fs/btrfs/free-space-cache.c

   1 /*
   2  * Copyright (C) 2008 Red Hat.  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/pagemap.h>
  20 #include <linux/sched.h>
  21 #include <linux/slab.h>
  22 #include <linux/math64.h>
  23 #include <linux/ratelimit.h>
  24 #include "ctree.h"
  25 #include "free-space-cache.h"
  26 #include "transaction.h"
  27 #include "disk-io.h"
  28 #include "extent_io.h"
  29 #include "inode-map.h"
  30
  31 #define BITS_PER_BITMAP         (PAGE_CACHE_SIZE * 8)
  32 #define MAX_CACHE_BYTES_PER_GIG (32 * 1024)
  33
  34 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  35                            struct btrfs_free_space *info);
  36 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  37                               struct btrfs_free_space *info);
  38
  39 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  40                                                struct btrfs_path *path,
  41                                                u64 offset)
  42 {
  43         struct btrfs_key key;
  44         struct btrfs_key location;
  45         struct btrfs_disk_key disk_key;
  46         struct btrfs_free_space_header *header;
  47         struct extent_buffer *leaf;
  48         struct inode *inode = NULL;
  49         int ret;
  50
  51         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  52         key.offset = offset;
  53         key.type = 0;
  54
  55         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  56         if (ret < 0)
  57                 return ERR_PTR(ret);
  58         if (ret > 0) {
  59                 btrfs_release_path(path);
  60                 return ERR_PTR(-ENOENT);
  61         }
  62
  63         leaf = path->nodes[0];
  64         header = btrfs_item_ptr(leaf, path->slots[0],
  65                                 struct btrfs_free_space_header);
  66         btrfs_free_space_key(leaf, header, &disk_key);
  67         btrfs_disk_key_to_cpu(&location, &disk_key);
  68         btrfs_release_path(path);
  69
  70         inode = btrfs_iget(root->fs_info->sb, &location, root, NULL);
  71         if (!inode)
  72                 return ERR_PTR(-ENOENT);
  73         if (IS_ERR(inode))
  74                 return inode;
  75         if (is_bad_inode(inode)) {
  76                 iput(inode);
  77                 return ERR_PTR(-ENOENT);
  78         }
  79
  80         mapping_set_gfp_mask(inode->i_mapping,
  81                         mapping_gfp_mask(inode->i_mapping) & ~__GFP_FS);
  82
  83         return inode;
  84 }
  85
  86 struct inode *lookup_free_space_inode(struct btrfs_root *root,
  87                                       struct btrfs_block_group_cache
  88                                       *block_group, struct btrfs_path *path)
  89 {
  90         struct inode *inode = NULL;
  91         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
  92
  93         spin_lock(&block_group->lock);
  94         if (block_group->inode)
  95                 inode = igrab(block_group->inode);
  96         spin_unlock(&block_group->lock);
  97         if (inode)
  98                 return inode;
  99
 100         inode = __lookup_free_space_inode(root, path,
 101                                           block_group->key.objectid);
 102         if (IS_ERR(inode))
 103                 return inode;
 104
 105         spin_lock(&block_group->lock);
 106         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 107                 btrfs_info(root->fs_info,
 108                         "Old style space inode found, converting.");
 109                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 110                         BTRFS_INODE_NODATACOW;
 111                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 112         }
 113
 114         if (!block_group->iref) {
 115                 block_group->inode = igrab(inode);
 116                 block_group->iref = 1;
 117         }
 118         spin_unlock(&block_group->lock);
 119
 120         return inode;
 121 }
 122
 123 static int __create_free_space_inode(struct btrfs_root *root,
 124                                      struct btrfs_trans_handle *trans,
 125                                      struct btrfs_path *path,
 126                                      u64 ino, u64 offset)
 127 {
 128         struct btrfs_key key;
 129         struct btrfs_disk_key disk_key;
 130         struct btrfs_free_space_header *header;
 131         struct btrfs_inode_item *inode_item;
 132         struct extent_buffer *leaf;
 133         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 134         int ret;
 135
 136         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 137         if (ret)
 138                 return ret;
 139
 140         /* We inline crc's for the free disk space cache */
 141         if (ino != BTRFS_FREE_INO_OBJECTID)
 142                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 143
 144         leaf = path->nodes[0];
 145         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 146                                     struct btrfs_inode_item);
 147         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 148         memset_extent_buffer(leaf, 0, (unsigned long)inode_item,
 149                              sizeof(*inode_item));
 150         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 151         btrfs_set_inode_size(leaf, inode_item, 0);
 152         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 153         btrfs_set_inode_uid(leaf, inode_item, 0);
 154         btrfs_set_inode_gid(leaf, inode_item, 0);
 155         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 156         btrfs_set_inode_flags(leaf, inode_item, flags);
 157         btrfs_set_inode_nlink(leaf, inode_item, 1);
 158         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 159         btrfs_set_inode_block_group(leaf, inode_item, offset);
 160         btrfs_mark_buffer_dirty(leaf);
 161         btrfs_release_path(path);
 162
 163         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 164         key.offset = offset;
 165         key.type = 0;
 166
 167         ret = btrfs_insert_empty_item(trans, root, path, &key,
 168                                       sizeof(struct btrfs_free_space_header));
 169         if (ret < 0) {
 170                 btrfs_release_path(path);
 171                 return ret;
 172         }
 173         leaf = path->nodes[0];
 174         header = btrfs_item_ptr(leaf, path->slots[0],
 175                                 struct btrfs_free_space_header);
 176         memset_extent_buffer(leaf, 0, (unsigned long)header, sizeof(*header));
 177         btrfs_set_free_space_key(leaf, header, &disk_key);
 178         btrfs_mark_buffer_dirty(leaf);
 179         btrfs_release_path(path);
 180
 181         return 0;
 182 }
 183
 184 int create_free_space_inode(struct btrfs_root *root,
 185                             struct btrfs_trans_handle *trans,
 186                             struct btrfs_block_group_cache *block_group,
 187                             struct btrfs_path *path)
 188 {
 189         int ret;
 190         u64 ino;
 191
 192         ret = btrfs_find_free_objectid(root, &ino);
 193         if (ret < 0)
 194                 return ret;
 195
 196         return __create_free_space_inode(root, trans, path, ino,
 197                                          block_group->key.objectid);
 198 }
 199
 200 int btrfs_check_trunc_cache_free_space(struct btrfs_root *root,
 201                                        struct btrfs_block_rsv *rsv)
 202 {
 203         u64 needed_bytes;
 204         int ret;
 205
 206         /* 1 for slack space, 1 for updating the inode */
 207         needed_bytes = btrfs_calc_trunc_metadata_size(root, 1) +
 208                 btrfs_calc_trans_metadata_size(root, 1);
 209
 210         spin_lock(&rsv->lock);
 211         if (rsv->reserved < needed_bytes)
 212                 ret = -ENOSPC;
 213         else
 214                 ret = 0;
 215         spin_unlock(&rsv->lock);
 216         return ret;
 217 }
 218
 219 int btrfs_truncate_free_space_cache(struct btrfs_root *root,
 220                                     struct btrfs_trans_handle *trans,
 221                                     struct btrfs_path *path,
 222                                     struct inode *inode)
 223 {
 224         loff_t oldsize;
 225         int ret = 0;
 226
 227         oldsize = i_size_read(inode);
 228         btrfs_i_size_write(inode, 0);
 229         truncate_pagecache(inode, oldsize, 0);
 230
 231         /*
 232          * We don't need an orphan item because truncating the free space cache
 233          * will never be split across transactions.
 234          */
 235         ret = btrfs_truncate_inode_items(trans, root, inode,
 236                                          0, BTRFS_EXTENT_DATA_KEY);
 237         if (ret) {
 238                 btrfs_abort_transaction(trans, root, ret);
 239                 return ret;
 240         }
 241
 242         ret = btrfs_update_inode(trans, root, inode);
 243         if (ret)
 244                 btrfs_abort_transaction(trans, root, ret);
 245
 246         return ret;
 247 }
 248
 249 static int readahead_cache(struct inode *inode)
 250 {
 251         struct file_ra_state *ra;
 252         unsigned long last_index;
 253
 254         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 255         if (!ra)
 256                 return -ENOMEM;
 257
 258         file_ra_state_init(ra, inode->i_mapping);
 259         last_index = (i_size_read(inode) - 1) >> PAGE_CACHE_SHIFT;
 260
 261         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 262
 263         kfree(ra);
 264
 265         return 0;
 266 }
 267
 268 struct io_ctl {
 269         void *cur, *orig;
 270         struct page *page;
 271         struct page **pages;
 272         struct btrfs_root *root;
 273         unsigned long size;
 274         int index;
 275         int num_pages;
 276         unsigned check_crcs:1;
 277 };
 278
 279 static int io_ctl_init(struct io_ctl *io_ctl, struct inode *inode,
 280                        struct btrfs_root *root)
 281 {
 282         memset(io_ctl, 0, sizeof(struct io_ctl));
 283         io_ctl->num_pages = (i_size_read(inode) + PAGE_CACHE_SIZE - 1) >>
 284                 PAGE_CACHE_SHIFT;
 285         io_ctl->pages = kzalloc(sizeof(struct page *) * io_ctl->num_pages,
 286                                 GFP_NOFS);
 287         if (!io_ctl->pages)
 288                 return -ENOMEM;
 289         io_ctl->root = root;
 290         if (btrfs_ino(inode) != BTRFS_FREE_INO_OBJECTID)
 291                 io_ctl->check_crcs = 1;
 292         return 0;
 293 }
 294
 295 static void io_ctl_free(struct io_ctl *io_ctl)
 296 {
 297         kfree(io_ctl->pages);
 298 }
 299
 300 static void io_ctl_unmap_page(struct io_ctl *io_ctl)
 301 {
 302         if (io_ctl->cur) {
 303                 kunmap(io_ctl->page);
 304                 io_ctl->cur = NULL;
 305                 io_ctl->orig = NULL;
 306         }
 307 }
 308
 309 static void io_ctl_map_page(struct io_ctl *io_ctl, int clear)
 310 {
 311         BUG_ON(io_ctl->index >= io_ctl->num_pages);
 312         io_ctl->page = io_ctl->pages[io_ctl->index++];
 313         io_ctl->cur = kmap(io_ctl->page);
 314         io_ctl->orig = io_ctl->cur;
 315         io_ctl->size = PAGE_CACHE_SIZE;
 316         if (clear)
 317                 memset(io_ctl->cur, 0, PAGE_CACHE_SIZE);
 318 }
 319
 320 static void io_ctl_drop_pages(struct io_ctl *io_ctl)
 321 {
 322         int i;
 323
 324         io_ctl_unmap_page(io_ctl);
 325
 326         for (i = 0; i < io_ctl->num_pages; i++) {
 327                 if (io_ctl->pages[i]) {
 328                         ClearPageChecked(io_ctl->pages[i]);
 329                         unlock_page(io_ctl->pages[i]);
 330                         page_cache_release(io_ctl->pages[i]);
 331                 }
 332         }
 333 }
 334
 335 static int io_ctl_prepare_pages(struct io_ctl *io_ctl, struct inode *inode,
 336                                 int uptodate)
 337 {
 338         struct page *page;
 339         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 340         int i;
 341
 342         for (i = 0; i < io_ctl->num_pages; i++) {
 343                 page = find_or_create_page(inode->i_mapping, i, mask);
 344                 if (!page) {
 345                         io_ctl_drop_pages(io_ctl);
 346                         return -ENOMEM;
 347                 }
 348                 io_ctl->pages[i] = page;
 349                 if (uptodate && !PageUptodate(page)) {
 350                         btrfs_readpage(NULL, page);
 351                         lock_page(page);
 352                         if (!PageUptodate(page)) {
 353                                 printk(KERN_ERR "btrfs: error reading free "
 354                                        "space cache\n");
 355                                 io_ctl_drop_pages(io_ctl);
 356                                 return -EIO;
 357                         }
 358                 }
 359         }
 360
 361         for (i = 0; i < io_ctl->num_pages; i++) {
 362                 clear_page_dirty_for_io(io_ctl->pages[i]);
 363                 set_page_extent_mapped(io_ctl->pages[i]);
 364         }
 365
 366         return 0;
 367 }
 368
 369 static void io_ctl_set_generation(struct io_ctl *io_ctl, u64 generation)
 370 {
 371         __le64 *val;
 372
 373         io_ctl_map_page(io_ctl, 1);
 374
 375         /*
 376          * Skip the csum areas.  If we don't check crcs then we just have a
 377          * 64bit chunk at the front of the first page.
 378          */
 379         if (io_ctl->check_crcs) {
 380                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 381                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 382         } else {
 383                 io_ctl->cur += sizeof(u64);
 384                 io_ctl->size -= sizeof(u64) * 2;
 385         }
 386
 387         val = io_ctl->cur;
 388         *val = cpu_to_le64(generation);
 389         io_ctl->cur += sizeof(u64);
 390 }
 391
 392 static int io_ctl_check_generation(struct io_ctl *io_ctl, u64 generation)
 393 {
 394         __le64 *gen;
 395
 396         /*
 397          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 398          * chunk at the front of the first page.
 399          */
 400         if (io_ctl->check_crcs) {
 401                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 402                 io_ctl->size -= sizeof(u64) +
 403                         (sizeof(u32) * io_ctl->num_pages);
 404         } else {
 405                 io_ctl->cur += sizeof(u64);
 406                 io_ctl->size -= sizeof(u64) * 2;
 407         }
 408
 409         gen = io_ctl->cur;
 410         if (le64_to_cpu(*gen) != generation) {
 411                 printk_ratelimited(KERN_ERR "btrfs: space cache generation "
 412                                    "(%Lu) does not match inode (%Lu)\n", *gen,
 413                                    generation);
 414                 io_ctl_unmap_page(io_ctl);
 415                 return -EIO;
 416         }
 417         io_ctl->cur += sizeof(u64);
 418         return 0;
 419 }
 420
 421 static void io_ctl_set_crc(struct io_ctl *io_ctl, int index)
 422 {
 423         u32 *tmp;
 424         u32 crc = ~(u32)0;
 425         unsigned offset = 0;
 426
 427         if (!io_ctl->check_crcs) {
 428                 io_ctl_unmap_page(io_ctl);
 429                 return;
 430         }
 431
 432         if (index == 0)
 433                 offset = sizeof(u32) * io_ctl->num_pages;
 434
 435         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 436                               PAGE_CACHE_SIZE - offset);
 437         btrfs_csum_final(crc, (char *)&crc);
 438         io_ctl_unmap_page(io_ctl);
 439         tmp = kmap(io_ctl->pages[0]);
 440         tmp += index;
 441         *tmp = crc;
 442         kunmap(io_ctl->pages[0]);
 443 }
 444
 445 static int io_ctl_check_crc(struct io_ctl *io_ctl, int index)
 446 {
 447         u32 *tmp, val;
 448         u32 crc = ~(u32)0;
 449         unsigned offset = 0;
 450
 451         if (!io_ctl->check_crcs) {
 452                 io_ctl_map_page(io_ctl, 0);
 453                 return 0;
 454         }
 455
 456         if (index == 0)
 457                 offset = sizeof(u32) * io_ctl->num_pages;
 458
 459         tmp = kmap(io_ctl->pages[0]);
 460         tmp += index;
 461         val = *tmp;
 462         kunmap(io_ctl->pages[0]);
 463
 464         io_ctl_map_page(io_ctl, 0);
 465         crc = btrfs_csum_data(io_ctl->orig + offset, crc,
 466                               PAGE_CACHE_SIZE - offset);
 467         btrfs_csum_final(crc, (char *)&crc);
 468         if (val != crc) {
 469                 printk_ratelimited(KERN_ERR "btrfs: csum mismatch on free "
 470                                    "space cache\n");
 471                 io_ctl_unmap_page(io_ctl);
 472                 return -EIO;
 473         }
 474
 475         return 0;
 476 }
 477
 478 static int io_ctl_add_entry(struct io_ctl *io_ctl, u64 offset, u64 bytes,
 479                             void *bitmap)
 480 {
 481         struct btrfs_free_space_entry *entry;
 482
 483         if (!io_ctl->cur)
 484                 return -ENOSPC;
 485
 486         entry = io_ctl->cur;
 487         entry->offset = cpu_to_le64(offset);
 488         entry->bytes = cpu_to_le64(bytes);
 489         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 490                 BTRFS_FREE_SPACE_EXTENT;
 491         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 492         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 493
 494         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 495                 return 0;
 496
 497         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 498
 499         /* No more pages to map */
 500         if (io_ctl->index >= io_ctl->num_pages)
 501                 return 0;
 502
 503         /* map the next page */
 504         io_ctl_map_page(io_ctl, 1);
 505         return 0;
 506 }
 507
 508 static int io_ctl_add_bitmap(struct io_ctl *io_ctl, void *bitmap)
 509 {
 510         if (!io_ctl->cur)
 511                 return -ENOSPC;
 512
 513         /*
 514          * If we aren't at the start of the current page, unmap this one and
 515          * map the next one if there is any left.
 516          */
 517         if (io_ctl->cur != io_ctl->orig) {
 518                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 519                 if (io_ctl->index >= io_ctl->num_pages)
 520                         return -ENOSPC;
 521                 io_ctl_map_page(io_ctl, 0);
 522         }
 523
 524         memcpy(io_ctl->cur, bitmap, PAGE_CACHE_SIZE);
 525         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 526         if (io_ctl->index < io_ctl->num_pages)
 527                 io_ctl_map_page(io_ctl, 0);
 528         return 0;
 529 }
 530
 531 static void io_ctl_zero_remaining_pages(struct io_ctl *io_ctl)
 532 {
 533         /*
 534          * If we're not on the boundary we know we've modified the page and we
 535          * need to crc the page.
 536          */
 537         if (io_ctl->cur != io_ctl->orig)
 538                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 539         else
 540                 io_ctl_unmap_page(io_ctl);
 541
 542         while (io_ctl->index < io_ctl->num_pages) {
 543                 io_ctl_map_page(io_ctl, 1);
 544                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 545         }
 546 }
 547
 548 static int io_ctl_read_entry(struct io_ctl *io_ctl,
 549                             struct btrfs_free_space *entry, u8 *type)
 550 {
 551         struct btrfs_free_space_entry *e;
 552         int ret;
 553
 554         if (!io_ctl->cur) {
 555                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 556                 if (ret)
 557                         return ret;
 558         }
 559
 560         e = io_ctl->cur;
 561         entry->offset = le64_to_cpu(e->offset);
 562         entry->bytes = le64_to_cpu(e->bytes);
 563         *type = e->type;
 564         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 565         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 566
 567         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 568                 return 0;
 569
 570         io_ctl_unmap_page(io_ctl);
 571
 572         return 0;
 573 }
 574
 575 static int io_ctl_read_bitmap(struct io_ctl *io_ctl,
 576                               struct btrfs_free_space *entry)
 577 {
 578         int ret;
 579
 580         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 581         if (ret)
 582                 return ret;
 583
 584         memcpy(entry->bitmap, io_ctl->cur, PAGE_CACHE_SIZE);
 585         io_ctl_unmap_page(io_ctl);
 586
 587         return 0;
 588 }
 589
 590 /*
 591  * Since we attach pinned extents after the fact we can have contiguous sections
 592  * of free space that are split up in entries.  This poses a problem with the
 593  * tree logging stuff since it could have allocated across what appears to be 2
 594  * entries since we would have merged the entries when adding the pinned extents
 595  * back to the free space cache.  So run through the space cache that we just
 596  * loaded and merge contiguous entries.  This will make the log replay stuff not
 597  * blow up and it will make for nicer allocator behavior.
 598  */
 599 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 600 {
 601         struct btrfs_free_space *e, *prev = NULL;
 602         struct rb_node *n;
 603
 604 again:
 605         spin_lock(&ctl->tree_lock);
 606         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 607                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 608                 if (!prev)
 609                         goto next;
 610                 if (e->bitmap || prev->bitmap)
 611                         goto next;
 612                 if (prev->offset + prev->bytes == e->offset) {
 613                         unlink_free_space(ctl, prev);
 614                         unlink_free_space(ctl, e);
 615                         prev->bytes += e->bytes;
 616                         kmem_cache_free(btrfs_free_space_cachep, e);
 617                         link_free_space(ctl, prev);
 618                         prev = NULL;
 619                         spin_unlock(&ctl->tree_lock);
 620                         goto again;
 621                 }
 622 next:
 623                 prev = e;
 624         }
 625         spin_unlock(&ctl->tree_lock);
 626 }
 627
 628 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 629                                    struct btrfs_free_space_ctl *ctl,
 630                                    struct btrfs_path *path, u64 offset)
 631 {
 632         struct btrfs_free_space_header *header;
 633         struct extent_buffer *leaf;
 634         struct io_ctl io_ctl;
 635         struct btrfs_key key;
 636         struct btrfs_free_space *e, *n;
 637         struct list_head bitmaps;
 638         u64 num_entries;
 639         u64 num_bitmaps;
 640         u64 generation;
 641         u8 type;
 642         int ret = 0;
 643
 644         INIT_LIST_HEAD(&bitmaps);
 645
 646         /* Nothing in the space cache, goodbye */
 647         if (!i_size_read(inode))
 648                 return 0;
 649
 650         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 651         key.offset = offset;
 652         key.type = 0;
 653
 654         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 655         if (ret < 0)
 656                 return 0;
 657         else if (ret > 0) {
 658                 btrfs_release_path(path);
 659                 return 0;
 660         }
 661
 662         ret = -1;
 663
 664         leaf = path->nodes[0];
 665         header = btrfs_item_ptr(leaf, path->slots[0],
 666                                 struct btrfs_free_space_header);
 667         num_entries = btrfs_free_space_entries(leaf, header);
 668         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 669         generation = btrfs_free_space_generation(leaf, header);
 670         btrfs_release_path(path);
 671
 672         if (BTRFS_I(inode)->generation != generation) {
 673                 btrfs_err(root->fs_info,
 674                         "free space inode generation (%llu) "
 675                         "did not match free space cache generation (%llu)",
 676                         BTRFS_I(inode)->generation, generation);
 677                 return 0;
 678         }
 679
 680         if (!num_entries)
 681                 return 0;
 682
 683         ret = io_ctl_init(&io_ctl, inode, root);
 684         if (ret)
 685                 return ret;
 686
 687         ret = readahead_cache(inode);
 688         if (ret)
 689                 goto out;
 690
 691         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 692         if (ret)
 693                 goto out;
 694
 695         ret = io_ctl_check_crc(&io_ctl, 0);
 696         if (ret)
 697                 goto free_cache;
 698
 699         ret = io_ctl_check_generation(&io_ctl, generation);
 700         if (ret)
 701                 goto free_cache;
 702
 703         while (num_entries) {
 704                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 705                                       GFP_NOFS);
 706                 if (!e)
 707                         goto free_cache;
 708
 709                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 710                 if (ret) {
 711                         kmem_cache_free(btrfs_free_space_cachep, e);
 712                         goto free_cache;
 713                 }
 714
 715                 if (!e->bytes) {
 716                         kmem_cache_free(btrfs_free_space_cachep, e);
 717                         goto free_cache;
 718                 }
 719
 720                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 721                         spin_lock(&ctl->tree_lock);
 722                         ret = link_free_space(ctl, e);
 723                         spin_unlock(&ctl->tree_lock);
 724                         if (ret) {
 725                                 btrfs_err(root->fs_info,
 726                                         "Duplicate entries in free space cache, dumping");
 727                                 kmem_cache_free(btrfs_free_space_cachep, e);
 728                                 goto free_cache;
 729                         }
 730                 } else {
 731                         BUG_ON(!num_bitmaps);
 732                         num_bitmaps--;
 733                         e->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
 734                         if (!e->bitmap) {
 735                                 kmem_cache_free(
 736                                         btrfs_free_space_cachep, e);
 737                                 goto free_cache;
 738                         }
 739                         spin_lock(&ctl->tree_lock);
 740                         ret = link_free_space(ctl, e);
 741                         ctl->total_bitmaps++;
 742                         ctl->op->recalc_thresholds(ctl);
 743                         spin_unlock(&ctl->tree_lock);
 744                         if (ret) {
 745                                 btrfs_err(root->fs_info,
 746                                         "Duplicate entries in free space cache, dumping");
 747                                 kmem_cache_free(btrfs_free_space_cachep, e);
 748                                 goto free_cache;
 749                         }
 750                         list_add_tail(&e->list, &bitmaps);
 751                 }
 752
 753                 num_entries--;
 754         }
 755
 756         io_ctl_unmap_page(&io_ctl);
 757
 758         /*
 759          * We add the bitmaps at the end of the entries in order that
 760          * the bitmap entries are added to the cache.
 761          */
 762         list_for_each_entry_safe(e, n, &bitmaps, list) {
 763                 list_del_init(&e->list);
 764                 ret = io_ctl_read_bitmap(&io_ctl, e);
 765                 if (ret)
 766                         goto free_cache;
 767         }
 768
 769         io_ctl_drop_pages(&io_ctl);
 770         merge_space_tree(ctl);
 771         ret = 1;
 772 out:
 773         io_ctl_free(&io_ctl);
 774         return ret;
 775 free_cache:
 776         io_ctl_drop_pages(&io_ctl);
 777         __btrfs_remove_free_space_cache(ctl);
 778         goto out;
 779 }
 780
 781 int load_free_space_cache(struct btrfs_fs_info *fs_info,
 782                           struct btrfs_block_group_cache *block_group)
 783 {
 784         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 785         struct btrfs_root *root = fs_info->tree_root;
 786         struct inode *inode;
 787         struct btrfs_path *path;
 788         int ret = 0;
 789         bool matched;
 790         u64 used = btrfs_block_group_used(&block_group->item);
 791
 792         /*
 793          * If this block group has been marked to be cleared for one reason or
 794          * another then we can't trust the on disk cache, so just return.
 795          */
 796         spin_lock(&block_group->lock);
 797         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 798                 spin_unlock(&block_group->lock);
 799                 return 0;
 800         }
 801         spin_unlock(&block_group->lock);
 802
 803         path = btrfs_alloc_path();
 804         if (!path)
 805                 return 0;
 806         path->search_commit_root = 1;
 807         path->skip_locking = 1;
 808
 809         inode = lookup_free_space_inode(root, block_group, path);
 810         if (IS_ERR(inode)) {
 811                 btrfs_free_path(path);
 812                 return 0;
 813         }
 814
 815         /* We may have converted the inode and made the cache invalid. */
 816         spin_lock(&block_group->lock);
 817         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 818                 spin_unlock(&block_group->lock);
 819                 btrfs_free_path(path);
 820                 goto out;
 821         }
 822         spin_unlock(&block_group->lock);
 823
 824         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 825                                       path, block_group->key.objectid);
 826         btrfs_free_path(path);
 827         if (ret <= 0)
 828                 goto out;
 829
 830         spin_lock(&ctl->tree_lock);
 831         matched = (ctl->free_space == (block_group->key.offset - used -
 832                                        block_group->bytes_super));
 833         spin_unlock(&ctl->tree_lock);
 834
 835         if (!matched) {
 836                 __btrfs_remove_free_space_cache(ctl);
 837                 btrfs_err(fs_info, "block group %llu has wrong amount of free space",
 838                         block_group->key.objectid);
 839                 ret = -1;
 840         }
 841 out:
 842         if (ret < 0) {
 843                 /* This cache is bogus, make sure it gets cleared */
 844                 spin_lock(&block_group->lock);
 845                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 846                 spin_unlock(&block_group->lock);
 847                 ret = 0;
 848
 849                 btrfs_err(fs_info, "failed to load free space cache for block group %llu",
 850                         block_group->key.objectid);
 851         }
 852
 853         iput(inode);
 854         return ret;
 855 }
 856
 857 /**
 858  * __btrfs_write_out_cache - write out cached info to an inode
 859  * @root - the root the inode belongs to
 860  * @ctl - the free space cache we are going to write out
 861  * @block_group - the block_group for this cache if it belongs to a block_group
 862  * @trans - the trans handle
 863  * @path - the path to use
 864  * @offset - the offset for the key we'll insert
 865  *
 866  * This function writes out a free space cache struct to disk for quick recovery
 867  * on mount.  This will return 0 if it was successfull in writing the cache out,
 868  * and -1 if it was not.
 869  */
 870 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
 871                                    struct btrfs_free_space_ctl *ctl,
 872                                    struct btrfs_block_group_cache *block_group,
 873                                    struct btrfs_trans_handle *trans,
 874                                    struct btrfs_path *path, u64 offset)
 875 {
 876         struct btrfs_free_space_header *header;
 877         struct extent_buffer *leaf;
 878         struct rb_node *node;
 879         struct list_head *pos, *n;
 880         struct extent_state *cached_state = NULL;
 881         struct btrfs_free_cluster *cluster = NULL;
 882         struct extent_io_tree *unpin = NULL;
 883         struct io_ctl io_ctl;
 884         struct list_head bitmap_list;
 885         struct btrfs_key key;
 886         u64 start, extent_start, extent_end, len;
 887         int entries = 0;
 888         int bitmaps = 0;
 889         int ret;
 890         int err = -1;
 891
 892         INIT_LIST_HEAD(&bitmap_list);
 893
 894         if (!i_size_read(inode))
 895                 return -1;
 896
 897         ret = io_ctl_init(&io_ctl, inode, root);
 898         if (ret)
 899                 return -1;
 900
 901         /* Get the cluster for this block_group if it exists */
 902         if (block_group && !list_empty(&block_group->cluster_list))
 903                 cluster = list_entry(block_group->cluster_list.next,
 904                                      struct btrfs_free_cluster,
 905                                      block_group_list);
 906
 907         /* Lock all pages first so we can lock the extent safely. */
 908         io_ctl_prepare_pages(&io_ctl, inode, 0);
 909
 910         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
 911                          0, &cached_state);
 912
 913         node = rb_first(&ctl->free_space_offset);
 914         if (!node && cluster) {
 915                 node = rb_first(&cluster->root);
 916                 cluster = NULL;
 917         }
 918
 919         /* Make sure we can fit our crcs into the first page */
 920         if (io_ctl.check_crcs &&
 921             (io_ctl.num_pages * sizeof(u32)) >= PAGE_CACHE_SIZE)
 922                 goto out_nospc;
 923
 924         io_ctl_set_generation(&io_ctl, trans->transid);
 925
 926         /* Write out the extent entries */
 927         while (node) {
 928                 struct btrfs_free_space *e;
 929
 930                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 931                 entries++;
 932
 933                 ret = io_ctl_add_entry(&io_ctl, e->offset, e->bytes,
 934                                        e->bitmap);
 935                 if (ret)
 936                         goto out_nospc;
 937
 938                 if (e->bitmap) {
 939                         list_add_tail(&e->list, &bitmap_list);
 940                         bitmaps++;
 941                 }
 942                 node = rb_next(node);
 943                 if (!node && cluster) {
 944                         node = rb_first(&cluster->root);
 945                         cluster = NULL;
 946                 }
 947         }
 948
 949         /*
 950          * We want to add any pinned extents to our free space cache
 951          * so we don't leak the space
 952          */
 953
 954         /*
 955          * We shouldn't have switched the pinned extents yet so this is the
 956          * right one
 957          */
 958         unpin = root->fs_info->pinned_extents;
 959
 960         if (block_group)
 961                 start = block_group->key.objectid;
 962
 963         while (block_group && (start < block_group->key.objectid +
 964                                block_group->key.offset)) {
 965                 ret = find_first_extent_bit(unpin, start,
 966                                             &extent_start, &extent_end,
 967                                             EXTENT_DIRTY, NULL);
 968                 if (ret) {
 969                         ret = 0;
 970                         break;
 971                 }
 972
 973                 /* This pinned extent is out of our range */
 974                 if (extent_start >= block_group->key.objectid +
 975                     block_group->key.offset)
 976                         break;
 977
 978                 extent_start = max(extent_start, start);
 979                 extent_end = min(block_group->key.objectid +
 980                                  block_group->key.offset, extent_end + 1);
 981                 len = extent_end - extent_start;
 982
 983                 entries++;
 984                 ret = io_ctl_add_entry(&io_ctl, extent_start, len, NULL);
 985                 if (ret)
 986                         goto out_nospc;
 987
 988                 start = extent_end;
 989         }
 990
 991         /* Write out the bitmaps */
 992         list_for_each_safe(pos, n, &bitmap_list) {
 993                 struct btrfs_free_space *entry =
 994                         list_entry(pos, struct btrfs_free_space, list);
 995
 996                 ret = io_ctl_add_bitmap(&io_ctl, entry->bitmap);
 997                 if (ret)
 998                         goto out_nospc;
 999                 list_del_init(&entry->list);
1000         }
1001
1002         /* Zero out the rest of the pages just to make sure */
1003         io_ctl_zero_remaining_pages(&io_ctl);
1004
1005         ret = btrfs_dirty_pages(root, inode, io_ctl.pages, io_ctl.num_pages,
1006                                 0, i_size_read(inode), &cached_state);
1007         io_ctl_drop_pages(&io_ctl);
1008         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1009                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1010
1011         if (ret)
1012                 goto out;
1013
1014
1015         btrfs_wait_ordered_range(inode, 0, (u64)-1);
1016
1017         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1018         key.offset = offset;
1019         key.type = 0;
1020
1021         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1022         if (ret < 0) {
1023                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1024                                  EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0, NULL,
1025                                  GFP_NOFS);
1026                 goto out;
1027         }
1028         leaf = path->nodes[0];
1029         if (ret > 0) {
1030                 struct btrfs_key found_key;
1031                 BUG_ON(!path->slots[0]);
1032                 path->slots[0]--;
1033                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1034                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1035                     found_key.offset != offset) {
1036                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1037                                          inode->i_size - 1,
1038                                          EXTENT_DIRTY | EXTENT_DELALLOC, 0, 0,
1039                                          NULL, GFP_NOFS);
1040                         btrfs_release_path(path);
1041                         goto out;
1042                 }
1043         }
1044
1045         BTRFS_I(inode)->generation = trans->transid;
1046         header = btrfs_item_ptr(leaf, path->slots[0],
1047                                 struct btrfs_free_space_header);
1048         btrfs_set_free_space_entries(leaf, header, entries);
1049         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1050         btrfs_set_free_space_generation(leaf, header, trans->transid);
1051         btrfs_mark_buffer_dirty(leaf);
1052         btrfs_release_path(path);
1053
1054         err = 0;
1055 out:
1056         io_ctl_free(&io_ctl);
1057         if (err) {
1058                 invalidate_inode_pages2(inode->i_mapping);
1059                 BTRFS_I(inode)->generation = 0;
1060         }
1061         btrfs_update_inode(trans, root, inode);
1062         return err;
1063
1064 out_nospc:
1065         list_for_each_safe(pos, n, &bitmap_list) {
1066                 struct btrfs_free_space *entry =
1067                         list_entry(pos, struct btrfs_free_space, list);
1068                 list_del_init(&entry->list);
1069         }
1070         io_ctl_drop_pages(&io_ctl);
1071         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1072                              i_size_read(inode) - 1, &cached_state, GFP_NOFS);
1073         goto out;
1074 }
1075
1076 int btrfs_write_out_cache(struct btrfs_root *root,
1077                           struct btrfs_trans_handle *trans,
1078                           struct btrfs_block_group_cache *block_group,
1079                           struct btrfs_path *path)
1080 {
1081         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1082         struct inode *inode;
1083         int ret = 0;
1084
1085         root = root->fs_info->tree_root;
1086
1087         spin_lock(&block_group->lock);
1088         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1089                 spin_unlock(&block_group->lock);
1090                 return 0;
1091         }
1092         spin_unlock(&block_group->lock);
1093
1094         inode = lookup_free_space_inode(root, block_group, path);
1095         if (IS_ERR(inode))
1096                 return 0;
1097
1098         ret = __btrfs_write_out_cache(root, inode, ctl, block_group, trans,
1099                                       path, block_group->key.objectid);
1100         if (ret) {
1101                 spin_lock(&block_group->lock);
1102                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1103                 spin_unlock(&block_group->lock);
1104                 ret = 0;
1105 #ifdef DEBUG
1106                 btrfs_err(root->fs_info,
1107                         "failed to write free space cache for block group %llu",
1108                         block_group->key.objectid);
1109 #endif
1110         }
1111
1112         iput(inode);
1113         return ret;
1114 }
1115
1116 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1117                                           u64 offset)
1118 {
1119         BUG_ON(offset < bitmap_start);
1120         offset -= bitmap_start;
1121         return (unsigned long)(div_u64(offset, unit));
1122 }
1123
1124 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1125 {
1126         return (unsigned long)(div_u64(bytes, unit));
1127 }
1128
1129 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1130                                    u64 offset)
1131 {
1132         u64 bitmap_start;
1133         u64 bytes_per_bitmap;
1134
1135         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1136         bitmap_start = offset - ctl->start;
1137         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1138         bitmap_start *= bytes_per_bitmap;
1139         bitmap_start += ctl->start;
1140
1141         return bitmap_start;
1142 }
1143
1144 static int tree_insert_offset(struct rb_root *root, u64 offset,
1145                               struct rb_node *node, int bitmap)
1146 {
1147         struct rb_node **p = &root->rb_node;
1148         struct rb_node *parent = NULL;
1149         struct btrfs_free_space *info;
1150
1151         while (*p) {
1152                 parent = *p;
1153                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1154
1155                 if (offset < info->offset) {
1156                         p = &(*p)->rb_left;
1157                 } else if (offset > info->offset) {
1158                         p = &(*p)->rb_right;
1159                 } else {
1160                         /*
1161                          * we could have a bitmap entry and an extent entry
1162                          * share the same offset.  If this is the case, we want
1163                          * the extent entry to always be found first if we do a
1164                          * linear search through the tree, since we want to have
1165                          * the quickest allocation time, and allocating from an
1166                          * extent is faster than allocating from a bitmap.  So
1167                          * if we're inserting a bitmap and we find an entry at
1168                          * this offset, we want to go right, or after this entry
1169                          * logically.  If we are inserting an extent and we've
1170                          * found a bitmap, we want to go left, or before
1171                          * logically.
1172                          */
1173                         if (bitmap) {
1174                                 if (info->bitmap) {
1175                                         WARN_ON_ONCE(1);
1176                                         return -EEXIST;
1177                                 }
1178                                 p = &(*p)->rb_right;
1179                         } else {
1180                                 if (!info->bitmap) {
1181                                         WARN_ON_ONCE(1);
1182                                         return -EEXIST;
1183                                 }
1184                                 p = &(*p)->rb_left;
1185                         }
1186                 }
1187         }
1188
1189         rb_link_node(node, parent, p);
1190         rb_insert_color(node, root);
1191
1192         return 0;
1193 }
1194
1195 /*
1196  * searches the tree for the given offset.
1197  *
1198  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1199  * want a section that has at least bytes size and comes at or after the given
1200  * offset.
1201  */
1202 static struct btrfs_free_space *
1203 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1204                    u64 offset, int bitmap_only, int fuzzy)
1205 {
1206         struct rb_node *n = ctl->free_space_offset.rb_node;
1207         struct btrfs_free_space *entry, *prev = NULL;
1208
1209         /* find entry that is closest to the 'offset' */
1210         while (1) {
1211                 if (!n) {
1212                         entry = NULL;
1213                         break;
1214                 }
1215
1216                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1217                 prev = entry;
1218
1219                 if (offset < entry->offset)
1220                         n = n->rb_left;
1221                 else if (offset > entry->offset)
1222                         n = n->rb_right;
1223                 else
1224                         break;
1225         }
1226
1227         if (bitmap_only) {
1228                 if (!entry)
1229                         return NULL;
1230                 if (entry->bitmap)
1231                         return entry;
1232
1233                 /*
1234                  * bitmap entry and extent entry may share same offset,
1235                  * in that case, bitmap entry comes after extent entry.
1236                  */
1237                 n = rb_next(n);
1238                 if (!n)
1239                         return NULL;
1240                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1241                 if (entry->offset != offset)
1242                         return NULL;
1243
1244                 WARN_ON(!entry->bitmap);
1245                 return entry;
1246         } else if (entry) {
1247                 if (entry->bitmap) {
1248                         /*
1249                          * if previous extent entry covers the offset,
1250                          * we should return it instead of the bitmap entry
1251                          */
1252                         n = rb_prev(&entry->offset_index);
1253                         if (n) {
1254                                 prev = rb_entry(n, struct btrfs_free_space,
1255                                                 offset_index);
1256                                 if (!prev->bitmap &&
1257                                     prev->offset + prev->bytes > offset)
1258                                         entry = prev;
1259                         }
1260                 }
1261                 return entry;
1262         }
1263
1264         if (!prev)
1265                 return NULL;
1266
1267         /* find last entry before the 'offset' */
1268         entry = prev;
1269         if (entry->offset > offset) {
1270                 n = rb_prev(&entry->offset_index);
1271                 if (n) {
1272                         entry = rb_entry(n, struct btrfs_free_space,
1273                                         offset_index);
1274                         BUG_ON(entry->offset > offset);
1275                 } else {
1276                         if (fuzzy)
1277                                 return entry;
1278                         else
1279                                 return NULL;
1280                 }
1281         }
1282
1283         if (entry->bitmap) {
1284                 n = rb_prev(&entry->offset_index);
1285                 if (n) {
1286                         prev = rb_entry(n, struct btrfs_free_space,
1287                                         offset_index);
1288                         if (!prev->bitmap &&
1289                             prev->offset + prev->bytes > offset)
1290                                 return prev;
1291                 }
1292                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1293                         return entry;
1294         } else if (entry->offset + entry->bytes > offset)
1295                 return entry;
1296
1297         if (!fuzzy)
1298                 return NULL;
1299
1300         while (1) {
1301                 if (entry->bitmap) {
1302                         if (entry->offset + BITS_PER_BITMAP *
1303                             ctl->unit > offset)
1304                                 break;
1305                 } else {
1306                         if (entry->offset + entry->bytes > offset)
1307                                 break;
1308                 }
1309
1310                 n = rb_next(&entry->offset_index);
1311                 if (!n)
1312                         return NULL;
1313                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1314         }
1315         return entry;
1316 }
1317
1318 static inline void
1319 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1320                     struct btrfs_free_space *info)
1321 {
1322         rb_erase(&info->offset_index, &ctl->free_space_offset);
1323         ctl->free_extents--;
1324 }
1325
1326 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1327                               struct btrfs_free_space *info)
1328 {
1329         __unlink_free_space(ctl, info);
1330         ctl->free_space -= info->bytes;
1331 }
1332
1333 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1334                            struct btrfs_free_space *info)
1335 {
1336         int ret = 0;
1337
1338         BUG_ON(!info->bitmap && !info->bytes);
1339         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1340                                  &info->offset_index, (info->bitmap != NULL));
1341         if (ret)
1342                 return ret;
1343
1344         ctl->free_space += info->bytes;
1345         ctl->free_extents++;
1346         return ret;
1347 }
1348
1349 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1350 {
1351         struct btrfs_block_group_cache *block_group = ctl->private;
1352         u64 max_bytes;
1353         u64 bitmap_bytes;
1354         u64 extent_bytes;
1355         u64 size = block_group->key.offset;
1356         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1357         int max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1358
1359         max_bitmaps = max(max_bitmaps, 1);
1360
1361         BUG_ON(ctl->total_bitmaps > max_bitmaps);
1362
1363         /*
1364          * The goal is to keep the total amount of memory used per 1gb of space
1365          * at or below 32k, so we need to adjust how much memory we allow to be
1366          * used by extent based free space tracking
1367          */
1368         if (size < 1024 * 1024 * 1024)
1369                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1370         else
1371                 max_bytes = MAX_CACHE_BYTES_PER_GIG *
1372                         div64_u64(size, 1024 * 1024 * 1024);
1373
1374         /*
1375          * we want to account for 1 more bitmap than what we have so we can make
1376          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1377          * we add more bitmaps.
1378          */
1379         bitmap_bytes = (ctl->total_bitmaps + 1) * PAGE_CACHE_SIZE;
1380
1381         if (bitmap_bytes >= max_bytes) {
1382                 ctl->extents_thresh = 0;
1383                 return;
1384         }
1385
1386         /*
1387          * we want the extent entry threshold to always be at most 1/2 the maxw
1388          * bytes we can have, or whatever is less than that.
1389          */
1390         extent_bytes = max_bytes - bitmap_bytes;
1391         extent_bytes = min_t(u64, extent_bytes, div64_u64(max_bytes, 2));
1392
1393         ctl->extents_thresh =
1394                 div64_u64(extent_bytes, (sizeof(struct btrfs_free_space)));
1395 }
1396
1397 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1398                                        struct btrfs_free_space *info,
1399                                        u64 offset, u64 bytes)
1400 {
1401         unsigned long start, count;
1402
1403         start = offset_to_bit(info->offset, ctl->unit, offset);
1404         count = bytes_to_bits(bytes, ctl->unit);
1405         BUG_ON(start + count > BITS_PER_BITMAP);
1406
1407         bitmap_clear(info->bitmap, start, count);
1408
1409         info->bytes -= bytes;
1410 }
1411
1412 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1413                               struct btrfs_free_space *info, u64 offset,
1414                               u64 bytes)
1415 {
1416         __bitmap_clear_bits(ctl, info, offset, bytes);
1417         ctl->free_space -= bytes;
1418 }
1419
1420 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1421                             struct btrfs_free_space *info, u64 offset,
1422                             u64 bytes)
1423 {
1424         unsigned long start, count;
1425
1426         start = offset_to_bit(info->offset, ctl->unit, offset);
1427         count = bytes_to_bits(bytes, ctl->unit);
1428         BUG_ON(start + count > BITS_PER_BITMAP);
1429
1430         bitmap_set(info->bitmap, start, count);
1431
1432         info->bytes += bytes;
1433         ctl->free_space += bytes;
1434 }
1435
1436 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1437                          struct btrfs_free_space *bitmap_info, u64 *offset,
1438                          u64 *bytes)
1439 {
1440         unsigned long found_bits = 0;
1441         unsigned long bits, i;
1442         unsigned long next_zero;
1443
1444         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1445                           max_t(u64, *offset, bitmap_info->offset));
1446         bits = bytes_to_bits(*bytes, ctl->unit);
1447
1448         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1449                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1450                                                BITS_PER_BITMAP, i);
1451                 if ((next_zero - i) >= bits) {
1452                         found_bits = next_zero - i;
1453                         break;
1454                 }
1455                 i = next_zero;
1456         }
1457
1458         if (found_bits) {
1459                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1460                 *bytes = (u64)(found_bits) * ctl->unit;
1461                 return 0;
1462         }
1463
1464         return -1;
1465 }
1466
1467 static struct btrfs_free_space *
1468 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1469                 unsigned long align)
1470 {
1471         struct btrfs_free_space *entry;
1472         struct rb_node *node;
1473         u64 ctl_off;
1474         u64 tmp;
1475         u64 align_off;
1476         int ret;
1477
1478         if (!ctl->free_space_offset.rb_node)
1479                 return NULL;
1480
1481         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1482         if (!entry)
1483                 return NULL;
1484
1485         for (node = &entry->offset_index; node; node = rb_next(node)) {
1486                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1487                 if (entry->bytes < *bytes)
1488                         continue;
1489
1490                 /* make sure the space returned is big enough
1491                  * to match our requested alignment
1492                  */
1493                 if (*bytes >= align) {
1494                         ctl_off = entry->offset - ctl->start;
1495                         tmp = ctl_off + align - 1;;
1496                         do_div(tmp, align);
1497                         tmp = tmp * align + ctl->start;
1498                         align_off = tmp - entry->offset;
1499                 } else {
1500                         align_off = 0;
1501                         tmp = entry->offset;
1502                 }
1503
1504                 if (entry->bytes < *bytes + align_off)
1505                         continue;
1506
1507                 if (entry->bitmap) {
1508                         ret = search_bitmap(ctl, entry, &tmp, bytes);
1509                         if (!ret) {
1510                                 *offset = tmp;
1511                                 return entry;
1512                         }
1513                         continue;
1514                 }
1515
1516                 *offset = tmp;
1517                 *bytes = entry->bytes - align_off;
1518                 return entry;
1519         }
1520
1521         return NULL;
1522 }
1523
1524 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1525                            struct btrfs_free_space *info, u64 offset)
1526 {
1527         info->offset = offset_to_bitmap(ctl, offset);
1528         info->bytes = 0;
1529         INIT_LIST_HEAD(&info->list);
1530         link_free_space(ctl, info);
1531         ctl->total_bitmaps++;
1532
1533         ctl->op->recalc_thresholds(ctl);
1534 }
1535
1536 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1537                         struct btrfs_free_space *bitmap_info)
1538 {
1539         unlink_free_space(ctl, bitmap_info);
1540         kfree(bitmap_info->bitmap);
1541         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1542         ctl->total_bitmaps--;
1543         ctl->op->recalc_thresholds(ctl);
1544 }
1545
1546 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1547                               struct btrfs_free_space *bitmap_info,
1548                               u64 *offset, u64 *bytes)
1549 {
1550         u64 end;
1551         u64 search_start, search_bytes;
1552         int ret;
1553
1554 again:
1555         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1556
1557         /*
1558          * We need to search for bits in this bitmap.  We could only cover some
1559          * of the extent in this bitmap thanks to how we add space, so we need
1560          * to search for as much as it as we can and clear that amount, and then
1561          * go searching for the next bit.
1562          */
1563         search_start = *offset;
1564         search_bytes = ctl->unit;
1565         search_bytes = min(search_bytes, end - search_start + 1);
1566         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes);
1567         if (ret < 0 || search_start != *offset)
1568                 return -EINVAL;
1569
1570         /* We may have found more bits than what we need */
1571         search_bytes = min(search_bytes, *bytes);
1572
1573         /* Cannot clear past the end of the bitmap */
1574         search_bytes = min(search_bytes, end - search_start + 1);
1575
1576         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1577         *offset += search_bytes;
1578         *bytes -= search_bytes;
1579
1580         if (*bytes) {
1581                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1582                 if (!bitmap_info->bytes)
1583                         free_bitmap(ctl, bitmap_info);
1584
1585                 /*
1586                  * no entry after this bitmap, but we still have bytes to
1587                  * remove, so something has gone wrong.
1588                  */
1589                 if (!next)
1590                         return -EINVAL;
1591
1592                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1593                                        offset_index);
1594
1595                 /*
1596                  * if the next entry isn't a bitmap we need to return to let the
1597                  * extent stuff do its work.
1598                  */
1599                 if (!bitmap_info->bitmap)
1600                         return -EAGAIN;
1601
1602                 /*
1603                  * Ok the next item is a bitmap, but it may not actually hold
1604                  * the information for the rest of this free space stuff, so
1605                  * look for it, and if we don't find it return so we can try
1606                  * everything over again.
1607                  */
1608                 search_start = *offset;
1609                 search_bytes = ctl->unit;
1610                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1611                                     &search_bytes);
1612                 if (ret < 0 || search_start != *offset)
1613                         return -EAGAIN;
1614
1615                 goto again;
1616         } else if (!bitmap_info->bytes)
1617                 free_bitmap(ctl, bitmap_info);
1618
1619         return 0;
1620 }
1621
1622 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1623                                struct btrfs_free_space *info, u64 offset,
1624                                u64 bytes)
1625 {
1626         u64 bytes_to_set = 0;
1627         u64 end;
1628
1629         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1630
1631         bytes_to_set = min(end - offset, bytes);
1632
1633         bitmap_set_bits(ctl, info, offset, bytes_to_set);
1634
1635         return bytes_to_set;
1636
1637 }
1638
1639 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
1640                       struct btrfs_free_space *info)
1641 {
1642         struct btrfs_block_group_cache *block_group = ctl->private;
1643
1644         /*
1645          * If we are below the extents threshold then we can add this as an
1646          * extent, and don't have to deal with the bitmap
1647          */
1648         if (ctl->free_extents < ctl->extents_thresh) {
1649                 /*
1650                  * If this block group has some small extents we don't want to
1651                  * use up all of our free slots in the cache with them, we want
1652                  * to reserve them to larger extents, however if we have plent
1653                  * of cache left then go ahead an dadd them, no sense in adding
1654                  * the overhead of a bitmap if we don't have to.
1655                  */
1656                 if (info->bytes <= block_group->sectorsize * 4) {
1657                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
1658                                 return false;
1659                 } else {
1660                         return false;
1661                 }
1662         }
1663
1664         /*
1665          * The original block groups from mkfs can be really small, like 8
1666          * megabytes, so don't bother with a bitmap for those entries.  However
1667          * some block groups can be smaller than what a bitmap would cover but
1668          * are still large enough that they could overflow the 32k memory limit,
1669          * so allow those block groups to still be allowed to have a bitmap
1670          * entry.
1671          */
1672         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->key.offset)
1673                 return false;
1674
1675         return true;
1676 }
1677
1678 static struct btrfs_free_space_op free_space_op = {
1679         .recalc_thresholds      = recalculate_thresholds,
1680         .use_bitmap             = use_bitmap,
1681 };
1682
1683 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
1684                               struct btrfs_free_space *info)
1685 {
1686         struct btrfs_free_space *bitmap_info;
1687         struct btrfs_block_group_cache *block_group = NULL;
1688         int added = 0;
1689         u64 bytes, offset, bytes_added;
1690         int ret;
1691
1692         bytes = info->bytes;
1693         offset = info->offset;
1694
1695         if (!ctl->op->use_bitmap(ctl, info))
1696                 return 0;
1697
1698         if (ctl->op == &free_space_op)
1699                 block_group = ctl->private;
1700 again:
1701         /*
1702          * Since we link bitmaps right into the cluster we need to see if we
1703          * have a cluster here, and if so and it has our bitmap we need to add
1704          * the free space to that bitmap.
1705          */
1706         if (block_group && !list_empty(&block_group->cluster_list)) {
1707                 struct btrfs_free_cluster *cluster;
1708                 struct rb_node *node;
1709                 struct btrfs_free_space *entry;
1710
1711                 cluster = list_entry(block_group->cluster_list.next,
1712                                      struct btrfs_free_cluster,
1713                                      block_group_list);
1714                 spin_lock(&cluster->lock);
1715                 node = rb_first(&cluster->root);
1716                 if (!node) {
1717                         spin_unlock(&cluster->lock);
1718                         goto no_cluster_bitmap;
1719                 }
1720
1721                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1722                 if (!entry->bitmap) {
1723                         spin_unlock(&cluster->lock);
1724                         goto no_cluster_bitmap;
1725                 }
1726
1727                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
1728                         bytes_added = add_bytes_to_bitmap(ctl, entry,
1729                                                           offset, bytes);
1730                         bytes -= bytes_added;
1731                         offset += bytes_added;
1732                 }
1733                 spin_unlock(&cluster->lock);
1734                 if (!bytes) {
1735                         ret = 1;
1736                         goto out;
1737                 }
1738         }
1739
1740 no_cluster_bitmap:
1741         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1742                                          1, 0);
1743         if (!bitmap_info) {
1744                 BUG_ON(added);
1745                 goto new_bitmap;
1746         }
1747
1748         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
1749         bytes -= bytes_added;
1750         offset += bytes_added;
1751         added = 0;
1752
1753         if (!bytes) {
1754                 ret = 1;
1755                 goto out;
1756         } else
1757                 goto again;
1758
1759 new_bitmap:
1760         if (info && info->bitmap) {
1761                 add_new_bitmap(ctl, info, offset);
1762                 added = 1;
1763                 info = NULL;
1764                 goto again;
1765         } else {
1766                 spin_unlock(&ctl->tree_lock);
1767
1768                 /* no pre-allocated info, allocate a new one */
1769                 if (!info) {
1770                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
1771                                                  GFP_NOFS);
1772                         if (!info) {
1773                                 spin_lock(&ctl->tree_lock);
1774                                 ret = -ENOMEM;
1775                                 goto out;
1776                         }
1777                 }
1778
1779                 /* allocate the bitmap */
1780                 info->bitmap = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
1781                 spin_lock(&ctl->tree_lock);
1782                 if (!info->bitmap) {
1783                         ret = -ENOMEM;
1784                         goto out;
1785                 }
1786                 goto again;
1787         }
1788
1789 out:
1790         if (info) {
1791                 if (info->bitmap)
1792                         kfree(info->bitmap);
1793                 kmem_cache_free(btrfs_free_space_cachep, info);
1794         }
1795
1796         return ret;
1797 }
1798
1799 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
1800                           struct btrfs_free_space *info, bool update_stat)
1801 {
1802         struct btrfs_free_space *left_info;
1803         struct btrfs_free_space *right_info;
1804         bool merged = false;
1805         u64 offset = info->offset;
1806         u64 bytes = info->bytes;
1807
1808         /*
1809          * first we want to see if there is free space adjacent to the range we
1810          * are adding, if there is remove that struct and add a new one to
1811          * cover the entire range
1812          */
1813         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
1814         if (right_info && rb_prev(&right_info->offset_index))
1815                 left_info = rb_entry(rb_prev(&right_info->offset_index),
1816                                      struct btrfs_free_space, offset_index);
1817         else
1818                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
1819
1820         if (right_info && !right_info->bitmap) {
1821                 if (update_stat)
1822                         unlink_free_space(ctl, right_info);
1823                 else
1824                         __unlink_free_space(ctl, right_info);
1825                 info->bytes += right_info->bytes;
1826                 kmem_cache_free(btrfs_free_space_cachep, right_info);
1827                 merged = true;
1828         }
1829
1830         if (left_info && !left_info->bitmap &&
1831             left_info->offset + left_info->bytes == offset) {
1832                 if (update_stat)
1833                         unlink_free_space(ctl, left_info);
1834                 else
1835                         __unlink_free_space(ctl, left_info);
1836                 info->offset = left_info->offset;
1837                 info->bytes += left_info->bytes;
1838                 kmem_cache_free(btrfs_free_space_cachep, left_info);
1839                 merged = true;
1840         }
1841
1842         return merged;
1843 }
1844
1845 int __btrfs_add_free_space(struct btrfs_free_space_ctl *ctl,
1846                            u64 offset, u64 bytes)
1847 {
1848         struct btrfs_free_space *info;
1849         int ret = 0;
1850
1851         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
1852         if (!info)
1853                 return -ENOMEM;
1854
1855         info->offset = offset;
1856         info->bytes = bytes;
1857
1858         spin_lock(&ctl->tree_lock);
1859
1860         if (try_merge_free_space(ctl, info, true))
1861                 goto link;
1862
1863         /*
1864          * There was no extent directly to the left or right of this new
1865          * extent then we know we're going to have to allocate a new extent, so
1866          * before we do that see if we need to drop this into a bitmap
1867          */
1868         ret = insert_into_bitmap(ctl, info);
1869         if (ret < 0) {
1870                 goto out;
1871         } else if (ret) {
1872                 ret = 0;
1873                 goto out;
1874         }
1875 link:
1876         ret = link_free_space(ctl, info);
1877         if (ret)
1878                 kmem_cache_free(btrfs_free_space_cachep, info);
1879 out:
1880         spin_unlock(&ctl->tree_lock);
1881
1882         if (ret) {
1883                 printk(KERN_CRIT "btrfs: unable to add free space :%d\n", ret);
1884                 BUG_ON(ret == -EEXIST);
1885         }
1886
1887         return ret;
1888 }
1889
1890 int btrfs_remove_free_space(struct btrfs_block_group_cache *block_group,
1891                             u64 offset, u64 bytes)
1892 {
1893         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1894         struct btrfs_free_space *info;
1895         int ret;
1896         bool re_search = false;
1897
1898         spin_lock(&ctl->tree_lock);
1899
1900 again:
1901         ret = 0;
1902         if (!bytes)
1903                 goto out_lock;
1904
1905         info = tree_search_offset(ctl, offset, 0, 0);
1906         if (!info) {
1907                 /*
1908                  * oops didn't find an extent that matched the space we wanted
1909                  * to remove, look for a bitmap instead
1910                  */
1911                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
1912                                           1, 0);
1913                 if (!info) {
1914                         /*
1915                          * If we found a partial bit of our free space in a
1916                          * bitmap but then couldn't find the other part this may
1917                          * be a problem, so WARN about it.
1918                          */
1919                         WARN_ON(re_search);
1920                         goto out_lock;
1921                 }
1922         }
1923
1924         re_search = false;
1925         if (!info->bitmap) {
1926                 unlink_free_space(ctl, info);
1927                 if (offset == info->offset) {
1928                         u64 to_free = min(bytes, info->bytes);
1929
1930                         info->bytes -= to_free;
1931                         info->offset += to_free;
1932                         if (info->bytes) {
1933                                 ret = link_free_space(ctl, info);
1934                                 WARN_ON(ret);
1935                         } else {
1936                                 kmem_cache_free(btrfs_free_space_cachep, info);
1937                         }
1938
1939                         offset += to_free;
1940                         bytes -= to_free;
1941                         goto again;
1942                 } else {
1943                         u64 old_end = info->bytes + info->offset;
1944
1945                         info->bytes = offset - info->offset;
1946                         ret = link_free_space(ctl, info);
1947                         WARN_ON(ret);
1948                         if (ret)
1949                                 goto out_lock;
1950
1951                         /* Not enough bytes in this entry to satisfy us */
1952                         if (old_end < offset + bytes) {
1953                                 bytes -= old_end - offset;
1954                                 offset = old_end;
1955                                 goto again;
1956                         } else if (old_end == offset + bytes) {
1957                                 /* all done */
1958                                 goto out_lock;
1959                         }
1960                         spin_unlock(&ctl->tree_lock);
1961
1962                         ret = btrfs_add_free_space(block_group, offset + bytes,
1963                                                    old_end - (offset + bytes));
1964                         WARN_ON(ret);
1965                         goto out;
1966                 }
1967         }
1968
1969         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
1970         if (ret == -EAGAIN) {
1971                 re_search = true;
1972                 goto again;
1973         }
1974 out_lock:
1975         spin_unlock(&ctl->tree_lock);
1976 out:
1977         return ret;
1978 }
1979
1980 void btrfs_dump_free_space(struct btrfs_block_group_cache *block_group,
1981                            u64 bytes)
1982 {
1983         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1984         struct btrfs_free_space *info;
1985         struct rb_node *n;
1986         int count = 0;
1987
1988         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
1989                 info = rb_entry(n, struct btrfs_free_space, offset_index);
1990                 if (info->bytes >= bytes && !block_group->ro)
1991                         count++;
1992                 printk(KERN_CRIT "entry offset %llu, bytes %llu, bitmap %s\n",
1993                        info->offset, info->bytes,
1994                        (info->bitmap) ? "yes" : "no");
1995         }
1996         printk(KERN_INFO "block group has cluster?: %s\n",
1997                list_empty(&block_group->cluster_list) ? "no" : "yes");
1998         printk(KERN_INFO "%d blocks of free space at or bigger than bytes is"
1999                "\n", count);
2000 }
2001
2002 void btrfs_init_free_space_ctl(struct btrfs_block_group_cache *block_group)
2003 {
2004         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2005
2006         spin_lock_init(&ctl->tree_lock);
2007         ctl->unit = block_group->sectorsize;
2008         ctl->start = block_group->key.objectid;
2009         ctl->private = block_group;
2010         ctl->op = &free_space_op;
2011
2012         /*
2013          * we only want to have 32k of ram per block group for keeping
2014          * track of free space, and if we pass 1/2 of that we want to
2015          * start converting things over to using bitmaps
2016          */
2017         ctl->extents_thresh = ((1024 * 32) / 2) /
2018                                 sizeof(struct btrfs_free_space);
2019 }
2020
2021 /*
2022  * for a given cluster, put all of its extents back into the free
2023  * space cache.  If the block group passed doesn't match the block group
2024  * pointed to by the cluster, someone else raced in and freed the
2025  * cluster already.  In that case, we just return without changing anything
2026  */
2027 static int
2028 __btrfs_return_cluster_to_free_space(
2029                              struct btrfs_block_group_cache *block_group,
2030                              struct btrfs_free_cluster *cluster)
2031 {
2032         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2033         struct btrfs_free_space *entry;
2034         struct rb_node *node;
2035
2036         spin_lock(&cluster->lock);
2037         if (cluster->block_group != block_group)
2038                 goto out;
2039
2040         cluster->block_group = NULL;
2041         cluster->window_start = 0;
2042         list_del_init(&cluster->block_group_list);
2043
2044         node = rb_first(&cluster->root);
2045         while (node) {
2046                 bool bitmap;
2047
2048                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2049                 node = rb_next(&entry->offset_index);
2050                 rb_erase(&entry->offset_index, &cluster->root);
2051
2052                 bitmap = (entry->bitmap != NULL);
2053                 if (!bitmap)
2054                         try_merge_free_space(ctl, entry, false);
2055                 tree_insert_offset(&ctl->free_space_offset,
2056                                    entry->offset, &entry->offset_index, bitmap);
2057         }
2058         cluster->root = RB_ROOT;
2059
2060 out:
2061         spin_unlock(&cluster->lock);
2062         btrfs_put_block_group(block_group);
2063         return 0;
2064 }
2065
2066 static void __btrfs_remove_free_space_cache_locked(
2067                                 struct btrfs_free_space_ctl *ctl)
2068 {
2069         struct btrfs_free_space *info;
2070         struct rb_node *node;
2071
2072         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2073                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2074                 if (!info->bitmap) {
2075                         unlink_free_space(ctl, info);
2076                         kmem_cache_free(btrfs_free_space_cachep, info);
2077                 } else {
2078                         free_bitmap(ctl, info);
2079                 }
2080                 if (need_resched()) {
2081                         spin_unlock(&ctl->tree_lock);
2082                         cond_resched();
2083                         spin_lock(&ctl->tree_lock);
2084                 }
2085         }
2086 }
2087
2088 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2089 {
2090         spin_lock(&ctl->tree_lock);
2091         __btrfs_remove_free_space_cache_locked(ctl);
2092         spin_unlock(&ctl->tree_lock);
2093 }
2094
2095 void btrfs_remove_free_space_cache(struct btrfs_block_group_cache *block_group)
2096 {
2097         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2098         struct btrfs_free_cluster *cluster;
2099         struct list_head *head;
2100
2101         spin_lock(&ctl->tree_lock);
2102         while ((head = block_group->cluster_list.next) !=
2103                &block_group->cluster_list) {
2104                 cluster = list_entry(head, struct btrfs_free_cluster,
2105                                      block_group_list);
2106
2107                 WARN_ON(cluster->block_group != block_group);
2108                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2109                 if (need_resched()) {
2110                         spin_unlock(&ctl->tree_lock);
2111                         cond_resched();
2112                         spin_lock(&ctl->tree_lock);
2113                 }
2114         }
2115         __btrfs_remove_free_space_cache_locked(ctl);
2116         spin_unlock(&ctl->tree_lock);
2117
2118 }
2119
2120 u64 btrfs_find_space_for_alloc(struct btrfs_block_group_cache *block_group,
2121                                u64 offset, u64 bytes, u64 empty_size)
2122 {
2123         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2124         struct btrfs_free_space *entry = NULL;
2125         u64 bytes_search = bytes + empty_size;
2126         u64 ret = 0;
2127         u64 align_gap = 0;
2128         u64 align_gap_len = 0;
2129
2130         spin_lock(&ctl->tree_lock);
2131         entry = find_free_space(ctl, &offset, &bytes_search,
2132                                 block_group->full_stripe_len);
2133         if (!entry)
2134                 goto out;
2135
2136         ret = offset;
2137         if (entry->bitmap) {
2138                 bitmap_clear_bits(ctl, entry, offset, bytes);
2139                 if (!entry->bytes)
2140                         free_bitmap(ctl, entry);
2141         } else {
2142
2143                 unlink_free_space(ctl, entry);
2144                 align_gap_len = offset - entry->offset;
2145                 align_gap = entry->offset;
2146
2147                 entry->offset = offset + bytes;
2148                 WARN_ON(entry->bytes < bytes + align_gap_len);
2149
2150                 entry->bytes -= bytes + align_gap_len;
2151                 if (!entry->bytes)
2152                         kmem_cache_free(btrfs_free_space_cachep, entry);
2153                 else
2154                         link_free_space(ctl, entry);
2155         }
2156
2157 out:
2158         spin_unlock(&ctl->tree_lock);
2159
2160         if (align_gap_len)
2161                 __btrfs_add_free_space(ctl, align_gap, align_gap_len);
2162         return ret;
2163 }
2164
2165 /*
2166  * given a cluster, put all of its extents back into the free space
2167  * cache.  If a block group is passed, this function will only free
2168  * a cluster that belongs to the passed block group.
2169  *
2170  * Otherwise, it'll get a reference on the block group pointed to by the
2171  * cluster and remove the cluster from it.
2172  */
2173 int btrfs_return_cluster_to_free_space(
2174                                struct btrfs_block_group_cache *block_group,
2175                                struct btrfs_free_cluster *cluster)
2176 {
2177         struct btrfs_free_space_ctl *ctl;
2178         int ret;
2179
2180         /* first, get a safe pointer to the block group */
2181         spin_lock(&cluster->lock);
2182         if (!block_group) {
2183                 block_group = cluster->block_group;
2184                 if (!block_group) {
2185                         spin_unlock(&cluster->lock);
2186                         return 0;
2187                 }
2188         } else if (cluster->block_group != block_group) {
2189                 /* someone else has already freed it don't redo their work */
2190                 spin_unlock(&cluster->lock);
2191                 return 0;
2192         }
2193         atomic_inc(&block_group->count);
2194         spin_unlock(&cluster->lock);
2195
2196         ctl = block_group->free_space_ctl;
2197
2198         /* now return any extents the cluster had on it */
2199         spin_lock(&ctl->tree_lock);
2200         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2201         spin_unlock(&ctl->tree_lock);
2202
2203         /* finally drop our ref */
2204         btrfs_put_block_group(block_group);
2205         return ret;
2206 }
2207
2208 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group_cache *block_group,
2209                                    struct btrfs_free_cluster *cluster,
2210                                    struct btrfs_free_space *entry,
2211                                    u64 bytes, u64 min_start)
2212 {
2213         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2214         int err;
2215         u64 search_start = cluster->window_start;
2216         u64 search_bytes = bytes;
2217         u64 ret = 0;
2218
2219         search_start = min_start;
2220         search_bytes = bytes;
2221
2222         err = search_bitmap(ctl, entry, &search_start, &search_bytes);
2223         if (err)
2224                 return 0;
2225
2226         ret = search_start;
2227         __bitmap_clear_bits(ctl, entry, ret, bytes);
2228
2229         return ret;
2230 }
2231
2232 /*
2233  * given a cluster, try to allocate 'bytes' from it, returns 0
2234  * if it couldn't find anything suitably large, or a logical disk offset
2235  * if things worked out
2236  */
2237 u64 btrfs_alloc_from_cluster(struct btrfs_block_group_cache *block_group,
2238                              struct btrfs_free_cluster *cluster, u64 bytes,
2239                              u64 min_start)
2240 {
2241         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2242         struct btrfs_free_space *entry = NULL;
2243         struct rb_node *node;
2244         u64 ret = 0;
2245
2246         spin_lock(&cluster->lock);
2247         if (bytes > cluster->max_size)
2248                 goto out;
2249
2250         if (cluster->block_group != block_group)
2251                 goto out;
2252
2253         node = rb_first(&cluster->root);
2254         if (!node)
2255                 goto out;
2256
2257         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2258         while(1) {
2259                 if (entry->bytes < bytes ||
2260                     (!entry->bitmap && entry->offset < min_start)) {
2261                         node = rb_next(&entry->offset_index);
2262                         if (!node)
2263                                 break;
2264                         entry = rb_entry(node, struct btrfs_free_space,
2265                                          offset_index);
2266                         continue;
2267                 }
2268
2269                 if (entry->bitmap) {
2270                         ret = btrfs_alloc_from_bitmap(block_group,
2271                                                       cluster, entry, bytes,
2272                                                       cluster->window_start);
2273                         if (ret == 0) {
2274                                 node = rb_next(&entry->offset_index);
2275                                 if (!node)
2276                                         break;
2277                                 entry = rb_entry(node, struct btrfs_free_space,
2278                                                  offset_index);
2279                                 continue;
2280                         }
2281                         cluster->window_start += bytes;
2282                 } else {
2283                         ret = entry->offset;
2284
2285                         entry->offset += bytes;
2286                         entry->bytes -= bytes;
2287                 }
2288
2289                 if (entry->bytes == 0)
2290                         rb_erase(&entry->offset_index, &cluster->root);
2291                 break;
2292         }
2293 out:
2294         spin_unlock(&cluster->lock);
2295
2296         if (!ret)
2297                 return 0;
2298
2299         spin_lock(&ctl->tree_lock);
2300
2301         ctl->free_space -= bytes;
2302         if (entry->bytes == 0) {
2303                 ctl->free_extents--;
2304                 if (entry->bitmap) {
2305                         kfree(entry->bitmap);
2306                         ctl->total_bitmaps--;
2307                         ctl->op->recalc_thresholds(ctl);
2308                 }
2309                 kmem_cache_free(btrfs_free_space_cachep, entry);
2310         }
2311
2312         spin_unlock(&ctl->tree_lock);
2313
2314         return ret;
2315 }
2316
2317 static int btrfs_bitmap_cluster(struct btrfs_block_group_cache *block_group,
2318                                 struct btrfs_free_space *entry,
2319                                 struct btrfs_free_cluster *cluster,
2320                                 u64 offset, u64 bytes,
2321                                 u64 cont1_bytes, u64 min_bytes)
2322 {
2323         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2324         unsigned long next_zero;
2325         unsigned long i;
2326         unsigned long want_bits;
2327         unsigned long min_bits;
2328         unsigned long found_bits;
2329         unsigned long start = 0;
2330         unsigned long total_found = 0;
2331         int ret;
2332
2333         i = offset_to_bit(entry->offset, ctl->unit,
2334                           max_t(u64, offset, entry->offset));
2335         want_bits = bytes_to_bits(bytes, ctl->unit);
2336         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2337
2338 again:
2339         found_bits = 0;
2340         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2341                 next_zero = find_next_zero_bit(entry->bitmap,
2342                                                BITS_PER_BITMAP, i);
2343                 if (next_zero - i >= min_bits) {
2344                         found_bits = next_zero - i;
2345                         break;
2346                 }
2347                 i = next_zero;
2348         }
2349
2350         if (!found_bits)
2351                 return -ENOSPC;
2352
2353         if (!total_found) {
2354                 start = i;
2355                 cluster->max_size = 0;
2356         }
2357
2358         total_found += found_bits;
2359
2360         if (cluster->max_size < found_bits * ctl->unit)
2361                 cluster->max_size = found_bits * ctl->unit;
2362
2363         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
2364                 i = next_zero + 1;
2365                 goto again;
2366         }
2367
2368         cluster->window_start = start * ctl->unit + entry->offset;
2369         rb_erase(&entry->offset_index, &ctl->free_space_offset);
2370         ret = tree_insert_offset(&cluster->root, entry->offset,
2371                                  &entry->offset_index, 1);
2372         BUG_ON(ret); /* -EEXIST; Logic error */
2373
2374         trace_btrfs_setup_cluster(block_group, cluster,
2375                                   total_found * ctl->unit, 1);
2376         return 0;
2377 }
2378
2379 /*
2380  * This searches the block group for just extents to fill the cluster with.
2381  * Try to find a cluster with at least bytes total bytes, at least one
2382  * extent of cont1_bytes, and other clusters of at least min_bytes.
2383  */
2384 static noinline int
2385 setup_cluster_no_bitmap(struct btrfs_block_group_cache *block_group,
2386                         struct btrfs_free_cluster *cluster,
2387                         struct list_head *bitmaps, u64 offset, u64 bytes,
2388                         u64 cont1_bytes, u64 min_bytes)
2389 {
2390         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2391         struct btrfs_free_space *first = NULL;
2392         struct btrfs_free_space *entry = NULL;
2393         struct btrfs_free_space *last;
2394         struct rb_node *node;
2395         u64 window_start;
2396         u64 window_free;
2397         u64 max_extent;
2398         u64 total_size = 0;
2399
2400         entry = tree_search_offset(ctl, offset, 0, 1);
2401         if (!entry)
2402                 return -ENOSPC;
2403
2404         /*
2405          * We don't want bitmaps, so just move along until we find a normal
2406          * extent entry.
2407          */
2408         while (entry->bitmap || entry->bytes < min_bytes) {
2409                 if (entry->bitmap && list_empty(&entry->list))
2410                         list_add_tail(&entry->list, bitmaps);
2411                 node = rb_next(&entry->offset_index);
2412                 if (!node)
2413                         return -ENOSPC;
2414                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2415         }
2416
2417         window_start = entry->offset;
2418         window_free = entry->bytes;
2419         max_extent = entry->bytes;
2420         first = entry;
2421         last = entry;
2422
2423         for (node = rb_next(&entry->offset_index); node;
2424              node = rb_next(&entry->offset_index)) {
2425                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2426
2427                 if (entry->bitmap) {
2428                         if (list_empty(&entry->list))
2429                                 list_add_tail(&entry->list, bitmaps);
2430                         continue;
2431                 }
2432
2433                 if (entry->bytes < min_bytes)
2434                         continue;
2435
2436                 last = entry;
2437                 window_free += entry->bytes;
2438                 if (entry->bytes > max_extent)
2439                         max_extent = entry->bytes;
2440         }
2441
2442         if (window_free < bytes || max_extent < cont1_bytes)
2443                 return -ENOSPC;
2444
2445         cluster->window_start = first->offset;
2446
2447         node = &first->offset_index;
2448
2449         /*
2450          * now we've found our entries, pull them out of the free space
2451          * cache and put them into the cluster rbtree
2452          */
2453         do {
2454                 int ret;
2455
2456                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2457                 node = rb_next(&entry->offset_index);
2458                 if (entry->bitmap || entry->bytes < min_bytes)
2459                         continue;
2460
2461                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
2462                 ret = tree_insert_offset(&cluster->root, entry->offset,
2463                                          &entry->offset_index, 0);
2464                 total_size += entry->bytes;
2465                 BUG_ON(ret); /* -EEXIST; Logic error */
2466         } while (node && entry != last);
2467
2468         cluster->max_size = max_extent;
2469         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
2470         return 0;
2471 }
2472
2473 /*
2474  * This specifically looks for bitmaps that may work in the cluster, we assume
2475  * that we have already failed to find extents that will work.
2476  */
2477 static noinline int
2478 setup_cluster_bitmap(struct btrfs_block_group_cache *block_group,
2479                      struct btrfs_free_cluster *cluster,
2480                      struct list_head *bitmaps, u64 offset, u64 bytes,
2481                      u64 cont1_bytes, u64 min_bytes)
2482 {
2483         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2484         struct btrfs_free_space *entry;
2485         int ret = -ENOSPC;
2486         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
2487
2488         if (ctl->total_bitmaps == 0)
2489                 return -ENOSPC;
2490
2491         /*
2492          * The bitmap that covers offset won't be in the list unless offset
2493          * is just its start offset.
2494          */
2495         entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
2496         if (entry->offset != bitmap_offset) {
2497                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
2498                 if (entry && list_empty(&entry->list))
2499                         list_add(&entry->list, bitmaps);
2500         }
2501
2502         list_for_each_entry(entry, bitmaps, list) {
2503                 if (entry->bytes < bytes)
2504                         continue;
2505                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
2506                                            bytes, cont1_bytes, min_bytes);
2507                 if (!ret)
2508                         return 0;
2509         }
2510
2511         /*
2512          * The bitmaps list has all the bitmaps that record free space
2513          * starting after offset, so no more search is required.
2514          */
2515         return -ENOSPC;
2516 }
2517
2518 /*
2519  * here we try to find a cluster of blocks in a block group.  The goal
2520  * is to find at least bytes+empty_size.
2521  * We might not find them all in one contiguous area.
2522  *
2523  * returns zero and sets up cluster if things worked out, otherwise
2524  * it returns -enospc
2525  */
2526 int btrfs_find_space_cluster(struct btrfs_root *root,
2527                              struct btrfs_block_group_cache *block_group,
2528                              struct btrfs_free_cluster *cluster,
2529                              u64 offset, u64 bytes, u64 empty_size)
2530 {
2531         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2532         struct btrfs_free_space *entry, *tmp;
2533         LIST_HEAD(bitmaps);
2534         u64 min_bytes;
2535         u64 cont1_bytes;
2536         int ret;
2537
2538         /*
2539          * Choose the minimum extent size we'll require for this
2540          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
2541          * For metadata, allow allocates with smaller extents.  For
2542          * data, keep it dense.
2543          */
2544         if (btrfs_test_opt(root, SSD_SPREAD)) {
2545                 cont1_bytes = min_bytes = bytes + empty_size;
2546         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
2547                 cont1_bytes = bytes;
2548                 min_bytes = block_group->sectorsize;
2549         } else {
2550                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
2551                 min_bytes = block_group->sectorsize;
2552         }
2553
2554         spin_lock(&ctl->tree_lock);
2555
2556         /*
2557          * If we know we don't have enough space to make a cluster don't even
2558          * bother doing all the work to try and find one.
2559          */
2560         if (ctl->free_space < bytes) {
2561                 spin_unlock(&ctl->tree_lock);
2562                 return -ENOSPC;
2563         }
2564
2565         spin_lock(&cluster->lock);
2566
2567         /* someone already found a cluster, hooray */
2568         if (cluster->block_group) {
2569                 ret = 0;
2570                 goto out;
2571         }
2572
2573         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
2574                                  min_bytes);
2575
2576         INIT_LIST_HEAD(&bitmaps);
2577         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
2578                                       bytes + empty_size,
2579                                       cont1_bytes, min_bytes);
2580         if (ret)
2581                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
2582                                            offset, bytes + empty_size,
2583                                            cont1_bytes, min_bytes);
2584
2585         /* Clear our temporary list */
2586         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
2587                 list_del_init(&entry->list);
2588
2589         if (!ret) {
2590                 atomic_inc(&block_group->count);
2591                 list_add_tail(&cluster->block_group_list,
2592                               &block_group->cluster_list);
2593                 cluster->block_group = block_group;
2594         } else {
2595                 trace_btrfs_failed_cluster_setup(block_group);
2596         }
2597 out:
2598         spin_unlock(&cluster->lock);
2599         spin_unlock(&ctl->tree_lock);
2600
2601         return ret;
2602 }
2603
2604 /*
2605  * simple code to zero out a cluster
2606  */
2607 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
2608 {
2609         spin_lock_init(&cluster->lock);
2610         spin_lock_init(&cluster->refill_lock);
2611         cluster->root = RB_ROOT;
2612         cluster->max_size = 0;
2613         INIT_LIST_HEAD(&cluster->block_group_list);
2614         cluster->block_group = NULL;
2615 }
2616
2617 static int do_trimming(struct btrfs_block_group_cache *block_group,
2618                        u64 *total_trimmed, u64 start, u64 bytes,
2619                        u64 reserved_start, u64 reserved_bytes)
2620 {
2621         struct btrfs_space_info *space_info = block_group->space_info;
2622         struct btrfs_fs_info *fs_info = block_group->fs_info;
2623         int ret;
2624         int update = 0;
2625         u64 trimmed = 0;
2626
2627         spin_lock(&space_info->lock);
2628         spin_lock(&block_group->lock);
2629         if (!block_group->ro) {
2630                 block_group->reserved += reserved_bytes;
2631                 space_info->bytes_reserved += reserved_bytes;
2632                 update = 1;
2633         }
2634         spin_unlock(&block_group->lock);
2635         spin_unlock(&space_info->lock);
2636
2637         ret = btrfs_error_discard_extent(fs_info->extent_root,
2638                                          start, bytes, &trimmed);
2639         if (!ret)
2640                 *total_trimmed += trimmed;
2641
2642         btrfs_add_free_space(block_group, reserved_start, reserved_bytes);
2643
2644         if (update) {
2645                 spin_lock(&space_info->lock);
2646                 spin_lock(&block_group->lock);
2647                 if (block_group->ro)
2648                         space_info->bytes_readonly += reserved_bytes;
2649                 block_group->reserved -= reserved_bytes;
2650                 space_info->bytes_reserved -= reserved_bytes;
2651                 spin_unlock(&space_info->lock);
2652                 spin_unlock(&block_group->lock);
2653         }
2654
2655         return ret;
2656 }
2657
2658 static int trim_no_bitmap(struct btrfs_block_group_cache *block_group,
2659                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
2660 {
2661         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2662         struct btrfs_free_space *entry;
2663         struct rb_node *node;
2664         int ret = 0;
2665         u64 extent_start;
2666         u64 extent_bytes;
2667         u64 bytes;
2668
2669         while (start < end) {
2670                 spin_lock(&ctl->tree_lock);
2671
2672                 if (ctl->free_space < minlen) {
2673                         spin_unlock(&ctl->tree_lock);
2674                         break;
2675                 }
2676
2677                 entry = tree_search_offset(ctl, start, 0, 1);
2678                 if (!entry) {
2679                         spin_unlock(&ctl->tree_lock);
2680                         break;
2681                 }
2682
2683                 /* skip bitmaps */
2684                 while (entry->bitmap) {
2685                         node = rb_next(&entry->offset_index);
2686                         if (!node) {
2687                                 spin_unlock(&ctl->tree_lock);
2688                                 goto out;
2689                         }
2690                         entry = rb_entry(node, struct btrfs_free_space,
2691                                          offset_index);
2692                 }
2693
2694                 if (entry->offset >= end) {
2695                         spin_unlock(&ctl->tree_lock);
2696                         break;
2697                 }
2698
2699                 extent_start = entry->offset;
2700                 extent_bytes = entry->bytes;
2701                 start = max(start, extent_start);
2702                 bytes = min(extent_start + extent_bytes, end) - start;
2703                 if (bytes < minlen) {
2704                         spin_unlock(&ctl->tree_lock);
2705                         goto next;
2706                 }
2707
2708                 unlink_free_space(ctl, entry);
2709                 kmem_cache_free(btrfs_free_space_cachep, entry);
2710
2711                 spin_unlock(&ctl->tree_lock);
2712
2713                 ret = do_trimming(block_group, total_trimmed, start, bytes,
2714                                   extent_start, extent_bytes);
2715                 if (ret)
2716                         break;
2717 next:
2718                 start += bytes;
2719
2720                 if (fatal_signal_pending(current)) {
2721                         ret = -ERESTARTSYS;
2722                         break;
2723                 }
2724
2725                 cond_resched();
2726         }
2727 out:
2728         return ret;
2729 }
2730
2731 static int trim_bitmaps(struct btrfs_block_group_cache *block_group,
2732                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
2733 {
2734         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2735         struct btrfs_free_space *entry;
2736         int ret = 0;
2737         int ret2;
2738         u64 bytes;
2739         u64 offset = offset_to_bitmap(ctl, start);
2740
2741         while (offset < end) {
2742                 bool next_bitmap = false;
2743
2744                 spin_lock(&ctl->tree_lock);
2745
2746                 if (ctl->free_space < minlen) {
2747                         spin_unlock(&ctl->tree_lock);
2748                         break;
2749                 }
2750
2751                 entry = tree_search_offset(ctl, offset, 1, 0);
2752                 if (!entry) {
2753                         spin_unlock(&ctl->tree_lock);
2754                         next_bitmap = true;
2755                         goto next;
2756                 }
2757
2758                 bytes = minlen;
2759                 ret2 = search_bitmap(ctl, entry, &start, &bytes);
2760                 if (ret2 || start >= end) {
2761                         spin_unlock(&ctl->tree_lock);
2762                         next_bitmap = true;
2763                         goto next;
2764                 }
2765
2766                 bytes = min(bytes, end - start);
2767                 if (bytes < minlen) {
2768                         spin_unlock(&ctl->tree_lock);
2769                         goto next;
2770                 }
2771
2772                 bitmap_clear_bits(ctl, entry, start, bytes);
2773                 if (entry->bytes == 0)
2774                         free_bitmap(ctl, entry);
2775
2776                 spin_unlock(&ctl->tree_lock);
2777
2778                 ret = do_trimming(block_group, total_trimmed, start, bytes,
2779                                   start, bytes);
2780                 if (ret)
2781                         break;
2782 next:
2783                 if (next_bitmap) {
2784                         offset += BITS_PER_BITMAP * ctl->unit;
2785                 } else {
2786                         start += bytes;
2787                         if (start >= offset + BITS_PER_BITMAP * ctl->unit)
2788                                 offset += BITS_PER_BITMAP * ctl->unit;
2789                 }
2790
2791                 if (fatal_signal_pending(current)) {
2792                         ret = -ERESTARTSYS;
2793                         break;
2794                 }
2795
2796                 cond_resched();
2797         }
2798
2799         return ret;
2800 }
2801
2802 int btrfs_trim_block_group(struct btrfs_block_group_cache *block_group,
2803                            u64 *trimmed, u64 start, u64 end, u64 minlen)
2804 {
2805         int ret;
2806
2807         *trimmed = 0;
2808
2809         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
2810         if (ret)
2811                 return ret;
2812
2813         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
2814
2815         return ret;
2816 }
2817
2818 /*
2819  * Find the left-most item in the cache tree, and then return the
2820  * smallest inode number in the item.
2821  *
2822  * Note: the returned inode number may not be the smallest one in
2823  * the tree, if the left-most item is a bitmap.
2824  */
2825 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
2826 {
2827         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
2828         struct btrfs_free_space *entry = NULL;
2829         u64 ino = 0;
2830
2831         spin_lock(&ctl->tree_lock);
2832
2833         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
2834                 goto out;
2835
2836         entry = rb_entry(rb_first(&ctl->free_space_offset),
2837                          struct btrfs_free_space, offset_index);
2838
2839         if (!entry->bitmap) {
2840                 ino = entry->offset;
2841
2842                 unlink_free_space(ctl, entry);
2843                 entry->offset++;
2844                 entry->bytes--;
2845                 if (!entry->bytes)
2846                         kmem_cache_free(btrfs_free_space_cachep, entry);
2847                 else
2848                         link_free_space(ctl, entry);
2849         } else {
2850                 u64 offset = 0;
2851                 u64 count = 1;
2852                 int ret;
2853
2854                 ret = search_bitmap(ctl, entry, &offset, &count);
2855                 /* Logic error; Should be empty if it can't find anything */
2856                 BUG_ON(ret);
2857
2858                 ino = offset;
2859                 bitmap_clear_bits(ctl, entry, offset, 1);
2860                 if (entry->bytes == 0)
2861                         free_bitmap(ctl, entry);
2862         }
2863 out:
2864         spin_unlock(&ctl->tree_lock);
2865
2866         return ino;
2867 }
2868
2869 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
2870                                     struct btrfs_path *path)
2871 {
2872         struct inode *inode = NULL;
2873
2874         spin_lock(&root->cache_lock);
2875         if (root->cache_inode)
2876                 inode = igrab(root->cache_inode);
2877         spin_unlock(&root->cache_lock);
2878         if (inode)
2879                 return inode;
2880
2881         inode = __lookup_free_space_inode(root, path, 0);
2882         if (IS_ERR(inode))
2883                 return inode;
2884
2885         spin_lock(&root->cache_lock);
2886         if (!btrfs_fs_closing(root->fs_info))
2887                 root->cache_inode = igrab(inode);
2888         spin_unlock(&root->cache_lock);
2889
2890         return inode;
2891 }
2892
2893 int create_free_ino_inode(struct btrfs_root *root,
2894                           struct btrfs_trans_handle *trans,
2895                           struct btrfs_path *path)
2896 {
2897         return __create_free_space_inode(root, trans, path,
2898                                          BTRFS_FREE_INO_OBJECTID, 0);
2899 }
2900
2901 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
2902 {
2903         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2904         struct btrfs_path *path;
2905         struct inode *inode;
2906         int ret = 0;
2907         u64 root_gen = btrfs_root_generation(&root->root_item);
2908
2909         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2910                 return 0;
2911
2912         /*
2913          * If we're unmounting then just return, since this does a search on the
2914          * normal root and not the commit root and we could deadlock.
2915          */
2916         if (btrfs_fs_closing(fs_info))
2917                 return 0;
2918
2919         path = btrfs_alloc_path();
2920         if (!path)
2921                 return 0;
2922
2923         inode = lookup_free_ino_inode(root, path);
2924         if (IS_ERR(inode))
2925                 goto out;
2926
2927         if (root_gen != BTRFS_I(inode)->generation)
2928                 goto out_put;
2929
2930         ret = __load_free_space_cache(root, inode, ctl, path, 0);
2931
2932         if (ret < 0)
2933                 btrfs_err(fs_info,
2934                         "failed to load free ino cache for root %llu",
2935                         root->root_key.objectid);
2936 out_put:
2937         iput(inode);
2938 out:
2939         btrfs_free_path(path);
2940         return ret;
2941 }
2942
2943 int btrfs_write_out_ino_cache(struct btrfs_root *root,
2944                               struct btrfs_trans_handle *trans,
2945                               struct btrfs_path *path)
2946 {
2947         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
2948         struct inode *inode;
2949         int ret;
2950
2951         if (!btrfs_test_opt(root, INODE_MAP_CACHE))
2952                 return 0;
2953
2954         inode = lookup_free_ino_inode(root, path);
2955         if (IS_ERR(inode))
2956                 return 0;
2957
2958         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, trans, path, 0);
2959         if (ret) {
2960                 btrfs_delalloc_release_metadata(inode, inode->i_size);
2961 #ifdef DEBUG
2962                 btrfs_err(root->fs_info,
2963                         "failed to write free ino cache for root %llu",
2964                         root->root_key.objectid);
2965 #endif
2966         }
2967
2968         iput(inode);
2969         return ret;
2970 }
2971
2972 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
2973 /*
2974  * Use this if you need to make a bitmap or extent entry specifically, it
2975  * doesn't do any of the merging that add_free_space does, this acts a lot like
2976  * how the free space cache loading stuff works, so you can get really weird
2977  * configurations.
2978  */
2979 int test_add_free_space_entry(struct btrfs_block_group_cache *cache,
2980                               u64 offset, u64 bytes, bool bitmap)
2981 {
2982         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
2983         struct btrfs_free_space *info = NULL, *bitmap_info;
2984         void *map = NULL;
2985         u64 bytes_added;
2986         int ret;
2987
2988 again:
2989         if (!info) {
2990                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2991                 if (!info)
2992                         return -ENOMEM;
2993         }
2994
2995         if (!bitmap) {
2996                 spin_lock(&ctl->tree_lock);
2997                 info->offset = offset;
2998                 info->bytes = bytes;
2999                 ret = link_free_space(ctl, info);
3000                 spin_unlock(&ctl->tree_lock);
3001                 if (ret)
3002                         kmem_cache_free(btrfs_free_space_cachep, info);
3003                 return ret;
3004         }
3005
3006         if (!map) {
3007                 map = kzalloc(PAGE_CACHE_SIZE, GFP_NOFS);
3008                 if (!map) {
3009                         kmem_cache_free(btrfs_free_space_cachep, info);
3010                         return -ENOMEM;
3011                 }
3012         }
3013
3014         spin_lock(&ctl->tree_lock);
3015         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3016                                          1, 0);
3017         if (!bitmap_info) {
3018                 info->bitmap = map;
3019                 map = NULL;
3020                 add_new_bitmap(ctl, info, offset);
3021                 bitmap_info = info;
3022         }
3023
3024         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes);
3025         bytes -= bytes_added;
3026         offset += bytes_added;
3027         spin_unlock(&ctl->tree_lock);
3028
3029         if (bytes)
3030                 goto again;
3031
3032         if (map)
3033                 kfree(map);
3034         return 0;
3035 }
3036
3037 /*
3038  * Checks to see if the given range is in the free space cache.  This is really
3039  * just used to check the absence of space, so if there is free space in the
3040  * range at all we will return 1.
3041  */
3042 int test_check_exists(struct btrfs_block_group_cache *cache,
3043                       u64 offset, u64 bytes)
3044 {
3045         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3046         struct btrfs_free_space *info;
3047         int ret = 0;
3048
3049         spin_lock(&ctl->tree_lock);
3050         info = tree_search_offset(ctl, offset, 0, 0);
3051         if (!info) {
3052                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3053                                           1, 0);
3054                 if (!info)
3055                         goto out;
3056         }
3057
3058 have_info:
3059         if (info->bitmap) {
3060                 u64 bit_off, bit_bytes;
3061                 struct rb_node *n;
3062                 struct btrfs_free_space *tmp;
3063
3064                 bit_off = offset;
3065                 bit_bytes = ctl->unit;
3066                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes);
3067                 if (!ret) {
3068                         if (bit_off == offset) {
3069                                 ret = 1;
3070                                 goto out;
3071                         } else if (bit_off > offset &&
3072                                    offset + bytes > bit_off) {
3073                                 ret = 1;
3074                                 goto out;
3075                         }
3076                 }
3077
3078                 n = rb_prev(&info->offset_index);
3079                 while (n) {
3080                         tmp = rb_entry(n, struct btrfs_free_space,
3081                                        offset_index);
3082                         if (tmp->offset + tmp->bytes < offset)
3083                                 break;
3084                         if (offset + bytes < tmp->offset) {
3085                                 n = rb_prev(&info->offset_index);
3086                                 continue;
3087                         }
3088                         info = tmp;
3089                         goto have_info;
3090                 }
3091
3092                 n = rb_next(&info->offset_index);
3093                 while (n) {
3094                         tmp = rb_entry(n, struct btrfs_free_space,
3095                                        offset_index);
3096                         if (offset + bytes < tmp->offset)
3097                                 break;
3098                         if (tmp->offset + tmp->bytes < offset) {
3099                                 n = rb_next(&info->offset_index);
3100                                 continue;
3101                         }
3102                         info = tmp;
3103                         goto have_info;
3104                 }
3105
3106                 goto out;
3107         }
3108
3109         if (info->offset == offset) {
3110                 ret = 1;
3111                 goto out;
3112         }
3113
3114         if (offset > info->offset && offset < info->offset + info->bytes)
3115                 ret = 1;
3116 out:
3117         spin_unlock(&ctl->tree_lock);
3118         return ret;
3119 }
3120 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */