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