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