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7792                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7793                     read_mode, failrec->this_mirror, failrec->in_validation);
7794
7795         ret = submit_dio_repair_bio(inode, bio, read_mode,
7796                                     failrec->this_mirror);
7797         if (ret) {
7798                 free_io_failure(inode, failrec);
7799                 bio_put(bio);
7800         }
7801
7802         return ret;
7803 }
7804
7805 struct btrfs_retry_complete {
7806         struct completion done;
7807         struct inode *inode;
7808         u64 start;
7809         int uptodate;
7810 };
7811
7812 static void btrfs_retry_endio_nocsum(struct bio *bio)
7813 {
7814         struct btrfs_retry_complete *done = bio->bi_private;
7815         struct bio_vec *bvec;
7816         int i;
7817
7818         if (bio->bi_error)
7819                 goto end;
7820
7821         done->uptodate = 1;
7822         bio_for_each_segment_all(bvec, bio, i)
7823                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7824 end:
7825         complete(&done->done);
7826         bio_put(bio);
7827 }
7828
7829 static int __btrfs_correct_data_nocsum(struct inode *inode,
7830                                        struct btrfs_io_bio *io_bio)
7831 {
7832         struct bio_vec *bvec;
7833         struct btrfs_retry_complete done;
7834         u64 start;
7835         int i;
7836         int ret;
7837
7838         start = io_bio->logical;
7839         done.inode = inode;
7840
7841         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7842 try_again:
7843                 done.uptodate = 0;
7844                 done.start = start;
7845                 init_completion(&done.done);
7846
7847                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7848                                      start + bvec->bv_len - 1,
7849                                      io_bio->mirror_num,
7850                                      btrfs_retry_endio_nocsum, &done);
7851                 if (ret)
7852                         return ret;
7853
7854                 wait_for_completion(&done.done);
7855
7856                 if (!done.uptodate) {
7857                         /* We might have another mirror, so try again */
7858                         goto try_again;
7859                 }
7860
7861                 start += bvec->bv_len;
7862         }
7863
7864         return 0;
7865 }
7866
7867 static void btrfs_retry_endio(struct bio *bio)
7868 {
7869         struct btrfs_retry_complete *done = bio->bi_private;
7870         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7871         struct bio_vec *bvec;
7872         int uptodate;
7873         int ret;
7874         int i;
7875
7876         if (bio->bi_error)
7877                 goto end;
7878
7879         uptodate = 1;
7880         bio_for_each_segment_all(bvec, bio, i) {
7881                 ret = __readpage_endio_check(done->inode, io_bio, i,
7882                                              bvec->bv_page, 0,
7883                                              done->start, bvec->bv_len);
7884                 if (!ret)
7885                         clean_io_failure(done->inode, done->start,
7886                                          bvec->bv_page, 0);
7887                 else
7888                         uptodate = 0;
7889         }
7890
7891         done->uptodate = uptodate;
7892 end:
7893         complete(&done->done);
7894         bio_put(bio);
7895 }
7896
7897 static int __btrfs_subio_endio_read(struct inode *inode,
7898                                     struct btrfs_io_bio *io_bio, int err)
7899 {
7900         struct bio_vec *bvec;
7901         struct btrfs_retry_complete done;
7902         u64 start;
7903         u64 offset = 0;
7904         int i;
7905         int ret;
7906
7907         err = 0;
7908         start = io_bio->logical;
7909         done.inode = inode;
7910
7911         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7912                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7913                                              0, start, bvec->bv_len);
7914                 if (likely(!ret))
7915                         goto next;
7916 try_again:
7917                 done.uptodate = 0;
7918                 done.start = start;
7919                 init_completion(&done.done);
7920
7921                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7922                                      start + bvec->bv_len - 1,
7923                                      io_bio->mirror_num,
7924                                      btrfs_retry_endio, &done);
7925                 if (ret) {
7926                         err = ret;
7927                         goto next;
7928                 }
7929
7930                 wait_for_completion(&done.done);
7931
7932                 if (!done.uptodate) {
7933                         /* We might have another mirror, so try again */
7934                         goto try_again;
7935                 }
7936 next:
7937                 offset += bvec->bv_len;
7938                 start += bvec->bv_len;
7939         }
7940
7941         return err;
7942 }
7943
7944 static int btrfs_subio_endio_read(struct inode *inode,
7945                                   struct btrfs_io_bio *io_bio, int err)
7946 {
7947         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7948
7949         if (skip_csum) {
7950                 if (unlikely(err))
7951                         return __btrfs_correct_data_nocsum(inode, io_bio);
7952                 else
7953                         return 0;
7954         } else {
7955                 return __btrfs_subio_endio_read(inode, io_bio, err);
7956         }
7957 }
7958
7959 static void btrfs_endio_direct_read(struct bio *bio)
7960 {
7961         struct btrfs_dio_private *dip = bio->bi_private;
7962         struct inode *inode = dip->inode;
7963         struct bio *dio_bio;
7964         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7965         int err = bio->bi_error;
7966
7967         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7968                 err = btrfs_subio_endio_read(inode, io_bio, err);
7969
7970         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7971                       dip->logical_offset + dip->bytes - 1);
7972         dio_bio = dip->dio_bio;
7973
7974         kfree(dip);
7975
7976         dio_end_io(dio_bio, bio->bi_error);
7977
7978         if (io_bio->end_io)
7979                 io_bio->end_io(io_bio, err);
7980         bio_put(bio);
7981 }
7982
7983 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
7984                                                     const u64 offset,
7985                                                     const u64 bytes,
7986                                                     const int uptodate)
7987 {
7988         struct btrfs_root *root = BTRFS_I(inode)->root;
7989         struct btrfs_ordered_extent *ordered = NULL;
7990         u64 ordered_offset = offset;
7991         u64 ordered_bytes = bytes;
7992         int ret;
7993
7994 again:
7995         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
7996                                                    &ordered_offset,
7997                                                    ordered_bytes,
7998                                                    uptodate);
7999         if (!ret)
8000                 goto out_test;
8001
8002         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8003                         finish_ordered_fn, NULL, NULL);
8004         btrfs_queue_work(root->fs_info->endio_write_workers,
8005                          &ordered->work);
8006 out_test:
8007         /*
8008          * our bio might span multiple ordered extents.  If we haven't
8009          * completed the accounting for the whole dio, go back and try again
8010          */
8011         if (ordered_offset < offset + bytes) {
8012                 ordered_bytes = offset + bytes - ordered_offset;
8013                 ordered = NULL;
8014                 goto again;
8015         }
8016 }
8017
8018 static void btrfs_endio_direct_write(struct bio *bio)
8019 {
8020         struct btrfs_dio_private *dip = bio->bi_private;
8021         struct bio *dio_bio = dip->dio_bio;
8022
8023         btrfs_endio_direct_write_update_ordered(dip->inode,
8024                                                 dip->logical_offset,
8025                                                 dip->bytes,
8026                                                 !bio->bi_error);
8027
8028         kfree(dip);
8029
8030         dio_end_io(dio_bio, bio->bi_error);
8031         bio_put(bio);
8032 }
8033
8034 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8035                                     struct bio *bio, int mirror_num,
8036                                     unsigned long bio_flags, u64 offset)
8037 {
8038         int ret;
8039         struct btrfs_root *root = BTRFS_I(inode)->root;
8040         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8041         BUG_ON(ret); /* -ENOMEM */
8042         return 0;
8043 }
8044
8045 static void btrfs_end_dio_bio(struct bio *bio)
8046 {
8047         struct btrfs_dio_private *dip = bio->bi_private;
8048         int err = bio->bi_error;
8049
8050         if (err)
8051                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8052                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8053                            btrfs_ino(dip->inode), bio->bi_rw,
8054                            (unsigned long long)bio->bi_iter.bi_sector,
8055                            bio->bi_iter.bi_size, err);
8056
8057         if (dip->subio_endio)
8058                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8059
8060         if (err) {
8061                 dip->errors = 1;
8062
8063                 /*
8064                  * before atomic variable goto zero, we must make sure
8065                  * dip->errors is perceived to be set.
8066                  */
8067                 smp_mb__before_atomic();
8068         }
8069
8070         /* if there are more bios still pending for this dio, just exit */
8071         if (!atomic_dec_and_test(&dip->pending_bios))
8072                 goto out;
8073
8074         if (dip->errors) {
8075                 bio_io_error(dip->orig_bio);
8076         } else {
8077                 dip->dio_bio->bi_error = 0;
8078                 bio_endio(dip->orig_bio);
8079         }
8080 out:
8081         bio_put(bio);
8082 }
8083
8084 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8085                                        u64 first_sector, gfp_t gfp_flags)
8086 {
8087         struct bio *bio;
8088         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8089         if (bio)
8090                 bio_associate_current(bio);
8091         return bio;
8092 }
8093
8094 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8095                                                  struct inode *inode,
8096                                                  struct btrfs_dio_private *dip,
8097                                                  struct bio *bio,
8098                                                  u64 file_offset)
8099 {
8100         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8101         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8102         int ret;
8103
8104         /*
8105          * We load all the csum data we need when we submit
8106          * the first bio to reduce the csum tree search and
8107          * contention.
8108          */
8109         if (dip->logical_offset == file_offset) {
8110                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8111                                                 file_offset);
8112                 if (ret)
8113                         return ret;
8114         }
8115
8116         if (bio == dip->orig_bio)
8117                 return 0;
8118
8119         file_offset -= dip->logical_offset;
8120         file_offset >>= inode->i_sb->s_blocksize_bits;
8121         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8122
8123         return 0;
8124 }
8125
8126 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8127                                          int rw, u64 file_offset, int skip_sum,
8128                                          int async_submit)
8129 {
8130         struct btrfs_dio_private *dip = bio->bi_private;
8131         int write = rw & REQ_WRITE;
8132         struct btrfs_root *root = BTRFS_I(inode)->root;
8133         int ret;
8134
8135         if (async_submit)
8136                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8137
8138         bio_get(bio);
8139
8140         if (!write) {
8141                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8142                                 BTRFS_WQ_ENDIO_DATA);
8143                 if (ret)
8144                         goto err;
8145         }
8146
8147         if (skip_sum)
8148                 goto map;
8149
8150         if (write && async_submit) {
8151                 ret = btrfs_wq_submit_bio(root->fs_info,
8152                                    inode, rw, bio, 0, 0,
8153                                    file_offset,
8154                                    __btrfs_submit_bio_start_direct_io,
8155                                    __btrfs_submit_bio_done);
8156                 goto err;
8157         } else if (write) {
8158                 /*
8159                  * If we aren't doing async submit, calculate the csum of the
8160                  * bio now.
8161                  */
8162                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8163                 if (ret)
8164                         goto err;
8165         } else {
8166                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8167                                                      file_offset);
8168                 if (ret)
8169                         goto err;
8170         }
8171 map:
8172         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8173 err:
8174         bio_put(bio);
8175         return ret;
8176 }
8177
8178 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8179                                     int skip_sum)
8180 {
8181         struct inode *inode = dip->inode;
8182         struct btrfs_root *root = BTRFS_I(inode)->root;
8183         struct bio *bio;
8184         struct bio *orig_bio = dip->orig_bio;
8185         struct bio_vec *bvec = orig_bio->bi_io_vec;
8186         u64 start_sector = orig_bio->bi_iter.bi_sector;
8187         u64 file_offset = dip->logical_offset;
8188         u64 submit_len = 0;
8189         u64 map_length;
8190         int nr_pages = 0;
8191         int ret;
8192         int async_submit = 0;
8193
8194         map_length = orig_bio->bi_iter.bi_size;
8195         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8196                               &map_length, NULL, 0);
8197         if (ret)
8198                 return -EIO;
8199
8200         if (map_length >= orig_bio->bi_iter.bi_size) {
8201                 bio = orig_bio;
8202                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8203                 goto submit;
8204         }
8205
8206         /* async crcs make it difficult to collect full stripe writes. */
8207         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8208                 async_submit = 0;
8209         else
8210                 async_submit = 1;
8211
8212         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8213         if (!bio)
8214                 return -ENOMEM;
8215
8216         bio->bi_private = dip;
8217         bio->bi_end_io = btrfs_end_dio_bio;
8218         btrfs_io_bio(bio)->logical = file_offset;
8219         atomic_inc(&dip->pending_bios);
8220
8221         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8222                 if (map_length < submit_len + bvec->bv_len ||
8223                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8224                                  bvec->bv_offset) < bvec->bv_len) {
8225                         /*
8226                          * inc the count before we submit the bio so
8227                          * we know the end IO handler won't happen before
8228                          * we inc the count. Otherwise, the dip might get freed
8229                          * before we're done setting it up
8230                          */
8231                         atomic_inc(&dip->pending_bios);
8232                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8233                                                      file_offset, skip_sum,
8234                                                      async_submit);
8235                         if (ret) {
8236                                 bio_put(bio);
8237                                 atomic_dec(&dip->pending_bios);
8238                                 goto out_err;
8239                         }
8240
8241                         start_sector += submit_len >> 9;
8242                         file_offset += submit_len;
8243
8244                         submit_len = 0;
8245                         nr_pages = 0;
8246
8247                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8248                                                   start_sector, GFP_NOFS);
8249                         if (!bio)
8250                                 goto out_err;
8251                         bio->bi_private = dip;
8252                         bio->bi_end_io = btrfs_end_dio_bio;
8253                         btrfs_io_bio(bio)->logical = file_offset;
8254
8255                         map_length = orig_bio->bi_iter.bi_size;
8256                         ret = btrfs_map_block(root->fs_info, rw,
8257                                               start_sector << 9,
8258                                               &map_length, NULL, 0);
8259                         if (ret) {
8260                                 bio_put(bio);
8261                                 goto out_err;
8262                         }
8263                 } else {
8264                         submit_len += bvec->bv_len;
8265                         nr_pages++;
8266                         bvec++;
8267                 }
8268         }
8269
8270 submit:
8271         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8272                                      async_submit);
8273         if (!ret)
8274                 return 0;
8275
8276         bio_put(bio);
8277 out_err:
8278         dip->errors = 1;
8279         /*
8280          * before atomic variable goto zero, we must
8281          * make sure dip->errors is perceived to be set.
8282          */
8283         smp_mb__before_atomic();
8284         if (atomic_dec_and_test(&dip->pending_bios))
8285                 bio_io_error(dip->orig_bio);
8286
8287         /* bio_end_io() will handle error, so we needn't return it */
8288         return 0;
8289 }
8290
8291 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8292                                 struct inode *inode, loff_t file_offset)
8293 {
8294         struct btrfs_dio_private *dip = NULL;
8295         struct bio *io_bio = NULL;
8296         struct btrfs_io_bio *btrfs_bio;
8297         int skip_sum;
8298         int write = rw & REQ_WRITE;
8299         int ret = 0;
8300
8301         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8302
8303         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8304         if (!io_bio) {
8305                 ret = -ENOMEM;
8306                 goto free_ordered;
8307         }
8308
8309         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8310         if (!dip) {
8311                 ret = -ENOMEM;
8312                 goto free_ordered;
8313         }
8314
8315         dip->private = dio_bio->bi_private;
8316         dip->inode = inode;
8317         dip->logical_offset = file_offset;
8318         dip->bytes = dio_bio->bi_iter.bi_size;
8319         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8320         io_bio->bi_private = dip;
8321         dip->orig_bio = io_bio;
8322         dip->dio_bio = dio_bio;
8323         atomic_set(&dip->pending_bios, 0);
8324         btrfs_bio = btrfs_io_bio(io_bio);
8325         btrfs_bio->logical = file_offset;
8326
8327         if (write) {
8328                 io_bio->bi_end_io = btrfs_endio_direct_write;
8329         } else {
8330                 io_bio->bi_end_io = btrfs_endio_direct_read;
8331                 dip->subio_endio = btrfs_subio_endio_read;
8332         }
8333
8334         /*
8335          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8336          * even if we fail to submit a bio, because in such case we do the
8337          * corresponding error handling below and it must not be done a second
8338          * time by btrfs_direct_IO().
8339          */
8340         if (write) {
8341                 struct btrfs_dio_data *dio_data = current->journal_info;
8342
8343                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8344                         dip->bytes;
8345                 dio_data->unsubmitted_oe_range_start =
8346                         dio_data->unsubmitted_oe_range_end;
8347         }
8348
8349         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8350         if (!ret)
8351                 return;
8352
8353         if (btrfs_bio->end_io)
8354                 btrfs_bio->end_io(btrfs_bio, ret);
8355
8356 free_ordered:
8357         /*
8358          * If we arrived here it means either we failed to submit the dip
8359          * or we either failed to clone the dio_bio or failed to allocate the
8360          * dip. If we cloned the dio_bio and allocated the dip, we can just
8361          * call bio_endio against our io_bio so that we get proper resource
8362          * cleanup if we fail to submit the dip, otherwise, we must do the
8363          * same as btrfs_endio_direct_[write|read] because we can't call these
8364          * callbacks - they require an allocated dip and a clone of dio_bio.
8365          */
8366         if (io_bio && dip) {
8367                 io_bio->bi_error = -EIO;
8368                 bio_endio(io_bio);
8369                 /*
8370                  * The end io callbacks free our dip, do the final put on io_bio
8371                  * and all the cleanup and final put for dio_bio (through
8372                  * dio_end_io()).
8373                  */
8374                 dip = NULL;
8375                 io_bio = NULL;
8376         } else {
8377                 if (write)
8378                         btrfs_endio_direct_write_update_ordered(inode,
8379                                                 file_offset,
8380                                                 dio_bio->bi_iter.bi_size,
8381                                                 0);
8382                 else
8383                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8384                               file_offset + dio_bio->bi_iter.bi_size - 1);
8385
8386                 dio_bio->bi_error = -EIO;
8387                 /*
8388                  * Releases and cleans up our dio_bio, no need to bio_put()
8389                  * nor bio_endio()/bio_io_error() against dio_bio.
8390                  */
8391                 dio_end_io(dio_bio, ret);
8392         }
8393         if (io_bio)
8394                 bio_put(io_bio);
8395         kfree(dip);
8396 }
8397
8398 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8399                         const struct iov_iter *iter, loff_t offset)
8400 {
8401         int seg;
8402         int i;
8403         unsigned blocksize_mask = root->sectorsize - 1;
8404         ssize_t retval = -EINVAL;
8405
8406         if (offset & blocksize_mask)
8407                 goto out;
8408
8409         if (iov_iter_alignment(iter) & blocksize_mask)
8410                 goto out;
8411
8412         /* If this is a write we don't need to check anymore */
8413         if (iov_iter_rw(iter) == WRITE)
8414                 return 0;
8415         /*
8416          * Check to make sure we don't have duplicate iov_base's in this
8417          * iovec, if so return EINVAL, otherwise we'll get csum errors
8418          * when reading back.
8419          */
8420         for (seg = 0; seg < iter->nr_segs; seg++) {
8421                 for (i = seg + 1; i < iter->nr_segs; i++) {
8422                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8423                                 goto out;
8424                 }
8425         }
8426         retval = 0;
8427 out:
8428         return retval;
8429 }
8430
8431 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8432                                loff_t offset)
8433 {
8434         struct file *file = iocb->ki_filp;
8435         struct inode *inode = file->f_mapping->host;
8436         struct btrfs_root *root = BTRFS_I(inode)->root;
8437         struct btrfs_dio_data dio_data = { 0 };
8438         size_t count = 0;
8439         int flags = 0;
8440         bool wakeup = true;
8441         bool relock = false;
8442         ssize_t ret;
8443
8444         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8445                 return 0;
8446
8447         inode_dio_begin(inode);
8448         smp_mb__after_atomic();
8449
8450         /*
8451          * The generic stuff only does filemap_write_and_wait_range, which
8452          * isn't enough if we've written compressed pages to this area, so
8453          * we need to flush the dirty pages again to make absolutely sure
8454          * that any outstanding dirty pages are on disk.
8455          */
8456         count = iov_iter_count(iter);
8457         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8458                      &BTRFS_I(inode)->runtime_flags))
8459                 filemap_fdatawrite_range(inode->i_mapping, offset,
8460                                          offset + count - 1);
8461
8462         if (iov_iter_rw(iter) == WRITE) {
8463                 /*
8464                  * If the write DIO is beyond the EOF, we need update
8465                  * the isize, but it is protected by i_mutex. So we can
8466                  * not unlock the i_mutex at this case.
8467                  */
8468                 if (offset + count <= inode->i_size) {
8469                         mutex_unlock(&inode->i_mutex);
8470                         relock = true;
8471                 }
8472                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8473                 if (ret)
8474                         goto out;
8475                 dio_data.outstanding_extents = div64_u64(count +
8476                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8477                                                 BTRFS_MAX_EXTENT_SIZE);
8478
8479                 /*
8480                  * We need to know how many extents we reserved so that we can
8481                  * do the accounting properly if we go over the number we
8482                  * originally calculated.  Abuse current->journal_info for this.
8483                  */
8484                 dio_data.reserve = round_up(count, root->sectorsize);
8485                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8486                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8487                 current->journal_info = &dio_data;
8488         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8489                                      &BTRFS_I(inode)->runtime_flags)) {
8490                 inode_dio_end(inode);
8491                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8492                 wakeup = false;
8493         }
8494
8495         ret = __blockdev_direct_IO(iocb, inode,
8496                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8497                                    iter, offset, btrfs_get_blocks_direct, NULL,
8498                                    btrfs_submit_direct, flags);
8499         if (iov_iter_rw(iter) == WRITE) {
8500                 current->journal_info = NULL;
8501                 if (ret < 0 && ret != -EIOCBQUEUED) {
8502                         if (dio_data.reserve)
8503                                 btrfs_delalloc_release_space(inode, offset,
8504                                                              dio_data.reserve);
8505                         /*
8506                          * On error we might have left some ordered extents
8507                          * without submitting corresponding bios for them, so
8508                          * cleanup them up to avoid other tasks getting them
8509                          * and waiting for them to complete forever.
8510                          */
8511                         if (dio_data.unsubmitted_oe_range_start <
8512                             dio_data.unsubmitted_oe_range_end)
8513                                 btrfs_endio_direct_write_update_ordered(inode,
8514                                         dio_data.unsubmitted_oe_range_start,
8515                                         dio_data.unsubmitted_oe_range_end -
8516                                         dio_data.unsubmitted_oe_range_start,
8517                                         0);
8518                 } else if (ret >= 0 && (size_t)ret < count)
8519                         btrfs_delalloc_release_space(inode, offset,
8520                                                      count - (size_t)ret);
8521         }
8522 out:
8523         if (wakeup)
8524                 inode_dio_end(inode);
8525         if (relock)
8526                 mutex_lock(&inode->i_mutex);
8527
8528         return ret;
8529 }
8530
8531 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8532
8533 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8534                 __u64 start, __u64 len)
8535 {
8536         int     ret;
8537
8538         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8539         if (ret)
8540                 return ret;
8541
8542         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8543 }
8544
8545 int btrfs_readpage(struct file *file, struct page *page)
8546 {
8547         struct extent_io_tree *tree;
8548         tree = &BTRFS_I(page->mapping->host)->io_tree;
8549         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8550 }
8551
8552 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8553 {
8554         struct extent_io_tree *tree;
8555         struct inode *inode = page->mapping->host;
8556         int ret;
8557
8558         if (current->flags & PF_MEMALLOC) {
8559                 redirty_page_for_writepage(wbc, page);
8560                 unlock_page(page);
8561                 return 0;
8562         }
8563
8564         /*
8565          * If we are under memory pressure we will call this directly from the
8566          * VM, we need to make sure we have the inode referenced for the ordered
8567          * extent.  If not just return like we didn't do anything.
8568          */
8569         if (!igrab(inode)) {
8570                 redirty_page_for_writepage(wbc, page);
8571                 return AOP_WRITEPAGE_ACTIVATE;
8572         }
8573         tree = &BTRFS_I(page->mapping->host)->io_tree;
8574         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8575         btrfs_add_delayed_iput(inode);
8576         return ret;
8577 }
8578
8579 static int btrfs_writepages(struct address_space *mapping,
8580                             struct writeback_control *wbc)
8581 {
8582         struct extent_io_tree *tree;
8583
8584         tree = &BTRFS_I(mapping->host)->io_tree;
8585         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8586 }
8587
8588 static int
8589 btrfs_readpages(struct file *file, struct address_space *mapping,
8590                 struct list_head *pages, unsigned nr_pages)
8591 {
8592         struct extent_io_tree *tree;
8593         tree = &BTRFS_I(mapping->host)->io_tree;
8594         return extent_readpages(tree, mapping, pages, nr_pages,
8595                                 btrfs_get_extent);
8596 }
8597 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8598 {
8599         struct extent_io_tree *tree;
8600         struct extent_map_tree *map;
8601         int ret;
8602
8603         tree = &BTRFS_I(page->mapping->host)->io_tree;
8604         map = &BTRFS_I(page->mapping->host)->extent_tree;
8605         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8606         if (ret == 1) {
8607                 ClearPagePrivate(page);
8608                 set_page_private(page, 0);
8609                 page_cache_release(page);
8610         }
8611         return ret;
8612 }
8613
8614 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8615 {
8616         if (PageWriteback(page) || PageDirty(page))
8617                 return 0;
8618         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8619 }
8620
8621 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8622                                  unsigned int length)
8623 {
8624         struct inode *inode = page->mapping->host;
8625         struct extent_io_tree *tree;
8626         struct btrfs_ordered_extent *ordered;
8627         struct extent_state *cached_state = NULL;
8628         u64 page_start = page_offset(page);
8629         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8630         int inode_evicting = inode->i_state & I_FREEING;
8631
8632         /*
8633          * we have the page locked, so new writeback can't start,
8634          * and the dirty bit won't be cleared while we are here.
8635          *
8636          * Wait for IO on this page so that we can safely clear
8637          * the PagePrivate2 bit and do ordered accounting
8638          */
8639         wait_on_page_writeback(page);
8640
8641         tree = &BTRFS_I(inode)->io_tree;
8642         if (offset) {
8643                 btrfs_releasepage(page, GFP_NOFS);
8644                 return;
8645         }
8646
8647         if (!inode_evicting)
8648                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8649         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8650         if (ordered) {
8651                 /*
8652                  * IO on this page will never be started, so we need
8653                  * to account for any ordered extents now
8654                  */
8655                 if (!inode_evicting)
8656                         clear_extent_bit(tree, page_start, page_end,
8657                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8658                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8659                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8660                                          GFP_NOFS);
8661                 /*
8662                  * whoever cleared the private bit is responsible
8663                  * for the finish_ordered_io
8664                  */
8665                 if (TestClearPagePrivate2(page)) {
8666                         struct btrfs_ordered_inode_tree *tree;
8667                         u64 new_len;
8668
8669                         tree = &BTRFS_I(inode)->ordered_tree;
8670
8671                         spin_lock_irq(&tree->lock);
8672                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8673                         new_len = page_start - ordered->file_offset;
8674                         if (new_len < ordered->truncated_len)
8675                                 ordered->truncated_len = new_len;
8676                         spin_unlock_irq(&tree->lock);
8677
8678                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8679                                                            page_start,
8680                                                            PAGE_CACHE_SIZE, 1))
8681                                 btrfs_finish_ordered_io(ordered);
8682                 }
8683                 btrfs_put_ordered_extent(ordered);
8684                 if (!inode_evicting) {
8685                         cached_state = NULL;
8686                         lock_extent_bits(tree, page_start, page_end,
8687                                          &cached_state);
8688                 }
8689         }
8690
8691         /*
8692          * Qgroup reserved space handler
8693          * Page here will be either
8694          * 1) Already written to disk
8695          *    In this case, its reserved space is released from data rsv map
8696          *    and will be freed by delayed_ref handler finally.
8697          *    So even we call qgroup_free_data(), it won't decrease reserved
8698          *    space.
8699          * 2) Not written to disk
8700          *    This means the reserved space should be freed here.
8701          */
8702         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8703         if (!inode_evicting) {
8704                 clear_extent_bit(tree, page_start, page_end,
8705                                  EXTENT_LOCKED | EXTENT_DIRTY |
8706                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8707                                  EXTENT_DEFRAG, 1, 1,
8708                                  &cached_state, GFP_NOFS);
8709
8710                 __btrfs_releasepage(page, GFP_NOFS);
8711         }
8712
8713         ClearPageChecked(page);
8714         if (PagePrivate(page)) {
8715                 ClearPagePrivate(page);
8716                 set_page_private(page, 0);
8717                 page_cache_release(page);
8718         }
8719 }
8720
8721 /*
8722  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8723  * called from a page fault handler when a page is first dirtied. Hence we must
8724  * be careful to check for EOF conditions here. We set the page up correctly
8725  * for a written page which means we get ENOSPC checking when writing into
8726  * holes and correct delalloc and unwritten extent mapping on filesystems that
8727  * support these features.
8728  *
8729  * We are not allowed to take the i_mutex here so we have to play games to
8730  * protect against truncate races as the page could now be beyond EOF.  Because
8731  * vmtruncate() writes the inode size before removing pages, once we have the
8732  * page lock we can determine safely if the page is beyond EOF. If it is not
8733  * beyond EOF, then the page is guaranteed safe against truncation until we
8734  * unlock the page.
8735  */
8736 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8737 {
8738         struct page *page = vmf->page;
8739         struct inode *inode = file_inode(vma->vm_file);
8740         struct btrfs_root *root = BTRFS_I(inode)->root;
8741         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8742         struct btrfs_ordered_extent *ordered;
8743         struct extent_state *cached_state = NULL;
8744         char *kaddr;
8745         unsigned long zero_start;
8746         loff_t size;
8747         int ret;
8748         int reserved = 0;
8749         u64 page_start;
8750         u64 page_end;
8751
8752         sb_start_pagefault(inode->i_sb);
8753         page_start = page_offset(page);
8754         page_end = page_start + PAGE_CACHE_SIZE - 1;
8755
8756         ret = btrfs_delalloc_reserve_space(inode, page_start,
8757                                            PAGE_CACHE_SIZE);
8758         if (!ret) {
8759                 ret = file_update_time(vma->vm_file);
8760                 reserved = 1;
8761         }
8762         if (ret) {
8763                 if (ret == -ENOMEM)
8764                         ret = VM_FAULT_OOM;
8765                 else /* -ENOSPC, -EIO, etc */
8766                         ret = VM_FAULT_SIGBUS;
8767                 if (reserved)
8768                         goto out;
8769                 goto out_noreserve;
8770         }
8771
8772         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8773 again:
8774         lock_page(page);
8775         size = i_size_read(inode);
8776
8777         if ((page->mapping != inode->i_mapping) ||
8778             (page_start >= size)) {
8779                 /* page got truncated out from underneath us */
8780                 goto out_unlock;
8781         }
8782         wait_on_page_writeback(page);
8783
8784         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8785         set_page_extent_mapped(page);
8786
8787         /*
8788          * we can't set the delalloc bits if there are pending ordered
8789          * extents.  Drop our locks and wait for them to finish
8790          */
8791         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8792         if (ordered) {
8793                 unlock_extent_cached(io_tree, page_start, page_end,
8794                                      &cached_state, GFP_NOFS);
8795                 unlock_page(page);
8796                 btrfs_start_ordered_extent(inode, ordered, 1);
8797                 btrfs_put_ordered_extent(ordered);
8798                 goto again;
8799         }
8800
8801         /*
8802          * XXX - page_mkwrite gets called every time the page is dirtied, even
8803          * if it was already dirty, so for space accounting reasons we need to
8804          * clear any delalloc bits for the range we are fixing to save.  There
8805          * is probably a better way to do this, but for now keep consistent with
8806          * prepare_pages in the normal write path.
8807          */
8808         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8809                           EXTENT_DIRTY | EXTENT_DELALLOC |
8810                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8811                           0, 0, &cached_state, GFP_NOFS);
8812
8813         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8814                                         &cached_state);
8815         if (ret) {
8816                 unlock_extent_cached(io_tree, page_start, page_end,
8817                                      &cached_state, GFP_NOFS);
8818                 ret = VM_FAULT_SIGBUS;
8819                 goto out_unlock;
8820         }
8821         ret = 0;
8822
8823         /* page is wholly or partially inside EOF */
8824         if (page_start + PAGE_CACHE_SIZE > size)
8825                 zero_start = size & ~PAGE_CACHE_MASK;
8826         else
8827                 zero_start = PAGE_CACHE_SIZE;
8828
8829         if (zero_start != PAGE_CACHE_SIZE) {
8830                 kaddr = kmap(page);
8831                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8832                 flush_dcache_page(page);
8833                 kunmap(page);
8834         }
8835         ClearPageChecked(page);
8836         set_page_dirty(page);
8837         SetPageUptodate(page);
8838
8839         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8840         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8841         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8842
8843         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8844
8845 out_unlock:
8846         if (!ret) {
8847                 sb_end_pagefault(inode->i_sb);
8848                 return VM_FAULT_LOCKED;
8849         }
8850         unlock_page(page);
8851 out:
8852         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8853 out_noreserve:
8854         sb_end_pagefault(inode->i_sb);
8855         return ret;
8856 }
8857
8858 static int btrfs_truncate(struct inode *inode)
8859 {
8860         struct btrfs_root *root = BTRFS_I(inode)->root;
8861         struct btrfs_block_rsv *rsv;
8862         int ret = 0;
8863         int err = 0;
8864         struct btrfs_trans_handle *trans;
8865         u64 mask = root->sectorsize - 1;
8866         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8867
8868         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8869                                        (u64)-1);
8870         if (ret)
8871                 return ret;
8872
8873         /*
8874          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8875          * 3 things going on here
8876          *
8877          * 1) We need to reserve space for our orphan item and the space to
8878          * delete our orphan item.  Lord knows we don't want to have a dangling
8879          * orphan item because we didn't reserve space to remove it.
8880          *
8881          * 2) We need to reserve space to update our inode.
8882          *
8883          * 3) We need to have something to cache all the space that is going to
8884          * be free'd up by the truncate operation, but also have some slack
8885          * space reserved in case it uses space during the truncate (thank you
8886          * very much snapshotting).
8887          *
8888          * And we need these to all be seperate.  The fact is we can use alot of
8889          * space doing the truncate, and we have no earthly idea how much space
8890          * we will use, so we need the truncate reservation to be seperate so it
8891          * doesn't end up using space reserved for updating the inode or
8892          * removing the orphan item.  We also need to be able to stop the
8893          * transaction and start a new one, which means we need to be able to
8894          * update the inode several times, and we have no idea of knowing how
8895          * many times that will be, so we can't just reserve 1 item for the
8896          * entirety of the opration, so that has to be done seperately as well.
8897          * Then there is the orphan item, which does indeed need to be held on
8898          * to for the whole operation, and we need nobody to touch this reserved
8899          * space except the orphan code.
8900          *
8901          * So that leaves us with
8902          *
8903          * 1) root->orphan_block_rsv - for the orphan deletion.
8904          * 2) rsv - for the truncate reservation, which we will steal from the
8905          * transaction reservation.
8906          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8907          * updating the inode.
8908          */
8909         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8910         if (!rsv)
8911                 return -ENOMEM;
8912         rsv->size = min_size;
8913         rsv->failfast = 1;
8914
8915         /*
8916          * 1 for the truncate slack space
8917          * 1 for updating the inode.
8918          */
8919         trans = btrfs_start_transaction(root, 2);
8920         if (IS_ERR(trans)) {
8921                 err = PTR_ERR(trans);
8922                 goto out;
8923         }
8924
8925         /* Migrate the slack space for the truncate to our reserve */
8926         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8927                                       min_size);
8928         BUG_ON(ret);
8929
8930         /*
8931          * So if we truncate and then write and fsync we normally would just
8932          * write the extents that changed, which is a problem if we need to
8933          * first truncate that entire inode.  So set this flag so we write out
8934          * all of the extents in the inode to the sync log so we're completely
8935          * safe.
8936          */
8937         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8938         trans->block_rsv = rsv;
8939
8940         while (1) {
8941                 ret = btrfs_truncate_inode_items(trans, root, inode,
8942                                                  inode->i_size,
8943                                                  BTRFS_EXTENT_DATA_KEY);
8944                 if (ret != -ENOSPC && ret != -EAGAIN) {
8945                         err = ret;
8946                         break;
8947                 }
8948
8949                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8950                 ret = btrfs_update_inode(trans, root, inode);
8951                 if (ret) {
8952                         err = ret;
8953                         break;
8954                 }
8955
8956                 btrfs_end_transaction(trans, root);
8957                 btrfs_btree_balance_dirty(root);
8958
8959                 trans = btrfs_start_transaction(root, 2);
8960                 if (IS_ERR(trans)) {
8961                         ret = err = PTR_ERR(trans);
8962                         trans = NULL;
8963                         break;
8964                 }
8965
8966                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8967                                               rsv, min_size);
8968                 BUG_ON(ret);    /* shouldn't happen */
8969                 trans->block_rsv = rsv;
8970         }
8971
8972         if (ret == 0 && inode->i_nlink > 0) {
8973                 trans->block_rsv = root->orphan_block_rsv;
8974                 ret = btrfs_orphan_del(trans, inode);
8975                 if (ret)
8976                         err = ret;
8977         }
8978
8979         if (trans) {
8980                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8981                 ret = btrfs_update_inode(trans, root, inode);
8982                 if (ret && !err)
8983                         err = ret;
8984
8985                 ret = btrfs_end_transaction(trans, root);
8986                 btrfs_btree_balance_dirty(root);
8987         }
8988
8989 out:
8990         btrfs_free_block_rsv(root, rsv);
8991
8992         if (ret && !err)
8993                 err = ret;
8994
8995         return err;
8996 }
8997
8998 /*
8999  * create a new subvolume directory/inode (helper for the ioctl).
9000  */
9001 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9002                              struct btrfs_root *new_root,
9003                              struct btrfs_root *parent_root,
9004                              u64 new_dirid)
9005 {
9006         struct inode *inode;
9007         int err;
9008         u64 index = 0;
9009
9010         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9011                                 new_dirid, new_dirid,
9012                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9013                                 &index);
9014         if (IS_ERR(inode))
9015                 return PTR_ERR(inode);
9016         inode->i_op = &btrfs_dir_inode_operations;
9017         inode->i_fop = &btrfs_dir_file_operations;
9018
9019         set_nlink(inode, 1);
9020         btrfs_i_size_write(inode, 0);
9021         unlock_new_inode(inode);
9022
9023         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9024         if (err)
9025                 btrfs_err(new_root->fs_info,
9026                           "error inheriting subvolume %llu properties: %d",
9027                           new_root->root_key.objectid, err);
9028
9029         err = btrfs_update_inode(trans, new_root, inode);
9030
9031         iput(inode);
9032         return err;
9033 }
9034
9035 struct inode *btrfs_alloc_inode(struct super_block *sb)
9036 {
9037         struct btrfs_inode *ei;
9038         struct inode *inode;
9039
9040         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9041         if (!ei)
9042                 return NULL;
9043
9044         ei->root = NULL;
9045         ei->generation = 0;
9046         ei->last_trans = 0;
9047         ei->last_sub_trans = 0;
9048         ei->logged_trans = 0;
9049         ei->delalloc_bytes = 0;
9050         ei->defrag_bytes = 0;
9051         ei->disk_i_size = 0;
9052         ei->flags = 0;
9053         ei->csum_bytes = 0;
9054         ei->index_cnt = (u64)-1;
9055         ei->dir_index = 0;
9056         ei->last_unlink_trans = 0;
9057         ei->last_log_commit = 0;
9058         ei->delayed_iput_count = 0;
9059
9060         spin_lock_init(&ei->lock);
9061         ei->outstanding_extents = 0;
9062         ei->reserved_extents = 0;
9063
9064         ei->runtime_flags = 0;
9065         ei->force_compress = BTRFS_COMPRESS_NONE;
9066
9067         ei->delayed_node = NULL;
9068
9069         ei->i_otime.tv_sec = 0;
9070         ei->i_otime.tv_nsec = 0;
9071
9072         inode = &ei->vfs_inode;
9073         extent_map_tree_init(&ei->extent_tree);
9074         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9075         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9076         ei->io_tree.track_uptodate = 1;
9077         ei->io_failure_tree.track_uptodate = 1;
9078         atomic_set(&ei->sync_writers, 0);
9079         mutex_init(&ei->log_mutex);
9080         mutex_init(&ei->delalloc_mutex);
9081         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9082         INIT_LIST_HEAD(&ei->delalloc_inodes);
9083         INIT_LIST_HEAD(&ei->delayed_iput);
9084         RB_CLEAR_NODE(&ei->rb_node);
9085
9086         return inode;
9087 }
9088
9089 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9090 void btrfs_test_destroy_inode(struct inode *inode)
9091 {
9092         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9093         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9094 }
9095 #endif
9096
9097 static void btrfs_i_callback(struct rcu_head *head)
9098 {
9099         struct inode *inode = container_of(head, struct inode, i_rcu);
9100         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9101 }
9102
9103 void btrfs_destroy_inode(struct inode *inode)
9104 {
9105         struct btrfs_ordered_extent *ordered;
9106         struct btrfs_root *root = BTRFS_I(inode)->root;
9107
9108         WARN_ON(!hlist_empty(&inode->i_dentry));
9109         WARN_ON(inode->i_data.nrpages);
9110         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9111         WARN_ON(BTRFS_I(inode)->reserved_extents);
9112         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9113         WARN_ON(BTRFS_I(inode)->csum_bytes);
9114         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9115
9116         /*
9117          * This can happen where we create an inode, but somebody else also
9118          * created the same inode and we need to destroy the one we already
9119          * created.
9120          */
9121         if (!root)
9122                 goto free;
9123
9124         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9125                      &BTRFS_I(inode)->runtime_flags)) {
9126                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9127                         btrfs_ino(inode));
9128                 atomic_dec(&root->orphan_inodes);
9129         }
9130
9131         while (1) {
9132                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9133                 if (!ordered)
9134                         break;
9135                 else {
9136                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9137                                 ordered->file_offset, ordered->len);
9138                         btrfs_remove_ordered_extent(inode, ordered);
9139                         btrfs_put_ordered_extent(ordered);
9140                         btrfs_put_ordered_extent(ordered);
9141                 }
9142         }
9143         btrfs_qgroup_check_reserved_leak(inode);
9144         inode_tree_del(inode);
9145         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9146 free:
9147         call_rcu(&inode->i_rcu, btrfs_i_callback);
9148 }
9149
9150 int btrfs_drop_inode(struct inode *inode)
9151 {
9152         struct btrfs_root *root = BTRFS_I(inode)->root;
9153
9154         if (root == NULL)
9155                 return 1;
9156
9157         /* the snap/subvol tree is on deleting */
9158         if (btrfs_root_refs(&root->root_item) == 0)
9159                 return 1;
9160         else
9161                 return generic_drop_inode(inode);
9162 }
9163
9164 static void init_once(void *foo)
9165 {
9166         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9167
9168         inode_init_once(&ei->vfs_inode);
9169 }
9170
9171 void btrfs_destroy_cachep(void)
9172 {
9173         /*
9174          * Make sure all delayed rcu free inodes are flushed before we
9175          * destroy cache.
9176          */
9177         rcu_barrier();
9178         if (btrfs_inode_cachep)
9179                 kmem_cache_destroy(btrfs_inode_cachep);
9180         if (btrfs_trans_handle_cachep)
9181                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9182         if (btrfs_transaction_cachep)
9183                 kmem_cache_destroy(btrfs_transaction_cachep);
9184         if (btrfs_path_cachep)
9185                 kmem_cache_destroy(btrfs_path_cachep);
9186         if (btrfs_free_space_cachep)
9187                 kmem_cache_destroy(btrfs_free_space_cachep);
9188 }
9189
9190 int btrfs_init_cachep(void)
9191 {
9192         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9193                         sizeof(struct btrfs_inode), 0,
9194                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
9195         if (!btrfs_inode_cachep)
9196                 goto fail;
9197
9198         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9199                         sizeof(struct btrfs_trans_handle), 0,
9200                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9201         if (!btrfs_trans_handle_cachep)
9202                 goto fail;
9203
9204         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9205                         sizeof(struct btrfs_transaction), 0,
9206                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9207         if (!btrfs_transaction_cachep)
9208                 goto fail;
9209
9210         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9211                         sizeof(struct btrfs_path), 0,
9212                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9213         if (!btrfs_path_cachep)
9214                 goto fail;
9215
9216         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9217                         sizeof(struct btrfs_free_space), 0,
9218                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9219         if (!btrfs_free_space_cachep)
9220                 goto fail;
9221
9222         return 0;
9223 fail:
9224         btrfs_destroy_cachep();
9225         return -ENOMEM;
9226 }
9227
9228 static int btrfs_getattr(struct vfsmount *mnt,
9229                          struct dentry *dentry, struct kstat *stat)
9230 {
9231         u64 delalloc_bytes;
9232         struct inode *inode = d_inode(dentry);
9233         u32 blocksize = inode->i_sb->s_blocksize;
9234
9235         generic_fillattr(inode, stat);
9236         stat->dev = BTRFS_I(inode)->root->anon_dev;
9237         stat->blksize = PAGE_CACHE_SIZE;
9238
9239         spin_lock(&BTRFS_I(inode)->lock);
9240         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9241         spin_unlock(&BTRFS_I(inode)->lock);
9242         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9243                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9244         return 0;
9245 }
9246
9247 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9248                            struct inode *new_dir, struct dentry *new_dentry)
9249 {
9250         struct btrfs_trans_handle *trans;
9251         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9252         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9253         struct inode *new_inode = d_inode(new_dentry);
9254         struct inode *old_inode = d_inode(old_dentry);
9255         struct timespec ctime = CURRENT_TIME;
9256         u64 index = 0;
9257         u64 root_objectid;
9258         int ret;
9259         u64 old_ino = btrfs_ino(old_inode);
9260
9261         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9262                 return -EPERM;
9263
9264         /* we only allow rename subvolume link between subvolumes */
9265         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9266                 return -EXDEV;
9267
9268         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9269             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9270                 return -ENOTEMPTY;
9271
9272         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9273             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9274                 return -ENOTEMPTY;
9275
9276
9277         /* check for collisions, even if the  name isn't there */
9278         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9279                              new_dentry->d_name.name,
9280                              new_dentry->d_name.len);
9281
9282         if (ret) {
9283                 if (ret == -EEXIST) {
9284                         /* we shouldn't get
9285                          * eexist without a new_inode */
9286                         if (WARN_ON(!new_inode)) {
9287                                 return ret;
9288                         }
9289                 } else {
9290                         /* maybe -EOVERFLOW */
9291                         return ret;
9292                 }
9293         }
9294         ret = 0;
9295
9296         /*
9297          * we're using rename to replace one file with another.  Start IO on it
9298          * now so  we don't add too much work to the end of the transaction
9299          */
9300         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9301                 filemap_flush(old_inode->i_mapping);
9302
9303         /* close the racy window with snapshot create/destroy ioctl */
9304         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9305                 down_read(&root->fs_info->subvol_sem);
9306         /*
9307          * We want to reserve the absolute worst case amount of items.  So if
9308          * both inodes are subvols and we need to unlink them then that would
9309          * require 4 item modifications, but if they are both normal inodes it
9310          * would require 5 item modifications, so we'll assume their normal
9311          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9312          * should cover the worst case number of items we'll modify.
9313          */
9314         trans = btrfs_start_transaction(root, 11);
9315         if (IS_ERR(trans)) {
9316                 ret = PTR_ERR(trans);
9317                 goto out_notrans;
9318         }
9319
9320         if (dest != root)
9321                 btrfs_record_root_in_trans(trans, dest);
9322
9323         ret = btrfs_set_inode_index(new_dir, &index);
9324         if (ret)
9325                 goto out_fail;
9326
9327         BTRFS_I(old_inode)->dir_index = 0ULL;
9328         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9329                 /* force full log commit if subvolume involved. */
9330                 btrfs_set_log_full_commit(root->fs_info, trans);
9331         } else {
9332                 ret = btrfs_insert_inode_ref(trans, dest,
9333                                              new_dentry->d_name.name,
9334                                              new_dentry->d_name.len,
9335                                              old_ino,
9336                                              btrfs_ino(new_dir), index);
9337                 if (ret)
9338                         goto out_fail;
9339                 /*
9340                  * this is an ugly little race, but the rename is required
9341                  * to make sure that if we crash, the inode is either at the
9342                  * old name or the new one.  pinning the log transaction lets
9343                  * us make sure we don't allow a log commit to come in after
9344                  * we unlink the name but before we add the new name back in.
9345                  */
9346                 btrfs_pin_log_trans(root);
9347         }
9348
9349         inode_inc_iversion(old_dir);
9350         inode_inc_iversion(new_dir);
9351         inode_inc_iversion(old_inode);
9352         old_dir->i_ctime = old_dir->i_mtime = ctime;
9353         new_dir->i_ctime = new_dir->i_mtime = ctime;
9354         old_inode->i_ctime = ctime;
9355
9356         if (old_dentry->d_parent != new_dentry->d_parent)
9357                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9358
9359         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9360                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9361                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9362                                         old_dentry->d_name.name,
9363                                         old_dentry->d_name.len);
9364         } else {
9365                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9366                                         d_inode(old_dentry),
9367                                         old_dentry->d_name.name,
9368                                         old_dentry->d_name.len);
9369                 if (!ret)
9370                         ret = btrfs_update_inode(trans, root, old_inode);
9371         }
9372         if (ret) {
9373                 btrfs_abort_transaction(trans, root, ret);
9374                 goto out_fail;
9375         }
9376
9377         if (new_inode) {
9378                 inode_inc_iversion(new_inode);
9379                 new_inode->i_ctime = CURRENT_TIME;
9380                 if (unlikely(btrfs_ino(new_inode) ==
9381                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9382                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9383                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9384                                                 root_objectid,
9385                                                 new_dentry->d_name.name,
9386                                                 new_dentry->d_name.len);
9387                         BUG_ON(new_inode->i_nlink == 0);
9388                 } else {
9389                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9390                                                  d_inode(new_dentry),
9391                                                  new_dentry->d_name.name,
9392                                                  new_dentry->d_name.len);
9393                 }
9394                 if (!ret && new_inode->i_nlink == 0)
9395                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9396                 if (ret) {
9397                         btrfs_abort_transaction(trans, root, ret);
9398                         goto out_fail;
9399                 }
9400         }
9401
9402         ret = btrfs_add_link(trans, new_dir, old_inode,
9403                              new_dentry->d_name.name,
9404                              new_dentry->d_name.len, 0, index);
9405         if (ret) {
9406                 btrfs_abort_transaction(trans, root, ret);
9407                 goto out_fail;
9408         }
9409
9410         if (old_inode->i_nlink == 1)
9411                 BTRFS_I(old_inode)->dir_index = index;
9412
9413         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9414                 struct dentry *parent = new_dentry->d_parent;
9415                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9416                 btrfs_end_log_trans(root);
9417         }
9418 out_fail:
9419         btrfs_end_transaction(trans, root);
9420 out_notrans:
9421         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9422                 up_read(&root->fs_info->subvol_sem);
9423
9424         return ret;
9425 }
9426
9427 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9428                          struct inode *new_dir, struct dentry *new_dentry,
9429                          unsigned int flags)
9430 {
9431         if (flags & ~RENAME_NOREPLACE)
9432                 return -EINVAL;
9433
9434         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9435 }
9436
9437 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9438 {
9439         struct btrfs_delalloc_work *delalloc_work;
9440         struct inode *inode;
9441
9442         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9443                                      work);
9444         inode = delalloc_work->inode;
9445         filemap_flush(inode->i_mapping);
9446         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9447                                 &BTRFS_I(inode)->runtime_flags))
9448                 filemap_flush(inode->i_mapping);
9449
9450         if (delalloc_work->delay_iput)
9451                 btrfs_add_delayed_iput(inode);
9452         else
9453                 iput(inode);
9454         complete(&delalloc_work->completion);
9455 }
9456
9457 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9458                                                     int delay_iput)
9459 {
9460         struct btrfs_delalloc_work *work;
9461
9462         work = kmalloc(sizeof(*work), GFP_NOFS);
9463         if (!work)
9464                 return NULL;
9465
9466         init_completion(&work->completion);
9467         INIT_LIST_HEAD(&work->list);
9468         work->inode = inode;
9469         work->delay_iput = delay_iput;
9470         WARN_ON_ONCE(!inode);
9471         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9472                         btrfs_run_delalloc_work, NULL, NULL);
9473
9474         return work;
9475 }
9476
9477 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9478 {
9479         wait_for_completion(&work->completion);
9480         kfree(work);
9481 }
9482
9483 /*
9484  * some fairly slow code that needs optimization. This walks the list
9485  * of all the inodes with pending delalloc and forces them to disk.
9486  */
9487 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9488                                    int nr)
9489 {
9490         struct btrfs_inode *binode;
9491         struct inode *inode;
9492         struct btrfs_delalloc_work *work, *next;
9493         struct list_head works;
9494         struct list_head splice;
9495         int ret = 0;
9496
9497         INIT_LIST_HEAD(&works);
9498         INIT_LIST_HEAD(&splice);
9499
9500         mutex_lock(&root->delalloc_mutex);
9501         spin_lock(&root->delalloc_lock);
9502         list_splice_init(&root->delalloc_inodes, &splice);
9503         while (!list_empty(&splice)) {
9504                 binode = list_entry(splice.next, struct btrfs_inode,
9505                                     delalloc_inodes);
9506
9507                 list_move_tail(&binode->delalloc_inodes,
9508                                &root->delalloc_inodes);
9509                 inode = igrab(&binode->vfs_inode);
9510                 if (!inode) {
9511                         cond_resched_lock(&root->delalloc_lock);
9512                         continue;
9513                 }
9514                 spin_unlock(&root->delalloc_lock);
9515
9516                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
9517                 if (!work) {
9518                         if (delay_iput)
9519                                 btrfs_add_delayed_iput(inode);
9520                         else
9521                                 iput(inode);
9522                         ret = -ENOMEM;
9523                         goto out;
9524                 }
9525                 list_add_tail(&work->list, &works);
9526                 btrfs_queue_work(root->fs_info->flush_workers,
9527                                  &work->work);
9528                 ret++;
9529                 if (nr != -1 && ret >= nr)
9530                         goto out;
9531                 cond_resched();
9532                 spin_lock(&root->delalloc_lock);
9533         }
9534         spin_unlock(&root->delalloc_lock);
9535
9536 out:
9537         list_for_each_entry_safe(work, next, &works, list) {
9538                 list_del_init(&work->list);
9539                 btrfs_wait_and_free_delalloc_work(work);
9540         }
9541
9542         if (!list_empty_careful(&splice)) {
9543                 spin_lock(&root->delalloc_lock);
9544                 list_splice_tail(&splice, &root->delalloc_inodes);
9545                 spin_unlock(&root->delalloc_lock);
9546         }
9547         mutex_unlock(&root->delalloc_mutex);
9548         return ret;
9549 }
9550
9551 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9552 {
9553         int ret;
9554
9555         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9556                 return -EROFS;
9557
9558         ret = __start_delalloc_inodes(root, delay_iput, -1);
9559         if (ret > 0)
9560                 ret = 0;
9561         /*
9562          * the filemap_flush will queue IO into the worker threads, but
9563          * we have to make sure the IO is actually started and that
9564          * ordered extents get created before we return
9565          */
9566         atomic_inc(&root->fs_info->async_submit_draining);
9567         while (atomic_read(&root->fs_info->nr_async_submits) ||
9568               atomic_read(&root->fs_info->async_delalloc_pages)) {
9569                 wait_event(root->fs_info->async_submit_wait,
9570                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9571                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9572         }
9573         atomic_dec(&root->fs_info->async_submit_draining);
9574         return ret;
9575 }
9576
9577 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9578                                int nr)
9579 {
9580         struct btrfs_root *root;
9581         struct list_head splice;
9582         int ret;
9583
9584         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9585                 return -EROFS;
9586
9587         INIT_LIST_HEAD(&splice);
9588
9589         mutex_lock(&fs_info->delalloc_root_mutex);
9590         spin_lock(&fs_info->delalloc_root_lock);
9591         list_splice_init(&fs_info->delalloc_roots, &splice);
9592         while (!list_empty(&splice) && nr) {
9593                 root = list_first_entry(&splice, struct btrfs_root,
9594                                         delalloc_root);
9595                 root = btrfs_grab_fs_root(root);
9596                 BUG_ON(!root);
9597                 list_move_tail(&root->delalloc_root,
9598                                &fs_info->delalloc_roots);
9599                 spin_unlock(&fs_info->delalloc_root_lock);
9600
9601                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9602                 btrfs_put_fs_root(root);
9603                 if (ret < 0)
9604                         goto out;
9605
9606                 if (nr != -1) {
9607                         nr -= ret;
9608                         WARN_ON(nr < 0);
9609                 }
9610                 spin_lock(&fs_info->delalloc_root_lock);
9611         }
9612         spin_unlock(&fs_info->delalloc_root_lock);
9613
9614         ret = 0;
9615         atomic_inc(&fs_info->async_submit_draining);
9616         while (atomic_read(&fs_info->nr_async_submits) ||
9617               atomic_read(&fs_info->async_delalloc_pages)) {
9618                 wait_event(fs_info->async_submit_wait,
9619                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9620                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9621         }
9622         atomic_dec(&fs_info->async_submit_draining);
9623 out:
9624         if (!list_empty_careful(&splice)) {
9625                 spin_lock(&fs_info->delalloc_root_lock);
9626                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9627                 spin_unlock(&fs_info->delalloc_root_lock);
9628         }
9629         mutex_unlock(&fs_info->delalloc_root_mutex);
9630         return ret;
9631 }
9632
9633 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9634                          const char *symname)
9635 {
9636         struct btrfs_trans_handle *trans;
9637         struct btrfs_root *root = BTRFS_I(dir)->root;
9638         struct btrfs_path *path;
9639         struct btrfs_key key;
9640         struct inode *inode = NULL;
9641         int err;
9642         int drop_inode = 0;
9643         u64 objectid;
9644         u64 index = 0;
9645         int name_len;
9646         int datasize;
9647         unsigned long ptr;
9648         struct btrfs_file_extent_item *ei;
9649         struct extent_buffer *leaf;
9650
9651         name_len = strlen(symname);
9652         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9653                 return -ENAMETOOLONG;
9654
9655         /*
9656          * 2 items for inode item and ref
9657          * 2 items for dir items
9658          * 1 item for updating parent inode item
9659          * 1 item for the inline extent item
9660          * 1 item for xattr if selinux is on
9661          */
9662         trans = btrfs_start_transaction(root, 7);
9663         if (IS_ERR(trans))
9664                 return PTR_ERR(trans);
9665
9666         err = btrfs_find_free_ino(root, &objectid);
9667         if (err)
9668                 goto out_unlock;
9669
9670         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9671                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9672                                 S_IFLNK|S_IRWXUGO, &index);
9673         if (IS_ERR(inode)) {
9674                 err = PTR_ERR(inode);
9675                 goto out_unlock;
9676         }
9677
9678         /*
9679         * If the active LSM wants to access the inode during
9680         * d_instantiate it needs these. Smack checks to see
9681         * if the filesystem supports xattrs by looking at the
9682         * ops vector.
9683         */
9684         inode->i_fop = &btrfs_file_operations;
9685         inode->i_op = &btrfs_file_inode_operations;
9686         inode->i_mapping->a_ops = &btrfs_aops;
9687         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9688
9689         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9690         if (err)
9691                 goto out_unlock_inode;
9692
9693         path = btrfs_alloc_path();
9694         if (!path) {
9695                 err = -ENOMEM;
9696                 goto out_unlock_inode;
9697         }
9698         key.objectid = btrfs_ino(inode);
9699         key.offset = 0;
9700         key.type = BTRFS_EXTENT_DATA_KEY;
9701         datasize = btrfs_file_extent_calc_inline_size(name_len);
9702         err = btrfs_insert_empty_item(trans, root, path, &key,
9703                                       datasize);
9704         if (err) {
9705                 btrfs_free_path(path);
9706                 goto out_unlock_inode;
9707         }
9708         leaf = path->nodes[0];
9709         ei = btrfs_item_ptr(leaf, path->slots[0],
9710                             struct btrfs_file_extent_item);
9711         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9712         btrfs_set_file_extent_type(leaf, ei,
9713                                    BTRFS_FILE_EXTENT_INLINE);
9714         btrfs_set_file_extent_encryption(leaf, ei, 0);
9715         btrfs_set_file_extent_compression(leaf, ei, 0);
9716         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9717         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9718
9719         ptr = btrfs_file_extent_inline_start(ei);
9720         write_extent_buffer(leaf, symname, ptr, name_len);
9721         btrfs_mark_buffer_dirty(leaf);
9722         btrfs_free_path(path);
9723
9724         inode->i_op = &btrfs_symlink_inode_operations;
9725         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9726         inode_set_bytes(inode, name_len);
9727         btrfs_i_size_write(inode, name_len);
9728         err = btrfs_update_inode(trans, root, inode);
9729         /*
9730          * Last step, add directory indexes for our symlink inode. This is the
9731          * last step to avoid extra cleanup of these indexes if an error happens
9732          * elsewhere above.
9733          */
9734         if (!err)
9735                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9736         if (err) {
9737                 drop_inode = 1;
9738                 goto out_unlock_inode;
9739         }
9740
9741         unlock_new_inode(inode);
9742         d_instantiate(dentry, inode);
9743
9744 out_unlock:
9745         btrfs_end_transaction(trans, root);
9746         if (drop_inode) {
9747                 inode_dec_link_count(inode);
9748                 iput(inode);
9749         }
9750         btrfs_btree_balance_dirty(root);
9751         return err;
9752
9753 out_unlock_inode:
9754         drop_inode = 1;
9755         unlock_new_inode(inode);
9756         goto out_unlock;
9757 }
9758
9759 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9760                                        u64 start, u64 num_bytes, u64 min_size,
9761                                        loff_t actual_len, u64 *alloc_hint,
9762                                        struct btrfs_trans_handle *trans)
9763 {
9764         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9765         struct extent_map *em;
9766         struct btrfs_root *root = BTRFS_I(inode)->root;
9767         struct btrfs_key ins;
9768         u64 cur_offset = start;
9769         u64 i_size;
9770         u64 cur_bytes;
9771         u64 last_alloc = (u64)-1;
9772         int ret = 0;
9773         bool own_trans = true;
9774
9775         if (trans)
9776                 own_trans = false;
9777         while (num_bytes > 0) {
9778                 if (own_trans) {
9779                         trans = btrfs_start_transaction(root, 3);
9780                         if (IS_ERR(trans)) {
9781                                 ret = PTR_ERR(trans);
9782                                 break;
9783                         }
9784                 }
9785
9786                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9787                 cur_bytes = max(cur_bytes, min_size);
9788                 /*
9789                  * If we are severely fragmented we could end up with really
9790                  * small allocations, so if the allocator is returning small
9791                  * chunks lets make its job easier by only searching for those
9792                  * sized chunks.
9793                  */
9794                 cur_bytes = min(cur_bytes, last_alloc);
9795                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9796                                            *alloc_hint, &ins, 1, 0);
9797                 if (ret) {
9798                         if (own_trans)
9799                                 btrfs_end_transaction(trans, root);
9800                         break;
9801                 }
9802
9803                 last_alloc = ins.offset;
9804                 ret = insert_reserved_file_extent(trans, inode,
9805                                                   cur_offset, ins.objectid,
9806                                                   ins.offset, ins.offset,
9807                                                   ins.offset, 0, 0, 0,
9808                                                   BTRFS_FILE_EXTENT_PREALLOC);
9809                 if (ret) {
9810                         btrfs_free_reserved_extent(root, ins.objectid,
9811                                                    ins.offset, 0);
9812                         btrfs_abort_transaction(trans, root, ret);
9813                         if (own_trans)
9814                                 btrfs_end_transaction(trans, root);
9815                         break;
9816                 }
9817
9818                 btrfs_drop_extent_cache(inode, cur_offset,
9819                                         cur_offset + ins.offset -1, 0);
9820
9821                 em = alloc_extent_map();
9822                 if (!em) {
9823                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9824                                 &BTRFS_I(inode)->runtime_flags);
9825                         goto next;
9826                 }
9827
9828                 em->start = cur_offset;
9829                 em->orig_start = cur_offset;
9830                 em->len = ins.offset;
9831                 em->block_start = ins.objectid;
9832                 em->block_len = ins.offset;
9833                 em->orig_block_len = ins.offset;
9834                 em->ram_bytes = ins.offset;
9835                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9836                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9837                 em->generation = trans->transid;
9838
9839                 while (1) {
9840                         write_lock(&em_tree->lock);
9841                         ret = add_extent_mapping(em_tree, em, 1);
9842                         write_unlock(&em_tree->lock);
9843                         if (ret != -EEXIST)
9844                                 break;
9845                         btrfs_drop_extent_cache(inode, cur_offset,
9846                                                 cur_offset + ins.offset - 1,
9847                                                 0);
9848                 }
9849                 free_extent_map(em);
9850 next:
9851                 num_bytes -= ins.offset;
9852                 cur_offset += ins.offset;
9853                 *alloc_hint = ins.objectid + ins.offset;
9854
9855                 inode_inc_iversion(inode);
9856                 inode->i_ctime = CURRENT_TIME;
9857                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9858                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9859                     (actual_len > inode->i_size) &&
9860                     (cur_offset > inode->i_size)) {
9861                         if (cur_offset > actual_len)
9862                                 i_size = actual_len;
9863                         else
9864                                 i_size = cur_offset;
9865                         i_size_write(inode, i_size);
9866                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9867                 }
9868
9869                 ret = btrfs_update_inode(trans, root, inode);
9870
9871                 if (ret) {
9872                         btrfs_abort_transaction(trans, root, ret);
9873                         if (own_trans)
9874                                 btrfs_end_transaction(trans, root);
9875                         break;
9876                 }
9877
9878                 if (own_trans)
9879                         btrfs_end_transaction(trans, root);
9880         }
9881         return ret;
9882 }
9883
9884 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9885                               u64 start, u64 num_bytes, u64 min_size,
9886                               loff_t actual_len, u64 *alloc_hint)
9887 {
9888         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9889                                            min_size, actual_len, alloc_hint,
9890                                            NULL);
9891 }
9892
9893 int btrfs_prealloc_file_range_trans(struct inode *inode,
9894                                     struct btrfs_trans_handle *trans, int mode,
9895                                     u64 start, u64 num_bytes, u64 min_size,
9896                                     loff_t actual_len, u64 *alloc_hint)
9897 {
9898         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9899                                            min_size, actual_len, alloc_hint, trans);
9900 }
9901
9902 static int btrfs_set_page_dirty(struct page *page)
9903 {
9904         return __set_page_dirty_nobuffers(page);
9905 }
9906
9907 static int btrfs_permission(struct inode *inode, int mask)
9908 {
9909         struct btrfs_root *root = BTRFS_I(inode)->root;
9910         umode_t mode = inode->i_mode;
9911
9912         if (mask & MAY_WRITE &&
9913             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9914                 if (btrfs_root_readonly(root))
9915                         return -EROFS;
9916                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9917                         return -EACCES;
9918         }
9919         return generic_permission(inode, mask);
9920 }
9921
9922 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9923 {
9924         struct btrfs_trans_handle *trans;
9925         struct btrfs_root *root = BTRFS_I(dir)->root;
9926         struct inode *inode = NULL;
9927         u64 objectid;
9928         u64 index;
9929         int ret = 0;
9930
9931         /*
9932          * 5 units required for adding orphan entry
9933          */
9934         trans = btrfs_start_transaction(root, 5);
9935         if (IS_ERR(trans))
9936                 return PTR_ERR(trans);
9937
9938         ret = btrfs_find_free_ino(root, &objectid);
9939         if (ret)
9940                 goto out;
9941
9942         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9943                                 btrfs_ino(dir), objectid, mode, &index);
9944         if (IS_ERR(inode)) {
9945                 ret = PTR_ERR(inode);
9946                 inode = NULL;
9947                 goto out;
9948         }
9949
9950         inode->i_fop = &btrfs_file_operations;
9951         inode->i_op = &btrfs_file_inode_operations;
9952
9953         inode->i_mapping->a_ops = &btrfs_aops;
9954         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9955
9956         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9957         if (ret)
9958                 goto out_inode;
9959
9960         ret = btrfs_update_inode(trans, root, inode);
9961         if (ret)
9962                 goto out_inode;
9963         ret = btrfs_orphan_add(trans, inode);
9964         if (ret)
9965                 goto out_inode;
9966
9967         /*
9968          * We set number of links to 0 in btrfs_new_inode(), and here we set
9969          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9970          * through:
9971          *
9972          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9973          */
9974         set_nlink(inode, 1);
9975         unlock_new_inode(inode);
9976         d_tmpfile(dentry, inode);
9977         mark_inode_dirty(inode);
9978
9979 out:
9980         btrfs_end_transaction(trans, root);
9981         if (ret)
9982                 iput(inode);
9983         btrfs_balance_delayed_items(root);
9984         btrfs_btree_balance_dirty(root);
9985         return ret;
9986
9987 out_inode:
9988         unlock_new_inode(inode);
9989         goto out;
9990
9991 }
9992
9993 /* Inspired by filemap_check_errors() */
9994 int btrfs_inode_check_errors(struct inode *inode)
9995 {
9996         int ret = 0;
9997
9998         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
9999             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10000                 ret = -ENOSPC;
10001         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10002             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10003                 ret = -EIO;
10004
10005         return ret;
10006 }
10007
10008 static const struct inode_operations btrfs_dir_inode_operations = {
10009         .getattr        = btrfs_getattr,
10010         .lookup         = btrfs_lookup,
10011         .create         = btrfs_create,
10012         .unlink         = btrfs_unlink,
10013         .link           = btrfs_link,
10014         .mkdir          = btrfs_mkdir,
10015         .rmdir          = btrfs_rmdir,
10016         .rename2        = btrfs_rename2,
10017         .symlink        = btrfs_symlink,
10018         .setattr        = btrfs_setattr,
10019         .mknod          = btrfs_mknod,
10020         .setxattr       = btrfs_setxattr,
10021         .getxattr       = btrfs_getxattr,
10022         .listxattr      = btrfs_listxattr,
10023         .removexattr    = btrfs_removexattr,
10024         .permission     = btrfs_permission,
10025         .get_acl        = btrfs_get_acl,
10026         .set_acl        = btrfs_set_acl,
10027         .update_time    = btrfs_update_time,
10028         .tmpfile        = btrfs_tmpfile,
10029 };
10030 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10031         .lookup         = btrfs_lookup,
10032         .permission     = btrfs_permission,
10033         .get_acl        = btrfs_get_acl,
10034         .set_acl        = btrfs_set_acl,
10035         .update_time    = btrfs_update_time,
10036 };
10037
10038 static const struct file_operations btrfs_dir_file_operations = {
10039         .llseek         = generic_file_llseek,
10040         .read           = generic_read_dir,
10041         .iterate        = btrfs_real_readdir,
10042         .unlocked_ioctl = btrfs_ioctl,
10043 #ifdef CONFIG_COMPAT
10044         .compat_ioctl   = btrfs_ioctl,
10045 #endif
10046         .release        = btrfs_release_file,
10047         .fsync          = btrfs_sync_file,
10048 };
10049
10050 static const struct extent_io_ops btrfs_extent_io_ops = {
10051         .fill_delalloc = run_delalloc_range,
10052         .submit_bio_hook = btrfs_submit_bio_hook,
10053         .merge_bio_hook = btrfs_merge_bio_hook,
10054         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10055         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10056         .writepage_start_hook = btrfs_writepage_start_hook,
10057         .set_bit_hook = btrfs_set_bit_hook,
10058         .clear_bit_hook = btrfs_clear_bit_hook,
10059         .merge_extent_hook = btrfs_merge_extent_hook,
10060         .split_extent_hook = btrfs_split_extent_hook,
10061 };
10062
10063 /*
10064  * btrfs doesn't support the bmap operation because swapfiles
10065  * use bmap to make a mapping of extents in the file.  They assume
10066  * these extents won't change over the life of the file and they
10067  * use the bmap result to do IO directly to the drive.
10068  *
10069  * the btrfs bmap call would return logical addresses that aren't
10070  * suitable for IO and they also will change frequently as COW
10071  * operations happen.  So, swapfile + btrfs == corruption.
10072  *
10073  * For now we're avoiding this by dropping bmap.
10074  */
10075 static const struct address_space_operations btrfs_aops = {
10076         .readpage       = btrfs_readpage,
10077         .writepage      = btrfs_writepage,
10078         .writepages     = btrfs_writepages,
10079         .readpages      = btrfs_readpages,
10080         .direct_IO      = btrfs_direct_IO,
10081         .invalidatepage = btrfs_invalidatepage,
10082         .releasepage    = btrfs_releasepage,
10083         .set_page_dirty = btrfs_set_page_dirty,
10084         .error_remove_page = generic_error_remove_page,
10085 };
10086
10087 static const struct address_space_operations btrfs_symlink_aops = {
10088         .readpage       = btrfs_readpage,
10089         .writepage      = btrfs_writepage,
10090         .invalidatepage = btrfs_invalidatepage,
10091         .releasepage    = btrfs_releasepage,
10092 };
10093
10094 static const struct inode_operations btrfs_file_inode_operations = {
10095         .getattr        = btrfs_getattr,
10096         .setattr        = btrfs_setattr,
10097         .setxattr       = btrfs_setxattr,
10098         .getxattr       = btrfs_getxattr,
10099         .listxattr      = btrfs_listxattr,
10100         .removexattr    = btrfs_removexattr,
10101         .permission     = btrfs_permission,
10102         .fiemap         = btrfs_fiemap,
10103         .get_acl        = btrfs_get_acl,
10104         .set_acl        = btrfs_set_acl,
10105         .update_time    = btrfs_update_time,
10106 };
10107 static const struct inode_operations btrfs_special_inode_operations = {
10108         .getattr        = btrfs_getattr,
10109         .setattr        = btrfs_setattr,
10110         .permission     = btrfs_permission,
10111         .setxattr       = btrfs_setxattr,
10112         .getxattr       = btrfs_getxattr,
10113         .listxattr      = btrfs_listxattr,
10114         .removexattr    = btrfs_removexattr,
10115         .get_acl        = btrfs_get_acl,
10116         .set_acl        = btrfs_set_acl,
10117         .update_time    = btrfs_update_time,
10118 };
10119 static const struct inode_operations btrfs_symlink_inode_operations = {
10120         .readlink       = generic_readlink,
10121         .follow_link    = page_follow_link_light,
10122         .put_link       = page_put_link,
10123         .getattr        = btrfs_getattr,
10124         .setattr        = btrfs_setattr,
10125         .permission     = btrfs_permission,
10126         .setxattr       = btrfs_setxattr,
10127         .getxattr       = btrfs_getxattr,
10128         .listxattr      = btrfs_listxattr,
10129         .removexattr    = btrfs_removexattr,
10130         .update_time    = btrfs_update_time,
10131 };
10132
10133 const struct dentry_operations btrfs_dentry_operations = {
10134         .d_delete       = btrfs_dentry_delete,
10135         .d_release      = btrfs_dentry_release,
10136 };