fs/btrfs/file.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/fs.h>
  20 #include <linux/pagemap.h>
  21 #include <linux/highmem.h>
  22 #include <linux/time.h>
  23 #include <linux/init.h>
  24 #include <linux/string.h>
  25 #include <linux/backing-dev.h>
  26 #include <linux/mpage.h>
  27 #include <linux/aio.h>
  28 #include <linux/falloc.h>
  29 #include <linux/swap.h>
  30 #include <linux/writeback.h>
  31 #include <linux/statfs.h>
  32 #include <linux/compat.h>
  33 #include <linux/slab.h>
  34 #include <linux/btrfs.h>
  35 #include "ctree.h"
  36 #include "disk-io.h"
  37 #include "transaction.h"
  38 #include "btrfs_inode.h"
  39 #include "print-tree.h"
  40 #include "tree-log.h"
  41 #include "locking.h"
  42 #include "volumes.h"
  43 #include "qgroup.h"
  44
  45 static struct kmem_cache *btrfs_inode_defrag_cachep;
  46 /*
  47  * when auto defrag is enabled we
  48  * queue up these defrag structs to remember which
  49  * inodes need defragging passes
  50  */
  51 struct inode_defrag {
  52         struct rb_node rb_node;
  53         /* objectid */
  54         u64 ino;
  55         /*
  56          * transid where the defrag was added, we search for
  57          * extents newer than this
  58          */
  59         u64 transid;
  60
  61         /* root objectid */
  62         u64 root;
  63
  64         /* last offset we were able to defrag */
  65         u64 last_offset;
  66
  67         /* if we've wrapped around back to zero once already */
  68         int cycled;
  69 };
  70
  71 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  72                                   struct inode_defrag *defrag2)
  73 {
  74         if (defrag1->root > defrag2->root)
  75                 return 1;
  76         else if (defrag1->root < defrag2->root)
  77                 return -1;
  78         else if (defrag1->ino > defrag2->ino)
  79                 return 1;
  80         else if (defrag1->ino < defrag2->ino)
  81                 return -1;
  82         else
  83                 return 0;
  84 }
  85
  86 /* pop a record for an inode into the defrag tree.  The lock
  87  * must be held already
  88  *
  89  * If you're inserting a record for an older transid than an
  90  * existing record, the transid already in the tree is lowered
  91  *
  92  * If an existing record is found the defrag item you
  93  * pass in is freed
  94  */
  95 static int __btrfs_add_inode_defrag(struct inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_root *root = BTRFS_I(inode)->root;
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &root->fs_info->defrag_inodes.rb_node;
 105         while (*p) {
 106                 parent = *p;
 107                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 108
 109                 ret = __compare_inode_defrag(defrag, entry);
 110                 if (ret < 0)
 111                         p = &parent->rb_left;
 112                 else if (ret > 0)
 113                         p = &parent->rb_right;
 114                 else {
 115                         /* if we're reinserting an entry for
 116                          * an old defrag run, make sure to
 117                          * lower the transid of our existing record
 118                          */
 119                         if (defrag->transid < entry->transid)
 120                                 entry->transid = defrag->transid;
 121                         if (defrag->last_offset > entry->last_offset)
 122                                 entry->last_offset = defrag->last_offset;
 123                         return -EEXIST;
 124                 }
 125         }
 126         set_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &root->fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_root *root)
 133 {
 134         if (!btrfs_test_opt(root, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(root->fs_info))
 138                 return 0;
 139
 140         return 1;
 141 }
 142
 143 /*
 144  * insert a defrag record for this inode if auto defrag is
 145  * enabled
 146  */
 147 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 148                            struct inode *inode)
 149 {
 150         struct btrfs_root *root = BTRFS_I(inode)->root;
 151         struct inode_defrag *defrag;
 152         u64 transid;
 153         int ret;
 154
 155         if (!__need_auto_defrag(root))
 156                 return 0;
 157
 158         if (test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags))
 159                 return 0;
 160
 161         if (trans)
 162                 transid = trans->transid;
 163         else
 164                 transid = BTRFS_I(inode)->root->last_trans;
 165
 166         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 167         if (!defrag)
 168                 return -ENOMEM;
 169
 170         defrag->ino = btrfs_ino(inode);
 171         defrag->transid = transid;
 172         defrag->root = root->root_key.objectid;
 173
 174         spin_lock(&root->fs_info->defrag_inodes_lock);
 175         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags)) {
 176                 /*
 177                  * If we set IN_DEFRAG flag and evict the inode from memory,
 178                  * and then re-read this inode, this new inode doesn't have
 179                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 180                  */
 181                 ret = __btrfs_add_inode_defrag(inode, defrag);
 182                 if (ret)
 183                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 184         } else {
 185                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 186         }
 187         spin_unlock(&root->fs_info->defrag_inodes_lock);
 188         return 0;
 189 }
 190
 191 /*
 192  * Requeue the defrag object. If there is a defrag object that points to
 193  * the same inode in the tree, we will merge them together (by
 194  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 195  */
 196 static void btrfs_requeue_inode_defrag(struct inode *inode,
 197                                        struct inode_defrag *defrag)
 198 {
 199         struct btrfs_root *root = BTRFS_I(inode)->root;
 200         int ret;
 201
 202         if (!__need_auto_defrag(root))
 203                 goto out;
 204
 205         /*
 206          * Here we don't check the IN_DEFRAG flag, because we need merge
 207          * them together.
 208          */
 209         spin_lock(&root->fs_info->defrag_inodes_lock);
 210         ret = __btrfs_add_inode_defrag(inode, defrag);
 211         spin_unlock(&root->fs_info->defrag_inodes_lock);
 212         if (ret)
 213                 goto out;
 214         return;
 215 out:
 216         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 217 }
 218
 219 /*
 220  * pick the defragable inode that we want, if it doesn't exist, we will get
 221  * the next one.
 222  */
 223 static struct inode_defrag *
 224 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 225 {
 226         struct inode_defrag *entry = NULL;
 227         struct inode_defrag tmp;
 228         struct rb_node *p;
 229         struct rb_node *parent = NULL;
 230         int ret;
 231
 232         tmp.ino = ino;
 233         tmp.root = root;
 234
 235         spin_lock(&fs_info->defrag_inodes_lock);
 236         p = fs_info->defrag_inodes.rb_node;
 237         while (p) {
 238                 parent = p;
 239                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 240
 241                 ret = __compare_inode_defrag(&tmp, entry);
 242                 if (ret < 0)
 243                         p = parent->rb_left;
 244                 else if (ret > 0)
 245                         p = parent->rb_right;
 246                 else
 247                         goto out;
 248         }
 249
 250         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 251                 parent = rb_next(parent);
 252                 if (parent)
 253                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 254                 else
 255                         entry = NULL;
 256         }
 257 out:
 258         if (entry)
 259                 rb_erase(parent, &fs_info->defrag_inodes);
 260         spin_unlock(&fs_info->defrag_inodes_lock);
 261         return entry;
 262 }
 263
 264 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 265 {
 266         struct inode_defrag *defrag;
 267         struct rb_node *node;
 268
 269         spin_lock(&fs_info->defrag_inodes_lock);
 270         node = rb_first(&fs_info->defrag_inodes);
 271         while (node) {
 272                 rb_erase(node, &fs_info->defrag_inodes);
 273                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 274                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 275
 276                 if (need_resched()) {
 277                         spin_unlock(&fs_info->defrag_inodes_lock);
 278                         cond_resched();
 279                         spin_lock(&fs_info->defrag_inodes_lock);
 280                 }
 281
 282                 node = rb_first(&fs_info->defrag_inodes);
 283         }
 284         spin_unlock(&fs_info->defrag_inodes_lock);
 285 }
 286
 287 #define BTRFS_DEFRAG_BATCH      1024
 288
 289 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 290                                     struct inode_defrag *defrag)
 291 {
 292         struct btrfs_root *inode_root;
 293         struct inode *inode;
 294         struct btrfs_key key;
 295         struct btrfs_ioctl_defrag_range_args range;
 296         int num_defrag;
 297         int index;
 298         int ret;
 299
 300         /* get the inode */
 301         key.objectid = defrag->root;
 302         btrfs_set_key_type(&key, BTRFS_ROOT_ITEM_KEY);
 303         key.offset = (u64)-1;
 304
 305         index = srcu_read_lock(&fs_info->subvol_srcu);
 306
 307         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 308         if (IS_ERR(inode_root)) {
 309                 ret = PTR_ERR(inode_root);
 310                 goto cleanup;
 311         }
 312
 313         key.objectid = defrag->ino;
 314         btrfs_set_key_type(&key, BTRFS_INODE_ITEM_KEY);
 315         key.offset = 0;
 316         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 317         if (IS_ERR(inode)) {
 318                 ret = PTR_ERR(inode);
 319                 goto cleanup;
 320         }
 321         srcu_read_unlock(&fs_info->subvol_srcu, index);
 322
 323         /* do a chunk of defrag */
 324         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 325         memset(&range, 0, sizeof(range));
 326         range.len = (u64)-1;
 327         range.start = defrag->last_offset;
 328
 329         sb_start_write(fs_info->sb);
 330         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 331                                        BTRFS_DEFRAG_BATCH);
 332         sb_end_write(fs_info->sb);
 333         /*
 334          * if we filled the whole defrag batch, there
 335          * must be more work to do.  Queue this defrag
 336          * again
 337          */
 338         if (num_defrag == BTRFS_DEFRAG_BATCH) {
 339                 defrag->last_offset = range.start;
 340                 btrfs_requeue_inode_defrag(inode, defrag);
 341         } else if (defrag->last_offset && !defrag->cycled) {
 342                 /*
 343                  * we didn't fill our defrag batch, but
 344                  * we didn't start at zero.  Make sure we loop
 345                  * around to the start of the file.
 346                  */
 347                 defrag->last_offset = 0;
 348                 defrag->cycled = 1;
 349                 btrfs_requeue_inode_defrag(inode, defrag);
 350         } else {
 351                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 352         }
 353
 354         iput(inode);
 355         return 0;
 356 cleanup:
 357         srcu_read_unlock(&fs_info->subvol_srcu, index);
 358         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 359         return ret;
 360 }
 361
 362 /*
 363  * run through the list of inodes in the FS that need
 364  * defragging
 365  */
 366 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 367 {
 368         struct inode_defrag *defrag;
 369         u64 first_ino = 0;
 370         u64 root_objectid = 0;
 371
 372         atomic_inc(&fs_info->defrag_running);
 373         while (1) {
 374                 /* Pause the auto defragger. */
 375                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 376                              &fs_info->fs_state))
 377                         break;
 378
 379                 if (!__need_auto_defrag(fs_info->tree_root))
 380                         break;
 381
 382                 /* find an inode to defrag */
 383                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 384                                                  first_ino);
 385                 if (!defrag) {
 386                         if (root_objectid || first_ino) {
 387                                 root_objectid = 0;
 388                                 first_ino = 0;
 389                                 continue;
 390                         } else {
 391                                 break;
 392                         }
 393                 }
 394
 395                 first_ino = defrag->ino + 1;
 396                 root_objectid = defrag->root;
 397
 398                 __btrfs_run_defrag_inode(fs_info, defrag);
 399         }
 400         atomic_dec(&fs_info->defrag_running);
 401
 402         /*
 403          * during unmount, we use the transaction_wait queue to
 404          * wait for the defragger to stop
 405          */
 406         wake_up(&fs_info->transaction_wait);
 407         return 0;
 408 }
 409
 410 /* simple helper to fault in pages and copy.  This should go away
 411  * and be replaced with calls into generic code.
 412  */
 413 static noinline int btrfs_copy_from_user(loff_t pos, int num_pages,
 414                                          size_t write_bytes,
 415                                          struct page **prepared_pages,
 416                                          struct iov_iter *i)
 417 {
 418         size_t copied = 0;
 419         size_t total_copied = 0;
 420         int pg = 0;
 421         int offset = pos & (PAGE_CACHE_SIZE - 1);
 422
 423         while (write_bytes > 0) {
 424                 size_t count = min_t(size_t,
 425                                      PAGE_CACHE_SIZE - offset, write_bytes);
 426                 struct page *page = prepared_pages[pg];
 427                 /*
 428                  * Copy data from userspace to the current page
 429                  */
 430                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 431
 432                 /* Flush processor's dcache for this page */
 433                 flush_dcache_page(page);
 434
 435                 /*
 436                  * if we get a partial write, we can end up with
 437                  * partially up to date pages.  These add
 438                  * a lot of complexity, so make sure they don't
 439                  * happen by forcing this copy to be retried.
 440                  *
 441                  * The rest of the btrfs_file_write code will fall
 442                  * back to page at a time copies after we return 0.
 443                  */
 444                 if (!PageUptodate(page) && copied < count)
 445                         copied = 0;
 446
 447                 iov_iter_advance(i, copied);
 448                 write_bytes -= copied;
 449                 total_copied += copied;
 450
 451                 /* Return to btrfs_file_aio_write to fault page */
 452                 if (unlikely(copied == 0))
 453                         break;
 454
 455                 if (unlikely(copied < PAGE_CACHE_SIZE - offset)) {
 456                         offset += copied;
 457                 } else {
 458                         pg++;
 459                         offset = 0;
 460                 }
 461         }
 462         return total_copied;
 463 }
 464
 465 /*
 466  * unlocks pages after btrfs_file_write is done with them
 467  */
 468 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 469 {
 470         size_t i;
 471         for (i = 0; i < num_pages; i++) {
 472                 /* page checked is some magic around finding pages that
 473                  * have been modified without going through btrfs_set_page_dirty
 474                  * clear it here
 475                  */
 476                 ClearPageChecked(pages[i]);
 477                 unlock_page(pages[i]);
 478                 mark_page_accessed(pages[i]);
 479                 page_cache_release(pages[i]);
 480         }
 481 }
 482
 483 /*
 484  * after copy_from_user, pages need to be dirtied and we need to make
 485  * sure holes are created between the current EOF and the start of
 486  * any next extents (if required).
 487  *
 488  * this also makes the decision about creating an inline extent vs
 489  * doing real data extents, marking pages dirty and delalloc as required.
 490  */
 491 int btrfs_dirty_pages(struct btrfs_root *root, struct inode *inode,
 492                              struct page **pages, size_t num_pages,
 493                              loff_t pos, size_t write_bytes,
 494                              struct extent_state **cached)
 495 {
 496         int err = 0;
 497         int i;
 498         u64 num_bytes;
 499         u64 start_pos;
 500         u64 end_of_last_block;
 501         u64 end_pos = pos + write_bytes;
 502         loff_t isize = i_size_read(inode);
 503
 504         start_pos = pos & ~((u64)root->sectorsize - 1);
 505         num_bytes = ALIGN(write_bytes + pos - start_pos, root->sectorsize);
 506
 507         end_of_last_block = start_pos + num_bytes - 1;
 508         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 509                                         cached);
 510         if (err)
 511                 return err;
 512
 513         for (i = 0; i < num_pages; i++) {
 514                 struct page *p = pages[i];
 515                 SetPageUptodate(p);
 516                 ClearPageChecked(p);
 517                 set_page_dirty(p);
 518         }
 519
 520         /*
 521          * we've only changed i_size in ram, and we haven't updated
 522          * the disk i_size.  There is no need to log the inode
 523          * at this time.
 524          */
 525         if (end_pos > isize)
 526                 i_size_write(inode, end_pos);
 527         return 0;
 528 }
 529
 530 /*
 531  * this drops all the extents in the cache that intersect the range
 532  * [start, end].  Existing extents are split as required.
 533  */
 534 void btrfs_drop_extent_cache(struct inode *inode, u64 start, u64 end,
 535                              int skip_pinned)
 536 {
 537         struct extent_map *em;
 538         struct extent_map *split = NULL;
 539         struct extent_map *split2 = NULL;
 540         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
 541         u64 len = end - start + 1;
 542         u64 gen;
 543         int ret;
 544         int testend = 1;
 545         unsigned long flags;
 546         int compressed = 0;
 547         bool modified;
 548
 549         WARN_ON(end < start);
 550         if (end == (u64)-1) {
 551                 len = (u64)-1;
 552                 testend = 0;
 553         }
 554         while (1) {
 555                 int no_splits = 0;
 556
 557                 modified = false;
 558                 if (!split)
 559                         split = alloc_extent_map();
 560                 if (!split2)
 561                         split2 = alloc_extent_map();
 562                 if (!split || !split2)
 563                         no_splits = 1;
 564
 565                 write_lock(&em_tree->lock);
 566                 em = lookup_extent_mapping(em_tree, start, len);
 567                 if (!em) {
 568                         write_unlock(&em_tree->lock);
 569                         break;
 570                 }
 571                 flags = em->flags;
 572                 gen = em->generation;
 573                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 574                         if (testend && em->start + em->len >= start + len) {
 575                                 free_extent_map(em);
 576                                 write_unlock(&em_tree->lock);
 577                                 break;
 578                         }
 579                         start = em->start + em->len;
 580                         if (testend)
 581                                 len = start + len - (em->start + em->len);
 582                         free_extent_map(em);
 583                         write_unlock(&em_tree->lock);
 584                         continue;
 585                 }
 586                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 587                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 588                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 589                 modified = !list_empty(&em->list);
 590                 if (no_splits)
 591                         goto next;
 592
 593                 if (em->start < start) {
 594                         split->start = em->start;
 595                         split->len = start - em->start;
 596
 597                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 598                                 split->orig_start = em->orig_start;
 599                                 split->block_start = em->block_start;
 600
 601                                 if (compressed)
 602                                         split->block_len = em->block_len;
 603                                 else
 604                                         split->block_len = split->len;
 605                                 split->orig_block_len = max(split->block_len,
 606                                                 em->orig_block_len);
 607                                 split->ram_bytes = em->ram_bytes;
 608                         } else {
 609                                 split->orig_start = split->start;
 610                                 split->block_len = 0;
 611                                 split->block_start = em->block_start;
 612                                 split->orig_block_len = 0;
 613                                 split->ram_bytes = split->len;
 614                         }
 615
 616                         split->generation = gen;
 617                         split->bdev = em->bdev;
 618                         split->flags = flags;
 619                         split->compress_type = em->compress_type;
 620                         replace_extent_mapping(em_tree, em, split, modified);
 621                         free_extent_map(split);
 622                         split = split2;
 623                         split2 = NULL;
 624                 }
 625                 if (testend && em->start + em->len > start + len) {
 626                         u64 diff = start + len - em->start;
 627
 628                         split->start = start + len;
 629                         split->len = em->start + em->len - (start + len);
 630                         split->bdev = em->bdev;
 631                         split->flags = flags;
 632                         split->compress_type = em->compress_type;
 633                         split->generation = gen;
 634
 635                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 636                                 split->orig_block_len = max(em->block_len,
 637                                                     em->orig_block_len);
 638
 639                                 split->ram_bytes = em->ram_bytes;
 640                                 if (compressed) {
 641                                         split->block_len = em->block_len;
 642                                         split->block_start = em->block_start;
 643                                         split->orig_start = em->orig_start;
 644                                 } else {
 645                                         split->block_len = split->len;
 646                                         split->block_start = em->block_start
 647                                                 + diff;
 648                                         split->orig_start = em->orig_start;
 649                                 }
 650                         } else {
 651                                 split->ram_bytes = split->len;
 652                                 split->orig_start = split->start;
 653                                 split->block_len = 0;
 654                                 split->block_start = em->block_start;
 655                                 split->orig_block_len = 0;
 656                         }
 657
 658                         if (extent_map_in_tree(em)) {
 659                                 replace_extent_mapping(em_tree, em, split,
 660                                                        modified);
 661                         } else {
 662                                 ret = add_extent_mapping(em_tree, split,
 663                                                          modified);
 664                                 ASSERT(ret == 0); /* Logic error */
 665                         }
 666                         free_extent_map(split);
 667                         split = NULL;
 668                 }
 669 next:
 670                 if (extent_map_in_tree(em))
 671                         remove_extent_mapping(em_tree, em);
 672                 write_unlock(&em_tree->lock);
 673
 674                 /* once for us */
 675                 free_extent_map(em);
 676                 /* once for the tree*/
 677                 free_extent_map(em);
 678         }
 679         if (split)
 680                 free_extent_map(split);
 681         if (split2)
 682                 free_extent_map(split2);
 683 }
 684
 685 /*
 686  * this is very complex, but the basic idea is to drop all extents
 687  * in the range start - end.  hint_block is filled in with a block number
 688  * that would be a good hint to the block allocator for this file.
 689  *
 690  * If an extent intersects the range but is not entirely inside the range
 691  * it is either truncated or split.  Anything entirely inside the range
 692  * is deleted from the tree.
 693  */
 694 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 695                          struct btrfs_root *root, struct inode *inode,
 696                          struct btrfs_path *path, u64 start, u64 end,
 697                          u64 *drop_end, int drop_cache,
 698                          int replace_extent,
 699                          u32 extent_item_size,
 700                          int *key_inserted)
 701 {
 702         struct extent_buffer *leaf;
 703         struct btrfs_file_extent_item *fi;
 704         struct btrfs_key key;
 705         struct btrfs_key new_key;
 706         u64 ino = btrfs_ino(inode);
 707         u64 search_start = start;
 708         u64 disk_bytenr = 0;
 709         u64 num_bytes = 0;
 710         u64 extent_offset = 0;
 711         u64 extent_end = 0;
 712         int del_nr = 0;
 713         int del_slot = 0;
 714         int extent_type;
 715         int recow;
 716         int ret;
 717         int modify_tree = -1;
 718         int update_refs;
 719         int found = 0;
 720         int leafs_visited = 0;
 721
 722         if (drop_cache)
 723                 btrfs_drop_extent_cache(inode, start, end - 1, 0);
 724
 725         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 726                 modify_tree = 0;
 727
 728         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 729                        root == root->fs_info->tree_root);
 730         while (1) {
 731                 recow = 0;
 732                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 733                                                search_start, modify_tree);
 734                 if (ret < 0)
 735                         break;
 736                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 737                         leaf = path->nodes[0];
 738                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 739                         if (key.objectid == ino &&
 740                             key.type == BTRFS_EXTENT_DATA_KEY)
 741                                 path->slots[0]--;
 742                 }
 743                 ret = 0;
 744                 leafs_visited++;
 745 next_slot:
 746                 leaf = path->nodes[0];
 747                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 748                         BUG_ON(del_nr > 0);
 749                         ret = btrfs_next_leaf(root, path);
 750                         if (ret < 0)
 751                                 break;
 752                         if (ret > 0) {
 753                                 ret = 0;
 754                                 break;
 755                         }
 756                         leafs_visited++;
 757                         leaf = path->nodes[0];
 758                         recow = 1;
 759                 }
 760
 761                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 762                 if (key.objectid > ino ||
 763                     key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 764                         break;
 765
 766                 fi = btrfs_item_ptr(leaf, path->slots[0],
 767                                     struct btrfs_file_extent_item);
 768                 extent_type = btrfs_file_extent_type(leaf, fi);
 769
 770                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 771                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 772                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 773                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 774                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 775                         extent_end = key.offset +
 776                                 btrfs_file_extent_num_bytes(leaf, fi);
 777                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 778                         extent_end = key.offset +
 779                                 btrfs_file_extent_inline_len(leaf,
 780                                                      path->slots[0], fi);
 781                 } else {
 782                         WARN_ON(1);
 783                         extent_end = search_start;
 784                 }
 785
 786                 /*
 787                  * Don't skip extent items representing 0 byte lengths. They
 788                  * used to be created (bug) if while punching holes we hit
 789                  * -ENOSPC condition. So if we find one here, just ensure we
 790                  * delete it, otherwise we would insert a new file extent item
 791                  * with the same key (offset) as that 0 bytes length file
 792                  * extent item in the call to setup_items_for_insert() later
 793                  * in this function.
 794                  */
 795                 if (extent_end == key.offset && extent_end >= search_start)
 796                         goto delete_extent_item;
 797
 798                 if (extent_end <= search_start) {
 799                         path->slots[0]++;
 800                         goto next_slot;
 801                 }
 802
 803                 found = 1;
 804                 search_start = max(key.offset, start);
 805                 if (recow || !modify_tree) {
 806                         modify_tree = -1;
 807                         btrfs_release_path(path);
 808                         continue;
 809                 }
 810
 811                 /*
 812                  *     | - range to drop - |
 813                  *  | -------- extent -------- |
 814                  */
 815                 if (start > key.offset && end < extent_end) {
 816                         BUG_ON(del_nr > 0);
 817                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 818                                 ret = -EOPNOTSUPP;
 819                                 break;
 820                         }
 821
 822                         memcpy(&new_key, &key, sizeof(new_key));
 823                         new_key.offset = start;
 824                         ret = btrfs_duplicate_item(trans, root, path,
 825                                                    &new_key);
 826                         if (ret == -EAGAIN) {
 827                                 btrfs_release_path(path);
 828                                 continue;
 829                         }
 830                         if (ret < 0)
 831                                 break;
 832
 833                         leaf = path->nodes[0];
 834                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 835                                             struct btrfs_file_extent_item);
 836                         btrfs_set_file_extent_num_bytes(leaf, fi,
 837                                                         start - key.offset);
 838
 839                         fi = btrfs_item_ptr(leaf, path->slots[0],
 840                                             struct btrfs_file_extent_item);
 841
 842                         extent_offset += start - key.offset;
 843                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 844                         btrfs_set_file_extent_num_bytes(leaf, fi,
 845                                                         extent_end - start);
 846                         btrfs_mark_buffer_dirty(leaf);
 847
 848                         if (update_refs && disk_bytenr > 0) {
 849                                 ret = btrfs_inc_extent_ref(trans, root,
 850                                                 disk_bytenr, num_bytes, 0,
 851                                                 root->root_key.objectid,
 852                                                 new_key.objectid,
 853                                                 start - extent_offset, 1);
 854                                 BUG_ON(ret); /* -ENOMEM */
 855                         }
 856                         key.offset = start;
 857                 }
 858                 /*
 859                  *  | ---- range to drop ----- |
 860                  *      | -------- extent -------- |
 861                  */
 862                 if (start <= key.offset && end < extent_end) {
 863                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 864                                 ret = -EOPNOTSUPP;
 865                                 break;
 866                         }
 867
 868                         memcpy(&new_key, &key, sizeof(new_key));
 869                         new_key.offset = end;
 870                         btrfs_set_item_key_safe(root, path, &new_key);
 871
 872                         extent_offset += end - key.offset;
 873                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 874                         btrfs_set_file_extent_num_bytes(leaf, fi,
 875                                                         extent_end - end);
 876                         btrfs_mark_buffer_dirty(leaf);
 877                         if (update_refs && disk_bytenr > 0)
 878                                 inode_sub_bytes(inode, end - key.offset);
 879                         break;
 880                 }
 881
 882                 search_start = extent_end;
 883                 /*
 884                  *       | ---- range to drop ----- |
 885                  *  | -------- extent -------- |
 886                  */
 887                 if (start > key.offset && end >= extent_end) {
 888                         BUG_ON(del_nr > 0);
 889                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 890                                 ret = -EOPNOTSUPP;
 891                                 break;
 892                         }
 893
 894                         btrfs_set_file_extent_num_bytes(leaf, fi,
 895                                                         start - key.offset);
 896                         btrfs_mark_buffer_dirty(leaf);
 897                         if (update_refs && disk_bytenr > 0)
 898                                 inode_sub_bytes(inode, extent_end - start);
 899                         if (end == extent_end)
 900                                 break;
 901
 902                         path->slots[0]++;
 903                         goto next_slot;
 904                 }
 905
 906                 /*
 907                  *  | ---- range to drop ----- |
 908                  *    | ------ extent ------ |
 909                  */
 910                 if (start <= key.offset && end >= extent_end) {
 911 delete_extent_item:
 912                         if (del_nr == 0) {
 913                                 del_slot = path->slots[0];
 914                                 del_nr = 1;
 915                         } else {
 916                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 917                                 del_nr++;
 918                         }
 919
 920                         if (update_refs &&
 921                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 922                                 inode_sub_bytes(inode,
 923                                                 extent_end - key.offset);
 924                                 extent_end = ALIGN(extent_end,
 925                                                    root->sectorsize);
 926                         } else if (update_refs && disk_bytenr > 0) {
 927                                 ret = btrfs_free_extent(trans, root,
 928                                                 disk_bytenr, num_bytes, 0,
 929                                                 root->root_key.objectid,
 930                                                 key.objectid, key.offset -
 931                                                 extent_offset, 0);
 932                                 BUG_ON(ret); /* -ENOMEM */
 933                                 inode_sub_bytes(inode,
 934                                                 extent_end - key.offset);
 935                         }
 936
 937                         if (end == extent_end)
 938                                 break;
 939
 940                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 941                                 path->slots[0]++;
 942                                 goto next_slot;
 943                         }
 944
 945                         ret = btrfs_del_items(trans, root, path, del_slot,
 946                                               del_nr);
 947                         if (ret) {
 948                                 btrfs_abort_transaction(trans, root, ret);
 949                                 break;
 950                         }
 951
 952                         del_nr = 0;
 953                         del_slot = 0;
 954
 955                         btrfs_release_path(path);
 956                         continue;
 957                 }
 958
 959                 BUG_ON(1);
 960         }
 961
 962         if (!ret && del_nr > 0) {
 963                 /*
 964                  * Set path->slots[0] to first slot, so that after the delete
 965                  * if items are move off from our leaf to its immediate left or
 966                  * right neighbor leafs, we end up with a correct and adjusted
 967                  * path->slots[0] for our insertion (if replace_extent != 0).
 968                  */
 969                 path->slots[0] = del_slot;
 970                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 971                 if (ret)
 972                         btrfs_abort_transaction(trans, root, ret);
 973         }
 974
 975         leaf = path->nodes[0];
 976         /*
 977          * If btrfs_del_items() was called, it might have deleted a leaf, in
 978          * which case it unlocked our path, so check path->locks[0] matches a
 979          * write lock.
 980          */
 981         if (!ret && replace_extent && leafs_visited == 1 &&
 982             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 983              path->locks[0] == BTRFS_WRITE_LOCK) &&
 984             btrfs_leaf_free_space(root, leaf) >=
 985             sizeof(struct btrfs_item) + extent_item_size) {
 986
 987                 key.objectid = ino;
 988                 key.type = BTRFS_EXTENT_DATA_KEY;
 989                 key.offset = start;
 990                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
 991                         struct btrfs_key slot_key;
 992
 993                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
 994                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
 995                                 path->slots[0]++;
 996                 }
 997                 setup_items_for_insert(root, path, &key,
 998                                        &extent_item_size,
 999                                        extent_item_size,
1000                                        sizeof(struct btrfs_item) +
1001                                        extent_item_size, 1);
1002                 *key_inserted = 1;
1003         }
1004
1005         if (!replace_extent || !(*key_inserted))
1006                 btrfs_release_path(path);
1007         if (drop_end)
1008                 *drop_end = found ? min(end, extent_end) : end;
1009         return ret;
1010 }
1011
1012 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1013                        struct btrfs_root *root, struct inode *inode, u64 start,
1014                        u64 end, int drop_cache)
1015 {
1016         struct btrfs_path *path;
1017         int ret;
1018
1019         path = btrfs_alloc_path();
1020         if (!path)
1021                 return -ENOMEM;
1022         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1023                                    drop_cache, 0, 0, NULL);
1024         btrfs_free_path(path);
1025         return ret;
1026 }
1027
1028 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1029                             u64 objectid, u64 bytenr, u64 orig_offset,
1030                             u64 *start, u64 *end)
1031 {
1032         struct btrfs_file_extent_item *fi;
1033         struct btrfs_key key;
1034         u64 extent_end;
1035
1036         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1037                 return 0;
1038
1039         btrfs_item_key_to_cpu(leaf, &key, slot);
1040         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1041                 return 0;
1042
1043         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1044         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1045             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1046             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1047             btrfs_file_extent_compression(leaf, fi) ||
1048             btrfs_file_extent_encryption(leaf, fi) ||
1049             btrfs_file_extent_other_encoding(leaf, fi))
1050                 return 0;
1051
1052         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1053         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1054                 return 0;
1055
1056         *start = key.offset;
1057         *end = extent_end;
1058         return 1;
1059 }
1060
1061 /*
1062  * Mark extent in the range start - end as written.
1063  *
1064  * This changes extent type from 'pre-allocated' to 'regular'. If only
1065  * part of extent is marked as written, the extent will be split into
1066  * two or three.
1067  */
1068 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1069                               struct inode *inode, u64 start, u64 end)
1070 {
1071         struct btrfs_root *root = BTRFS_I(inode)->root;
1072         struct extent_buffer *leaf;
1073         struct btrfs_path *path;
1074         struct btrfs_file_extent_item *fi;
1075         struct btrfs_key key;
1076         struct btrfs_key new_key;
1077         u64 bytenr;
1078         u64 num_bytes;
1079         u64 extent_end;
1080         u64 orig_offset;
1081         u64 other_start;
1082         u64 other_end;
1083         u64 split;
1084         int del_nr = 0;
1085         int del_slot = 0;
1086         int recow;
1087         int ret;
1088         u64 ino = btrfs_ino(inode);
1089
1090         path = btrfs_alloc_path();
1091         if (!path)
1092                 return -ENOMEM;
1093 again:
1094         recow = 0;
1095         split = start;
1096         key.objectid = ino;
1097         key.type = BTRFS_EXTENT_DATA_KEY;
1098         key.offset = split;
1099
1100         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1101         if (ret < 0)
1102                 goto out;
1103         if (ret > 0 && path->slots[0] > 0)
1104                 path->slots[0]--;
1105
1106         leaf = path->nodes[0];
1107         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1108         BUG_ON(key.objectid != ino || key.type != BTRFS_EXTENT_DATA_KEY);
1109         fi = btrfs_item_ptr(leaf, path->slots[0],
1110                             struct btrfs_file_extent_item);
1111         BUG_ON(btrfs_file_extent_type(leaf, fi) !=
1112                BTRFS_FILE_EXTENT_PREALLOC);
1113         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1114         BUG_ON(key.offset > start || extent_end < end);
1115
1116         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1117         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1118         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1119         memcpy(&new_key, &key, sizeof(new_key));
1120
1121         if (start == key.offset && end < extent_end) {
1122                 other_start = 0;
1123                 other_end = start;
1124                 if (extent_mergeable(leaf, path->slots[0] - 1,
1125                                      ino, bytenr, orig_offset,
1126                                      &other_start, &other_end)) {
1127                         new_key.offset = end;
1128                         btrfs_set_item_key_safe(root, path, &new_key);
1129                         fi = btrfs_item_ptr(leaf, path->slots[0],
1130                                             struct btrfs_file_extent_item);
1131                         btrfs_set_file_extent_generation(leaf, fi,
1132                                                          trans->transid);
1133                         btrfs_set_file_extent_num_bytes(leaf, fi,
1134                                                         extent_end - end);
1135                         btrfs_set_file_extent_offset(leaf, fi,
1136                                                      end - orig_offset);
1137                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1138                                             struct btrfs_file_extent_item);
1139                         btrfs_set_file_extent_generation(leaf, fi,
1140                                                          trans->transid);
1141                         btrfs_set_file_extent_num_bytes(leaf, fi,
1142                                                         end - other_start);
1143                         btrfs_mark_buffer_dirty(leaf);
1144                         goto out;
1145                 }
1146         }
1147
1148         if (start > key.offset && end == extent_end) {
1149                 other_start = end;
1150                 other_end = 0;
1151                 if (extent_mergeable(leaf, path->slots[0] + 1,
1152                                      ino, bytenr, orig_offset,
1153                                      &other_start, &other_end)) {
1154                         fi = btrfs_item_ptr(leaf, path->slots[0],
1155                                             struct btrfs_file_extent_item);
1156                         btrfs_set_file_extent_num_bytes(leaf, fi,
1157                                                         start - key.offset);
1158                         btrfs_set_file_extent_generation(leaf, fi,
1159                                                          trans->transid);
1160                         path->slots[0]++;
1161                         new_key.offset = start;
1162                         btrfs_set_item_key_safe(root, path, &new_key);
1163
1164                         fi = btrfs_item_ptr(leaf, path->slots[0],
1165                                             struct btrfs_file_extent_item);
1166                         btrfs_set_file_extent_generation(leaf, fi,
1167                                                          trans->transid);
1168                         btrfs_set_file_extent_num_bytes(leaf, fi,
1169                                                         other_end - start);
1170                         btrfs_set_file_extent_offset(leaf, fi,
1171                                                      start - orig_offset);
1172                         btrfs_mark_buffer_dirty(leaf);
1173                         goto out;
1174                 }
1175         }
1176
1177         while (start > key.offset || end < extent_end) {
1178                 if (key.offset == start)
1179                         split = end;
1180
1181                 new_key.offset = split;
1182                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1183                 if (ret == -EAGAIN) {
1184                         btrfs_release_path(path);
1185                         goto again;
1186                 }
1187                 if (ret < 0) {
1188                         btrfs_abort_transaction(trans, root, ret);
1189                         goto out;
1190                 }
1191
1192                 leaf = path->nodes[0];
1193                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1194                                     struct btrfs_file_extent_item);
1195                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1196                 btrfs_set_file_extent_num_bytes(leaf, fi,
1197                                                 split - key.offset);
1198
1199                 fi = btrfs_item_ptr(leaf, path->slots[0],
1200                                     struct btrfs_file_extent_item);
1201
1202                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1203                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1204                 btrfs_set_file_extent_num_bytes(leaf, fi,
1205                                                 extent_end - split);
1206                 btrfs_mark_buffer_dirty(leaf);
1207
1208                 ret = btrfs_inc_extent_ref(trans, root, bytenr, num_bytes, 0,
1209                                            root->root_key.objectid,
1210                                            ino, orig_offset, 1);
1211                 BUG_ON(ret); /* -ENOMEM */
1212
1213                 if (split == start) {
1214                         key.offset = start;
1215                 } else {
1216                         BUG_ON(start != key.offset);
1217                         path->slots[0]--;
1218                         extent_end = end;
1219                 }
1220                 recow = 1;
1221         }
1222
1223         other_start = end;
1224         other_end = 0;
1225         if (extent_mergeable(leaf, path->slots[0] + 1,
1226                              ino, bytenr, orig_offset,
1227                              &other_start, &other_end)) {
1228                 if (recow) {
1229                         btrfs_release_path(path);
1230                         goto again;
1231                 }
1232                 extent_end = other_end;
1233                 del_slot = path->slots[0] + 1;
1234                 del_nr++;
1235                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1236                                         0, root->root_key.objectid,
1237                                         ino, orig_offset, 0);
1238                 BUG_ON(ret); /* -ENOMEM */
1239         }
1240         other_start = 0;
1241         other_end = start;
1242         if (extent_mergeable(leaf, path->slots[0] - 1,
1243                              ino, bytenr, orig_offset,
1244                              &other_start, &other_end)) {
1245                 if (recow) {
1246                         btrfs_release_path(path);
1247                         goto again;
1248                 }
1249                 key.offset = other_start;
1250                 del_slot = path->slots[0];
1251                 del_nr++;
1252                 ret = btrfs_free_extent(trans, root, bytenr, num_bytes,
1253                                         0, root->root_key.objectid,
1254                                         ino, orig_offset, 0);
1255                 BUG_ON(ret); /* -ENOMEM */
1256         }
1257         if (del_nr == 0) {
1258                 fi = btrfs_item_ptr(leaf, path->slots[0],
1259                            struct btrfs_file_extent_item);
1260                 btrfs_set_file_extent_type(leaf, fi,
1261                                            BTRFS_FILE_EXTENT_REG);
1262                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1263                 btrfs_mark_buffer_dirty(leaf);
1264         } else {
1265                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1266                            struct btrfs_file_extent_item);
1267                 btrfs_set_file_extent_type(leaf, fi,
1268                                            BTRFS_FILE_EXTENT_REG);
1269                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1270                 btrfs_set_file_extent_num_bytes(leaf, fi,
1271                                                 extent_end - key.offset);
1272                 btrfs_mark_buffer_dirty(leaf);
1273
1274                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1275                 if (ret < 0) {
1276                         btrfs_abort_transaction(trans, root, ret);
1277                         goto out;
1278                 }
1279         }
1280 out:
1281         btrfs_free_path(path);
1282         return 0;
1283 }
1284
1285 /*
1286  * on error we return an unlocked page and the error value
1287  * on success we return a locked page and 0
1288  */
1289 static int prepare_uptodate_page(struct page *page, u64 pos,
1290                                  bool force_uptodate)
1291 {
1292         int ret = 0;
1293
1294         if (((pos & (PAGE_CACHE_SIZE - 1)) || force_uptodate) &&
1295             !PageUptodate(page)) {
1296                 ret = btrfs_readpage(NULL, page);
1297                 if (ret)
1298                         return ret;
1299                 lock_page(page);
1300                 if (!PageUptodate(page)) {
1301                         unlock_page(page);
1302                         return -EIO;
1303                 }
1304         }
1305         return 0;
1306 }
1307
1308 /*
1309  * this just gets pages into the page cache and locks them down.
1310  */
1311 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1312                                   size_t num_pages, loff_t pos,
1313                                   size_t write_bytes, bool force_uptodate)
1314 {
1315         int i;
1316         unsigned long index = pos >> PAGE_CACHE_SHIFT;
1317         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1318         int err = 0;
1319         int faili;
1320
1321         for (i = 0; i < num_pages; i++) {
1322                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1323                                                mask | __GFP_WRITE);
1324                 if (!pages[i]) {
1325                         faili = i - 1;
1326                         err = -ENOMEM;
1327                         goto fail;
1328                 }
1329
1330                 if (i == 0)
1331                         err = prepare_uptodate_page(pages[i], pos,
1332                                                     force_uptodate);
1333                 if (i == num_pages - 1)
1334                         err = prepare_uptodate_page(pages[i],
1335                                                     pos + write_bytes, false);
1336                 if (err) {
1337                         page_cache_release(pages[i]);
1338                         faili = i - 1;
1339                         goto fail;
1340                 }
1341                 wait_on_page_writeback(pages[i]);
1342         }
1343
1344         return 0;
1345 fail:
1346         while (faili >= 0) {
1347                 unlock_page(pages[faili]);
1348                 page_cache_release(pages[faili]);
1349                 faili--;
1350         }
1351         return err;
1352
1353 }
1354
1355 /*
1356  * This function locks the extent and properly waits for data=ordered extents
1357  * to finish before allowing the pages to be modified if need.
1358  *
1359  * The return value:
1360  * 1 - the extent is locked
1361  * 0 - the extent is not locked, and everything is OK
1362  * -EAGAIN - need re-prepare the pages
1363  * the other < 0 number - Something wrong happens
1364  */
1365 static noinline int
1366 lock_and_cleanup_extent_if_need(struct inode *inode, struct page **pages,
1367                                 size_t num_pages, loff_t pos,
1368                                 u64 *lockstart, u64 *lockend,
1369                                 struct extent_state **cached_state)
1370 {
1371         u64 start_pos;
1372         u64 last_pos;
1373         int i;
1374         int ret = 0;
1375
1376         start_pos = pos & ~((u64)PAGE_CACHE_SIZE - 1);
1377         last_pos = start_pos + ((u64)num_pages << PAGE_CACHE_SHIFT) - 1;
1378
1379         if (start_pos < inode->i_size) {
1380                 struct btrfs_ordered_extent *ordered;
1381                 lock_extent_bits(&BTRFS_I(inode)->io_tree,
1382                                  start_pos, last_pos, 0, cached_state);
1383                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1384                                                      last_pos - start_pos + 1);
1385                 if (ordered &&
1386                     ordered->file_offset + ordered->len > start_pos &&
1387                     ordered->file_offset <= last_pos) {
1388                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1389                                              start_pos, last_pos,
1390                                              cached_state, GFP_NOFS);
1391                         for (i = 0; i < num_pages; i++) {
1392                                 unlock_page(pages[i]);
1393                                 page_cache_release(pages[i]);
1394                         }
1395                         btrfs_start_ordered_extent(inode, ordered, 1);
1396                         btrfs_put_ordered_extent(ordered);
1397                         return -EAGAIN;
1398                 }
1399                 if (ordered)
1400                         btrfs_put_ordered_extent(ordered);
1401
1402                 clear_extent_bit(&BTRFS_I(inode)->io_tree, start_pos,
1403                                   last_pos, EXTENT_DIRTY | EXTENT_DELALLOC |
1404                                   EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG,
1405                                   0, 0, cached_state, GFP_NOFS);
1406                 *lockstart = start_pos;
1407                 *lockend = last_pos;
1408                 ret = 1;
1409         }
1410
1411         for (i = 0; i < num_pages; i++) {
1412                 if (clear_page_dirty_for_io(pages[i]))
1413                         account_page_redirty(pages[i]);
1414                 set_page_extent_mapped(pages[i]);
1415                 WARN_ON(!PageLocked(pages[i]));
1416         }
1417
1418         return ret;
1419 }
1420
1421 static noinline int check_can_nocow(struct inode *inode, loff_t pos,
1422                                     size_t *write_bytes)
1423 {
1424         struct btrfs_root *root = BTRFS_I(inode)->root;
1425         struct btrfs_ordered_extent *ordered;
1426         u64 lockstart, lockend;
1427         u64 num_bytes;
1428         int ret;
1429
1430         ret = btrfs_start_nocow_write(root);
1431         if (!ret)
1432                 return -ENOSPC;
1433
1434         lockstart = round_down(pos, root->sectorsize);
1435         lockend = round_up(pos + *write_bytes, root->sectorsize) - 1;
1436
1437         while (1) {
1438                 lock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1439                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1440                                                      lockend - lockstart + 1);
1441                 if (!ordered) {
1442                         break;
1443                 }
1444                 unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1445                 btrfs_start_ordered_extent(inode, ordered, 1);
1446                 btrfs_put_ordered_extent(ordered);
1447         }
1448
1449         num_bytes = lockend - lockstart + 1;
1450         ret = can_nocow_extent(inode, lockstart, &num_bytes, NULL, NULL, NULL);
1451         if (ret <= 0) {
1452                 ret = 0;
1453                 btrfs_end_nocow_write(root);
1454         } else {
1455                 *write_bytes = min_t(size_t, *write_bytes ,
1456                                      num_bytes - pos + lockstart);
1457         }
1458
1459         unlock_extent(&BTRFS_I(inode)->io_tree, lockstart, lockend);
1460
1461         return ret;
1462 }
1463
1464 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1465                                                struct iov_iter *i,
1466                                                loff_t pos)
1467 {
1468         struct inode *inode = file_inode(file);
1469         struct btrfs_root *root = BTRFS_I(inode)->root;
1470         struct page **pages = NULL;
1471         struct extent_state *cached_state = NULL;
1472         u64 release_bytes = 0;
1473         u64 lockstart;
1474         u64 lockend;
1475         unsigned long first_index;
1476         size_t num_written = 0;
1477         int nrptrs;
1478         int ret = 0;
1479         bool only_release_metadata = false;
1480         bool force_page_uptodate = false;
1481         bool need_unlock;
1482
1483         nrptrs = min((iov_iter_count(i) + PAGE_CACHE_SIZE - 1) /
1484                      PAGE_CACHE_SIZE, PAGE_CACHE_SIZE /
1485                      (sizeof(struct page *)));
1486         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1487         nrptrs = max(nrptrs, 8);
1488         pages = kmalloc(nrptrs * sizeof(struct page *), GFP_KERNEL);
1489         if (!pages)
1490                 return -ENOMEM;
1491
1492         first_index = pos >> PAGE_CACHE_SHIFT;
1493
1494         while (iov_iter_count(i) > 0) {
1495                 size_t offset = pos & (PAGE_CACHE_SIZE - 1);
1496                 size_t write_bytes = min(iov_iter_count(i),
1497                                          nrptrs * (size_t)PAGE_CACHE_SIZE -
1498                                          offset);
1499                 size_t num_pages = (write_bytes + offset +
1500                                     PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
1501                 size_t reserve_bytes;
1502                 size_t dirty_pages;
1503                 size_t copied;
1504
1505                 WARN_ON(num_pages > nrptrs);
1506
1507                 /*
1508                  * Fault pages before locking them in prepare_pages
1509                  * to avoid recursive lock
1510                  */
1511                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1512                         ret = -EFAULT;
1513                         break;
1514                 }
1515
1516                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1517                 ret = btrfs_check_data_free_space(inode, reserve_bytes);
1518                 if (ret == -ENOSPC &&
1519                     (BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1520                                               BTRFS_INODE_PREALLOC))) {
1521                         ret = check_can_nocow(inode, pos, &write_bytes);
1522                         if (ret > 0) {
1523                                 only_release_metadata = true;
1524                                 /*
1525                                  * our prealloc extent may be smaller than
1526                                  * write_bytes, so scale down.
1527                                  */
1528                                 num_pages = (write_bytes + offset +
1529                                              PAGE_CACHE_SIZE - 1) >>
1530                                         PAGE_CACHE_SHIFT;
1531                                 reserve_bytes = num_pages << PAGE_CACHE_SHIFT;
1532                                 ret = 0;
1533                         } else {
1534                                 ret = -ENOSPC;
1535                         }
1536                 }
1537
1538                 if (ret)
1539                         break;
1540
1541                 ret = btrfs_delalloc_reserve_metadata(inode, reserve_bytes);
1542                 if (ret) {
1543                         if (!only_release_metadata)
1544                                 btrfs_free_reserved_data_space(inode,
1545                                                                reserve_bytes);
1546                         else
1547                                 btrfs_end_nocow_write(root);
1548                         break;
1549                 }
1550
1551                 release_bytes = reserve_bytes;
1552                 need_unlock = false;
1553 again:
1554                 /*
1555                  * This is going to setup the pages array with the number of
1556                  * pages we want, so we don't really need to worry about the
1557                  * contents of pages from loop to loop
1558                  */
1559                 ret = prepare_pages(inode, pages, num_pages,
1560                                     pos, write_bytes,
1561                                     force_page_uptodate);
1562                 if (ret)
1563                         break;
1564
1565                 ret = lock_and_cleanup_extent_if_need(inode, pages, num_pages,
1566                                                       pos, &lockstart, &lockend,
1567                                                       &cached_state);
1568                 if (ret < 0) {
1569                         if (ret == -EAGAIN)
1570                                 goto again;
1571                         break;
1572                 } else if (ret > 0) {
1573                         need_unlock = true;
1574                         ret = 0;
1575                 }
1576
1577                 copied = btrfs_copy_from_user(pos, num_pages,
1578                                            write_bytes, pages, i);
1579
1580                 /*
1581                  * if we have trouble faulting in the pages, fall
1582                  * back to one page at a time
1583                  */
1584                 if (copied < write_bytes)
1585                         nrptrs = 1;
1586
1587                 if (copied == 0) {
1588                         force_page_uptodate = true;
1589                         dirty_pages = 0;
1590                 } else {
1591                         force_page_uptodate = false;
1592                         dirty_pages = (copied + offset +
1593                                        PAGE_CACHE_SIZE - 1) >>
1594                                        PAGE_CACHE_SHIFT;
1595                 }
1596
1597                 /*
1598                  * If we had a short copy we need to release the excess delaloc
1599                  * bytes we reserved.  We need to increment outstanding_extents
1600                  * because btrfs_delalloc_release_space will decrement it, but
1601                  * we still have an outstanding extent for the chunk we actually
1602                  * managed to copy.
1603                  */
1604                 if (num_pages > dirty_pages) {
1605                         release_bytes = (num_pages - dirty_pages) <<
1606                                 PAGE_CACHE_SHIFT;
1607                         if (copied > 0) {
1608                                 spin_lock(&BTRFS_I(inode)->lock);
1609                                 BTRFS_I(inode)->outstanding_extents++;
1610                                 spin_unlock(&BTRFS_I(inode)->lock);
1611                         }
1612                         if (only_release_metadata)
1613                                 btrfs_delalloc_release_metadata(inode,
1614                                                                 release_bytes);
1615                         else
1616                                 btrfs_delalloc_release_space(inode,
1617                                                              release_bytes);
1618                 }
1619
1620                 release_bytes = dirty_pages << PAGE_CACHE_SHIFT;
1621
1622                 if (copied > 0)
1623                         ret = btrfs_dirty_pages(root, inode, pages,
1624                                                 dirty_pages, pos, copied,
1625                                                 NULL);
1626                 if (need_unlock)
1627                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1628                                              lockstart, lockend, &cached_state,
1629                                              GFP_NOFS);
1630                 if (ret) {
1631                         btrfs_drop_pages(pages, num_pages);
1632                         break;
1633                 }
1634
1635                 release_bytes = 0;
1636                 if (only_release_metadata)
1637                         btrfs_end_nocow_write(root);
1638
1639                 if (only_release_metadata && copied > 0) {
1640                         u64 lockstart = round_down(pos, root->sectorsize);
1641                         u64 lockend = lockstart +
1642                                 (dirty_pages << PAGE_CACHE_SHIFT) - 1;
1643
1644                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1645                                        lockend, EXTENT_NORESERVE, NULL,
1646                                        NULL, GFP_NOFS);
1647                         only_release_metadata = false;
1648                 }
1649
1650                 btrfs_drop_pages(pages, num_pages);
1651
1652                 cond_resched();
1653
1654                 balance_dirty_pages_ratelimited(inode->i_mapping);
1655                 if (dirty_pages < (root->leafsize >> PAGE_CACHE_SHIFT) + 1)
1656                         btrfs_btree_balance_dirty(root);
1657
1658                 pos += copied;
1659                 num_written += copied;
1660         }
1661
1662         kfree(pages);
1663
1664         if (release_bytes) {
1665                 if (only_release_metadata) {
1666                         btrfs_end_nocow_write(root);
1667                         btrfs_delalloc_release_metadata(inode, release_bytes);
1668                 } else {
1669                         btrfs_delalloc_release_space(inode, release_bytes);
1670                 }
1671         }
1672
1673         return num_written ? num_written : ret;
1674 }
1675
1676 static ssize_t __btrfs_direct_write(struct kiocb *iocb,
1677                                     const struct iovec *iov,
1678                                     unsigned long nr_segs, loff_t pos,
1679                                     size_t count, size_t ocount)
1680 {
1681         struct file *file = iocb->ki_filp;
1682         struct iov_iter i;
1683         ssize_t written;
1684         ssize_t written_buffered;
1685         loff_t endbyte;
1686         int err;
1687
1688         written = generic_file_direct_write(iocb, iov, &nr_segs, pos,
1689                                             count, ocount);
1690
1691         if (written < 0 || written == count)
1692                 return written;
1693
1694         pos += written;
1695         count -= written;
1696         iov_iter_init(&i, iov, nr_segs, count, written);
1697         written_buffered = __btrfs_buffered_write(file, &i, pos);
1698         if (written_buffered < 0) {
1699                 err = written_buffered;
1700                 goto out;
1701         }
1702         endbyte = pos + written_buffered - 1;
1703         err = filemap_write_and_wait_range(file->f_mapping, pos, endbyte);
1704         if (err)
1705                 goto out;
1706         written += written_buffered;
1707         iocb->ki_pos = pos + written_buffered;
1708         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_CACHE_SHIFT,
1709                                  endbyte >> PAGE_CACHE_SHIFT);
1710 out:
1711         return written ? written : err;
1712 }
1713
1714 static void update_time_for_write(struct inode *inode)
1715 {
1716         struct timespec now;
1717
1718         if (IS_NOCMTIME(inode))
1719                 return;
1720
1721         now = current_fs_time(inode->i_sb);
1722         if (!timespec_equal(&inode->i_mtime, &now))
1723                 inode->i_mtime = now;
1724
1725         if (!timespec_equal(&inode->i_ctime, &now))
1726                 inode->i_ctime = now;
1727
1728         if (IS_I_VERSION(inode))
1729                 inode_inc_iversion(inode);
1730 }
1731
1732 static ssize_t btrfs_file_aio_write(struct kiocb *iocb,
1733                                     const struct iovec *iov,
1734                                     unsigned long nr_segs, loff_t pos)
1735 {
1736         struct file *file = iocb->ki_filp;
1737         struct inode *inode = file_inode(file);
1738         struct btrfs_root *root = BTRFS_I(inode)->root;
1739         u64 start_pos;
1740         u64 end_pos;
1741         ssize_t num_written = 0;
1742         ssize_t err = 0;
1743         size_t count, ocount;
1744         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1745
1746         mutex_lock(&inode->i_mutex);
1747
1748         err = generic_segment_checks(iov, &nr_segs, &ocount, VERIFY_READ);
1749         if (err) {
1750                 mutex_unlock(&inode->i_mutex);
1751                 goto out;
1752         }
1753         count = ocount;
1754
1755         current->backing_dev_info = inode->i_mapping->backing_dev_info;
1756         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
1757         if (err) {
1758                 mutex_unlock(&inode->i_mutex);
1759                 goto out;
1760         }
1761
1762         if (count == 0) {
1763                 mutex_unlock(&inode->i_mutex);
1764                 goto out;
1765         }
1766
1767         err = file_remove_suid(file);
1768         if (err) {
1769                 mutex_unlock(&inode->i_mutex);
1770                 goto out;
1771         }
1772
1773         /*
1774          * If BTRFS flips readonly due to some impossible error
1775          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1776          * although we have opened a file as writable, we have
1777          * to stop this write operation to ensure FS consistency.
1778          */
1779         if (test_bit(BTRFS_FS_STATE_ERROR, &root->fs_info->fs_state)) {
1780                 mutex_unlock(&inode->i_mutex);
1781                 err = -EROFS;
1782                 goto out;
1783         }
1784
1785         /*
1786          * We reserve space for updating the inode when we reserve space for the
1787          * extent we are going to write, so we will enospc out there.  We don't
1788          * need to start yet another transaction to update the inode as we will
1789          * update the inode when we finish writing whatever data we write.
1790          */
1791         update_time_for_write(inode);
1792
1793         start_pos = round_down(pos, root->sectorsize);
1794         if (start_pos > i_size_read(inode)) {
1795                 /* Expand hole size to cover write data, preventing empty gap */
1796                 end_pos = round_up(pos + count, root->sectorsize);
1797                 err = btrfs_cont_expand(inode, i_size_read(inode), end_pos);
1798                 if (err) {
1799                         mutex_unlock(&inode->i_mutex);
1800                         goto out;
1801                 }
1802         }
1803
1804         if (sync)
1805                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1806
1807         if (unlikely(file->f_flags & O_DIRECT)) {
1808                 num_written = __btrfs_direct_write(iocb, iov, nr_segs,
1809                                                    pos, count, ocount);
1810         } else {
1811                 struct iov_iter i;
1812
1813                 iov_iter_init(&i, iov, nr_segs, count, num_written);
1814
1815                 num_written = __btrfs_buffered_write(file, &i, pos);
1816                 if (num_written > 0)
1817                         iocb->ki_pos = pos + num_written;
1818         }
1819
1820         mutex_unlock(&inode->i_mutex);
1821
1822         /*
1823          * we want to make sure fsync finds this change
1824          * but we haven't joined a transaction running right now.
1825          *
1826          * Later on, someone is sure to update the inode and get the
1827          * real transid recorded.
1828          *
1829          * We set last_trans now to the fs_info generation + 1,
1830          * this will either be one more than the running transaction
1831          * or the generation used for the next transaction if there isn't
1832          * one running right now.
1833          *
1834          * We also have to set last_sub_trans to the current log transid,
1835          * otherwise subsequent syncs to a file that's been synced in this
1836          * transaction will appear to have already occured.
1837          */
1838         BTRFS_I(inode)->last_trans = root->fs_info->generation + 1;
1839         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1840         if (num_written > 0) {
1841                 err = generic_write_sync(file, pos, num_written);
1842                 if (err < 0)
1843                         num_written = err;
1844         }
1845
1846         if (sync)
1847                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1848 out:
1849         current->backing_dev_info = NULL;
1850         return num_written ? num_written : err;
1851 }
1852
1853 int btrfs_release_file(struct inode *inode, struct file *filp)
1854 {
1855         /*
1856          * ordered_data_close is set by settattr when we are about to truncate
1857          * a file from a non-zero size to a zero size.  This tries to
1858          * flush down new bytes that may have been written if the
1859          * application were using truncate to replace a file in place.
1860          */
1861         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1862                                &BTRFS_I(inode)->runtime_flags)) {
1863                 struct btrfs_trans_handle *trans;
1864                 struct btrfs_root *root = BTRFS_I(inode)->root;
1865
1866                 /*
1867                  * We need to block on a committing transaction to keep us from
1868                  * throwing a ordered operation on to the list and causing
1869                  * something like sync to deadlock trying to flush out this
1870                  * inode.
1871                  */
1872                 trans = btrfs_start_transaction(root, 0);
1873                 if (IS_ERR(trans))
1874                         return PTR_ERR(trans);
1875                 btrfs_add_ordered_operation(trans, BTRFS_I(inode)->root, inode);
1876                 btrfs_end_transaction(trans, root);
1877                 if (inode->i_size > BTRFS_ORDERED_OPERATIONS_FLUSH_LIMIT)
1878                         filemap_flush(inode->i_mapping);
1879         }
1880         if (filp->private_data)
1881                 btrfs_ioctl_trans_end(filp);
1882         return 0;
1883 }
1884
1885 /*
1886  * fsync call for both files and directories.  This logs the inode into
1887  * the tree log instead of forcing full commits whenever possible.
1888  *
1889  * It needs to call filemap_fdatawait so that all ordered extent updates are
1890  * in the metadata btree are up to date for copying to the log.
1891  *
1892  * It drops the inode mutex before doing the tree log commit.  This is an
1893  * important optimization for directories because holding the mutex prevents
1894  * new operations on the dir while we write to disk.
1895  */
1896 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
1897 {
1898         struct dentry *dentry = file->f_path.dentry;
1899         struct inode *inode = dentry->d_inode;
1900         struct btrfs_root *root = BTRFS_I(inode)->root;
1901         struct btrfs_trans_handle *trans;
1902         struct btrfs_log_ctx ctx;
1903         int ret = 0;
1904         bool full_sync = 0;
1905
1906         trace_btrfs_sync_file(file, datasync);
1907
1908         /*
1909          * We write the dirty pages in the range and wait until they complete
1910          * out of the ->i_mutex. If so, we can flush the dirty pages by
1911          * multi-task, and make the performance up.  See
1912          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
1913          */
1914         atomic_inc(&BTRFS_I(inode)->sync_writers);
1915         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1916         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
1917                              &BTRFS_I(inode)->runtime_flags))
1918                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
1919         atomic_dec(&BTRFS_I(inode)->sync_writers);
1920         if (ret)
1921                 return ret;
1922
1923         mutex_lock(&inode->i_mutex);
1924
1925         /*
1926          * We flush the dirty pages again to avoid some dirty pages in the
1927          * range being left.
1928          */
1929         atomic_inc(&root->log_batch);
1930         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1931                              &BTRFS_I(inode)->runtime_flags);
1932         if (full_sync) {
1933                 ret = btrfs_wait_ordered_range(inode, start, end - start + 1);
1934                 if (ret) {
1935                         mutex_unlock(&inode->i_mutex);
1936                         goto out;
1937                 }
1938         }
1939         atomic_inc(&root->log_batch);
1940
1941         /*
1942          * check the transaction that last modified this inode
1943          * and see if its already been committed
1944          */
1945         if (!BTRFS_I(inode)->last_trans) {
1946                 mutex_unlock(&inode->i_mutex);
1947                 goto out;
1948         }
1949
1950         /*
1951          * if the last transaction that changed this file was before
1952          * the current transaction, we can bail out now without any
1953          * syncing
1954          */
1955         smp_mb();
1956         if (btrfs_inode_in_log(inode, root->fs_info->generation) ||
1957             BTRFS_I(inode)->last_trans <=
1958             root->fs_info->last_trans_committed) {
1959                 BTRFS_I(inode)->last_trans = 0;
1960
1961                 /*
1962                  * We'v had everything committed since the last time we were
1963                  * modified so clear this flag in case it was set for whatever
1964                  * reason, it's no longer relevant.
1965                  */
1966                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
1967                           &BTRFS_I(inode)->runtime_flags);
1968                 mutex_unlock(&inode->i_mutex);
1969                 goto out;
1970         }
1971
1972         /*
1973          * ok we haven't committed the transaction yet, lets do a commit
1974          */
1975         if (file->private_data)
1976                 btrfs_ioctl_trans_end(file);
1977
1978         /*
1979          * We use start here because we will need to wait on the IO to complete
1980          * in btrfs_sync_log, which could require joining a transaction (for
1981          * example checking cross references in the nocow path).  If we use join
1982          * here we could get into a situation where we're waiting on IO to
1983          * happen that is blocked on a transaction trying to commit.  With start
1984          * we inc the extwriter counter, so we wait for all extwriters to exit
1985          * before we start blocking join'ers.  This comment is to keep somebody
1986          * from thinking they are super smart and changing this to
1987          * btrfs_join_transaction *cough*Josef*cough*.
1988          */
1989         trans = btrfs_start_transaction(root, 0);
1990         if (IS_ERR(trans)) {
1991                 ret = PTR_ERR(trans);
1992                 mutex_unlock(&inode->i_mutex);
1993                 goto out;
1994         }
1995         trans->sync = true;
1996
1997         btrfs_init_log_ctx(&ctx);
1998
1999         ret = btrfs_log_dentry_safe(trans, root, dentry, &ctx);
2000         if (ret < 0) {
2001                 /* Fallthrough and commit/free transaction. */
2002                 ret = 1;
2003         }
2004
2005         /* we've logged all the items and now have a consistent
2006          * version of the file in the log.  It is possible that
2007          * someone will come in and modify the file, but that's
2008          * fine because the log is consistent on disk, and we
2009          * have references to all of the file's extents
2010          *
2011          * It is possible that someone will come in and log the
2012          * file again, but that will end up using the synchronization
2013          * inside btrfs_sync_log to keep things safe.
2014          */
2015         mutex_unlock(&inode->i_mutex);
2016
2017         if (ret != BTRFS_NO_LOG_SYNC) {
2018                 if (!ret) {
2019                         ret = btrfs_sync_log(trans, root, &ctx);
2020                         if (!ret) {
2021                                 ret = btrfs_end_transaction(trans, root);
2022                                 goto out;
2023                         }
2024                 }
2025                 if (!full_sync) {
2026                         ret = btrfs_wait_ordered_range(inode, start,
2027                                                        end - start + 1);
2028                         if (ret) {
2029                                 btrfs_end_transaction(trans, root);
2030                                 goto out;
2031                         }
2032                 }
2033                 ret = btrfs_commit_transaction(trans, root);
2034         } else {
2035                 ret = btrfs_end_transaction(trans, root);
2036         }
2037 out:
2038         return ret > 0 ? -EIO : ret;
2039 }
2040
2041 static const struct vm_operations_struct btrfs_file_vm_ops = {
2042         .fault          = filemap_fault,
2043         .map_pages      = filemap_map_pages,
2044         .page_mkwrite   = btrfs_page_mkwrite,
2045         .remap_pages    = generic_file_remap_pages,
2046 };
2047
2048 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2049 {
2050         struct address_space *mapping = filp->f_mapping;
2051
2052         if (!mapping->a_ops->readpage)
2053                 return -ENOEXEC;
2054
2055         file_accessed(filp);
2056         vma->vm_ops = &btrfs_file_vm_ops;
2057
2058         return 0;
2059 }
2060
2061 static int hole_mergeable(struct inode *inode, struct extent_buffer *leaf,
2062                           int slot, u64 start, u64 end)
2063 {
2064         struct btrfs_file_extent_item *fi;
2065         struct btrfs_key key;
2066
2067         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2068                 return 0;
2069
2070         btrfs_item_key_to_cpu(leaf, &key, slot);
2071         if (key.objectid != btrfs_ino(inode) ||
2072             key.type != BTRFS_EXTENT_DATA_KEY)
2073                 return 0;
2074
2075         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2076
2077         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2078                 return 0;
2079
2080         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2081                 return 0;
2082
2083         if (key.offset == end)
2084                 return 1;
2085         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2086                 return 1;
2087         return 0;
2088 }
2089
2090 static int fill_holes(struct btrfs_trans_handle *trans, struct inode *inode,
2091                       struct btrfs_path *path, u64 offset, u64 end)
2092 {
2093         struct btrfs_root *root = BTRFS_I(inode)->root;
2094         struct extent_buffer *leaf;
2095         struct btrfs_file_extent_item *fi;
2096         struct extent_map *hole_em;
2097         struct extent_map_tree *em_tree = &BTRFS_I(inode)->extent_tree;
2098         struct btrfs_key key;
2099         int ret;
2100
2101         if (btrfs_fs_incompat(root->fs_info, NO_HOLES))
2102                 goto out;
2103
2104         key.objectid = btrfs_ino(inode);
2105         key.type = BTRFS_EXTENT_DATA_KEY;
2106         key.offset = offset;
2107
2108         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2109         if (ret < 0)
2110                 return ret;
2111         BUG_ON(!ret);
2112
2113         leaf = path->nodes[0];
2114         if (hole_mergeable(inode, leaf, path->slots[0]-1, offset, end)) {
2115                 u64 num_bytes;
2116
2117                 path->slots[0]--;
2118                 fi = btrfs_item_ptr(leaf, path->slots[0],
2119                                     struct btrfs_file_extent_item);
2120                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2121                         end - offset;
2122                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2123                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2124                 btrfs_set_file_extent_offset(leaf, fi, 0);
2125                 btrfs_mark_buffer_dirty(leaf);
2126                 goto out;
2127         }
2128
2129         if (hole_mergeable(inode, leaf, path->slots[0]+1, offset, end)) {
2130                 u64 num_bytes;
2131
2132                 path->slots[0]++;
2133                 key.offset = offset;
2134                 btrfs_set_item_key_safe(root, path, &key);
2135                 fi = btrfs_item_ptr(leaf, path->slots[0],
2136                                     struct btrfs_file_extent_item);
2137                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2138                         offset;
2139                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2140                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2141                 btrfs_set_file_extent_offset(leaf, fi, 0);
2142                 btrfs_mark_buffer_dirty(leaf);
2143                 goto out;
2144         }
2145         btrfs_release_path(path);
2146
2147         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode), offset,
2148                                        0, 0, end - offset, 0, end - offset,
2149                                        0, 0, 0);
2150         if (ret)
2151                 return ret;
2152
2153 out:
2154         btrfs_release_path(path);
2155
2156         hole_em = alloc_extent_map();
2157         if (!hole_em) {
2158                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2159                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2160                         &BTRFS_I(inode)->runtime_flags);
2161         } else {
2162                 hole_em->start = offset;
2163                 hole_em->len = end - offset;
2164                 hole_em->ram_bytes = hole_em->len;
2165                 hole_em->orig_start = offset;
2166
2167                 hole_em->block_start = EXTENT_MAP_HOLE;
2168                 hole_em->block_len = 0;
2169                 hole_em->orig_block_len = 0;
2170                 hole_em->bdev = root->fs_info->fs_devices->latest_bdev;
2171                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2172                 hole_em->generation = trans->transid;
2173
2174                 do {
2175                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2176                         write_lock(&em_tree->lock);
2177                         ret = add_extent_mapping(em_tree, hole_em, 1);
2178                         write_unlock(&em_tree->lock);
2179                 } while (ret == -EEXIST);
2180                 free_extent_map(hole_em);
2181                 if (ret)
2182                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2183                                 &BTRFS_I(inode)->runtime_flags);
2184         }
2185
2186         return 0;
2187 }
2188
2189 /*
2190  * Find a hole extent on given inode and change start/len to the end of hole
2191  * extent.(hole/vacuum extent whose em->start <= start &&
2192  *         em->start + em->len > start)
2193  * When a hole extent is found, return 1 and modify start/len.
2194  */
2195 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2196 {
2197         struct extent_map *em;
2198         int ret = 0;
2199
2200         em = btrfs_get_extent(inode, NULL, 0, *start, *len, 0);
2201         if (IS_ERR_OR_NULL(em)) {
2202                 if (!em)
2203                         ret = -ENOMEM;
2204                 else
2205                         ret = PTR_ERR(em);
2206                 return ret;
2207         }
2208
2209         /* Hole or vacuum extent(only exists in no-hole mode) */
2210         if (em->block_start == EXTENT_MAP_HOLE) {
2211                 ret = 1;
2212                 *len = em->start + em->len > *start + *len ?
2213                        0 : *start + *len - em->start - em->len;
2214                 *start = em->start + em->len;
2215         }
2216         free_extent_map(em);
2217         return ret;
2218 }
2219
2220 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2221 {
2222         struct btrfs_root *root = BTRFS_I(inode)->root;
2223         struct extent_state *cached_state = NULL;
2224         struct btrfs_path *path;
2225         struct btrfs_block_rsv *rsv;
2226         struct btrfs_trans_handle *trans;
2227         u64 lockstart;
2228         u64 lockend;
2229         u64 tail_start;
2230         u64 tail_len;
2231         u64 orig_start = offset;
2232         u64 cur_offset;
2233         u64 min_size = btrfs_calc_trunc_metadata_size(root, 1);
2234         u64 drop_end;
2235         int ret = 0;
2236         int err = 0;
2237         int rsv_count;
2238         bool same_page;
2239         bool no_holes = btrfs_fs_incompat(root->fs_info, NO_HOLES);
2240         u64 ino_size;
2241
2242         ret = btrfs_wait_ordered_range(inode, offset, len);
2243         if (ret)
2244                 return ret;
2245
2246         mutex_lock(&inode->i_mutex);
2247         ino_size = round_up(inode->i_size, PAGE_CACHE_SIZE);
2248         ret = find_first_non_hole(inode, &offset, &len);
2249         if (ret < 0)
2250                 goto out_only_mutex;
2251         if (ret && !len) {
2252                 /* Already in a large hole */
2253                 ret = 0;
2254                 goto out_only_mutex;
2255         }
2256
2257         lockstart = round_up(offset , BTRFS_I(inode)->root->sectorsize);
2258         lockend = round_down(offset + len,
2259                              BTRFS_I(inode)->root->sectorsize) - 1;
2260         same_page = ((offset >> PAGE_CACHE_SHIFT) ==
2261                     ((offset + len - 1) >> PAGE_CACHE_SHIFT));
2262
2263         /*
2264          * We needn't truncate any page which is beyond the end of the file
2265          * because we are sure there is no data there.
2266          */
2267         /*
2268          * Only do this if we are in the same page and we aren't doing the
2269          * entire page.
2270          */
2271         if (same_page && len < PAGE_CACHE_SIZE) {
2272                 if (offset < ino_size)
2273                         ret = btrfs_truncate_page(inode, offset, len, 0);
2274                 goto out_only_mutex;
2275         }
2276
2277         /* zero back part of the first page */
2278         if (offset < ino_size) {
2279                 ret = btrfs_truncate_page(inode, offset, 0, 0);
2280                 if (ret) {
2281                         mutex_unlock(&inode->i_mutex);
2282                         return ret;
2283                 }
2284         }
2285
2286         /* Check the aligned pages after the first unaligned page,
2287          * if offset != orig_start, which means the first unaligned page
2288          * including serveral following pages are already in holes,
2289          * the extra check can be skipped */
2290         if (offset == orig_start) {
2291                 /* after truncate page, check hole again */
2292                 len = offset + len - lockstart;
2293                 offset = lockstart;
2294                 ret = find_first_non_hole(inode, &offset, &len);
2295                 if (ret < 0)
2296                         goto out_only_mutex;
2297                 if (ret && !len) {
2298                         ret = 0;
2299                         goto out_only_mutex;
2300                 }
2301                 lockstart = offset;
2302         }
2303
2304         /* Check the tail unaligned part is in a hole */
2305         tail_start = lockend + 1;
2306         tail_len = offset + len - tail_start;
2307         if (tail_len) {
2308                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2309                 if (unlikely(ret < 0))
2310                         goto out_only_mutex;
2311                 if (!ret) {
2312                         /* zero the front end of the last page */
2313                         if (tail_start + tail_len < ino_size) {
2314                                 ret = btrfs_truncate_page(inode,
2315                                                 tail_start + tail_len, 0, 1);
2316                                 if (ret)
2317                                         goto out_only_mutex;
2318                                 }
2319                 }
2320         }
2321
2322         if (lockend < lockstart) {
2323                 mutex_unlock(&inode->i_mutex);
2324                 return 0;
2325         }
2326
2327         while (1) {
2328                 struct btrfs_ordered_extent *ordered;
2329
2330                 truncate_pagecache_range(inode, lockstart, lockend);
2331
2332                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2333                                  0, &cached_state);
2334                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2335
2336                 /*
2337                  * We need to make sure we have no ordered extents in this range
2338                  * and nobody raced in and read a page in this range, if we did
2339                  * we need to try again.
2340                  */
2341                 if ((!ordered ||
2342                     (ordered->file_offset + ordered->len <= lockstart ||
2343                      ordered->file_offset > lockend)) &&
2344                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2345                         if (ordered)
2346                                 btrfs_put_ordered_extent(ordered);
2347                         break;
2348                 }
2349                 if (ordered)
2350                         btrfs_put_ordered_extent(ordered);
2351                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2352                                      lockend, &cached_state, GFP_NOFS);
2353                 ret = btrfs_wait_ordered_range(inode, lockstart,
2354                                                lockend - lockstart + 1);
2355                 if (ret) {
2356                         mutex_unlock(&inode->i_mutex);
2357                         return ret;
2358                 }
2359         }
2360
2361         path = btrfs_alloc_path();
2362         if (!path) {
2363                 ret = -ENOMEM;
2364                 goto out;
2365         }
2366
2367         rsv = btrfs_alloc_block_rsv(root, BTRFS_BLOCK_RSV_TEMP);
2368         if (!rsv) {
2369                 ret = -ENOMEM;
2370                 goto out_free;
2371         }
2372         rsv->size = btrfs_calc_trunc_metadata_size(root, 1);
2373         rsv->failfast = 1;
2374
2375         /*
2376          * 1 - update the inode
2377          * 1 - removing the extents in the range
2378          * 1 - adding the hole extent if no_holes isn't set
2379          */
2380         rsv_count = no_holes ? 2 : 3;
2381         trans = btrfs_start_transaction(root, rsv_count);
2382         if (IS_ERR(trans)) {
2383                 err = PTR_ERR(trans);
2384                 goto out_free;
2385         }
2386
2387         ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv, rsv,
2388                                       min_size);
2389         BUG_ON(ret);
2390         trans->block_rsv = rsv;
2391
2392         cur_offset = lockstart;
2393         len = lockend - cur_offset;
2394         while (cur_offset < lockend) {
2395                 ret = __btrfs_drop_extents(trans, root, inode, path,
2396                                            cur_offset, lockend + 1,
2397                                            &drop_end, 1, 0, 0, NULL);
2398                 if (ret != -ENOSPC)
2399                         break;
2400
2401                 trans->block_rsv = &root->fs_info->trans_block_rsv;
2402
2403                 if (cur_offset < ino_size) {
2404                         ret = fill_holes(trans, inode, path, cur_offset,
2405                                          drop_end);
2406                         if (ret) {
2407                                 err = ret;
2408                                 break;
2409                         }
2410                 }
2411
2412                 cur_offset = drop_end;
2413
2414                 ret = btrfs_update_inode(trans, root, inode);
2415                 if (ret) {
2416                         err = ret;
2417                         break;
2418                 }
2419
2420                 btrfs_end_transaction(trans, root);
2421                 btrfs_btree_balance_dirty(root);
2422
2423                 trans = btrfs_start_transaction(root, rsv_count);
2424                 if (IS_ERR(trans)) {
2425                         ret = PTR_ERR(trans);
2426                         trans = NULL;
2427                         break;
2428                 }
2429
2430                 ret = btrfs_block_rsv_migrate(&root->fs_info->trans_block_rsv,
2431                                               rsv, min_size);
2432                 BUG_ON(ret);    /* shouldn't happen */
2433                 trans->block_rsv = rsv;
2434
2435                 ret = find_first_non_hole(inode, &cur_offset, &len);
2436                 if (unlikely(ret < 0))
2437                         break;
2438                 if (ret && !len) {
2439                         ret = 0;
2440                         break;
2441                 }
2442         }
2443
2444         if (ret) {
2445                 err = ret;
2446                 goto out_trans;
2447         }
2448
2449         trans->block_rsv = &root->fs_info->trans_block_rsv;
2450         /*
2451          * Don't insert file hole extent item if it's for a range beyond eof
2452          * (because it's useless) or if it represents a 0 bytes range (when
2453          * cur_offset == drop_end).
2454          */
2455         if (cur_offset < ino_size && cur_offset < drop_end) {
2456                 ret = fill_holes(trans, inode, path, cur_offset, drop_end);
2457                 if (ret) {
2458                         err = ret;
2459                         goto out_trans;
2460                 }
2461         }
2462
2463 out_trans:
2464         if (!trans)
2465                 goto out_free;
2466
2467         inode_inc_iversion(inode);
2468         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
2469
2470         trans->block_rsv = &root->fs_info->trans_block_rsv;
2471         ret = btrfs_update_inode(trans, root, inode);
2472         btrfs_end_transaction(trans, root);
2473         btrfs_btree_balance_dirty(root);
2474 out_free:
2475         btrfs_free_path(path);
2476         btrfs_free_block_rsv(root, rsv);
2477 out:
2478         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2479                              &cached_state, GFP_NOFS);
2480 out_only_mutex:
2481         mutex_unlock(&inode->i_mutex);
2482         if (ret && !err)
2483                 err = ret;
2484         return err;
2485 }
2486
2487 static long btrfs_fallocate(struct file *file, int mode,
2488                             loff_t offset, loff_t len)
2489 {
2490         struct inode *inode = file_inode(file);
2491         struct extent_state *cached_state = NULL;
2492         struct btrfs_root *root = BTRFS_I(inode)->root;
2493         u64 cur_offset;
2494         u64 last_byte;
2495         u64 alloc_start;
2496         u64 alloc_end;
2497         u64 alloc_hint = 0;
2498         u64 locked_end;
2499         struct extent_map *em;
2500         int blocksize = BTRFS_I(inode)->root->sectorsize;
2501         int ret;
2502
2503         alloc_start = round_down(offset, blocksize);
2504         alloc_end = round_up(offset + len, blocksize);
2505
2506         /* Make sure we aren't being give some crap mode */
2507         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2508                 return -EOPNOTSUPP;
2509
2510         if (mode & FALLOC_FL_PUNCH_HOLE)
2511                 return btrfs_punch_hole(inode, offset, len);
2512
2513         /*
2514          * Make sure we have enough space before we do the
2515          * allocation.
2516          */
2517         ret = btrfs_check_data_free_space(inode, alloc_end - alloc_start);
2518         if (ret)
2519                 return ret;
2520         if (root->fs_info->quota_enabled) {
2521                 ret = btrfs_qgroup_reserve(root, alloc_end - alloc_start);
2522                 if (ret)
2523                         goto out_reserve_fail;
2524         }
2525
2526         mutex_lock(&inode->i_mutex);
2527         ret = inode_newsize_ok(inode, alloc_end);
2528         if (ret)
2529                 goto out;
2530
2531         if (alloc_start > inode->i_size) {
2532                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2533                                         alloc_start);
2534                 if (ret)
2535                         goto out;
2536         } else {
2537                 /*
2538                  * If we are fallocating from the end of the file onward we
2539                  * need to zero out the end of the page if i_size lands in the
2540                  * middle of a page.
2541                  */
2542                 ret = btrfs_truncate_page(inode, inode->i_size, 0, 0);
2543                 if (ret)
2544                         goto out;
2545         }
2546
2547         /*
2548          * wait for ordered IO before we have any locks.  We'll loop again
2549          * below with the locks held.
2550          */
2551         ret = btrfs_wait_ordered_range(inode, alloc_start,
2552                                        alloc_end - alloc_start);
2553         if (ret)
2554                 goto out;
2555
2556         locked_end = alloc_end - 1;
2557         while (1) {
2558                 struct btrfs_ordered_extent *ordered;
2559
2560                 /* the extent lock is ordered inside the running
2561                  * transaction
2562                  */
2563                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2564                                  locked_end, 0, &cached_state);
2565                 ordered = btrfs_lookup_first_ordered_extent(inode,
2566                                                             alloc_end - 1);
2567                 if (ordered &&
2568                     ordered->file_offset + ordered->len > alloc_start &&
2569                     ordered->file_offset < alloc_end) {
2570                         btrfs_put_ordered_extent(ordered);
2571                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2572                                              alloc_start, locked_end,
2573                                              &cached_state, GFP_NOFS);
2574                         /*
2575                          * we can't wait on the range with the transaction
2576                          * running or with the extent lock held
2577                          */
2578                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2579                                                        alloc_end - alloc_start);
2580                         if (ret)
2581                                 goto out;
2582                 } else {
2583                         if (ordered)
2584                                 btrfs_put_ordered_extent(ordered);
2585                         break;
2586                 }
2587         }
2588
2589         cur_offset = alloc_start;
2590         while (1) {
2591                 u64 actual_end;
2592
2593                 em = btrfs_get_extent(inode, NULL, 0, cur_offset,
2594                                       alloc_end - cur_offset, 0);
2595                 if (IS_ERR_OR_NULL(em)) {
2596                         if (!em)
2597                                 ret = -ENOMEM;
2598                         else
2599                                 ret = PTR_ERR(em);
2600                         break;
2601                 }
2602                 last_byte = min(extent_map_end(em), alloc_end);
2603                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2604                 last_byte = ALIGN(last_byte, blocksize);
2605
2606                 if (em->block_start == EXTENT_MAP_HOLE ||
2607                     (cur_offset >= inode->i_size &&
2608                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2609                         ret = btrfs_prealloc_file_range(inode, mode, cur_offset,
2610                                                         last_byte - cur_offset,
2611                                                         1 << inode->i_blkbits,
2612                                                         offset + len,
2613                                                         &alloc_hint);
2614
2615                         if (ret < 0) {
2616                                 free_extent_map(em);
2617                                 break;
2618                         }
2619                 } else if (actual_end > inode->i_size &&
2620                            !(mode & FALLOC_FL_KEEP_SIZE)) {
2621                         /*
2622                          * We didn't need to allocate any more space, but we
2623                          * still extended the size of the file so we need to
2624                          * update i_size.
2625                          */
2626                         inode->i_ctime = CURRENT_TIME;
2627                         i_size_write(inode, actual_end);
2628                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2629                 }
2630                 free_extent_map(em);
2631
2632                 cur_offset = last_byte;
2633                 if (cur_offset >= alloc_end) {
2634                         ret = 0;
2635                         break;
2636                 }
2637         }
2638         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2639                              &cached_state, GFP_NOFS);
2640 out:
2641         mutex_unlock(&inode->i_mutex);
2642         if (root->fs_info->quota_enabled)
2643                 btrfs_qgroup_free(root, alloc_end - alloc_start);
2644 out_reserve_fail:
2645         /* Let go of our reservation. */
2646         btrfs_free_reserved_data_space(inode, alloc_end - alloc_start);
2647         return ret;
2648 }
2649
2650 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2651 {
2652         struct btrfs_root *root = BTRFS_I(inode)->root;
2653         struct extent_map *em = NULL;
2654         struct extent_state *cached_state = NULL;
2655         u64 lockstart = *offset;
2656         u64 lockend = i_size_read(inode);
2657         u64 start = *offset;
2658         u64 len = i_size_read(inode);
2659         int ret = 0;
2660
2661         lockend = max_t(u64, root->sectorsize, lockend);
2662         if (lockend <= lockstart)
2663                 lockend = lockstart + root->sectorsize;
2664
2665         lockend--;
2666         len = lockend - lockstart + 1;
2667
2668         len = max_t(u64, len, root->sectorsize);
2669         if (inode->i_size == 0)
2670                 return -ENXIO;
2671
2672         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend, 0,
2673                          &cached_state);
2674
2675         while (start < inode->i_size) {
2676                 em = btrfs_get_extent_fiemap(inode, NULL, 0, start, len, 0);
2677                 if (IS_ERR(em)) {
2678                         ret = PTR_ERR(em);
2679                         em = NULL;
2680                         break;
2681                 }
2682
2683                 if (whence == SEEK_HOLE &&
2684                     (em->block_start == EXTENT_MAP_HOLE ||
2685                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2686                         break;
2687                 else if (whence == SEEK_DATA &&
2688                            (em->block_start != EXTENT_MAP_HOLE &&
2689                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
2690                         break;
2691
2692                 start = em->start + em->len;
2693                 free_extent_map(em);
2694                 em = NULL;
2695                 cond_resched();
2696         }
2697         free_extent_map(em);
2698         if (!ret) {
2699                 if (whence == SEEK_DATA && start >= inode->i_size)
2700                         ret = -ENXIO;
2701                 else
2702                         *offset = min_t(loff_t, start, inode->i_size);
2703         }
2704         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2705                              &cached_state, GFP_NOFS);
2706         return ret;
2707 }
2708
2709 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
2710 {
2711         struct inode *inode = file->f_mapping->host;
2712         int ret;
2713
2714         mutex_lock(&inode->i_mutex);
2715         switch (whence) {
2716         case SEEK_END:
2717         case SEEK_CUR:
2718                 offset = generic_file_llseek(file, offset, whence);
2719                 goto out;
2720         case SEEK_DATA:
2721         case SEEK_HOLE:
2722                 if (offset >= i_size_read(inode)) {
2723                         mutex_unlock(&inode->i_mutex);
2724                         return -ENXIO;
2725                 }
2726
2727                 ret = find_desired_extent(inode, &offset, whence);
2728                 if (ret) {
2729                         mutex_unlock(&inode->i_mutex);
2730                         return ret;
2731                 }
2732         }
2733
2734         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
2735 out:
2736         mutex_unlock(&inode->i_mutex);
2737         return offset;
2738 }
2739
2740 const struct file_operations btrfs_file_operations = {
2741         .llseek         = btrfs_file_llseek,
2742         .read           = do_sync_read,
2743         .write          = do_sync_write,
2744         .aio_read       = generic_file_aio_read,
2745         .splice_read    = generic_file_splice_read,
2746         .aio_write      = btrfs_file_aio_write,
2747         .mmap           = btrfs_file_mmap,
2748         .open           = generic_file_open,
2749         .release        = btrfs_release_file,
2750         .fsync          = btrfs_sync_file,
2751         .fallocate      = btrfs_fallocate,
2752         .unlocked_ioctl = btrfs_ioctl,
2753 #ifdef CONFIG_COMPAT
2754         .compat_ioctl   = btrfs_ioctl,
2755 #endif
2756 };
2757
2758 void btrfs_auto_defrag_exit(void)
2759 {
2760         if (btrfs_inode_defrag_cachep)
2761                 kmem_cache_destroy(btrfs_inode_defrag_cachep);
2762 }
2763
2764 int btrfs_auto_defrag_init(void)
2765 {
2766         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
2767                                         sizeof(struct inode_defrag), 0,
2768                                         SLAB_RECLAIM_ACCOUNT | SLAB_MEM_SPREAD,
2769                                         NULL);
2770         if (!btrfs_inode_defrag_cachep)
2771                 return -ENOMEM;
2772
2773         return 0;
2774 }