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         down_read(&fs_info->delayed_iput_sem);
3138         spin_lock(&fs_info->delayed_iput_lock);
3139         while (!list_empty(&fs_info->delayed_iputs)) {
3140                 struct btrfs_inode *inode;
3141
3142                 inode = list_first_entry(&fs_info->delayed_iputs,
3143                                 struct btrfs_inode, delayed_iput);
3144                 if (inode->delayed_iput_count) {
3145                         inode->delayed_iput_count--;
3146                         list_move_tail(&inode->delayed_iput,
3147                                         &fs_info->delayed_iputs);
3148                 } else {
3149                         list_del_init(&inode->delayed_iput);
3150                 }
3151                 spin_unlock(&fs_info->delayed_iput_lock);
3152                 iput(&inode->vfs_inode);
3153                 spin_lock(&fs_info->delayed_iput_lock);
3154         }
3155         spin_unlock(&fs_info->delayed_iput_lock);
3156         up_read(&root->fs_info->delayed_iput_sem);
3157 }
3158
3159 /*
3160  * This is called in transaction commit time. If there are no orphan
3161  * files in the subvolume, it removes orphan item and frees block_rsv
3162  * structure.
3163  */
3164 void btrfs_orphan_commit_root(struct btrfs_trans_handle *trans,
3165                               struct btrfs_root *root)
3166 {
3167         struct btrfs_block_rsv *block_rsv;
3168         int ret;
3169
3170         if (atomic_read(&root->orphan_inodes) ||
3171             root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE)
3172                 return;
3173
3174         spin_lock(&root->orphan_lock);
3175         if (atomic_read(&root->orphan_inodes)) {
3176                 spin_unlock(&root->orphan_lock);
3177                 return;
3178         }
3179
3180         if (root->orphan_cleanup_state != ORPHAN_CLEANUP_DONE) {
3181                 spin_unlock(&root->orphan_lock);
3182                 return;
3183         }
3184
3185         block_rsv = root->orphan_block_rsv;
3186         root->orphan_block_rsv = NULL;
3187         spin_unlock(&root->orphan_lock);
3188
3189         if (test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state) &&
3190             btrfs_root_refs(&root->root_item) > 0) {
3191                 ret = btrfs_del_orphan_item(trans, root->fs_info->tree_root,
3192                                             root->root_key.objectid);
3193                 if (ret)
3194                         btrfs_abort_transaction(trans, root, ret);
3195                 else
3196                         clear_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED,
3197                                   &root->state);
3198         }
3199
3200         if (block_rsv) {
3201                 WARN_ON(block_rsv->size > 0);
3202                 btrfs_free_block_rsv(root, block_rsv);
3203         }
3204 }
3205
3206 /*
3207  * This creates an orphan entry for the given inode in case something goes
3208  * wrong in the middle of an unlink/truncate.
3209  *
3210  * NOTE: caller of this function should reserve 5 units of metadata for
3211  *       this function.
3212  */
3213 int btrfs_orphan_add(struct btrfs_trans_handle *trans, struct inode *inode)
3214 {
3215         struct btrfs_root *root = BTRFS_I(inode)->root;
3216         struct btrfs_block_rsv *block_rsv = NULL;
3217         int reserve = 0;
3218         int insert = 0;
3219         int ret;
3220
3221         if (!root->orphan_block_rsv) {
3222                 block_rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
3223                 if (!block_rsv)
3224                         return -ENOMEM;
3225         }
3226
3227         spin_lock(&root->orphan_lock);
3228         if (!root->orphan_block_rsv) {
3229                 root->orphan_block_rsv = block_rsv;
3230         } else if (block_rsv) {
3231                 btrfs_free_block_rsv(root, block_rsv);
3232                 block_rsv = NULL;
3233         }
3234
3235         if (!test_and_set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3236                               &BTRFS_I(inode)->runtime_flags)) {
3237 #if 0
3238                 /*
3239                  * For proper ENOSPC handling, we should do orphan
3240                  * cleanup when mounting. But this introduces backward
3241                  * compatibility issue.
3242                  */
3243                 if (!xchg(&root->orphan_item_inserted, 1))
3244                         insert = 2;
3245                 else
3246                         insert = 1;
3247 #endif
3248                 insert = 1;
3249                 atomic_inc(&root->orphan_inodes);
3250         }
3251
3252         if (!test_and_set_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3253                               &BTRFS_I(inode)->runtime_flags))
3254                 reserve = 1;
3255         spin_unlock(&root->orphan_lock);
3256
3257         /* grab metadata reservation from transaction handle */
3258         if (reserve) {
3259                 ret = btrfs_orphan_reserve_metadata(trans, inode);
3260                 BUG_ON(ret); /* -ENOSPC in reservation; Logic error? JDM */
3261         }
3262
3263         /* insert an orphan item to track this unlinked/truncated file */
3264         if (insert >= 1) {
3265                 ret = btrfs_insert_orphan_item(trans, root, btrfs_ino(inode));
3266                 if (ret) {
3267                         atomic_dec(&root->orphan_inodes);
3268                         if (reserve) {
3269                                 clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3270                                           &BTRFS_I(inode)->runtime_flags);
3271                                 btrfs_orphan_release_metadata(inode);
3272                         }
3273                         if (ret != -EEXIST) {
3274                                 clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3275                                           &BTRFS_I(inode)->runtime_flags);
3276                                 btrfs_abort_transaction(trans, root, ret);
3277                                 return ret;
3278                         }
3279                 }
3280                 ret = 0;
3281         }
3282
3283         /* insert an orphan item to track subvolume contains orphan files */
3284         if (insert >= 2) {
3285                 ret = btrfs_insert_orphan_item(trans, root->fs_info->tree_root,
3286                                                root->root_key.objectid);
3287                 if (ret && ret != -EEXIST) {
3288                         btrfs_abort_transaction(trans, root, ret);
3289                         return ret;
3290                 }
3291         }
3292         return 0;
3293 }
3294
3295 /*
3296  * We have done the truncate/delete so we can go ahead and remove the orphan
3297  * item for this particular inode.
3298  */
3299 static int btrfs_orphan_del(struct btrfs_trans_handle *trans,
3300                             struct inode *inode)
3301 {
3302         struct btrfs_root *root = BTRFS_I(inode)->root;
3303         int delete_item = 0;
3304         int release_rsv = 0;
3305         int ret = 0;
3306
3307         spin_lock(&root->orphan_lock);
3308         if (test_and_clear_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3309                                &BTRFS_I(inode)->runtime_flags))
3310                 delete_item = 1;
3311
3312         if (test_and_clear_bit(BTRFS_INODE_ORPHAN_META_RESERVED,
3313                                &BTRFS_I(inode)->runtime_flags))
3314                 release_rsv = 1;
3315         spin_unlock(&root->orphan_lock);
3316
3317         if (delete_item) {
3318                 atomic_dec(&root->orphan_inodes);
3319                 if (trans)
3320                         ret = btrfs_del_orphan_item(trans, root,
3321                                                     btrfs_ino(inode));
3322         }
3323
3324         if (release_rsv)
3325                 btrfs_orphan_release_metadata(inode);
3326
3327         return ret;
3328 }
3329
3330 /*
3331  * this cleans up any orphans that may be left on the list from the last use
3332  * of this root.
3333  */
3334 int btrfs_orphan_cleanup(struct btrfs_root *root)
3335 {
3336         struct btrfs_path *path;
3337         struct extent_buffer *leaf;
3338         struct btrfs_key key, found_key;
3339         struct btrfs_trans_handle *trans;
3340         struct inode *inode;
3341         u64 last_objectid = 0;
3342         int ret = 0, nr_unlink = 0, nr_truncate = 0;
3343
3344         if (cmpxchg(&root->orphan_cleanup_state, 0, ORPHAN_CLEANUP_STARTED))
3345                 return 0;
3346
3347         path = btrfs_alloc_path();
3348         if (!path) {
3349                 ret = -ENOMEM;
3350                 goto out;
3351         }
3352         path->reada = READA_BACK;
3353
3354         key.objectid = BTRFS_ORPHAN_OBJECTID;
3355         key.type = BTRFS_ORPHAN_ITEM_KEY;
3356         key.offset = (u64)-1;
3357
3358         while (1) {
3359                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
3360                 if (ret < 0)
3361                         goto out;
3362
3363                 /*
3364                  * if ret == 0 means we found what we were searching for, which
3365                  * is weird, but possible, so only screw with path if we didn't
3366                  * find the key and see if we have stuff that matches
3367                  */
3368                 if (ret > 0) {
3369                         ret = 0;
3370                         if (path->slots[0] == 0)
3371                                 break;
3372                         path->slots[0]--;
3373                 }
3374
3375                 /* pull out the item */
3376                 leaf = path->nodes[0];
3377                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
3378
3379                 /* make sure the item matches what we want */
3380                 if (found_key.objectid != BTRFS_ORPHAN_OBJECTID)
3381                         break;
3382                 if (found_key.type != BTRFS_ORPHAN_ITEM_KEY)
3383                         break;
3384
3385                 /* release the path since we're done with it */
3386                 btrfs_release_path(path);
3387
3388                 /*
3389                  * this is where we are basically btrfs_lookup, without the
3390                  * crossing root thing.  we store the inode number in the
3391                  * offset of the orphan item.
3392                  */
3393
3394                 if (found_key.offset == last_objectid) {
3395                         btrfs_err(root->fs_info,
3396                                 "Error removing orphan entry, stopping orphan cleanup");
3397                         ret = -EINVAL;
3398                         goto out;
3399                 }
3400
3401                 last_objectid = found_key.offset;
3402
3403                 found_key.objectid = found_key.offset;
3404                 found_key.type = BTRFS_INODE_ITEM_KEY;
3405                 found_key.offset = 0;
3406                 inode = btrfs_iget(root->fs_info->sb, &found_key, root, NULL);
3407                 ret = PTR_ERR_OR_ZERO(inode);
3408                 if (ret && ret != -ESTALE)
3409                         goto out;
3410
3411                 if (ret == -ESTALE && root == root->fs_info->tree_root) {
3412                         struct btrfs_root *dead_root;
3413                         struct btrfs_fs_info *fs_info = root->fs_info;
3414                         int is_dead_root = 0;
3415
3416                         /*
3417                          * this is an orphan in the tree root. Currently these
3418                          * could come from 2 sources:
3419                          *  a) a snapshot deletion in progress
3420                          *  b) a free space cache inode
3421                          * We need to distinguish those two, as the snapshot
3422                          * orphan must not get deleted.
3423                          * find_dead_roots already ran before us, so if this
3424                          * is a snapshot deletion, we should find the root
3425                          * in the dead_roots list
3426                          */
3427                         spin_lock(&fs_info->trans_lock);
3428                         list_for_each_entry(dead_root, &fs_info->dead_roots,
3429                                             root_list) {
3430                                 if (dead_root->root_key.objectid ==
3431                                     found_key.objectid) {
3432                                         is_dead_root = 1;
3433                                         break;
3434                                 }
3435                         }
3436                         spin_unlock(&fs_info->trans_lock);
3437                         if (is_dead_root) {
3438                                 /* prevent this orphan from being found again */
3439                                 key.offset = found_key.objectid - 1;
3440                                 continue;
3441                         }
3442                 }
3443                 /*
3444                  * Inode is already gone but the orphan item is still there,
3445                  * kill the orphan item.
3446                  */
3447                 if (ret == -ESTALE) {
3448                         trans = btrfs_start_transaction(root, 1);
3449                         if (IS_ERR(trans)) {
3450                                 ret = PTR_ERR(trans);
3451                                 goto out;
3452                         }
3453                         btrfs_debug(root->fs_info, "auto deleting %Lu",
3454                                 found_key.objectid);
3455                         ret = btrfs_del_orphan_item(trans, root,
3456                                                     found_key.objectid);
3457                         btrfs_end_transaction(trans, root);
3458                         if (ret)
3459                                 goto out;
3460                         continue;
3461                 }
3462
3463                 /*
3464                  * add this inode to the orphan list so btrfs_orphan_del does
3465                  * the proper thing when we hit it
3466                  */
3467                 set_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
3468                         &BTRFS_I(inode)->runtime_flags);
3469                 atomic_inc(&root->orphan_inodes);
3470
3471                 /* if we have links, this was a truncate, lets do that */
3472                 if (inode->i_nlink) {
3473                         if (WARN_ON(!S_ISREG(inode->i_mode))) {
3474                                 iput(inode);
3475                                 continue;
3476                         }
3477                         nr_truncate++;
3478
3479                         /* 1 for the orphan item deletion. */
3480                         trans = btrfs_start_transaction(root, 1);
3481                         if (IS_ERR(trans)) {
3482                                 iput(inode);
3483                                 ret = PTR_ERR(trans);
3484                                 goto out;
3485                         }
3486                         ret = btrfs_orphan_add(trans, inode);
3487                         btrfs_end_transaction(trans, root);
3488                         if (ret) {
3489                                 iput(inode);
3490                                 goto out;
3491                         }
3492
3493                         ret = btrfs_truncate(inode);
3494                         if (ret)
3495                                 btrfs_orphan_del(NULL, inode);
3496                 } else {
3497                         nr_unlink++;
3498                 }
3499
3500                 /* this will do delete_inode and everything for us */
3501                 iput(inode);
3502                 if (ret)
3503                         goto out;
3504         }
3505         /* release the path since we're done with it */
3506         btrfs_release_path(path);
3507
3508         root->orphan_cleanup_state = ORPHAN_CLEANUP_DONE;
3509
3510         if (root->orphan_block_rsv)
3511                 btrfs_block_rsv_release(root, root->orphan_block_rsv,
3512                                         (u64)-1);
3513
3514         if (root->orphan_block_rsv ||
3515             test_bit(BTRFS_ROOT_ORPHAN_ITEM_INSERTED, &root->state)) {
3516                 trans = btrfs_join_transaction(root);
3517                 if (!IS_ERR(trans))
3518                         btrfs_end_transaction(trans, root);
3519         }
3520
3521         if (nr_unlink)
3522                 btrfs_debug(root->fs_info, "unlinked %d orphans", nr_unlink);
3523         if (nr_truncate)
3524                 btrfs_debug(root->fs_info, "truncated %d orphans", nr_truncate);
3525
3526 out:
3527         if (ret)
3528                 btrfs_err(root->fs_info,
3529                         "could not do orphan cleanup %d", ret);
3530         btrfs_free_path(path);
3531         return ret;
3532 }
3533
3534 /*
3535  * very simple check to peek ahead in the leaf looking for xattrs.  If we
3536  * don't find any xattrs, we know there can't be any acls.
3537  *
3538  * slot is the slot the inode is in, objectid is the objectid of the inode
3539  */
3540 static noinline int acls_after_inode_item(struct extent_buffer *leaf,
3541                                           int slot, u64 objectid,
3542                                           int *first_xattr_slot)
3543 {
3544         u32 nritems = btrfs_header_nritems(leaf);
3545         struct btrfs_key found_key;
3546         static u64 xattr_access = 0;
3547         static u64 xattr_default = 0;
3548         int scanned = 0;
3549
3550         if (!xattr_access) {
3551                 xattr_access = btrfs_name_hash(POSIX_ACL_XATTR_ACCESS,
3552                                         strlen(POSIX_ACL_XATTR_ACCESS));
3553                 xattr_default = btrfs_name_hash(POSIX_ACL_XATTR_DEFAULT,
3554                                         strlen(POSIX_ACL_XATTR_DEFAULT));
3555         }
3556
3557         slot++;
3558         *first_xattr_slot = -1;
3559         while (slot < nritems) {
3560                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
3561
3562                 /* we found a different objectid, there must not be acls */
3563                 if (found_key.objectid != objectid)
3564                         return 0;
3565
3566                 /* we found an xattr, assume we've got an acl */
3567                 if (found_key.type == BTRFS_XATTR_ITEM_KEY) {
3568                         if (*first_xattr_slot == -1)
3569                                 *first_xattr_slot = slot;
3570                         if (found_key.offset == xattr_access ||
3571                             found_key.offset == xattr_default)
3572                                 return 1;
3573                 }
3574
3575                 /*
3576                  * we found a key greater than an xattr key, there can't
3577                  * be any acls later on
3578                  */
3579                 if (found_key.type > BTRFS_XATTR_ITEM_KEY)
3580                         return 0;
3581
3582                 slot++;
3583                 scanned++;
3584
3585                 /*
3586                  * it goes inode, inode backrefs, xattrs, extents,
3587                  * so if there are a ton of hard links to an inode there can
3588                  * be a lot of backrefs.  Don't waste time searching too hard,
3589                  * this is just an optimization
3590                  */
3591                 if (scanned >= 8)
3592                         break;
3593         }
3594         /* we hit the end of the leaf before we found an xattr or
3595          * something larger than an xattr.  We have to assume the inode
3596          * has acls
3597          */
3598         if (*first_xattr_slot == -1)
3599                 *first_xattr_slot = slot;
3600         return 1;
3601 }
3602
3603 /*
3604  * read an inode from the btree into the in-memory inode
3605  */
3606 static void btrfs_read_locked_inode(struct inode *inode)
3607 {
3608         struct btrfs_path *path;
3609         struct extent_buffer *leaf;
3610         struct btrfs_inode_item *inode_item;
3611         struct btrfs_root *root = BTRFS_I(inode)->root;
3612         struct btrfs_key location;
3613         unsigned long ptr;
3614         int maybe_acls;
3615         u32 rdev;
3616         int ret;
3617         bool filled = false;
3618         int first_xattr_slot;
3619
3620         ret = btrfs_fill_inode(inode, &rdev);
3621         if (!ret)
3622                 filled = true;
3623
3624         path = btrfs_alloc_path();
3625         if (!path)
3626                 goto make_bad;
3627
3628         memcpy(&location, &BTRFS_I(inode)->location, sizeof(location));
3629
3630         ret = btrfs_lookup_inode(NULL, root, path, &location, 0);
3631         if (ret)
3632                 goto make_bad;
3633
3634         leaf = path->nodes[0];
3635
3636         if (filled)
3637                 goto cache_index;
3638
3639         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3640                                     struct btrfs_inode_item);
3641         inode->i_mode = btrfs_inode_mode(leaf, inode_item);
3642         set_nlink(inode, btrfs_inode_nlink(leaf, inode_item));
3643         i_uid_write(inode, btrfs_inode_uid(leaf, inode_item));
3644         i_gid_write(inode, btrfs_inode_gid(leaf, inode_item));
3645         btrfs_i_size_write(inode, btrfs_inode_size(leaf, inode_item));
3646
3647         inode->i_atime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->atime);
3648         inode->i_atime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->atime);
3649
3650         inode->i_mtime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->mtime);
3651         inode->i_mtime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->mtime);
3652
3653         inode->i_ctime.tv_sec = btrfs_timespec_sec(leaf, &inode_item->ctime);
3654         inode->i_ctime.tv_nsec = btrfs_timespec_nsec(leaf, &inode_item->ctime);
3655
3656         BTRFS_I(inode)->i_otime.tv_sec =
3657                 btrfs_timespec_sec(leaf, &inode_item->otime);
3658         BTRFS_I(inode)->i_otime.tv_nsec =
3659                 btrfs_timespec_nsec(leaf, &inode_item->otime);
3660
3661         inode_set_bytes(inode, btrfs_inode_nbytes(leaf, inode_item));
3662         BTRFS_I(inode)->generation = btrfs_inode_generation(leaf, inode_item);
3663         BTRFS_I(inode)->last_trans = btrfs_inode_transid(leaf, inode_item);
3664
3665         inode->i_version = btrfs_inode_sequence(leaf, inode_item);
3666         inode->i_generation = BTRFS_I(inode)->generation;
3667         inode->i_rdev = 0;
3668         rdev = btrfs_inode_rdev(leaf, inode_item);
3669
3670         BTRFS_I(inode)->index_cnt = (u64)-1;
3671         BTRFS_I(inode)->flags = btrfs_inode_flags(leaf, inode_item);
3672
3673 cache_index:
3674         /*
3675          * If we were modified in the current generation and evicted from memory
3676          * and then re-read we need to do a full sync since we don't have any
3677          * idea about which extents were modified before we were evicted from
3678          * cache.
3679          *
3680          * This is required for both inode re-read from disk and delayed inode
3681          * in delayed_nodes_tree.
3682          */
3683         if (BTRFS_I(inode)->last_trans == root->fs_info->generation)
3684                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
3685                         &BTRFS_I(inode)->runtime_flags);
3686
3687         /*
3688          * We don't persist the id of the transaction where an unlink operation
3689          * against the inode was last made. So here we assume the inode might
3690          * have been evicted, and therefore the exact value of last_unlink_trans
3691          * lost, and set it to last_trans to avoid metadata inconsistencies
3692          * between the inode and its parent if the inode is fsync'ed and the log
3693          * replayed. For example, in the scenario:
3694          *
3695          * touch mydir/foo
3696          * ln mydir/foo mydir/bar
3697          * sync
3698          * unlink mydir/bar
3699          * echo 2 > /proc/sys/vm/drop_caches   # evicts inode
3700          * xfs_io -c fsync mydir/foo
3701          * <power failure>
3702          * mount fs, triggers fsync log replay
3703          *
3704          * We must make sure that when we fsync our inode foo we also log its
3705          * parent inode, otherwise after log replay the parent still has the
3706          * dentry with the "bar" name but our inode foo has a link count of 1
3707          * and doesn't have an inode ref with the name "bar" anymore.
3708          *
3709          * Setting last_unlink_trans to last_trans is a pessimistic approach,
3710          * but it guarantees correctness at the expense of ocassional full
3711          * transaction commits on fsync if our inode is a directory, or if our
3712          * inode is not a directory, logging its parent unnecessarily.
3713          */
3714         BTRFS_I(inode)->last_unlink_trans = BTRFS_I(inode)->last_trans;
3715
3716         path->slots[0]++;
3717         if (inode->i_nlink != 1 ||
3718             path->slots[0] >= btrfs_header_nritems(leaf))
3719                 goto cache_acl;
3720
3721         btrfs_item_key_to_cpu(leaf, &location, path->slots[0]);
3722         if (location.objectid != btrfs_ino(inode))
3723                 goto cache_acl;
3724
3725         ptr = btrfs_item_ptr_offset(leaf, path->slots[0]);
3726         if (location.type == BTRFS_INODE_REF_KEY) {
3727                 struct btrfs_inode_ref *ref;
3728
3729                 ref = (struct btrfs_inode_ref *)ptr;
3730                 BTRFS_I(inode)->dir_index = btrfs_inode_ref_index(leaf, ref);
3731         } else if (location.type == BTRFS_INODE_EXTREF_KEY) {
3732                 struct btrfs_inode_extref *extref;
3733
3734                 extref = (struct btrfs_inode_extref *)ptr;
3735                 BTRFS_I(inode)->dir_index = btrfs_inode_extref_index(leaf,
3736                                                                      extref);
3737         }
3738 cache_acl:
3739         /*
3740          * try to precache a NULL acl entry for files that don't have
3741          * any xattrs or acls
3742          */
3743         maybe_acls = acls_after_inode_item(leaf, path->slots[0],
3744                                            btrfs_ino(inode), &first_xattr_slot);
3745         if (first_xattr_slot != -1) {
3746                 path->slots[0] = first_xattr_slot;
3747                 ret = btrfs_load_inode_props(inode, path);
3748                 if (ret)
3749                         btrfs_err(root->fs_info,
3750                                   "error loading props for ino %llu (root %llu): %d",
3751                                   btrfs_ino(inode),
3752                                   root->root_key.objectid, ret);
3753         }
3754         btrfs_free_path(path);
3755
3756         if (!maybe_acls)
3757                 cache_no_acl(inode);
3758
3759         switch (inode->i_mode & S_IFMT) {
3760         case S_IFREG:
3761                 inode->i_mapping->a_ops = &btrfs_aops;
3762                 BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
3763                 inode->i_fop = &btrfs_file_operations;
3764                 inode->i_op = &btrfs_file_inode_operations;
3765                 break;
3766         case S_IFDIR:
3767                 inode->i_fop = &btrfs_dir_file_operations;
3768                 if (root == root->fs_info->tree_root)
3769                         inode->i_op = &btrfs_dir_ro_inode_operations;
3770                 else
3771                         inode->i_op = &btrfs_dir_inode_operations;
3772                 break;
3773         case S_IFLNK:
3774                 inode->i_op = &btrfs_symlink_inode_operations;
3775                 inode->i_mapping->a_ops = &btrfs_symlink_aops;
3776                 break;
3777         default:
3778                 inode->i_op = &btrfs_special_inode_operations;
3779                 init_special_inode(inode, inode->i_mode, rdev);
3780                 break;
3781         }
3782
3783         btrfs_update_iflags(inode);
3784         return;
3785
3786 make_bad:
3787         btrfs_free_path(path);
3788         make_bad_inode(inode);
3789 }
3790
3791 /*
3792  * given a leaf and an inode, copy the inode fields into the leaf
3793  */
3794 static void fill_inode_item(struct btrfs_trans_handle *trans,
3795                             struct extent_buffer *leaf,
3796                             struct btrfs_inode_item *item,
3797                             struct inode *inode)
3798 {
3799         struct btrfs_map_token token;
3800
3801         btrfs_init_map_token(&token);
3802
3803         btrfs_set_token_inode_uid(leaf, item, i_uid_read(inode), &token);
3804         btrfs_set_token_inode_gid(leaf, item, i_gid_read(inode), &token);
3805         btrfs_set_token_inode_size(leaf, item, BTRFS_I(inode)->disk_i_size,
3806                                    &token);
3807         btrfs_set_token_inode_mode(leaf, item, inode->i_mode, &token);
3808         btrfs_set_token_inode_nlink(leaf, item, inode->i_nlink, &token);
3809
3810         btrfs_set_token_timespec_sec(leaf, &item->atime,
3811                                      inode->i_atime.tv_sec, &token);
3812         btrfs_set_token_timespec_nsec(leaf, &item->atime,
3813                                       inode->i_atime.tv_nsec, &token);
3814
3815         btrfs_set_token_timespec_sec(leaf, &item->mtime,
3816                                      inode->i_mtime.tv_sec, &token);
3817         btrfs_set_token_timespec_nsec(leaf, &item->mtime,
3818                                       inode->i_mtime.tv_nsec, &token);
3819
3820         btrfs_set_token_timespec_sec(leaf, &item->ctime,
3821                                      inode->i_ctime.tv_sec, &token);
3822         btrfs_set_token_timespec_nsec(leaf, &item->ctime,
3823                                       inode->i_ctime.tv_nsec, &token);
3824
3825         btrfs_set_token_timespec_sec(leaf, &item->otime,
3826                                      BTRFS_I(inode)->i_otime.tv_sec, &token);
3827         btrfs_set_token_timespec_nsec(leaf, &item->otime,
3828                                       BTRFS_I(inode)->i_otime.tv_nsec, &token);
3829
3830         btrfs_set_token_inode_nbytes(leaf, item, inode_get_bytes(inode),
3831                                      &token);
3832         btrfs_set_token_inode_generation(leaf, item, BTRFS_I(inode)->generation,
3833                                          &token);
3834         btrfs_set_token_inode_sequence(leaf, item, inode->i_version, &token);
3835         btrfs_set_token_inode_transid(leaf, item, trans->transid, &token);
3836         btrfs_set_token_inode_rdev(leaf, item, inode->i_rdev, &token);
3837         btrfs_set_token_inode_flags(leaf, item, BTRFS_I(inode)->flags, &token);
3838         btrfs_set_token_inode_block_group(leaf, item, 0, &token);
3839 }
3840
3841 /*
3842  * copy everything in the in-memory inode into the btree.
3843  */
3844 static noinline int btrfs_update_inode_item(struct btrfs_trans_handle *trans,
3845                                 struct btrfs_root *root, struct inode *inode)
3846 {
3847         struct btrfs_inode_item *inode_item;
3848         struct btrfs_path *path;
3849         struct extent_buffer *leaf;
3850         int ret;
3851
3852         path = btrfs_alloc_path();
3853         if (!path)
3854                 return -ENOMEM;
3855
3856         path->leave_spinning = 1;
3857         ret = btrfs_lookup_inode(trans, root, path, &BTRFS_I(inode)->location,
3858                                  1);
3859         if (ret) {
3860                 if (ret > 0)
3861                         ret = -ENOENT;
3862                 goto failed;
3863         }
3864
3865         leaf = path->nodes[0];
3866         inode_item = btrfs_item_ptr(leaf, path->slots[0],
3867                                     struct btrfs_inode_item);
3868
3869         fill_inode_item(trans, leaf, inode_item, inode);
3870         btrfs_mark_buffer_dirty(leaf);
3871         btrfs_set_inode_last_trans(trans, inode);
3872         ret = 0;
3873 failed:
3874         btrfs_free_path(path);
3875         return ret;
3876 }
3877
3878 /*
3879  * copy everything in the in-memory inode into the btree.
3880  */
3881 noinline int btrfs_update_inode(struct btrfs_trans_handle *trans,
3882                                 struct btrfs_root *root, struct inode *inode)
3883 {
3884         int ret;
3885
3886         /*
3887          * If the inode is a free space inode, we can deadlock during commit
3888          * if we put it into the delayed code.
3889          *
3890          * The data relocation inode should also be directly updated
3891          * without delay
3892          */
3893         if (!btrfs_is_free_space_inode(inode)
3894             && root->root_key.objectid != BTRFS_DATA_RELOC_TREE_OBJECTID
3895             && !root->fs_info->log_root_recovering) {
3896                 btrfs_update_root_times(trans, root);
3897
3898                 ret = btrfs_delayed_update_inode(trans, root, inode);
3899                 if (!ret)
3900                         btrfs_set_inode_last_trans(trans, inode);
3901                 return ret;
3902         }
3903
3904         return btrfs_update_inode_item(trans, root, inode);
3905 }
3906
3907 noinline int btrfs_update_inode_fallback(struct btrfs_trans_handle *trans,
3908                                          struct btrfs_root *root,
3909                                          struct inode *inode)
3910 {
3911         int ret;
3912
3913         ret = btrfs_update_inode(trans, root, inode);
3914         if (ret == -ENOSPC)
3915                 return btrfs_update_inode_item(trans, root, inode);
3916         return ret;
3917 }
3918
3919 /*
3920  * unlink helper that gets used here in inode.c and in the tree logging
3921  * recovery code.  It remove a link in a directory with a given name, and
3922  * also drops the back refs in the inode to the directory
3923  */
3924 static int __btrfs_unlink_inode(struct btrfs_trans_handle *trans,
3925                                 struct btrfs_root *root,
3926                                 struct inode *dir, struct inode *inode,
3927                                 const char *name, int name_len)
3928 {
3929         struct btrfs_path *path;
3930         int ret = 0;
3931         struct extent_buffer *leaf;
3932         struct btrfs_dir_item *di;
3933         struct btrfs_key key;
3934         u64 index;
3935         u64 ino = btrfs_ino(inode);
3936         u64 dir_ino = btrfs_ino(dir);
3937
3938         path = btrfs_alloc_path();
3939         if (!path) {
3940                 ret = -ENOMEM;
3941                 goto out;
3942         }
3943
3944         path->leave_spinning = 1;
3945         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
3946                                     name, name_len, -1);
3947         if (IS_ERR(di)) {
3948                 ret = PTR_ERR(di);
3949                 goto err;
3950         }
3951         if (!di) {
3952                 ret = -ENOENT;
3953                 goto err;
3954         }
3955         leaf = path->nodes[0];
3956         btrfs_dir_item_key_to_cpu(leaf, di, &key);
3957         ret = btrfs_delete_one_dir_name(trans, root, path, di);
3958         if (ret)
3959                 goto err;
3960         btrfs_release_path(path);
3961
3962         /*
3963          * If we don't have dir index, we have to get it by looking up
3964          * the inode ref, since we get the inode ref, remove it directly,
3965          * it is unnecessary to do delayed deletion.
3966          *
3967          * But if we have dir index, needn't search inode ref to get it.
3968          * Since the inode ref is close to the inode item, it is better
3969          * that we delay to delete it, and just do this deletion when
3970          * we update the inode item.
3971          */
3972         if (BTRFS_I(inode)->dir_index) {
3973                 ret = btrfs_delayed_delete_inode_ref(inode);
3974                 if (!ret) {
3975                         index = BTRFS_I(inode)->dir_index;
3976                         goto skip_backref;
3977                 }
3978         }
3979
3980         ret = btrfs_del_inode_ref(trans, root, name, name_len, ino,
3981                                   dir_ino, &index);
3982         if (ret) {
3983                 btrfs_info(root->fs_info,
3984                         "failed to delete reference to %.*s, inode %llu parent %llu",
3985                         name_len, name, ino, dir_ino);
3986                 btrfs_abort_transaction(trans, root, ret);
3987                 goto err;
3988         }
3989 skip_backref:
3990         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
3991         if (ret) {
3992                 btrfs_abort_transaction(trans, root, ret);
3993                 goto err;
3994         }
3995
3996         ret = btrfs_del_inode_ref_in_log(trans, root, name, name_len,
3997                                          inode, dir_ino);
3998         if (ret != 0 && ret != -ENOENT) {
3999                 btrfs_abort_transaction(trans, root, ret);
4000                 goto err;
4001         }
4002
4003         ret = btrfs_del_dir_entries_in_log(trans, root, name, name_len,
4004                                            dir, index);
4005         if (ret == -ENOENT)
4006                 ret = 0;
4007         else if (ret)
4008                 btrfs_abort_transaction(trans, root, ret);
4009 err:
4010         btrfs_free_path(path);
4011         if (ret)
4012                 goto out;
4013
4014         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4015         inode_inc_iversion(inode);
4016         inode_inc_iversion(dir);
4017         inode->i_ctime = dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4018         ret = btrfs_update_inode(trans, root, dir);
4019 out:
4020         return ret;
4021 }
4022
4023 int btrfs_unlink_inode(struct btrfs_trans_handle *trans,
4024                        struct btrfs_root *root,
4025                        struct inode *dir, struct inode *inode,
4026                        const char *name, int name_len)
4027 {
4028         int ret;
4029         ret = __btrfs_unlink_inode(trans, root, dir, inode, name, name_len);
4030         if (!ret) {
4031                 drop_nlink(inode);
4032                 ret = btrfs_update_inode(trans, root, inode);
4033         }
4034         return ret;
4035 }
4036
4037 /*
4038  * helper to start transaction for unlink and rmdir.
4039  *
4040  * unlink and rmdir are special in btrfs, they do not always free space, so
4041  * if we cannot make our reservations the normal way try and see if there is
4042  * plenty of slack room in the global reserve to migrate, otherwise we cannot
4043  * allow the unlink to occur.
4044  */
4045 static struct btrfs_trans_handle *__unlink_start_trans(struct inode *dir)
4046 {
4047         struct btrfs_root *root = BTRFS_I(dir)->root;
4048
4049         /*
4050          * 1 for the possible orphan item
4051          * 1 for the dir item
4052          * 1 for the dir index
4053          * 1 for the inode ref
4054          * 1 for the inode
4055          */
4056         return btrfs_start_transaction_fallback_global_rsv(root, 5, 5);
4057 }
4058
4059 static int btrfs_unlink(struct inode *dir, struct dentry *dentry)
4060 {
4061         struct btrfs_root *root = BTRFS_I(dir)->root;
4062         struct btrfs_trans_handle *trans;
4063         struct inode *inode = d_inode(dentry);
4064         int ret;
4065
4066         trans = __unlink_start_trans(dir);
4067         if (IS_ERR(trans))
4068                 return PTR_ERR(trans);
4069
4070         btrfs_record_unlink_dir(trans, dir, d_inode(dentry), 0);
4071
4072         ret = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4073                                  dentry->d_name.name, dentry->d_name.len);
4074         if (ret)
4075                 goto out;
4076
4077         if (inode->i_nlink == 0) {
4078                 ret = btrfs_orphan_add(trans, inode);
4079                 if (ret)
4080                         goto out;
4081         }
4082
4083 out:
4084         btrfs_end_transaction(trans, root);
4085         btrfs_btree_balance_dirty(root);
4086         return ret;
4087 }
4088
4089 int btrfs_unlink_subvol(struct btrfs_trans_handle *trans,
4090                         struct btrfs_root *root,
4091                         struct inode *dir, u64 objectid,
4092                         const char *name, int name_len)
4093 {
4094         struct btrfs_path *path;
4095         struct extent_buffer *leaf;
4096         struct btrfs_dir_item *di;
4097         struct btrfs_key key;
4098         u64 index;
4099         int ret;
4100         u64 dir_ino = btrfs_ino(dir);
4101
4102         path = btrfs_alloc_path();
4103         if (!path)
4104                 return -ENOMEM;
4105
4106         di = btrfs_lookup_dir_item(trans, root, path, dir_ino,
4107                                    name, name_len, -1);
4108         if (IS_ERR_OR_NULL(di)) {
4109                 if (!di)
4110                         ret = -ENOENT;
4111                 else
4112                         ret = PTR_ERR(di);
4113                 goto out;
4114         }
4115
4116         leaf = path->nodes[0];
4117         btrfs_dir_item_key_to_cpu(leaf, di, &key);
4118         WARN_ON(key.type != BTRFS_ROOT_ITEM_KEY || key.objectid != objectid);
4119         ret = btrfs_delete_one_dir_name(trans, root, path, di);
4120         if (ret) {
4121                 btrfs_abort_transaction(trans, root, ret);
4122                 goto out;
4123         }
4124         btrfs_release_path(path);
4125
4126         ret = btrfs_del_root_ref(trans, root->fs_info->tree_root,
4127                                  objectid, root->root_key.objectid,
4128                                  dir_ino, &index, name, name_len);
4129         if (ret < 0) {
4130                 if (ret != -ENOENT) {
4131                         btrfs_abort_transaction(trans, root, ret);
4132                         goto out;
4133                 }
4134                 di = btrfs_search_dir_index_item(root, path, dir_ino,
4135                                                  name, name_len);
4136                 if (IS_ERR_OR_NULL(di)) {
4137                         if (!di)
4138                                 ret = -ENOENT;
4139                         else
4140                                 ret = PTR_ERR(di);
4141                         btrfs_abort_transaction(trans, root, ret);
4142                         goto out;
4143                 }
4144
4145                 leaf = path->nodes[0];
4146                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
4147                 btrfs_release_path(path);
4148                 index = key.offset;
4149         }
4150         btrfs_release_path(path);
4151
4152         ret = btrfs_delete_delayed_dir_index(trans, root, dir, index);
4153         if (ret) {
4154                 btrfs_abort_transaction(trans, root, ret);
4155                 goto out;
4156         }
4157
4158         btrfs_i_size_write(dir, dir->i_size - name_len * 2);
4159         inode_inc_iversion(dir);
4160         dir->i_mtime = dir->i_ctime = CURRENT_TIME;
4161         ret = btrfs_update_inode_fallback(trans, root, dir);
4162         if (ret)
4163                 btrfs_abort_transaction(trans, root, ret);
4164 out:
4165         btrfs_free_path(path);
4166         return ret;
4167 }
4168
4169 static int btrfs_rmdir(struct inode *dir, struct dentry *dentry)
4170 {
4171         struct inode *inode = d_inode(dentry);
4172         int err = 0;
4173         struct btrfs_root *root = BTRFS_I(dir)->root;
4174         struct btrfs_trans_handle *trans;
4175
4176         if (inode->i_size > BTRFS_EMPTY_DIR_SIZE)
4177                 return -ENOTEMPTY;
4178         if (btrfs_ino(inode) == BTRFS_FIRST_FREE_OBJECTID)
4179                 return -EPERM;
4180
4181         trans = __unlink_start_trans(dir);
4182         if (IS_ERR(trans))
4183                 return PTR_ERR(trans);
4184
4185         if (unlikely(btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
4186                 err = btrfs_unlink_subvol(trans, root, dir,
4187                                           BTRFS_I(inode)->location.objectid,
4188                                           dentry->d_name.name,
4189                                           dentry->d_name.len);
4190                 goto out;
4191         }
4192
4193         err = btrfs_orphan_add(trans, inode);
4194         if (err)
4195                 goto out;
4196
4197         /* now the directory is empty */
4198         err = btrfs_unlink_inode(trans, root, dir, d_inode(dentry),
4199                                  dentry->d_name.name, dentry->d_name.len);
4200         if (!err)
4201                 btrfs_i_size_write(inode, 0);
4202 out:
4203         btrfs_end_transaction(trans, root);
4204         btrfs_btree_balance_dirty(root);
4205
4206         return err;
4207 }
4208
4209 static int truncate_space_check(struct btrfs_trans_handle *trans,
4210                                 struct btrfs_root *root,
4211                                 u64 bytes_deleted)
4212 {
4213         int ret;
4214
4215         bytes_deleted = btrfs_csum_bytes_to_leaves(root, bytes_deleted);
4216         ret = btrfs_block_rsv_add(root, &root->fs_info->trans_block_rsv,
4217                                   bytes_deleted, BTRFS_RESERVE_NO_FLUSH);
4218         if (!ret)
4219                 trans->bytes_reserved += bytes_deleted;
4220         return ret;
4221
4222 }
4223
4224 static int truncate_inline_extent(struct inode *inode,
4225                                   struct btrfs_path *path,
4226                                   struct btrfs_key *found_key,
4227                                   const u64 item_end,
4228                                   const u64 new_size)
4229 {
4230         struct extent_buffer *leaf = path->nodes[0];
4231         int slot = path->slots[0];
4232         struct btrfs_file_extent_item *fi;
4233         u32 size = (u32)(new_size - found_key->offset);
4234         struct btrfs_root *root = BTRFS_I(inode)->root;
4235
4236         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
4237
4238         if (btrfs_file_extent_compression(leaf, fi) != BTRFS_COMPRESS_NONE) {
4239                 loff_t offset = new_size;
4240                 loff_t page_end = ALIGN(offset, PAGE_CACHE_SIZE);
4241
4242                 /*
4243                  * Zero out the remaining of the last page of our inline extent,
4244                  * instead of directly truncating our inline extent here - that
4245                  * would be much more complex (decompressing all the data, then
4246                  * compressing the truncated data, which might be bigger than
4247                  * the size of the inline extent, resize the extent, etc).
4248                  * We release the path because to get the page we might need to
4249                  * read the extent item from disk (data not in the page cache).
4250                  */
4251                 btrfs_release_path(path);
4252                 return btrfs_truncate_page(inode, offset, page_end - offset, 0);
4253         }
4254
4255         btrfs_set_file_extent_ram_bytes(leaf, fi, size);
4256         size = btrfs_file_extent_calc_inline_size(size);
4257         btrfs_truncate_item(root, path, size, 1);
4258
4259         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4260                 inode_sub_bytes(inode, item_end + 1 - new_size);
4261
4262         return 0;
4263 }
4264
4265 /*
4266  * this can truncate away extent items, csum items and directory items.
4267  * It starts at a high offset and removes keys until it can't find
4268  * any higher than new_size
4269  *
4270  * csum items that cross the new i_size are truncated to the new size
4271  * as well.
4272  *
4273  * min_type is the minimum key type to truncate down to.  If set to 0, this
4274  * will kill all the items on this inode, including the INODE_ITEM_KEY.
4275  */
4276 int btrfs_truncate_inode_items(struct btrfs_trans_handle *trans,
4277                                struct btrfs_root *root,
4278                                struct inode *inode,
4279                                u64 new_size, u32 min_type)
4280 {
4281         struct btrfs_path *path;
4282         struct extent_buffer *leaf;
4283         struct btrfs_file_extent_item *fi;
4284         struct btrfs_key key;
4285         struct btrfs_key found_key;
4286         u64 extent_start = 0;
4287         u64 extent_num_bytes = 0;
4288         u64 extent_offset = 0;
4289         u64 item_end = 0;
4290         u64 last_size = new_size;
4291         u32 found_type = (u8)-1;
4292         int found_extent;
4293         int del_item;
4294         int pending_del_nr = 0;
4295         int pending_del_slot = 0;
4296         int extent_type = -1;
4297         int ret;
4298         int err = 0;
4299         u64 ino = btrfs_ino(inode);
4300         u64 bytes_deleted = 0;
4301         bool be_nice = 0;
4302         bool should_throttle = 0;
4303         bool should_end = 0;
4304
4305         BUG_ON(new_size > 0 && min_type != BTRFS_EXTENT_DATA_KEY);
4306
4307         /*
4308          * for non-free space inodes and ref cows, we want to back off from
4309          * time to time
4310          */
4311         if (!btrfs_is_free_space_inode(inode) &&
4312             test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4313                 be_nice = 1;
4314
4315         path = btrfs_alloc_path();
4316         if (!path)
4317                 return -ENOMEM;
4318         path->reada = READA_BACK;
4319
4320         /*
4321          * We want to drop from the next block forward in case this new size is
4322          * not block aligned since we will be keeping the last block of the
4323          * extent just the way it is.
4324          */
4325         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4326             root == root->fs_info->tree_root)
4327                 btrfs_drop_extent_cache(inode, ALIGN(new_size,
4328                                         root->sectorsize), (u64)-1, 0);
4329
4330         /*
4331          * This function is also used to drop the items in the log tree before
4332          * we relog the inode, so if root != BTRFS_I(inode)->root, it means
4333          * it is used to drop the loged items. So we shouldn't kill the delayed
4334          * items.
4335          */
4336         if (min_type == 0 && root == BTRFS_I(inode)->root)
4337                 btrfs_kill_delayed_inode_items(inode);
4338
4339         key.objectid = ino;
4340         key.offset = (u64)-1;
4341         key.type = (u8)-1;
4342
4343 search_again:
4344         /*
4345          * with a 16K leaf size and 128MB extents, you can actually queue
4346          * up a huge file in a single leaf.  Most of the time that
4347          * bytes_deleted is > 0, it will be huge by the time we get here
4348          */
4349         if (be_nice && bytes_deleted > SZ_32M) {
4350                 if (btrfs_should_end_transaction(trans, root)) {
4351                         err = -EAGAIN;
4352                         goto error;
4353                 }
4354         }
4355
4356
4357         path->leave_spinning = 1;
4358         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
4359         if (ret < 0) {
4360                 err = ret;
4361                 goto out;
4362         }
4363
4364         if (ret > 0) {
4365                 /* there are no items in the tree for us to truncate, we're
4366                  * done
4367                  */
4368                 if (path->slots[0] == 0)
4369                         goto out;
4370                 path->slots[0]--;
4371         }
4372
4373         while (1) {
4374                 fi = NULL;
4375                 leaf = path->nodes[0];
4376                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
4377                 found_type = found_key.type;
4378
4379                 if (found_key.objectid != ino)
4380                         break;
4381
4382                 if (found_type < min_type)
4383                         break;
4384
4385                 item_end = found_key.offset;
4386                 if (found_type == BTRFS_EXTENT_DATA_KEY) {
4387                         fi = btrfs_item_ptr(leaf, path->slots[0],
4388                                             struct btrfs_file_extent_item);
4389                         extent_type = btrfs_file_extent_type(leaf, fi);
4390                         if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4391                                 item_end +=
4392                                     btrfs_file_extent_num_bytes(leaf, fi);
4393                         } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4394                                 item_end += btrfs_file_extent_inline_len(leaf,
4395                                                          path->slots[0], fi);
4396                         }
4397                         item_end--;
4398                 }
4399                 if (found_type > min_type) {
4400                         del_item = 1;
4401                 } else {
4402                         if (item_end < new_size)
4403                                 break;
4404                         if (found_key.offset >= new_size)
4405                                 del_item = 1;
4406                         else
4407                                 del_item = 0;
4408                 }
4409                 found_extent = 0;
4410                 /* FIXME, shrink the extent if the ref count is only 1 */
4411                 if (found_type != BTRFS_EXTENT_DATA_KEY)
4412                         goto delete;
4413
4414                 if (del_item)
4415                         last_size = found_key.offset;
4416                 else
4417                         last_size = new_size;
4418
4419                 if (extent_type != BTRFS_FILE_EXTENT_INLINE) {
4420                         u64 num_dec;
4421                         extent_start = btrfs_file_extent_disk_bytenr(leaf, fi);
4422                         if (!del_item) {
4423                                 u64 orig_num_bytes =
4424                                         btrfs_file_extent_num_bytes(leaf, fi);
4425                                 extent_num_bytes = ALIGN(new_size -
4426                                                 found_key.offset,
4427                                                 root->sectorsize);
4428                                 btrfs_set_file_extent_num_bytes(leaf, fi,
4429                                                          extent_num_bytes);
4430                                 num_dec = (orig_num_bytes -
4431                                            extent_num_bytes);
4432                                 if (test_bit(BTRFS_ROOT_REF_COWS,
4433                                              &root->state) &&
4434                                     extent_start != 0)
4435                                         inode_sub_bytes(inode, num_dec);
4436                                 btrfs_mark_buffer_dirty(leaf);
4437                         } else {
4438                                 extent_num_bytes =
4439                                         btrfs_file_extent_disk_num_bytes(leaf,
4440                                                                          fi);
4441                                 extent_offset = found_key.offset -
4442                                         btrfs_file_extent_offset(leaf, fi);
4443
4444                                 /* FIXME blocksize != 4096 */
4445                                 num_dec = btrfs_file_extent_num_bytes(leaf, fi);
4446                                 if (extent_start != 0) {
4447                                         found_extent = 1;
4448                                         if (test_bit(BTRFS_ROOT_REF_COWS,
4449                                                      &root->state))
4450                                                 inode_sub_bytes(inode, num_dec);
4451                                 }
4452                         }
4453                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
4454                         /*
4455                          * we can't truncate inline items that have had
4456                          * special encodings
4457                          */
4458                         if (!del_item &&
4459                             btrfs_file_extent_encryption(leaf, fi) == 0 &&
4460                             btrfs_file_extent_other_encoding(leaf, fi) == 0) {
4461
4462                                 /*
4463                                  * Need to release path in order to truncate a
4464                                  * compressed extent. So delete any accumulated
4465                                  * extent items so far.
4466                                  */
4467                                 if (btrfs_file_extent_compression(leaf, fi) !=
4468                                     BTRFS_COMPRESS_NONE && pending_del_nr) {
4469                                         err = btrfs_del_items(trans, root, path,
4470                                                               pending_del_slot,
4471                                                               pending_del_nr);
4472                                         if (err) {
4473                                                 btrfs_abort_transaction(trans,
4474                                                                         root,
4475                                                                         err);
4476                                                 goto error;
4477                                         }
4478                                         pending_del_nr = 0;
4479                                 }
4480
4481                                 err = truncate_inline_extent(inode, path,
4482                                                              &found_key,
4483                                                              item_end,
4484                                                              new_size);
4485                                 if (err) {
4486                                         btrfs_abort_transaction(trans,
4487                                                                 root, err);
4488                                         goto error;
4489                                 }
4490                         } else if (test_bit(BTRFS_ROOT_REF_COWS,
4491                                             &root->state)) {
4492                                 inode_sub_bytes(inode, item_end + 1 - new_size);
4493                         }
4494                 }
4495 delete:
4496                 if (del_item) {
4497                         if (!pending_del_nr) {
4498                                 /* no pending yet, add ourselves */
4499                                 pending_del_slot = path->slots[0];
4500                                 pending_del_nr = 1;
4501                         } else if (pending_del_nr &&
4502                                    path->slots[0] + 1 == pending_del_slot) {
4503                                 /* hop on the pending chunk */
4504                                 pending_del_nr++;
4505                                 pending_del_slot = path->slots[0];
4506                         } else {
4507                                 BUG();
4508                         }
4509                 } else {
4510                         break;
4511                 }
4512                 should_throttle = 0;
4513
4514                 if (found_extent &&
4515                     (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
4516                      root == root->fs_info->tree_root)) {
4517                         btrfs_set_path_blocking(path);
4518                         bytes_deleted += extent_num_bytes;
4519                         ret = btrfs_free_extent(trans, root, extent_start,
4520                                                 extent_num_bytes, 0,
4521                                                 btrfs_header_owner(leaf),
4522                                                 ino, extent_offset);
4523                         BUG_ON(ret);
4524                         if (btrfs_should_throttle_delayed_refs(trans, root))
4525                                 btrfs_async_run_delayed_refs(root,
4526                                         trans->delayed_ref_updates * 2, 0);
4527                         if (be_nice) {
4528                                 if (truncate_space_check(trans, root,
4529                                                          extent_num_bytes)) {
4530                                         should_end = 1;
4531                                 }
4532                                 if (btrfs_should_throttle_delayed_refs(trans,
4533                                                                        root)) {
4534                                         should_throttle = 1;
4535                                 }
4536                         }
4537                 }
4538
4539                 if (found_type == BTRFS_INODE_ITEM_KEY)
4540                         break;
4541
4542                 if (path->slots[0] == 0 ||
4543                     path->slots[0] != pending_del_slot ||
4544                     should_throttle || should_end) {
4545                         if (pending_del_nr) {
4546                                 ret = btrfs_del_items(trans, root, path,
4547                                                 pending_del_slot,
4548                                                 pending_del_nr);
4549                                 if (ret) {
4550                                         btrfs_abort_transaction(trans,
4551                                                                 root, ret);
4552                                         goto error;
4553                                 }
4554                                 pending_del_nr = 0;
4555                         }
4556                         btrfs_release_path(path);
4557                         if (should_throttle) {
4558                                 unsigned long updates = trans->delayed_ref_updates;
4559                                 if (updates) {
4560                                         trans->delayed_ref_updates = 0;
4561                                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4562                                         if (ret && !err)
4563                                                 err = ret;
4564                                 }
4565                         }
4566                         /*
4567                          * if we failed to refill our space rsv, bail out
4568                          * and let the transaction restart
4569                          */
4570                         if (should_end) {
4571                                 err = -EAGAIN;
4572                                 goto error;
4573                         }
4574                         goto search_again;
4575                 } else {
4576                         path->slots[0]--;
4577                 }
4578         }
4579 out:
4580         if (pending_del_nr) {
4581                 ret = btrfs_del_items(trans, root, path, pending_del_slot,
4582                                       pending_del_nr);
4583                 if (ret)
4584                         btrfs_abort_transaction(trans, root, ret);
4585         }
4586 error:
4587         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID)
4588                 btrfs_ordered_update_i_size(inode, last_size, NULL);
4589
4590         btrfs_free_path(path);
4591
4592         if (be_nice && bytes_deleted > SZ_32M) {
4593                 unsigned long updates = trans->delayed_ref_updates;
4594                 if (updates) {
4595                         trans->delayed_ref_updates = 0;
4596                         ret = btrfs_run_delayed_refs(trans, root, updates * 2);
4597                         if (ret && !err)
4598                                 err = ret;
4599                 }
4600         }
4601         return err;
4602 }
4603
4604 /*
4605  * btrfs_truncate_page - read, zero a chunk and write a page
4606  * @inode - inode that we're zeroing
4607  * @from - the offset to start zeroing
4608  * @len - the length to zero, 0 to zero the entire range respective to the
4609  *      offset
4610  * @front - zero up to the offset instead of from the offset on
4611  *
4612  * This will find the page for the "from" offset and cow the page and zero the
4613  * part we want to zero.  This is used with truncate and hole punching.
4614  */
4615 int btrfs_truncate_page(struct inode *inode, loff_t from, loff_t len,
4616                         int front)
4617 {
4618         struct address_space *mapping = inode->i_mapping;
4619         struct btrfs_root *root = BTRFS_I(inode)->root;
4620         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4621         struct btrfs_ordered_extent *ordered;
4622         struct extent_state *cached_state = NULL;
4623         char *kaddr;
4624         u32 blocksize = root->sectorsize;
4625         pgoff_t index = from >> PAGE_CACHE_SHIFT;
4626         unsigned offset = from & (PAGE_CACHE_SIZE-1);
4627         struct page *page;
4628         gfp_t mask = btrfs_alloc_write_mask(mapping);
4629         int ret = 0;
4630         u64 page_start;
4631         u64 page_end;
4632
4633         if ((offset & (blocksize - 1)) == 0 &&
4634             (!len || ((len & (blocksize - 1)) == 0)))
4635                 goto out;
4636         ret = btrfs_delalloc_reserve_space(inode,
4637                         round_down(from, PAGE_CACHE_SIZE), PAGE_CACHE_SIZE);
4638         if (ret)
4639                 goto out;
4640
4641 again:
4642         page = find_or_create_page(mapping, index, mask);
4643         if (!page) {
4644                 btrfs_delalloc_release_space(inode,
4645                                 round_down(from, PAGE_CACHE_SIZE),
4646                                 PAGE_CACHE_SIZE);
4647                 ret = -ENOMEM;
4648                 goto out;
4649         }
4650
4651         page_start = page_offset(page);
4652         page_end = page_start + PAGE_CACHE_SIZE - 1;
4653
4654         if (!PageUptodate(page)) {
4655                 ret = btrfs_readpage(NULL, page);
4656                 lock_page(page);
4657                 if (page->mapping != mapping) {
4658                         unlock_page(page);
4659                         page_cache_release(page);
4660                         goto again;
4661                 }
4662                 if (!PageUptodate(page)) {
4663                         ret = -EIO;
4664                         goto out_unlock;
4665                 }
4666         }
4667         wait_on_page_writeback(page);
4668
4669         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
4670         set_page_extent_mapped(page);
4671
4672         ordered = btrfs_lookup_ordered_extent(inode, page_start);
4673         if (ordered) {
4674                 unlock_extent_cached(io_tree, page_start, page_end,
4675                                      &cached_state, GFP_NOFS);
4676                 unlock_page(page);
4677                 page_cache_release(page);
4678                 btrfs_start_ordered_extent(inode, ordered, 1);
4679                 btrfs_put_ordered_extent(ordered);
4680                 goto again;
4681         }
4682
4683         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
4684                           EXTENT_DIRTY | EXTENT_DELALLOC |
4685                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
4686                           0, 0, &cached_state, GFP_NOFS);
4687
4688         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
4689                                         &cached_state);
4690         if (ret) {
4691                 unlock_extent_cached(io_tree, page_start, page_end,
4692                                      &cached_state, GFP_NOFS);
4693                 goto out_unlock;
4694         }
4695
4696         if (offset != PAGE_CACHE_SIZE) {
4697                 if (!len)
4698                         len = PAGE_CACHE_SIZE - offset;
4699                 kaddr = kmap(page);
4700                 if (front)
4701                         memset(kaddr, 0, offset);
4702                 else
4703                         memset(kaddr + offset, 0, len);
4704                 flush_dcache_page(page);
4705                 kunmap(page);
4706         }
4707         ClearPageChecked(page);
4708         set_page_dirty(page);
4709         unlock_extent_cached(io_tree, page_start, page_end, &cached_state,
4710                              GFP_NOFS);
4711
4712 out_unlock:
4713         if (ret)
4714                 btrfs_delalloc_release_space(inode, page_start,
4715                                              PAGE_CACHE_SIZE);
4716         unlock_page(page);
4717         page_cache_release(page);
4718 out:
4719         return ret;
4720 }
4721
4722 static int maybe_insert_hole(struct btrfs_root *root, struct inode *inode,
4723                              u64 offset, u64 len)
4724 {
4725         struct btrfs_trans_handle *trans;
4726         int ret;
4727
4728         /*
4729          * Still need to make sure the inode looks like it's been updated so
4730          * that any holes get logged if we fsync.
4731          */
4732         if (btrfs_fs_incompat(root->fs_info, NO_HOLES)) {
4733                 BTRFS_I(inode)->last_trans = root->fs_info->generation;
4734                 BTRFS_I(inode)->last_sub_trans = root->log_transid;
4735                 BTRFS_I(inode)->last_log_commit = root->last_log_commit;
4736                 return 0;
4737         }
4738
4739         /*
4740          * 1 - for the one we're dropping
4741          * 1 - for the one we're adding
4742          * 1 - for updating the inode.
4743          */
4744         trans = btrfs_start_transaction(root, 3);
4745         if (IS_ERR(trans))
4746                 return PTR_ERR(trans);
4747
4748         ret = btrfs_drop_extents(trans, root, inode, offset, offset + len, 1);
4749         if (ret) {
4750                 btrfs_abort_transaction(trans, root, ret);
4751                 btrfs_end_transaction(trans, root);
4752                 return ret;
4753         }
4754
4755         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
4756                                        0, 0, len, 0, len, 0, 0, 0);
4757         if (ret)
4758                 btrfs_abort_transaction(trans, root, ret);
4759         else
4760                 btrfs_update_inode(trans, root, inode);
4761         btrfs_end_transaction(trans, root);
4762         return ret;
4763 }
4764
4765 /*
4766  * This function puts in dummy file extents for the area we're creating a hole
4767  * for.  So if we are truncating this file to a larger size we need to insert
4768  * these file extents so that btrfs_get_extent will return a EXTENT_MAP_HOLE for
4769  * the range between oldsize and size
4770  */
4771 int btrfs_cont_expand(struct inode *inode, loff_t oldsize, loff_t size)
4772 {
4773         struct btrfs_root *root = BTRFS_I(inode)->root;
4774         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
4775         struct extent_map *em = NULL;
4776         struct extent_state *cached_state = NULL;
4777         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
4778         u64 hole_start = ALIGN(oldsize, root->sectorsize);
4779         u64 block_end = ALIGN(size, root->sectorsize);
4780         u64 last_byte;
4781         u64 cur_offset;
4782         u64 hole_size;
4783         int err = 0;
4784
4785         /*
4786          * If our size started in the middle of a page we need to zero out the
4787          * rest of the page before we expand the i_size, otherwise we could
4788          * expose stale data.
4789          */
4790         err = btrfs_truncate_page(inode, oldsize, 0, 0);
4791         if (err)
4792                 return err;
4793
4794         if (size <= hole_start)
4795                 return 0;
4796
4797         while (1) {
4798                 struct btrfs_ordered_extent *ordered;
4799
4800                 lock_extent_bits(io_tree, hole_start, block_end - 1,
4801                                  &cached_state);
4802                 ordered = btrfs_lookup_ordered_range(inode, hole_start,
4803                                                      block_end - hole_start);
4804                 if (!ordered)
4805                         break;
4806                 unlock_extent_cached(io_tree, hole_start, block_end - 1,
4807                                      &cached_state, GFP_NOFS);
4808                 btrfs_start_ordered_extent(inode, ordered, 1);
4809                 btrfs_put_ordered_extent(ordered);
4810         }
4811
4812         cur_offset = hole_start;
4813         while (1) {
4814                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
4815                                 block_end - cur_offset, 0);
4816                 if (IS_ERR(em)) {
4817                         err = PTR_ERR(em);
4818                         em = NULL;
4819                         break;
4820                 }
4821                 last_byte = min(extent_map_end(em), block_end);
4822                 last_byte = ALIGN(last_byte , root->sectorsize);
4823                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags)) {
4824                         struct extent_map *hole_em;
4825                         hole_size = last_byte - cur_offset;
4826
4827                         err = maybe_insert_hole(root, inode, cur_offset,
4828                                                 hole_size);
4829                         if (err)
4830                                 break;
4831                         btrfs_drop_extent_cache(inode, cur_offset,
4832                                                 cur_offset + hole_size - 1, 0);
4833                         hole_em = alloc_extent_map();
4834                         if (!hole_em) {
4835                                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
4836                                         &BTRFS_I(inode)->runtime_flags);
4837                                 goto next;
4838                         }
4839                         hole_em->start = cur_offset;
4840                         hole_em->len = hole_size;
4841                         hole_em->orig_start = cur_offset;
4842
4843                         hole_em->block_start = EXTENT_MAP_HOLE;
4844                         hole_em->block_len = 0;
4845                         hole_em->orig_block_len = 0;
4846                         hole_em->ram_bytes = hole_size;
4847                         hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
4848                         hole_em->compress_type = BTRFS_COMPRESS_NONE;
4849                         hole_em->generation = root->fs_info->generation;
4850
4851                         while (1) {
4852                                 write_lock(&em_tree->lock);
4853                                 err = add_extent_mapping(em_tree, hole_em, 1);
4854                                 write_unlock(&em_tree->lock);
4855                                 if (err != -EEXIST)
4856                                         break;
4857                                 btrfs_drop_extent_cache(inode, cur_offset,
4858                                                         cur_offset +
4859                                                         hole_size - 1, 0);
4860                         }
4861                         free_extent_map(hole_em);
4862                 }
4863 next:
4864                 free_extent_map(em);
4865                 em = NULL;
4866                 cur_offset = last_byte;
4867                 if (cur_offset >= block_end)
4868                         break;
4869         }
4870         free_extent_map(em);
4871         unlock_extent_cached(io_tree, hole_start, block_end - 1, &cached_state,
4872                              GFP_NOFS);
4873         return err;
4874 }
4875
4876 static int wait_snapshoting_atomic_t(atomic_t *a)
4877 {
4878         schedule();
4879         return 0;
4880 }
4881
4882 static void wait_for_snapshot_creation(struct btrfs_root *root)
4883 {
4884         while (true) {
4885                 int ret;
4886
4887                 ret = btrfs_start_write_no_snapshoting(root);
4888                 if (ret)
4889                         break;
4890                 wait_on_atomic_t(&root->will_be_snapshoted,
4891                                  wait_snapshoting_atomic_t,
4892                                  TASK_UNINTERRUPTIBLE);
4893         }
4894 }
4895
4896 static int btrfs_setsize(struct inode *inode, struct iattr *attr)
4897 {
4898         struct btrfs_root *root = BTRFS_I(inode)->root;
4899         struct btrfs_trans_handle *trans;
4900         loff_t oldsize = i_size_read(inode);
4901         loff_t newsize = attr->ia_size;
4902         int mask = attr->ia_valid;
4903         int ret;
4904
4905         /*
4906          * The regular truncate() case without ATTR_CTIME and ATTR_MTIME is a
4907          * special case where we need to update the times despite not having
4908          * these flags set.  For all other operations the VFS set these flags
4909          * explicitly if it wants a timestamp update.
4910          */
4911         if (newsize != oldsize) {
4912                 inode_inc_iversion(inode);
4913                 if (!(mask & (ATTR_CTIME | ATTR_MTIME)))
4914                         inode->i_ctime = inode->i_mtime =
4915                                 current_fs_time(inode->i_sb);
4916         }
4917
4918         if (newsize > oldsize) {
4919                 truncate_pagecache(inode, newsize);
4920                 /*
4921                  * Don't do an expanding truncate while snapshoting is ongoing.
4922                  * This is to ensure the snapshot captures a fully consistent
4923                  * state of this file - if the snapshot captures this expanding
4924                  * truncation, it must capture all writes that happened before
4925                  * this truncation.
4926                  */
4927                 wait_for_snapshot_creation(root);
4928                 ret = btrfs_cont_expand(inode, oldsize, newsize);
4929                 if (ret) {
4930                         btrfs_end_write_no_snapshoting(root);
4931                         return ret;
4932                 }
4933
4934                 trans = btrfs_start_transaction(root, 1);
4935                 if (IS_ERR(trans)) {
4936                         btrfs_end_write_no_snapshoting(root);
4937                         return PTR_ERR(trans);
4938                 }
4939
4940                 i_size_write(inode, newsize);
4941                 btrfs_ordered_update_i_size(inode, i_size_read(inode), NULL);
4942                 ret = btrfs_update_inode(trans, root, inode);
4943                 btrfs_end_write_no_snapshoting(root);
4944                 btrfs_end_transaction(trans, root);
4945         } else {
4946
4947                 /*
4948                  * We're truncating a file that used to have good data down to
4949                  * zero. Make sure it gets into the ordered flush list so that
4950                  * any new writes get down to disk quickly.
4951                  */
4952                 if (newsize == 0)
4953                         set_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
4954                                 &BTRFS_I(inode)->runtime_flags);
4955
4956                 /*
4957                  * 1 for the orphan item we're going to add
4958                  * 1 for the orphan item deletion.
4959                  */
4960                 trans = btrfs_start_transaction(root, 2);
4961                 if (IS_ERR(trans))
4962                         return PTR_ERR(trans);
4963
4964                 /*
4965                  * We need to do this in case we fail at _any_ point during the
4966                  * actual truncate.  Once we do the truncate_setsize we could
4967                  * invalidate pages which forces any outstanding ordered io to
4968                  * be instantly completed which will give us extents that need
4969                  * to be truncated.  If we fail to get an orphan inode down we
4970                  * could have left over extents that were never meant to live,
4971                  * so we need to garuntee from this point on that everything
4972                  * will be consistent.
4973                  */
4974                 ret = btrfs_orphan_add(trans, inode);
4975                 btrfs_end_transaction(trans, root);
4976                 if (ret)
4977                         return ret;
4978
4979                 /* we don't support swapfiles, so vmtruncate shouldn't fail */
4980                 truncate_setsize(inode, newsize);
4981
4982                 /* Disable nonlocked read DIO to avoid the end less truncate */
4983                 btrfs_inode_block_unlocked_dio(inode);
4984                 inode_dio_wait(inode);
4985                 btrfs_inode_resume_unlocked_dio(inode);
4986
4987                 ret = btrfs_truncate(inode);
4988                 if (ret && inode->i_nlink) {
4989                         int err;
4990
4991                         /*
4992                          * failed to truncate, disk_i_size is only adjusted down
4993                          * as we remove extents, so it should represent the true
4994                          * size of the inode, so reset the in memory size and
4995                          * delete our orphan entry.
4996                          */
4997                         trans = btrfs_join_transaction(root);
4998                         if (IS_ERR(trans)) {
4999                                 btrfs_orphan_del(NULL, inode);
5000                                 return ret;
5001                         }
5002                         i_size_write(inode, BTRFS_I(inode)->disk_i_size);
5003                         err = btrfs_orphan_del(trans, inode);
5004                         if (err)
5005                                 btrfs_abort_transaction(trans, root, err);
5006                         btrfs_end_transaction(trans, root);
5007                 }
5008         }
5009
5010         return ret;
5011 }
5012
5013 static int btrfs_setattr(struct dentry *dentry, struct iattr *attr)
5014 {
5015         struct inode *inode = d_inode(dentry);
5016         struct btrfs_root *root = BTRFS_I(inode)->root;
5017         int err;
5018
5019         if (btrfs_root_readonly(root))
5020                 return -EROFS;
5021
5022         err = inode_change_ok(inode, attr);
5023         if (err)
5024                 return err;
5025
5026         if (S_ISREG(inode->i_mode) && (attr->ia_valid & ATTR_SIZE)) {
5027                 err = btrfs_setsize(inode, attr);
5028                 if (err)
5029                         return err;
5030         }
5031
5032         if (attr->ia_valid) {
5033                 setattr_copy(inode, attr);
5034                 inode_inc_iversion(inode);
5035                 err = btrfs_dirty_inode(inode);
5036
5037                 if (!err && attr->ia_valid & ATTR_MODE)
5038                         err = posix_acl_chmod(inode, inode->i_mode);
5039         }
5040
5041         return err;
5042 }
5043
5044 /*
5045  * While truncating the inode pages during eviction, we get the VFS calling
5046  * btrfs_invalidatepage() against each page of the inode. This is slow because
5047  * the calls to btrfs_invalidatepage() result in a huge amount of calls to
5048  * lock_extent_bits() and clear_extent_bit(), which keep merging and splitting
5049  * extent_state structures over and over, wasting lots of time.
5050  *
5051  * Therefore if the inode is being evicted, let btrfs_invalidatepage() skip all
5052  * those expensive operations on a per page basis and do only the ordered io
5053  * finishing, while we release here the extent_map and extent_state structures,
5054  * without the excessive merging and splitting.
5055  */
5056 static void evict_inode_truncate_pages(struct inode *inode)
5057 {
5058         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
5059         struct extent_map_tree *map_tree = &BTRFS_I(inode)->extent_tree;
5060         struct rb_node *node;
5061
5062         ASSERT(inode->i_state & I_FREEING);
5063         truncate_inode_pages_final(&inode->i_data);
5064
5065         write_lock(&map_tree->lock);
5066         while (!RB_EMPTY_ROOT(&map_tree->map)) {
5067                 struct extent_map *em;
5068
5069                 node = rb_first(&map_tree->map);
5070                 em = rb_entry(node, struct extent_map, rb_node);
5071                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
5072                 clear_bit(EXTENT_FLAG_LOGGING, &em->flags);
5073                 remove_extent_mapping(map_tree, em);
5074                 free_extent_map(em);
5075                 if (need_resched()) {
5076                         write_unlock(&map_tree->lock);
5077                         cond_resched();
5078                         write_lock(&map_tree->lock);
5079                 }
5080         }
5081         write_unlock(&map_tree->lock);
5082
5083         /*
5084          * Keep looping until we have no more ranges in the io tree.
5085          * We can have ongoing bios started by readpages (called from readahead)
5086          * that have their endio callback (extent_io.c:end_bio_extent_readpage)
5087          * still in progress (unlocked the pages in the bio but did not yet
5088          * unlocked the ranges in the io tree). Therefore this means some
5089          * ranges can still be locked and eviction started because before
5090          * submitting those bios, which are executed by a separate task (work
5091          * queue kthread), inode references (inode->i_count) were not taken
5092          * (which would be dropped in the end io callback of each bio).
5093          * Therefore here we effectively end up waiting for those bios and
5094          * anyone else holding locked ranges without having bumped the inode's
5095          * reference count - if we don't do it, when they access the inode's
5096          * io_tree to unlock a range it may be too late, leading to an
5097          * use-after-free issue.
5098          */
5099         spin_lock(&io_tree->lock);
5100         while (!RB_EMPTY_ROOT(&io_tree->state)) {
5101                 struct extent_state *state;
5102                 struct extent_state *cached_state = NULL;
5103                 u64 start;
5104                 u64 end;
5105
5106                 node = rb_first(&io_tree->state);
5107                 state = rb_entry(node, struct extent_state, rb_node);
5108                 start = state->start;
5109                 end = state->end;
5110                 spin_unlock(&io_tree->lock);
5111
5112                 lock_extent_bits(io_tree, start, end, &cached_state);
5113
5114                 /*
5115                  * If still has DELALLOC flag, the extent didn't reach disk,
5116                  * and its reserved space won't be freed by delayed_ref.
5117                  * So we need to free its reserved space here.
5118                  * (Refer to comment in btrfs_invalidatepage, case 2)
5119                  *
5120                  * Note, end is the bytenr of last byte, so we need + 1 here.
5121                  */
5122                 if (state->state & EXTENT_DELALLOC)
5123                         btrfs_qgroup_free_data(inode, start, end - start + 1);
5124
5125                 clear_extent_bit(io_tree, start, end,
5126                                  EXTENT_LOCKED | EXTENT_DIRTY |
5127                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
5128                                  EXTENT_DEFRAG, 1, 1,
5129                                  &cached_state, GFP_NOFS);
5130
5131                 cond_resched();
5132                 spin_lock(&io_tree->lock);
5133         }
5134         spin_unlock(&io_tree->lock);
5135 }
5136
5137 void btrfs_evict_inode(struct inode *inode)
5138 {
5139         struct btrfs_trans_handle *trans;
5140         struct btrfs_root *root = BTRFS_I(inode)->root;
5141         struct btrfs_block_rsv *rsv, *global_rsv;
5142         int steal_from_global = 0;
5143         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
5144         int ret;
5145
5146         trace_btrfs_inode_evict(inode);
5147
5148         evict_inode_truncate_pages(inode);
5149
5150         if (inode->i_nlink &&
5151             ((btrfs_root_refs(&root->root_item) != 0 &&
5152               root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID) ||
5153              btrfs_is_free_space_inode(inode)))
5154                 goto no_delete;
5155
5156         if (is_bad_inode(inode)) {
5157                 btrfs_orphan_del(NULL, inode);
5158                 goto no_delete;
5159         }
5160         /* do we really want it for ->i_nlink > 0 and zero btrfs_root_refs? */
5161         if (!special_file(inode->i_mode))
5162                 btrfs_wait_ordered_range(inode, 0, (u64)-1);
5163
5164         btrfs_free_io_failure_record(inode, 0, (u64)-1);
5165
5166         if (root->fs_info->log_root_recovering) {
5167                 BUG_ON(test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
5168                                  &BTRFS_I(inode)->runtime_flags));
5169                 goto no_delete;
5170         }
5171
5172         if (inode->i_nlink > 0) {
5173                 BUG_ON(btrfs_root_refs(&root->root_item) != 0 &&
5174                        root->root_key.objectid != BTRFS_ROOT_TREE_OBJECTID);
5175                 goto no_delete;
5176         }
5177
5178         ret = btrfs_commit_inode_delayed_inode(inode);
5179         if (ret) {
5180                 btrfs_orphan_del(NULL, inode);
5181                 goto no_delete;
5182         }
5183
5184         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
5185         if (!rsv) {
5186                 btrfs_orphan_del(NULL, inode);
5187                 goto no_delete;
5188         }
5189         rsv->size = min_size;
5190         rsv->failfast = 1;
5191         global_rsv = &root->fs_info->global_block_rsv;
5192
5193         btrfs_i_size_write(inode, 0);
5194
5195         /*
5196          * This is a bit simpler than btrfs_truncate since we've already
5197          * reserved our space for our orphan item in the unlink, so we just
5198          * need to reserve some slack space in case we add bytes and update
5199          * inode item when doing the truncate.
5200          */
5201         while (1) {
5202                 ret = btrfs_block_rsv_refill(root, rsv, min_size,
5203                                              BTRFS_RESERVE_FLUSH_LIMIT);
5204
5205                 /*
5206                  * Try and steal from the global reserve since we will
5207                  * likely not use this space anyway, we want to try as
5208                  * hard as possible to get this to work.
5209                  */
5210                 if (ret)
5211                         steal_from_global++;
5212                 else
5213                         steal_from_global = 0;
5214                 ret = 0;
5215
5216                 /*
5217                  * steal_from_global == 0: we reserved stuff, hooray!
5218                  * steal_from_global == 1: we didn't reserve stuff, boo!
5219                  * steal_from_global == 2: we've committed, still not a lot of
5220                  * room but maybe we'll have room in the global reserve this
5221                  * time.
5222                  * steal_from_global == 3: abandon all hope!
5223                  */
5224                 if (steal_from_global > 2) {
5225                         btrfs_warn(root->fs_info,
5226                                 "Could not get space for a delete, will truncate on mount %d",
5227                                 ret);
5228                         btrfs_orphan_del(NULL, inode);
5229                         btrfs_free_block_rsv(root, rsv);
5230                         goto no_delete;
5231                 }
5232
5233                 trans = btrfs_join_transaction(root);
5234                 if (IS_ERR(trans)) {
5235                         btrfs_orphan_del(NULL, inode);
5236                         btrfs_free_block_rsv(root, rsv);
5237                         goto no_delete;
5238                 }
5239
5240                 /*
5241                  * We can't just steal from the global reserve, we need tomake
5242                  * sure there is room to do it, if not we need to commit and try
5243                  * again.
5244                  */
5245                 if (steal_from_global) {
5246                         if (!btrfs_check_space_for_delayed_refs(trans, root))
5247                                 ret = btrfs_block_rsv_migrate(global_rsv, rsv,
5248                                                               min_size);
5249                         else
5250                                 ret = -ENOSPC;
5251                 }
5252
5253                 /*
5254                  * Couldn't steal from the global reserve, we have too much
5255                  * pending stuff built up, commit the transaction and try it
5256                  * again.
5257                  */
5258                 if (ret) {
5259                         ret = btrfs_commit_transaction(trans, root);
5260                         if (ret) {
5261                                 btrfs_orphan_del(NULL, inode);
5262                                 btrfs_free_block_rsv(root, rsv);
5263                                 goto no_delete;
5264                         }
5265                         continue;
5266                 } else {
5267                         steal_from_global = 0;
5268                 }
5269
5270                 trans->block_rsv = rsv;
5271
5272                 ret = btrfs_truncate_inode_items(trans, root, inode, 0, 0);
5273                 if (ret != -ENOSPC && ret != -EAGAIN)
5274                         break;
5275
5276                 trans->block_rsv = &root->fs_info->trans_block_rsv;
5277                 btrfs_end_transaction(trans, root);
5278                 trans = NULL;
5279                 btrfs_btree_balance_dirty(root);
5280         }
5281
5282         btrfs_free_block_rsv(root, rsv);
5283
5284         /*
5285          * Errors here aren't a big deal, it just means we leave orphan items
5286          * in the tree.  They will be cleaned up on the next mount.
5287          */
5288         if (ret == 0) {
5289                 trans->block_rsv = root->orphan_block_rsv;
5290                 btrfs_orphan_del(trans, inode);
5291         } else {
5292                 btrfs_orphan_del(NULL, inode);
5293         }
5294
5295         trans->block_rsv = &root->fs_info->trans_block_rsv;
5296         if (!(root == root->fs_info->tree_root ||
5297               root->root_key.objectid == BTRFS_TREE_RELOC_OBJECTID))
5298                 btrfs_return_ino(root, btrfs_ino(inode));
5299
5300         btrfs_end_transaction(trans, root);
5301         btrfs_btree_balance_dirty(root);
5302 no_delete:
5303         btrfs_remove_delayed_node(inode);
5304         clear_inode(inode);
5305 }
5306
5307 /*
5308  * this returns the key found in the dir entry in the location pointer.
5309  * If no dir entries were found, location->objectid is 0.
5310  */
5311 static int btrfs_inode_by_name(struct inode *dir, struct dentry *dentry,
5312                                struct btrfs_key *location)
5313 {
5314         const char *name = dentry->d_name.name;
5315         int namelen = dentry->d_name.len;
5316         struct btrfs_dir_item *di;
5317         struct btrfs_path *path;
5318         struct btrfs_root *root = BTRFS_I(dir)->root;
5319         int ret = 0;
5320
5321         path = btrfs_alloc_path();
5322         if (!path)
5323                 return -ENOMEM;
5324
5325         di = btrfs_lookup_dir_item(NULL, root, path, btrfs_ino(dir), name,
5326                                     namelen, 0);
5327         if (IS_ERR(di))
5328                 ret = PTR_ERR(di);
5329
5330         if (IS_ERR_OR_NULL(di))
5331                 goto out_err;
5332
5333         btrfs_dir_item_key_to_cpu(path->nodes[0], di, location);
5334 out:
5335         btrfs_free_path(path);
5336         return ret;
5337 out_err:
5338         location->objectid = 0;
5339         goto out;
5340 }
5341
5342 /*
5343  * when we hit a tree root in a directory, the btrfs part of the inode
5344  * needs to be changed to reflect the root directory of the tree root.  This
5345  * is kind of like crossing a mount point.
5346  */
5347 static int fixup_tree_root_location(struct btrfs_root *root,
5348                                     struct inode *dir,
5349                                     struct dentry *dentry,
5350                                     struct btrfs_key *location,
5351                                     struct btrfs_root **sub_root)
5352 {
5353         struct btrfs_path *path;
5354         struct btrfs_root *new_root;
5355         struct btrfs_root_ref *ref;
5356         struct extent_buffer *leaf;
5357         struct btrfs_key key;
5358         int ret;
5359         int err = 0;
5360
5361         path = btrfs_alloc_path();
5362         if (!path) {
5363                 err = -ENOMEM;
5364                 goto out;
5365         }
5366
5367         err = -ENOENT;
5368         key.objectid = BTRFS_I(dir)->root->root_key.objectid;
5369         key.type = BTRFS_ROOT_REF_KEY;
5370         key.offset = location->objectid;
5371
5372         ret = btrfs_search_slot(NULL, root->fs_info->tree_root, &key, path,
5373                                 0, 0);
5374         if (ret) {
5375                 if (ret < 0)
5376                         err = ret;
5377                 goto out;
5378         }
5379
5380         leaf = path->nodes[0];
5381         ref = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_root_ref);
5382         if (btrfs_root_ref_dirid(leaf, ref) != btrfs_ino(dir) ||
5383             btrfs_root_ref_name_len(leaf, ref) != dentry->d_name.len)
5384                 goto out;
5385
5386         ret = memcmp_extent_buffer(leaf, dentry->d_name.name,
5387                                    (unsigned long)(ref + 1),
5388                                    dentry->d_name.len);
5389         if (ret)
5390                 goto out;
5391
5392         btrfs_release_path(path);
5393
5394         new_root = btrfs_read_fs_root_no_name(root->fs_info, location);
5395         if (IS_ERR(new_root)) {
5396                 err = PTR_ERR(new_root);
5397                 goto out;
5398         }
5399
5400         *sub_root = new_root;
5401         location->objectid = btrfs_root_dirid(&new_root->root_item);
5402         location->type = BTRFS_INODE_ITEM_KEY;
5403         location->offset = 0;
5404         err = 0;
5405 out:
5406         btrfs_free_path(path);
5407         return err;
5408 }
5409
5410 static void inode_tree_add(struct inode *inode)
5411 {
5412         struct btrfs_root *root = BTRFS_I(inode)->root;
5413         struct btrfs_inode *entry;
5414         struct rb_node **p;
5415         struct rb_node *parent;
5416         struct rb_node *new = &BTRFS_I(inode)->rb_node;
5417         u64 ino = btrfs_ino(inode);
5418
5419         if (inode_unhashed(inode))
5420                 return;
5421         parent = NULL;
5422         spin_lock(&root->inode_lock);
5423         p = &root->inode_tree.rb_node;
5424         while (*p) {
5425                 parent = *p;
5426                 entry = rb_entry(parent, struct btrfs_inode, rb_node);
5427
5428                 if (ino < btrfs_ino(&entry->vfs_inode))
5429                         p = &parent->rb_left;
5430                 else if (ino > btrfs_ino(&entry->vfs_inode))
5431                         p = &parent->rb_right;
5432                 else {
5433                         WARN_ON(!(entry->vfs_inode.i_state &
5434                                   (I_WILL_FREE | I_FREEING)));
5435                         rb_replace_node(parent, new, &root->inode_tree);
5436                         RB_CLEAR_NODE(parent);
5437                         spin_unlock(&root->inode_lock);
5438                         return;
5439                 }
5440         }
5441         rb_link_node(new, parent, p);
5442         rb_insert_color(new, &root->inode_tree);
5443         spin_unlock(&root->inode_lock);
5444 }
5445
5446 static void inode_tree_del(struct inode *inode)
5447 {
5448         struct btrfs_root *root = BTRFS_I(inode)->root;
5449         int empty = 0;
5450
5451         spin_lock(&root->inode_lock);
5452         if (!RB_EMPTY_NODE(&BTRFS_I(inode)->rb_node)) {
5453                 rb_erase(&BTRFS_I(inode)->rb_node, &root->inode_tree);
5454                 RB_CLEAR_NODE(&BTRFS_I(inode)->rb_node);
5455                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5456         }
5457         spin_unlock(&root->inode_lock);
5458
5459         if (empty && btrfs_root_refs(&root->root_item) == 0) {
5460                 synchronize_srcu(&root->fs_info->subvol_srcu);
5461                 spin_lock(&root->inode_lock);
5462                 empty = RB_EMPTY_ROOT(&root->inode_tree);
5463                 spin_unlock(&root->inode_lock);
5464                 if (empty)
5465                         btrfs_add_dead_root(root);
5466         }
5467 }
5468
5469 void btrfs_invalidate_inodes(struct btrfs_root *root)
5470 {
5471         struct rb_node *node;
5472         struct rb_node *prev;
5473         struct btrfs_inode *entry;
5474         struct inode *inode;
5475         u64 objectid = 0;
5476
5477         if (!test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
5478                 WARN_ON(btrfs_root_refs(&root->root_item) != 0);
5479
5480         spin_lock(&root->inode_lock);
5481 again:
5482         node = root->inode_tree.rb_node;
5483         prev = NULL;
5484         while (node) {
5485                 prev = node;
5486                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5487
5488                 if (objectid < btrfs_ino(&entry->vfs_inode))
5489                         node = node->rb_left;
5490                 else if (objectid > btrfs_ino(&entry->vfs_inode))
5491                         node = node->rb_right;
5492                 else
5493                         break;
5494         }
5495         if (!node) {
5496                 while (prev) {
5497                         entry = rb_entry(prev, struct btrfs_inode, rb_node);
5498                         if (objectid <= btrfs_ino(&entry->vfs_inode)) {
5499                                 node = prev;
5500                                 break;
5501                         }
5502                         prev = rb_next(prev);
5503                 }
5504         }
5505         while (node) {
5506                 entry = rb_entry(node, struct btrfs_inode, rb_node);
5507                 objectid = btrfs_ino(&entry->vfs_inode) + 1;
5508                 inode = igrab(&entry->vfs_inode);
5509                 if (inode) {
5510                         spin_unlock(&root->inode_lock);
5511                         if (atomic_read(&inode->i_count) > 1)
5512                                 d_prune_aliases(inode);
5513                         /*
5514                          * btrfs_drop_inode will have it removed from
5515                          * the inode cache when its usage count
5516                          * hits zero.
5517                          */
5518                         iput(inode);
5519                         cond_resched();
5520                         spin_lock(&root->inode_lock);
5521                         goto again;
5522                 }
5523
5524                 if (cond_resched_lock(&root->inode_lock))
5525                         goto again;
5526
5527                 node = rb_next(node);
5528         }
5529         spin_unlock(&root->inode_lock);
5530 }
5531
5532 static int btrfs_init_locked_inode(struct inode *inode, void *p)
5533 {
5534         struct btrfs_iget_args *args = p;
5535         inode->i_ino = args->location->objectid;
5536         memcpy(&BTRFS_I(inode)->location, args->location,
5537                sizeof(*args->location));
5538         BTRFS_I(inode)->root = args->root;
5539         return 0;
5540 }
5541
5542 static int btrfs_find_actor(struct inode *inode, void *opaque)
5543 {
5544         struct btrfs_iget_args *args = opaque;
5545         return args->location->objectid == BTRFS_I(inode)->location.objectid &&
5546                 args->root == BTRFS_I(inode)->root;
5547 }
5548
5549 static struct inode *btrfs_iget_locked(struct super_block *s,
5550                                        struct btrfs_key *location,
5551                                        struct btrfs_root *root)
5552 {
5553         struct inode *inode;
5554         struct btrfs_iget_args args;
5555         unsigned long hashval = btrfs_inode_hash(location->objectid, root);
5556
5557         args.location = location;
5558         args.root = root;
5559
5560         inode = iget5_locked(s, hashval, btrfs_find_actor,
5561                              btrfs_init_locked_inode,
5562                              (void *)&args);
5563         return inode;
5564 }
5565
5566 /* Get an inode object given its location and corresponding root.
5567  * Returns in *is_new if the inode was read from disk
5568  */
5569 struct inode *btrfs_iget(struct super_block *s, struct btrfs_key *location,
5570                          struct btrfs_root *root, int *new)
5571 {
5572         struct inode *inode;
5573
5574         inode = btrfs_iget_locked(s, location, root);
5575         if (!inode)
5576                 return ERR_PTR(-ENOMEM);
5577
5578         if (inode->i_state & I_NEW) {
5579                 btrfs_read_locked_inode(inode);
5580                 if (!is_bad_inode(inode)) {
5581                         inode_tree_add(inode);
5582                         unlock_new_inode(inode);
5583                         if (new)
5584                                 *new = 1;
5585                 } else {
5586                         unlock_new_inode(inode);
5587                         iput(inode);
5588                         inode = ERR_PTR(-ESTALE);
5589                 }
5590         }
5591
5592         return inode;
5593 }
5594
5595 static struct inode *new_simple_dir(struct super_block *s,
5596                                     struct btrfs_key *key,
5597                                     struct btrfs_root *root)
5598 {
5599         struct inode *inode = new_inode(s);
5600
5601         if (!inode)
5602                 return ERR_PTR(-ENOMEM);
5603
5604         BTRFS_I(inode)->root = root;
5605         memcpy(&BTRFS_I(inode)->location, key, sizeof(*key));
5606         set_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags);
5607
5608         inode->i_ino = BTRFS_EMPTY_SUBVOL_DIR_OBJECTID;
5609         inode->i_op = &btrfs_dir_ro_inode_operations;
5610         inode->i_fop = &simple_dir_operations;
5611         inode->i_mode = S_IFDIR | S_IRUGO | S_IWUSR | S_IXUGO;
5612         inode->i_mtime = CURRENT_TIME;
5613         inode->i_atime = inode->i_mtime;
5614         inode->i_ctime = inode->i_mtime;
5615         BTRFS_I(inode)->i_otime = inode->i_mtime;
5616
5617         return inode;
5618 }
5619
5620 struct inode *btrfs_lookup_dentry(struct inode *dir, struct dentry *dentry)
5621 {
5622         struct inode *inode;
5623         struct btrfs_root *root = BTRFS_I(dir)->root;
5624         struct btrfs_root *sub_root = root;
5625         struct btrfs_key location;
5626         int index;
5627         int ret = 0;
5628
5629         if (dentry->d_name.len > BTRFS_NAME_LEN)
5630                 return ERR_PTR(-ENAMETOOLONG);
5631
5632         ret = btrfs_inode_by_name(dir, dentry, &location);
5633         if (ret < 0)
5634                 return ERR_PTR(ret);
5635
5636         if (location.objectid == 0)
5637                 return ERR_PTR(-ENOENT);
5638
5639         if (location.type == BTRFS_INODE_ITEM_KEY) {
5640                 inode = btrfs_iget(dir->i_sb, &location, root, NULL);
5641                 return inode;
5642         }
5643
5644         BUG_ON(location.type != BTRFS_ROOT_ITEM_KEY);
5645
5646         index = srcu_read_lock(&root->fs_info->subvol_srcu);
5647         ret = fixup_tree_root_location(root, dir, dentry,
5648                                        &location, &sub_root);
5649         if (ret < 0) {
5650                 if (ret != -ENOENT)
5651                         inode = ERR_PTR(ret);
5652                 else
5653                         inode = new_simple_dir(dir->i_sb, &location, sub_root);
5654         } else {
5655                 inode = btrfs_iget(dir->i_sb, &location, sub_root, NULL);
5656         }
5657         srcu_read_unlock(&root->fs_info->subvol_srcu, index);
5658
5659         if (!IS_ERR(inode) && root != sub_root) {
5660                 down_read(&root->fs_info->cleanup_work_sem);
5661                 if (!(inode->i_sb->s_flags & MS_RDONLY))
5662                         ret = btrfs_orphan_cleanup(sub_root);
5663                 up_read(&root->fs_info->cleanup_work_sem);
5664                 if (ret) {
5665                         iput(inode);
5666                         inode = ERR_PTR(ret);
5667                 }
5668         }
5669
5670         return inode;
5671 }
5672
5673 static int btrfs_dentry_delete(const struct dentry *dentry)
5674 {
5675         struct btrfs_root *root;
5676         struct inode *inode = d_inode(dentry);
5677
5678         if (!inode && !IS_ROOT(dentry))
5679                 inode = d_inode(dentry->d_parent);
5680
5681         if (inode) {
5682                 root = BTRFS_I(inode)->root;
5683                 if (btrfs_root_refs(&root->root_item) == 0)
5684                         return 1;
5685
5686                 if (btrfs_ino(inode) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
5687                         return 1;
5688         }
5689         return 0;
5690 }
5691
5692 static void btrfs_dentry_release(struct dentry *dentry)
5693 {
5694         kfree(dentry->d_fsdata);
5695 }
5696
5697 static struct dentry *btrfs_lookup(struct inode *dir, struct dentry *dentry,
5698                                    unsigned int flags)
5699 {
5700         struct inode *inode;
5701
5702         inode = btrfs_lookup_dentry(dir, dentry);
5703         if (IS_ERR(inode)) {
5704                 if (PTR_ERR(inode) == -ENOENT)
5705                         inode = NULL;
5706                 else
5707                         return ERR_CAST(inode);
5708         }
5709
5710         return d_splice_alias(inode, dentry);
5711 }
5712
5713 unsigned char btrfs_filetype_table[] = {
5714         DT_UNKNOWN, DT_REG, DT_DIR, DT_CHR, DT_BLK, DT_FIFO, DT_SOCK, DT_LNK
5715 };
5716
5717 static int btrfs_real_readdir(struct file *file, struct dir_context *ctx)
5718 {
5719         struct inode *inode = file_inode(file);
5720         struct btrfs_root *root = BTRFS_I(inode)->root;
5721         struct btrfs_item *item;
5722         struct btrfs_dir_item *di;
5723         struct btrfs_key key;
5724         struct btrfs_key found_key;
5725         struct btrfs_path *path;
5726         struct list_head ins_list;
5727         struct list_head del_list;
5728         int ret;
5729         struct extent_buffer *leaf;
5730         int slot;
5731         unsigned char d_type;
5732         int over = 0;
5733         u32 di_cur;
5734         u32 di_total;
5735         u32 di_len;
5736         int key_type = BTRFS_DIR_INDEX_KEY;
5737         char tmp_name[32];
5738         char *name_ptr;
5739         int name_len;
5740         int is_curr = 0;        /* ctx->pos points to the current index? */
5741
5742         /* FIXME, use a real flag for deciding about the key type */
5743         if (root->fs_info->tree_root == root)
5744                 key_type = BTRFS_DIR_ITEM_KEY;
5745
5746         if (!dir_emit_dots(file, ctx))
5747                 return 0;
5748
5749         path = btrfs_alloc_path();
5750         if (!path)
5751                 return -ENOMEM;
5752
5753         path->reada = READA_FORWARD;
5754
5755         if (key_type == BTRFS_DIR_INDEX_KEY) {
5756                 INIT_LIST_HEAD(&ins_list);
5757                 INIT_LIST_HEAD(&del_list);
5758                 btrfs_get_delayed_items(inode, &ins_list, &del_list);
5759         }
5760
5761         key.type = key_type;
5762         key.offset = ctx->pos;
5763         key.objectid = btrfs_ino(inode);
5764
5765         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
5766         if (ret < 0)
5767                 goto err;
5768
5769         while (1) {
5770                 leaf = path->nodes[0];
5771                 slot = path->slots[0];
5772                 if (slot >= btrfs_header_nritems(leaf)) {
5773                         ret = btrfs_next_leaf(root, path);
5774                         if (ret < 0)
5775                                 goto err;
5776                         else if (ret > 0)
5777                                 break;
5778                         continue;
5779                 }
5780
5781                 item = btrfs_item_nr(slot);
5782                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
5783
5784                 if (found_key.objectid != key.objectid)
5785                         break;
5786                 if (found_key.type != key_type)
5787                         break;
5788                 if (found_key.offset < ctx->pos)
5789                         goto next;
5790                 if (key_type == BTRFS_DIR_INDEX_KEY &&
5791                     btrfs_should_delete_dir_index(&del_list,
5792                                                   found_key.offset))
5793                         goto next;
5794
5795                 ctx->pos = found_key.offset;
5796                 is_curr = 1;
5797
5798                 di = btrfs_item_ptr(leaf, slot, struct btrfs_dir_item);
5799                 di_cur = 0;
5800                 di_total = btrfs_item_size(leaf, item);
5801
5802                 while (di_cur < di_total) {
5803                         struct btrfs_key location;
5804
5805                         if (verify_dir_item(root, leaf, di))
5806                                 break;
5807
5808                         name_len = btrfs_dir_name_len(leaf, di);
5809                         if (name_len <= sizeof(tmp_name)) {
5810                                 name_ptr = tmp_name;
5811                         } else {
5812                                 name_ptr = kmalloc(name_len, GFP_NOFS);
5813                                 if (!name_ptr) {
5814                                         ret = -ENOMEM;
5815                                         goto err;
5816                                 }
5817                         }
5818                         read_extent_buffer(leaf, name_ptr,
5819                                            (unsigned long)(di + 1), name_len);
5820
5821                         d_type = btrfs_filetype_table[btrfs_dir_type(leaf, di)];
5822                         btrfs_dir_item_key_to_cpu(leaf, di, &location);
5823
5824
5825                         /* is this a reference to our own snapshot? If so
5826                          * skip it.
5827                          *
5828                          * In contrast to old kernels, we insert the snapshot's
5829                          * dir item and dir index after it has been created, so
5830                          * we won't find a reference to our own snapshot. We
5831                          * still keep the following code for backward
5832                          * compatibility.
5833                          */
5834                         if (location.type == BTRFS_ROOT_ITEM_KEY &&
5835                             location.objectid == root->root_key.objectid) {
5836                                 over = 0;
5837                                 goto skip;
5838                         }
5839                         over = !dir_emit(ctx, name_ptr, name_len,
5840                                        location.objectid, d_type);
5841
5842 skip:
5843                         if (name_ptr != tmp_name)
5844                                 kfree(name_ptr);
5845
5846                         if (over)
5847                                 goto nopos;
5848                         di_len = btrfs_dir_name_len(leaf, di) +
5849                                  btrfs_dir_data_len(leaf, di) + sizeof(*di);
5850                         di_cur += di_len;
5851                         di = (struct btrfs_dir_item *)((char *)di + di_len);
5852                 }
5853 next:
5854                 path->slots[0]++;
5855         }
5856
5857         if (key_type == BTRFS_DIR_INDEX_KEY) {
5858                 if (is_curr)
5859                         ctx->pos++;
5860                 ret = btrfs_readdir_delayed_dir_index(ctx, &ins_list);
5861                 if (ret)
5862                         goto nopos;
5863         }
5864
5865         /* Reached end of directory/root. Bump pos past the last item. */
5866         ctx->pos++;
5867
5868         /*
5869          * Stop new entries from being returned after we return the last
5870          * entry.
5871          *
5872          * New directory entries are assigned a strictly increasing
5873          * offset.  This means that new entries created during readdir
5874          * are *guaranteed* to be seen in the future by that readdir.
5875          * This has broken buggy programs which operate on names as
5876          * they're returned by readdir.  Until we re-use freed offsets
5877          * we have this hack to stop new entries from being returned
5878          * under the assumption that they'll never reach this huge
5879          * offset.
5880          *
5881          * This is being careful not to overflow 32bit loff_t unless the
5882          * last entry requires it because doing so has broken 32bit apps
5883          * in the past.
5884          */
5885         if (key_type == BTRFS_DIR_INDEX_KEY) {
5886                 if (ctx->pos >= INT_MAX)
5887                         ctx->pos = LLONG_MAX;
5888                 else
5889                         ctx->pos = INT_MAX;
5890         }
5891 nopos:
5892         ret = 0;
5893 err:
5894         if (key_type == BTRFS_DIR_INDEX_KEY)
5895                 btrfs_put_delayed_items(&ins_list, &del_list);
5896         btrfs_free_path(path);
5897         return ret;
5898 }
5899
5900 int btrfs_write_inode(struct inode *inode, struct writeback_control *wbc)
5901 {
5902         struct btrfs_root *root = BTRFS_I(inode)->root;
5903         struct btrfs_trans_handle *trans;
5904         int ret = 0;
5905         bool nolock = false;
5906
5907         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5908                 return 0;
5909
5910         if (btrfs_fs_closing(root->fs_info) && btrfs_is_free_space_inode(inode))
5911                 nolock = true;
5912
5913         if (wbc->sync_mode == WB_SYNC_ALL) {
5914                 if (nolock)
5915                         trans = btrfs_join_transaction_nolock(root);
5916                 else
5917                         trans = btrfs_join_transaction(root);
5918                 if (IS_ERR(trans))
5919                         return PTR_ERR(trans);
5920                 ret = btrfs_commit_transaction(trans, root);
5921         }
5922         return ret;
5923 }
5924
5925 /*
5926  * This is somewhat expensive, updating the tree every time the
5927  * inode changes.  But, it is most likely to find the inode in cache.
5928  * FIXME, needs more benchmarking...there are no reasons other than performance
5929  * to keep or drop this code.
5930  */
5931 static int btrfs_dirty_inode(struct inode *inode)
5932 {
5933         struct btrfs_root *root = BTRFS_I(inode)->root;
5934         struct btrfs_trans_handle *trans;
5935         int ret;
5936
5937         if (test_bit(BTRFS_INODE_DUMMY, &BTRFS_I(inode)->runtime_flags))
5938                 return 0;
5939
5940         trans = btrfs_join_transaction(root);
5941         if (IS_ERR(trans))
5942                 return PTR_ERR(trans);
5943
5944         ret = btrfs_update_inode(trans, root, inode);
5945         if (ret && ret == -ENOSPC) {
5946                 /* whoops, lets try again with the full transaction */
5947                 btrfs_end_transaction(trans, root);
5948                 trans = btrfs_start_transaction(root, 1);
5949                 if (IS_ERR(trans))
5950                         return PTR_ERR(trans);
5951
5952                 ret = btrfs_update_inode(trans, root, inode);
5953         }
5954         btrfs_end_transaction(trans, root);
5955         if (BTRFS_I(inode)->delayed_node)
5956                 btrfs_balance_delayed_items(root);
5957
5958         return ret;
5959 }
5960
5961 /*
5962  * This is a copy of file_update_time.  We need this so we can return error on
5963  * ENOSPC for updating the inode in the case of file write and mmap writes.
5964  */
5965 static int btrfs_update_time(struct inode *inode, struct timespec *now,
5966                              int flags)
5967 {
5968         struct btrfs_root *root = BTRFS_I(inode)->root;
5969
5970         if (btrfs_root_readonly(root))
5971                 return -EROFS;
5972
5973         if (flags & S_VERSION)
5974                 inode_inc_iversion(inode);
5975         if (flags & S_CTIME)
5976                 inode->i_ctime = *now;
5977         if (flags & S_MTIME)
5978                 inode->i_mtime = *now;
5979         if (flags & S_ATIME)
5980                 inode->i_atime = *now;
5981         return btrfs_dirty_inode(inode);
5982 }
5983
5984 /*
5985  * find the highest existing sequence number in a directory
5986  * and then set the in-memory index_cnt variable to reflect
5987  * free sequence numbers
5988  */
5989 static int btrfs_set_inode_index_count(struct inode *inode)
5990 {
5991         struct btrfs_root *root = BTRFS_I(inode)->root;
5992         struct btrfs_key key, found_key;
5993         struct btrfs_path *path;
5994         struct extent_buffer *leaf;
5995         int ret;
5996
5997         key.objectid = btrfs_ino(inode);
5998         key.type = BTRFS_DIR_INDEX_KEY;
5999         key.offset = (u64)-1;
6000
6001         path = btrfs_alloc_path();
6002         if (!path)
6003                 return -ENOMEM;
6004
6005         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6006         if (ret < 0)
6007                 goto out;
6008         /* FIXME: we should be able to handle this */
6009         if (ret == 0)
6010                 goto out;
6011         ret = 0;
6012
6013         /*
6014          * MAGIC NUMBER EXPLANATION:
6015          * since we search a directory based on f_pos we have to start at 2
6016          * since '.' and '..' have f_pos of 0 and 1 respectively, so everybody
6017          * else has to start at 2
6018          */
6019         if (path->slots[0] == 0) {
6020                 BTRFS_I(inode)->index_cnt = 2;
6021                 goto out;
6022         }
6023
6024         path->slots[0]--;
6025
6026         leaf = path->nodes[0];
6027         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6028
6029         if (found_key.objectid != btrfs_ino(inode) ||
6030             found_key.type != BTRFS_DIR_INDEX_KEY) {
6031                 BTRFS_I(inode)->index_cnt = 2;
6032                 goto out;
6033         }
6034
6035         BTRFS_I(inode)->index_cnt = found_key.offset + 1;
6036 out:
6037         btrfs_free_path(path);
6038         return ret;
6039 }
6040
6041 /*
6042  * helper to find a free sequence number in a given directory.  This current
6043  * code is very simple, later versions will do smarter things in the btree
6044  */
6045 int btrfs_set_inode_index(struct inode *dir, u64 *index)
6046 {
6047         int ret = 0;
6048
6049         if (BTRFS_I(dir)->index_cnt == (u64)-1) {
6050                 ret = btrfs_inode_delayed_dir_index_count(dir);
6051                 if (ret) {
6052                         ret = btrfs_set_inode_index_count(dir);
6053                         if (ret)
6054                                 return ret;
6055                 }
6056         }
6057
6058         *index = BTRFS_I(dir)->index_cnt;
6059         BTRFS_I(dir)->index_cnt++;
6060
6061         return ret;
6062 }
6063
6064 static int btrfs_insert_inode_locked(struct inode *inode)
6065 {
6066         struct btrfs_iget_args args;
6067         args.location = &BTRFS_I(inode)->location;
6068         args.root = BTRFS_I(inode)->root;
6069
6070         return insert_inode_locked4(inode,
6071                    btrfs_inode_hash(inode->i_ino, BTRFS_I(inode)->root),
6072                    btrfs_find_actor, &args);
6073 }
6074
6075 static struct inode *btrfs_new_inode(struct btrfs_trans_handle *trans,
6076                                      struct btrfs_root *root,
6077                                      struct inode *dir,
6078                                      const char *name, int name_len,
6079                                      u64 ref_objectid, u64 objectid,
6080                                      umode_t mode, u64 *index)
6081 {
6082         struct inode *inode;
6083         struct btrfs_inode_item *inode_item;
6084         struct btrfs_key *location;
6085         struct btrfs_path *path;
6086         struct btrfs_inode_ref *ref;
6087         struct btrfs_key key[2];
6088         u32 sizes[2];
6089         int nitems = name ? 2 : 1;
6090         unsigned long ptr;
6091         int ret;
6092
6093         path = btrfs_alloc_path();
6094         if (!path)
6095                 return ERR_PTR(-ENOMEM);
6096
6097         inode = new_inode(root->fs_info->sb);
6098         if (!inode) {
6099                 btrfs_free_path(path);
6100                 return ERR_PTR(-ENOMEM);
6101         }
6102
6103         /*
6104          * O_TMPFILE, set link count to 0, so that after this point,
6105          * we fill in an inode item with the correct link count.
6106          */
6107         if (!name)
6108                 set_nlink(inode, 0);
6109
6110         /*
6111          * we have to initialize this early, so we can reclaim the inode
6112          * number if we fail afterwards in this function.
6113          */
6114         inode->i_ino = objectid;
6115
6116         if (dir && name) {
6117                 trace_btrfs_inode_request(dir);
6118
6119                 ret = btrfs_set_inode_index(dir, index);
6120                 if (ret) {
6121                         btrfs_free_path(path);
6122                         iput(inode);
6123                         return ERR_PTR(ret);
6124                 }
6125         } else if (dir) {
6126                 *index = 0;
6127         }
6128         /*
6129          * index_cnt is ignored for everything but a dir,
6130          * btrfs_get_inode_index_count has an explanation for the magic
6131          * number
6132          */
6133         BTRFS_I(inode)->index_cnt = 2;
6134         BTRFS_I(inode)->dir_index = *index;
6135         BTRFS_I(inode)->root = root;
6136         BTRFS_I(inode)->generation = trans->transid;
6137         inode->i_generation = BTRFS_I(inode)->generation;
6138
6139         /*
6140          * We could have gotten an inode number from somebody who was fsynced
6141          * and then removed in this same transaction, so let's just set full
6142          * sync since it will be a full sync anyway and this will blow away the
6143          * old info in the log.
6144          */
6145         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
6146
6147         key[0].objectid = objectid;
6148         key[0].type = BTRFS_INODE_ITEM_KEY;
6149         key[0].offset = 0;
6150
6151         sizes[0] = sizeof(struct btrfs_inode_item);
6152
6153         if (name) {
6154                 /*
6155                  * Start new inodes with an inode_ref. This is slightly more
6156                  * efficient for small numbers of hard links since they will
6157                  * be packed into one item. Extended refs will kick in if we
6158                  * add more hard links than can fit in the ref item.
6159                  */
6160                 key[1].objectid = objectid;
6161                 key[1].type = BTRFS_INODE_REF_KEY;
6162                 key[1].offset = ref_objectid;
6163
6164                 sizes[1] = name_len + sizeof(*ref);
6165         }
6166
6167         location = &BTRFS_I(inode)->location;
6168         location->objectid = objectid;
6169         location->offset = 0;
6170         location->type = BTRFS_INODE_ITEM_KEY;
6171
6172         ret = btrfs_insert_inode_locked(inode);
6173         if (ret < 0)
6174                 goto fail;
6175
6176         path->leave_spinning = 1;
6177         ret = btrfs_insert_empty_items(trans, root, path, key, sizes, nitems);
6178         if (ret != 0)
6179                 goto fail_unlock;
6180
6181         inode_init_owner(inode, dir, mode);
6182         inode_set_bytes(inode, 0);
6183
6184         inode->i_mtime = CURRENT_TIME;
6185         inode->i_atime = inode->i_mtime;
6186         inode->i_ctime = inode->i_mtime;
6187         BTRFS_I(inode)->i_otime = inode->i_mtime;
6188
6189         inode_item = btrfs_item_ptr(path->nodes[0], path->slots[0],
6190                                   struct btrfs_inode_item);
6191         memset_extent_buffer(path->nodes[0], 0, (unsigned long)inode_item,
6192                              sizeof(*inode_item));
6193         fill_inode_item(trans, path->nodes[0], inode_item, inode);
6194
6195         if (name) {
6196                 ref = btrfs_item_ptr(path->nodes[0], path->slots[0] + 1,
6197                                      struct btrfs_inode_ref);
6198                 btrfs_set_inode_ref_name_len(path->nodes[0], ref, name_len);
6199                 btrfs_set_inode_ref_index(path->nodes[0], ref, *index);
6200                 ptr = (unsigned long)(ref + 1);
6201                 write_extent_buffer(path->nodes[0], name, ptr, name_len);
6202         }
6203
6204         btrfs_mark_buffer_dirty(path->nodes[0]);
6205         btrfs_free_path(path);
6206
6207         btrfs_inherit_iflags(inode, dir);
6208
6209         if (S_ISREG(mode)) {
6210                 if (btrfs_test_opt(root, NODATASUM))
6211                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM;
6212                 if (btrfs_test_opt(root, NODATACOW))
6213                         BTRFS_I(inode)->flags |= BTRFS_INODE_NODATACOW |
6214                                 BTRFS_INODE_NODATASUM;
6215         }
6216
6217         inode_tree_add(inode);
6218
6219         trace_btrfs_inode_new(inode);
6220         btrfs_set_inode_last_trans(trans, inode);
6221
6222         btrfs_update_root_times(trans, root);
6223
6224         ret = btrfs_inode_inherit_props(trans, inode, dir);
6225         if (ret)
6226                 btrfs_err(root->fs_info,
6227                           "error inheriting props for ino %llu (root %llu): %d",
6228                           btrfs_ino(inode), root->root_key.objectid, ret);
6229
6230         return inode;
6231
6232 fail_unlock:
6233         unlock_new_inode(inode);
6234 fail:
6235         if (dir && name)
6236                 BTRFS_I(dir)->index_cnt--;
6237         btrfs_free_path(path);
6238         iput(inode);
6239         return ERR_PTR(ret);
6240 }
6241
6242 static inline u8 btrfs_inode_type(struct inode *inode)
6243 {
6244         return btrfs_type_by_mode[(inode->i_mode & S_IFMT) >> S_SHIFT];
6245 }
6246
6247 /*
6248  * utility function to add 'inode' into 'parent_inode' with
6249  * a give name and a given sequence number.
6250  * if 'add_backref' is true, also insert a backref from the
6251  * inode to the parent directory.
6252  */
6253 int btrfs_add_link(struct btrfs_trans_handle *trans,
6254                    struct inode *parent_inode, struct inode *inode,
6255                    const char *name, int name_len, int add_backref, u64 index)
6256 {
6257         int ret = 0;
6258         struct btrfs_key key;
6259         struct btrfs_root *root = BTRFS_I(parent_inode)->root;
6260         u64 ino = btrfs_ino(inode);
6261         u64 parent_ino = btrfs_ino(parent_inode);
6262
6263         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6264                 memcpy(&key, &BTRFS_I(inode)->root->root_key, sizeof(key));
6265         } else {
6266                 key.objectid = ino;
6267                 key.type = BTRFS_INODE_ITEM_KEY;
6268                 key.offset = 0;
6269         }
6270
6271         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6272                 ret = btrfs_add_root_ref(trans, root->fs_info->tree_root,
6273                                          key.objectid, root->root_key.objectid,
6274                                          parent_ino, index, name, name_len);
6275         } else if (add_backref) {
6276                 ret = btrfs_insert_inode_ref(trans, root, name, name_len, ino,
6277                                              parent_ino, index);
6278         }
6279
6280         /* Nothing to clean up yet */
6281         if (ret)
6282                 return ret;
6283
6284         ret = btrfs_insert_dir_item(trans, root, name, name_len,
6285                                     parent_inode, &key,
6286                                     btrfs_inode_type(inode), index);
6287         if (ret == -EEXIST || ret == -EOVERFLOW)
6288                 goto fail_dir_item;
6289         else if (ret) {
6290                 btrfs_abort_transaction(trans, root, ret);
6291                 return ret;
6292         }
6293
6294         btrfs_i_size_write(parent_inode, parent_inode->i_size +
6295                            name_len * 2);
6296         inode_inc_iversion(parent_inode);
6297         parent_inode->i_mtime = parent_inode->i_ctime = CURRENT_TIME;
6298         ret = btrfs_update_inode(trans, root, parent_inode);
6299         if (ret)
6300                 btrfs_abort_transaction(trans, root, ret);
6301         return ret;
6302
6303 fail_dir_item:
6304         if (unlikely(ino == BTRFS_FIRST_FREE_OBJECTID)) {
6305                 u64 local_index;
6306                 int err;
6307                 err = btrfs_del_root_ref(trans, root->fs_info->tree_root,
6308                                  key.objectid, root->root_key.objectid,
6309                                  parent_ino, &local_index, name, name_len);
6310
6311         } else if (add_backref) {
6312                 u64 local_index;
6313                 int err;
6314
6315                 err = btrfs_del_inode_ref(trans, root, name, name_len,
6316                                           ino, parent_ino, &local_index);
6317         }
6318         return ret;
6319 }
6320
6321 static int btrfs_add_nondir(struct btrfs_trans_handle *trans,
6322                             struct inode *dir, struct dentry *dentry,
6323                             struct inode *inode, int backref, u64 index)
6324 {
6325         int err = btrfs_add_link(trans, dir, inode,
6326                                  dentry->d_name.name, dentry->d_name.len,
6327                                  backref, index);
6328         if (err > 0)
6329                 err = -EEXIST;
6330         return err;
6331 }
6332
6333 static int btrfs_mknod(struct inode *dir, struct dentry *dentry,
6334                         umode_t mode, dev_t rdev)
6335 {
6336         struct btrfs_trans_handle *trans;
6337         struct btrfs_root *root = BTRFS_I(dir)->root;
6338         struct inode *inode = NULL;
6339         int err;
6340         int drop_inode = 0;
6341         u64 objectid;
6342         u64 index = 0;
6343
6344         /*
6345          * 2 for inode item and ref
6346          * 2 for dir items
6347          * 1 for xattr if selinux is on
6348          */
6349         trans = btrfs_start_transaction(root, 5);
6350         if (IS_ERR(trans))
6351                 return PTR_ERR(trans);
6352
6353         err = btrfs_find_free_ino(root, &objectid);
6354         if (err)
6355                 goto out_unlock;
6356
6357         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6358                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6359                                 mode, &index);
6360         if (IS_ERR(inode)) {
6361                 err = PTR_ERR(inode);
6362                 goto out_unlock;
6363         }
6364
6365         /*
6366         * If the active LSM wants to access the inode during
6367         * d_instantiate it needs these. Smack checks to see
6368         * if the filesystem supports xattrs by looking at the
6369         * ops vector.
6370         */
6371         inode->i_op = &btrfs_special_inode_operations;
6372         init_special_inode(inode, inode->i_mode, rdev);
6373
6374         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6375         if (err)
6376                 goto out_unlock_inode;
6377
6378         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6379         if (err) {
6380                 goto out_unlock_inode;
6381         } else {
6382                 btrfs_update_inode(trans, root, inode);
6383                 unlock_new_inode(inode);
6384                 d_instantiate(dentry, inode);
6385         }
6386
6387 out_unlock:
6388         btrfs_end_transaction(trans, root);
6389         btrfs_balance_delayed_items(root);
6390         btrfs_btree_balance_dirty(root);
6391         if (drop_inode) {
6392                 inode_dec_link_count(inode);
6393                 iput(inode);
6394         }
6395         return err;
6396
6397 out_unlock_inode:
6398         drop_inode = 1;
6399         unlock_new_inode(inode);
6400         goto out_unlock;
6401
6402 }
6403
6404 static int btrfs_create(struct inode *dir, struct dentry *dentry,
6405                         umode_t mode, bool excl)
6406 {
6407         struct btrfs_trans_handle *trans;
6408         struct btrfs_root *root = BTRFS_I(dir)->root;
6409         struct inode *inode = NULL;
6410         int drop_inode_on_err = 0;
6411         int err;
6412         u64 objectid;
6413         u64 index = 0;
6414
6415         /*
6416          * 2 for inode item and ref
6417          * 2 for dir items
6418          * 1 for xattr if selinux is on
6419          */
6420         trans = btrfs_start_transaction(root, 5);
6421         if (IS_ERR(trans))
6422                 return PTR_ERR(trans);
6423
6424         err = btrfs_find_free_ino(root, &objectid);
6425         if (err)
6426                 goto out_unlock;
6427
6428         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6429                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6430                                 mode, &index);
6431         if (IS_ERR(inode)) {
6432                 err = PTR_ERR(inode);
6433                 goto out_unlock;
6434         }
6435         drop_inode_on_err = 1;
6436         /*
6437         * If the active LSM wants to access the inode during
6438         * d_instantiate it needs these. Smack checks to see
6439         * if the filesystem supports xattrs by looking at the
6440         * ops vector.
6441         */
6442         inode->i_fop = &btrfs_file_operations;
6443         inode->i_op = &btrfs_file_inode_operations;
6444         inode->i_mapping->a_ops = &btrfs_aops;
6445
6446         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6447         if (err)
6448                 goto out_unlock_inode;
6449
6450         err = btrfs_update_inode(trans, root, inode);
6451         if (err)
6452                 goto out_unlock_inode;
6453
6454         err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
6455         if (err)
6456                 goto out_unlock_inode;
6457
6458         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
6459         unlock_new_inode(inode);
6460         d_instantiate(dentry, inode);
6461
6462 out_unlock:
6463         btrfs_end_transaction(trans, root);
6464         if (err && drop_inode_on_err) {
6465                 inode_dec_link_count(inode);
6466                 iput(inode);
6467         }
6468         btrfs_balance_delayed_items(root);
6469         btrfs_btree_balance_dirty(root);
6470         return err;
6471
6472 out_unlock_inode:
6473         unlock_new_inode(inode);
6474         goto out_unlock;
6475
6476 }
6477
6478 static int btrfs_link(struct dentry *old_dentry, struct inode *dir,
6479                       struct dentry *dentry)
6480 {
6481         struct btrfs_trans_handle *trans = NULL;
6482         struct btrfs_root *root = BTRFS_I(dir)->root;
6483         struct inode *inode = d_inode(old_dentry);
6484         u64 index;
6485         int err;
6486         int drop_inode = 0;
6487
6488         /* do not allow sys_link's with other subvols of the same device */
6489         if (root->objectid != BTRFS_I(inode)->root->objectid)
6490                 return -EXDEV;
6491
6492         if (inode->i_nlink >= BTRFS_LINK_MAX)
6493                 return -EMLINK;
6494
6495         err = btrfs_set_inode_index(dir, &index);
6496         if (err)
6497                 goto fail;
6498
6499         /*
6500          * 2 items for inode and inode ref
6501          * 2 items for dir items
6502          * 1 item for parent inode
6503          */
6504         trans = btrfs_start_transaction(root, 5);
6505         if (IS_ERR(trans)) {
6506                 err = PTR_ERR(trans);
6507                 trans = NULL;
6508                 goto fail;
6509         }
6510
6511         /* There are several dir indexes for this inode, clear the cache. */
6512         BTRFS_I(inode)->dir_index = 0ULL;
6513         inc_nlink(inode);
6514         inode_inc_iversion(inode);
6515         inode->i_ctime = CURRENT_TIME;
6516         ihold(inode);
6517         set_bit(BTRFS_INODE_COPY_EVERYTHING, &BTRFS_I(inode)->runtime_flags);
6518
6519         err = btrfs_add_nondir(trans, dir, dentry, inode, 1, index);
6520
6521         if (err) {
6522                 drop_inode = 1;
6523         } else {
6524                 struct dentry *parent = dentry->d_parent;
6525                 err = btrfs_update_inode(trans, root, inode);
6526                 if (err)
6527                         goto fail;
6528                 if (inode->i_nlink == 1) {
6529                         /*
6530                          * If new hard link count is 1, it's a file created
6531                          * with open(2) O_TMPFILE flag.
6532                          */
6533                         err = btrfs_orphan_del(trans, inode);
6534                         if (err)
6535                                 goto fail;
6536                 }
6537                 d_instantiate(dentry, inode);
6538                 btrfs_log_new_name(trans, inode, NULL, parent);
6539         }
6540
6541         btrfs_balance_delayed_items(root);
6542 fail:
6543         if (trans)
6544                 btrfs_end_transaction(trans, root);
6545         if (drop_inode) {
6546                 inode_dec_link_count(inode);
6547                 iput(inode);
6548         }
6549         btrfs_btree_balance_dirty(root);
6550         return err;
6551 }
6552
6553 static int btrfs_mkdir(struct inode *dir, struct dentry *dentry, umode_t mode)
6554 {
6555         struct inode *inode = NULL;
6556         struct btrfs_trans_handle *trans;
6557         struct btrfs_root *root = BTRFS_I(dir)->root;
6558         int err = 0;
6559         int drop_on_err = 0;
6560         u64 objectid = 0;
6561         u64 index = 0;
6562
6563         /*
6564          * 2 items for inode and ref
6565          * 2 items for dir items
6566          * 1 for xattr if selinux is on
6567          */
6568         trans = btrfs_start_transaction(root, 5);
6569         if (IS_ERR(trans))
6570                 return PTR_ERR(trans);
6571
6572         err = btrfs_find_free_ino(root, &objectid);
6573         if (err)
6574                 goto out_fail;
6575
6576         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
6577                                 dentry->d_name.len, btrfs_ino(dir), objectid,
6578                                 S_IFDIR | mode, &index);
6579         if (IS_ERR(inode)) {
6580                 err = PTR_ERR(inode);
6581                 goto out_fail;
6582         }
6583
6584         drop_on_err = 1;
6585         /* these must be set before we unlock the inode */
6586         inode->i_op = &btrfs_dir_inode_operations;
6587         inode->i_fop = &btrfs_dir_file_operations;
6588
6589         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
6590         if (err)
6591                 goto out_fail_inode;
6592
6593         btrfs_i_size_write(inode, 0);
6594         err = btrfs_update_inode(trans, root, inode);
6595         if (err)
6596                 goto out_fail_inode;
6597
6598         err = btrfs_add_link(trans, dir, inode, dentry->d_name.name,
6599                              dentry->d_name.len, 0, index);
6600         if (err)
6601                 goto out_fail_inode;
6602
6603         d_instantiate(dentry, inode);
6604         /*
6605          * mkdir is special.  We're unlocking after we call d_instantiate
6606          * to avoid a race with nfsd calling d_instantiate.
6607          */
6608         unlock_new_inode(inode);
6609         drop_on_err = 0;
6610
6611 out_fail:
6612         btrfs_end_transaction(trans, root);
6613         if (drop_on_err) {
6614                 inode_dec_link_count(inode);
6615                 iput(inode);
6616         }
6617         btrfs_balance_delayed_items(root);
6618         btrfs_btree_balance_dirty(root);
6619         return err;
6620
6621 out_fail_inode:
6622         unlock_new_inode(inode);
6623         goto out_fail;
6624 }
6625
6626 /* Find next extent map of a given extent map, caller needs to ensure locks */
6627 static struct extent_map *next_extent_map(struct extent_map *em)
6628 {
6629         struct rb_node *next;
6630
6631         next = rb_next(&em->rb_node);
6632         if (!next)
6633                 return NULL;
6634         return container_of(next, struct extent_map, rb_node);
6635 }
6636
6637 static struct extent_map *prev_extent_map(struct extent_map *em)
6638 {
6639         struct rb_node *prev;
6640
6641         prev = rb_prev(&em->rb_node);
6642         if (!prev)
6643                 return NULL;
6644         return container_of(prev, struct extent_map, rb_node);
6645 }
6646
6647 /* helper for btfs_get_extent.  Given an existing extent in the tree,
6648  * the existing extent is the nearest extent to map_start,
6649  * and an extent that you want to insert, deal with overlap and insert
6650  * the best fitted new extent into the tree.
6651  */
6652 static int merge_extent_mapping(struct extent_map_tree *em_tree,
6653                                 struct extent_map *existing,
6654                                 struct extent_map *em,
6655                                 u64 map_start)
6656 {
6657         struct extent_map *prev;
6658         struct extent_map *next;
6659         u64 start;
6660         u64 end;
6661         u64 start_diff;
6662
6663         BUG_ON(map_start < em->start || map_start >= extent_map_end(em));
6664
6665         if (existing->start > map_start) {
6666                 next = existing;
6667                 prev = prev_extent_map(next);
6668         } else {
6669                 prev = existing;
6670                 next = next_extent_map(prev);
6671         }
6672
6673         start = prev ? extent_map_end(prev) : em->start;
6674         start = max_t(u64, start, em->start);
6675         end = next ? next->start : extent_map_end(em);
6676         end = min_t(u64, end, extent_map_end(em));
6677         start_diff = start - em->start;
6678         em->start = start;
6679         em->len = end - start;
6680         if (em->block_start < EXTENT_MAP_LAST_BYTE &&
6681             !test_bit(EXTENT_FLAG_COMPRESSED, &em->flags)) {
6682                 em->block_start += start_diff;
6683                 em->block_len -= start_diff;
6684         }
6685         return add_extent_mapping(em_tree, em, 0);
6686 }
6687
6688 static noinline int uncompress_inline(struct btrfs_path *path,
6689                                       struct page *page,
6690                                       size_t pg_offset, u64 extent_offset,
6691                                       struct btrfs_file_extent_item *item)
6692 {
6693         int ret;
6694         struct extent_buffer *leaf = path->nodes[0];
6695         char *tmp;
6696         size_t max_size;
6697         unsigned long inline_size;
6698         unsigned long ptr;
6699         int compress_type;
6700
6701         WARN_ON(pg_offset != 0);
6702         compress_type = btrfs_file_extent_compression(leaf, item);
6703         max_size = btrfs_file_extent_ram_bytes(leaf, item);
6704         inline_size = btrfs_file_extent_inline_item_len(leaf,
6705                                         btrfs_item_nr(path->slots[0]));
6706         tmp = kmalloc(inline_size, GFP_NOFS);
6707         if (!tmp)
6708                 return -ENOMEM;
6709         ptr = btrfs_file_extent_inline_start(item);
6710
6711         read_extent_buffer(leaf, tmp, ptr, inline_size);
6712
6713         max_size = min_t(unsigned long, PAGE_CACHE_SIZE, max_size);
6714         ret = btrfs_decompress(compress_type, tmp, page,
6715                                extent_offset, inline_size, max_size);
6716         kfree(tmp);
6717         return ret;
6718 }
6719
6720 /*
6721  * a bit scary, this does extent mapping from logical file offset to the disk.
6722  * the ugly parts come from merging extents from the disk with the in-ram
6723  * representation.  This gets more complex because of the data=ordered code,
6724  * where the in-ram extents might be locked pending data=ordered completion.
6725  *
6726  * This also copies inline extents directly into the page.
6727  */
6728
6729 struct extent_map *btrfs_get_extent(struct inode *inode, struct page *page,
6730                                     size_t pg_offset, u64 start, u64 len,
6731                                     int create)
6732 {
6733         int ret;
6734         int err = 0;
6735         u64 extent_start = 0;
6736         u64 extent_end = 0;
6737         u64 objectid = btrfs_ino(inode);
6738         u32 found_type;
6739         struct btrfs_path *path = NULL;
6740         struct btrfs_root *root = BTRFS_I(inode)->root;
6741         struct btrfs_file_extent_item *item;
6742         struct extent_buffer *leaf;
6743         struct btrfs_key found_key;
6744         struct extent_map *em = NULL;
6745         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
6746         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6747         struct btrfs_trans_handle *trans = NULL;
6748         const bool new_inline = !page || create;
6749
6750 again:
6751         read_lock(&em_tree->lock);
6752         em = lookup_extent_mapping(em_tree, start, len);
6753         if (em)
6754                 em->bdev = root->fs_info->fs_devices->latest_bdev;
6755         read_unlock(&em_tree->lock);
6756
6757         if (em) {
6758                 if (em->start > start || em->start + em->len <= start)
6759                         free_extent_map(em);
6760                 else if (em->block_start == EXTENT_MAP_INLINE && page)
6761                         free_extent_map(em);
6762                 else
6763                         goto out;
6764         }
6765         em = alloc_extent_map();
6766         if (!em) {
6767                 err = -ENOMEM;
6768                 goto out;
6769         }
6770         em->bdev = root->fs_info->fs_devices->latest_bdev;
6771         em->start = EXTENT_MAP_HOLE;
6772         em->orig_start = EXTENT_MAP_HOLE;
6773         em->len = (u64)-1;
6774         em->block_len = (u64)-1;
6775
6776         if (!path) {
6777                 path = btrfs_alloc_path();
6778                 if (!path) {
6779                         err = -ENOMEM;
6780                         goto out;
6781                 }
6782                 /*
6783                  * Chances are we'll be called again, so go ahead and do
6784                  * readahead
6785                  */
6786                 path->reada = READA_FORWARD;
6787         }
6788
6789         ret = btrfs_lookup_file_extent(trans, root, path,
6790                                        objectid, start, trans != NULL);
6791         if (ret < 0) {
6792                 err = ret;
6793                 goto out;
6794         }
6795
6796         if (ret != 0) {
6797                 if (path->slots[0] == 0)
6798                         goto not_found;
6799                 path->slots[0]--;
6800         }
6801
6802         leaf = path->nodes[0];
6803         item = btrfs_item_ptr(leaf, path->slots[0],
6804                               struct btrfs_file_extent_item);
6805         /* are we inside the extent that was found? */
6806         btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6807         found_type = found_key.type;
6808         if (found_key.objectid != objectid ||
6809             found_type != BTRFS_EXTENT_DATA_KEY) {
6810                 /*
6811                  * If we backup past the first extent we want to move forward
6812                  * and see if there is an extent in front of us, otherwise we'll
6813                  * say there is a hole for our whole search range which can
6814                  * cause problems.
6815                  */
6816                 extent_end = start;
6817                 goto next;
6818         }
6819
6820         found_type = btrfs_file_extent_type(leaf, item);
6821         extent_start = found_key.offset;
6822         if (found_type == BTRFS_FILE_EXTENT_REG ||
6823             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6824                 extent_end = extent_start +
6825                        btrfs_file_extent_num_bytes(leaf, item);
6826         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6827                 size_t size;
6828                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6829                 extent_end = ALIGN(extent_start + size, root->sectorsize);
6830         }
6831 next:
6832         if (start >= extent_end) {
6833                 path->slots[0]++;
6834                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
6835                         ret = btrfs_next_leaf(root, path);
6836                         if (ret < 0) {
6837                                 err = ret;
6838                                 goto out;
6839                         }
6840                         if (ret > 0)
6841                                 goto not_found;
6842                         leaf = path->nodes[0];
6843                 }
6844                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6845                 if (found_key.objectid != objectid ||
6846                     found_key.type != BTRFS_EXTENT_DATA_KEY)
6847                         goto not_found;
6848                 if (start + len <= found_key.offset)
6849                         goto not_found;
6850                 if (start > found_key.offset)
6851                         goto next;
6852                 em->start = start;
6853                 em->orig_start = start;
6854                 em->len = found_key.offset - start;
6855                 goto not_found_em;
6856         }
6857
6858         btrfs_extent_item_to_extent_map(inode, path, item, new_inline, em);
6859
6860         if (found_type == BTRFS_FILE_EXTENT_REG ||
6861             found_type == BTRFS_FILE_EXTENT_PREALLOC) {
6862                 goto insert;
6863         } else if (found_type == BTRFS_FILE_EXTENT_INLINE) {
6864                 unsigned long ptr;
6865                 char *map;
6866                 size_t size;
6867                 size_t extent_offset;
6868                 size_t copy_size;
6869
6870                 if (new_inline)
6871                         goto out;
6872
6873                 size = btrfs_file_extent_inline_len(leaf, path->slots[0], item);
6874                 extent_offset = page_offset(page) + pg_offset - extent_start;
6875                 copy_size = min_t(u64, PAGE_CACHE_SIZE - pg_offset,
6876                                 size - extent_offset);
6877                 em->start = extent_start + extent_offset;
6878                 em->len = ALIGN(copy_size, root->sectorsize);
6879                 em->orig_block_len = em->len;
6880                 em->orig_start = em->start;
6881                 ptr = btrfs_file_extent_inline_start(item) + extent_offset;
6882                 if (create == 0 && !PageUptodate(page)) {
6883                         if (btrfs_file_extent_compression(leaf, item) !=
6884                             BTRFS_COMPRESS_NONE) {
6885                                 ret = uncompress_inline(path, page, pg_offset,
6886                                                         extent_offset, item);
6887                                 if (ret) {
6888                                         err = ret;
6889                                         goto out;
6890                                 }
6891                         } else {
6892                                 map = kmap(page);
6893                                 read_extent_buffer(leaf, map + pg_offset, ptr,
6894                                                    copy_size);
6895                                 if (pg_offset + copy_size < PAGE_CACHE_SIZE) {
6896                                         memset(map + pg_offset + copy_size, 0,
6897                                                PAGE_CACHE_SIZE - pg_offset -
6898                                                copy_size);
6899                                 }
6900                                 kunmap(page);
6901                         }
6902                         flush_dcache_page(page);
6903                 } else if (create && PageUptodate(page)) {
6904                         BUG();
6905                         if (!trans) {
6906                                 kunmap(page);
6907                                 free_extent_map(em);
6908                                 em = NULL;
6909
6910                                 btrfs_release_path(path);
6911                                 trans = btrfs_join_transaction(root);
6912
6913                                 if (IS_ERR(trans))
6914                                         return ERR_CAST(trans);
6915                                 goto again;
6916                         }
6917                         map = kmap(page);
6918                         write_extent_buffer(leaf, map + pg_offset, ptr,
6919                                             copy_size);
6920                         kunmap(page);
6921                         btrfs_mark_buffer_dirty(leaf);
6922                 }
6923                 set_extent_uptodate(io_tree, em->start,
6924                                     extent_map_end(em) - 1, NULL, GFP_NOFS);
6925                 goto insert;
6926         }
6927 not_found:
6928         em->start = start;
6929         em->orig_start = start;
6930         em->len = len;
6931 not_found_em:
6932         em->block_start = EXTENT_MAP_HOLE;
6933         set_bit(EXTENT_FLAG_VACANCY, &em->flags);
6934 insert:
6935         btrfs_release_path(path);
6936         if (em->start > start || extent_map_end(em) <= start) {
6937                 btrfs_err(root->fs_info, "bad extent! em: [%llu %llu] passed [%llu %llu]",
6938                         em->start, em->len, start, len);
6939                 err = -EIO;
6940                 goto out;
6941         }
6942
6943         err = 0;
6944         write_lock(&em_tree->lock);
6945         ret = add_extent_mapping(em_tree, em, 0);
6946         /* it is possible that someone inserted the extent into the tree
6947          * while we had the lock dropped.  It is also possible that
6948          * an overlapping map exists in the tree
6949          */
6950         if (ret == -EEXIST) {
6951                 struct extent_map *existing;
6952
6953                 ret = 0;
6954
6955                 existing = search_extent_mapping(em_tree, start, len);
6956                 /*
6957                  * existing will always be non-NULL, since there must be
6958                  * extent causing the -EEXIST.
6959                  */
6960                 if (start >= extent_map_end(existing) ||
6961                     start <= existing->start) {
6962                         /*
6963                          * The existing extent map is the one nearest to
6964                          * the [start, start + len) range which overlaps
6965                          */
6966                         err = merge_extent_mapping(em_tree, existing,
6967                                                    em, start);
6968                         free_extent_map(existing);
6969                         if (err) {
6970                                 free_extent_map(em);
6971                                 em = NULL;
6972                         }
6973                 } else {
6974                         free_extent_map(em);
6975                         em = existing;
6976                         err = 0;
6977                 }
6978         }
6979         write_unlock(&em_tree->lock);
6980 out:
6981
6982         trace_btrfs_get_extent(root, em);
6983
6984         btrfs_free_path(path);
6985         if (trans) {
6986                 ret = btrfs_end_transaction(trans, root);
6987                 if (!err)
6988                         err = ret;
6989         }
6990         if (err) {
6991                 free_extent_map(em);
6992                 return ERR_PTR(err);
6993         }
6994         BUG_ON(!em); /* Error is always set */
6995         return em;
6996 }
6997
6998 struct extent_map *btrfs_get_extent_fiemap(struct inode *inode, struct page *page,
6999                                            size_t pg_offset, u64 start, u64 len,
7000                                            int create)
7001 {
7002         struct extent_map *em;
7003         struct extent_map *hole_em = NULL;
7004         u64 range_start = start;
7005         u64 end;
7006         u64 found;
7007         u64 found_end;
7008         int err = 0;
7009
7010         em = btrfs_get_extent(inode, page, pg_offset, start, len, create);
7011         if (IS_ERR(em))
7012                 return em;
7013         if (em) {
7014                 /*
7015                  * if our em maps to
7016                  * -  a hole or
7017                  * -  a pre-alloc extent,
7018                  * there might actually be delalloc bytes behind it.
7019                  */
7020                 if (em->block_start != EXTENT_MAP_HOLE &&
7021                     !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7022                         return em;
7023                 else
7024                         hole_em = em;
7025         }
7026
7027         /* check to see if we've wrapped (len == -1 or similar) */
7028         end = start + len;
7029         if (end < start)
7030                 end = (u64)-1;
7031         else
7032                 end -= 1;
7033
7034         em = NULL;
7035
7036         /* ok, we didn't find anything, lets look for delalloc */
7037         found = count_range_bits(&BTRFS_I(inode)->io_tree, &range_start,
7038                                  end, len, EXTENT_DELALLOC, 1);
7039         found_end = range_start + found;
7040         if (found_end < range_start)
7041                 found_end = (u64)-1;
7042
7043         /*
7044          * we didn't find anything useful, return
7045          * the original results from get_extent()
7046          */
7047         if (range_start > end || found_end <= start) {
7048                 em = hole_em;
7049                 hole_em = NULL;
7050                 goto out;
7051         }
7052
7053         /* adjust the range_start to make sure it doesn't
7054          * go backwards from the start they passed in
7055          */
7056         range_start = max(start, range_start);
7057         found = found_end - range_start;
7058
7059         if (found > 0) {
7060                 u64 hole_start = start;
7061                 u64 hole_len = len;
7062
7063                 em = alloc_extent_map();
7064                 if (!em) {
7065                         err = -ENOMEM;
7066                         goto out;
7067                 }
7068                 /*
7069                  * when btrfs_get_extent can't find anything it
7070                  * returns one huge hole
7071                  *
7072                  * make sure what it found really fits our range, and
7073                  * adjust to make sure it is based on the start from
7074                  * the caller
7075                  */
7076                 if (hole_em) {
7077                         u64 calc_end = extent_map_end(hole_em);
7078
7079                         if (calc_end <= start || (hole_em->start > end)) {
7080                                 free_extent_map(hole_em);
7081                                 hole_em = NULL;
7082                         } else {
7083                                 hole_start = max(hole_em->start, start);
7084                                 hole_len = calc_end - hole_start;
7085                         }
7086                 }
7087                 em->bdev = NULL;
7088                 if (hole_em && range_start > hole_start) {
7089                         /* our hole starts before our delalloc, so we
7090                          * have to return just the parts of the hole
7091                          * that go until  the delalloc starts
7092                          */
7093                         em->len = min(hole_len,
7094                                       range_start - hole_start);
7095                         em->start = hole_start;
7096                         em->orig_start = hole_start;
7097                         /*
7098                          * don't adjust block start at all,
7099                          * it is fixed at EXTENT_MAP_HOLE
7100                          */
7101                         em->block_start = hole_em->block_start;
7102                         em->block_len = hole_len;
7103                         if (test_bit(EXTENT_FLAG_PREALLOC, &hole_em->flags))
7104                                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
7105                 } else {
7106                         em->start = range_start;
7107                         em->len = found;
7108                         em->orig_start = range_start;
7109                         em->block_start = EXTENT_MAP_DELALLOC;
7110                         em->block_len = found;
7111                 }
7112         } else if (hole_em) {
7113                 return hole_em;
7114         }
7115 out:
7116
7117         free_extent_map(hole_em);
7118         if (err) {
7119                 free_extent_map(em);
7120                 return ERR_PTR(err);
7121         }
7122         return em;
7123 }
7124
7125 static struct extent_map *btrfs_new_extent_direct(struct inode *inode,
7126                                                   u64 start, u64 len)
7127 {
7128         struct btrfs_root *root = BTRFS_I(inode)->root;
7129         struct extent_map *em;
7130         struct btrfs_key ins;
7131         u64 alloc_hint;
7132         int ret;
7133
7134         alloc_hint = get_extent_allocation_hint(inode, start, len);
7135         ret = btrfs_reserve_extent(root, len, root->sectorsize, 0,
7136                                    alloc_hint, &ins, 1, 1);
7137         if (ret)
7138                 return ERR_PTR(ret);
7139
7140         em = create_pinned_em(inode, start, ins.offset, start, ins.objectid,
7141                               ins.offset, ins.offset, ins.offset, 0);
7142         if (IS_ERR(em)) {
7143                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7144                 return em;
7145         }
7146
7147         ret = btrfs_add_ordered_extent_dio(inode, start, ins.objectid,
7148                                            ins.offset, ins.offset, 0);
7149         if (ret) {
7150                 btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 1);
7151                 free_extent_map(em);
7152                 return ERR_PTR(ret);
7153         }
7154
7155         return em;
7156 }
7157
7158 /*
7159  * returns 1 when the nocow is safe, < 1 on error, 0 if the
7160  * block must be cow'd
7161  */
7162 noinline int can_nocow_extent(struct inode *inode, u64 offset, u64 *len,
7163                               u64 *orig_start, u64 *orig_block_len,
7164                               u64 *ram_bytes)
7165 {
7166         struct btrfs_trans_handle *trans;
7167         struct btrfs_path *path;
7168         int ret;
7169         struct extent_buffer *leaf;
7170         struct btrfs_root *root = BTRFS_I(inode)->root;
7171         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
7172         struct btrfs_file_extent_item *fi;
7173         struct btrfs_key key;
7174         u64 disk_bytenr;
7175         u64 backref_offset;
7176         u64 extent_end;
7177         u64 num_bytes;
7178         int slot;
7179         int found_type;
7180         bool nocow = (BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW);
7181
7182         path = btrfs_alloc_path();
7183         if (!path)
7184                 return -ENOMEM;
7185
7186         ret = btrfs_lookup_file_extent(NULL, root, path, btrfs_ino(inode),
7187                                        offset, 0);
7188         if (ret < 0)
7189                 goto out;
7190
7191         slot = path->slots[0];
7192         if (ret == 1) {
7193                 if (slot == 0) {
7194                         /* can't find the item, must cow */
7195                         ret = 0;
7196                         goto out;
7197                 }
7198                 slot--;
7199         }
7200         ret = 0;
7201         leaf = path->nodes[0];
7202         btrfs_item_key_to_cpu(leaf, &key, slot);
7203         if (key.objectid != btrfs_ino(inode) ||
7204             key.type != BTRFS_EXTENT_DATA_KEY) {
7205                 /* not our file or wrong item type, must cow */
7206                 goto out;
7207         }
7208
7209         if (key.offset > offset) {
7210                 /* Wrong offset, must cow */
7211                 goto out;
7212         }
7213
7214         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
7215         found_type = btrfs_file_extent_type(leaf, fi);
7216         if (found_type != BTRFS_FILE_EXTENT_REG &&
7217             found_type != BTRFS_FILE_EXTENT_PREALLOC) {
7218                 /* not a regular extent, must cow */
7219                 goto out;
7220         }
7221
7222         if (!nocow && found_type == BTRFS_FILE_EXTENT_REG)
7223                 goto out;
7224
7225         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
7226         if (extent_end <= offset)
7227                 goto out;
7228
7229         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7230         if (disk_bytenr == 0)
7231                 goto out;
7232
7233         if (btrfs_file_extent_compression(leaf, fi) ||
7234             btrfs_file_extent_encryption(leaf, fi) ||
7235             btrfs_file_extent_other_encoding(leaf, fi))
7236                 goto out;
7237
7238         backref_offset = btrfs_file_extent_offset(leaf, fi);
7239
7240         if (orig_start) {
7241                 *orig_start = key.offset - backref_offset;
7242                 *orig_block_len = btrfs_file_extent_disk_num_bytes(leaf, fi);
7243                 *ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
7244         }
7245
7246         if (btrfs_extent_readonly(root, disk_bytenr))
7247                 goto out;
7248
7249         num_bytes = min(offset + *len, extent_end) - offset;
7250         if (!nocow && found_type == BTRFS_FILE_EXTENT_PREALLOC) {
7251                 u64 range_end;
7252
7253                 range_end = round_up(offset + num_bytes, root->sectorsize) - 1;
7254                 ret = test_range_bit(io_tree, offset, range_end,
7255                                      EXTENT_DELALLOC, 0, NULL);
7256                 if (ret) {
7257                         ret = -EAGAIN;
7258                         goto out;
7259                 }
7260         }
7261
7262         btrfs_release_path(path);
7263
7264         /*
7265          * look for other files referencing this extent, if we
7266          * find any we must cow
7267          */
7268         trans = btrfs_join_transaction(root);
7269         if (IS_ERR(trans)) {
7270                 ret = 0;
7271                 goto out;
7272         }
7273
7274         ret = btrfs_cross_ref_exist(trans, root, btrfs_ino(inode),
7275                                     key.offset - backref_offset, disk_bytenr);
7276         btrfs_end_transaction(trans, root);
7277         if (ret) {
7278                 ret = 0;
7279                 goto out;
7280         }
7281
7282         /*
7283          * adjust disk_bytenr and num_bytes to cover just the bytes
7284          * in this extent we are about to write.  If there
7285          * are any csums in that range we have to cow in order
7286          * to keep the csums correct
7287          */
7288         disk_bytenr += backref_offset;
7289         disk_bytenr += offset - key.offset;
7290         if (csum_exist_in_range(root, disk_bytenr, num_bytes))
7291                                 goto out;
7292         /*
7293          * all of the above have passed, it is safe to overwrite this extent
7294          * without cow
7295          */
7296         *len = num_bytes;
7297         ret = 1;
7298 out:
7299         btrfs_free_path(path);
7300         return ret;
7301 }
7302
7303 bool btrfs_page_exists_in_range(struct inode *inode, loff_t start, loff_t end)
7304 {
7305         struct radix_tree_root *root = &inode->i_mapping->page_tree;
7306         int found = false;
7307         void **pagep = NULL;
7308         struct page *page = NULL;
7309         int start_idx;
7310         int end_idx;
7311
7312         start_idx = start >> PAGE_CACHE_SHIFT;
7313
7314         /*
7315          * end is the last byte in the last page.  end == start is legal
7316          */
7317         end_idx = end >> PAGE_CACHE_SHIFT;
7318
7319         rcu_read_lock();
7320
7321         /* Most of the code in this while loop is lifted from
7322          * find_get_page.  It's been modified to begin searching from a
7323          * page and return just the first page found in that range.  If the
7324          * found idx is less than or equal to the end idx then we know that
7325          * a page exists.  If no pages are found or if those pages are
7326          * outside of the range then we're fine (yay!) */
7327         while (page == NULL &&
7328                radix_tree_gang_lookup_slot(root, &pagep, NULL, start_idx, 1)) {
7329                 page = radix_tree_deref_slot(pagep);
7330                 if (unlikely(!page))
7331                         break;
7332
7333                 if (radix_tree_exception(page)) {
7334                         if (radix_tree_deref_retry(page)) {
7335                                 page = NULL;
7336                                 continue;
7337                         }
7338                         /*
7339                          * Otherwise, shmem/tmpfs must be storing a swap entry
7340                          * here as an exceptional entry: so return it without
7341                          * attempting to raise page count.
7342                          */
7343                         page = NULL;
7344                         break; /* TODO: Is this relevant for this use case? */
7345                 }
7346
7347                 if (!page_cache_get_speculative(page)) {
7348                         page = NULL;
7349                         continue;
7350                 }
7351
7352                 /*
7353                  * Has the page moved?
7354                  * This is part of the lockless pagecache protocol. See
7355                  * include/linux/pagemap.h for details.
7356                  */
7357                 if (unlikely(page != *pagep)) {
7358                         page_cache_release(page);
7359                         page = NULL;
7360                 }
7361         }
7362
7363         if (page) {
7364                 if (page->index <= end_idx)
7365                         found = true;
7366                 page_cache_release(page);
7367         }
7368
7369         rcu_read_unlock();
7370         return found;
7371 }
7372
7373 static int lock_extent_direct(struct inode *inode, u64 lockstart, u64 lockend,
7374                               struct extent_state **cached_state, int writing)
7375 {
7376         struct btrfs_ordered_extent *ordered;
7377         int ret = 0;
7378
7379         while (1) {
7380                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7381                                  cached_state);
7382                 /*
7383                  * We're concerned with the entire range that we're going to be
7384                  * doing DIO to, so we need to make sure theres no ordered
7385                  * extents in this range.
7386                  */
7387                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
7388                                                      lockend - lockstart + 1);
7389
7390                 /*
7391                  * We need to make sure there are no buffered pages in this
7392                  * range either, we could have raced between the invalidate in
7393                  * generic_file_direct_write and locking the extent.  The
7394                  * invalidate needs to happen so that reads after a write do not
7395                  * get stale data.
7396                  */
7397                 if (!ordered &&
7398                     (!writing ||
7399                      !btrfs_page_exists_in_range(inode, lockstart, lockend)))
7400                         break;
7401
7402                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7403                                      cached_state, GFP_NOFS);
7404
7405                 if (ordered) {
7406                         btrfs_start_ordered_extent(inode, ordered, 1);
7407                         btrfs_put_ordered_extent(ordered);
7408                 } else {
7409                         /*
7410                          * We could trigger writeback for this range (and wait
7411                          * for it to complete) and then invalidate the pages for
7412                          * this range (through invalidate_inode_pages2_range()),
7413                          * but that can lead us to a deadlock with a concurrent
7414                          * call to readpages() (a buffered read or a defrag call
7415                          * triggered a readahead) on a page lock due to an
7416                          * ordered dio extent we created before but did not have
7417                          * yet a corresponding bio submitted (whence it can not
7418                          * complete), which makes readpages() wait for that
7419                          * ordered extent to complete while holding a lock on
7420                          * that page.
7421                          */
7422                         ret = -ENOTBLK;
7423                         break;
7424                 }
7425
7426                 cond_resched();
7427         }
7428
7429         return ret;
7430 }
7431
7432 static struct extent_map *create_pinned_em(struct inode *inode, u64 start,
7433                                            u64 len, u64 orig_start,
7434                                            u64 block_start, u64 block_len,
7435                                            u64 orig_block_len, u64 ram_bytes,
7436                                            int type)
7437 {
7438         struct extent_map_tree *em_tree;
7439         struct extent_map *em;
7440         struct btrfs_root *root = BTRFS_I(inode)->root;
7441         int ret;
7442
7443         em_tree = &BTRFS_I(inode)->extent_tree;
7444         em = alloc_extent_map();
7445         if (!em)
7446                 return ERR_PTR(-ENOMEM);
7447
7448         em->start = start;
7449         em->orig_start = orig_start;
7450         em->mod_start = start;
7451         em->mod_len = len;
7452         em->len = len;
7453         em->block_len = block_len;
7454         em->block_start = block_start;
7455         em->bdev = root->fs_info->fs_devices->latest_bdev;
7456         em->orig_block_len = orig_block_len;
7457         em->ram_bytes = ram_bytes;
7458         em->generation = -1;
7459         set_bit(EXTENT_FLAG_PINNED, &em->flags);
7460         if (type == BTRFS_ORDERED_PREALLOC)
7461                 set_bit(EXTENT_FLAG_FILLING, &em->flags);
7462
7463         do {
7464                 btrfs_drop_extent_cache(inode, em->start,
7465                                 em->start + em->len - 1, 0);
7466                 write_lock(&em_tree->lock);
7467                 ret = add_extent_mapping(em_tree, em, 1);
7468                 write_unlock(&em_tree->lock);
7469         } while (ret == -EEXIST);
7470
7471         if (ret) {
7472                 free_extent_map(em);
7473                 return ERR_PTR(ret);
7474         }
7475
7476         return em;
7477 }
7478
7479 static void adjust_dio_outstanding_extents(struct inode *inode,
7480                                            struct btrfs_dio_data *dio_data,
7481                                            const u64 len)
7482 {
7483         unsigned num_extents;
7484
7485         num_extents = (unsigned) div64_u64(len + BTRFS_MAX_EXTENT_SIZE - 1,
7486                                            BTRFS_MAX_EXTENT_SIZE);
7487         /*
7488          * If we have an outstanding_extents count still set then we're
7489          * within our reservation, otherwise we need to adjust our inode
7490          * counter appropriately.
7491          */
7492         if (dio_data->outstanding_extents) {
7493                 dio_data->outstanding_extents -= num_extents;
7494         } else {
7495                 spin_lock(&BTRFS_I(inode)->lock);
7496                 BTRFS_I(inode)->outstanding_extents += num_extents;
7497                 spin_unlock(&BTRFS_I(inode)->lock);
7498         }
7499 }
7500
7501 static int btrfs_get_blocks_direct(struct inode *inode, sector_t iblock,
7502                                    struct buffer_head *bh_result, int create)
7503 {
7504         struct extent_map *em;
7505         struct btrfs_root *root = BTRFS_I(inode)->root;
7506         struct extent_state *cached_state = NULL;
7507         struct btrfs_dio_data *dio_data = NULL;
7508         u64 start = iblock << inode->i_blkbits;
7509         u64 lockstart, lockend;
7510         u64 len = bh_result->b_size;
7511         int unlock_bits = EXTENT_LOCKED;
7512         int ret = 0;
7513
7514         if (create)
7515                 unlock_bits |= EXTENT_DIRTY;
7516         else
7517                 len = min_t(u64, len, root->sectorsize);
7518
7519         lockstart = start;
7520         lockend = start + len - 1;
7521
7522         if (current->journal_info) {
7523                 /*
7524                  * Need to pull our outstanding extents and set journal_info to NULL so
7525                  * that anything that needs to check if there's a transction doesn't get
7526                  * confused.
7527                  */
7528                 dio_data = current->journal_info;
7529                 current->journal_info = NULL;
7530         }
7531
7532         /*
7533          * If this errors out it's because we couldn't invalidate pagecache for
7534          * this range and we need to fallback to buffered.
7535          */
7536         if (lock_extent_direct(inode, lockstart, lockend, &cached_state,
7537                                create)) {
7538                 ret = -ENOTBLK;
7539                 goto err;
7540         }
7541
7542         em = btrfs_get_extent(inode, NULL, 0, start, len, 0);
7543         if (IS_ERR(em)) {
7544                 ret = PTR_ERR(em);
7545                 goto unlock_err;
7546         }
7547
7548         /*
7549          * Ok for INLINE and COMPRESSED extents we need to fallback on buffered
7550          * io.  INLINE is special, and we could probably kludge it in here, but
7551          * it's still buffered so for safety lets just fall back to the generic
7552          * buffered path.
7553          *
7554          * For COMPRESSED we _have_ to read the entire extent in so we can
7555          * decompress it, so there will be buffering required no matter what we
7556          * do, so go ahead and fallback to buffered.
7557          *
7558          * We return -ENOTBLK because thats what makes DIO go ahead and go back
7559          * to buffered IO.  Don't blame me, this is the price we pay for using
7560          * the generic code.
7561          */
7562         if (test_bit(EXTENT_FLAG_COMPRESSED, &em->flags) ||
7563             em->block_start == EXTENT_MAP_INLINE) {
7564                 free_extent_map(em);
7565                 ret = -ENOTBLK;
7566                 goto unlock_err;
7567         }
7568
7569         /* Just a good old fashioned hole, return */
7570         if (!create && (em->block_start == EXTENT_MAP_HOLE ||
7571                         test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
7572                 free_extent_map(em);
7573                 goto unlock_err;
7574         }
7575
7576         /*
7577          * We don't allocate a new extent in the following cases
7578          *
7579          * 1) The inode is marked as NODATACOW.  In this case we'll just use the
7580          * existing extent.
7581          * 2) The extent is marked as PREALLOC.  We're good to go here and can
7582          * just use the extent.
7583          *
7584          */
7585         if (!create) {
7586                 len = min(len, em->len - (start - em->start));
7587                 lockstart = start + len;
7588                 goto unlock;
7589         }
7590
7591         if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags) ||
7592             ((BTRFS_I(inode)->flags & BTRFS_INODE_NODATACOW) &&
7593              em->block_start != EXTENT_MAP_HOLE)) {
7594                 int type;
7595                 u64 block_start, orig_start, orig_block_len, ram_bytes;
7596
7597                 if (test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7598                         type = BTRFS_ORDERED_PREALLOC;
7599                 else
7600                         type = BTRFS_ORDERED_NOCOW;
7601                 len = min(len, em->len - (start - em->start));
7602                 block_start = em->block_start + (start - em->start);
7603
7604                 if (can_nocow_extent(inode, start, &len, &orig_start,
7605                                      &orig_block_len, &ram_bytes) == 1) {
7606                         if (type == BTRFS_ORDERED_PREALLOC) {
7607                                 free_extent_map(em);
7608                                 em = create_pinned_em(inode, start, len,
7609                                                        orig_start,
7610                                                        block_start, len,
7611                                                        orig_block_len,
7612                                                        ram_bytes, type);
7613                                 if (IS_ERR(em)) {
7614                                         ret = PTR_ERR(em);
7615                                         goto unlock_err;
7616                                 }
7617                         }
7618
7619                         ret = btrfs_add_ordered_extent_dio(inode, start,
7620                                            block_start, len, len, type);
7621                         if (ret) {
7622                                 free_extent_map(em);
7623                                 goto unlock_err;
7624                         }
7625                         goto unlock;
7626                 }
7627         }
7628
7629         /*
7630          * this will cow the extent, reset the len in case we changed
7631          * it above
7632          */
7633         len = bh_result->b_size;
7634         free_extent_map(em);
7635         em = btrfs_new_extent_direct(inode, start, len);
7636         if (IS_ERR(em)) {
7637                 ret = PTR_ERR(em);
7638                 goto unlock_err;
7639         }
7640         len = min(len, em->len - (start - em->start));
7641 unlock:
7642         bh_result->b_blocknr = (em->block_start + (start - em->start)) >>
7643                 inode->i_blkbits;
7644         bh_result->b_size = len;
7645         bh_result->b_bdev = em->bdev;
7646         set_buffer_mapped(bh_result);
7647         if (create) {
7648                 if (!test_bit(EXTENT_FLAG_PREALLOC, &em->flags))
7649                         set_buffer_new(bh_result);
7650
7651                 /*
7652                  * Need to update the i_size under the extent lock so buffered
7653                  * readers will get the updated i_size when we unlock.
7654                  */
7655                 if (start + len > i_size_read(inode))
7656                         i_size_write(inode, start + len);
7657
7658                 adjust_dio_outstanding_extents(inode, dio_data, len);
7659                 btrfs_free_reserved_data_space(inode, start, len);
7660                 WARN_ON(dio_data->reserve < len);
7661                 dio_data->reserve -= len;
7662                 dio_data->unsubmitted_oe_range_end = start + len;
7663                 current->journal_info = dio_data;
7664         }
7665
7666         /*
7667          * In the case of write we need to clear and unlock the entire range,
7668          * in the case of read we need to unlock only the end area that we
7669          * aren't using if there is any left over space.
7670          */
7671         if (lockstart < lockend) {
7672                 clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
7673                                  lockend, unlock_bits, 1, 0,
7674                                  &cached_state, GFP_NOFS);
7675         } else {
7676                 free_extent_state(cached_state);
7677         }
7678
7679         free_extent_map(em);
7680
7681         return 0;
7682
7683 unlock_err:
7684         clear_extent_bit(&BTRFS_I(inode)->io_tree, lockstart, lockend,
7685                          unlock_bits, 1, 0, &cached_state, GFP_NOFS);
7686 err:
7687         if (dio_data)
7688                 current->journal_info = dio_data;
7689         /*
7690          * Compensate the delalloc release we do in btrfs_direct_IO() when we
7691          * write less data then expected, so that we don't underflow our inode's
7692          * outstanding extents counter.
7693          */
7694         if (create && dio_data)
7695                 adjust_dio_outstanding_extents(inode, dio_data, len);
7696
7697         return ret;
7698 }
7699
7700 static inline int submit_dio_repair_bio(struct inode *inode, struct bio *bio,
7701                                         int rw, int mirror_num)
7702 {
7703         struct btrfs_root *root = BTRFS_I(inode)->root;
7704         int ret;
7705
7706         BUG_ON(rw & REQ_WRITE);
7707
7708         bio_get(bio);
7709
7710         ret = btrfs_bio_wq_end_io(root->fs_info, bio,
7711                                   BTRFS_WQ_ENDIO_DIO_REPAIR);
7712         if (ret)
7713                 goto err;
7714
7715         ret = btrfs_map_bio(root, rw, bio, mirror_num, 0);
7716 err:
7717         bio_put(bio);
7718         return ret;
7719 }
7720
7721 static int btrfs_check_dio_repairable(struct inode *inode,
7722                                       struct bio *failed_bio,
7723                                       struct io_failure_record *failrec,
7724                                       int failed_mirror)
7725 {
7726         int num_copies;
7727
7728         num_copies = btrfs_num_copies(BTRFS_I(inode)->root->fs_info,
7729                                       failrec->logical, failrec->len);
7730         if (num_copies == 1) {
7731                 /*
7732                  * we only have a single copy of the data, so don't bother with
7733                  * all the retry and error correction code that follows. no
7734                  * matter what the error is, it is very likely to persist.
7735                  */
7736                 pr_debug("Check DIO Repairable: cannot repair, num_copies=%d, next_mirror %d, failed_mirror %d\n",
7737                          num_copies, failrec->this_mirror, failed_mirror);
7738                 return 0;
7739         }
7740
7741         failrec->failed_mirror = failed_mirror;
7742         failrec->this_mirror++;
7743         if (failrec->this_mirror == failed_mirror)
7744                 failrec->this_mirror++;
7745
7746         if (failrec->this_mirror > num_copies) {
7747                 pr_debug("Check DIO Repairable: (fail) num_copies=%d, next_mirror %d, failed_mirror %d\n",
7748                          num_copies, failrec->this_mirror, failed_mirror);
7749                 return 0;
7750         }
7751
7752         return 1;
7753 }
7754
7755 static int dio_read_error(struct inode *inode, struct bio *failed_bio,
7756                           struct page *page, u64 start, u64 end,
7757                           int failed_mirror, bio_end_io_t *repair_endio,
7758                           void *repair_arg)
7759 {
7760         struct io_failure_record *failrec;
7761         struct bio *bio;
7762         int isector;
7763         int read_mode;
7764         int ret;
7765
7766         BUG_ON(failed_bio->bi_rw & REQ_WRITE);
7767
7768         ret = btrfs_get_io_failure_record(inode, start, end, &failrec);
7769         if (ret)
7770                 return ret;
7771
7772         ret = btrfs_check_dio_repairable(inode, failed_bio, failrec,
7773                                          failed_mirror);
7774         if (!ret) {
7775                 free_io_failure(inode, failrec);
7776                 return -EIO;
7777         }
7778
7779         if (failed_bio->bi_vcnt > 1)
7780                 read_mode = READ_SYNC | REQ_FAILFAST_DEV;
7781         else
7782                 read_mode = READ_SYNC;
7783
7784         isector = start - btrfs_io_bio(failed_bio)->logical;
7785         isector >>= inode->i_sb->s_blocksize_bits;
7786         bio = btrfs_create_repair_bio(inode, failed_bio, failrec, page,
7787                                       0, isector, repair_endio, repair_arg);
7788         if (!bio) {
7789                 free_io_failure(inode, failrec);
7790                 return -EIO;
7791         }
7792
7793         btrfs_debug(BTRFS_I(inode)->root->fs_info,
7794                     "Repair DIO Read Error: submitting new dio read[%#x] to this_mirror=%d, in_validation=%d\n",
7795                     read_mode, failrec->this_mirror, failrec->in_validation);
7796
7797         ret = submit_dio_repair_bio(inode, bio, read_mode,
7798                                     failrec->this_mirror);
7799         if (ret) {
7800                 free_io_failure(inode, failrec);
7801                 bio_put(bio);
7802         }
7803
7804         return ret;
7805 }
7806
7807 struct btrfs_retry_complete {
7808         struct completion done;
7809         struct inode *inode;
7810         u64 start;
7811         int uptodate;
7812 };
7813
7814 static void btrfs_retry_endio_nocsum(struct bio *bio)
7815 {
7816         struct btrfs_retry_complete *done = bio->bi_private;
7817         struct bio_vec *bvec;
7818         int i;
7819
7820         if (bio->bi_error)
7821                 goto end;
7822
7823         done->uptodate = 1;
7824         bio_for_each_segment_all(bvec, bio, i)
7825                 clean_io_failure(done->inode, done->start, bvec->bv_page, 0);
7826 end:
7827         complete(&done->done);
7828         bio_put(bio);
7829 }
7830
7831 static int __btrfs_correct_data_nocsum(struct inode *inode,
7832                                        struct btrfs_io_bio *io_bio)
7833 {
7834         struct bio_vec *bvec;
7835         struct btrfs_retry_complete done;
7836         u64 start;
7837         int i;
7838         int ret;
7839
7840         start = io_bio->logical;
7841         done.inode = inode;
7842
7843         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7844 try_again:
7845                 done.uptodate = 0;
7846                 done.start = start;
7847                 init_completion(&done.done);
7848
7849                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7850                                      start + bvec->bv_len - 1,
7851                                      io_bio->mirror_num,
7852                                      btrfs_retry_endio_nocsum, &done);
7853                 if (ret)
7854                         return ret;
7855
7856                 wait_for_completion(&done.done);
7857
7858                 if (!done.uptodate) {
7859                         /* We might have another mirror, so try again */
7860                         goto try_again;
7861                 }
7862
7863                 start += bvec->bv_len;
7864         }
7865
7866         return 0;
7867 }
7868
7869 static void btrfs_retry_endio(struct bio *bio)
7870 {
7871         struct btrfs_retry_complete *done = bio->bi_private;
7872         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7873         struct bio_vec *bvec;
7874         int uptodate;
7875         int ret;
7876         int i;
7877
7878         if (bio->bi_error)
7879                 goto end;
7880
7881         uptodate = 1;
7882         bio_for_each_segment_all(bvec, bio, i) {
7883                 ret = __readpage_endio_check(done->inode, io_bio, i,
7884                                              bvec->bv_page, 0,
7885                                              done->start, bvec->bv_len);
7886                 if (!ret)
7887                         clean_io_failure(done->inode, done->start,
7888                                          bvec->bv_page, 0);
7889                 else
7890                         uptodate = 0;
7891         }
7892
7893         done->uptodate = uptodate;
7894 end:
7895         complete(&done->done);
7896         bio_put(bio);
7897 }
7898
7899 static int __btrfs_subio_endio_read(struct inode *inode,
7900                                     struct btrfs_io_bio *io_bio, int err)
7901 {
7902         struct bio_vec *bvec;
7903         struct btrfs_retry_complete done;
7904         u64 start;
7905         u64 offset = 0;
7906         int i;
7907         int ret;
7908
7909         err = 0;
7910         start = io_bio->logical;
7911         done.inode = inode;
7912
7913         bio_for_each_segment_all(bvec, &io_bio->bio, i) {
7914                 ret = __readpage_endio_check(inode, io_bio, i, bvec->bv_page,
7915                                              0, start, bvec->bv_len);
7916                 if (likely(!ret))
7917                         goto next;
7918 try_again:
7919                 done.uptodate = 0;
7920                 done.start = start;
7921                 init_completion(&done.done);
7922
7923                 ret = dio_read_error(inode, &io_bio->bio, bvec->bv_page, start,
7924                                      start + bvec->bv_len - 1,
7925                                      io_bio->mirror_num,
7926                                      btrfs_retry_endio, &done);
7927                 if (ret) {
7928                         err = ret;
7929                         goto next;
7930                 }
7931
7932                 wait_for_completion(&done.done);
7933
7934                 if (!done.uptodate) {
7935                         /* We might have another mirror, so try again */
7936                         goto try_again;
7937                 }
7938 next:
7939                 offset += bvec->bv_len;
7940                 start += bvec->bv_len;
7941         }
7942
7943         return err;
7944 }
7945
7946 static int btrfs_subio_endio_read(struct inode *inode,
7947                                   struct btrfs_io_bio *io_bio, int err)
7948 {
7949         bool skip_csum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
7950
7951         if (skip_csum) {
7952                 if (unlikely(err))
7953                         return __btrfs_correct_data_nocsum(inode, io_bio);
7954                 else
7955                         return 0;
7956         } else {
7957                 return __btrfs_subio_endio_read(inode, io_bio, err);
7958         }
7959 }
7960
7961 static void btrfs_endio_direct_read(struct bio *bio)
7962 {
7963         struct btrfs_dio_private *dip = bio->bi_private;
7964         struct inode *inode = dip->inode;
7965         struct bio *dio_bio;
7966         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
7967         int err = bio->bi_error;
7968
7969         if (dip->flags & BTRFS_DIO_ORIG_BIO_SUBMITTED)
7970                 err = btrfs_subio_endio_read(inode, io_bio, err);
7971
7972         unlock_extent(&BTRFS_I(inode)->io_tree, dip->logical_offset,
7973                       dip->logical_offset + dip->bytes - 1);
7974         dio_bio = dip->dio_bio;
7975
7976         kfree(dip);
7977
7978         dio_end_io(dio_bio, bio->bi_error);
7979
7980         if (io_bio->end_io)
7981                 io_bio->end_io(io_bio, err);
7982         bio_put(bio);
7983 }
7984
7985 static void btrfs_endio_direct_write_update_ordered(struct inode *inode,
7986                                                     const u64 offset,
7987                                                     const u64 bytes,
7988                                                     const int uptodate)
7989 {
7990         struct btrfs_root *root = BTRFS_I(inode)->root;
7991         struct btrfs_ordered_extent *ordered = NULL;
7992         u64 ordered_offset = offset;
7993         u64 ordered_bytes = bytes;
7994         int ret;
7995
7996 again:
7997         ret = btrfs_dec_test_first_ordered_pending(inode, &ordered,
7998                                                    &ordered_offset,
7999                                                    ordered_bytes,
8000                                                    uptodate);
8001         if (!ret)
8002                 goto out_test;
8003
8004         btrfs_init_work(&ordered->work, btrfs_endio_write_helper,
8005                         finish_ordered_fn, NULL, NULL);
8006         btrfs_queue_work(root->fs_info->endio_write_workers,
8007                          &ordered->work);
8008 out_test:
8009         /*
8010          * our bio might span multiple ordered extents.  If we haven't
8011          * completed the accounting for the whole dio, go back and try again
8012          */
8013         if (ordered_offset < offset + bytes) {
8014                 ordered_bytes = offset + bytes - ordered_offset;
8015                 ordered = NULL;
8016                 goto again;
8017         }
8018 }
8019
8020 static void btrfs_endio_direct_write(struct bio *bio)
8021 {
8022         struct btrfs_dio_private *dip = bio->bi_private;
8023         struct bio *dio_bio = dip->dio_bio;
8024
8025         btrfs_endio_direct_write_update_ordered(dip->inode,
8026                                                 dip->logical_offset,
8027                                                 dip->bytes,
8028                                                 !bio->bi_error);
8029
8030         kfree(dip);
8031
8032         dio_end_io(dio_bio, bio->bi_error);
8033         bio_put(bio);
8034 }
8035
8036 static int __btrfs_submit_bio_start_direct_io(struct inode *inode, int rw,
8037                                     struct bio *bio, int mirror_num,
8038                                     unsigned long bio_flags, u64 offset)
8039 {
8040         int ret;
8041         struct btrfs_root *root = BTRFS_I(inode)->root;
8042         ret = btrfs_csum_one_bio(root, inode, bio, offset, 1);
8043         BUG_ON(ret); /* -ENOMEM */
8044         return 0;
8045 }
8046
8047 static void btrfs_end_dio_bio(struct bio *bio)
8048 {
8049         struct btrfs_dio_private *dip = bio->bi_private;
8050         int err = bio->bi_error;
8051
8052         if (err)
8053                 btrfs_warn(BTRFS_I(dip->inode)->root->fs_info,
8054                            "direct IO failed ino %llu rw %lu sector %#Lx len %u err no %d",
8055                            btrfs_ino(dip->inode), bio->bi_rw,
8056                            (unsigned long long)bio->bi_iter.bi_sector,
8057                            bio->bi_iter.bi_size, err);
8058
8059         if (dip->subio_endio)
8060                 err = dip->subio_endio(dip->inode, btrfs_io_bio(bio), err);
8061
8062         if (err) {
8063                 dip->errors = 1;
8064
8065                 /*
8066                  * before atomic variable goto zero, we must make sure
8067                  * dip->errors is perceived to be set.
8068                  */
8069                 smp_mb__before_atomic();
8070         }
8071
8072         /* if there are more bios still pending for this dio, just exit */
8073         if (!atomic_dec_and_test(&dip->pending_bios))
8074                 goto out;
8075
8076         if (dip->errors) {
8077                 bio_io_error(dip->orig_bio);
8078         } else {
8079                 dip->dio_bio->bi_error = 0;
8080                 bio_endio(dip->orig_bio);
8081         }
8082 out:
8083         bio_put(bio);
8084 }
8085
8086 static struct bio *btrfs_dio_bio_alloc(struct block_device *bdev,
8087                                        u64 first_sector, gfp_t gfp_flags)
8088 {
8089         struct bio *bio;
8090         bio = btrfs_bio_alloc(bdev, first_sector, BIO_MAX_PAGES, gfp_flags);
8091         if (bio)
8092                 bio_associate_current(bio);
8093         return bio;
8094 }
8095
8096 static inline int btrfs_lookup_and_bind_dio_csum(struct btrfs_root *root,
8097                                                  struct inode *inode,
8098                                                  struct btrfs_dio_private *dip,
8099                                                  struct bio *bio,
8100                                                  u64 file_offset)
8101 {
8102         struct btrfs_io_bio *io_bio = btrfs_io_bio(bio);
8103         struct btrfs_io_bio *orig_io_bio = btrfs_io_bio(dip->orig_bio);
8104         int ret;
8105
8106         /*
8107          * We load all the csum data we need when we submit
8108          * the first bio to reduce the csum tree search and
8109          * contention.
8110          */
8111         if (dip->logical_offset == file_offset) {
8112                 ret = btrfs_lookup_bio_sums_dio(root, inode, dip->orig_bio,
8113                                                 file_offset);
8114                 if (ret)
8115                         return ret;
8116         }
8117
8118         if (bio == dip->orig_bio)
8119                 return 0;
8120
8121         file_offset -= dip->logical_offset;
8122         file_offset >>= inode->i_sb->s_blocksize_bits;
8123         io_bio->csum = (u8 *)(((u32 *)orig_io_bio->csum) + file_offset);
8124
8125         return 0;
8126 }
8127
8128 static inline int __btrfs_submit_dio_bio(struct bio *bio, struct inode *inode,
8129                                          int rw, u64 file_offset, int skip_sum,
8130                                          int async_submit)
8131 {
8132         struct btrfs_dio_private *dip = bio->bi_private;
8133         int write = rw & REQ_WRITE;
8134         struct btrfs_root *root = BTRFS_I(inode)->root;
8135         int ret;
8136
8137         if (async_submit)
8138                 async_submit = !atomic_read(&BTRFS_I(inode)->sync_writers);
8139
8140         bio_get(bio);
8141
8142         if (!write) {
8143                 ret = btrfs_bio_wq_end_io(root->fs_info, bio,
8144                                 BTRFS_WQ_ENDIO_DATA);
8145                 if (ret)
8146                         goto err;
8147         }
8148
8149         if (skip_sum)
8150                 goto map;
8151
8152         if (write && async_submit) {
8153                 ret = btrfs_wq_submit_bio(root->fs_info,
8154                                    inode, rw, bio, 0, 0,
8155                                    file_offset,
8156                                    __btrfs_submit_bio_start_direct_io,
8157                                    __btrfs_submit_bio_done);
8158                 goto err;
8159         } else if (write) {
8160                 /*
8161                  * If we aren't doing async submit, calculate the csum of the
8162                  * bio now.
8163                  */
8164                 ret = btrfs_csum_one_bio(root, inode, bio, file_offset, 1);
8165                 if (ret)
8166                         goto err;
8167         } else {
8168                 ret = btrfs_lookup_and_bind_dio_csum(root, inode, dip, bio,
8169                                                      file_offset);
8170                 if (ret)
8171                         goto err;
8172         }
8173 map:
8174         ret = btrfs_map_bio(root, rw, bio, 0, async_submit);
8175 err:
8176         bio_put(bio);
8177         return ret;
8178 }
8179
8180 static int btrfs_submit_direct_hook(int rw, struct btrfs_dio_private *dip,
8181                                     int skip_sum)
8182 {
8183         struct inode *inode = dip->inode;
8184         struct btrfs_root *root = BTRFS_I(inode)->root;
8185         struct bio *bio;
8186         struct bio *orig_bio = dip->orig_bio;
8187         struct bio_vec *bvec = orig_bio->bi_io_vec;
8188         u64 start_sector = orig_bio->bi_iter.bi_sector;
8189         u64 file_offset = dip->logical_offset;
8190         u64 submit_len = 0;
8191         u64 map_length;
8192         int nr_pages = 0;
8193         int ret;
8194         int async_submit = 0;
8195
8196         map_length = orig_bio->bi_iter.bi_size;
8197         ret = btrfs_map_block(root->fs_info, rw, start_sector << 9,
8198                               &map_length, NULL, 0);
8199         if (ret)
8200                 return -EIO;
8201
8202         if (map_length >= orig_bio->bi_iter.bi_size) {
8203                 bio = orig_bio;
8204                 dip->flags |= BTRFS_DIO_ORIG_BIO_SUBMITTED;
8205                 goto submit;
8206         }
8207
8208         /* async crcs make it difficult to collect full stripe writes. */
8209         if (btrfs_get_alloc_profile(root, 1) & BTRFS_BLOCK_GROUP_RAID56_MASK)
8210                 async_submit = 0;
8211         else
8212                 async_submit = 1;
8213
8214         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev, start_sector, GFP_NOFS);
8215         if (!bio)
8216                 return -ENOMEM;
8217
8218         bio->bi_private = dip;
8219         bio->bi_end_io = btrfs_end_dio_bio;
8220         btrfs_io_bio(bio)->logical = file_offset;
8221         atomic_inc(&dip->pending_bios);
8222
8223         while (bvec <= (orig_bio->bi_io_vec + orig_bio->bi_vcnt - 1)) {
8224                 if (map_length < submit_len + bvec->bv_len ||
8225                     bio_add_page(bio, bvec->bv_page, bvec->bv_len,
8226                                  bvec->bv_offset) < bvec->bv_len) {
8227                         /*
8228                          * inc the count before we submit the bio so
8229                          * we know the end IO handler won't happen before
8230                          * we inc the count. Otherwise, the dip might get freed
8231                          * before we're done setting it up
8232                          */
8233                         atomic_inc(&dip->pending_bios);
8234                         ret = __btrfs_submit_dio_bio(bio, inode, rw,
8235                                                      file_offset, skip_sum,
8236                                                      async_submit);
8237                         if (ret) {
8238                                 bio_put(bio);
8239                                 atomic_dec(&dip->pending_bios);
8240                                 goto out_err;
8241                         }
8242
8243                         start_sector += submit_len >> 9;
8244                         file_offset += submit_len;
8245
8246                         submit_len = 0;
8247                         nr_pages = 0;
8248
8249                         bio = btrfs_dio_bio_alloc(orig_bio->bi_bdev,
8250                                                   start_sector, GFP_NOFS);
8251                         if (!bio)
8252                                 goto out_err;
8253                         bio->bi_private = dip;
8254                         bio->bi_end_io = btrfs_end_dio_bio;
8255                         btrfs_io_bio(bio)->logical = file_offset;
8256
8257                         map_length = orig_bio->bi_iter.bi_size;
8258                         ret = btrfs_map_block(root->fs_info, rw,
8259                                               start_sector << 9,
8260                                               &map_length, NULL, 0);
8261                         if (ret) {
8262                                 bio_put(bio);
8263                                 goto out_err;
8264                         }
8265                 } else {
8266                         submit_len += bvec->bv_len;
8267                         nr_pages++;
8268                         bvec++;
8269                 }
8270         }
8271
8272 submit:
8273         ret = __btrfs_submit_dio_bio(bio, inode, rw, file_offset, skip_sum,
8274                                      async_submit);
8275         if (!ret)
8276                 return 0;
8277
8278         bio_put(bio);
8279 out_err:
8280         dip->errors = 1;
8281         /*
8282          * before atomic variable goto zero, we must
8283          * make sure dip->errors is perceived to be set.
8284          */
8285         smp_mb__before_atomic();
8286         if (atomic_dec_and_test(&dip->pending_bios))
8287                 bio_io_error(dip->orig_bio);
8288
8289         /* bio_end_io() will handle error, so we needn't return it */
8290         return 0;
8291 }
8292
8293 static void btrfs_submit_direct(int rw, struct bio *dio_bio,
8294                                 struct inode *inode, loff_t file_offset)
8295 {
8296         struct btrfs_dio_private *dip = NULL;
8297         struct bio *io_bio = NULL;
8298         struct btrfs_io_bio *btrfs_bio;
8299         int skip_sum;
8300         int write = rw & REQ_WRITE;
8301         int ret = 0;
8302
8303         skip_sum = BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM;
8304
8305         io_bio = btrfs_bio_clone(dio_bio, GFP_NOFS);
8306         if (!io_bio) {
8307                 ret = -ENOMEM;
8308                 goto free_ordered;
8309         }
8310
8311         dip = kzalloc(sizeof(*dip), GFP_NOFS);
8312         if (!dip) {
8313                 ret = -ENOMEM;
8314                 goto free_ordered;
8315         }
8316
8317         dip->private = dio_bio->bi_private;
8318         dip->inode = inode;
8319         dip->logical_offset = file_offset;
8320         dip->bytes = dio_bio->bi_iter.bi_size;
8321         dip->disk_bytenr = (u64)dio_bio->bi_iter.bi_sector << 9;
8322         io_bio->bi_private = dip;
8323         dip->orig_bio = io_bio;
8324         dip->dio_bio = dio_bio;
8325         atomic_set(&dip->pending_bios, 0);
8326         btrfs_bio = btrfs_io_bio(io_bio);
8327         btrfs_bio->logical = file_offset;
8328
8329         if (write) {
8330                 io_bio->bi_end_io = btrfs_endio_direct_write;
8331         } else {
8332                 io_bio->bi_end_io = btrfs_endio_direct_read;
8333                 dip->subio_endio = btrfs_subio_endio_read;
8334         }
8335
8336         /*
8337          * Reset the range for unsubmitted ordered extents (to a 0 length range)
8338          * even if we fail to submit a bio, because in such case we do the
8339          * corresponding error handling below and it must not be done a second
8340          * time by btrfs_direct_IO().
8341          */
8342         if (write) {
8343                 struct btrfs_dio_data *dio_data = current->journal_info;
8344
8345                 dio_data->unsubmitted_oe_range_end = dip->logical_offset +
8346                         dip->bytes;
8347                 dio_data->unsubmitted_oe_range_start =
8348                         dio_data->unsubmitted_oe_range_end;
8349         }
8350
8351         ret = btrfs_submit_direct_hook(rw, dip, skip_sum);
8352         if (!ret)
8353                 return;
8354
8355         if (btrfs_bio->end_io)
8356                 btrfs_bio->end_io(btrfs_bio, ret);
8357
8358 free_ordered:
8359         /*
8360          * If we arrived here it means either we failed to submit the dip
8361          * or we either failed to clone the dio_bio or failed to allocate the
8362          * dip. If we cloned the dio_bio and allocated the dip, we can just
8363          * call bio_endio against our io_bio so that we get proper resource
8364          * cleanup if we fail to submit the dip, otherwise, we must do the
8365          * same as btrfs_endio_direct_[write|read] because we can't call these
8366          * callbacks - they require an allocated dip and a clone of dio_bio.
8367          */
8368         if (io_bio && dip) {
8369                 io_bio->bi_error = -EIO;
8370                 bio_endio(io_bio);
8371                 /*
8372                  * The end io callbacks free our dip, do the final put on io_bio
8373                  * and all the cleanup and final put for dio_bio (through
8374                  * dio_end_io()).
8375                  */
8376                 dip = NULL;
8377                 io_bio = NULL;
8378         } else {
8379                 if (write)
8380                         btrfs_endio_direct_write_update_ordered(inode,
8381                                                 file_offset,
8382                                                 dio_bio->bi_iter.bi_size,
8383                                                 0);
8384                 else
8385                         unlock_extent(&BTRFS_I(inode)->io_tree, file_offset,
8386                               file_offset + dio_bio->bi_iter.bi_size - 1);
8387
8388                 dio_bio->bi_error = -EIO;
8389                 /*
8390                  * Releases and cleans up our dio_bio, no need to bio_put()
8391                  * nor bio_endio()/bio_io_error() against dio_bio.
8392                  */
8393                 dio_end_io(dio_bio, ret);
8394         }
8395         if (io_bio)
8396                 bio_put(io_bio);
8397         kfree(dip);
8398 }
8399
8400 static ssize_t check_direct_IO(struct btrfs_root *root, struct kiocb *iocb,
8401                         const struct iov_iter *iter, loff_t offset)
8402 {
8403         int seg;
8404         int i;
8405         unsigned blocksize_mask = root->sectorsize - 1;
8406         ssize_t retval = -EINVAL;
8407
8408         if (offset & blocksize_mask)
8409                 goto out;
8410
8411         if (iov_iter_alignment(iter) & blocksize_mask)
8412                 goto out;
8413
8414         /* If this is a write we don't need to check anymore */
8415         if (iov_iter_rw(iter) == WRITE)
8416                 return 0;
8417         /*
8418          * Check to make sure we don't have duplicate iov_base's in this
8419          * iovec, if so return EINVAL, otherwise we'll get csum errors
8420          * when reading back.
8421          */
8422         for (seg = 0; seg < iter->nr_segs; seg++) {
8423                 for (i = seg + 1; i < iter->nr_segs; i++) {
8424                         if (iter->iov[seg].iov_base == iter->iov[i].iov_base)
8425                                 goto out;
8426                 }
8427         }
8428         retval = 0;
8429 out:
8430         return retval;
8431 }
8432
8433 static ssize_t btrfs_direct_IO(struct kiocb *iocb, struct iov_iter *iter,
8434                                loff_t offset)
8435 {
8436         struct file *file = iocb->ki_filp;
8437         struct inode *inode = file->f_mapping->host;
8438         struct btrfs_root *root = BTRFS_I(inode)->root;
8439         struct btrfs_dio_data dio_data = { 0 };
8440         size_t count = 0;
8441         int flags = 0;
8442         bool wakeup = true;
8443         bool relock = false;
8444         ssize_t ret;
8445
8446         if (check_direct_IO(BTRFS_I(inode)->root, iocb, iter, offset))
8447                 return 0;
8448
8449         inode_dio_begin(inode);
8450         smp_mb__after_atomic();
8451
8452         /*
8453          * The generic stuff only does filemap_write_and_wait_range, which
8454          * isn't enough if we've written compressed pages to this area, so
8455          * we need to flush the dirty pages again to make absolutely sure
8456          * that any outstanding dirty pages are on disk.
8457          */
8458         count = iov_iter_count(iter);
8459         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
8460                      &BTRFS_I(inode)->runtime_flags))
8461                 filemap_fdatawrite_range(inode->i_mapping, offset,
8462                                          offset + count - 1);
8463
8464         if (iov_iter_rw(iter) == WRITE) {
8465                 /*
8466                  * If the write DIO is beyond the EOF, we need update
8467                  * the isize, but it is protected by i_mutex. So we can
8468                  * not unlock the i_mutex at this case.
8469                  */
8470                 if (offset + count <= inode->i_size) {
8471                         mutex_unlock(&inode->i_mutex);
8472                         relock = true;
8473                 }
8474                 ret = btrfs_delalloc_reserve_space(inode, offset, count);
8475                 if (ret)
8476                         goto out;
8477                 dio_data.outstanding_extents = div64_u64(count +
8478                                                 BTRFS_MAX_EXTENT_SIZE - 1,
8479                                                 BTRFS_MAX_EXTENT_SIZE);
8480
8481                 /*
8482                  * We need to know how many extents we reserved so that we can
8483                  * do the accounting properly if we go over the number we
8484                  * originally calculated.  Abuse current->journal_info for this.
8485                  */
8486                 dio_data.reserve = round_up(count, root->sectorsize);
8487                 dio_data.unsubmitted_oe_range_start = (u64)offset;
8488                 dio_data.unsubmitted_oe_range_end = (u64)offset;
8489                 current->journal_info = &dio_data;
8490         } else if (test_bit(BTRFS_INODE_READDIO_NEED_LOCK,
8491                                      &BTRFS_I(inode)->runtime_flags)) {
8492                 inode_dio_end(inode);
8493                 flags = DIO_LOCKING | DIO_SKIP_HOLES;
8494                 wakeup = false;
8495         }
8496
8497         ret = __blockdev_direct_IO(iocb, inode,
8498                                    BTRFS_I(inode)->root->fs_info->fs_devices->latest_bdev,
8499                                    iter, offset, btrfs_get_blocks_direct, NULL,
8500                                    btrfs_submit_direct, flags);
8501         if (iov_iter_rw(iter) == WRITE) {
8502                 current->journal_info = NULL;
8503                 if (ret < 0 && ret != -EIOCBQUEUED) {
8504                         if (dio_data.reserve)
8505                                 btrfs_delalloc_release_space(inode, offset,
8506                                                              dio_data.reserve);
8507                         /*
8508                          * On error we might have left some ordered extents
8509                          * without submitting corresponding bios for them, so
8510                          * cleanup them up to avoid other tasks getting them
8511                          * and waiting for them to complete forever.
8512                          */
8513                         if (dio_data.unsubmitted_oe_range_start <
8514                             dio_data.unsubmitted_oe_range_end)
8515                                 btrfs_endio_direct_write_update_ordered(inode,
8516                                         dio_data.unsubmitted_oe_range_start,
8517                                         dio_data.unsubmitted_oe_range_end -
8518                                         dio_data.unsubmitted_oe_range_start,
8519                                         0);
8520                 } else if (ret >= 0 && (size_t)ret < count)
8521                         btrfs_delalloc_release_space(inode, offset,
8522                                                      count - (size_t)ret);
8523         }
8524 out:
8525         if (wakeup)
8526                 inode_dio_end(inode);
8527         if (relock)
8528                 mutex_lock(&inode->i_mutex);
8529
8530         return ret;
8531 }
8532
8533 #define BTRFS_FIEMAP_FLAGS      (FIEMAP_FLAG_SYNC)
8534
8535 static int btrfs_fiemap(struct inode *inode, struct fiemap_extent_info *fieinfo,
8536                 __u64 start, __u64 len)
8537 {
8538         int     ret;
8539
8540         ret = fiemap_check_flags(fieinfo, BTRFS_FIEMAP_FLAGS);
8541         if (ret)
8542                 return ret;
8543
8544         return extent_fiemap(inode, fieinfo, start, len, btrfs_get_extent_fiemap);
8545 }
8546
8547 int btrfs_readpage(struct file *file, struct page *page)
8548 {
8549         struct extent_io_tree *tree;
8550         tree = &BTRFS_I(page->mapping->host)->io_tree;
8551         return extent_read_full_page(tree, page, btrfs_get_extent, 0);
8552 }
8553
8554 static int btrfs_writepage(struct page *page, struct writeback_control *wbc)
8555 {
8556         struct extent_io_tree *tree;
8557         struct inode *inode = page->mapping->host;
8558         int ret;
8559
8560         if (current->flags & PF_MEMALLOC) {
8561                 redirty_page_for_writepage(wbc, page);
8562                 unlock_page(page);
8563                 return 0;
8564         }
8565
8566         /*
8567          * If we are under memory pressure we will call this directly from the
8568          * VM, we need to make sure we have the inode referenced for the ordered
8569          * extent.  If not just return like we didn't do anything.
8570          */
8571         if (!igrab(inode)) {
8572                 redirty_page_for_writepage(wbc, page);
8573                 return AOP_WRITEPAGE_ACTIVATE;
8574         }
8575         tree = &BTRFS_I(page->mapping->host)->io_tree;
8576         ret = extent_write_full_page(tree, page, btrfs_get_extent, wbc);
8577         btrfs_add_delayed_iput(inode);
8578         return ret;
8579 }
8580
8581 static int btrfs_writepages(struct address_space *mapping,
8582                             struct writeback_control *wbc)
8583 {
8584         struct extent_io_tree *tree;
8585
8586         tree = &BTRFS_I(mapping->host)->io_tree;
8587         return extent_writepages(tree, mapping, btrfs_get_extent, wbc);
8588 }
8589
8590 static int
8591 btrfs_readpages(struct file *file, struct address_space *mapping,
8592                 struct list_head *pages, unsigned nr_pages)
8593 {
8594         struct extent_io_tree *tree;
8595         tree = &BTRFS_I(mapping->host)->io_tree;
8596         return extent_readpages(tree, mapping, pages, nr_pages,
8597                                 btrfs_get_extent);
8598 }
8599 static int __btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8600 {
8601         struct extent_io_tree *tree;
8602         struct extent_map_tree *map;
8603         int ret;
8604
8605         tree = &BTRFS_I(page->mapping->host)->io_tree;
8606         map = &BTRFS_I(page->mapping->host)->extent_tree;
8607         ret = try_release_extent_mapping(map, tree, page, gfp_flags);
8608         if (ret == 1) {
8609                 ClearPagePrivate(page);
8610                 set_page_private(page, 0);
8611                 page_cache_release(page);
8612         }
8613         return ret;
8614 }
8615
8616 static int btrfs_releasepage(struct page *page, gfp_t gfp_flags)
8617 {
8618         if (PageWriteback(page) || PageDirty(page))
8619                 return 0;
8620         return __btrfs_releasepage(page, gfp_flags & GFP_NOFS);
8621 }
8622
8623 static void btrfs_invalidatepage(struct page *page, unsigned int offset,
8624                                  unsigned int length)
8625 {
8626         struct inode *inode = page->mapping->host;
8627         struct extent_io_tree *tree;
8628         struct btrfs_ordered_extent *ordered;
8629         struct extent_state *cached_state = NULL;
8630         u64 page_start = page_offset(page);
8631         u64 page_end = page_start + PAGE_CACHE_SIZE - 1;
8632         int inode_evicting = inode->i_state & I_FREEING;
8633
8634         /*
8635          * we have the page locked, so new writeback can't start,
8636          * and the dirty bit won't be cleared while we are here.
8637          *
8638          * Wait for IO on this page so that we can safely clear
8639          * the PagePrivate2 bit and do ordered accounting
8640          */
8641         wait_on_page_writeback(page);
8642
8643         tree = &BTRFS_I(inode)->io_tree;
8644         if (offset) {
8645                 btrfs_releasepage(page, GFP_NOFS);
8646                 return;
8647         }
8648
8649         if (!inode_evicting)
8650                 lock_extent_bits(tree, page_start, page_end, &cached_state);
8651         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8652         if (ordered) {
8653                 /*
8654                  * IO on this page will never be started, so we need
8655                  * to account for any ordered extents now
8656                  */
8657                 if (!inode_evicting)
8658                         clear_extent_bit(tree, page_start, page_end,
8659                                          EXTENT_DIRTY | EXTENT_DELALLOC |
8660                                          EXTENT_LOCKED | EXTENT_DO_ACCOUNTING |
8661                                          EXTENT_DEFRAG, 1, 0, &cached_state,
8662                                          GFP_NOFS);
8663                 /*
8664                  * whoever cleared the private bit is responsible
8665                  * for the finish_ordered_io
8666                  */
8667                 if (TestClearPagePrivate2(page)) {
8668                         struct btrfs_ordered_inode_tree *tree;
8669                         u64 new_len;
8670
8671                         tree = &BTRFS_I(inode)->ordered_tree;
8672
8673                         spin_lock_irq(&tree->lock);
8674                         set_bit(BTRFS_ORDERED_TRUNCATED, &ordered->flags);
8675                         new_len = page_start - ordered->file_offset;
8676                         if (new_len < ordered->truncated_len)
8677                                 ordered->truncated_len = new_len;
8678                         spin_unlock_irq(&tree->lock);
8679
8680                         if (btrfs_dec_test_ordered_pending(inode, &ordered,
8681                                                            page_start,
8682                                                            PAGE_CACHE_SIZE, 1))
8683                                 btrfs_finish_ordered_io(ordered);
8684                 }
8685                 btrfs_put_ordered_extent(ordered);
8686                 if (!inode_evicting) {
8687                         cached_state = NULL;
8688                         lock_extent_bits(tree, page_start, page_end,
8689                                          &cached_state);
8690                 }
8691         }
8692
8693         /*
8694          * Qgroup reserved space handler
8695          * Page here will be either
8696          * 1) Already written to disk
8697          *    In this case, its reserved space is released from data rsv map
8698          *    and will be freed by delayed_ref handler finally.
8699          *    So even we call qgroup_free_data(), it won't decrease reserved
8700          *    space.
8701          * 2) Not written to disk
8702          *    This means the reserved space should be freed here.
8703          */
8704         btrfs_qgroup_free_data(inode, page_start, PAGE_CACHE_SIZE);
8705         if (!inode_evicting) {
8706                 clear_extent_bit(tree, page_start, page_end,
8707                                  EXTENT_LOCKED | EXTENT_DIRTY |
8708                                  EXTENT_DELALLOC | EXTENT_DO_ACCOUNTING |
8709                                  EXTENT_DEFRAG, 1, 1,
8710                                  &cached_state, GFP_NOFS);
8711
8712                 __btrfs_releasepage(page, GFP_NOFS);
8713         }
8714
8715         ClearPageChecked(page);
8716         if (PagePrivate(page)) {
8717                 ClearPagePrivate(page);
8718                 set_page_private(page, 0);
8719                 page_cache_release(page);
8720         }
8721 }
8722
8723 /*
8724  * btrfs_page_mkwrite() is not allowed to change the file size as it gets
8725  * called from a page fault handler when a page is first dirtied. Hence we must
8726  * be careful to check for EOF conditions here. We set the page up correctly
8727  * for a written page which means we get ENOSPC checking when writing into
8728  * holes and correct delalloc and unwritten extent mapping on filesystems that
8729  * support these features.
8730  *
8731  * We are not allowed to take the i_mutex here so we have to play games to
8732  * protect against truncate races as the page could now be beyond EOF.  Because
8733  * vmtruncate() writes the inode size before removing pages, once we have the
8734  * page lock we can determine safely if the page is beyond EOF. If it is not
8735  * beyond EOF, then the page is guaranteed safe against truncation until we
8736  * unlock the page.
8737  */
8738 int btrfs_page_mkwrite(struct vm_area_struct *vma, struct vm_fault *vmf)
8739 {
8740         struct page *page = vmf->page;
8741         struct inode *inode = file_inode(vma->vm_file);
8742         struct btrfs_root *root = BTRFS_I(inode)->root;
8743         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
8744         struct btrfs_ordered_extent *ordered;
8745         struct extent_state *cached_state = NULL;
8746         char *kaddr;
8747         unsigned long zero_start;
8748         loff_t size;
8749         int ret;
8750         int reserved = 0;
8751         u64 page_start;
8752         u64 page_end;
8753
8754         sb_start_pagefault(inode->i_sb);
8755         page_start = page_offset(page);
8756         page_end = page_start + PAGE_CACHE_SIZE - 1;
8757
8758         ret = btrfs_delalloc_reserve_space(inode, page_start,
8759                                            PAGE_CACHE_SIZE);
8760         if (!ret) {
8761                 ret = file_update_time(vma->vm_file);
8762                 reserved = 1;
8763         }
8764         if (ret) {
8765                 if (ret == -ENOMEM)
8766                         ret = VM_FAULT_OOM;
8767                 else /* -ENOSPC, -EIO, etc */
8768                         ret = VM_FAULT_SIGBUS;
8769                 if (reserved)
8770                         goto out;
8771                 goto out_noreserve;
8772         }
8773
8774         ret = VM_FAULT_NOPAGE; /* make the VM retry the fault */
8775 again:
8776         lock_page(page);
8777         size = i_size_read(inode);
8778
8779         if ((page->mapping != inode->i_mapping) ||
8780             (page_start >= size)) {
8781                 /* page got truncated out from underneath us */
8782                 goto out_unlock;
8783         }
8784         wait_on_page_writeback(page);
8785
8786         lock_extent_bits(io_tree, page_start, page_end, &cached_state);
8787         set_page_extent_mapped(page);
8788
8789         /*
8790          * we can't set the delalloc bits if there are pending ordered
8791          * extents.  Drop our locks and wait for them to finish
8792          */
8793         ordered = btrfs_lookup_ordered_extent(inode, page_start);
8794         if (ordered) {
8795                 unlock_extent_cached(io_tree, page_start, page_end,
8796                                      &cached_state, GFP_NOFS);
8797                 unlock_page(page);
8798                 btrfs_start_ordered_extent(inode, ordered, 1);
8799                 btrfs_put_ordered_extent(ordered);
8800                 goto again;
8801         }
8802
8803         /*
8804          * XXX - page_mkwrite gets called every time the page is dirtied, even
8805          * if it was already dirty, so for space accounting reasons we need to
8806          * clear any delalloc bits for the range we are fixing to save.  There
8807          * is probably a better way to do this, but for now keep consistent with
8808          * prepare_pages in the normal write path.
8809          */
8810         clear_extent_bit(&BTRFS_I(inode)->io_tree, page_start, page_end,
8811                           EXTENT_DIRTY | EXTENT_DELALLOC |
8812                           EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
8813                           0, 0, &cached_state, GFP_NOFS);
8814
8815         ret = btrfs_set_extent_delalloc(inode, page_start, page_end,
8816                                         &cached_state);
8817         if (ret) {
8818                 unlock_extent_cached(io_tree, page_start, page_end,
8819                                      &cached_state, GFP_NOFS);
8820                 ret = VM_FAULT_SIGBUS;
8821                 goto out_unlock;
8822         }
8823         ret = 0;
8824
8825         /* page is wholly or partially inside EOF */
8826         if (page_start + PAGE_CACHE_SIZE > size)
8827                 zero_start = size & ~PAGE_CACHE_MASK;
8828         else
8829                 zero_start = PAGE_CACHE_SIZE;
8830
8831         if (zero_start != PAGE_CACHE_SIZE) {
8832                 kaddr = kmap(page);
8833                 memset(kaddr + zero_start, 0, PAGE_CACHE_SIZE - zero_start);
8834                 flush_dcache_page(page);
8835                 kunmap(page);
8836         }
8837         ClearPageChecked(page);
8838         set_page_dirty(page);
8839         SetPageUptodate(page);
8840
8841         BTRFS_I(inode)->last_trans = root->fs_info->generation;
8842         BTRFS_I(inode)->last_sub_trans = BTRFS_I(inode)->root->log_transid;
8843         BTRFS_I(inode)->last_log_commit = BTRFS_I(inode)->root->last_log_commit;
8844
8845         unlock_extent_cached(io_tree, page_start, page_end, &cached_state, GFP_NOFS);
8846
8847 out_unlock:
8848         if (!ret) {
8849                 sb_end_pagefault(inode->i_sb);
8850                 return VM_FAULT_LOCKED;
8851         }
8852         unlock_page(page);
8853 out:
8854         btrfs_delalloc_release_space(inode, page_start, PAGE_CACHE_SIZE);
8855 out_noreserve:
8856         sb_end_pagefault(inode->i_sb);
8857         return ret;
8858 }
8859
8860 static int btrfs_truncate(struct inode *inode)
8861 {
8862         struct btrfs_root *root = BTRFS_I(inode)->root;
8863         struct btrfs_block_rsv *rsv;
8864         int ret = 0;
8865         int err = 0;
8866         struct btrfs_trans_handle *trans;
8867         u64 mask = root->sectorsize - 1;
8868         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
8869
8870         ret = btrfs_wait_ordered_range(inode, inode->i_size & (~mask),
8871                                        (u64)-1);
8872         if (ret)
8873                 return ret;
8874
8875         /*
8876          * Yes ladies and gentelment, this is indeed ugly.  The fact is we have
8877          * 3 things going on here
8878          *
8879          * 1) We need to reserve space for our orphan item and the space to
8880          * delete our orphan item.  Lord knows we don't want to have a dangling
8881          * orphan item because we didn't reserve space to remove it.
8882          *
8883          * 2) We need to reserve space to update our inode.
8884          *
8885          * 3) We need to have something to cache all the space that is going to
8886          * be free'd up by the truncate operation, but also have some slack
8887          * space reserved in case it uses space during the truncate (thank you
8888          * very much snapshotting).
8889          *
8890          * And we need these to all be seperate.  The fact is we can use alot of
8891          * space doing the truncate, and we have no earthly idea how much space
8892          * we will use, so we need the truncate reservation to be seperate so it
8893          * doesn't end up using space reserved for updating the inode or
8894          * removing the orphan item.  We also need to be able to stop the
8895          * transaction and start a new one, which means we need to be able to
8896          * update the inode several times, and we have no idea of knowing how
8897          * many times that will be, so we can't just reserve 1 item for the
8898          * entirety of the opration, so that has to be done seperately as well.
8899          * Then there is the orphan item, which does indeed need to be held on
8900          * to for the whole operation, and we need nobody to touch this reserved
8901          * space except the orphan code.
8902          *
8903          * So that leaves us with
8904          *
8905          * 1) root->orphan_block_rsv - for the orphan deletion.
8906          * 2) rsv - for the truncate reservation, which we will steal from the
8907          * transaction reservation.
8908          * 3) fs_info->trans_block_rsv - this will have 1 items worth left for
8909          * updating the inode.
8910          */
8911         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
8912         if (!rsv)
8913                 return -ENOMEM;
8914         rsv->size = min_size;
8915         rsv->failfast = 1;
8916
8917         /*
8918          * 1 for the truncate slack space
8919          * 1 for updating the inode.
8920          */
8921         trans = btrfs_start_transaction(root, 2);
8922         if (IS_ERR(trans)) {
8923                 err = PTR_ERR(trans);
8924                 goto out;
8925         }
8926
8927         /* Migrate the slack space for the truncate to our reserve */
8928         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
8929                                       min_size);
8930         BUG_ON(ret);
8931
8932         /*
8933          * So if we truncate and then write and fsync we normally would just
8934          * write the extents that changed, which is a problem if we need to
8935          * first truncate that entire inode.  So set this flag so we write out
8936          * all of the extents in the inode to the sync log so we're completely
8937          * safe.
8938          */
8939         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &BTRFS_I(inode)->runtime_flags);
8940         trans->block_rsv = rsv;
8941
8942         while (1) {
8943                 ret = btrfs_truncate_inode_items(trans, root, inode,
8944                                                  inode->i_size,
8945                                                  BTRFS_EXTENT_DATA_KEY);
8946                 if (ret != -ENOSPC && ret != -EAGAIN) {
8947                         err = ret;
8948                         break;
8949                 }
8950
8951                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8952                 ret = btrfs_update_inode(trans, root, inode);
8953                 if (ret) {
8954                         err = ret;
8955                         break;
8956                 }
8957
8958                 btrfs_end_transaction(trans, root);
8959                 btrfs_btree_balance_dirty(root);
8960
8961                 trans = btrfs_start_transaction(root, 2);
8962                 if (IS_ERR(trans)) {
8963                         ret = err = PTR_ERR(trans);
8964                         trans = NULL;
8965                         break;
8966                 }
8967
8968                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
8969                                               rsv, min_size);
8970                 BUG_ON(ret);    /* shouldn't happen */
8971                 trans->block_rsv = rsv;
8972         }
8973
8974         if (ret == 0 && inode->i_nlink > 0) {
8975                 trans->block_rsv = root->orphan_block_rsv;
8976                 ret = btrfs_orphan_del(trans, inode);
8977                 if (ret)
8978                         err = ret;
8979         }
8980
8981         if (trans) {
8982                 trans->block_rsv = &root->fs_info->trans_block_rsv;
8983                 ret = btrfs_update_inode(trans, root, inode);
8984                 if (ret && !err)
8985                         err = ret;
8986
8987                 ret = btrfs_end_transaction(trans, root);
8988                 btrfs_btree_balance_dirty(root);
8989         }
8990
8991 out:
8992         btrfs_free_block_rsv(root, rsv);
8993
8994         if (ret && !err)
8995                 err = ret;
8996
8997         return err;
8998 }
8999
9000 /*
9001  * create a new subvolume directory/inode (helper for the ioctl).
9002  */
9003 int btrfs_create_subvol_root(struct btrfs_trans_handle *trans,
9004                              struct btrfs_root *new_root,
9005                              struct btrfs_root *parent_root,
9006                              u64 new_dirid)
9007 {
9008         struct inode *inode;
9009         int err;
9010         u64 index = 0;
9011
9012         inode = btrfs_new_inode(trans, new_root, NULL, "..", 2,
9013                                 new_dirid, new_dirid,
9014                                 S_IFDIR | (~current_umask() & S_IRWXUGO),
9015                                 &index);
9016         if (IS_ERR(inode))
9017                 return PTR_ERR(inode);
9018         inode->i_op = &btrfs_dir_inode_operations;
9019         inode->i_fop = &btrfs_dir_file_operations;
9020
9021         set_nlink(inode, 1);
9022         btrfs_i_size_write(inode, 0);
9023         unlock_new_inode(inode);
9024
9025         err = btrfs_subvol_inherit_props(trans, new_root, parent_root);
9026         if (err)
9027                 btrfs_err(new_root->fs_info,
9028                           "error inheriting subvolume %llu properties: %d",
9029                           new_root->root_key.objectid, err);
9030
9031         err = btrfs_update_inode(trans, new_root, inode);
9032
9033         iput(inode);
9034         return err;
9035 }
9036
9037 struct inode *btrfs_alloc_inode(struct super_block *sb)
9038 {
9039         struct btrfs_inode *ei;
9040         struct inode *inode;
9041
9042         ei = kmem_cache_alloc(btrfs_inode_cachep, GFP_NOFS);
9043         if (!ei)
9044                 return NULL;
9045
9046         ei->root = NULL;
9047         ei->generation = 0;
9048         ei->last_trans = 0;
9049         ei->last_sub_trans = 0;
9050         ei->logged_trans = 0;
9051         ei->delalloc_bytes = 0;
9052         ei->defrag_bytes = 0;
9053         ei->disk_i_size = 0;
9054         ei->flags = 0;
9055         ei->csum_bytes = 0;
9056         ei->index_cnt = (u64)-1;
9057         ei->dir_index = 0;
9058         ei->last_unlink_trans = 0;
9059         ei->last_log_commit = 0;
9060         ei->delayed_iput_count = 0;
9061
9062         spin_lock_init(&ei->lock);
9063         ei->outstanding_extents = 0;
9064         ei->reserved_extents = 0;
9065
9066         ei->runtime_flags = 0;
9067         ei->force_compress = BTRFS_COMPRESS_NONE;
9068
9069         ei->delayed_node = NULL;
9070
9071         ei->i_otime.tv_sec = 0;
9072         ei->i_otime.tv_nsec = 0;
9073
9074         inode = &ei->vfs_inode;
9075         extent_map_tree_init(&ei->extent_tree);
9076         extent_io_tree_init(&ei->io_tree, &inode->i_data);
9077         extent_io_tree_init(&ei->io_failure_tree, &inode->i_data);
9078         ei->io_tree.track_uptodate = 1;
9079         ei->io_failure_tree.track_uptodate = 1;
9080         atomic_set(&ei->sync_writers, 0);
9081         mutex_init(&ei->log_mutex);
9082         mutex_init(&ei->delalloc_mutex);
9083         btrfs_ordered_inode_tree_init(&ei->ordered_tree);
9084         INIT_LIST_HEAD(&ei->delalloc_inodes);
9085         INIT_LIST_HEAD(&ei->delayed_iput);
9086         RB_CLEAR_NODE(&ei->rb_node);
9087
9088         return inode;
9089 }
9090
9091 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
9092 void btrfs_test_destroy_inode(struct inode *inode)
9093 {
9094         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9095         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9096 }
9097 #endif
9098
9099 static void btrfs_i_callback(struct rcu_head *head)
9100 {
9101         struct inode *inode = container_of(head, struct inode, i_rcu);
9102         kmem_cache_free(btrfs_inode_cachep, BTRFS_I(inode));
9103 }
9104
9105 void btrfs_destroy_inode(struct inode *inode)
9106 {
9107         struct btrfs_ordered_extent *ordered;
9108         struct btrfs_root *root = BTRFS_I(inode)->root;
9109
9110         WARN_ON(!hlist_empty(&inode->i_dentry));
9111         WARN_ON(inode->i_data.nrpages);
9112         WARN_ON(BTRFS_I(inode)->outstanding_extents);
9113         WARN_ON(BTRFS_I(inode)->reserved_extents);
9114         WARN_ON(BTRFS_I(inode)->delalloc_bytes);
9115         WARN_ON(BTRFS_I(inode)->csum_bytes);
9116         WARN_ON(BTRFS_I(inode)->defrag_bytes);
9117
9118         /*
9119          * This can happen where we create an inode, but somebody else also
9120          * created the same inode and we need to destroy the one we already
9121          * created.
9122          */
9123         if (!root)
9124                 goto free;
9125
9126         if (test_bit(BTRFS_INODE_HAS_ORPHAN_ITEM,
9127                      &BTRFS_I(inode)->runtime_flags)) {
9128                 btrfs_info(root->fs_info, "inode %llu still on the orphan list",
9129                         btrfs_ino(inode));
9130                 atomic_dec(&root->orphan_inodes);
9131         }
9132
9133         while (1) {
9134                 ordered = btrfs_lookup_first_ordered_extent(inode, (u64)-1);
9135                 if (!ordered)
9136                         break;
9137                 else {
9138                         btrfs_err(root->fs_info, "found ordered extent %llu %llu on inode cleanup",
9139                                 ordered->file_offset, ordered->len);
9140                         btrfs_remove_ordered_extent(inode, ordered);
9141                         btrfs_put_ordered_extent(ordered);
9142                         btrfs_put_ordered_extent(ordered);
9143                 }
9144         }
9145         btrfs_qgroup_check_reserved_leak(inode);
9146         inode_tree_del(inode);
9147         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
9148 free:
9149         call_rcu(&inode->i_rcu, btrfs_i_callback);
9150 }
9151
9152 int btrfs_drop_inode(struct inode *inode)
9153 {
9154         struct btrfs_root *root = BTRFS_I(inode)->root;
9155
9156         if (root == NULL)
9157                 return 1;
9158
9159         /* the snap/subvol tree is on deleting */
9160         if (btrfs_root_refs(&root->root_item) == 0)
9161                 return 1;
9162         else
9163                 return generic_drop_inode(inode);
9164 }
9165
9166 static void init_once(void *foo)
9167 {
9168         struct btrfs_inode *ei = (struct btrfs_inode *) foo;
9169
9170         inode_init_once(&ei->vfs_inode);
9171 }
9172
9173 void btrfs_destroy_cachep(void)
9174 {
9175         /*
9176          * Make sure all delayed rcu free inodes are flushed before we
9177          * destroy cache.
9178          */
9179         rcu_barrier();
9180         if (btrfs_inode_cachep)
9181                 kmem_cache_destroy(btrfs_inode_cachep);
9182         if (btrfs_trans_handle_cachep)
9183                 kmem_cache_destroy(btrfs_trans_handle_cachep);
9184         if (btrfs_transaction_cachep)
9185                 kmem_cache_destroy(btrfs_transaction_cachep);
9186         if (btrfs_path_cachep)
9187                 kmem_cache_destroy(btrfs_path_cachep);
9188         if (btrfs_free_space_cachep)
9189                 kmem_cache_destroy(btrfs_free_space_cachep);
9190 }
9191
9192 int btrfs_init_cachep(void)
9193 {
9194         btrfs_inode_cachep = kmem_cache_create("btrfs_inode",
9195                         sizeof(struct btrfs_inode), 0,
9196                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, init_once);
9197         if (!btrfs_inode_cachep)
9198                 goto fail;
9199
9200         btrfs_trans_handle_cachep = kmem_cache_create("btrfs_trans_handle",
9201                         sizeof(struct btrfs_trans_handle), 0,
9202                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9203         if (!btrfs_trans_handle_cachep)
9204                 goto fail;
9205
9206         btrfs_transaction_cachep = kmem_cache_create("btrfs_transaction",
9207                         sizeof(struct btrfs_transaction), 0,
9208                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9209         if (!btrfs_transaction_cachep)
9210                 goto fail;
9211
9212         btrfs_path_cachep = kmem_cache_create("btrfs_path",
9213                         sizeof(struct btrfs_path), 0,
9214                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9215         if (!btrfs_path_cachep)
9216                 goto fail;
9217
9218         btrfs_free_space_cachep = kmem_cache_create("btrfs_free_space",
9219                         sizeof(struct btrfs_free_space), 0,
9220                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD, NULL);
9221         if (!btrfs_free_space_cachep)
9222                 goto fail;
9223
9224         return 0;
9225 fail:
9226         btrfs_destroy_cachep();
9227         return -ENOMEM;
9228 }
9229
9230 static int btrfs_getattr(struct vfsmount *mnt,
9231                          struct dentry *dentry, struct kstat *stat)
9232 {
9233         u64 delalloc_bytes;
9234         struct inode *inode = d_inode(dentry);
9235         u32 blocksize = inode->i_sb->s_blocksize;
9236
9237         generic_fillattr(inode, stat);
9238         stat->dev = BTRFS_I(inode)->root->anon_dev;
9239         stat->blksize = PAGE_CACHE_SIZE;
9240
9241         spin_lock(&BTRFS_I(inode)->lock);
9242         delalloc_bytes = BTRFS_I(inode)->delalloc_bytes;
9243         spin_unlock(&BTRFS_I(inode)->lock);
9244         stat->blocks = (ALIGN(inode_get_bytes(inode), blocksize) +
9245                         ALIGN(delalloc_bytes, blocksize)) >> 9;
9246         return 0;
9247 }
9248
9249 static int btrfs_rename(struct inode *old_dir, struct dentry *old_dentry,
9250                            struct inode *new_dir, struct dentry *new_dentry)
9251 {
9252         struct btrfs_trans_handle *trans;
9253         struct btrfs_root *root = BTRFS_I(old_dir)->root;
9254         struct btrfs_root *dest = BTRFS_I(new_dir)->root;
9255         struct inode *new_inode = d_inode(new_dentry);
9256         struct inode *old_inode = d_inode(old_dentry);
9257         struct timespec ctime = CURRENT_TIME;
9258         u64 index = 0;
9259         u64 root_objectid;
9260         int ret;
9261         u64 old_ino = btrfs_ino(old_inode);
9262
9263         if (btrfs_ino(new_dir) == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)
9264                 return -EPERM;
9265
9266         /* we only allow rename subvolume link between subvolumes */
9267         if (old_ino != BTRFS_FIRST_FREE_OBJECTID && root != dest)
9268                 return -EXDEV;
9269
9270         if (old_ino == BTRFS_EMPTY_SUBVOL_DIR_OBJECTID ||
9271             (new_inode && btrfs_ino(new_inode) == BTRFS_FIRST_FREE_OBJECTID))
9272                 return -ENOTEMPTY;
9273
9274         if (S_ISDIR(old_inode->i_mode) && new_inode &&
9275             new_inode->i_size > BTRFS_EMPTY_DIR_SIZE)
9276                 return -ENOTEMPTY;
9277
9278
9279         /* check for collisions, even if the  name isn't there */
9280         ret = btrfs_check_dir_item_collision(dest, new_dir->i_ino,
9281                              new_dentry->d_name.name,
9282                              new_dentry->d_name.len);
9283
9284         if (ret) {
9285                 if (ret == -EEXIST) {
9286                         /* we shouldn't get
9287                          * eexist without a new_inode */
9288                         if (WARN_ON(!new_inode)) {
9289                                 return ret;
9290                         }
9291                 } else {
9292                         /* maybe -EOVERFLOW */
9293                         return ret;
9294                 }
9295         }
9296         ret = 0;
9297
9298         /*
9299          * we're using rename to replace one file with another.  Start IO on it
9300          * now so  we don't add too much work to the end of the transaction
9301          */
9302         if (new_inode && S_ISREG(old_inode->i_mode) && new_inode->i_size)
9303                 filemap_flush(old_inode->i_mapping);
9304
9305         /* close the racy window with snapshot create/destroy ioctl */
9306         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9307                 down_read(&root->fs_info->subvol_sem);
9308         /*
9309          * We want to reserve the absolute worst case amount of items.  So if
9310          * both inodes are subvols and we need to unlink them then that would
9311          * require 4 item modifications, but if they are both normal inodes it
9312          * would require 5 item modifications, so we'll assume their normal
9313          * inodes.  So 5 * 2 is 10, plus 1 for the new link, so 11 total items
9314          * should cover the worst case number of items we'll modify.
9315          */
9316         trans = btrfs_start_transaction(root, 11);
9317         if (IS_ERR(trans)) {
9318                 ret = PTR_ERR(trans);
9319                 goto out_notrans;
9320         }
9321
9322         if (dest != root)
9323                 btrfs_record_root_in_trans(trans, dest);
9324
9325         ret = btrfs_set_inode_index(new_dir, &index);
9326         if (ret)
9327                 goto out_fail;
9328
9329         BTRFS_I(old_inode)->dir_index = 0ULL;
9330         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9331                 /* force full log commit if subvolume involved. */
9332                 btrfs_set_log_full_commit(root->fs_info, trans);
9333         } else {
9334                 ret = btrfs_insert_inode_ref(trans, dest,
9335                                              new_dentry->d_name.name,
9336                                              new_dentry->d_name.len,
9337                                              old_ino,
9338                                              btrfs_ino(new_dir), index);
9339                 if (ret)
9340                         goto out_fail;
9341                 /*
9342                  * this is an ugly little race, but the rename is required
9343                  * to make sure that if we crash, the inode is either at the
9344                  * old name or the new one.  pinning the log transaction lets
9345                  * us make sure we don't allow a log commit to come in after
9346                  * we unlink the name but before we add the new name back in.
9347                  */
9348                 btrfs_pin_log_trans(root);
9349         }
9350
9351         inode_inc_iversion(old_dir);
9352         inode_inc_iversion(new_dir);
9353         inode_inc_iversion(old_inode);
9354         old_dir->i_ctime = old_dir->i_mtime = ctime;
9355         new_dir->i_ctime = new_dir->i_mtime = ctime;
9356         old_inode->i_ctime = ctime;
9357
9358         if (old_dentry->d_parent != new_dentry->d_parent)
9359                 btrfs_record_unlink_dir(trans, old_dir, old_inode, 1);
9360
9361         if (unlikely(old_ino == BTRFS_FIRST_FREE_OBJECTID)) {
9362                 root_objectid = BTRFS_I(old_inode)->root->root_key.objectid;
9363                 ret = btrfs_unlink_subvol(trans, root, old_dir, root_objectid,
9364                                         old_dentry->d_name.name,
9365                                         old_dentry->d_name.len);
9366         } else {
9367                 ret = __btrfs_unlink_inode(trans, root, old_dir,
9368                                         d_inode(old_dentry),
9369                                         old_dentry->d_name.name,
9370                                         old_dentry->d_name.len);
9371                 if (!ret)
9372                         ret = btrfs_update_inode(trans, root, old_inode);
9373         }
9374         if (ret) {
9375                 btrfs_abort_transaction(trans, root, ret);
9376                 goto out_fail;
9377         }
9378
9379         if (new_inode) {
9380                 inode_inc_iversion(new_inode);
9381                 new_inode->i_ctime = CURRENT_TIME;
9382                 if (unlikely(btrfs_ino(new_inode) ==
9383                              BTRFS_EMPTY_SUBVOL_DIR_OBJECTID)) {
9384                         root_objectid = BTRFS_I(new_inode)->location.objectid;
9385                         ret = btrfs_unlink_subvol(trans, dest, new_dir,
9386                                                 root_objectid,
9387                                                 new_dentry->d_name.name,
9388                                                 new_dentry->d_name.len);
9389                         BUG_ON(new_inode->i_nlink == 0);
9390                 } else {
9391                         ret = btrfs_unlink_inode(trans, dest, new_dir,
9392                                                  d_inode(new_dentry),
9393                                                  new_dentry->d_name.name,
9394                                                  new_dentry->d_name.len);
9395                 }
9396                 if (!ret && new_inode->i_nlink == 0)
9397                         ret = btrfs_orphan_add(trans, d_inode(new_dentry));
9398                 if (ret) {
9399                         btrfs_abort_transaction(trans, root, ret);
9400                         goto out_fail;
9401                 }
9402         }
9403
9404         ret = btrfs_add_link(trans, new_dir, old_inode,
9405                              new_dentry->d_name.name,
9406                              new_dentry->d_name.len, 0, index);
9407         if (ret) {
9408                 btrfs_abort_transaction(trans, root, ret);
9409                 goto out_fail;
9410         }
9411
9412         if (old_inode->i_nlink == 1)
9413                 BTRFS_I(old_inode)->dir_index = index;
9414
9415         if (old_ino != BTRFS_FIRST_FREE_OBJECTID) {
9416                 struct dentry *parent = new_dentry->d_parent;
9417                 btrfs_log_new_name(trans, old_inode, old_dir, parent);
9418                 btrfs_end_log_trans(root);
9419         }
9420 out_fail:
9421         btrfs_end_transaction(trans, root);
9422 out_notrans:
9423         if (old_ino == BTRFS_FIRST_FREE_OBJECTID)
9424                 up_read(&root->fs_info->subvol_sem);
9425
9426         return ret;
9427 }
9428
9429 static int btrfs_rename2(struct inode *old_dir, struct dentry *old_dentry,
9430                          struct inode *new_dir, struct dentry *new_dentry,
9431                          unsigned int flags)
9432 {
9433         if (flags & ~RENAME_NOREPLACE)
9434                 return -EINVAL;
9435
9436         return btrfs_rename(old_dir, old_dentry, new_dir, new_dentry);
9437 }
9438
9439 static void btrfs_run_delalloc_work(struct btrfs_work *work)
9440 {
9441         struct btrfs_delalloc_work *delalloc_work;
9442         struct inode *inode;
9443
9444         delalloc_work = container_of(work, struct btrfs_delalloc_work,
9445                                      work);
9446         inode = delalloc_work->inode;
9447         filemap_flush(inode->i_mapping);
9448         if (test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
9449                                 &BTRFS_I(inode)->runtime_flags))
9450                 filemap_flush(inode->i_mapping);
9451
9452         if (delalloc_work->delay_iput)
9453                 btrfs_add_delayed_iput(inode);
9454         else
9455                 iput(inode);
9456         complete(&delalloc_work->completion);
9457 }
9458
9459 struct btrfs_delalloc_work *btrfs_alloc_delalloc_work(struct inode *inode,
9460                                                     int delay_iput)
9461 {
9462         struct btrfs_delalloc_work *work;
9463
9464         work = kmalloc(sizeof(*work), GFP_NOFS);
9465         if (!work)
9466                 return NULL;
9467
9468         init_completion(&work->completion);
9469         INIT_LIST_HEAD(&work->list);
9470         work->inode = inode;
9471         work->delay_iput = delay_iput;
9472         WARN_ON_ONCE(!inode);
9473         btrfs_init_work(&work->work, btrfs_flush_delalloc_helper,
9474                         btrfs_run_delalloc_work, NULL, NULL);
9475
9476         return work;
9477 }
9478
9479 void btrfs_wait_and_free_delalloc_work(struct btrfs_delalloc_work *work)
9480 {
9481         wait_for_completion(&work->completion);
9482         kfree(work);
9483 }
9484
9485 /*
9486  * some fairly slow code that needs optimization. This walks the list
9487  * of all the inodes with pending delalloc and forces them to disk.
9488  */
9489 static int __start_delalloc_inodes(struct btrfs_root *root, int delay_iput,
9490                                    int nr)
9491 {
9492         struct btrfs_inode *binode;
9493         struct inode *inode;
9494         struct btrfs_delalloc_work *work, *next;
9495         struct list_head works;
9496         struct list_head splice;
9497         int ret = 0;
9498
9499         INIT_LIST_HEAD(&works);
9500         INIT_LIST_HEAD(&splice);
9501
9502         mutex_lock(&root->delalloc_mutex);
9503         spin_lock(&root->delalloc_lock);
9504         list_splice_init(&root->delalloc_inodes, &splice);
9505         while (!list_empty(&splice)) {
9506                 binode = list_entry(splice.next, struct btrfs_inode,
9507                                     delalloc_inodes);
9508
9509                 list_move_tail(&binode->delalloc_inodes,
9510                                &root->delalloc_inodes);
9511                 inode = igrab(&binode->vfs_inode);
9512                 if (!inode) {
9513                         cond_resched_lock(&root->delalloc_lock);
9514                         continue;
9515                 }
9516                 spin_unlock(&root->delalloc_lock);
9517
9518                 work = btrfs_alloc_delalloc_work(inode, delay_iput);
9519                 if (!work) {
9520                         if (delay_iput)
9521                                 btrfs_add_delayed_iput(inode);
9522                         else
9523                                 iput(inode);
9524                         ret = -ENOMEM;
9525                         goto out;
9526                 }
9527                 list_add_tail(&work->list, &works);
9528                 btrfs_queue_work(root->fs_info->flush_workers,
9529                                  &work->work);
9530                 ret++;
9531                 if (nr != -1 && ret >= nr)
9532                         goto out;
9533                 cond_resched();
9534                 spin_lock(&root->delalloc_lock);
9535         }
9536         spin_unlock(&root->delalloc_lock);
9537
9538 out:
9539         list_for_each_entry_safe(work, next, &works, list) {
9540                 list_del_init(&work->list);
9541                 btrfs_wait_and_free_delalloc_work(work);
9542         }
9543
9544         if (!list_empty_careful(&splice)) {
9545                 spin_lock(&root->delalloc_lock);
9546                 list_splice_tail(&splice, &root->delalloc_inodes);
9547                 spin_unlock(&root->delalloc_lock);
9548         }
9549         mutex_unlock(&root->delalloc_mutex);
9550         return ret;
9551 }
9552
9553 int btrfs_start_delalloc_inodes(struct btrfs_root *root, int delay_iput)
9554 {
9555         int ret;
9556
9557         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state))
9558                 return -EROFS;
9559
9560         ret = __start_delalloc_inodes(root, delay_iput, -1);
9561         if (ret > 0)
9562                 ret = 0;
9563         /*
9564          * the filemap_flush will queue IO into the worker threads, but
9565          * we have to make sure the IO is actually started and that
9566          * ordered extents get created before we return
9567          */
9568         atomic_inc(&root->fs_info->async_submit_draining);
9569         while (atomic_read(&root->fs_info->nr_async_submits) ||
9570               atomic_read(&root->fs_info->async_delalloc_pages)) {
9571                 wait_event(root->fs_info->async_submit_wait,
9572                    (atomic_read(&root->fs_info->nr_async_submits) == 0 &&
9573                     atomic_read(&root->fs_info->async_delalloc_pages) == 0));
9574         }
9575         atomic_dec(&root->fs_info->async_submit_draining);
9576         return ret;
9577 }
9578
9579 int btrfs_start_delalloc_roots(struct btrfs_fs_info *fs_info, int delay_iput,
9580                                int nr)
9581 {
9582         struct btrfs_root *root;
9583         struct list_head splice;
9584         int ret;
9585
9586         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state))
9587                 return -EROFS;
9588
9589         INIT_LIST_HEAD(&splice);
9590
9591         mutex_lock(&fs_info->delalloc_root_mutex);
9592         spin_lock(&fs_info->delalloc_root_lock);
9593         list_splice_init(&fs_info->delalloc_roots, &splice);
9594         while (!list_empty(&splice) && nr) {
9595                 root = list_first_entry(&splice, struct btrfs_root,
9596                                         delalloc_root);
9597                 root = btrfs_grab_fs_root(root);
9598                 BUG_ON(!root);
9599                 list_move_tail(&root->delalloc_root,
9600                                &fs_info->delalloc_roots);
9601                 spin_unlock(&fs_info->delalloc_root_lock);
9602
9603                 ret = __start_delalloc_inodes(root, delay_iput, nr);
9604                 btrfs_put_fs_root(root);
9605                 if (ret < 0)
9606                         goto out;
9607
9608                 if (nr != -1) {
9609                         nr -= ret;
9610                         WARN_ON(nr < 0);
9611                 }
9612                 spin_lock(&fs_info->delalloc_root_lock);
9613         }
9614         spin_unlock(&fs_info->delalloc_root_lock);
9615
9616         ret = 0;
9617         atomic_inc(&fs_info->async_submit_draining);
9618         while (atomic_read(&fs_info->nr_async_submits) ||
9619               atomic_read(&fs_info->async_delalloc_pages)) {
9620                 wait_event(fs_info->async_submit_wait,
9621                    (atomic_read(&fs_info->nr_async_submits) == 0 &&
9622                     atomic_read(&fs_info->async_delalloc_pages) == 0));
9623         }
9624         atomic_dec(&fs_info->async_submit_draining);
9625 out:
9626         if (!list_empty_careful(&splice)) {
9627                 spin_lock(&fs_info->delalloc_root_lock);
9628                 list_splice_tail(&splice, &fs_info->delalloc_roots);
9629                 spin_unlock(&fs_info->delalloc_root_lock);
9630         }
9631         mutex_unlock(&fs_info->delalloc_root_mutex);
9632         return ret;
9633 }
9634
9635 static int btrfs_symlink(struct inode *dir, struct dentry *dentry,
9636                          const char *symname)
9637 {
9638         struct btrfs_trans_handle *trans;
9639         struct btrfs_root *root = BTRFS_I(dir)->root;
9640         struct btrfs_path *path;
9641         struct btrfs_key key;
9642         struct inode *inode = NULL;
9643         int err;
9644         int drop_inode = 0;
9645         u64 objectid;
9646         u64 index = 0;
9647         int name_len;
9648         int datasize;
9649         unsigned long ptr;
9650         struct btrfs_file_extent_item *ei;
9651         struct extent_buffer *leaf;
9652
9653         name_len = strlen(symname);
9654         if (name_len > BTRFS_MAX_INLINE_DATA_SIZE(root))
9655                 return -ENAMETOOLONG;
9656
9657         /*
9658          * 2 items for inode item and ref
9659          * 2 items for dir items
9660          * 1 item for updating parent inode item
9661          * 1 item for the inline extent item
9662          * 1 item for xattr if selinux is on
9663          */
9664         trans = btrfs_start_transaction(root, 7);
9665         if (IS_ERR(trans))
9666                 return PTR_ERR(trans);
9667
9668         err = btrfs_find_free_ino(root, &objectid);
9669         if (err)
9670                 goto out_unlock;
9671
9672         inode = btrfs_new_inode(trans, root, dir, dentry->d_name.name,
9673                                 dentry->d_name.len, btrfs_ino(dir), objectid,
9674                                 S_IFLNK|S_IRWXUGO, &index);
9675         if (IS_ERR(inode)) {
9676                 err = PTR_ERR(inode);
9677                 goto out_unlock;
9678         }
9679
9680         /*
9681         * If the active LSM wants to access the inode during
9682         * d_instantiate it needs these. Smack checks to see
9683         * if the filesystem supports xattrs by looking at the
9684         * ops vector.
9685         */
9686         inode->i_fop = &btrfs_file_operations;
9687         inode->i_op = &btrfs_file_inode_operations;
9688         inode->i_mapping->a_ops = &btrfs_aops;
9689         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9690
9691         err = btrfs_init_inode_security(trans, inode, dir, &dentry->d_name);
9692         if (err)
9693                 goto out_unlock_inode;
9694
9695         path = btrfs_alloc_path();
9696         if (!path) {
9697                 err = -ENOMEM;
9698                 goto out_unlock_inode;
9699         }
9700         key.objectid = btrfs_ino(inode);
9701         key.offset = 0;
9702         key.type = BTRFS_EXTENT_DATA_KEY;
9703         datasize = btrfs_file_extent_calc_inline_size(name_len);
9704         err = btrfs_insert_empty_item(trans, root, path, &key,
9705                                       datasize);
9706         if (err) {
9707                 btrfs_free_path(path);
9708                 goto out_unlock_inode;
9709         }
9710         leaf = path->nodes[0];
9711         ei = btrfs_item_ptr(leaf, path->slots[0],
9712                             struct btrfs_file_extent_item);
9713         btrfs_set_file_extent_generation(leaf, ei, trans->transid);
9714         btrfs_set_file_extent_type(leaf, ei,
9715                                    BTRFS_FILE_EXTENT_INLINE);
9716         btrfs_set_file_extent_encryption(leaf, ei, 0);
9717         btrfs_set_file_extent_compression(leaf, ei, 0);
9718         btrfs_set_file_extent_other_encoding(leaf, ei, 0);
9719         btrfs_set_file_extent_ram_bytes(leaf, ei, name_len);
9720
9721         ptr = btrfs_file_extent_inline_start(ei);
9722         write_extent_buffer(leaf, symname, ptr, name_len);
9723         btrfs_mark_buffer_dirty(leaf);
9724         btrfs_free_path(path);
9725
9726         inode->i_op = &btrfs_symlink_inode_operations;
9727         inode->i_mapping->a_ops = &btrfs_symlink_aops;
9728         inode_set_bytes(inode, name_len);
9729         btrfs_i_size_write(inode, name_len);
9730         err = btrfs_update_inode(trans, root, inode);
9731         /*
9732          * Last step, add directory indexes for our symlink inode. This is the
9733          * last step to avoid extra cleanup of these indexes if an error happens
9734          * elsewhere above.
9735          */
9736         if (!err)
9737                 err = btrfs_add_nondir(trans, dir, dentry, inode, 0, index);
9738         if (err) {
9739                 drop_inode = 1;
9740                 goto out_unlock_inode;
9741         }
9742
9743         unlock_new_inode(inode);
9744         d_instantiate(dentry, inode);
9745
9746 out_unlock:
9747         btrfs_end_transaction(trans, root);
9748         if (drop_inode) {
9749                 inode_dec_link_count(inode);
9750                 iput(inode);
9751         }
9752         btrfs_btree_balance_dirty(root);
9753         return err;
9754
9755 out_unlock_inode:
9756         drop_inode = 1;
9757         unlock_new_inode(inode);
9758         goto out_unlock;
9759 }
9760
9761 static int __btrfs_prealloc_file_range(struct inode *inode, int mode,
9762                                        u64 start, u64 num_bytes, u64 min_size,
9763                                        loff_t actual_len, u64 *alloc_hint,
9764                                        struct btrfs_trans_handle *trans)
9765 {
9766         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
9767         struct extent_map *em;
9768         struct btrfs_root *root = BTRFS_I(inode)->root;
9769         struct btrfs_key ins;
9770         u64 cur_offset = start;
9771         u64 i_size;
9772         u64 cur_bytes;
9773         u64 last_alloc = (u64)-1;
9774         int ret = 0;
9775         bool own_trans = true;
9776
9777         if (trans)
9778                 own_trans = false;
9779         while (num_bytes > 0) {
9780                 if (own_trans) {
9781                         trans = btrfs_start_transaction(root, 3);
9782                         if (IS_ERR(trans)) {
9783                                 ret = PTR_ERR(trans);
9784                                 break;
9785                         }
9786                 }
9787
9788                 cur_bytes = min_t(u64, num_bytes, SZ_256M);
9789                 cur_bytes = max(cur_bytes, min_size);
9790                 /*
9791                  * If we are severely fragmented we could end up with really
9792                  * small allocations, so if the allocator is returning small
9793                  * chunks lets make its job easier by only searching for those
9794                  * sized chunks.
9795                  */
9796                 cur_bytes = min(cur_bytes, last_alloc);
9797                 ret = btrfs_reserve_extent(root, cur_bytes, min_size, 0,
9798                                            *alloc_hint, &ins, 1, 0);
9799                 if (ret) {
9800                         if (own_trans)
9801                                 btrfs_end_transaction(trans, root);
9802                         break;
9803                 }
9804
9805                 last_alloc = ins.offset;
9806                 ret = insert_reserved_file_extent(trans, inode,
9807                                                   cur_offset, ins.objectid,
9808                                                   ins.offset, ins.offset,
9809                                                   ins.offset, 0, 0, 0,
9810                                                   BTRFS_FILE_EXTENT_PREALLOC);
9811                 if (ret) {
9812                         btrfs_free_reserved_extent(root, ins.objectid,
9813                                                    ins.offset, 0);
9814                         btrfs_abort_transaction(trans, root, ret);
9815                         if (own_trans)
9816                                 btrfs_end_transaction(trans, root);
9817                         break;
9818                 }
9819
9820                 btrfs_drop_extent_cache(inode, cur_offset,
9821                                         cur_offset + ins.offset -1, 0);
9822
9823                 em = alloc_extent_map();
9824                 if (!em) {
9825                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
9826                                 &BTRFS_I(inode)->runtime_flags);
9827                         goto next;
9828                 }
9829
9830                 em->start = cur_offset;
9831                 em->orig_start = cur_offset;
9832                 em->len = ins.offset;
9833                 em->block_start = ins.objectid;
9834                 em->block_len = ins.offset;
9835                 em->orig_block_len = ins.offset;
9836                 em->ram_bytes = ins.offset;
9837                 em->bdev = root->fs_info->fs_devices->latest_bdev;
9838                 set_bit(EXTENT_FLAG_PREALLOC, &em->flags);
9839                 em->generation = trans->transid;
9840
9841                 while (1) {
9842                         write_lock(&em_tree->lock);
9843                         ret = add_extent_mapping(em_tree, em, 1);
9844                         write_unlock(&em_tree->lock);
9845                         if (ret != -EEXIST)
9846                                 break;
9847                         btrfs_drop_extent_cache(inode, cur_offset,
9848                                                 cur_offset + ins.offset - 1,
9849                                                 0);
9850                 }
9851                 free_extent_map(em);
9852 next:
9853                 num_bytes -= ins.offset;
9854                 cur_offset += ins.offset;
9855                 *alloc_hint = ins.objectid + ins.offset;
9856
9857                 inode_inc_iversion(inode);
9858                 inode->i_ctime = CURRENT_TIME;
9859                 BTRFS_I(inode)->flags |= BTRFS_INODE_PREALLOC;
9860                 if (!(mode & FALLOC_FL_KEEP_SIZE) &&
9861                     (actual_len > inode->i_size) &&
9862                     (cur_offset > inode->i_size)) {
9863                         if (cur_offset > actual_len)
9864                                 i_size = actual_len;
9865                         else
9866                                 i_size = cur_offset;
9867                         i_size_write(inode, i_size);
9868                         btrfs_ordered_update_i_size(inode, i_size, NULL);
9869                 }
9870
9871                 ret = btrfs_update_inode(trans, root, inode);
9872
9873                 if (ret) {
9874                         btrfs_abort_transaction(trans, root, ret);
9875                         if (own_trans)
9876                                 btrfs_end_transaction(trans, root);
9877                         break;
9878                 }
9879
9880                 if (own_trans)
9881                         btrfs_end_transaction(trans, root);
9882         }
9883         return ret;
9884 }
9885
9886 int btrfs_prealloc_file_range(struct inode *inode, int mode,
9887                               u64 start, u64 num_bytes, u64 min_size,
9888                               loff_t actual_len, u64 *alloc_hint)
9889 {
9890         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9891                                            min_size, actual_len, alloc_hint,
9892                                            NULL);
9893 }
9894
9895 int btrfs_prealloc_file_range_trans(struct inode *inode,
9896                                     struct btrfs_trans_handle *trans, int mode,
9897                                     u64 start, u64 num_bytes, u64 min_size,
9898                                     loff_t actual_len, u64 *alloc_hint)
9899 {
9900         return __btrfs_prealloc_file_range(inode, mode, start, num_bytes,
9901                                            min_size, actual_len, alloc_hint, trans);
9902 }
9903
9904 static int btrfs_set_page_dirty(struct page *page)
9905 {
9906         return __set_page_dirty_nobuffers(page);
9907 }
9908
9909 static int btrfs_permission(struct inode *inode, int mask)
9910 {
9911         struct btrfs_root *root = BTRFS_I(inode)->root;
9912         umode_t mode = inode->i_mode;
9913
9914         if (mask & MAY_WRITE &&
9915             (S_ISREG(mode) || S_ISDIR(mode) || S_ISLNK(mode))) {
9916                 if (btrfs_root_readonly(root))
9917                         return -EROFS;
9918                 if (BTRFS_I(inode)->flags & BTRFS_INODE_READONLY)
9919                         return -EACCES;
9920         }
9921         return generic_permission(inode, mask);
9922 }
9923
9924 static int btrfs_tmpfile(struct inode *dir, struct dentry *dentry, umode_t mode)
9925 {
9926         struct btrfs_trans_handle *trans;
9927         struct btrfs_root *root = BTRFS_I(dir)->root;
9928         struct inode *inode = NULL;
9929         u64 objectid;
9930         u64 index;
9931         int ret = 0;
9932
9933         /*
9934          * 5 units required for adding orphan entry
9935          */
9936         trans = btrfs_start_transaction(root, 5);
9937         if (IS_ERR(trans))
9938                 return PTR_ERR(trans);
9939
9940         ret = btrfs_find_free_ino(root, &objectid);
9941         if (ret)
9942                 goto out;
9943
9944         inode = btrfs_new_inode(trans, root, dir, NULL, 0,
9945                                 btrfs_ino(dir), objectid, mode, &index);
9946         if (IS_ERR(inode)) {
9947                 ret = PTR_ERR(inode);
9948                 inode = NULL;
9949                 goto out;
9950         }
9951
9952         inode->i_fop = &btrfs_file_operations;
9953         inode->i_op = &btrfs_file_inode_operations;
9954
9955         inode->i_mapping->a_ops = &btrfs_aops;
9956         BTRFS_I(inode)->io_tree.ops = &btrfs_extent_io_ops;
9957
9958         ret = btrfs_init_inode_security(trans, inode, dir, NULL);
9959         if (ret)
9960                 goto out_inode;
9961
9962         ret = btrfs_update_inode(trans, root, inode);
9963         if (ret)
9964                 goto out_inode;
9965         ret = btrfs_orphan_add(trans, inode);
9966         if (ret)
9967                 goto out_inode;
9968
9969         /*
9970          * We set number of links to 0 in btrfs_new_inode(), and here we set
9971          * it to 1 because d_tmpfile() will issue a warning if the count is 0,
9972          * through:
9973          *
9974          *    d_tmpfile() -> inode_dec_link_count() -> drop_nlink()
9975          */
9976         set_nlink(inode, 1);
9977         unlock_new_inode(inode);
9978         d_tmpfile(dentry, inode);
9979         mark_inode_dirty(inode);
9980
9981 out:
9982         btrfs_end_transaction(trans, root);
9983         if (ret)
9984                 iput(inode);
9985         btrfs_balance_delayed_items(root);
9986         btrfs_btree_balance_dirty(root);
9987         return ret;
9988
9989 out_inode:
9990         unlock_new_inode(inode);
9991         goto out;
9992
9993 }
9994
9995 /* Inspired by filemap_check_errors() */
9996 int btrfs_inode_check_errors(struct inode *inode)
9997 {
9998         int ret = 0;
9999
10000         if (test_bit(AS_ENOSPC, &inode->i_mapping->flags) &&
10001             test_and_clear_bit(AS_ENOSPC, &inode->i_mapping->flags))
10002                 ret = -ENOSPC;
10003         if (test_bit(AS_EIO, &inode->i_mapping->flags) &&
10004             test_and_clear_bit(AS_EIO, &inode->i_mapping->flags))
10005                 ret = -EIO;
10006
10007         return ret;
10008 }
10009
10010 static const struct inode_operations btrfs_dir_inode_operations = {
10011         .getattr        = btrfs_getattr,
10012         .lookup         = btrfs_lookup,
10013         .create         = btrfs_create,
10014         .unlink         = btrfs_unlink,
10015         .link           = btrfs_link,
10016         .mkdir          = btrfs_mkdir,
10017         .rmdir          = btrfs_rmdir,
10018         .rename2        = btrfs_rename2,
10019         .symlink        = btrfs_symlink,
10020         .setattr        = btrfs_setattr,
10021         .mknod          = btrfs_mknod,
10022         .setxattr       = btrfs_setxattr,
10023         .getxattr       = btrfs_getxattr,
10024         .listxattr      = btrfs_listxattr,
10025         .removexattr    = btrfs_removexattr,
10026         .permission     = btrfs_permission,
10027         .get_acl        = btrfs_get_acl,
10028         .set_acl        = btrfs_set_acl,
10029         .update_time    = btrfs_update_time,
10030         .tmpfile        = btrfs_tmpfile,
10031 };
10032 static const struct inode_operations btrfs_dir_ro_inode_operations = {
10033         .lookup         = btrfs_lookup,
10034         .permission     = btrfs_permission,
10035         .get_acl        = btrfs_get_acl,
10036         .set_acl        = btrfs_set_acl,
10037         .update_time    = btrfs_update_time,
10038 };
10039
10040 static const struct file_operations btrfs_dir_file_operations = {
10041         .llseek         = generic_file_llseek,
10042         .read           = generic_read_dir,
10043         .iterate        = btrfs_real_readdir,
10044         .unlocked_ioctl = btrfs_ioctl,
10045 #ifdef CONFIG_COMPAT
10046         .compat_ioctl   = btrfs_ioctl,
10047 #endif
10048         .release        = btrfs_release_file,
10049         .fsync          = btrfs_sync_file,
10050 };
10051
10052 static const struct extent_io_ops btrfs_extent_io_ops = {
10053         .fill_delalloc = run_delalloc_range,
10054         .submit_bio_hook = btrfs_submit_bio_hook,
10055         .merge_bio_hook = btrfs_merge_bio_hook,
10056         .readpage_end_io_hook = btrfs_readpage_end_io_hook,
10057         .writepage_end_io_hook = btrfs_writepage_end_io_hook,
10058         .writepage_start_hook = btrfs_writepage_start_hook,
10059         .set_bit_hook = btrfs_set_bit_hook,
10060         .clear_bit_hook = btrfs_clear_bit_hook,
10061         .merge_extent_hook = btrfs_merge_extent_hook,
10062         .split_extent_hook = btrfs_split_extent_hook,
10063 };
10064
10065 /*
10066  * btrfs doesn't support the bmap operation because swapfiles
10067  * use bmap to make a mapping of extents in the file.  They assume
10068  * these extents won't change over the life of the file and they
10069  * use the bmap result to do IO directly to the drive.
10070  *
10071  * the btrfs bmap call would return logical addresses that aren't
10072  * suitable for IO and they also will change frequently as COW
10073  * operations happen.  So, swapfile + btrfs == corruption.
10074  *
10075  * For now we're avoiding this by dropping bmap.
10076  */
10077 static const struct address_space_operations btrfs_aops = {
10078         .readpage       = btrfs_readpage,
10079         .writepage      = btrfs_writepage,
10080         .writepages     = btrfs_writepages,
10081         .readpages      = btrfs_readpages,
10082         .direct_IO      = btrfs_direct_IO,
10083         .invalidatepage = btrfs_invalidatepage,
10084         .releasepage    = btrfs_releasepage,
10085         .set_page_dirty = btrfs_set_page_dirty,
10086         .error_remove_page = generic_error_remove_page,
10087 };
10088
10089 static const struct address_space_operations btrfs_symlink_aops = {
10090         .readpage       = btrfs_readpage,
10091         .writepage      = btrfs_writepage,
10092         .invalidatepage = btrfs_invalidatepage,
10093         .releasepage    = btrfs_releasepage,
10094 };
10095
10096 static const struct inode_operations btrfs_file_inode_operations = {
10097         .getattr        = btrfs_getattr,
10098         .setattr        = btrfs_setattr,
10099         .setxattr       = btrfs_setxattr,
10100         .getxattr       = btrfs_getxattr,
10101         .listxattr      = btrfs_listxattr,
10102         .removexattr    = btrfs_removexattr,
10103         .permission     = btrfs_permission,
10104         .fiemap         = btrfs_fiemap,
10105         .get_acl        = btrfs_get_acl,
10106         .set_acl        = btrfs_set_acl,
10107         .update_time    = btrfs_update_time,
10108 };
10109 static const struct inode_operations btrfs_special_inode_operations = {
10110         .getattr        = btrfs_getattr,
10111         .setattr        = btrfs_setattr,
10112         .permission     = btrfs_permission,
10113         .setxattr       = btrfs_setxattr,
10114         .getxattr       = btrfs_getxattr,
10115         .listxattr      = btrfs_listxattr,
10116         .removexattr    = btrfs_removexattr,
10117         .get_acl        = btrfs_get_acl,
10118         .set_acl        = btrfs_set_acl,
10119         .update_time    = btrfs_update_time,
10120 };
10121 static const struct inode_operations btrfs_symlink_inode_operations = {
10122         .readlink       = generic_readlink,
10123         .follow_link    = page_follow_link_light,
10124         .put_link       = page_put_link,
10125         .getattr        = btrfs_getattr,
10126         .setattr        = btrfs_setattr,
10127         .permission     = btrfs_permission,
10128         .setxattr       = btrfs_setxattr,
10129         .getxattr       = btrfs_getxattr,
10130         .listxattr      = btrfs_listxattr,
10131         .removexattr    = btrfs_removexattr,
10132         .update_time    = btrfs_update_time,
10133 };
10134
10135 const struct dentry_operations btrfs_dentry_operations = {
10136         .d_delete       = btrfs_dentry_delete,
10137         .d_release      = btrfs_dentry_release,
10138 };