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/compat.h>
  31 #include <linux/slab.h>
  32 #include <linux/btrfs.h>
  33 #include <linux/uio.h>
  34 #include "ctree.h"
  35 #include "disk-io.h"
  36 #include "transaction.h"
  37 #include "btrfs_inode.h"
  38 #include "print-tree.h"
  39 #include "tree-log.h"
  40 #include "locking.h"
  41 #include "volumes.h"
  42 #include "qgroup.h"
  43 #include "compression.h"
  44
  45 static struct kmem_cache *btrfs_inode_defrag_cachep;
  46 /*
  47  * when auto defrag is enabled we
  48  * queue up these defrag structs to remember which
  49  * inodes need defragging passes
  50  */
  51 struct inode_defrag {
  52         struct rb_node rb_node;
  53         /* objectid */
  54         u64 ino;
  55         /*
  56          * transid where the defrag was added, we search for
  57          * extents newer than this
  58          */
  59         u64 transid;
  60
  61         /* root objectid */
  62         u64 root;
  63
  64         /* last offset we were able to defrag */
  65         u64 last_offset;
  66
  67         /* if we've wrapped around back to zero once already */
  68         int cycled;
  69 };
  70
  71 static int __compare_inode_defrag(struct inode_defrag *defrag1,
  72                                   struct inode_defrag *defrag2)
  73 {
  74         if (defrag1->root > defrag2->root)
  75                 return 1;
  76         else if (defrag1->root < defrag2->root)
  77                 return -1;
  78         else if (defrag1->ino > defrag2->ino)
  79                 return 1;
  80         else if (defrag1->ino < defrag2->ino)
  81                 return -1;
  82         else
  83                 return 0;
  84 }
  85
  86 /* pop a record for an inode into the defrag tree.  The lock
  87  * must be held already
  88  *
  89  * If you're inserting a record for an older transid than an
  90  * existing record, the transid already in the tree is lowered
  91  *
  92  * If an existing record is found the defrag item you
  93  * pass in is freed
  94  */
  95 static int __btrfs_add_inode_defrag(struct btrfs_inode *inode,
  96                                     struct inode_defrag *defrag)
  97 {
  98         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
  99         struct inode_defrag *entry;
 100         struct rb_node **p;
 101         struct rb_node *parent = NULL;
 102         int ret;
 103
 104         p = &fs_info->defrag_inodes.rb_node;
 105         while (*p) {
 106                 parent = *p;
 107                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 108
 109                 ret = __compare_inode_defrag(defrag, entry);
 110                 if (ret < 0)
 111                         p = &parent->rb_left;
 112                 else if (ret > 0)
 113                         p = &parent->rb_right;
 114                 else {
 115                         /* if we're reinserting an entry for
 116                          * an old defrag run, make sure to
 117                          * lower the transid of our existing record
 118                          */
 119                         if (defrag->transid < entry->transid)
 120                                 entry->transid = defrag->transid;
 121                         if (defrag->last_offset > entry->last_offset)
 122                                 entry->last_offset = defrag->last_offset;
 123                         return -EEXIST;
 124                 }
 125         }
 126         set_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags);
 127         rb_link_node(&defrag->rb_node, parent, p);
 128         rb_insert_color(&defrag->rb_node, &fs_info->defrag_inodes);
 129         return 0;
 130 }
 131
 132 static inline int __need_auto_defrag(struct btrfs_fs_info *fs_info)
 133 {
 134         if (!btrfs_test_opt(fs_info, AUTO_DEFRAG))
 135                 return 0;
 136
 137         if (btrfs_fs_closing(fs_info))
 138                 return 0;
 139
 140         return 1;
 141 }
 142
 143 /*
 144  * insert a defrag record for this inode if auto defrag is
 145  * enabled
 146  */
 147 int btrfs_add_inode_defrag(struct btrfs_trans_handle *trans,
 148                            struct btrfs_inode *inode)
 149 {
 150         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 151         struct btrfs_root *root = inode->root;
 152         struct inode_defrag *defrag;
 153         u64 transid;
 154         int ret;
 155
 156         if (!__need_auto_defrag(fs_info))
 157                 return 0;
 158
 159         if (test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags))
 160                 return 0;
 161
 162         if (trans)
 163                 transid = trans->transid;
 164         else
 165                 transid = inode->root->last_trans;
 166
 167         defrag = kmem_cache_zalloc(btrfs_inode_defrag_cachep, GFP_NOFS);
 168         if (!defrag)
 169                 return -ENOMEM;
 170
 171         defrag->ino = btrfs_ino(inode);
 172         defrag->transid = transid;
 173         defrag->root = root->root_key.objectid;
 174
 175         spin_lock(&fs_info->defrag_inodes_lock);
 176         if (!test_bit(BTRFS_INODE_IN_DEFRAG, &inode->runtime_flags)) {
 177                 /*
 178                  * If we set IN_DEFRAG flag and evict the inode from memory,
 179                  * and then re-read this inode, this new inode doesn't have
 180                  * IN_DEFRAG flag. At the case, we may find the existed defrag.
 181                  */
 182                 ret = __btrfs_add_inode_defrag(inode, defrag);
 183                 if (ret)
 184                         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 185         } else {
 186                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 187         }
 188         spin_unlock(&fs_info->defrag_inodes_lock);
 189         return 0;
 190 }
 191
 192 /*
 193  * Requeue the defrag object. If there is a defrag object that points to
 194  * the same inode in the tree, we will merge them together (by
 195  * __btrfs_add_inode_defrag()) and free the one that we want to requeue.
 196  */
 197 static void btrfs_requeue_inode_defrag(struct btrfs_inode *inode,
 198                                        struct inode_defrag *defrag)
 199 {
 200         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
 201         int ret;
 202
 203         if (!__need_auto_defrag(fs_info))
 204                 goto out;
 205
 206         /*
 207          * Here we don't check the IN_DEFRAG flag, because we need merge
 208          * them together.
 209          */
 210         spin_lock(&fs_info->defrag_inodes_lock);
 211         ret = __btrfs_add_inode_defrag(inode, defrag);
 212         spin_unlock(&fs_info->defrag_inodes_lock);
 213         if (ret)
 214                 goto out;
 215         return;
 216 out:
 217         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 218 }
 219
 220 /*
 221  * pick the defragable inode that we want, if it doesn't exist, we will get
 222  * the next one.
 223  */
 224 static struct inode_defrag *
 225 btrfs_pick_defrag_inode(struct btrfs_fs_info *fs_info, u64 root, u64 ino)
 226 {
 227         struct inode_defrag *entry = NULL;
 228         struct inode_defrag tmp;
 229         struct rb_node *p;
 230         struct rb_node *parent = NULL;
 231         int ret;
 232
 233         tmp.ino = ino;
 234         tmp.root = root;
 235
 236         spin_lock(&fs_info->defrag_inodes_lock);
 237         p = fs_info->defrag_inodes.rb_node;
 238         while (p) {
 239                 parent = p;
 240                 entry = rb_entry(parent, struct inode_defrag, rb_node);
 241
 242                 ret = __compare_inode_defrag(&tmp, entry);
 243                 if (ret < 0)
 244                         p = parent->rb_left;
 245                 else if (ret > 0)
 246                         p = parent->rb_right;
 247                 else
 248                         goto out;
 249         }
 250
 251         if (parent && __compare_inode_defrag(&tmp, entry) > 0) {
 252                 parent = rb_next(parent);
 253                 if (parent)
 254                         entry = rb_entry(parent, struct inode_defrag, rb_node);
 255                 else
 256                         entry = NULL;
 257         }
 258 out:
 259         if (entry)
 260                 rb_erase(parent, &fs_info->defrag_inodes);
 261         spin_unlock(&fs_info->defrag_inodes_lock);
 262         return entry;
 263 }
 264
 265 void btrfs_cleanup_defrag_inodes(struct btrfs_fs_info *fs_info)
 266 {
 267         struct inode_defrag *defrag;
 268         struct rb_node *node;
 269
 270         spin_lock(&fs_info->defrag_inodes_lock);
 271         node = rb_first(&fs_info->defrag_inodes);
 272         while (node) {
 273                 rb_erase(node, &fs_info->defrag_inodes);
 274                 defrag = rb_entry(node, struct inode_defrag, rb_node);
 275                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 276
 277                 cond_resched_lock(&fs_info->defrag_inodes_lock);
 278
 279                 node = rb_first(&fs_info->defrag_inodes);
 280         }
 281         spin_unlock(&fs_info->defrag_inodes_lock);
 282 }
 283
 284 #define BTRFS_DEFRAG_BATCH      1024
 285
 286 static int __btrfs_run_defrag_inode(struct btrfs_fs_info *fs_info,
 287                                     struct inode_defrag *defrag)
 288 {
 289         struct btrfs_root *inode_root;
 290         struct inode *inode;
 291         struct btrfs_key key;
 292         struct btrfs_ioctl_defrag_range_args range;
 293         int num_defrag;
 294         int index;
 295         int ret;
 296
 297         /* get the inode */
 298         key.objectid = defrag->root;
 299         key.type = BTRFS_ROOT_ITEM_KEY;
 300         key.offset = (u64)-1;
 301
 302         index = srcu_read_lock(&fs_info->subvol_srcu);
 303
 304         inode_root = btrfs_read_fs_root_no_name(fs_info, &key);
 305         if (IS_ERR(inode_root)) {
 306                 ret = PTR_ERR(inode_root);
 307                 goto cleanup;
 308         }
 309
 310         key.objectid = defrag->ino;
 311         key.type = BTRFS_INODE_ITEM_KEY;
 312         key.offset = 0;
 313         inode = btrfs_iget(fs_info->sb, &key, inode_root, NULL);
 314         if (IS_ERR(inode)) {
 315                 ret = PTR_ERR(inode);
 316                 goto cleanup;
 317         }
 318         srcu_read_unlock(&fs_info->subvol_srcu, index);
 319
 320         /* do a chunk of defrag */
 321         clear_bit(BTRFS_INODE_IN_DEFRAG, &BTRFS_I(inode)->runtime_flags);
 322         memset(&range, 0, sizeof(range));
 323         range.len = (u64)-1;
 324         range.start = defrag->last_offset;
 325
 326         sb_start_write(fs_info->sb);
 327         num_defrag = btrfs_defrag_file(inode, NULL, &range, defrag->transid,
 328                                        BTRFS_DEFRAG_BATCH);
 329         sb_end_write(fs_info->sb);
 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 == BTRFS_DEFRAG_BATCH) {
 336                 defrag->last_offset = range.start;
 337                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 338         } else if (defrag->last_offset && !defrag->cycled) {
 339                 /*
 340                  * we didn't fill our defrag batch, but
 341                  * we didn't start at zero.  Make sure we loop
 342                  * around to the start of the file.
 343                  */
 344                 defrag->last_offset = 0;
 345                 defrag->cycled = 1;
 346                 btrfs_requeue_inode_defrag(BTRFS_I(inode), defrag);
 347         } else {
 348                 kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 349         }
 350
 351         iput(inode);
 352         return 0;
 353 cleanup:
 354         srcu_read_unlock(&fs_info->subvol_srcu, index);
 355         kmem_cache_free(btrfs_inode_defrag_cachep, defrag);
 356         return ret;
 357 }
 358
 359 /*
 360  * run through the list of inodes in the FS that need
 361  * defragging
 362  */
 363 int btrfs_run_defrag_inodes(struct btrfs_fs_info *fs_info)
 364 {
 365         struct inode_defrag *defrag;
 366         u64 first_ino = 0;
 367         u64 root_objectid = 0;
 368
 369         atomic_inc(&fs_info->defrag_running);
 370         while (1) {
 371                 /* Pause the auto defragger. */
 372                 if (test_bit(BTRFS_FS_STATE_REMOUNTING,
 373                              &fs_info->fs_state))
 374                         break;
 375
 376                 if (!__need_auto_defrag(fs_info))
 377                         break;
 378
 379                 /* find an inode to defrag */
 380                 defrag = btrfs_pick_defrag_inode(fs_info, root_objectid,
 381                                                  first_ino);
 382                 if (!defrag) {
 383                         if (root_objectid || first_ino) {
 384                                 root_objectid = 0;
 385                                 first_ino = 0;
 386                                 continue;
 387                         } else {
 388                                 break;
 389                         }
 390                 }
 391
 392                 first_ino = defrag->ino + 1;
 393                 root_objectid = defrag->root;
 394
 395                 __btrfs_run_defrag_inode(fs_info, defrag);
 396         }
 397         atomic_dec(&fs_info->defrag_running);
 398
 399         /*
 400          * during unmount, we use the transaction_wait queue to
 401          * wait for the defragger to stop
 402          */
 403         wake_up(&fs_info->transaction_wait);
 404         return 0;
 405 }
 406
 407 /* simple helper to fault in pages and copy.  This should go away
 408  * and be replaced with calls into generic code.
 409  */
 410 static noinline int btrfs_copy_from_user(loff_t pos, size_t write_bytes,
 411                                          struct page **prepared_pages,
 412                                          struct iov_iter *i)
 413 {
 414         size_t copied = 0;
 415         size_t total_copied = 0;
 416         int pg = 0;
 417         int offset = pos & (PAGE_SIZE - 1);
 418
 419         while (write_bytes > 0) {
 420                 size_t count = min_t(size_t,
 421                                      PAGE_SIZE - offset, write_bytes);
 422                 struct page *page = prepared_pages[pg];
 423                 /*
 424                  * Copy data from userspace to the current page
 425                  */
 426                 copied = iov_iter_copy_from_user_atomic(page, i, offset, count);
 427
 428                 /* Flush processor's dcache for this page */
 429                 flush_dcache_page(page);
 430
 431                 /*
 432                  * if we get a partial write, we can end up with
 433                  * partially up to date pages.  These add
 434                  * a lot of complexity, so make sure they don't
 435                  * happen by forcing this copy to be retried.
 436                  *
 437                  * The rest of the btrfs_file_write code will fall
 438                  * back to page at a time copies after we return 0.
 439                  */
 440                 if (!PageUptodate(page) && copied < count)
 441                         copied = 0;
 442
 443                 iov_iter_advance(i, copied);
 444                 write_bytes -= copied;
 445                 total_copied += copied;
 446
 447                 /* Return to btrfs_file_write_iter to fault page */
 448                 if (unlikely(copied == 0))
 449                         break;
 450
 451                 if (copied < PAGE_SIZE - offset) {
 452                         offset += copied;
 453                 } else {
 454                         pg++;
 455                         offset = 0;
 456                 }
 457         }
 458         return total_copied;
 459 }
 460
 461 /*
 462  * unlocks pages after btrfs_file_write is done with them
 463  */
 464 static void btrfs_drop_pages(struct page **pages, size_t num_pages)
 465 {
 466         size_t i;
 467         for (i = 0; i < num_pages; i++) {
 468                 /* page checked is some magic around finding pages that
 469                  * have been modified without going through btrfs_set_page_dirty
 470                  * clear it here. There should be no need to mark the pages
 471                  * accessed as prepare_pages should have marked them accessed
 472                  * in prepare_pages via find_or_create_page()
 473                  */
 474                 ClearPageChecked(pages[i]);
 475                 unlock_page(pages[i]);
 476                 put_page(pages[i]);
 477         }
 478 }
 479
 480 /*
 481  * after copy_from_user, pages need to be dirtied and we need to make
 482  * sure holes are created between the current EOF and the start of
 483  * any next extents (if required).
 484  *
 485  * this also makes the decision about creating an inline extent vs
 486  * doing real data extents, marking pages dirty and delalloc as required.
 487  */
 488 int btrfs_dirty_pages(struct inode *inode, struct page **pages,
 489                       size_t num_pages, loff_t pos, size_t write_bytes,
 490                       struct extent_state **cached)
 491 {
 492         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
 493         int err = 0;
 494         int i;
 495         u64 num_bytes;
 496         u64 start_pos;
 497         u64 end_of_last_block;
 498         u64 end_pos = pos + write_bytes;
 499         loff_t isize = i_size_read(inode);
 500
 501         start_pos = pos & ~((u64) fs_info->sectorsize - 1);
 502         num_bytes = round_up(write_bytes + pos - start_pos,
 503                              fs_info->sectorsize);
 504
 505         end_of_last_block = start_pos + num_bytes - 1;
 506         err = btrfs_set_extent_delalloc(inode, start_pos, end_of_last_block,
 507                                         cached, 0);
 508         if (err)
 509                 return err;
 510
 511         for (i = 0; i < num_pages; i++) {
 512                 struct page *p = pages[i];
 513                 SetPageUptodate(p);
 514                 ClearPageChecked(p);
 515                 set_page_dirty(p);
 516         }
 517
 518         /*
 519          * we've only changed i_size in ram, and we haven't updated
 520          * the disk i_size.  There is no need to log the inode
 521          * at this time.
 522          */
 523         if (end_pos > isize)
 524                 i_size_write(inode, end_pos);
 525         return 0;
 526 }
 527
 528 /*
 529  * this drops all the extents in the cache that intersect the range
 530  * [start, end].  Existing extents are split as required.
 531  */
 532 void btrfs_drop_extent_cache(struct btrfs_inode *inode, u64 start, u64 end,
 533                              int skip_pinned)
 534 {
 535         struct extent_map *em;
 536         struct extent_map *split = NULL;
 537         struct extent_map *split2 = NULL;
 538         struct extent_map_tree *em_tree = &inode->extent_tree;
 539         u64 len = end - start + 1;
 540         u64 gen;
 541         int ret;
 542         int testend = 1;
 543         unsigned long flags;
 544         int compressed = 0;
 545         bool modified;
 546
 547         WARN_ON(end < start);
 548         if (end == (u64)-1) {
 549                 len = (u64)-1;
 550                 testend = 0;
 551         }
 552         while (1) {
 553                 int no_splits = 0;
 554
 555                 modified = false;
 556                 if (!split)
 557                         split = alloc_extent_map();
 558                 if (!split2)
 559                         split2 = alloc_extent_map();
 560                 if (!split || !split2)
 561                         no_splits = 1;
 562
 563                 write_lock(&em_tree->lock);
 564                 em = lookup_extent_mapping(em_tree, start, len);
 565                 if (!em) {
 566                         write_unlock(&em_tree->lock);
 567                         break;
 568                 }
 569                 flags = em->flags;
 570                 gen = em->generation;
 571                 if (skip_pinned && test_bit(EXTENT_FLAG_PINNED, &em->flags)) {
 572                         if (testend && em->start + em->len >= start + len) {
 573                                 free_extent_map(em);
 574                                 write_unlock(&em_tree->lock);
 575                                 break;
 576                         }
 577                         start = em->start + em->len;
 578                         if (testend)
 579                                 len = start + len - (em->start + em->len);
 580                         free_extent_map(em);
 581                         write_unlock(&em_tree->lock);
 582                         continue;
 583                 }
 584                 compressed = test_bit(EXTENT_FLAG_COMPRESSED, &em->flags);
 585                 clear_bit(EXTENT_FLAG_PINNED, &em->flags);
 586                 clear_bit(EXTENT_FLAG_LOGGING, &flags);
 587                 modified = !list_empty(&em->list);
 588                 if (no_splits)
 589                         goto next;
 590
 591                 if (em->start < start) {
 592                         split->start = em->start;
 593                         split->len = start - em->start;
 594
 595                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 596                                 split->orig_start = em->orig_start;
 597                                 split->block_start = em->block_start;
 598
 599                                 if (compressed)
 600                                         split->block_len = em->block_len;
 601                                 else
 602                                         split->block_len = split->len;
 603                                 split->orig_block_len = max(split->block_len,
 604                                                 em->orig_block_len);
 605                                 split->ram_bytes = em->ram_bytes;
 606                         } else {
 607                                 split->orig_start = split->start;
 608                                 split->block_len = 0;
 609                                 split->block_start = em->block_start;
 610                                 split->orig_block_len = 0;
 611                                 split->ram_bytes = split->len;
 612                         }
 613
 614                         split->generation = gen;
 615                         split->bdev = em->bdev;
 616                         split->flags = flags;
 617                         split->compress_type = em->compress_type;
 618                         replace_extent_mapping(em_tree, em, split, modified);
 619                         free_extent_map(split);
 620                         split = split2;
 621                         split2 = NULL;
 622                 }
 623                 if (testend && em->start + em->len > start + len) {
 624                         u64 diff = start + len - em->start;
 625
 626                         split->start = start + len;
 627                         split->len = em->start + em->len - (start + len);
 628                         split->bdev = em->bdev;
 629                         split->flags = flags;
 630                         split->compress_type = em->compress_type;
 631                         split->generation = gen;
 632
 633                         if (em->block_start < EXTENT_MAP_LAST_BYTE) {
 634                                 split->orig_block_len = max(em->block_len,
 635                                                     em->orig_block_len);
 636
 637                                 split->ram_bytes = em->ram_bytes;
 638                                 if (compressed) {
 639                                         split->block_len = em->block_len;
 640                                         split->block_start = em->block_start;
 641                                         split->orig_start = em->orig_start;
 642                                 } else {
 643                                         split->block_len = split->len;
 644                                         split->block_start = em->block_start
 645                                                 + diff;
 646                                         split->orig_start = em->orig_start;
 647                                 }
 648                         } else {
 649                                 split->ram_bytes = split->len;
 650                                 split->orig_start = split->start;
 651                                 split->block_len = 0;
 652                                 split->block_start = em->block_start;
 653                                 split->orig_block_len = 0;
 654                         }
 655
 656                         if (extent_map_in_tree(em)) {
 657                                 replace_extent_mapping(em_tree, em, split,
 658                                                        modified);
 659                         } else {
 660                                 ret = add_extent_mapping(em_tree, split,
 661                                                          modified);
 662                                 ASSERT(ret == 0); /* Logic error */
 663                         }
 664                         free_extent_map(split);
 665                         split = NULL;
 666                 }
 667 next:
 668                 if (extent_map_in_tree(em))
 669                         remove_extent_mapping(em_tree, em);
 670                 write_unlock(&em_tree->lock);
 671
 672                 /* once for us */
 673                 free_extent_map(em);
 674                 /* once for the tree*/
 675                 free_extent_map(em);
 676         }
 677         if (split)
 678                 free_extent_map(split);
 679         if (split2)
 680                 free_extent_map(split2);
 681 }
 682
 683 /*
 684  * this is very complex, but the basic idea is to drop all extents
 685  * in the range start - end.  hint_block is filled in with a block number
 686  * that would be a good hint to the block allocator for this file.
 687  *
 688  * If an extent intersects the range but is not entirely inside the range
 689  * it is either truncated or split.  Anything entirely inside the range
 690  * is deleted from the tree.
 691  */
 692 int __btrfs_drop_extents(struct btrfs_trans_handle *trans,
 693                          struct btrfs_root *root, struct inode *inode,
 694                          struct btrfs_path *path, u64 start, u64 end,
 695                          u64 *drop_end, int drop_cache,
 696                          int replace_extent,
 697                          u32 extent_item_size,
 698                          int *key_inserted)
 699 {
 700         struct btrfs_fs_info *fs_info = root->fs_info;
 701         struct extent_buffer *leaf;
 702         struct btrfs_file_extent_item *fi;
 703         struct btrfs_key key;
 704         struct btrfs_key new_key;
 705         u64 ino = btrfs_ino(BTRFS_I(inode));
 706         u64 search_start = start;
 707         u64 disk_bytenr = 0;
 708         u64 num_bytes = 0;
 709         u64 extent_offset = 0;
 710         u64 extent_end = 0;
 711         u64 last_end = start;
 712         int del_nr = 0;
 713         int del_slot = 0;
 714         int extent_type;
 715         int recow;
 716         int ret;
 717         int modify_tree = -1;
 718         int update_refs;
 719         int found = 0;
 720         int leafs_visited = 0;
 721
 722         if (drop_cache)
 723                 btrfs_drop_extent_cache(BTRFS_I(inode), start, end - 1, 0);
 724
 725         if (start >= BTRFS_I(inode)->disk_i_size && !replace_extent)
 726                 modify_tree = 0;
 727
 728         update_refs = (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
 729                        root == fs_info->tree_root);
 730         while (1) {
 731                 recow = 0;
 732                 ret = btrfs_lookup_file_extent(trans, root, path, ino,
 733                                                search_start, modify_tree);
 734                 if (ret < 0)
 735                         break;
 736                 if (ret > 0 && path->slots[0] > 0 && search_start == start) {
 737                         leaf = path->nodes[0];
 738                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0] - 1);
 739                         if (key.objectid == ino &&
 740                             key.type == BTRFS_EXTENT_DATA_KEY)
 741                                 path->slots[0]--;
 742                 }
 743                 ret = 0;
 744                 leafs_visited++;
 745 next_slot:
 746                 leaf = path->nodes[0];
 747                 if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 748                         BUG_ON(del_nr > 0);
 749                         ret = btrfs_next_leaf(root, path);
 750                         if (ret < 0)
 751                                 break;
 752                         if (ret > 0) {
 753                                 ret = 0;
 754                                 break;
 755                         }
 756                         leafs_visited++;
 757                         leaf = path->nodes[0];
 758                         recow = 1;
 759                 }
 760
 761                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 762
 763                 if (key.objectid > ino)
 764                         break;
 765                 if (WARN_ON_ONCE(key.objectid < ino) ||
 766                     key.type < BTRFS_EXTENT_DATA_KEY) {
 767                         ASSERT(del_nr == 0);
 768                         path->slots[0]++;
 769                         goto next_slot;
 770                 }
 771                 if (key.type > BTRFS_EXTENT_DATA_KEY || key.offset >= end)
 772                         break;
 773
 774                 fi = btrfs_item_ptr(leaf, path->slots[0],
 775                                     struct btrfs_file_extent_item);
 776                 extent_type = btrfs_file_extent_type(leaf, fi);
 777
 778                 if (extent_type == BTRFS_FILE_EXTENT_REG ||
 779                     extent_type == BTRFS_FILE_EXTENT_PREALLOC) {
 780                         disk_bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
 781                         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
 782                         extent_offset = btrfs_file_extent_offset(leaf, fi);
 783                         extent_end = key.offset +
 784                                 btrfs_file_extent_num_bytes(leaf, fi);
 785                 } else if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 786                         extent_end = key.offset +
 787                                 btrfs_file_extent_inline_len(leaf,
 788                                                      path->slots[0], fi);
 789                 } else {
 790                         /* can't happen */
 791                         BUG();
 792                 }
 793
 794                 /*
 795                  * Don't skip extent items representing 0 byte lengths. They
 796                  * used to be created (bug) if while punching holes we hit
 797                  * -ENOSPC condition. So if we find one here, just ensure we
 798                  * delete it, otherwise we would insert a new file extent item
 799                  * with the same key (offset) as that 0 bytes length file
 800                  * extent item in the call to setup_items_for_insert() later
 801                  * in this function.
 802                  */
 803                 if (extent_end == key.offset && extent_end >= search_start) {
 804                         last_end = extent_end;
 805                         goto delete_extent_item;
 806                 }
 807
 808                 if (extent_end <= search_start) {
 809                         path->slots[0]++;
 810                         goto next_slot;
 811                 }
 812
 813                 found = 1;
 814                 search_start = max(key.offset, start);
 815                 if (recow || !modify_tree) {
 816                         modify_tree = -1;
 817                         btrfs_release_path(path);
 818                         continue;
 819                 }
 820
 821                 /*
 822                  *     | - range to drop - |
 823                  *  | -------- extent -------- |
 824                  */
 825                 if (start > key.offset && end < extent_end) {
 826                         BUG_ON(del_nr > 0);
 827                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 828                                 ret = -EOPNOTSUPP;
 829                                 break;
 830                         }
 831
 832                         memcpy(&new_key, &key, sizeof(new_key));
 833                         new_key.offset = start;
 834                         ret = btrfs_duplicate_item(trans, root, path,
 835                                                    &new_key);
 836                         if (ret == -EAGAIN) {
 837                                 btrfs_release_path(path);
 838                                 continue;
 839                         }
 840                         if (ret < 0)
 841                                 break;
 842
 843                         leaf = path->nodes[0];
 844                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
 845                                             struct btrfs_file_extent_item);
 846                         btrfs_set_file_extent_num_bytes(leaf, fi,
 847                                                         start - key.offset);
 848
 849                         fi = btrfs_item_ptr(leaf, path->slots[0],
 850                                             struct btrfs_file_extent_item);
 851
 852                         extent_offset += start - key.offset;
 853                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 854                         btrfs_set_file_extent_num_bytes(leaf, fi,
 855                                                         extent_end - start);
 856                         btrfs_mark_buffer_dirty(leaf);
 857
 858                         if (update_refs && disk_bytenr > 0) {
 859                                 ret = btrfs_inc_extent_ref(trans, fs_info,
 860                                                 disk_bytenr, num_bytes, 0,
 861                                                 root->root_key.objectid,
 862                                                 new_key.objectid,
 863                                                 start - extent_offset);
 864                                 BUG_ON(ret); /* -ENOMEM */
 865                         }
 866                         key.offset = start;
 867                 }
 868                 /*
 869                  * From here on out we will have actually dropped something, so
 870                  * last_end can be updated.
 871                  */
 872                 last_end = extent_end;
 873
 874                 /*
 875                  *  | ---- range to drop ----- |
 876                  *      | -------- extent -------- |
 877                  */
 878                 if (start <= key.offset && end < extent_end) {
 879                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 880                                 ret = -EOPNOTSUPP;
 881                                 break;
 882                         }
 883
 884                         memcpy(&new_key, &key, sizeof(new_key));
 885                         new_key.offset = end;
 886                         btrfs_set_item_key_safe(fs_info, path, &new_key);
 887
 888                         extent_offset += end - key.offset;
 889                         btrfs_set_file_extent_offset(leaf, fi, extent_offset);
 890                         btrfs_set_file_extent_num_bytes(leaf, fi,
 891                                                         extent_end - end);
 892                         btrfs_mark_buffer_dirty(leaf);
 893                         if (update_refs && disk_bytenr > 0)
 894                                 inode_sub_bytes(inode, end - key.offset);
 895                         break;
 896                 }
 897
 898                 search_start = extent_end;
 899                 /*
 900                  *       | ---- range to drop ----- |
 901                  *  | -------- extent -------- |
 902                  */
 903                 if (start > key.offset && end >= extent_end) {
 904                         BUG_ON(del_nr > 0);
 905                         if (extent_type == BTRFS_FILE_EXTENT_INLINE) {
 906                                 ret = -EOPNOTSUPP;
 907                                 break;
 908                         }
 909
 910                         btrfs_set_file_extent_num_bytes(leaf, fi,
 911                                                         start - key.offset);
 912                         btrfs_mark_buffer_dirty(leaf);
 913                         if (update_refs && disk_bytenr > 0)
 914                                 inode_sub_bytes(inode, extent_end - start);
 915                         if (end == extent_end)
 916                                 break;
 917
 918                         path->slots[0]++;
 919                         goto next_slot;
 920                 }
 921
 922                 /*
 923                  *  | ---- range to drop ----- |
 924                  *    | ------ extent ------ |
 925                  */
 926                 if (start <= key.offset && end >= extent_end) {
 927 delete_extent_item:
 928                         if (del_nr == 0) {
 929                                 del_slot = path->slots[0];
 930                                 del_nr = 1;
 931                         } else {
 932                                 BUG_ON(del_slot + del_nr != path->slots[0]);
 933                                 del_nr++;
 934                         }
 935
 936                         if (update_refs &&
 937                             extent_type == BTRFS_FILE_EXTENT_INLINE) {
 938                                 inode_sub_bytes(inode,
 939                                                 extent_end - key.offset);
 940                                 extent_end = ALIGN(extent_end,
 941                                                    fs_info->sectorsize);
 942                         } else if (update_refs && disk_bytenr > 0) {
 943                                 ret = btrfs_free_extent(trans, fs_info,
 944                                                 disk_bytenr, num_bytes, 0,
 945                                                 root->root_key.objectid,
 946                                                 key.objectid, key.offset -
 947                                                 extent_offset);
 948                                 BUG_ON(ret); /* -ENOMEM */
 949                                 inode_sub_bytes(inode,
 950                                                 extent_end - key.offset);
 951                         }
 952
 953                         if (end == extent_end)
 954                                 break;
 955
 956                         if (path->slots[0] + 1 < btrfs_header_nritems(leaf)) {
 957                                 path->slots[0]++;
 958                                 goto next_slot;
 959                         }
 960
 961                         ret = btrfs_del_items(trans, root, path, del_slot,
 962                                               del_nr);
 963                         if (ret) {
 964                                 btrfs_abort_transaction(trans, ret);
 965                                 break;
 966                         }
 967
 968                         del_nr = 0;
 969                         del_slot = 0;
 970
 971                         btrfs_release_path(path);
 972                         continue;
 973                 }
 974
 975                 BUG_ON(1);
 976         }
 977
 978         if (!ret && del_nr > 0) {
 979                 /*
 980                  * Set path->slots[0] to first slot, so that after the delete
 981                  * if items are move off from our leaf to its immediate left or
 982                  * right neighbor leafs, we end up with a correct and adjusted
 983                  * path->slots[0] for our insertion (if replace_extent != 0).
 984                  */
 985                 path->slots[0] = del_slot;
 986                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
 987                 if (ret)
 988                         btrfs_abort_transaction(trans, ret);
 989         }
 990
 991         leaf = path->nodes[0];
 992         /*
 993          * If btrfs_del_items() was called, it might have deleted a leaf, in
 994          * which case it unlocked our path, so check path->locks[0] matches a
 995          * write lock.
 996          */
 997         if (!ret && replace_extent && leafs_visited == 1 &&
 998             (path->locks[0] == BTRFS_WRITE_LOCK_BLOCKING ||
 999              path->locks[0] == BTRFS_WRITE_LOCK) &&
1000             btrfs_leaf_free_space(fs_info, leaf) >=
1001             sizeof(struct btrfs_item) + extent_item_size) {
1002
1003                 key.objectid = ino;
1004                 key.type = BTRFS_EXTENT_DATA_KEY;
1005                 key.offset = start;
1006                 if (!del_nr && path->slots[0] < btrfs_header_nritems(leaf)) {
1007                         struct btrfs_key slot_key;
1008
1009                         btrfs_item_key_to_cpu(leaf, &slot_key, path->slots[0]);
1010                         if (btrfs_comp_cpu_keys(&key, &slot_key) > 0)
1011                                 path->slots[0]++;
1012                 }
1013                 setup_items_for_insert(root, path, &key,
1014                                        &extent_item_size,
1015                                        extent_item_size,
1016                                        sizeof(struct btrfs_item) +
1017                                        extent_item_size, 1);
1018                 *key_inserted = 1;
1019         }
1020
1021         if (!replace_extent || !(*key_inserted))
1022                 btrfs_release_path(path);
1023         if (drop_end)
1024                 *drop_end = found ? min(end, last_end) : end;
1025         return ret;
1026 }
1027
1028 int btrfs_drop_extents(struct btrfs_trans_handle *trans,
1029                        struct btrfs_root *root, struct inode *inode, u64 start,
1030                        u64 end, int drop_cache)
1031 {
1032         struct btrfs_path *path;
1033         int ret;
1034
1035         path = btrfs_alloc_path();
1036         if (!path)
1037                 return -ENOMEM;
1038         ret = __btrfs_drop_extents(trans, root, inode, path, start, end, NULL,
1039                                    drop_cache, 0, 0, NULL);
1040         btrfs_free_path(path);
1041         return ret;
1042 }
1043
1044 static int extent_mergeable(struct extent_buffer *leaf, int slot,
1045                             u64 objectid, u64 bytenr, u64 orig_offset,
1046                             u64 *start, u64 *end)
1047 {
1048         struct btrfs_file_extent_item *fi;
1049         struct btrfs_key key;
1050         u64 extent_end;
1051
1052         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
1053                 return 0;
1054
1055         btrfs_item_key_to_cpu(leaf, &key, slot);
1056         if (key.objectid != objectid || key.type != BTRFS_EXTENT_DATA_KEY)
1057                 return 0;
1058
1059         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
1060         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG ||
1061             btrfs_file_extent_disk_bytenr(leaf, fi) != bytenr ||
1062             btrfs_file_extent_offset(leaf, fi) != key.offset - orig_offset ||
1063             btrfs_file_extent_compression(leaf, fi) ||
1064             btrfs_file_extent_encryption(leaf, fi) ||
1065             btrfs_file_extent_other_encoding(leaf, fi))
1066                 return 0;
1067
1068         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1069         if ((*start && *start != key.offset) || (*end && *end != extent_end))
1070                 return 0;
1071
1072         *start = key.offset;
1073         *end = extent_end;
1074         return 1;
1075 }
1076
1077 /*
1078  * Mark extent in the range start - end as written.
1079  *
1080  * This changes extent type from 'pre-allocated' to 'regular'. If only
1081  * part of extent is marked as written, the extent will be split into
1082  * two or three.
1083  */
1084 int btrfs_mark_extent_written(struct btrfs_trans_handle *trans,
1085                               struct btrfs_inode *inode, u64 start, u64 end)
1086 {
1087         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1088         struct btrfs_root *root = inode->root;
1089         struct extent_buffer *leaf;
1090         struct btrfs_path *path;
1091         struct btrfs_file_extent_item *fi;
1092         struct btrfs_key key;
1093         struct btrfs_key new_key;
1094         u64 bytenr;
1095         u64 num_bytes;
1096         u64 extent_end;
1097         u64 orig_offset;
1098         u64 other_start;
1099         u64 other_end;
1100         u64 split;
1101         int del_nr = 0;
1102         int del_slot = 0;
1103         int recow;
1104         int ret;
1105         u64 ino = btrfs_ino(inode);
1106
1107         path = btrfs_alloc_path();
1108         if (!path)
1109                 return -ENOMEM;
1110 again:
1111         recow = 0;
1112         split = start;
1113         key.objectid = ino;
1114         key.type = BTRFS_EXTENT_DATA_KEY;
1115         key.offset = split;
1116
1117         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1118         if (ret < 0)
1119                 goto out;
1120         if (ret > 0 && path->slots[0] > 0)
1121                 path->slots[0]--;
1122
1123         leaf = path->nodes[0];
1124         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1125         if (key.objectid != ino ||
1126             key.type != BTRFS_EXTENT_DATA_KEY) {
1127                 ret = -EINVAL;
1128                 btrfs_abort_transaction(trans, ret);
1129                 goto out;
1130         }
1131         fi = btrfs_item_ptr(leaf, path->slots[0],
1132                             struct btrfs_file_extent_item);
1133         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_PREALLOC) {
1134                 ret = -EINVAL;
1135                 btrfs_abort_transaction(trans, ret);
1136                 goto out;
1137         }
1138         extent_end = key.offset + btrfs_file_extent_num_bytes(leaf, fi);
1139         if (key.offset > start || extent_end < end) {
1140                 ret = -EINVAL;
1141                 btrfs_abort_transaction(trans, ret);
1142                 goto out;
1143         }
1144
1145         bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
1146         num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
1147         orig_offset = key.offset - btrfs_file_extent_offset(leaf, fi);
1148         memcpy(&new_key, &key, sizeof(new_key));
1149
1150         if (start == key.offset && end < extent_end) {
1151                 other_start = 0;
1152                 other_end = start;
1153                 if (extent_mergeable(leaf, path->slots[0] - 1,
1154                                      ino, bytenr, orig_offset,
1155                                      &other_start, &other_end)) {
1156                         new_key.offset = end;
1157                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1158                         fi = btrfs_item_ptr(leaf, path->slots[0],
1159                                             struct btrfs_file_extent_item);
1160                         btrfs_set_file_extent_generation(leaf, fi,
1161                                                          trans->transid);
1162                         btrfs_set_file_extent_num_bytes(leaf, fi,
1163                                                         extent_end - end);
1164                         btrfs_set_file_extent_offset(leaf, fi,
1165                                                      end - orig_offset);
1166                         fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1167                                             struct btrfs_file_extent_item);
1168                         btrfs_set_file_extent_generation(leaf, fi,
1169                                                          trans->transid);
1170                         btrfs_set_file_extent_num_bytes(leaf, fi,
1171                                                         end - other_start);
1172                         btrfs_mark_buffer_dirty(leaf);
1173                         goto out;
1174                 }
1175         }
1176
1177         if (start > key.offset && end == extent_end) {
1178                 other_start = end;
1179                 other_end = 0;
1180                 if (extent_mergeable(leaf, path->slots[0] + 1,
1181                                      ino, bytenr, orig_offset,
1182                                      &other_start, &other_end)) {
1183                         fi = btrfs_item_ptr(leaf, path->slots[0],
1184                                             struct btrfs_file_extent_item);
1185                         btrfs_set_file_extent_num_bytes(leaf, fi,
1186                                                         start - key.offset);
1187                         btrfs_set_file_extent_generation(leaf, fi,
1188                                                          trans->transid);
1189                         path->slots[0]++;
1190                         new_key.offset = start;
1191                         btrfs_set_item_key_safe(fs_info, path, &new_key);
1192
1193                         fi = btrfs_item_ptr(leaf, path->slots[0],
1194                                             struct btrfs_file_extent_item);
1195                         btrfs_set_file_extent_generation(leaf, fi,
1196                                                          trans->transid);
1197                         btrfs_set_file_extent_num_bytes(leaf, fi,
1198                                                         other_end - start);
1199                         btrfs_set_file_extent_offset(leaf, fi,
1200                                                      start - orig_offset);
1201                         btrfs_mark_buffer_dirty(leaf);
1202                         goto out;
1203                 }
1204         }
1205
1206         while (start > key.offset || end < extent_end) {
1207                 if (key.offset == start)
1208                         split = end;
1209
1210                 new_key.offset = split;
1211                 ret = btrfs_duplicate_item(trans, root, path, &new_key);
1212                 if (ret == -EAGAIN) {
1213                         btrfs_release_path(path);
1214                         goto again;
1215                 }
1216                 if (ret < 0) {
1217                         btrfs_abort_transaction(trans, ret);
1218                         goto out;
1219                 }
1220
1221                 leaf = path->nodes[0];
1222                 fi = btrfs_item_ptr(leaf, path->slots[0] - 1,
1223                                     struct btrfs_file_extent_item);
1224                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1225                 btrfs_set_file_extent_num_bytes(leaf, fi,
1226                                                 split - key.offset);
1227
1228                 fi = btrfs_item_ptr(leaf, path->slots[0],
1229                                     struct btrfs_file_extent_item);
1230
1231                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1232                 btrfs_set_file_extent_offset(leaf, fi, split - orig_offset);
1233                 btrfs_set_file_extent_num_bytes(leaf, fi,
1234                                                 extent_end - split);
1235                 btrfs_mark_buffer_dirty(leaf);
1236
1237                 ret = btrfs_inc_extent_ref(trans, fs_info, bytenr, num_bytes,
1238                                            0, root->root_key.objectid,
1239                                            ino, orig_offset);
1240                 if (ret) {
1241                         btrfs_abort_transaction(trans, ret);
1242                         goto out;
1243                 }
1244
1245                 if (split == start) {
1246                         key.offset = start;
1247                 } else {
1248                         if (start != key.offset) {
1249                                 ret = -EINVAL;
1250                                 btrfs_abort_transaction(trans, ret);
1251                                 goto out;
1252                         }
1253                         path->slots[0]--;
1254                         extent_end = end;
1255                 }
1256                 recow = 1;
1257         }
1258
1259         other_start = end;
1260         other_end = 0;
1261         if (extent_mergeable(leaf, path->slots[0] + 1,
1262                              ino, bytenr, orig_offset,
1263                              &other_start, &other_end)) {
1264                 if (recow) {
1265                         btrfs_release_path(path);
1266                         goto again;
1267                 }
1268                 extent_end = other_end;
1269                 del_slot = path->slots[0] + 1;
1270                 del_nr++;
1271                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1272                                         0, root->root_key.objectid,
1273                                         ino, orig_offset);
1274                 if (ret) {
1275                         btrfs_abort_transaction(trans, ret);
1276                         goto out;
1277                 }
1278         }
1279         other_start = 0;
1280         other_end = start;
1281         if (extent_mergeable(leaf, path->slots[0] - 1,
1282                              ino, bytenr, orig_offset,
1283                              &other_start, &other_end)) {
1284                 if (recow) {
1285                         btrfs_release_path(path);
1286                         goto again;
1287                 }
1288                 key.offset = other_start;
1289                 del_slot = path->slots[0];
1290                 del_nr++;
1291                 ret = btrfs_free_extent(trans, fs_info, bytenr, num_bytes,
1292                                         0, root->root_key.objectid,
1293                                         ino, orig_offset);
1294                 if (ret) {
1295                         btrfs_abort_transaction(trans, ret);
1296                         goto out;
1297                 }
1298         }
1299         if (del_nr == 0) {
1300                 fi = btrfs_item_ptr(leaf, path->slots[0],
1301                            struct btrfs_file_extent_item);
1302                 btrfs_set_file_extent_type(leaf, fi,
1303                                            BTRFS_FILE_EXTENT_REG);
1304                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1305                 btrfs_mark_buffer_dirty(leaf);
1306         } else {
1307                 fi = btrfs_item_ptr(leaf, del_slot - 1,
1308                            struct btrfs_file_extent_item);
1309                 btrfs_set_file_extent_type(leaf, fi,
1310                                            BTRFS_FILE_EXTENT_REG);
1311                 btrfs_set_file_extent_generation(leaf, fi, trans->transid);
1312                 btrfs_set_file_extent_num_bytes(leaf, fi,
1313                                                 extent_end - key.offset);
1314                 btrfs_mark_buffer_dirty(leaf);
1315
1316                 ret = btrfs_del_items(trans, root, path, del_slot, del_nr);
1317                 if (ret < 0) {
1318                         btrfs_abort_transaction(trans, ret);
1319                         goto out;
1320                 }
1321         }
1322 out:
1323         btrfs_free_path(path);
1324         return 0;
1325 }
1326
1327 /*
1328  * on error we return an unlocked page and the error value
1329  * on success we return a locked page and 0
1330  */
1331 static int prepare_uptodate_page(struct inode *inode,
1332                                  struct page *page, u64 pos,
1333                                  bool force_uptodate)
1334 {
1335         int ret = 0;
1336
1337         if (((pos & (PAGE_SIZE - 1)) || force_uptodate) &&
1338             !PageUptodate(page)) {
1339                 ret = btrfs_readpage(NULL, page);
1340                 if (ret)
1341                         return ret;
1342                 lock_page(page);
1343                 if (!PageUptodate(page)) {
1344                         unlock_page(page);
1345                         return -EIO;
1346                 }
1347                 if (page->mapping != inode->i_mapping) {
1348                         unlock_page(page);
1349                         return -EAGAIN;
1350                 }
1351         }
1352         return 0;
1353 }
1354
1355 /*
1356  * this just gets pages into the page cache and locks them down.
1357  */
1358 static noinline int prepare_pages(struct inode *inode, struct page **pages,
1359                                   size_t num_pages, loff_t pos,
1360                                   size_t write_bytes, bool force_uptodate)
1361 {
1362         int i;
1363         unsigned long index = pos >> PAGE_SHIFT;
1364         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
1365         int err = 0;
1366         int faili;
1367
1368         for (i = 0; i < num_pages; i++) {
1369 again:
1370                 pages[i] = find_or_create_page(inode->i_mapping, index + i,
1371                                                mask | __GFP_WRITE);
1372                 if (!pages[i]) {
1373                         faili = i - 1;
1374                         err = -ENOMEM;
1375                         goto fail;
1376                 }
1377
1378                 if (i == 0)
1379                         err = prepare_uptodate_page(inode, pages[i], pos,
1380                                                     force_uptodate);
1381                 if (!err && i == num_pages - 1)
1382                         err = prepare_uptodate_page(inode, pages[i],
1383                                                     pos + write_bytes, false);
1384                 if (err) {
1385                         put_page(pages[i]);
1386                         if (err == -EAGAIN) {
1387                                 err = 0;
1388                                 goto again;
1389                         }
1390                         faili = i - 1;
1391                         goto fail;
1392                 }
1393                 wait_on_page_writeback(pages[i]);
1394         }
1395
1396         return 0;
1397 fail:
1398         while (faili >= 0) {
1399                 unlock_page(pages[faili]);
1400                 put_page(pages[faili]);
1401                 faili--;
1402         }
1403         return err;
1404
1405 }
1406
1407 static int btrfs_find_new_delalloc_bytes(struct btrfs_inode *inode,
1408                                          const u64 start,
1409                                          const u64 len,
1410                                          struct extent_state **cached_state)
1411 {
1412         u64 search_start = start;
1413         const u64 end = start + len - 1;
1414
1415         while (search_start < end) {
1416                 const u64 search_len = end - search_start + 1;
1417                 struct extent_map *em;
1418                 u64 em_len;
1419                 int ret = 0;
1420
1421                 em = btrfs_get_extent(inode, NULL, 0, search_start,
1422                                       search_len, 0);
1423                 if (IS_ERR(em))
1424                         return PTR_ERR(em);
1425
1426                 if (em->block_start != EXTENT_MAP_HOLE)
1427                         goto next;
1428
1429                 em_len = em->len;
1430                 if (em->start < search_start)
1431                         em_len -= search_start - em->start;
1432                 if (em_len > search_len)
1433                         em_len = search_len;
1434
1435                 ret = set_extent_bit(&inode->io_tree, search_start,
1436                                      search_start + em_len - 1,
1437                                      EXTENT_DELALLOC_NEW,
1438                                      NULL, cached_state, GFP_NOFS);
1439 next:
1440                 search_start = extent_map_end(em);
1441                 free_extent_map(em);
1442                 if (ret)
1443                         return ret;
1444         }
1445         return 0;
1446 }
1447
1448 /*
1449  * This function locks the extent and properly waits for data=ordered extents
1450  * to finish before allowing the pages to be modified if need.
1451  *
1452  * The return value:
1453  * 1 - the extent is locked
1454  * 0 - the extent is not locked, and everything is OK
1455  * -EAGAIN - need re-prepare the pages
1456  * the other < 0 number - Something wrong happens
1457  */
1458 static noinline int
1459 lock_and_cleanup_extent_if_need(struct btrfs_inode *inode, struct page **pages,
1460                                 size_t num_pages, loff_t pos,
1461                                 size_t write_bytes,
1462                                 u64 *lockstart, u64 *lockend,
1463                                 struct extent_state **cached_state)
1464 {
1465         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1466         u64 start_pos;
1467         u64 last_pos;
1468         int i;
1469         int ret = 0;
1470
1471         start_pos = round_down(pos, fs_info->sectorsize);
1472         last_pos = start_pos
1473                 + round_up(pos + write_bytes - start_pos,
1474                            fs_info->sectorsize) - 1;
1475
1476         if (start_pos < inode->vfs_inode.i_size ||
1477             (inode->flags & BTRFS_INODE_PREALLOC)) {
1478                 struct btrfs_ordered_extent *ordered;
1479                 unsigned int clear_bits;
1480
1481                 lock_extent_bits(&inode->io_tree, start_pos, last_pos,
1482                                 cached_state);
1483                 ordered = btrfs_lookup_ordered_range(inode, start_pos,
1484                                                      last_pos - start_pos + 1);
1485                 if (ordered &&
1486                     ordered->file_offset + ordered->len > start_pos &&
1487                     ordered->file_offset <= last_pos) {
1488                         unlock_extent_cached(&inode->io_tree, start_pos,
1489                                         last_pos, cached_state, GFP_NOFS);
1490                         for (i = 0; i < num_pages; i++) {
1491                                 unlock_page(pages[i]);
1492                                 put_page(pages[i]);
1493                         }
1494                         btrfs_start_ordered_extent(&inode->vfs_inode,
1495                                         ordered, 1);
1496                         btrfs_put_ordered_extent(ordered);
1497                         return -EAGAIN;
1498                 }
1499                 if (ordered)
1500                         btrfs_put_ordered_extent(ordered);
1501                 ret = btrfs_find_new_delalloc_bytes(inode, start_pos,
1502                                                     last_pos - start_pos + 1,
1503                                                     cached_state);
1504                 clear_bits = EXTENT_DIRTY | EXTENT_DELALLOC |
1505                         EXTENT_DO_ACCOUNTING | EXTENT_DEFRAG;
1506                 if (ret)
1507                         clear_bits |= EXTENT_DELALLOC_NEW | EXTENT_LOCKED;
1508                 clear_extent_bit(&inode->io_tree, start_pos,
1509                                  last_pos, clear_bits,
1510                                  (clear_bits & EXTENT_LOCKED) ? 1 : 0,
1511                                  0, cached_state, GFP_NOFS);
1512                 if (ret)
1513                         return ret;
1514                 *lockstart = start_pos;
1515                 *lockend = last_pos;
1516                 ret = 1;
1517         }
1518
1519         for (i = 0; i < num_pages; i++) {
1520                 if (clear_page_dirty_for_io(pages[i]))
1521                         account_page_redirty(pages[i]);
1522                 set_page_extent_mapped(pages[i]);
1523                 WARN_ON(!PageLocked(pages[i]));
1524         }
1525
1526         return ret;
1527 }
1528
1529 static noinline int check_can_nocow(struct btrfs_inode *inode, loff_t pos,
1530                                     size_t *write_bytes)
1531 {
1532         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
1533         struct btrfs_root *root = inode->root;
1534         struct btrfs_ordered_extent *ordered;
1535         u64 lockstart, lockend;
1536         u64 num_bytes;
1537         int ret;
1538
1539         ret = btrfs_start_write_no_snapshoting(root);
1540         if (!ret)
1541                 return -ENOSPC;
1542
1543         lockstart = round_down(pos, fs_info->sectorsize);
1544         lockend = round_up(pos + *write_bytes,
1545                            fs_info->sectorsize) - 1;
1546
1547         while (1) {
1548                 lock_extent(&inode->io_tree, lockstart, lockend);
1549                 ordered = btrfs_lookup_ordered_range(inode, lockstart,
1550                                                      lockend - lockstart + 1);
1551                 if (!ordered) {
1552                         break;
1553                 }
1554                 unlock_extent(&inode->io_tree, lockstart, lockend);
1555                 btrfs_start_ordered_extent(&inode->vfs_inode, ordered, 1);
1556                 btrfs_put_ordered_extent(ordered);
1557         }
1558
1559         num_bytes = lockend - lockstart + 1;
1560         ret = can_nocow_extent(&inode->vfs_inode, lockstart, &num_bytes,
1561                         NULL, NULL, NULL);
1562         if (ret <= 0) {
1563                 ret = 0;
1564                 btrfs_end_write_no_snapshoting(root);
1565         } else {
1566                 *write_bytes = min_t(size_t, *write_bytes ,
1567                                      num_bytes - pos + lockstart);
1568         }
1569
1570         unlock_extent(&inode->io_tree, lockstart, lockend);
1571
1572         return ret;
1573 }
1574
1575 static noinline ssize_t __btrfs_buffered_write(struct file *file,
1576                                                struct iov_iter *i,
1577                                                loff_t pos)
1578 {
1579         struct inode *inode = file_inode(file);
1580         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1581         struct btrfs_root *root = BTRFS_I(inode)->root;
1582         struct page **pages = NULL;
1583         struct extent_state *cached_state = NULL;
1584         u64 release_bytes = 0;
1585         u64 lockstart;
1586         u64 lockend;
1587         size_t num_written = 0;
1588         int nrptrs;
1589         int ret = 0;
1590         bool only_release_metadata = false;
1591         bool force_page_uptodate = false;
1592         bool need_unlock;
1593
1594         nrptrs = min(DIV_ROUND_UP(iov_iter_count(i), PAGE_SIZE),
1595                         PAGE_SIZE / (sizeof(struct page *)));
1596         nrptrs = min(nrptrs, current->nr_dirtied_pause - current->nr_dirtied);
1597         nrptrs = max(nrptrs, 8);
1598         pages = kmalloc_array(nrptrs, sizeof(struct page *), GFP_KERNEL);
1599         if (!pages)
1600                 return -ENOMEM;
1601
1602         while (iov_iter_count(i) > 0) {
1603                 size_t offset = pos & (PAGE_SIZE - 1);
1604                 size_t sector_offset;
1605                 size_t write_bytes = min(iov_iter_count(i),
1606                                          nrptrs * (size_t)PAGE_SIZE -
1607                                          offset);
1608                 size_t num_pages = DIV_ROUND_UP(write_bytes + offset,
1609                                                 PAGE_SIZE);
1610                 size_t reserve_bytes;
1611                 size_t dirty_pages;
1612                 size_t copied;
1613                 size_t dirty_sectors;
1614                 size_t num_sectors;
1615
1616                 WARN_ON(num_pages > nrptrs);
1617
1618                 /*
1619                  * Fault pages before locking them in prepare_pages
1620                  * to avoid recursive lock
1621                  */
1622                 if (unlikely(iov_iter_fault_in_readable(i, write_bytes))) {
1623                         ret = -EFAULT;
1624                         break;
1625                 }
1626
1627                 sector_offset = pos & (fs_info->sectorsize - 1);
1628                 reserve_bytes = round_up(write_bytes + sector_offset,
1629                                 fs_info->sectorsize);
1630
1631                 ret = btrfs_check_data_free_space(inode, pos, write_bytes);
1632                 if (ret < 0) {
1633                         if ((BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1634                                                       BTRFS_INODE_PREALLOC)) &&
1635                             check_can_nocow(BTRFS_I(inode), pos,
1636                                         &write_bytes) > 0) {
1637                                 /*
1638                                  * For nodata cow case, no need to reserve
1639                                  * data space.
1640                                  */
1641                                 only_release_metadata = true;
1642                                 /*
1643                                  * our prealloc extent may be smaller than
1644                                  * write_bytes, so scale down.
1645                                  */
1646                                 num_pages = DIV_ROUND_UP(write_bytes + offset,
1647                                                          PAGE_SIZE);
1648                                 reserve_bytes = round_up(write_bytes +
1649                                                          sector_offset,
1650                                                          fs_info->sectorsize);
1651                         } else {
1652                                 break;
1653                         }
1654                 }
1655
1656                 ret = btrfs_delalloc_reserve_metadata(BTRFS_I(inode),
1657                                 reserve_bytes);
1658                 if (ret) {
1659                         if (!only_release_metadata)
1660                                 btrfs_free_reserved_data_space(inode, pos,
1661                                                                write_bytes);
1662                         else
1663                                 btrfs_end_write_no_snapshoting(root);
1664                         break;
1665                 }
1666
1667                 release_bytes = reserve_bytes;
1668                 need_unlock = false;
1669 again:
1670                 /*
1671                  * This is going to setup the pages array with the number of
1672                  * pages we want, so we don't really need to worry about the
1673                  * contents of pages from loop to loop
1674                  */
1675                 ret = prepare_pages(inode, pages, num_pages,
1676                                     pos, write_bytes,
1677                                     force_page_uptodate);
1678                 if (ret)
1679                         break;
1680
1681                 ret = lock_and_cleanup_extent_if_need(BTRFS_I(inode), pages,
1682                                 num_pages, pos, write_bytes, &lockstart,
1683                                 &lockend, &cached_state);
1684                 if (ret < 0) {
1685                         if (ret == -EAGAIN)
1686                                 goto again;
1687                         break;
1688                 } else if (ret > 0) {
1689                         need_unlock = true;
1690                         ret = 0;
1691                 }
1692
1693                 copied = btrfs_copy_from_user(pos, write_bytes, pages, i);
1694
1695                 num_sectors = BTRFS_BYTES_TO_BLKS(fs_info, reserve_bytes);
1696                 dirty_sectors = round_up(copied + sector_offset,
1697                                         fs_info->sectorsize);
1698                 dirty_sectors = BTRFS_BYTES_TO_BLKS(fs_info, dirty_sectors);
1699
1700                 /*
1701                  * if we have trouble faulting in the pages, fall
1702                  * back to one page at a time
1703                  */
1704                 if (copied < write_bytes)
1705                         nrptrs = 1;
1706
1707                 if (copied == 0) {
1708                         force_page_uptodate = true;
1709                         dirty_sectors = 0;
1710                         dirty_pages = 0;
1711                 } else {
1712                         force_page_uptodate = false;
1713                         dirty_pages = DIV_ROUND_UP(copied + offset,
1714                                                    PAGE_SIZE);
1715                 }
1716
1717                 /*
1718                  * If we had a short copy we need to release the excess delaloc
1719                  * bytes we reserved.  We need to increment outstanding_extents
1720                  * because btrfs_delalloc_release_space and
1721                  * btrfs_delalloc_release_metadata will decrement it, but
1722                  * we still have an outstanding extent for the chunk we actually
1723                  * managed to copy.
1724                  */
1725                 if (num_sectors > dirty_sectors) {
1726                         /* release everything except the sectors we dirtied */
1727                         release_bytes -= dirty_sectors <<
1728                                                 fs_info->sb->s_blocksize_bits;
1729                         if (copied > 0) {
1730                                 spin_lock(&BTRFS_I(inode)->lock);
1731                                 BTRFS_I(inode)->outstanding_extents++;
1732                                 spin_unlock(&BTRFS_I(inode)->lock);
1733                         }
1734                         if (only_release_metadata) {
1735                                 btrfs_delalloc_release_metadata(BTRFS_I(inode),
1736                                                                 release_bytes);
1737                         } else {
1738                                 u64 __pos;
1739
1740                                 __pos = round_down(pos,
1741                                                    fs_info->sectorsize) +
1742                                         (dirty_pages << PAGE_SHIFT);
1743                                 btrfs_delalloc_release_space(inode, __pos,
1744                                                              release_bytes);
1745                         }
1746                 }
1747
1748                 release_bytes = round_up(copied + sector_offset,
1749                                         fs_info->sectorsize);
1750
1751                 if (copied > 0)
1752                         ret = btrfs_dirty_pages(inode, pages, dirty_pages,
1753                                                 pos, copied, NULL);
1754                 if (need_unlock)
1755                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
1756                                              lockstart, lockend, &cached_state,
1757                                              GFP_NOFS);
1758                 if (ret) {
1759                         btrfs_drop_pages(pages, num_pages);
1760                         break;
1761                 }
1762
1763                 release_bytes = 0;
1764                 if (only_release_metadata)
1765                         btrfs_end_write_no_snapshoting(root);
1766
1767                 if (only_release_metadata && copied > 0) {
1768                         lockstart = round_down(pos,
1769                                                fs_info->sectorsize);
1770                         lockend = round_up(pos + copied,
1771                                            fs_info->sectorsize) - 1;
1772
1773                         set_extent_bit(&BTRFS_I(inode)->io_tree, lockstart,
1774                                        lockend, EXTENT_NORESERVE, NULL,
1775                                        NULL, GFP_NOFS);
1776                         only_release_metadata = false;
1777                 }
1778
1779                 btrfs_drop_pages(pages, num_pages);
1780
1781                 cond_resched();
1782
1783                 balance_dirty_pages_ratelimited(inode->i_mapping);
1784                 if (dirty_pages < (fs_info->nodesize >> PAGE_SHIFT) + 1)
1785                         btrfs_btree_balance_dirty(fs_info);
1786
1787                 pos += copied;
1788                 num_written += copied;
1789         }
1790
1791         kfree(pages);
1792
1793         if (release_bytes) {
1794                 if (only_release_metadata) {
1795                         btrfs_end_write_no_snapshoting(root);
1796                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
1797                                         release_bytes);
1798                 } else {
1799                         btrfs_delalloc_release_space(inode,
1800                                                 round_down(pos, fs_info->sectorsize),
1801                                                 release_bytes);
1802                 }
1803         }
1804
1805         return num_written ? num_written : ret;
1806 }
1807
1808 static ssize_t __btrfs_direct_write(struct kiocb *iocb, struct iov_iter *from)
1809 {
1810         struct file *file = iocb->ki_filp;
1811         struct inode *inode = file_inode(file);
1812         loff_t pos = iocb->ki_pos;
1813         ssize_t written;
1814         ssize_t written_buffered;
1815         loff_t endbyte;
1816         int err;
1817
1818         written = generic_file_direct_write(iocb, from);
1819
1820         if (written < 0 || !iov_iter_count(from))
1821                 return written;
1822
1823         pos += written;
1824         written_buffered = __btrfs_buffered_write(file, from, pos);
1825         if (written_buffered < 0) {
1826                 err = written_buffered;
1827                 goto out;
1828         }
1829         /*
1830          * Ensure all data is persisted. We want the next direct IO read to be
1831          * able to read what was just written.
1832          */
1833         endbyte = pos + written_buffered - 1;
1834         err = btrfs_fdatawrite_range(inode, pos, endbyte);
1835         if (err)
1836                 goto out;
1837         err = filemap_fdatawait_range(inode->i_mapping, pos, endbyte);
1838         if (err)
1839                 goto out;
1840         written += written_buffered;
1841         iocb->ki_pos = pos + written_buffered;
1842         invalidate_mapping_pages(file->f_mapping, pos >> PAGE_SHIFT,
1843                                  endbyte >> PAGE_SHIFT);
1844 out:
1845         return written ? written : err;
1846 }
1847
1848 static void update_time_for_write(struct inode *inode)
1849 {
1850         struct timespec now;
1851
1852         if (IS_NOCMTIME(inode))
1853                 return;
1854
1855         now = current_time(inode);
1856         if (!timespec_equal(&inode->i_mtime, &now))
1857                 inode->i_mtime = now;
1858
1859         if (!timespec_equal(&inode->i_ctime, &now))
1860                 inode->i_ctime = now;
1861
1862         if (IS_I_VERSION(inode))
1863                 inode_inc_iversion(inode);
1864 }
1865
1866 static ssize_t btrfs_file_write_iter(struct kiocb *iocb,
1867                                     struct iov_iter *from)
1868 {
1869         struct file *file = iocb->ki_filp;
1870         struct inode *inode = file_inode(file);
1871         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
1872         struct btrfs_root *root = BTRFS_I(inode)->root;
1873         u64 start_pos;
1874         u64 end_pos;
1875         ssize_t num_written = 0;
1876         bool sync = (file->f_flags & O_DSYNC) || IS_SYNC(file->f_mapping->host);
1877         ssize_t err;
1878         loff_t pos = iocb->ki_pos;
1879         size_t count = iov_iter_count(from);
1880         loff_t oldsize;
1881         int clean_page = 0;
1882
1883         if ((iocb->ki_flags & IOCB_NOWAIT) &&
1884                         (iocb->ki_flags & IOCB_DIRECT)) {
1885                 /* Don't sleep on inode rwsem */
1886                 if (!inode_trylock(inode))
1887                         return -EAGAIN;
1888                 /*
1889                  * We will allocate space in case nodatacow is not set,
1890                  * so bail
1891                  */
1892                 if (!(BTRFS_I(inode)->flags & (BTRFS_INODE_NODATACOW |
1893                                               BTRFS_INODE_PREALLOC)) ||
1894                     check_can_nocow(BTRFS_I(inode), pos, &count) <= 0) {
1895                         inode_unlock(inode);
1896                         return -EAGAIN;
1897                 }
1898         } else
1899                 inode_lock(inode);
1900
1901         err = generic_write_checks(iocb, from);
1902         if (err <= 0) {
1903                 inode_unlock(inode);
1904                 return err;
1905         }
1906
1907         current->backing_dev_info = inode_to_bdi(inode);
1908         err = file_remove_privs(file);
1909         if (err) {
1910                 inode_unlock(inode);
1911                 goto out;
1912         }
1913
1914         /*
1915          * If BTRFS flips readonly due to some impossible error
1916          * (fs_info->fs_state now has BTRFS_SUPER_FLAG_ERROR),
1917          * although we have opened a file as writable, we have
1918          * to stop this write operation to ensure FS consistency.
1919          */
1920         if (test_bit(BTRFS_FS_STATE_ERROR, &fs_info->fs_state)) {
1921                 inode_unlock(inode);
1922                 err = -EROFS;
1923                 goto out;
1924         }
1925
1926         /*
1927          * We reserve space for updating the inode when we reserve space for the
1928          * extent we are going to write, so we will enospc out there.  We don't
1929          * need to start yet another transaction to update the inode as we will
1930          * update the inode when we finish writing whatever data we write.
1931          */
1932         update_time_for_write(inode);
1933
1934         start_pos = round_down(pos, fs_info->sectorsize);
1935         oldsize = i_size_read(inode);
1936         if (start_pos > oldsize) {
1937                 /* Expand hole size to cover write data, preventing empty gap */
1938                 end_pos = round_up(pos + count,
1939                                    fs_info->sectorsize);
1940                 err = btrfs_cont_expand(inode, oldsize, end_pos);
1941                 if (err) {
1942                         inode_unlock(inode);
1943                         goto out;
1944                 }
1945                 if (start_pos > round_up(oldsize, fs_info->sectorsize))
1946                         clean_page = 1;
1947         }
1948
1949         if (sync)
1950                 atomic_inc(&BTRFS_I(inode)->sync_writers);
1951
1952         if (iocb->ki_flags & IOCB_DIRECT) {
1953                 num_written = __btrfs_direct_write(iocb, from);
1954         } else {
1955                 num_written = __btrfs_buffered_write(file, from, pos);
1956                 if (num_written > 0)
1957                         iocb->ki_pos = pos + num_written;
1958                 if (clean_page)
1959                         pagecache_isize_extended(inode, oldsize,
1960                                                 i_size_read(inode));
1961         }
1962
1963         inode_unlock(inode);
1964
1965         /*
1966          * We also have to set last_sub_trans to the current log transid,
1967          * otherwise subsequent syncs to a file that's been synced in this
1968          * transaction will appear to have already occurred.
1969          */
1970         spin_lock(&BTRFS_I(inode)->lock);
1971         BTRFS_I(inode)->last_sub_trans = root->log_transid;
1972         spin_unlock(&BTRFS_I(inode)->lock);
1973         if (num_written > 0)
1974                 num_written = generic_write_sync(iocb, num_written);
1975
1976         if (sync)
1977                 atomic_dec(&BTRFS_I(inode)->sync_writers);
1978 out:
1979         current->backing_dev_info = NULL;
1980         return num_written ? num_written : err;
1981 }
1982
1983 int btrfs_release_file(struct inode *inode, struct file *filp)
1984 {
1985         if (filp->private_data)
1986                 btrfs_ioctl_trans_end(filp);
1987         /*
1988          * ordered_data_close is set by settattr when we are about to truncate
1989          * a file from a non-zero size to a zero size.  This tries to
1990          * flush down new bytes that may have been written if the
1991          * application were using truncate to replace a file in place.
1992          */
1993         if (test_and_clear_bit(BTRFS_INODE_ORDERED_DATA_CLOSE,
1994                                &BTRFS_I(inode)->runtime_flags))
1995                         filemap_flush(inode->i_mapping);
1996         return 0;
1997 }
1998
1999 static int start_ordered_ops(struct inode *inode, loff_t start, loff_t end)
2000 {
2001         int ret;
2002
2003         atomic_inc(&BTRFS_I(inode)->sync_writers);
2004         ret = btrfs_fdatawrite_range(inode, start, end);
2005         atomic_dec(&BTRFS_I(inode)->sync_writers);
2006
2007         return ret;
2008 }
2009
2010 /*
2011  * fsync call for both files and directories.  This logs the inode into
2012  * the tree log instead of forcing full commits whenever possible.
2013  *
2014  * It needs to call filemap_fdatawait so that all ordered extent updates are
2015  * in the metadata btree are up to date for copying to the log.
2016  *
2017  * It drops the inode mutex before doing the tree log commit.  This is an
2018  * important optimization for directories because holding the mutex prevents
2019  * new operations on the dir while we write to disk.
2020  */
2021 int btrfs_sync_file(struct file *file, loff_t start, loff_t end, int datasync)
2022 {
2023         struct dentry *dentry = file_dentry(file);
2024         struct inode *inode = d_inode(dentry);
2025         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2026         struct btrfs_root *root = BTRFS_I(inode)->root;
2027         struct btrfs_trans_handle *trans;
2028         struct btrfs_log_ctx ctx;
2029         int ret = 0;
2030         bool full_sync = 0;
2031         u64 len;
2032
2033         /*
2034          * The range length can be represented by u64, we have to do the typecasts
2035          * to avoid signed overflow if it's [0, LLONG_MAX] eg. from fsync()
2036          */
2037         len = (u64)end - (u64)start + 1;
2038         trace_btrfs_sync_file(file, datasync);
2039
2040         /*
2041          * We write the dirty pages in the range and wait until they complete
2042          * out of the ->i_mutex. If so, we can flush the dirty pages by
2043          * multi-task, and make the performance up.  See
2044          * btrfs_wait_ordered_range for an explanation of the ASYNC check.
2045          */
2046         ret = start_ordered_ops(inode, start, end);
2047         if (ret)
2048                 return ret;
2049
2050         inode_lock(inode);
2051         atomic_inc(&root->log_batch);
2052         full_sync = test_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2053                              &BTRFS_I(inode)->runtime_flags);
2054         /*
2055          * We might have have had more pages made dirty after calling
2056          * start_ordered_ops and before acquiring the inode's i_mutex.
2057          */
2058         if (full_sync) {
2059                 /*
2060                  * For a full sync, we need to make sure any ordered operations
2061                  * start and finish before we start logging the inode, so that
2062                  * all extents are persisted and the respective file extent
2063                  * items are in the fs/subvol btree.
2064                  */
2065                 ret = btrfs_wait_ordered_range(inode, start, len);
2066         } else {
2067                 /*
2068                  * Start any new ordered operations before starting to log the
2069                  * inode. We will wait for them to finish in btrfs_sync_log().
2070                  *
2071                  * Right before acquiring the inode's mutex, we might have new
2072                  * writes dirtying pages, which won't immediately start the
2073                  * respective ordered operations - that is done through the
2074                  * fill_delalloc callbacks invoked from the writepage and
2075                  * writepages address space operations. So make sure we start
2076                  * all ordered operations before starting to log our inode. Not
2077                  * doing this means that while logging the inode, writeback
2078                  * could start and invoke writepage/writepages, which would call
2079                  * the fill_delalloc callbacks (cow_file_range,
2080                  * submit_compressed_extents). These callbacks add first an
2081                  * extent map to the modified list of extents and then create
2082                  * the respective ordered operation, which means in
2083                  * tree-log.c:btrfs_log_inode() we might capture all existing
2084                  * ordered operations (with btrfs_get_logged_extents()) before
2085                  * the fill_delalloc callback adds its ordered operation, and by
2086                  * the time we visit the modified list of extent maps (with
2087                  * btrfs_log_changed_extents()), we see and process the extent
2088                  * map they created. We then use the extent map to construct a
2089                  * file extent item for logging without waiting for the
2090                  * respective ordered operation to finish - this file extent
2091                  * item points to a disk location that might not have yet been
2092                  * written to, containing random data - so after a crash a log
2093                  * replay will make our inode have file extent items that point
2094                  * to disk locations containing invalid data, as we returned
2095                  * success to userspace without waiting for the respective
2096                  * ordered operation to finish, because it wasn't captured by
2097                  * btrfs_get_logged_extents().
2098                  */
2099                 ret = start_ordered_ops(inode, start, end);
2100         }
2101         if (ret) {
2102                 inode_unlock(inode);
2103                 goto out;
2104         }
2105         atomic_inc(&root->log_batch);
2106
2107         /*
2108          * If the last transaction that changed this file was before the current
2109          * transaction and we have the full sync flag set in our inode, we can
2110          * bail out now without any syncing.
2111          *
2112          * Note that we can't bail out if the full sync flag isn't set. This is
2113          * because when the full sync flag is set we start all ordered extents
2114          * and wait for them to fully complete - when they complete they update
2115          * the inode's last_trans field through:
2116          *
2117          *     btrfs_finish_ordered_io() ->
2118          *         btrfs_update_inode_fallback() ->
2119          *             btrfs_update_inode() ->
2120          *                 btrfs_set_inode_last_trans()
2121          *
2122          * So we are sure that last_trans is up to date and can do this check to
2123          * bail out safely. For the fast path, when the full sync flag is not
2124          * set in our inode, we can not do it because we start only our ordered
2125          * extents and don't wait for them to complete (that is when
2126          * btrfs_finish_ordered_io runs), so here at this point their last_trans
2127          * value might be less than or equals to fs_info->last_trans_committed,
2128          * and setting a speculative last_trans for an inode when a buffered
2129          * write is made (such as fs_info->generation + 1 for example) would not
2130          * be reliable since after setting the value and before fsync is called
2131          * any number of transactions can start and commit (transaction kthread
2132          * commits the current transaction periodically), and a transaction
2133          * commit does not start nor waits for ordered extents to complete.
2134          */
2135         smp_mb();
2136         if (btrfs_inode_in_log(BTRFS_I(inode), fs_info->generation) ||
2137             (full_sync && BTRFS_I(inode)->last_trans <=
2138              fs_info->last_trans_committed) ||
2139             (!btrfs_have_ordered_extents_in_range(inode, start, len) &&
2140              BTRFS_I(inode)->last_trans
2141              <= fs_info->last_trans_committed)) {
2142                 /*
2143                  * We've had everything committed since the last time we were
2144                  * modified so clear this flag in case it was set for whatever
2145                  * reason, it's no longer relevant.
2146                  */
2147                 clear_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2148                           &BTRFS_I(inode)->runtime_flags);
2149                 /*
2150                  * An ordered extent might have started before and completed
2151                  * already with io errors, in which case the inode was not
2152                  * updated and we end up here. So check the inode's mapping
2153                  * flags for any errors that might have happened while doing
2154                  * writeback of file data.
2155                  */
2156                 ret = filemap_check_errors(inode->i_mapping);
2157                 inode_unlock(inode);
2158                 goto out;
2159         }
2160
2161         /*
2162          * ok we haven't committed the transaction yet, lets do a commit
2163          */
2164         if (file->private_data)
2165                 btrfs_ioctl_trans_end(file);
2166
2167         /*
2168          * We use start here because we will need to wait on the IO to complete
2169          * in btrfs_sync_log, which could require joining a transaction (for
2170          * example checking cross references in the nocow path).  If we use join
2171          * here we could get into a situation where we're waiting on IO to
2172          * happen that is blocked on a transaction trying to commit.  With start
2173          * we inc the extwriter counter, so we wait for all extwriters to exit
2174          * before we start blocking join'ers.  This comment is to keep somebody
2175          * from thinking they are super smart and changing this to
2176          * btrfs_join_transaction *cough*Josef*cough*.
2177          */
2178         trans = btrfs_start_transaction(root, 0);
2179         if (IS_ERR(trans)) {
2180                 ret = PTR_ERR(trans);
2181                 inode_unlock(inode);
2182                 goto out;
2183         }
2184         trans->sync = true;
2185
2186         btrfs_init_log_ctx(&ctx, inode);
2187
2188         ret = btrfs_log_dentry_safe(trans, root, dentry, start, end, &ctx);
2189         if (ret < 0) {
2190                 /* Fallthrough and commit/free transaction. */
2191                 ret = 1;
2192         }
2193
2194         /* we've logged all the items and now have a consistent
2195          * version of the file in the log.  It is possible that
2196          * someone will come in and modify the file, but that's
2197          * fine because the log is consistent on disk, and we
2198          * have references to all of the file's extents
2199          *
2200          * It is possible that someone will come in and log the
2201          * file again, but that will end up using the synchronization
2202          * inside btrfs_sync_log to keep things safe.
2203          */
2204         inode_unlock(inode);
2205
2206         /*
2207          * If any of the ordered extents had an error, just return it to user
2208          * space, so that the application knows some writes didn't succeed and
2209          * can take proper action (retry for e.g.). Blindly committing the
2210          * transaction in this case, would fool userspace that everything was
2211          * successful. And we also want to make sure our log doesn't contain
2212          * file extent items pointing to extents that weren't fully written to -
2213          * just like in the non fast fsync path, where we check for the ordered
2214          * operation's error flag before writing to the log tree and return -EIO
2215          * if any of them had this flag set (btrfs_wait_ordered_range) -
2216          * therefore we need to check for errors in the ordered operations,
2217          * which are indicated by ctx.io_err.
2218          */
2219         if (ctx.io_err) {
2220                 btrfs_end_transaction(trans);
2221                 ret = ctx.io_err;
2222                 goto out;
2223         }
2224
2225         if (ret != BTRFS_NO_LOG_SYNC) {
2226                 if (!ret) {
2227                         ret = btrfs_sync_log(trans, root, &ctx);
2228                         if (!ret) {
2229                                 ret = btrfs_end_transaction(trans);
2230                                 goto out;
2231                         }
2232                 }
2233                 if (!full_sync) {
2234                         ret = btrfs_wait_ordered_range(inode, start, len);
2235                         if (ret) {
2236                                 btrfs_end_transaction(trans);
2237                                 goto out;
2238                         }
2239                 }
2240                 ret = btrfs_commit_transaction(trans);
2241         } else {
2242                 ret = btrfs_end_transaction(trans);
2243         }
2244 out:
2245         return ret > 0 ? -EIO : ret;
2246 }
2247
2248 static const struct vm_operations_struct btrfs_file_vm_ops = {
2249         .fault          = filemap_fault,
2250         .map_pages      = filemap_map_pages,
2251         .page_mkwrite   = btrfs_page_mkwrite,
2252 };
2253
2254 static int btrfs_file_mmap(struct file  *filp, struct vm_area_struct *vma)
2255 {
2256         struct address_space *mapping = filp->f_mapping;
2257
2258         if (!mapping->a_ops->readpage)
2259                 return -ENOEXEC;
2260
2261         file_accessed(filp);
2262         vma->vm_ops = &btrfs_file_vm_ops;
2263
2264         return 0;
2265 }
2266
2267 static int hole_mergeable(struct btrfs_inode *inode, struct extent_buffer *leaf,
2268                           int slot, u64 start, u64 end)
2269 {
2270         struct btrfs_file_extent_item *fi;
2271         struct btrfs_key key;
2272
2273         if (slot < 0 || slot >= btrfs_header_nritems(leaf))
2274                 return 0;
2275
2276         btrfs_item_key_to_cpu(leaf, &key, slot);
2277         if (key.objectid != btrfs_ino(inode) ||
2278             key.type != BTRFS_EXTENT_DATA_KEY)
2279                 return 0;
2280
2281         fi = btrfs_item_ptr(leaf, slot, struct btrfs_file_extent_item);
2282
2283         if (btrfs_file_extent_type(leaf, fi) != BTRFS_FILE_EXTENT_REG)
2284                 return 0;
2285
2286         if (btrfs_file_extent_disk_bytenr(leaf, fi))
2287                 return 0;
2288
2289         if (key.offset == end)
2290                 return 1;
2291         if (key.offset + btrfs_file_extent_num_bytes(leaf, fi) == start)
2292                 return 1;
2293         return 0;
2294 }
2295
2296 static int fill_holes(struct btrfs_trans_handle *trans,
2297                 struct btrfs_inode *inode,
2298                 struct btrfs_path *path, u64 offset, u64 end)
2299 {
2300         struct btrfs_fs_info *fs_info = btrfs_sb(inode->vfs_inode.i_sb);
2301         struct btrfs_root *root = inode->root;
2302         struct extent_buffer *leaf;
2303         struct btrfs_file_extent_item *fi;
2304         struct extent_map *hole_em;
2305         struct extent_map_tree *em_tree = &inode->extent_tree;
2306         struct btrfs_key key;
2307         int ret;
2308
2309         if (btrfs_fs_incompat(fs_info, NO_HOLES))
2310                 goto out;
2311
2312         key.objectid = btrfs_ino(inode);
2313         key.type = BTRFS_EXTENT_DATA_KEY;
2314         key.offset = offset;
2315
2316         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
2317         if (ret <= 0) {
2318                 /*
2319                  * We should have dropped this offset, so if we find it then
2320                  * something has gone horribly wrong.
2321                  */
2322                 if (ret == 0)
2323                         ret = -EINVAL;
2324                 return ret;
2325         }
2326
2327         leaf = path->nodes[0];
2328         if (hole_mergeable(inode, leaf, path->slots[0] - 1, offset, end)) {
2329                 u64 num_bytes;
2330
2331                 path->slots[0]--;
2332                 fi = btrfs_item_ptr(leaf, path->slots[0],
2333                                     struct btrfs_file_extent_item);
2334                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) +
2335                         end - offset;
2336                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2337                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2338                 btrfs_set_file_extent_offset(leaf, fi, 0);
2339                 btrfs_mark_buffer_dirty(leaf);
2340                 goto out;
2341         }
2342
2343         if (hole_mergeable(inode, leaf, path->slots[0], offset, end)) {
2344                 u64 num_bytes;
2345
2346                 key.offset = offset;
2347                 btrfs_set_item_key_safe(fs_info, path, &key);
2348                 fi = btrfs_item_ptr(leaf, path->slots[0],
2349                                     struct btrfs_file_extent_item);
2350                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi) + end -
2351                         offset;
2352                 btrfs_set_file_extent_num_bytes(leaf, fi, num_bytes);
2353                 btrfs_set_file_extent_ram_bytes(leaf, fi, num_bytes);
2354                 btrfs_set_file_extent_offset(leaf, fi, 0);
2355                 btrfs_mark_buffer_dirty(leaf);
2356                 goto out;
2357         }
2358         btrfs_release_path(path);
2359
2360         ret = btrfs_insert_file_extent(trans, root, btrfs_ino(inode),
2361                         offset, 0, 0, end - offset, 0, end - offset, 0, 0, 0);
2362         if (ret)
2363                 return ret;
2364
2365 out:
2366         btrfs_release_path(path);
2367
2368         hole_em = alloc_extent_map();
2369         if (!hole_em) {
2370                 btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2371                 set_bit(BTRFS_INODE_NEEDS_FULL_SYNC, &inode->runtime_flags);
2372         } else {
2373                 hole_em->start = offset;
2374                 hole_em->len = end - offset;
2375                 hole_em->ram_bytes = hole_em->len;
2376                 hole_em->orig_start = offset;
2377
2378                 hole_em->block_start = EXTENT_MAP_HOLE;
2379                 hole_em->block_len = 0;
2380                 hole_em->orig_block_len = 0;
2381                 hole_em->bdev = fs_info->fs_devices->latest_bdev;
2382                 hole_em->compress_type = BTRFS_COMPRESS_NONE;
2383                 hole_em->generation = trans->transid;
2384
2385                 do {
2386                         btrfs_drop_extent_cache(inode, offset, end - 1, 0);
2387                         write_lock(&em_tree->lock);
2388                         ret = add_extent_mapping(em_tree, hole_em, 1);
2389                         write_unlock(&em_tree->lock);
2390                 } while (ret == -EEXIST);
2391                 free_extent_map(hole_em);
2392                 if (ret)
2393                         set_bit(BTRFS_INODE_NEEDS_FULL_SYNC,
2394                                         &inode->runtime_flags);
2395         }
2396
2397         return 0;
2398 }
2399
2400 /*
2401  * Find a hole extent on given inode and change start/len to the end of hole
2402  * extent.(hole/vacuum extent whose em->start <= start &&
2403  *         em->start + em->len > start)
2404  * When a hole extent is found, return 1 and modify start/len.
2405  */
2406 static int find_first_non_hole(struct inode *inode, u64 *start, u64 *len)
2407 {
2408         struct extent_map *em;
2409         int ret = 0;
2410
2411         em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, *start, *len, 0);
2412         if (IS_ERR(em))
2413                 return PTR_ERR(em);
2414
2415         /* Hole or vacuum extent(only exists in no-hole mode) */
2416         if (em->block_start == EXTENT_MAP_HOLE) {
2417                 ret = 1;
2418                 *len = em->start + em->len > *start + *len ?
2419                        0 : *start + *len - em->start - em->len;
2420                 *start = em->start + em->len;
2421         }
2422         free_extent_map(em);
2423         return ret;
2424 }
2425
2426 static int btrfs_punch_hole(struct inode *inode, loff_t offset, loff_t len)
2427 {
2428         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2429         struct btrfs_root *root = BTRFS_I(inode)->root;
2430         struct extent_state *cached_state = NULL;
2431         struct btrfs_path *path;
2432         struct btrfs_block_rsv *rsv;
2433         struct btrfs_trans_handle *trans;
2434         u64 lockstart;
2435         u64 lockend;
2436         u64 tail_start;
2437         u64 tail_len;
2438         u64 orig_start = offset;
2439         u64 cur_offset;
2440         u64 min_size = btrfs_calc_trans_metadata_size(fs_info, 1);
2441         u64 drop_end;
2442         int ret = 0;
2443         int err = 0;
2444         unsigned int rsv_count;
2445         bool same_block;
2446         bool no_holes = btrfs_fs_incompat(fs_info, NO_HOLES);
2447         u64 ino_size;
2448         bool truncated_block = false;
2449         bool updated_inode = false;
2450
2451         ret = btrfs_wait_ordered_range(inode, offset, len);
2452         if (ret)
2453                 return ret;
2454
2455         inode_lock(inode);
2456         ino_size = round_up(inode->i_size, fs_info->sectorsize);
2457         ret = find_first_non_hole(inode, &offset, &len);
2458         if (ret < 0)
2459                 goto out_only_mutex;
2460         if (ret && !len) {
2461                 /* Already in a large hole */
2462                 ret = 0;
2463                 goto out_only_mutex;
2464         }
2465
2466         lockstart = round_up(offset, btrfs_inode_sectorsize(inode));
2467         lockend = round_down(offset + len,
2468                              btrfs_inode_sectorsize(inode)) - 1;
2469         same_block = (BTRFS_BYTES_TO_BLKS(fs_info, offset))
2470                 == (BTRFS_BYTES_TO_BLKS(fs_info, offset + len - 1));
2471         /*
2472          * We needn't truncate any block which is beyond the end of the file
2473          * because we are sure there is no data there.
2474          */
2475         /*
2476          * Only do this if we are in the same block and we aren't doing the
2477          * entire block.
2478          */
2479         if (same_block && len < fs_info->sectorsize) {
2480                 if (offset < ino_size) {
2481                         truncated_block = true;
2482                         ret = btrfs_truncate_block(inode, offset, len, 0);
2483                 } else {
2484                         ret = 0;
2485                 }
2486                 goto out_only_mutex;
2487         }
2488
2489         /* zero back part of the first block */
2490         if (offset < ino_size) {
2491                 truncated_block = true;
2492                 ret = btrfs_truncate_block(inode, offset, 0, 0);
2493                 if (ret) {
2494                         inode_unlock(inode);
2495                         return ret;
2496                 }
2497         }
2498
2499         /* Check the aligned pages after the first unaligned page,
2500          * if offset != orig_start, which means the first unaligned page
2501          * including several following pages are already in holes,
2502          * the extra check can be skipped */
2503         if (offset == orig_start) {
2504                 /* after truncate page, check hole again */
2505                 len = offset + len - lockstart;
2506                 offset = lockstart;
2507                 ret = find_first_non_hole(inode, &offset, &len);
2508                 if (ret < 0)
2509                         goto out_only_mutex;
2510                 if (ret && !len) {
2511                         ret = 0;
2512                         goto out_only_mutex;
2513                 }
2514                 lockstart = offset;
2515         }
2516
2517         /* Check the tail unaligned part is in a hole */
2518         tail_start = lockend + 1;
2519         tail_len = offset + len - tail_start;
2520         if (tail_len) {
2521                 ret = find_first_non_hole(inode, &tail_start, &tail_len);
2522                 if (unlikely(ret < 0))
2523                         goto out_only_mutex;
2524                 if (!ret) {
2525                         /* zero the front end of the last page */
2526                         if (tail_start + tail_len < ino_size) {
2527                                 truncated_block = true;
2528                                 ret = btrfs_truncate_block(inode,
2529                                                         tail_start + tail_len,
2530                                                         0, 1);
2531                                 if (ret)
2532                                         goto out_only_mutex;
2533                         }
2534                 }
2535         }
2536
2537         if (lockend < lockstart) {
2538                 ret = 0;
2539                 goto out_only_mutex;
2540         }
2541
2542         while (1) {
2543                 struct btrfs_ordered_extent *ordered;
2544
2545                 truncate_pagecache_range(inode, lockstart, lockend);
2546
2547                 lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2548                                  &cached_state);
2549                 ordered = btrfs_lookup_first_ordered_extent(inode, lockend);
2550
2551                 /*
2552                  * We need to make sure we have no ordered extents in this range
2553                  * and nobody raced in and read a page in this range, if we did
2554                  * we need to try again.
2555                  */
2556                 if ((!ordered ||
2557                     (ordered->file_offset + ordered->len <= lockstart ||
2558                      ordered->file_offset > lockend)) &&
2559                      !btrfs_page_exists_in_range(inode, lockstart, lockend)) {
2560                         if (ordered)
2561                                 btrfs_put_ordered_extent(ordered);
2562                         break;
2563                 }
2564                 if (ordered)
2565                         btrfs_put_ordered_extent(ordered);
2566                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart,
2567                                      lockend, &cached_state, GFP_NOFS);
2568                 ret = btrfs_wait_ordered_range(inode, lockstart,
2569                                                lockend - lockstart + 1);
2570                 if (ret) {
2571                         inode_unlock(inode);
2572                         return ret;
2573                 }
2574         }
2575
2576         path = btrfs_alloc_path();
2577         if (!path) {
2578                 ret = -ENOMEM;
2579                 goto out;
2580         }
2581
2582         rsv = btrfs_alloc_block_rsv(fs_info, BTRFS_BLOCK_RSV_TEMP);
2583         if (!rsv) {
2584                 ret = -ENOMEM;
2585                 goto out_free;
2586         }
2587         rsv->size = btrfs_calc_trans_metadata_size(fs_info, 1);
2588         rsv->failfast = 1;
2589
2590         /*
2591          * 1 - update the inode
2592          * 1 - removing the extents in the range
2593          * 1 - adding the hole extent if no_holes isn't set
2594          */
2595         rsv_count = no_holes ? 2 : 3;
2596         trans = btrfs_start_transaction(root, rsv_count);
2597         if (IS_ERR(trans)) {
2598                 err = PTR_ERR(trans);
2599                 goto out_free;
2600         }
2601
2602         ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv, rsv,
2603                                       min_size, 0);
2604         BUG_ON(ret);
2605         trans->block_rsv = rsv;
2606
2607         cur_offset = lockstart;
2608         len = lockend - cur_offset;
2609         while (cur_offset < lockend) {
2610                 ret = __btrfs_drop_extents(trans, root, inode, path,
2611                                            cur_offset, lockend + 1,
2612                                            &drop_end, 1, 0, 0, NULL);
2613                 if (ret != -ENOSPC)
2614                         break;
2615
2616                 trans->block_rsv = &fs_info->trans_block_rsv;
2617
2618                 if (cur_offset < drop_end && cur_offset < ino_size) {
2619                         ret = fill_holes(trans, BTRFS_I(inode), path,
2620                                         cur_offset, drop_end);
2621                         if (ret) {
2622                                 /*
2623                                  * If we failed then we didn't insert our hole
2624                                  * entries for the area we dropped, so now the
2625                                  * fs is corrupted, so we must abort the
2626                                  * transaction.
2627                                  */
2628                                 btrfs_abort_transaction(trans, ret);
2629                                 err = ret;
2630                                 break;
2631                         }
2632                 }
2633
2634                 cur_offset = drop_end;
2635
2636                 ret = btrfs_update_inode(trans, root, inode);
2637                 if (ret) {
2638                         err = ret;
2639                         break;
2640                 }
2641
2642                 btrfs_end_transaction(trans);
2643                 btrfs_btree_balance_dirty(fs_info);
2644
2645                 trans = btrfs_start_transaction(root, rsv_count);
2646                 if (IS_ERR(trans)) {
2647                         ret = PTR_ERR(trans);
2648                         trans = NULL;
2649                         break;
2650                 }
2651
2652                 ret = btrfs_block_rsv_migrate(&fs_info->trans_block_rsv,
2653                                               rsv, min_size, 0);
2654                 BUG_ON(ret);    /* shouldn't happen */
2655                 trans->block_rsv = rsv;
2656
2657                 ret = find_first_non_hole(inode, &cur_offset, &len);
2658                 if (unlikely(ret < 0))
2659                         break;
2660                 if (ret && !len) {
2661                         ret = 0;
2662                         break;
2663                 }
2664         }
2665
2666         if (ret) {
2667                 err = ret;
2668                 goto out_trans;
2669         }
2670
2671         trans->block_rsv = &fs_info->trans_block_rsv;
2672         /*
2673          * If we are using the NO_HOLES feature we might have had already an
2674          * hole that overlaps a part of the region [lockstart, lockend] and
2675          * ends at (or beyond) lockend. Since we have no file extent items to
2676          * represent holes, drop_end can be less than lockend and so we must
2677          * make sure we have an extent map representing the existing hole (the
2678          * call to __btrfs_drop_extents() might have dropped the existing extent
2679          * map representing the existing hole), otherwise the fast fsync path
2680          * will not record the existence of the hole region
2681          * [existing_hole_start, lockend].
2682          */
2683         if (drop_end <= lockend)
2684                 drop_end = lockend + 1;
2685         /*
2686          * Don't insert file hole extent item if it's for a range beyond eof
2687          * (because it's useless) or if it represents a 0 bytes range (when
2688          * cur_offset == drop_end).
2689          */
2690         if (cur_offset < ino_size && cur_offset < drop_end) {
2691                 ret = fill_holes(trans, BTRFS_I(inode), path,
2692                                 cur_offset, drop_end);
2693                 if (ret) {
2694                         /* Same comment as above. */
2695                         btrfs_abort_transaction(trans, ret);
2696                         err = ret;
2697                         goto out_trans;
2698                 }
2699         }
2700
2701 out_trans:
2702         if (!trans)
2703                 goto out_free;
2704
2705         inode_inc_iversion(inode);
2706         inode->i_mtime = inode->i_ctime = current_time(inode);
2707
2708         trans->block_rsv = &fs_info->trans_block_rsv;
2709         ret = btrfs_update_inode(trans, root, inode);
2710         updated_inode = true;
2711         btrfs_end_transaction(trans);
2712         btrfs_btree_balance_dirty(fs_info);
2713 out_free:
2714         btrfs_free_path(path);
2715         btrfs_free_block_rsv(fs_info, rsv);
2716 out:
2717         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
2718                              &cached_state, GFP_NOFS);
2719 out_only_mutex:
2720         if (!updated_inode && truncated_block && !ret && !err) {
2721                 /*
2722                  * If we only end up zeroing part of a page, we still need to
2723                  * update the inode item, so that all the time fields are
2724                  * updated as well as the necessary btrfs inode in memory fields
2725                  * for detecting, at fsync time, if the inode isn't yet in the
2726                  * log tree or it's there but not up to date.
2727                  */
2728                 trans = btrfs_start_transaction(root, 1);
2729                 if (IS_ERR(trans)) {
2730                         err = PTR_ERR(trans);
2731                 } else {
2732                         err = btrfs_update_inode(trans, root, inode);
2733                         ret = btrfs_end_transaction(trans);
2734                 }
2735         }
2736         inode_unlock(inode);
2737         if (ret && !err)
2738                 err = ret;
2739         return err;
2740 }
2741
2742 /* Helper structure to record which range is already reserved */
2743 struct falloc_range {
2744         struct list_head list;
2745         u64 start;
2746         u64 len;
2747 };
2748
2749 /*
2750  * Helper function to add falloc range
2751  *
2752  * Caller should have locked the larger range of extent containing
2753  * [start, len)
2754  */
2755 static int add_falloc_range(struct list_head *head, u64 start, u64 len)
2756 {
2757         struct falloc_range *prev = NULL;
2758         struct falloc_range *range = NULL;
2759
2760         if (list_empty(head))
2761                 goto insert;
2762
2763         /*
2764          * As fallocate iterate by bytenr order, we only need to check
2765          * the last range.
2766          */
2767         prev = list_entry(head->prev, struct falloc_range, list);
2768         if (prev->start + prev->len == start) {
2769                 prev->len += len;
2770                 return 0;
2771         }
2772 insert:
2773         range = kmalloc(sizeof(*range), GFP_KERNEL);
2774         if (!range)
2775                 return -ENOMEM;
2776         range->start = start;
2777         range->len = len;
2778         list_add_tail(&range->list, head);
2779         return 0;
2780 }
2781
2782 static long btrfs_fallocate(struct file *file, int mode,
2783                             loff_t offset, loff_t len)
2784 {
2785         struct inode *inode = file_inode(file);
2786         struct extent_state *cached_state = NULL;
2787         struct falloc_range *range;
2788         struct falloc_range *tmp;
2789         struct list_head reserve_list;
2790         u64 cur_offset;
2791         u64 last_byte;
2792         u64 alloc_start;
2793         u64 alloc_end;
2794         u64 alloc_hint = 0;
2795         u64 locked_end;
2796         u64 actual_end = 0;
2797         struct extent_map *em;
2798         int blocksize = btrfs_inode_sectorsize(inode);
2799         int ret;
2800
2801         alloc_start = round_down(offset, blocksize);
2802         alloc_end = round_up(offset + len, blocksize);
2803         cur_offset = alloc_start;
2804
2805         /* Make sure we aren't being give some crap mode */
2806         if (mode & ~(FALLOC_FL_KEEP_SIZE | FALLOC_FL_PUNCH_HOLE))
2807                 return -EOPNOTSUPP;
2808
2809         if (mode & FALLOC_FL_PUNCH_HOLE)
2810                 return btrfs_punch_hole(inode, offset, len);
2811
2812         /*
2813          * Only trigger disk allocation, don't trigger qgroup reserve
2814          *
2815          * For qgroup space, it will be checked later.
2816          */
2817         ret = btrfs_alloc_data_chunk_ondemand(BTRFS_I(inode),
2818                         alloc_end - alloc_start);
2819         if (ret < 0)
2820                 return ret;
2821
2822         inode_lock(inode);
2823
2824         if (!(mode & FALLOC_FL_KEEP_SIZE) && offset + len > inode->i_size) {
2825                 ret = inode_newsize_ok(inode, offset + len);
2826                 if (ret)
2827                         goto out;
2828         }
2829
2830         /*
2831          * TODO: Move these two operations after we have checked
2832          * accurate reserved space, or fallocate can still fail but
2833          * with page truncated or size expanded.
2834          *
2835          * But that's a minor problem and won't do much harm BTW.
2836          */
2837         if (alloc_start > inode->i_size) {
2838                 ret = btrfs_cont_expand(inode, i_size_read(inode),
2839                                         alloc_start);
2840                 if (ret)
2841                         goto out;
2842         } else if (offset + len > inode->i_size) {
2843                 /*
2844                  * If we are fallocating from the end of the file onward we
2845                  * need to zero out the end of the block if i_size lands in the
2846                  * middle of a block.
2847                  */
2848                 ret = btrfs_truncate_block(inode, inode->i_size, 0, 0);
2849                 if (ret)
2850                         goto out;
2851         }
2852
2853         /*
2854          * wait for ordered IO before we have any locks.  We'll loop again
2855          * below with the locks held.
2856          */
2857         ret = btrfs_wait_ordered_range(inode, alloc_start,
2858                                        alloc_end - alloc_start);
2859         if (ret)
2860                 goto out;
2861
2862         locked_end = alloc_end - 1;
2863         while (1) {
2864                 struct btrfs_ordered_extent *ordered;
2865
2866                 /* the extent lock is ordered inside the running
2867                  * transaction
2868                  */
2869                 lock_extent_bits(&BTRFS_I(inode)->io_tree, alloc_start,
2870                                  locked_end, &cached_state);
2871                 ordered = btrfs_lookup_first_ordered_extent(inode,
2872                                                             alloc_end - 1);
2873                 if (ordered &&
2874                     ordered->file_offset + ordered->len > alloc_start &&
2875                     ordered->file_offset < alloc_end) {
2876                         btrfs_put_ordered_extent(ordered);
2877                         unlock_extent_cached(&BTRFS_I(inode)->io_tree,
2878                                              alloc_start, locked_end,
2879                                              &cached_state, GFP_KERNEL);
2880                         /*
2881                          * we can't wait on the range with the transaction
2882                          * running or with the extent lock held
2883                          */
2884                         ret = btrfs_wait_ordered_range(inode, alloc_start,
2885                                                        alloc_end - alloc_start);
2886                         if (ret)
2887                                 goto out;
2888                 } else {
2889                         if (ordered)
2890                                 btrfs_put_ordered_extent(ordered);
2891                         break;
2892                 }
2893         }
2894
2895         /* First, check if we exceed the qgroup limit */
2896         INIT_LIST_HEAD(&reserve_list);
2897         while (1) {
2898                 em = btrfs_get_extent(BTRFS_I(inode), NULL, 0, cur_offset,
2899                                       alloc_end - cur_offset, 0);
2900                 if (IS_ERR(em)) {
2901                         ret = PTR_ERR(em);
2902                         break;
2903                 }
2904                 last_byte = min(extent_map_end(em), alloc_end);
2905                 actual_end = min_t(u64, extent_map_end(em), offset + len);
2906                 last_byte = ALIGN(last_byte, blocksize);
2907                 if (em->block_start == EXTENT_MAP_HOLE ||
2908                     (cur_offset >= inode->i_size &&
2909                      !test_bit(EXTENT_FLAG_PREALLOC, &em->flags))) {
2910                         ret = add_falloc_range(&reserve_list, cur_offset,
2911                                                last_byte - cur_offset);
2912                         if (ret < 0) {
2913                                 free_extent_map(em);
2914                                 break;
2915                         }
2916                         ret = btrfs_qgroup_reserve_data(inode, cur_offset,
2917                                         last_byte - cur_offset);
2918                         if (ret < 0) {
2919                                 free_extent_map(em);
2920                                 break;
2921                         }
2922                 } else {
2923                         /*
2924                          * Do not need to reserve unwritten extent for this
2925                          * range, free reserved data space first, otherwise
2926                          * it'll result in false ENOSPC error.
2927                          */
2928                         btrfs_free_reserved_data_space(inode, cur_offset,
2929                                 last_byte - cur_offset);
2930                 }
2931                 free_extent_map(em);
2932                 cur_offset = last_byte;
2933                 if (cur_offset >= alloc_end)
2934                         break;
2935         }
2936
2937         /*
2938          * If ret is still 0, means we're OK to fallocate.
2939          * Or just cleanup the list and exit.
2940          */
2941         list_for_each_entry_safe(range, tmp, &reserve_list, list) {
2942                 if (!ret)
2943                         ret = btrfs_prealloc_file_range(inode, mode,
2944                                         range->start,
2945                                         range->len, i_blocksize(inode),
2946                                         offset + len, &alloc_hint);
2947                 else
2948                         btrfs_free_reserved_data_space(inode, range->start,
2949                                                        range->len);
2950                 list_del(&range->list);
2951                 kfree(range);
2952         }
2953         if (ret < 0)
2954                 goto out_unlock;
2955
2956         if (actual_end > inode->i_size &&
2957             !(mode & FALLOC_FL_KEEP_SIZE)) {
2958                 struct btrfs_trans_handle *trans;
2959                 struct btrfs_root *root = BTRFS_I(inode)->root;
2960
2961                 /*
2962                  * We didn't need to allocate any more space, but we
2963                  * still extended the size of the file so we need to
2964                  * update i_size and the inode item.
2965                  */
2966                 trans = btrfs_start_transaction(root, 1);
2967                 if (IS_ERR(trans)) {
2968                         ret = PTR_ERR(trans);
2969                 } else {
2970                         inode->i_ctime = current_time(inode);
2971                         i_size_write(inode, actual_end);
2972                         btrfs_ordered_update_i_size(inode, actual_end, NULL);
2973                         ret = btrfs_update_inode(trans, root, inode);
2974                         if (ret)
2975                                 btrfs_end_transaction(trans);
2976                         else
2977                                 ret = btrfs_end_transaction(trans);
2978                 }
2979         }
2980 out_unlock:
2981         unlock_extent_cached(&BTRFS_I(inode)->io_tree, alloc_start, locked_end,
2982                              &cached_state, GFP_KERNEL);
2983 out:
2984         inode_unlock(inode);
2985         /* Let go of our reservation. */
2986         if (ret != 0)
2987                 btrfs_free_reserved_data_space(inode, alloc_start,
2988                                        alloc_end - cur_offset);
2989         return ret;
2990 }
2991
2992 static int find_desired_extent(struct inode *inode, loff_t *offset, int whence)
2993 {
2994         struct btrfs_fs_info *fs_info = btrfs_sb(inode->i_sb);
2995         struct extent_map *em = NULL;
2996         struct extent_state *cached_state = NULL;
2997         u64 lockstart;
2998         u64 lockend;
2999         u64 start;
3000         u64 len;
3001         int ret = 0;
3002
3003         if (inode->i_size == 0)
3004                 return -ENXIO;
3005
3006         /*
3007          * *offset can be negative, in this case we start finding DATA/HOLE from
3008          * the very start of the file.
3009          */
3010         start = max_t(loff_t, 0, *offset);
3011
3012         lockstart = round_down(start, fs_info->sectorsize);
3013         lockend = round_up(i_size_read(inode),
3014                            fs_info->sectorsize);
3015         if (lockend <= lockstart)
3016                 lockend = lockstart + fs_info->sectorsize;
3017         lockend--;
3018         len = lockend - lockstart + 1;
3019
3020         lock_extent_bits(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3021                          &cached_state);
3022
3023         while (start < inode->i_size) {
3024                 em = btrfs_get_extent_fiemap(BTRFS_I(inode), NULL, 0,
3025                                 start, len, 0);
3026                 if (IS_ERR(em)) {
3027                         ret = PTR_ERR(em);
3028                         em = NULL;
3029                         break;
3030                 }
3031
3032                 if (whence == SEEK_HOLE &&
3033                     (em->block_start == EXTENT_MAP_HOLE ||
3034                      test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3035                         break;
3036                 else if (whence == SEEK_DATA &&
3037                            (em->block_start != EXTENT_MAP_HOLE &&
3038                             !test_bit(EXTENT_FLAG_PREALLOC, &em->flags)))
3039                         break;
3040
3041                 start = em->start + em->len;
3042                 free_extent_map(em);
3043                 em = NULL;
3044                 cond_resched();
3045         }
3046         free_extent_map(em);
3047         if (!ret) {
3048                 if (whence == SEEK_DATA && start >= inode->i_size)
3049                         ret = -ENXIO;
3050                 else
3051                         *offset = min_t(loff_t, start, inode->i_size);
3052         }
3053         unlock_extent_cached(&BTRFS_I(inode)->io_tree, lockstart, lockend,
3054                              &cached_state, GFP_NOFS);
3055         return ret;
3056 }
3057
3058 static loff_t btrfs_file_llseek(struct file *file, loff_t offset, int whence)
3059 {
3060         struct inode *inode = file->f_mapping->host;
3061         int ret;
3062
3063         inode_lock(inode);
3064         switch (whence) {
3065         case SEEK_END:
3066         case SEEK_CUR:
3067                 offset = generic_file_llseek(file, offset, whence);
3068                 goto out;
3069         case SEEK_DATA:
3070         case SEEK_HOLE:
3071                 if (offset >= i_size_read(inode)) {
3072                         inode_unlock(inode);
3073                         return -ENXIO;
3074                 }
3075
3076                 ret = find_desired_extent(inode, &offset, whence);
3077                 if (ret) {
3078                         inode_unlock(inode);
3079                         return ret;
3080                 }
3081         }
3082
3083         offset = vfs_setpos(file, offset, inode->i_sb->s_maxbytes);
3084 out:
3085         inode_unlock(inode);
3086         return offset;
3087 }
3088
3089 static int btrfs_file_open(struct inode *inode, struct file *filp)
3090 {
3091         filp->f_mode |= FMODE_AIO_NOWAIT;
3092         return generic_file_open(inode, filp);
3093 }
3094
3095 const struct file_operations btrfs_file_operations = {
3096         .llseek         = btrfs_file_llseek,
3097         .read_iter      = generic_file_read_iter,
3098         .splice_read    = generic_file_splice_read,
3099         .write_iter     = btrfs_file_write_iter,
3100         .mmap           = btrfs_file_mmap,
3101         .open           = btrfs_file_open,
3102         .release        = btrfs_release_file,
3103         .fsync          = btrfs_sync_file,
3104         .fallocate      = btrfs_fallocate,
3105         .unlocked_ioctl = btrfs_ioctl,
3106 #ifdef CONFIG_COMPAT
3107         .compat_ioctl   = btrfs_compat_ioctl,
3108 #endif
3109         .clone_file_range = btrfs_clone_file_range,
3110         .dedupe_file_range = btrfs_dedupe_file_range,
3111 };
3112
3113 void btrfs_auto_defrag_exit(void)
3114 {
3115         kmem_cache_destroy(btrfs_inode_defrag_cachep);
3116 }
3117
3118 int btrfs_auto_defrag_init(void)
3119 {
3120         btrfs_inode_defrag_cachep = kmem_cache_create("btrfs_inode_defrag",
3121                                         sizeof(struct inode_defrag), 0,
3122                                         SLAB_MEM_SPREAD,
3123                                         NULL);
3124         if (!btrfs_inode_defrag_cachep)
3125                 return -ENOMEM;
3126
3127         return 0;
3128 }
3129
3130 int btrfs_fdatawrite_range(struct inode *inode, loff_t start, loff_t end)
3131 {
3132         int ret;
3133
3134         /*
3135          * So with compression we will find and lock a dirty page and clear the
3136          * first one as dirty, setup an async extent, and immediately return
3137          * with the entire range locked but with nobody actually marked with
3138          * writeback.  So we can't just filemap_write_and_wait_range() and
3139          * expect it to work since it will just kick off a thread to do the
3140          * actual work.  So we need to call filemap_fdatawrite_range _again_
3141          * since it will wait on the page lock, which won't be unlocked until
3142          * after the pages have been marked as writeback and so we're good to go
3143          * from there.  We have to do this otherwise we'll miss the ordered
3144          * extents and that results in badness.  Please Josef, do not think you
3145          * know better and pull this out at some point in the future, it is
3146          * right and you are wrong.
3147          */
3148         ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3149         if (!ret && test_bit(BTRFS_INODE_HAS_ASYNC_EXTENT,
3150                              &BTRFS_I(inode)->runtime_flags))
3151                 ret = filemap_fdatawrite_range(inode->i_mapping, start, end);
3152
3153         return ret;
3154 }