fs/btrfs/free-space-cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2008 Red Hat.  All rights reserved.
   4  */
   5
   6 #include <linux/pagemap.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/slab.h>
  10 #include <linux/math64.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/error-injection.h>
  13 #include <linux/sched/mm.h>
  14 #include "ctree.h"
  15 #include "free-space-cache.h"
  16 #include "transaction.h"
  17 #include "disk-io.h"
  18 #include "extent_io.h"
  19 #include "inode-map.h"
  20 #include "volumes.h"
  21 #include "space-info.h"
  22 #include "delalloc-space.h"
  23 #include "block-group.h"
  24 #include "discard.h"
  25
  26 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  27 #define MAX_CACHE_BYTES_PER_GIG SZ_32K
  28
  29 struct btrfs_trim_range {
  30         u64 start;
  31         u64 bytes;
  32         struct list_head list;
  33 };
  34
  35 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  36                            struct btrfs_free_space *info);
  37 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  38                               struct btrfs_free_space *info);
  39 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
  40                              struct btrfs_trans_handle *trans,
  41                              struct btrfs_io_ctl *io_ctl,
  42                              struct btrfs_path *path);
  43
  44 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  45                                                struct btrfs_path *path,
  46                                                u64 offset)
  47 {
  48         struct btrfs_fs_info *fs_info = root->fs_info;
  49         struct btrfs_key key;
  50         struct btrfs_key location;
  51         struct btrfs_disk_key disk_key;
  52         struct btrfs_free_space_header *header;
  53         struct extent_buffer *leaf;
  54         struct inode *inode = NULL;
  55         unsigned nofs_flag;
  56         int ret;
  57
  58         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  59         key.offset = offset;
  60         key.type = 0;
  61
  62         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  63         if (ret < 0)
  64                 return ERR_PTR(ret);
  65         if (ret > 0) {
  66                 btrfs_release_path(path);
  67                 return ERR_PTR(-ENOENT);
  68         }
  69
  70         leaf = path->nodes[0];
  71         header = btrfs_item_ptr(leaf, path->slots[0],
  72                                 struct btrfs_free_space_header);
  73         btrfs_free_space_key(leaf, header, &disk_key);
  74         btrfs_disk_key_to_cpu(&location, &disk_key);
  75         btrfs_release_path(path);
  76
  77         /*
  78          * We are often under a trans handle at this point, so we need to make
  79          * sure NOFS is set to keep us from deadlocking.
  80          */
  81         nofs_flag = memalloc_nofs_save();
  82         inode = btrfs_iget_path(fs_info->sb, &location, root, path);
  83         btrfs_release_path(path);
  84         memalloc_nofs_restore(nofs_flag);
  85         if (IS_ERR(inode))
  86                 return inode;
  87
  88         mapping_set_gfp_mask(inode->i_mapping,
  89                         mapping_gfp_constraint(inode->i_mapping,
  90                         ~(__GFP_FS | __GFP_HIGHMEM)));
  91
  92         return inode;
  93 }
  94
  95 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
  96                 struct btrfs_path *path)
  97 {
  98         struct btrfs_fs_info *fs_info = block_group->fs_info;
  99         struct inode *inode = NULL;
 100         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 101
 102         spin_lock(&block_group->lock);
 103         if (block_group->inode)
 104                 inode = igrab(block_group->inode);
 105         spin_unlock(&block_group->lock);
 106         if (inode)
 107                 return inode;
 108
 109         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 110                                           block_group->start);
 111         if (IS_ERR(inode))
 112                 return inode;
 113
 114         spin_lock(&block_group->lock);
 115         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 116                 btrfs_info(fs_info, "Old style space inode found, converting.");
 117                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 118                         BTRFS_INODE_NODATACOW;
 119                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 120         }
 121
 122         if (!block_group->iref) {
 123                 block_group->inode = igrab(inode);
 124                 block_group->iref = 1;
 125         }
 126         spin_unlock(&block_group->lock);
 127
 128         return inode;
 129 }
 130
 131 static int __create_free_space_inode(struct btrfs_root *root,
 132                                      struct btrfs_trans_handle *trans,
 133                                      struct btrfs_path *path,
 134                                      u64 ino, u64 offset)
 135 {
 136         struct btrfs_key key;
 137         struct btrfs_disk_key disk_key;
 138         struct btrfs_free_space_header *header;
 139         struct btrfs_inode_item *inode_item;
 140         struct extent_buffer *leaf;
 141         u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC;
 142         int ret;
 143
 144         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 145         if (ret)
 146                 return ret;
 147
 148         /* We inline crc's for the free disk space cache */
 149         if (ino != BTRFS_FREE_INO_OBJECTID)
 150                 flags |= BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 151
 152         leaf = path->nodes[0];
 153         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 154                                     struct btrfs_inode_item);
 155         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 156         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 157                              sizeof(*inode_item));
 158         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 159         btrfs_set_inode_size(leaf, inode_item, 0);
 160         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 161         btrfs_set_inode_uid(leaf, inode_item, 0);
 162         btrfs_set_inode_gid(leaf, inode_item, 0);
 163         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 164         btrfs_set_inode_flags(leaf, inode_item, flags);
 165         btrfs_set_inode_nlink(leaf, inode_item, 1);
 166         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 167         btrfs_set_inode_block_group(leaf, inode_item, offset);
 168         btrfs_mark_buffer_dirty(leaf);
 169         btrfs_release_path(path);
 170
 171         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 172         key.offset = offset;
 173         key.type = 0;
 174         ret = btrfs_insert_empty_item(trans, root, path, &key,
 175                                       sizeof(struct btrfs_free_space_header));
 176         if (ret < 0) {
 177                 btrfs_release_path(path);
 178                 return ret;
 179         }
 180
 181         leaf = path->nodes[0];
 182         header = btrfs_item_ptr(leaf, path->slots[0],
 183                                 struct btrfs_free_space_header);
 184         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 185         btrfs_set_free_space_key(leaf, header, &disk_key);
 186         btrfs_mark_buffer_dirty(leaf);
 187         btrfs_release_path(path);
 188
 189         return 0;
 190 }
 191
 192 int create_free_space_inode(struct btrfs_trans_handle *trans,
 193                             struct btrfs_block_group *block_group,
 194                             struct btrfs_path *path)
 195 {
 196         int ret;
 197         u64 ino;
 198
 199         ret = btrfs_find_free_objectid(trans->fs_info->tree_root, &ino);
 200         if (ret < 0)
 201                 return ret;
 202
 203         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 204                                          ino, block_group->start);
 205 }
 206
 207 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 208                                        struct btrfs_block_rsv *rsv)
 209 {
 210         u64 needed_bytes;
 211         int ret;
 212
 213         /* 1 for slack space, 1 for updating the inode */
 214         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 215                 btrfs_calc_metadata_size(fs_info, 1);
 216
 217         spin_lock(&rsv->lock);
 218         if (rsv->reserved < needed_bytes)
 219                 ret = -ENOSPC;
 220         else
 221                 ret = 0;
 222         spin_unlock(&rsv->lock);
 223         return ret;
 224 }
 225
 226 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 227                                     struct btrfs_block_group *block_group,
 228                                     struct inode *inode)
 229 {
 230         struct btrfs_root *root = BTRFS_I(inode)->root;
 231         int ret = 0;
 232         bool locked = false;
 233
 234         if (block_group) {
 235                 struct btrfs_path *path = btrfs_alloc_path();
 236
 237                 if (!path) {
 238                         ret = -ENOMEM;
 239                         goto fail;
 240                 }
 241                 locked = true;
 242                 mutex_lock(&trans->transaction->cache_write_mutex);
 243                 if (!list_empty(&block_group->io_list)) {
 244                         list_del_init(&block_group->io_list);
 245
 246                         btrfs_wait_cache_io(trans, block_group, path);
 247                         btrfs_put_block_group(block_group);
 248                 }
 249
 250                 /*
 251                  * now that we've truncated the cache away, its no longer
 252                  * setup or written
 253                  */
 254                 spin_lock(&block_group->lock);
 255                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 256                 spin_unlock(&block_group->lock);
 257                 btrfs_free_path(path);
 258         }
 259
 260         btrfs_i_size_write(BTRFS_I(inode), 0);
 261         truncate_pagecache(inode, 0);
 262
 263         /*
 264          * We skip the throttling logic for free space cache inodes, so we don't
 265          * need to check for -EAGAIN.
 266          */
 267         ret = btrfs_truncate_inode_items(trans, root, inode,
 268                                          0, BTRFS_EXTENT_DATA_KEY);
 269         if (ret)
 270                 goto fail;
 271
 272         ret = btrfs_update_inode(trans, root, inode);
 273
 274 fail:
 275         if (locked)
 276                 mutex_unlock(&trans->transaction->cache_write_mutex);
 277         if (ret)
 278                 btrfs_abort_transaction(trans, ret);
 279
 280         return ret;
 281 }
 282
 283 static void readahead_cache(struct inode *inode)
 284 {
 285         struct file_ra_state *ra;
 286         unsigned long last_index;
 287
 288         ra = kzalloc(sizeof(*ra), GFP_NOFS);
 289         if (!ra)
 290                 return;
 291
 292         file_ra_state_init(ra, inode->i_mapping);
 293         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 294
 295         page_cache_sync_readahead(inode->i_mapping, ra, NULL, 0, last_index);
 296
 297         kfree(ra);
 298 }
 299
 300 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 301                        int write)
 302 {
 303         int num_pages;
 304         int check_crcs = 0;
 305
 306         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 307
 308         if (btrfs_ino(BTRFS_I(inode)) != BTRFS_FREE_INO_OBJECTID)
 309                 check_crcs = 1;
 310
 311         /* Make sure we can fit our crcs and generation into the first page */
 312         if (write && check_crcs &&
 313             (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 314                 return -ENOSPC;
 315
 316         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 317
 318         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 319         if (!io_ctl->pages)
 320                 return -ENOMEM;
 321
 322         io_ctl->num_pages = num_pages;
 323         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 324         io_ctl->check_crcs = check_crcs;
 325         io_ctl->inode = inode;
 326
 327         return 0;
 328 }
 329 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 330
 331 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 332 {
 333         kfree(io_ctl->pages);
 334         io_ctl->pages = NULL;
 335 }
 336
 337 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 338 {
 339         if (io_ctl->cur) {
 340                 io_ctl->cur = NULL;
 341                 io_ctl->orig = NULL;
 342         }
 343 }
 344
 345 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 346 {
 347         ASSERT(io_ctl->index < io_ctl->num_pages);
 348         io_ctl->page = io_ctl->pages[io_ctl->index++];
 349         io_ctl->cur = page_address(io_ctl->page);
 350         io_ctl->orig = io_ctl->cur;
 351         io_ctl->size = PAGE_SIZE;
 352         if (clear)
 353                 clear_page(io_ctl->cur);
 354 }
 355
 356 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 357 {
 358         int i;
 359
 360         io_ctl_unmap_page(io_ctl);
 361
 362         for (i = 0; i < io_ctl->num_pages; i++) {
 363                 if (io_ctl->pages[i]) {
 364                         ClearPageChecked(io_ctl->pages[i]);
 365                         unlock_page(io_ctl->pages[i]);
 366                         put_page(io_ctl->pages[i]);
 367                 }
 368         }
 369 }
 370
 371 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 372                                 int uptodate)
 373 {
 374         struct page *page;
 375         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 376         int i;
 377
 378         for (i = 0; i < io_ctl->num_pages; i++) {
 379                 page = find_or_create_page(inode->i_mapping, i, mask);
 380                 if (!page) {
 381                         io_ctl_drop_pages(io_ctl);
 382                         return -ENOMEM;
 383                 }
 384                 io_ctl->pages[i] = page;
 385                 if (uptodate && !PageUptodate(page)) {
 386                         btrfs_readpage(NULL, page);
 387                         lock_page(page);
 388                         if (page->mapping != inode->i_mapping) {
 389                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 390                                           "free space cache page truncated");
 391                                 io_ctl_drop_pages(io_ctl);
 392                                 return -EIO;
 393                         }
 394                         if (!PageUptodate(page)) {
 395                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 396                                            "error reading free space cache");
 397                                 io_ctl_drop_pages(io_ctl);
 398                                 return -EIO;
 399                         }
 400                 }
 401         }
 402
 403         for (i = 0; i < io_ctl->num_pages; i++) {
 404                 clear_page_dirty_for_io(io_ctl->pages[i]);
 405                 set_page_extent_mapped(io_ctl->pages[i]);
 406         }
 407
 408         return 0;
 409 }
 410
 411 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 412 {
 413         __le64 *val;
 414
 415         io_ctl_map_page(io_ctl, 1);
 416
 417         /*
 418          * Skip the csum areas.  If we don't check crcs then we just have a
 419          * 64bit chunk at the front of the first page.
 420          */
 421         if (io_ctl->check_crcs) {
 422                 io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 423                 io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 424         } else {
 425                 io_ctl->cur += sizeof(u64);
 426                 io_ctl->size -= sizeof(u64) * 2;
 427         }
 428
 429         val = io_ctl->cur;
 430         *val = cpu_to_le64(generation);
 431         io_ctl->cur += sizeof(u64);
 432 }
 433
 434 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 435 {
 436         __le64 *gen;
 437
 438         /*
 439          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 440          * chunk at the front of the first page.
 441          */
 442         if (io_ctl->check_crcs) {
 443                 io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 444                 io_ctl->size -= sizeof(u64) +
 445                         (sizeof(u32) * io_ctl->num_pages);
 446         } else {
 447                 io_ctl->cur += sizeof(u64);
 448                 io_ctl->size -= sizeof(u64) * 2;
 449         }
 450
 451         gen = io_ctl->cur;
 452         if (le64_to_cpu(*gen) != generation) {
 453                 btrfs_err_rl(io_ctl->fs_info,
 454                         "space cache generation (%llu) does not match inode (%llu)",
 455                                 *gen, generation);
 456                 io_ctl_unmap_page(io_ctl);
 457                 return -EIO;
 458         }
 459         io_ctl->cur += sizeof(u64);
 460         return 0;
 461 }
 462
 463 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 464 {
 465         u32 *tmp;
 466         u32 crc = ~(u32)0;
 467         unsigned offset = 0;
 468
 469         if (!io_ctl->check_crcs) {
 470                 io_ctl_unmap_page(io_ctl);
 471                 return;
 472         }
 473
 474         if (index == 0)
 475                 offset = sizeof(u32) * io_ctl->num_pages;
 476
 477         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 478         btrfs_crc32c_final(crc, (u8 *)&crc);
 479         io_ctl_unmap_page(io_ctl);
 480         tmp = page_address(io_ctl->pages[0]);
 481         tmp += index;
 482         *tmp = crc;
 483 }
 484
 485 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 486 {
 487         u32 *tmp, val;
 488         u32 crc = ~(u32)0;
 489         unsigned offset = 0;
 490
 491         if (!io_ctl->check_crcs) {
 492                 io_ctl_map_page(io_ctl, 0);
 493                 return 0;
 494         }
 495
 496         if (index == 0)
 497                 offset = sizeof(u32) * io_ctl->num_pages;
 498
 499         tmp = page_address(io_ctl->pages[0]);
 500         tmp += index;
 501         val = *tmp;
 502
 503         io_ctl_map_page(io_ctl, 0);
 504         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 505         btrfs_crc32c_final(crc, (u8 *)&crc);
 506         if (val != crc) {
 507                 btrfs_err_rl(io_ctl->fs_info,
 508                         "csum mismatch on free space cache");
 509                 io_ctl_unmap_page(io_ctl);
 510                 return -EIO;
 511         }
 512
 513         return 0;
 514 }
 515
 516 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 517                             void *bitmap)
 518 {
 519         struct btrfs_free_space_entry *entry;
 520
 521         if (!io_ctl->cur)
 522                 return -ENOSPC;
 523
 524         entry = io_ctl->cur;
 525         entry->offset = cpu_to_le64(offset);
 526         entry->bytes = cpu_to_le64(bytes);
 527         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 528                 BTRFS_FREE_SPACE_EXTENT;
 529         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 530         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 531
 532         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 533                 return 0;
 534
 535         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 536
 537         /* No more pages to map */
 538         if (io_ctl->index >= io_ctl->num_pages)
 539                 return 0;
 540
 541         /* map the next page */
 542         io_ctl_map_page(io_ctl, 1);
 543         return 0;
 544 }
 545
 546 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 547 {
 548         if (!io_ctl->cur)
 549                 return -ENOSPC;
 550
 551         /*
 552          * If we aren't at the start of the current page, unmap this one and
 553          * map the next one if there is any left.
 554          */
 555         if (io_ctl->cur != io_ctl->orig) {
 556                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 557                 if (io_ctl->index >= io_ctl->num_pages)
 558                         return -ENOSPC;
 559                 io_ctl_map_page(io_ctl, 0);
 560         }
 561
 562         copy_page(io_ctl->cur, bitmap);
 563         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 564         if (io_ctl->index < io_ctl->num_pages)
 565                 io_ctl_map_page(io_ctl, 0);
 566         return 0;
 567 }
 568
 569 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 570 {
 571         /*
 572          * If we're not on the boundary we know we've modified the page and we
 573          * need to crc the page.
 574          */
 575         if (io_ctl->cur != io_ctl->orig)
 576                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 577         else
 578                 io_ctl_unmap_page(io_ctl);
 579
 580         while (io_ctl->index < io_ctl->num_pages) {
 581                 io_ctl_map_page(io_ctl, 1);
 582                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 583         }
 584 }
 585
 586 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 587                             struct btrfs_free_space *entry, u8 *type)
 588 {
 589         struct btrfs_free_space_entry *e;
 590         int ret;
 591
 592         if (!io_ctl->cur) {
 593                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 594                 if (ret)
 595                         return ret;
 596         }
 597
 598         e = io_ctl->cur;
 599         entry->offset = le64_to_cpu(e->offset);
 600         entry->bytes = le64_to_cpu(e->bytes);
 601         *type = e->type;
 602         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 603         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 604
 605         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 606                 return 0;
 607
 608         io_ctl_unmap_page(io_ctl);
 609
 610         return 0;
 611 }
 612
 613 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 614                               struct btrfs_free_space *entry)
 615 {
 616         int ret;
 617
 618         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 619         if (ret)
 620                 return ret;
 621
 622         copy_page(entry->bitmap, io_ctl->cur);
 623         io_ctl_unmap_page(io_ctl);
 624
 625         return 0;
 626 }
 627
 628 /*
 629  * Since we attach pinned extents after the fact we can have contiguous sections
 630  * of free space that are split up in entries.  This poses a problem with the
 631  * tree logging stuff since it could have allocated across what appears to be 2
 632  * entries since we would have merged the entries when adding the pinned extents
 633  * back to the free space cache.  So run through the space cache that we just
 634  * loaded and merge contiguous entries.  This will make the log replay stuff not
 635  * blow up and it will make for nicer allocator behavior.
 636  */
 637 static void merge_space_tree(struct btrfs_free_space_ctl *ctl)
 638 {
 639         struct btrfs_free_space *e, *prev = NULL;
 640         struct rb_node *n;
 641
 642 again:
 643         spin_lock(&ctl->tree_lock);
 644         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
 645                 e = rb_entry(n, struct btrfs_free_space, offset_index);
 646                 if (!prev)
 647                         goto next;
 648                 if (e->bitmap || prev->bitmap)
 649                         goto next;
 650                 if (prev->offset + prev->bytes == e->offset) {
 651                         unlink_free_space(ctl, prev);
 652                         unlink_free_space(ctl, e);
 653                         prev->bytes += e->bytes;
 654                         kmem_cache_free(btrfs_free_space_cachep, e);
 655                         link_free_space(ctl, prev);
 656                         prev = NULL;
 657                         spin_unlock(&ctl->tree_lock);
 658                         goto again;
 659                 }
 660 next:
 661                 prev = e;
 662         }
 663         spin_unlock(&ctl->tree_lock);
 664 }
 665
 666 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 667                                    struct btrfs_free_space_ctl *ctl,
 668                                    struct btrfs_path *path, u64 offset)
 669 {
 670         struct btrfs_fs_info *fs_info = root->fs_info;
 671         struct btrfs_free_space_header *header;
 672         struct extent_buffer *leaf;
 673         struct btrfs_io_ctl io_ctl;
 674         struct btrfs_key key;
 675         struct btrfs_free_space *e, *n;
 676         LIST_HEAD(bitmaps);
 677         u64 num_entries;
 678         u64 num_bitmaps;
 679         u64 generation;
 680         u8 type;
 681         int ret = 0;
 682
 683         /* Nothing in the space cache, goodbye */
 684         if (!i_size_read(inode))
 685                 return 0;
 686
 687         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 688         key.offset = offset;
 689         key.type = 0;
 690
 691         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 692         if (ret < 0)
 693                 return 0;
 694         else if (ret > 0) {
 695                 btrfs_release_path(path);
 696                 return 0;
 697         }
 698
 699         ret = -1;
 700
 701         leaf = path->nodes[0];
 702         header = btrfs_item_ptr(leaf, path->slots[0],
 703                                 struct btrfs_free_space_header);
 704         num_entries = btrfs_free_space_entries(leaf, header);
 705         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 706         generation = btrfs_free_space_generation(leaf, header);
 707         btrfs_release_path(path);
 708
 709         if (!BTRFS_I(inode)->generation) {
 710                 btrfs_info(fs_info,
 711                            "the free space cache file (%llu) is invalid, skip it",
 712                            offset);
 713                 return 0;
 714         }
 715
 716         if (BTRFS_I(inode)->generation != generation) {
 717                 btrfs_err(fs_info,
 718                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 719                           BTRFS_I(inode)->generation, generation);
 720                 return 0;
 721         }
 722
 723         if (!num_entries)
 724                 return 0;
 725
 726         ret = io_ctl_init(&io_ctl, inode, 0);
 727         if (ret)
 728                 return ret;
 729
 730         readahead_cache(inode);
 731
 732         ret = io_ctl_prepare_pages(&io_ctl, inode, 1);
 733         if (ret)
 734                 goto out;
 735
 736         ret = io_ctl_check_crc(&io_ctl, 0);
 737         if (ret)
 738                 goto free_cache;
 739
 740         ret = io_ctl_check_generation(&io_ctl, generation);
 741         if (ret)
 742                 goto free_cache;
 743
 744         while (num_entries) {
 745                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 746                                       GFP_NOFS);
 747                 if (!e)
 748                         goto free_cache;
 749
 750                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 751                 if (ret) {
 752                         kmem_cache_free(btrfs_free_space_cachep, e);
 753                         goto free_cache;
 754                 }
 755
 756                 /*
 757                  * Sync discard ensures that the free space cache is always
 758                  * trimmed.  So when reading this in, the state should reflect
 759                  * that.  We also do this for async as a stop gap for lack of
 760                  * persistence.
 761                  */
 762                 if (btrfs_test_opt(fs_info, DISCARD_SYNC) ||
 763                     btrfs_test_opt(fs_info, DISCARD_ASYNC))
 764                         e->trim_state = BTRFS_TRIM_STATE_TRIMMED;
 765
 766                 if (!e->bytes) {
 767                         kmem_cache_free(btrfs_free_space_cachep, e);
 768                         goto free_cache;
 769                 }
 770
 771                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 772                         spin_lock(&ctl->tree_lock);
 773                         ret = link_free_space(ctl, e);
 774                         spin_unlock(&ctl->tree_lock);
 775                         if (ret) {
 776                                 btrfs_err(fs_info,
 777                                         "Duplicate entries in free space cache, dumping");
 778                                 kmem_cache_free(btrfs_free_space_cachep, e);
 779                                 goto free_cache;
 780                         }
 781                 } else {
 782                         ASSERT(num_bitmaps);
 783                         num_bitmaps--;
 784                         e->bitmap = kmem_cache_zalloc(
 785                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 786                         if (!e->bitmap) {
 787                                 kmem_cache_free(
 788                                         btrfs_free_space_cachep, e);
 789                                 goto free_cache;
 790                         }
 791                         spin_lock(&ctl->tree_lock);
 792                         ret = link_free_space(ctl, e);
 793                         ctl->total_bitmaps++;
 794                         ctl->op->recalc_thresholds(ctl);
 795                         spin_unlock(&ctl->tree_lock);
 796                         if (ret) {
 797                                 btrfs_err(fs_info,
 798                                         "Duplicate entries in free space cache, dumping");
 799                                 kmem_cache_free(btrfs_free_space_cachep, e);
 800                                 goto free_cache;
 801                         }
 802                         list_add_tail(&e->list, &bitmaps);
 803                 }
 804
 805                 num_entries--;
 806         }
 807
 808         io_ctl_unmap_page(&io_ctl);
 809
 810         /*
 811          * We add the bitmaps at the end of the entries in order that
 812          * the bitmap entries are added to the cache.
 813          */
 814         list_for_each_entry_safe(e, n, &bitmaps, list) {
 815                 list_del_init(&e->list);
 816                 ret = io_ctl_read_bitmap(&io_ctl, e);
 817                 if (ret)
 818                         goto free_cache;
 819         }
 820
 821         io_ctl_drop_pages(&io_ctl);
 822         merge_space_tree(ctl);
 823         ret = 1;
 824 out:
 825         io_ctl_free(&io_ctl);
 826         return ret;
 827 free_cache:
 828         io_ctl_drop_pages(&io_ctl);
 829         __btrfs_remove_free_space_cache(ctl);
 830         goto out;
 831 }
 832
 833 int load_free_space_cache(struct btrfs_block_group *block_group)
 834 {
 835         struct btrfs_fs_info *fs_info = block_group->fs_info;
 836         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 837         struct inode *inode;
 838         struct btrfs_path *path;
 839         int ret = 0;
 840         bool matched;
 841         u64 used = block_group->used;
 842
 843         /*
 844          * If this block group has been marked to be cleared for one reason or
 845          * another then we can't trust the on disk cache, so just return.
 846          */
 847         spin_lock(&block_group->lock);
 848         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 849                 spin_unlock(&block_group->lock);
 850                 return 0;
 851         }
 852         spin_unlock(&block_group->lock);
 853
 854         path = btrfs_alloc_path();
 855         if (!path)
 856                 return 0;
 857         path->search_commit_root = 1;
 858         path->skip_locking = 1;
 859
 860         /*
 861          * We must pass a path with search_commit_root set to btrfs_iget in
 862          * order to avoid a deadlock when allocating extents for the tree root.
 863          *
 864          * When we are COWing an extent buffer from the tree root, when looking
 865          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 866          * block group without its free space cache loaded. When we find one
 867          * we must load its space cache which requires reading its free space
 868          * cache's inode item from the root tree. If this inode item is located
 869          * in the same leaf that we started COWing before, then we end up in
 870          * deadlock on the extent buffer (trying to read lock it when we
 871          * previously write locked it).
 872          *
 873          * It's safe to read the inode item using the commit root because
 874          * block groups, once loaded, stay in memory forever (until they are
 875          * removed) as well as their space caches once loaded. New block groups
 876          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 877          * we will never try to read their inode item while the fs is mounted.
 878          */
 879         inode = lookup_free_space_inode(block_group, path);
 880         if (IS_ERR(inode)) {
 881                 btrfs_free_path(path);
 882                 return 0;
 883         }
 884
 885         /* We may have converted the inode and made the cache invalid. */
 886         spin_lock(&block_group->lock);
 887         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 888                 spin_unlock(&block_group->lock);
 889                 btrfs_free_path(path);
 890                 goto out;
 891         }
 892         spin_unlock(&block_group->lock);
 893
 894         ret = __load_free_space_cache(fs_info->tree_root, inode, ctl,
 895                                       path, block_group->start);
 896         btrfs_free_path(path);
 897         if (ret <= 0)
 898                 goto out;
 899
 900         spin_lock(&ctl->tree_lock);
 901         matched = (ctl->free_space == (block_group->length - used -
 902                                        block_group->bytes_super));
 903         spin_unlock(&ctl->tree_lock);
 904
 905         if (!matched) {
 906                 __btrfs_remove_free_space_cache(ctl);
 907                 btrfs_warn(fs_info,
 908                            "block group %llu has wrong amount of free space",
 909                            block_group->start);
 910                 ret = -1;
 911         }
 912 out:
 913         if (ret < 0) {
 914                 /* This cache is bogus, make sure it gets cleared */
 915                 spin_lock(&block_group->lock);
 916                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 917                 spin_unlock(&block_group->lock);
 918                 ret = 0;
 919
 920                 btrfs_warn(fs_info,
 921                            "failed to load free space cache for block group %llu, rebuilding it now",
 922                            block_group->start);
 923         }
 924
 925         iput(inode);
 926         return ret;
 927 }
 928
 929 static noinline_for_stack
 930 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
 931                               struct btrfs_free_space_ctl *ctl,
 932                               struct btrfs_block_group *block_group,
 933                               int *entries, int *bitmaps,
 934                               struct list_head *bitmap_list)
 935 {
 936         int ret;
 937         struct btrfs_free_cluster *cluster = NULL;
 938         struct btrfs_free_cluster *cluster_locked = NULL;
 939         struct rb_node *node = rb_first(&ctl->free_space_offset);
 940         struct btrfs_trim_range *trim_entry;
 941
 942         /* Get the cluster for this block_group if it exists */
 943         if (block_group && !list_empty(&block_group->cluster_list)) {
 944                 cluster = list_entry(block_group->cluster_list.next,
 945                                      struct btrfs_free_cluster,
 946                                      block_group_list);
 947         }
 948
 949         if (!node && cluster) {
 950                 cluster_locked = cluster;
 951                 spin_lock(&cluster_locked->lock);
 952                 node = rb_first(&cluster->root);
 953                 cluster = NULL;
 954         }
 955
 956         /* Write out the extent entries */
 957         while (node) {
 958                 struct btrfs_free_space *e;
 959
 960                 e = rb_entry(node, struct btrfs_free_space, offset_index);
 961                 *entries += 1;
 962
 963                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
 964                                        e->bitmap);
 965                 if (ret)
 966                         goto fail;
 967
 968                 if (e->bitmap) {
 969                         list_add_tail(&e->list, bitmap_list);
 970                         *bitmaps += 1;
 971                 }
 972                 node = rb_next(node);
 973                 if (!node && cluster) {
 974                         node = rb_first(&cluster->root);
 975                         cluster_locked = cluster;
 976                         spin_lock(&cluster_locked->lock);
 977                         cluster = NULL;
 978                 }
 979         }
 980         if (cluster_locked) {
 981                 spin_unlock(&cluster_locked->lock);
 982                 cluster_locked = NULL;
 983         }
 984
 985         /*
 986          * Make sure we don't miss any range that was removed from our rbtree
 987          * because trimming is running. Otherwise after a umount+mount (or crash
 988          * after committing the transaction) we would leak free space and get
 989          * an inconsistent free space cache report from fsck.
 990          */
 991         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
 992                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
 993                                        trim_entry->bytes, NULL);
 994                 if (ret)
 995                         goto fail;
 996                 *entries += 1;
 997         }
 998
 999         return 0;
1000 fail:
1001         if (cluster_locked)
1002                 spin_unlock(&cluster_locked->lock);
1003         return -ENOSPC;
1004 }
1005
1006 static noinline_for_stack int
1007 update_cache_item(struct btrfs_trans_handle *trans,
1008                   struct btrfs_root *root,
1009                   struct inode *inode,
1010                   struct btrfs_path *path, u64 offset,
1011                   int entries, int bitmaps)
1012 {
1013         struct btrfs_key key;
1014         struct btrfs_free_space_header *header;
1015         struct extent_buffer *leaf;
1016         int ret;
1017
1018         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1019         key.offset = offset;
1020         key.type = 0;
1021
1022         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1023         if (ret < 0) {
1024                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1025                                  EXTENT_DELALLOC, 0, 0, NULL);
1026                 goto fail;
1027         }
1028         leaf = path->nodes[0];
1029         if (ret > 0) {
1030                 struct btrfs_key found_key;
1031                 ASSERT(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, EXTENT_DELALLOC, 0,
1038                                          0, NULL);
1039                         btrfs_release_path(path);
1040                         goto fail;
1041                 }
1042         }
1043
1044         BTRFS_I(inode)->generation = trans->transid;
1045         header = btrfs_item_ptr(leaf, path->slots[0],
1046                                 struct btrfs_free_space_header);
1047         btrfs_set_free_space_entries(leaf, header, entries);
1048         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1049         btrfs_set_free_space_generation(leaf, header, trans->transid);
1050         btrfs_mark_buffer_dirty(leaf);
1051         btrfs_release_path(path);
1052
1053         return 0;
1054
1055 fail:
1056         return -1;
1057 }
1058
1059 static noinline_for_stack int write_pinned_extent_entries(
1060                             struct btrfs_block_group *block_group,
1061                             struct btrfs_io_ctl *io_ctl,
1062                             int *entries)
1063 {
1064         u64 start, extent_start, extent_end, len;
1065         struct extent_io_tree *unpin = NULL;
1066         int ret;
1067
1068         if (!block_group)
1069                 return 0;
1070
1071         /*
1072          * We want to add any pinned extents to our free space cache
1073          * so we don't leak the space
1074          *
1075          * We shouldn't have switched the pinned extents yet so this is the
1076          * right one
1077          */
1078         unpin = block_group->fs_info->pinned_extents;
1079
1080         start = block_group->start;
1081
1082         while (start < block_group->start + block_group->length) {
1083                 ret = find_first_extent_bit(unpin, start,
1084                                             &extent_start, &extent_end,
1085                                             EXTENT_DIRTY, NULL);
1086                 if (ret)
1087                         return 0;
1088
1089                 /* This pinned extent is out of our range */
1090                 if (extent_start >= block_group->start + block_group->length)
1091                         return 0;
1092
1093                 extent_start = max(extent_start, start);
1094                 extent_end = min(block_group->start + block_group->length,
1095                                  extent_end + 1);
1096                 len = extent_end - extent_start;
1097
1098                 *entries += 1;
1099                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1100                 if (ret)
1101                         return -ENOSPC;
1102
1103                 start = extent_end;
1104         }
1105
1106         return 0;
1107 }
1108
1109 static noinline_for_stack int
1110 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1111 {
1112         struct btrfs_free_space *entry, *next;
1113         int ret;
1114
1115         /* Write out the bitmaps */
1116         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1117                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1118                 if (ret)
1119                         return -ENOSPC;
1120                 list_del_init(&entry->list);
1121         }
1122
1123         return 0;
1124 }
1125
1126 static int flush_dirty_cache(struct inode *inode)
1127 {
1128         int ret;
1129
1130         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1131         if (ret)
1132                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1133                                  EXTENT_DELALLOC, 0, 0, NULL);
1134
1135         return ret;
1136 }
1137
1138 static void noinline_for_stack
1139 cleanup_bitmap_list(struct list_head *bitmap_list)
1140 {
1141         struct btrfs_free_space *entry, *next;
1142
1143         list_for_each_entry_safe(entry, next, bitmap_list, list)
1144                 list_del_init(&entry->list);
1145 }
1146
1147 static void noinline_for_stack
1148 cleanup_write_cache_enospc(struct inode *inode,
1149                            struct btrfs_io_ctl *io_ctl,
1150                            struct extent_state **cached_state)
1151 {
1152         io_ctl_drop_pages(io_ctl);
1153         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1154                              i_size_read(inode) - 1, cached_state);
1155 }
1156
1157 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1158                                  struct btrfs_trans_handle *trans,
1159                                  struct btrfs_block_group *block_group,
1160                                  struct btrfs_io_ctl *io_ctl,
1161                                  struct btrfs_path *path, u64 offset)
1162 {
1163         int ret;
1164         struct inode *inode = io_ctl->inode;
1165
1166         if (!inode)
1167                 return 0;
1168
1169         /* Flush the dirty pages in the cache file. */
1170         ret = flush_dirty_cache(inode);
1171         if (ret)
1172                 goto out;
1173
1174         /* Update the cache item to tell everyone this cache file is valid. */
1175         ret = update_cache_item(trans, root, inode, path, offset,
1176                                 io_ctl->entries, io_ctl->bitmaps);
1177 out:
1178         io_ctl_free(io_ctl);
1179         if (ret) {
1180                 invalidate_inode_pages2(inode->i_mapping);
1181                 BTRFS_I(inode)->generation = 0;
1182                 if (block_group) {
1183 #ifdef DEBUG
1184                         btrfs_err(root->fs_info,
1185                                   "failed to write free space cache for block group %llu",
1186                                   block_group->start);
1187 #endif
1188                 }
1189         }
1190         btrfs_update_inode(trans, root, inode);
1191
1192         if (block_group) {
1193                 /* the dirty list is protected by the dirty_bgs_lock */
1194                 spin_lock(&trans->transaction->dirty_bgs_lock);
1195
1196                 /* the disk_cache_state is protected by the block group lock */
1197                 spin_lock(&block_group->lock);
1198
1199                 /*
1200                  * only mark this as written if we didn't get put back on
1201                  * the dirty list while waiting for IO.   Otherwise our
1202                  * cache state won't be right, and we won't get written again
1203                  */
1204                 if (!ret && list_empty(&block_group->dirty_list))
1205                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1206                 else if (ret)
1207                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1208
1209                 spin_unlock(&block_group->lock);
1210                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1211                 io_ctl->inode = NULL;
1212                 iput(inode);
1213         }
1214
1215         return ret;
1216
1217 }
1218
1219 static int btrfs_wait_cache_io_root(struct btrfs_root *root,
1220                                     struct btrfs_trans_handle *trans,
1221                                     struct btrfs_io_ctl *io_ctl,
1222                                     struct btrfs_path *path)
1223 {
1224         return __btrfs_wait_cache_io(root, trans, NULL, io_ctl, path, 0);
1225 }
1226
1227 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1228                         struct btrfs_block_group *block_group,
1229                         struct btrfs_path *path)
1230 {
1231         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1232                                      block_group, &block_group->io_ctl,
1233                                      path, block_group->start);
1234 }
1235
1236 /**
1237  * __btrfs_write_out_cache - write out cached info to an inode
1238  * @root - the root the inode belongs to
1239  * @ctl - the free space cache we are going to write out
1240  * @block_group - the block_group for this cache if it belongs to a block_group
1241  * @trans - the trans handle
1242  *
1243  * This function writes out a free space cache struct to disk for quick recovery
1244  * on mount.  This will return 0 if it was successful in writing the cache out,
1245  * or an errno if it was not.
1246  */
1247 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1248                                    struct btrfs_free_space_ctl *ctl,
1249                                    struct btrfs_block_group *block_group,
1250                                    struct btrfs_io_ctl *io_ctl,
1251                                    struct btrfs_trans_handle *trans)
1252 {
1253         struct extent_state *cached_state = NULL;
1254         LIST_HEAD(bitmap_list);
1255         int entries = 0;
1256         int bitmaps = 0;
1257         int ret;
1258         int must_iput = 0;
1259
1260         if (!i_size_read(inode))
1261                 return -EIO;
1262
1263         WARN_ON(io_ctl->pages);
1264         ret = io_ctl_init(io_ctl, inode, 1);
1265         if (ret)
1266                 return ret;
1267
1268         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1269                 down_write(&block_group->data_rwsem);
1270                 spin_lock(&block_group->lock);
1271                 if (block_group->delalloc_bytes) {
1272                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1273                         spin_unlock(&block_group->lock);
1274                         up_write(&block_group->data_rwsem);
1275                         BTRFS_I(inode)->generation = 0;
1276                         ret = 0;
1277                         must_iput = 1;
1278                         goto out;
1279                 }
1280                 spin_unlock(&block_group->lock);
1281         }
1282
1283         /* Lock all pages first so we can lock the extent safely. */
1284         ret = io_ctl_prepare_pages(io_ctl, inode, 0);
1285         if (ret)
1286                 goto out_unlock;
1287
1288         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1289                          &cached_state);
1290
1291         io_ctl_set_generation(io_ctl, trans->transid);
1292
1293         mutex_lock(&ctl->cache_writeout_mutex);
1294         /* Write out the extent entries in the free space cache */
1295         spin_lock(&ctl->tree_lock);
1296         ret = write_cache_extent_entries(io_ctl, ctl,
1297                                          block_group, &entries, &bitmaps,
1298                                          &bitmap_list);
1299         if (ret)
1300                 goto out_nospc_locked;
1301
1302         /*
1303          * Some spaces that are freed in the current transaction are pinned,
1304          * they will be added into free space cache after the transaction is
1305          * committed, we shouldn't lose them.
1306          *
1307          * If this changes while we are working we'll get added back to
1308          * the dirty list and redo it.  No locking needed
1309          */
1310         ret = write_pinned_extent_entries(block_group, io_ctl, &entries);
1311         if (ret)
1312                 goto out_nospc_locked;
1313
1314         /*
1315          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1316          * locked while doing it because a concurrent trim can be manipulating
1317          * or freeing the bitmap.
1318          */
1319         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1320         spin_unlock(&ctl->tree_lock);
1321         mutex_unlock(&ctl->cache_writeout_mutex);
1322         if (ret)
1323                 goto out_nospc;
1324
1325         /* Zero out the rest of the pages just to make sure */
1326         io_ctl_zero_remaining_pages(io_ctl);
1327
1328         /* Everything is written out, now we dirty the pages in the file. */
1329         ret = btrfs_dirty_pages(inode, io_ctl->pages, io_ctl->num_pages, 0,
1330                                 i_size_read(inode), &cached_state);
1331         if (ret)
1332                 goto out_nospc;
1333
1334         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1335                 up_write(&block_group->data_rwsem);
1336         /*
1337          * Release the pages and unlock the extent, we will flush
1338          * them out later
1339          */
1340         io_ctl_drop_pages(io_ctl);
1341
1342         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1343                              i_size_read(inode) - 1, &cached_state);
1344
1345         /*
1346          * at this point the pages are under IO and we're happy,
1347          * The caller is responsible for waiting on them and updating the
1348          * the cache and the inode
1349          */
1350         io_ctl->entries = entries;
1351         io_ctl->bitmaps = bitmaps;
1352
1353         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1354         if (ret)
1355                 goto out;
1356
1357         return 0;
1358
1359 out:
1360         io_ctl->inode = NULL;
1361         io_ctl_free(io_ctl);
1362         if (ret) {
1363                 invalidate_inode_pages2(inode->i_mapping);
1364                 BTRFS_I(inode)->generation = 0;
1365         }
1366         btrfs_update_inode(trans, root, inode);
1367         if (must_iput)
1368                 iput(inode);
1369         return ret;
1370
1371 out_nospc_locked:
1372         cleanup_bitmap_list(&bitmap_list);
1373         spin_unlock(&ctl->tree_lock);
1374         mutex_unlock(&ctl->cache_writeout_mutex);
1375
1376 out_nospc:
1377         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1378
1379 out_unlock:
1380         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1381                 up_write(&block_group->data_rwsem);
1382
1383         goto out;
1384 }
1385
1386 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1387                           struct btrfs_block_group *block_group,
1388                           struct btrfs_path *path)
1389 {
1390         struct btrfs_fs_info *fs_info = trans->fs_info;
1391         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1392         struct inode *inode;
1393         int ret = 0;
1394
1395         spin_lock(&block_group->lock);
1396         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1397                 spin_unlock(&block_group->lock);
1398                 return 0;
1399         }
1400         spin_unlock(&block_group->lock);
1401
1402         inode = lookup_free_space_inode(block_group, path);
1403         if (IS_ERR(inode))
1404                 return 0;
1405
1406         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1407                                 block_group, &block_group->io_ctl, trans);
1408         if (ret) {
1409 #ifdef DEBUG
1410                 btrfs_err(fs_info,
1411                           "failed to write free space cache for block group %llu",
1412                           block_group->start);
1413 #endif
1414                 spin_lock(&block_group->lock);
1415                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1416                 spin_unlock(&block_group->lock);
1417
1418                 block_group->io_ctl.inode = NULL;
1419                 iput(inode);
1420         }
1421
1422         /*
1423          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1424          * to wait for IO and put the inode
1425          */
1426
1427         return ret;
1428 }
1429
1430 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1431                                           u64 offset)
1432 {
1433         ASSERT(offset >= bitmap_start);
1434         offset -= bitmap_start;
1435         return (unsigned long)(div_u64(offset, unit));
1436 }
1437
1438 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1439 {
1440         return (unsigned long)(div_u64(bytes, unit));
1441 }
1442
1443 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1444                                    u64 offset)
1445 {
1446         u64 bitmap_start;
1447         u64 bytes_per_bitmap;
1448
1449         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1450         bitmap_start = offset - ctl->start;
1451         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1452         bitmap_start *= bytes_per_bitmap;
1453         bitmap_start += ctl->start;
1454
1455         return bitmap_start;
1456 }
1457
1458 static int tree_insert_offset(struct rb_root *root, u64 offset,
1459                               struct rb_node *node, int bitmap)
1460 {
1461         struct rb_node **p = &root->rb_node;
1462         struct rb_node *parent = NULL;
1463         struct btrfs_free_space *info;
1464
1465         while (*p) {
1466                 parent = *p;
1467                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1468
1469                 if (offset < info->offset) {
1470                         p = &(*p)->rb_left;
1471                 } else if (offset > info->offset) {
1472                         p = &(*p)->rb_right;
1473                 } else {
1474                         /*
1475                          * we could have a bitmap entry and an extent entry
1476                          * share the same offset.  If this is the case, we want
1477                          * the extent entry to always be found first if we do a
1478                          * linear search through the tree, since we want to have
1479                          * the quickest allocation time, and allocating from an
1480                          * extent is faster than allocating from a bitmap.  So
1481                          * if we're inserting a bitmap and we find an entry at
1482                          * this offset, we want to go right, or after this entry
1483                          * logically.  If we are inserting an extent and we've
1484                          * found a bitmap, we want to go left, or before
1485                          * logically.
1486                          */
1487                         if (bitmap) {
1488                                 if (info->bitmap) {
1489                                         WARN_ON_ONCE(1);
1490                                         return -EEXIST;
1491                                 }
1492                                 p = &(*p)->rb_right;
1493                         } else {
1494                                 if (!info->bitmap) {
1495                                         WARN_ON_ONCE(1);
1496                                         return -EEXIST;
1497                                 }
1498                                 p = &(*p)->rb_left;
1499                         }
1500                 }
1501         }
1502
1503         rb_link_node(node, parent, p);
1504         rb_insert_color(node, root);
1505
1506         return 0;
1507 }
1508
1509 /*
1510  * searches the tree for the given offset.
1511  *
1512  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1513  * want a section that has at least bytes size and comes at or after the given
1514  * offset.
1515  */
1516 static struct btrfs_free_space *
1517 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1518                    u64 offset, int bitmap_only, int fuzzy)
1519 {
1520         struct rb_node *n = ctl->free_space_offset.rb_node;
1521         struct btrfs_free_space *entry, *prev = NULL;
1522
1523         /* find entry that is closest to the 'offset' */
1524         while (1) {
1525                 if (!n) {
1526                         entry = NULL;
1527                         break;
1528                 }
1529
1530                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1531                 prev = entry;
1532
1533                 if (offset < entry->offset)
1534                         n = n->rb_left;
1535                 else if (offset > entry->offset)
1536                         n = n->rb_right;
1537                 else
1538                         break;
1539         }
1540
1541         if (bitmap_only) {
1542                 if (!entry)
1543                         return NULL;
1544                 if (entry->bitmap)
1545                         return entry;
1546
1547                 /*
1548                  * bitmap entry and extent entry may share same offset,
1549                  * in that case, bitmap entry comes after extent entry.
1550                  */
1551                 n = rb_next(n);
1552                 if (!n)
1553                         return NULL;
1554                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1555                 if (entry->offset != offset)
1556                         return NULL;
1557
1558                 WARN_ON(!entry->bitmap);
1559                 return entry;
1560         } else if (entry) {
1561                 if (entry->bitmap) {
1562                         /*
1563                          * if previous extent entry covers the offset,
1564                          * we should return it instead of the bitmap entry
1565                          */
1566                         n = rb_prev(&entry->offset_index);
1567                         if (n) {
1568                                 prev = rb_entry(n, struct btrfs_free_space,
1569                                                 offset_index);
1570                                 if (!prev->bitmap &&
1571                                     prev->offset + prev->bytes > offset)
1572                                         entry = prev;
1573                         }
1574                 }
1575                 return entry;
1576         }
1577
1578         if (!prev)
1579                 return NULL;
1580
1581         /* find last entry before the 'offset' */
1582         entry = prev;
1583         if (entry->offset > offset) {
1584                 n = rb_prev(&entry->offset_index);
1585                 if (n) {
1586                         entry = rb_entry(n, struct btrfs_free_space,
1587                                         offset_index);
1588                         ASSERT(entry->offset <= offset);
1589                 } else {
1590                         if (fuzzy)
1591                                 return entry;
1592                         else
1593                                 return NULL;
1594                 }
1595         }
1596
1597         if (entry->bitmap) {
1598                 n = rb_prev(&entry->offset_index);
1599                 if (n) {
1600                         prev = rb_entry(n, struct btrfs_free_space,
1601                                         offset_index);
1602                         if (!prev->bitmap &&
1603                             prev->offset + prev->bytes > offset)
1604                                 return prev;
1605                 }
1606                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1607                         return entry;
1608         } else if (entry->offset + entry->bytes > offset)
1609                 return entry;
1610
1611         if (!fuzzy)
1612                 return NULL;
1613
1614         while (1) {
1615                 if (entry->bitmap) {
1616                         if (entry->offset + BITS_PER_BITMAP *
1617                             ctl->unit > offset)
1618                                 break;
1619                 } else {
1620                         if (entry->offset + entry->bytes > offset)
1621                                 break;
1622                 }
1623
1624                 n = rb_next(&entry->offset_index);
1625                 if (!n)
1626                         return NULL;
1627                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1628         }
1629         return entry;
1630 }
1631
1632 static inline void
1633 __unlink_free_space(struct btrfs_free_space_ctl *ctl,
1634                     struct btrfs_free_space *info)
1635 {
1636         rb_erase(&info->offset_index, &ctl->free_space_offset);
1637         ctl->free_extents--;
1638 }
1639
1640 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1641                               struct btrfs_free_space *info)
1642 {
1643         __unlink_free_space(ctl, info);
1644         ctl->free_space -= info->bytes;
1645 }
1646
1647 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1648                            struct btrfs_free_space *info)
1649 {
1650         int ret = 0;
1651
1652         ASSERT(info->bytes || info->bitmap);
1653         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1654                                  &info->offset_index, (info->bitmap != NULL));
1655         if (ret)
1656                 return ret;
1657
1658         ctl->free_space += info->bytes;
1659         ctl->free_extents++;
1660         return ret;
1661 }
1662
1663 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
1664 {
1665         struct btrfs_block_group *block_group = ctl->private;
1666         u64 max_bytes;
1667         u64 bitmap_bytes;
1668         u64 extent_bytes;
1669         u64 size = block_group->length;
1670         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
1671         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
1672
1673         max_bitmaps = max_t(u64, max_bitmaps, 1);
1674
1675         ASSERT(ctl->total_bitmaps <= max_bitmaps);
1676
1677         /*
1678          * The goal is to keep the total amount of memory used per 1gb of space
1679          * at or below 32k, so we need to adjust how much memory we allow to be
1680          * used by extent based free space tracking
1681          */
1682         if (size < SZ_1G)
1683                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
1684         else
1685                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
1686
1687         /*
1688          * we want to account for 1 more bitmap than what we have so we can make
1689          * sure we don't go over our overall goal of MAX_CACHE_BYTES_PER_GIG as
1690          * we add more bitmaps.
1691          */
1692         bitmap_bytes = (ctl->total_bitmaps + 1) * ctl->unit;
1693
1694         if (bitmap_bytes >= max_bytes) {
1695                 ctl->extents_thresh = 0;
1696                 return;
1697         }
1698
1699         /*
1700          * we want the extent entry threshold to always be at most 1/2 the max
1701          * bytes we can have, or whatever is less than that.
1702          */
1703         extent_bytes = max_bytes - bitmap_bytes;
1704         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
1705
1706         ctl->extents_thresh =
1707                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
1708 }
1709
1710 static inline void __bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1711                                        struct btrfs_free_space *info,
1712                                        u64 offset, u64 bytes)
1713 {
1714         unsigned long start, count;
1715
1716         start = offset_to_bit(info->offset, ctl->unit, offset);
1717         count = bytes_to_bits(bytes, ctl->unit);
1718         ASSERT(start + count <= BITS_PER_BITMAP);
1719
1720         bitmap_clear(info->bitmap, start, count);
1721
1722         info->bytes -= bytes;
1723         if (info->max_extent_size > ctl->unit)
1724                 info->max_extent_size = 0;
1725 }
1726
1727 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1728                               struct btrfs_free_space *info, u64 offset,
1729                               u64 bytes)
1730 {
1731         __bitmap_clear_bits(ctl, info, offset, bytes);
1732         ctl->free_space -= bytes;
1733 }
1734
1735 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1736                             struct btrfs_free_space *info, u64 offset,
1737                             u64 bytes)
1738 {
1739         unsigned long start, count;
1740
1741         start = offset_to_bit(info->offset, ctl->unit, offset);
1742         count = bytes_to_bits(bytes, ctl->unit);
1743         ASSERT(start + count <= BITS_PER_BITMAP);
1744
1745         bitmap_set(info->bitmap, start, count);
1746
1747         info->bytes += bytes;
1748         ctl->free_space += bytes;
1749 }
1750
1751 /*
1752  * If we can not find suitable extent, we will use bytes to record
1753  * the size of the max extent.
1754  */
1755 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1756                          struct btrfs_free_space *bitmap_info, u64 *offset,
1757                          u64 *bytes, bool for_alloc)
1758 {
1759         unsigned long found_bits = 0;
1760         unsigned long max_bits = 0;
1761         unsigned long bits, i;
1762         unsigned long next_zero;
1763         unsigned long extent_bits;
1764
1765         /*
1766          * Skip searching the bitmap if we don't have a contiguous section that
1767          * is large enough for this allocation.
1768          */
1769         if (for_alloc &&
1770             bitmap_info->max_extent_size &&
1771             bitmap_info->max_extent_size < *bytes) {
1772                 *bytes = bitmap_info->max_extent_size;
1773                 return -1;
1774         }
1775
1776         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1777                           max_t(u64, *offset, bitmap_info->offset));
1778         bits = bytes_to_bits(*bytes, ctl->unit);
1779
1780         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1781                 if (for_alloc && bits == 1) {
1782                         found_bits = 1;
1783                         break;
1784                 }
1785                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1786                                                BITS_PER_BITMAP, i);
1787                 extent_bits = next_zero - i;
1788                 if (extent_bits >= bits) {
1789                         found_bits = extent_bits;
1790                         break;
1791                 } else if (extent_bits > max_bits) {
1792                         max_bits = extent_bits;
1793                 }
1794                 i = next_zero;
1795         }
1796
1797         if (found_bits) {
1798                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1799                 *bytes = (u64)(found_bits) * ctl->unit;
1800                 return 0;
1801         }
1802
1803         *bytes = (u64)(max_bits) * ctl->unit;
1804         bitmap_info->max_extent_size = *bytes;
1805         return -1;
1806 }
1807
1808 static inline u64 get_max_extent_size(struct btrfs_free_space *entry)
1809 {
1810         if (entry->bitmap)
1811                 return entry->max_extent_size;
1812         return entry->bytes;
1813 }
1814
1815 /* Cache the size of the max extent in bytes */
1816 static struct btrfs_free_space *
1817 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1818                 unsigned long align, u64 *max_extent_size)
1819 {
1820         struct btrfs_free_space *entry;
1821         struct rb_node *node;
1822         u64 tmp;
1823         u64 align_off;
1824         int ret;
1825
1826         if (!ctl->free_space_offset.rb_node)
1827                 goto out;
1828
1829         entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset), 0, 1);
1830         if (!entry)
1831                 goto out;
1832
1833         for (node = &entry->offset_index; node; node = rb_next(node)) {
1834                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
1835                 if (entry->bytes < *bytes) {
1836                         *max_extent_size = max(get_max_extent_size(entry),
1837                                                *max_extent_size);
1838                         continue;
1839                 }
1840
1841                 /* make sure the space returned is big enough
1842                  * to match our requested alignment
1843                  */
1844                 if (*bytes >= align) {
1845                         tmp = entry->offset - ctl->start + align - 1;
1846                         tmp = div64_u64(tmp, align);
1847                         tmp = tmp * align + ctl->start;
1848                         align_off = tmp - entry->offset;
1849                 } else {
1850                         align_off = 0;
1851                         tmp = entry->offset;
1852                 }
1853
1854                 if (entry->bytes < *bytes + align_off) {
1855                         *max_extent_size = max(get_max_extent_size(entry),
1856                                                *max_extent_size);
1857                         continue;
1858                 }
1859
1860                 if (entry->bitmap) {
1861                         u64 size = *bytes;
1862
1863                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
1864                         if (!ret) {
1865                                 *offset = tmp;
1866                                 *bytes = size;
1867                                 return entry;
1868                         } else {
1869                                 *max_extent_size =
1870                                         max(get_max_extent_size(entry),
1871                                             *max_extent_size);
1872                         }
1873                         continue;
1874                 }
1875
1876                 *offset = tmp;
1877                 *bytes = entry->bytes - align_off;
1878                 return entry;
1879         }
1880 out:
1881         return NULL;
1882 }
1883
1884 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
1885                            struct btrfs_free_space *info, u64 offset)
1886 {
1887         info->offset = offset_to_bitmap(ctl, offset);
1888         info->bytes = 0;
1889         INIT_LIST_HEAD(&info->list);
1890         link_free_space(ctl, info);
1891         ctl->total_bitmaps++;
1892
1893         ctl->op->recalc_thresholds(ctl);
1894 }
1895
1896 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
1897                         struct btrfs_free_space *bitmap_info)
1898 {
1899         unlink_free_space(ctl, bitmap_info);
1900         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
1901         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
1902         ctl->total_bitmaps--;
1903         ctl->op->recalc_thresholds(ctl);
1904 }
1905
1906 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
1907                               struct btrfs_free_space *bitmap_info,
1908                               u64 *offset, u64 *bytes)
1909 {
1910         u64 end;
1911         u64 search_start, search_bytes;
1912         int ret;
1913
1914 again:
1915         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
1916
1917         /*
1918          * We need to search for bits in this bitmap.  We could only cover some
1919          * of the extent in this bitmap thanks to how we add space, so we need
1920          * to search for as much as it as we can and clear that amount, and then
1921          * go searching for the next bit.
1922          */
1923         search_start = *offset;
1924         search_bytes = ctl->unit;
1925         search_bytes = min(search_bytes, end - search_start + 1);
1926         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
1927                             false);
1928         if (ret < 0 || search_start != *offset)
1929                 return -EINVAL;
1930
1931         /* We may have found more bits than what we need */
1932         search_bytes = min(search_bytes, *bytes);
1933
1934         /* Cannot clear past the end of the bitmap */
1935         search_bytes = min(search_bytes, end - search_start + 1);
1936
1937         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes);
1938         *offset += search_bytes;
1939         *bytes -= search_bytes;
1940
1941         if (*bytes) {
1942                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
1943                 if (!bitmap_info->bytes)
1944                         free_bitmap(ctl, bitmap_info);
1945
1946                 /*
1947                  * no entry after this bitmap, but we still have bytes to
1948                  * remove, so something has gone wrong.
1949                  */
1950                 if (!next)
1951                         return -EINVAL;
1952
1953                 bitmap_info = rb_entry(next, struct btrfs_free_space,
1954                                        offset_index);
1955
1956                 /*
1957                  * if the next entry isn't a bitmap we need to return to let the
1958                  * extent stuff do its work.
1959                  */
1960                 if (!bitmap_info->bitmap)
1961                         return -EAGAIN;
1962
1963                 /*
1964                  * Ok the next item is a bitmap, but it may not actually hold
1965                  * the information for the rest of this free space stuff, so
1966                  * look for it, and if we don't find it return so we can try
1967                  * everything over again.
1968                  */
1969                 search_start = *offset;
1970                 search_bytes = ctl->unit;
1971                 ret = search_bitmap(ctl, bitmap_info, &search_start,
1972                                     &search_bytes, false);
1973                 if (ret < 0 || search_start != *offset)
1974                         return -EAGAIN;
1975
1976                 goto again;
1977         } else if (!bitmap_info->bytes)
1978                 free_bitmap(ctl, bitmap_info);
1979
1980         return 0;
1981 }
1982
1983 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
1984                                struct btrfs_free_space *info, u64 offset,
1985                                u64 bytes, enum btrfs_trim_state trim_state)
1986 {
1987         u64 bytes_to_set = 0;
1988         u64 end;
1989
1990         /*
1991          * This is a tradeoff to make bitmap trim state minimal.  We mark the
1992          * whole bitmap untrimmed if at any point we add untrimmed regions.
1993          */
1994         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED)
1995                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
1996
1997         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
1998
1999         bytes_to_set = min(end - offset, bytes);
2000
2001         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2002
2003         /*
2004          * We set some bytes, we have no idea what the max extent size is
2005          * anymore.
2006          */
2007         info->max_extent_size = 0;
2008
2009         return bytes_to_set;
2010
2011 }
2012
2013 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2014                       struct btrfs_free_space *info)
2015 {
2016         struct btrfs_block_group *block_group = ctl->private;
2017         struct btrfs_fs_info *fs_info = block_group->fs_info;
2018         bool forced = false;
2019
2020 #ifdef CONFIG_BTRFS_DEBUG
2021         if (btrfs_should_fragment_free_space(block_group))
2022                 forced = true;
2023 #endif
2024
2025         /*
2026          * If we are below the extents threshold then we can add this as an
2027          * extent, and don't have to deal with the bitmap
2028          */
2029         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2030                 /*
2031                  * If this block group has some small extents we don't want to
2032                  * use up all of our free slots in the cache with them, we want
2033                  * to reserve them to larger extents, however if we have plenty
2034                  * of cache left then go ahead an dadd them, no sense in adding
2035                  * the overhead of a bitmap if we don't have to.
2036                  */
2037                 if (info->bytes <= fs_info->sectorsize * 4) {
2038                         if (ctl->free_extents * 2 <= ctl->extents_thresh)
2039                                 return false;
2040                 } else {
2041                         return false;
2042                 }
2043         }
2044
2045         /*
2046          * The original block groups from mkfs can be really small, like 8
2047          * megabytes, so don't bother with a bitmap for those entries.  However
2048          * some block groups can be smaller than what a bitmap would cover but
2049          * are still large enough that they could overflow the 32k memory limit,
2050          * so allow those block groups to still be allowed to have a bitmap
2051          * entry.
2052          */
2053         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2054                 return false;
2055
2056         return true;
2057 }
2058
2059 static const struct btrfs_free_space_op free_space_op = {
2060         .recalc_thresholds      = recalculate_thresholds,
2061         .use_bitmap             = use_bitmap,
2062 };
2063
2064 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2065                               struct btrfs_free_space *info)
2066 {
2067         struct btrfs_free_space *bitmap_info;
2068         struct btrfs_block_group *block_group = NULL;
2069         int added = 0;
2070         u64 bytes, offset, bytes_added;
2071         enum btrfs_trim_state trim_state;
2072         int ret;
2073
2074         bytes = info->bytes;
2075         offset = info->offset;
2076         trim_state = info->trim_state;
2077
2078         if (!ctl->op->use_bitmap(ctl, info))
2079                 return 0;
2080
2081         if (ctl->op == &free_space_op)
2082                 block_group = ctl->private;
2083 again:
2084         /*
2085          * Since we link bitmaps right into the cluster we need to see if we
2086          * have a cluster here, and if so and it has our bitmap we need to add
2087          * the free space to that bitmap.
2088          */
2089         if (block_group && !list_empty(&block_group->cluster_list)) {
2090                 struct btrfs_free_cluster *cluster;
2091                 struct rb_node *node;
2092                 struct btrfs_free_space *entry;
2093
2094                 cluster = list_entry(block_group->cluster_list.next,
2095                                      struct btrfs_free_cluster,
2096                                      block_group_list);
2097                 spin_lock(&cluster->lock);
2098                 node = rb_first(&cluster->root);
2099                 if (!node) {
2100                         spin_unlock(&cluster->lock);
2101                         goto no_cluster_bitmap;
2102                 }
2103
2104                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2105                 if (!entry->bitmap) {
2106                         spin_unlock(&cluster->lock);
2107                         goto no_cluster_bitmap;
2108                 }
2109
2110                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2111                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2112                                                           bytes, trim_state);
2113                         bytes -= bytes_added;
2114                         offset += bytes_added;
2115                 }
2116                 spin_unlock(&cluster->lock);
2117                 if (!bytes) {
2118                         ret = 1;
2119                         goto out;
2120                 }
2121         }
2122
2123 no_cluster_bitmap:
2124         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2125                                          1, 0);
2126         if (!bitmap_info) {
2127                 ASSERT(added == 0);
2128                 goto new_bitmap;
2129         }
2130
2131         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2132                                           trim_state);
2133         bytes -= bytes_added;
2134         offset += bytes_added;
2135         added = 0;
2136
2137         if (!bytes) {
2138                 ret = 1;
2139                 goto out;
2140         } else
2141                 goto again;
2142
2143 new_bitmap:
2144         if (info && info->bitmap) {
2145                 add_new_bitmap(ctl, info, offset);
2146                 added = 1;
2147                 info = NULL;
2148                 goto again;
2149         } else {
2150                 spin_unlock(&ctl->tree_lock);
2151
2152                 /* no pre-allocated info, allocate a new one */
2153                 if (!info) {
2154                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2155                                                  GFP_NOFS);
2156                         if (!info) {
2157                                 spin_lock(&ctl->tree_lock);
2158                                 ret = -ENOMEM;
2159                                 goto out;
2160                         }
2161                 }
2162
2163                 /* allocate the bitmap */
2164                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2165                                                  GFP_NOFS);
2166                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2167                 spin_lock(&ctl->tree_lock);
2168                 if (!info->bitmap) {
2169                         ret = -ENOMEM;
2170                         goto out;
2171                 }
2172                 goto again;
2173         }
2174
2175 out:
2176         if (info) {
2177                 if (info->bitmap)
2178                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2179                                         info->bitmap);
2180                 kmem_cache_free(btrfs_free_space_cachep, info);
2181         }
2182
2183         return ret;
2184 }
2185
2186 /*
2187  * Free space merging rules:
2188  *  1) Merge trimmed areas together
2189  *  2) Let untrimmed areas coalesce with trimmed areas
2190  *  3) Always pull neighboring regions from bitmaps
2191  *
2192  * The above rules are for when we merge free space based on btrfs_trim_state.
2193  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2194  * same reason: to promote larger extent regions which makes life easier for
2195  * find_free_extent().  Rule 2 enables coalescing based on the common path
2196  * being returning free space from btrfs_finish_extent_commit().  So when free
2197  * space is trimmed, it will prevent aggregating trimmed new region and
2198  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2199  * and provide find_free_extent() with the largest extents possible hoping for
2200  * the reuse path.
2201  */
2202 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2203                           struct btrfs_free_space *info, bool update_stat)
2204 {
2205         struct btrfs_free_space *left_info;
2206         struct btrfs_free_space *right_info;
2207         bool merged = false;
2208         u64 offset = info->offset;
2209         u64 bytes = info->bytes;
2210         const bool is_trimmed = btrfs_free_space_trimmed(info);
2211
2212         /*
2213          * first we want to see if there is free space adjacent to the range we
2214          * are adding, if there is remove that struct and add a new one to
2215          * cover the entire range
2216          */
2217         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2218         if (right_info && rb_prev(&right_info->offset_index))
2219                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2220                                      struct btrfs_free_space, offset_index);
2221         else
2222                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2223
2224         /* See try_merge_free_space() comment. */
2225         if (right_info && !right_info->bitmap &&
2226             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2227                 if (update_stat)
2228                         unlink_free_space(ctl, right_info);
2229                 else
2230                         __unlink_free_space(ctl, right_info);
2231                 info->bytes += right_info->bytes;
2232                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2233                 merged = true;
2234         }
2235
2236         /* See try_merge_free_space() comment. */
2237         if (left_info && !left_info->bitmap &&
2238             left_info->offset + left_info->bytes == offset &&
2239             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2240                 if (update_stat)
2241                         unlink_free_space(ctl, left_info);
2242                 else
2243                         __unlink_free_space(ctl, left_info);
2244                 info->offset = left_info->offset;
2245                 info->bytes += left_info->bytes;
2246                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2247                 merged = true;
2248         }
2249
2250         return merged;
2251 }
2252
2253 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2254                                      struct btrfs_free_space *info,
2255                                      bool update_stat)
2256 {
2257         struct btrfs_free_space *bitmap;
2258         unsigned long i;
2259         unsigned long j;
2260         const u64 end = info->offset + info->bytes;
2261         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2262         u64 bytes;
2263
2264         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2265         if (!bitmap)
2266                 return false;
2267
2268         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2269         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2270         if (j == i)
2271                 return false;
2272         bytes = (j - i) * ctl->unit;
2273         info->bytes += bytes;
2274
2275         /* See try_merge_free_space() comment. */
2276         if (!btrfs_free_space_trimmed(bitmap))
2277                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2278
2279         if (update_stat)
2280                 bitmap_clear_bits(ctl, bitmap, end, bytes);
2281         else
2282                 __bitmap_clear_bits(ctl, bitmap, end, bytes);
2283
2284         if (!bitmap->bytes)
2285                 free_bitmap(ctl, bitmap);
2286
2287         return true;
2288 }
2289
2290 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2291                                        struct btrfs_free_space *info,
2292                                        bool update_stat)
2293 {
2294         struct btrfs_free_space *bitmap;
2295         u64 bitmap_offset;
2296         unsigned long i;
2297         unsigned long j;
2298         unsigned long prev_j;
2299         u64 bytes;
2300
2301         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2302         /* If we're on a boundary, try the previous logical bitmap. */
2303         if (bitmap_offset == info->offset) {
2304                 if (info->offset == 0)
2305                         return false;
2306                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2307         }
2308
2309         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2310         if (!bitmap)
2311                 return false;
2312
2313         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2314         j = 0;
2315         prev_j = (unsigned long)-1;
2316         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2317                 if (j > i)
2318                         break;
2319                 prev_j = j;
2320         }
2321         if (prev_j == i)
2322                 return false;
2323
2324         if (prev_j == (unsigned long)-1)
2325                 bytes = (i + 1) * ctl->unit;
2326         else
2327                 bytes = (i - prev_j) * ctl->unit;
2328
2329         info->offset -= bytes;
2330         info->bytes += bytes;
2331
2332         /* See try_merge_free_space() comment. */
2333         if (!btrfs_free_space_trimmed(bitmap))
2334                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2335
2336         if (update_stat)
2337                 bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2338         else
2339                 __bitmap_clear_bits(ctl, bitmap, info->offset, bytes);
2340
2341         if (!bitmap->bytes)
2342                 free_bitmap(ctl, bitmap);
2343
2344         return true;
2345 }
2346
2347 /*
2348  * We prefer always to allocate from extent entries, both for clustered and
2349  * non-clustered allocation requests. So when attempting to add a new extent
2350  * entry, try to see if there's adjacent free space in bitmap entries, and if
2351  * there is, migrate that space from the bitmaps to the extent.
2352  * Like this we get better chances of satisfying space allocation requests
2353  * because we attempt to satisfy them based on a single cache entry, and never
2354  * on 2 or more entries - even if the entries represent a contiguous free space
2355  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2356  * ends).
2357  */
2358 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2359                               struct btrfs_free_space *info,
2360                               bool update_stat)
2361 {
2362         /*
2363          * Only work with disconnected entries, as we can change their offset,
2364          * and must be extent entries.
2365          */
2366         ASSERT(!info->bitmap);
2367         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2368
2369         if (ctl->total_bitmaps > 0) {
2370                 bool stole_end;
2371                 bool stole_front = false;
2372
2373                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2374                 if (ctl->total_bitmaps > 0)
2375                         stole_front = steal_from_bitmap_to_front(ctl, info,
2376                                                                  update_stat);
2377
2378                 if (stole_end || stole_front)
2379                         try_merge_free_space(ctl, info, update_stat);
2380         }
2381 }
2382
2383 int __btrfs_add_free_space(struct btrfs_fs_info *fs_info,
2384                            struct btrfs_free_space_ctl *ctl,
2385                            u64 offset, u64 bytes,
2386                            enum btrfs_trim_state trim_state)
2387 {
2388         struct btrfs_block_group *block_group = ctl->private;
2389         struct btrfs_free_space *info;
2390         int ret = 0;
2391
2392         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2393         if (!info)
2394                 return -ENOMEM;
2395
2396         info->offset = offset;
2397         info->bytes = bytes;
2398         info->trim_state = trim_state;
2399         RB_CLEAR_NODE(&info->offset_index);
2400
2401         spin_lock(&ctl->tree_lock);
2402
2403         if (try_merge_free_space(ctl, info, true))
2404                 goto link;
2405
2406         /*
2407          * There was no extent directly to the left or right of this new
2408          * extent then we know we're going to have to allocate a new extent, so
2409          * before we do that see if we need to drop this into a bitmap
2410          */
2411         ret = insert_into_bitmap(ctl, info);
2412         if (ret < 0) {
2413                 goto out;
2414         } else if (ret) {
2415                 ret = 0;
2416                 goto out;
2417         }
2418 link:
2419         /*
2420          * Only steal free space from adjacent bitmaps if we're sure we're not
2421          * going to add the new free space to existing bitmap entries - because
2422          * that would mean unnecessary work that would be reverted. Therefore
2423          * attempt to steal space from bitmaps if we're adding an extent entry.
2424          */
2425         steal_from_bitmap(ctl, info, true);
2426
2427         ret = link_free_space(ctl, info);
2428         if (ret)
2429                 kmem_cache_free(btrfs_free_space_cachep, info);
2430 out:
2431         spin_unlock(&ctl->tree_lock);
2432
2433         if (ret) {
2434                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2435                 ASSERT(ret != -EEXIST);
2436         }
2437
2438         if (trim_state != BTRFS_TRIM_STATE_TRIMMED)
2439                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2440
2441         return ret;
2442 }
2443
2444 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2445                          u64 bytenr, u64 size)
2446 {
2447         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2448
2449         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2450                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2451
2452         return __btrfs_add_free_space(block_group->fs_info,
2453                                       block_group->free_space_ctl,
2454                                       bytenr, size, trim_state);
2455 }
2456
2457 /*
2458  * This is a subtle distinction because when adding free space back in general,
2459  * we want it to be added as untrimmed for async. But in the case where we add
2460  * it on loading of a block group, we want to consider it trimmed.
2461  */
2462 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2463                                        u64 bytenr, u64 size)
2464 {
2465         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2466
2467         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2468             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2469                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2470
2471         return __btrfs_add_free_space(block_group->fs_info,
2472                                       block_group->free_space_ctl,
2473                                       bytenr, size, trim_state);
2474 }
2475
2476 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2477                             u64 offset, u64 bytes)
2478 {
2479         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2480         struct btrfs_free_space *info;
2481         int ret;
2482         bool re_search = false;
2483
2484         spin_lock(&ctl->tree_lock);
2485
2486 again:
2487         ret = 0;
2488         if (!bytes)
2489                 goto out_lock;
2490
2491         info = tree_search_offset(ctl, offset, 0, 0);
2492         if (!info) {
2493                 /*
2494                  * oops didn't find an extent that matched the space we wanted
2495                  * to remove, look for a bitmap instead
2496                  */
2497                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2498                                           1, 0);
2499                 if (!info) {
2500                         /*
2501                          * If we found a partial bit of our free space in a
2502                          * bitmap but then couldn't find the other part this may
2503                          * be a problem, so WARN about it.
2504                          */
2505                         WARN_ON(re_search);
2506                         goto out_lock;
2507                 }
2508         }
2509
2510         re_search = false;
2511         if (!info->bitmap) {
2512                 unlink_free_space(ctl, info);
2513                 if (offset == info->offset) {
2514                         u64 to_free = min(bytes, info->bytes);
2515
2516                         info->bytes -= to_free;
2517                         info->offset += to_free;
2518                         if (info->bytes) {
2519                                 ret = link_free_space(ctl, info);
2520                                 WARN_ON(ret);
2521                         } else {
2522                                 kmem_cache_free(btrfs_free_space_cachep, info);
2523                         }
2524
2525                         offset += to_free;
2526                         bytes -= to_free;
2527                         goto again;
2528                 } else {
2529                         u64 old_end = info->bytes + info->offset;
2530
2531                         info->bytes = offset - info->offset;
2532                         ret = link_free_space(ctl, info);
2533                         WARN_ON(ret);
2534                         if (ret)
2535                                 goto out_lock;
2536
2537                         /* Not enough bytes in this entry to satisfy us */
2538                         if (old_end < offset + bytes) {
2539                                 bytes -= old_end - offset;
2540                                 offset = old_end;
2541                                 goto again;
2542                         } else if (old_end == offset + bytes) {
2543                                 /* all done */
2544                                 goto out_lock;
2545                         }
2546                         spin_unlock(&ctl->tree_lock);
2547
2548                         ret = __btrfs_add_free_space(block_group->fs_info, ctl,
2549                                                      offset + bytes,
2550                                                      old_end - (offset + bytes),
2551                                                      info->trim_state);
2552                         WARN_ON(ret);
2553                         goto out;
2554                 }
2555         }
2556
2557         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2558         if (ret == -EAGAIN) {
2559                 re_search = true;
2560                 goto again;
2561         }
2562 out_lock:
2563         spin_unlock(&ctl->tree_lock);
2564 out:
2565         return ret;
2566 }
2567
2568 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2569                            u64 bytes)
2570 {
2571         struct btrfs_fs_info *fs_info = block_group->fs_info;
2572         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2573         struct btrfs_free_space *info;
2574         struct rb_node *n;
2575         int count = 0;
2576
2577         spin_lock(&ctl->tree_lock);
2578         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2579                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2580                 if (info->bytes >= bytes && !block_group->ro)
2581                         count++;
2582                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2583                            info->offset, info->bytes,
2584                        (info->bitmap) ? "yes" : "no");
2585         }
2586         spin_unlock(&ctl->tree_lock);
2587         btrfs_info(fs_info, "block group has cluster?: %s",
2588                list_empty(&block_group->cluster_list) ? "no" : "yes");
2589         btrfs_info(fs_info,
2590                    "%d blocks of free space at or bigger than bytes is", count);
2591 }
2592
2593 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group)
2594 {
2595         struct btrfs_fs_info *fs_info = block_group->fs_info;
2596         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2597
2598         spin_lock_init(&ctl->tree_lock);
2599         ctl->unit = fs_info->sectorsize;
2600         ctl->start = block_group->start;
2601         ctl->private = block_group;
2602         ctl->op = &free_space_op;
2603         INIT_LIST_HEAD(&ctl->trimming_ranges);
2604         mutex_init(&ctl->cache_writeout_mutex);
2605
2606         /*
2607          * we only want to have 32k of ram per block group for keeping
2608          * track of free space, and if we pass 1/2 of that we want to
2609          * start converting things over to using bitmaps
2610          */
2611         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2612 }
2613
2614 /*
2615  * for a given cluster, put all of its extents back into the free
2616  * space cache.  If the block group passed doesn't match the block group
2617  * pointed to by the cluster, someone else raced in and freed the
2618  * cluster already.  In that case, we just return without changing anything
2619  */
2620 static int
2621 __btrfs_return_cluster_to_free_space(
2622                              struct btrfs_block_group *block_group,
2623                              struct btrfs_free_cluster *cluster)
2624 {
2625         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2626         struct btrfs_free_space *entry;
2627         struct rb_node *node;
2628
2629         spin_lock(&cluster->lock);
2630         if (cluster->block_group != block_group)
2631                 goto out;
2632
2633         cluster->block_group = NULL;
2634         cluster->window_start = 0;
2635         list_del_init(&cluster->block_group_list);
2636
2637         node = rb_first(&cluster->root);
2638         while (node) {
2639                 bool bitmap;
2640
2641                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2642                 node = rb_next(&entry->offset_index);
2643                 rb_erase(&entry->offset_index, &cluster->root);
2644                 RB_CLEAR_NODE(&entry->offset_index);
2645
2646                 bitmap = (entry->bitmap != NULL);
2647                 if (!bitmap) {
2648                         try_merge_free_space(ctl, entry, false);
2649                         steal_from_bitmap(ctl, entry, false);
2650                 }
2651                 tree_insert_offset(&ctl->free_space_offset,
2652                                    entry->offset, &entry->offset_index, bitmap);
2653         }
2654         cluster->root = RB_ROOT;
2655
2656 out:
2657         spin_unlock(&cluster->lock);
2658         btrfs_put_block_group(block_group);
2659         return 0;
2660 }
2661
2662 static void __btrfs_remove_free_space_cache_locked(
2663                                 struct btrfs_free_space_ctl *ctl)
2664 {
2665         struct btrfs_free_space *info;
2666         struct rb_node *node;
2667
2668         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2669                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2670                 if (!info->bitmap) {
2671                         unlink_free_space(ctl, info);
2672                         kmem_cache_free(btrfs_free_space_cachep, info);
2673                 } else {
2674                         free_bitmap(ctl, info);
2675                 }
2676
2677                 cond_resched_lock(&ctl->tree_lock);
2678         }
2679 }
2680
2681 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2682 {
2683         spin_lock(&ctl->tree_lock);
2684         __btrfs_remove_free_space_cache_locked(ctl);
2685         spin_unlock(&ctl->tree_lock);
2686 }
2687
2688 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2689 {
2690         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2691         struct btrfs_free_cluster *cluster;
2692         struct list_head *head;
2693
2694         spin_lock(&ctl->tree_lock);
2695         while ((head = block_group->cluster_list.next) !=
2696                &block_group->cluster_list) {
2697                 cluster = list_entry(head, struct btrfs_free_cluster,
2698                                      block_group_list);
2699
2700                 WARN_ON(cluster->block_group != block_group);
2701                 __btrfs_return_cluster_to_free_space(block_group, cluster);
2702
2703                 cond_resched_lock(&ctl->tree_lock);
2704         }
2705         __btrfs_remove_free_space_cache_locked(ctl);
2706         spin_unlock(&ctl->tree_lock);
2707
2708 }
2709
2710 /**
2711  * btrfs_is_free_space_trimmed - see if everything is trimmed
2712  * @block_group: block_group of interest
2713  *
2714  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
2715  */
2716 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
2717 {
2718         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2719         struct btrfs_free_space *info;
2720         struct rb_node *node;
2721         bool ret = true;
2722
2723         spin_lock(&ctl->tree_lock);
2724         node = rb_first(&ctl->free_space_offset);
2725
2726         while (node) {
2727                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2728
2729                 if (!btrfs_free_space_trimmed(info)) {
2730                         ret = false;
2731                         break;
2732                 }
2733
2734                 node = rb_next(node);
2735         }
2736
2737         spin_unlock(&ctl->tree_lock);
2738         return ret;
2739 }
2740
2741 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
2742                                u64 offset, u64 bytes, u64 empty_size,
2743                                u64 *max_extent_size)
2744 {
2745         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2746         struct btrfs_free_space *entry = NULL;
2747         u64 bytes_search = bytes + empty_size;
2748         u64 ret = 0;
2749         u64 align_gap = 0;
2750         u64 align_gap_len = 0;
2751         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2752
2753         spin_lock(&ctl->tree_lock);
2754         entry = find_free_space(ctl, &offset, &bytes_search,
2755                                 block_group->full_stripe_len, max_extent_size);
2756         if (!entry)
2757                 goto out;
2758
2759         ret = offset;
2760         if (entry->bitmap) {
2761                 bitmap_clear_bits(ctl, entry, offset, bytes);
2762                 if (!entry->bytes)
2763                         free_bitmap(ctl, entry);
2764         } else {
2765                 unlink_free_space(ctl, entry);
2766                 align_gap_len = offset - entry->offset;
2767                 align_gap = entry->offset;
2768                 align_gap_trim_state = entry->trim_state;
2769
2770                 entry->offset = offset + bytes;
2771                 WARN_ON(entry->bytes < bytes + align_gap_len);
2772
2773                 entry->bytes -= bytes + align_gap_len;
2774                 if (!entry->bytes)
2775                         kmem_cache_free(btrfs_free_space_cachep, entry);
2776                 else
2777                         link_free_space(ctl, entry);
2778         }
2779 out:
2780         spin_unlock(&ctl->tree_lock);
2781
2782         if (align_gap_len)
2783                 __btrfs_add_free_space(block_group->fs_info, ctl,
2784                                        align_gap, align_gap_len,
2785                                        align_gap_trim_state);
2786         return ret;
2787 }
2788
2789 /*
2790  * given a cluster, put all of its extents back into the free space
2791  * cache.  If a block group is passed, this function will only free
2792  * a cluster that belongs to the passed block group.
2793  *
2794  * Otherwise, it'll get a reference on the block group pointed to by the
2795  * cluster and remove the cluster from it.
2796  */
2797 int btrfs_return_cluster_to_free_space(
2798                                struct btrfs_block_group *block_group,
2799                                struct btrfs_free_cluster *cluster)
2800 {
2801         struct btrfs_free_space_ctl *ctl;
2802         int ret;
2803
2804         /* first, get a safe pointer to the block group */
2805         spin_lock(&cluster->lock);
2806         if (!block_group) {
2807                 block_group = cluster->block_group;
2808                 if (!block_group) {
2809                         spin_unlock(&cluster->lock);
2810                         return 0;
2811                 }
2812         } else if (cluster->block_group != block_group) {
2813                 /* someone else has already freed it don't redo their work */
2814                 spin_unlock(&cluster->lock);
2815                 return 0;
2816         }
2817         atomic_inc(&block_group->count);
2818         spin_unlock(&cluster->lock);
2819
2820         ctl = block_group->free_space_ctl;
2821
2822         /* now return any extents the cluster had on it */
2823         spin_lock(&ctl->tree_lock);
2824         ret = __btrfs_return_cluster_to_free_space(block_group, cluster);
2825         spin_unlock(&ctl->tree_lock);
2826
2827         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
2828
2829         /* finally drop our ref */
2830         btrfs_put_block_group(block_group);
2831         return ret;
2832 }
2833
2834 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
2835                                    struct btrfs_free_cluster *cluster,
2836                                    struct btrfs_free_space *entry,
2837                                    u64 bytes, u64 min_start,
2838                                    u64 *max_extent_size)
2839 {
2840         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2841         int err;
2842         u64 search_start = cluster->window_start;
2843         u64 search_bytes = bytes;
2844         u64 ret = 0;
2845
2846         search_start = min_start;
2847         search_bytes = bytes;
2848
2849         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
2850         if (err) {
2851                 *max_extent_size = max(get_max_extent_size(entry),
2852                                        *max_extent_size);
2853                 return 0;
2854         }
2855
2856         ret = search_start;
2857         __bitmap_clear_bits(ctl, entry, ret, bytes);
2858
2859         return ret;
2860 }
2861
2862 /*
2863  * given a cluster, try to allocate 'bytes' from it, returns 0
2864  * if it couldn't find anything suitably large, or a logical disk offset
2865  * if things worked out
2866  */
2867 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
2868                              struct btrfs_free_cluster *cluster, u64 bytes,
2869                              u64 min_start, u64 *max_extent_size)
2870 {
2871         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2872         struct btrfs_free_space *entry = NULL;
2873         struct rb_node *node;
2874         u64 ret = 0;
2875
2876         spin_lock(&cluster->lock);
2877         if (bytes > cluster->max_size)
2878                 goto out;
2879
2880         if (cluster->block_group != block_group)
2881                 goto out;
2882
2883         node = rb_first(&cluster->root);
2884         if (!node)
2885                 goto out;
2886
2887         entry = rb_entry(node, struct btrfs_free_space, offset_index);
2888         while (1) {
2889                 if (entry->bytes < bytes)
2890                         *max_extent_size = max(get_max_extent_size(entry),
2891                                                *max_extent_size);
2892
2893                 if (entry->bytes < bytes ||
2894                     (!entry->bitmap && entry->offset < min_start)) {
2895                         node = rb_next(&entry->offset_index);
2896                         if (!node)
2897                                 break;
2898                         entry = rb_entry(node, struct btrfs_free_space,
2899                                          offset_index);
2900                         continue;
2901                 }
2902
2903                 if (entry->bitmap) {
2904                         ret = btrfs_alloc_from_bitmap(block_group,
2905                                                       cluster, entry, bytes,
2906                                                       cluster->window_start,
2907                                                       max_extent_size);
2908                         if (ret == 0) {
2909                                 node = rb_next(&entry->offset_index);
2910                                 if (!node)
2911                                         break;
2912                                 entry = rb_entry(node, struct btrfs_free_space,
2913                                                  offset_index);
2914                                 continue;
2915                         }
2916                         cluster->window_start += bytes;
2917                 } else {
2918                         ret = entry->offset;
2919
2920                         entry->offset += bytes;
2921                         entry->bytes -= bytes;
2922                 }
2923
2924                 if (entry->bytes == 0)
2925                         rb_erase(&entry->offset_index, &cluster->root);
2926                 break;
2927         }
2928 out:
2929         spin_unlock(&cluster->lock);
2930
2931         if (!ret)
2932                 return 0;
2933
2934         spin_lock(&ctl->tree_lock);
2935
2936         ctl->free_space -= bytes;
2937         if (entry->bytes == 0) {
2938                 ctl->free_extents--;
2939                 if (entry->bitmap) {
2940                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2941                                         entry->bitmap);
2942                         ctl->total_bitmaps--;
2943                         ctl->op->recalc_thresholds(ctl);
2944                 }
2945                 kmem_cache_free(btrfs_free_space_cachep, entry);
2946         }
2947
2948         spin_unlock(&ctl->tree_lock);
2949
2950         return ret;
2951 }
2952
2953 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
2954                                 struct btrfs_free_space *entry,
2955                                 struct btrfs_free_cluster *cluster,
2956                                 u64 offset, u64 bytes,
2957                                 u64 cont1_bytes, u64 min_bytes)
2958 {
2959         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2960         unsigned long next_zero;
2961         unsigned long i;
2962         unsigned long want_bits;
2963         unsigned long min_bits;
2964         unsigned long found_bits;
2965         unsigned long max_bits = 0;
2966         unsigned long start = 0;
2967         unsigned long total_found = 0;
2968         int ret;
2969
2970         i = offset_to_bit(entry->offset, ctl->unit,
2971                           max_t(u64, offset, entry->offset));
2972         want_bits = bytes_to_bits(bytes, ctl->unit);
2973         min_bits = bytes_to_bits(min_bytes, ctl->unit);
2974
2975         /*
2976          * Don't bother looking for a cluster in this bitmap if it's heavily
2977          * fragmented.
2978          */
2979         if (entry->max_extent_size &&
2980             entry->max_extent_size < cont1_bytes)
2981                 return -ENOSPC;
2982 again:
2983         found_bits = 0;
2984         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
2985                 next_zero = find_next_zero_bit(entry->bitmap,
2986                                                BITS_PER_BITMAP, i);
2987                 if (next_zero - i >= min_bits) {
2988                         found_bits = next_zero - i;
2989                         if (found_bits > max_bits)
2990                                 max_bits = found_bits;
2991                         break;
2992                 }
2993                 if (next_zero - i > max_bits)
2994                         max_bits = next_zero - i;
2995                 i = next_zero;
2996         }
2997
2998         if (!found_bits) {
2999                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3000                 return -ENOSPC;
3001         }
3002
3003         if (!total_found) {
3004                 start = i;
3005                 cluster->max_size = 0;
3006         }
3007
3008         total_found += found_bits;
3009
3010         if (cluster->max_size < found_bits * ctl->unit)
3011                 cluster->max_size = found_bits * ctl->unit;
3012
3013         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3014                 i = next_zero + 1;
3015                 goto again;
3016         }
3017
3018         cluster->window_start = start * ctl->unit + entry->offset;
3019         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3020         ret = tree_insert_offset(&cluster->root, entry->offset,
3021                                  &entry->offset_index, 1);
3022         ASSERT(!ret); /* -EEXIST; Logic error */
3023
3024         trace_btrfs_setup_cluster(block_group, cluster,
3025                                   total_found * ctl->unit, 1);
3026         return 0;
3027 }
3028
3029 /*
3030  * This searches the block group for just extents to fill the cluster with.
3031  * Try to find a cluster with at least bytes total bytes, at least one
3032  * extent of cont1_bytes, and other clusters of at least min_bytes.
3033  */
3034 static noinline int
3035 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3036                         struct btrfs_free_cluster *cluster,
3037                         struct list_head *bitmaps, u64 offset, u64 bytes,
3038                         u64 cont1_bytes, u64 min_bytes)
3039 {
3040         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3041         struct btrfs_free_space *first = NULL;
3042         struct btrfs_free_space *entry = NULL;
3043         struct btrfs_free_space *last;
3044         struct rb_node *node;
3045         u64 window_free;
3046         u64 max_extent;
3047         u64 total_size = 0;
3048
3049         entry = tree_search_offset(ctl, offset, 0, 1);
3050         if (!entry)
3051                 return -ENOSPC;
3052
3053         /*
3054          * We don't want bitmaps, so just move along until we find a normal
3055          * extent entry.
3056          */
3057         while (entry->bitmap || entry->bytes < min_bytes) {
3058                 if (entry->bitmap && list_empty(&entry->list))
3059                         list_add_tail(&entry->list, bitmaps);
3060                 node = rb_next(&entry->offset_index);
3061                 if (!node)
3062                         return -ENOSPC;
3063                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3064         }
3065
3066         window_free = entry->bytes;
3067         max_extent = entry->bytes;
3068         first = entry;
3069         last = entry;
3070
3071         for (node = rb_next(&entry->offset_index); node;
3072              node = rb_next(&entry->offset_index)) {
3073                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3074
3075                 if (entry->bitmap) {
3076                         if (list_empty(&entry->list))
3077                                 list_add_tail(&entry->list, bitmaps);
3078                         continue;
3079                 }
3080
3081                 if (entry->bytes < min_bytes)
3082                         continue;
3083
3084                 last = entry;
3085                 window_free += entry->bytes;
3086                 if (entry->bytes > max_extent)
3087                         max_extent = entry->bytes;
3088         }
3089
3090         if (window_free < bytes || max_extent < cont1_bytes)
3091                 return -ENOSPC;
3092
3093         cluster->window_start = first->offset;
3094
3095         node = &first->offset_index;
3096
3097         /*
3098          * now we've found our entries, pull them out of the free space
3099          * cache and put them into the cluster rbtree
3100          */
3101         do {
3102                 int ret;
3103
3104                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3105                 node = rb_next(&entry->offset_index);
3106                 if (entry->bitmap || entry->bytes < min_bytes)
3107                         continue;
3108
3109                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3110                 ret = tree_insert_offset(&cluster->root, entry->offset,
3111                                          &entry->offset_index, 0);
3112                 total_size += entry->bytes;
3113                 ASSERT(!ret); /* -EEXIST; Logic error */
3114         } while (node && entry != last);
3115
3116         cluster->max_size = max_extent;
3117         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3118         return 0;
3119 }
3120
3121 /*
3122  * This specifically looks for bitmaps that may work in the cluster, we assume
3123  * that we have already failed to find extents that will work.
3124  */
3125 static noinline int
3126 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3127                      struct btrfs_free_cluster *cluster,
3128                      struct list_head *bitmaps, u64 offset, u64 bytes,
3129                      u64 cont1_bytes, u64 min_bytes)
3130 {
3131         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3132         struct btrfs_free_space *entry = NULL;
3133         int ret = -ENOSPC;
3134         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3135
3136         if (ctl->total_bitmaps == 0)
3137                 return -ENOSPC;
3138
3139         /*
3140          * The bitmap that covers offset won't be in the list unless offset
3141          * is just its start offset.
3142          */
3143         if (!list_empty(bitmaps))
3144                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3145
3146         if (!entry || entry->offset != bitmap_offset) {
3147                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3148                 if (entry && list_empty(&entry->list))
3149                         list_add(&entry->list, bitmaps);
3150         }
3151
3152         list_for_each_entry(entry, bitmaps, list) {
3153                 if (entry->bytes < bytes)
3154                         continue;
3155                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3156                                            bytes, cont1_bytes, min_bytes);
3157                 if (!ret)
3158                         return 0;
3159         }
3160
3161         /*
3162          * The bitmaps list has all the bitmaps that record free space
3163          * starting after offset, so no more search is required.
3164          */
3165         return -ENOSPC;
3166 }
3167
3168 /*
3169  * here we try to find a cluster of blocks in a block group.  The goal
3170  * is to find at least bytes+empty_size.
3171  * We might not find them all in one contiguous area.
3172  *
3173  * returns zero and sets up cluster if things worked out, otherwise
3174  * it returns -enospc
3175  */
3176 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3177                              struct btrfs_free_cluster *cluster,
3178                              u64 offset, u64 bytes, u64 empty_size)
3179 {
3180         struct btrfs_fs_info *fs_info = block_group->fs_info;
3181         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3182         struct btrfs_free_space *entry, *tmp;
3183         LIST_HEAD(bitmaps);
3184         u64 min_bytes;
3185         u64 cont1_bytes;
3186         int ret;
3187
3188         /*
3189          * Choose the minimum extent size we'll require for this
3190          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3191          * For metadata, allow allocates with smaller extents.  For
3192          * data, keep it dense.
3193          */
3194         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3195                 cont1_bytes = min_bytes = bytes + empty_size;
3196         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3197                 cont1_bytes = bytes;
3198                 min_bytes = fs_info->sectorsize;
3199         } else {
3200                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3201                 min_bytes = fs_info->sectorsize;
3202         }
3203
3204         spin_lock(&ctl->tree_lock);
3205
3206         /*
3207          * If we know we don't have enough space to make a cluster don't even
3208          * bother doing all the work to try and find one.
3209          */
3210         if (ctl->free_space < bytes) {
3211                 spin_unlock(&ctl->tree_lock);
3212                 return -ENOSPC;
3213         }
3214
3215         spin_lock(&cluster->lock);
3216
3217         /* someone already found a cluster, hooray */
3218         if (cluster->block_group) {
3219                 ret = 0;
3220                 goto out;
3221         }
3222
3223         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3224                                  min_bytes);
3225
3226         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3227                                       bytes + empty_size,
3228                                       cont1_bytes, min_bytes);
3229         if (ret)
3230                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3231                                            offset, bytes + empty_size,
3232                                            cont1_bytes, min_bytes);
3233
3234         /* Clear our temporary list */
3235         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3236                 list_del_init(&entry->list);
3237
3238         if (!ret) {
3239                 atomic_inc(&block_group->count);
3240                 list_add_tail(&cluster->block_group_list,
3241                               &block_group->cluster_list);
3242                 cluster->block_group = block_group;
3243         } else {
3244                 trace_btrfs_failed_cluster_setup(block_group);
3245         }
3246 out:
3247         spin_unlock(&cluster->lock);
3248         spin_unlock(&ctl->tree_lock);
3249
3250         return ret;
3251 }
3252
3253 /*
3254  * simple code to zero out a cluster
3255  */
3256 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3257 {
3258         spin_lock_init(&cluster->lock);
3259         spin_lock_init(&cluster->refill_lock);
3260         cluster->root = RB_ROOT;
3261         cluster->max_size = 0;
3262         cluster->fragmented = false;
3263         INIT_LIST_HEAD(&cluster->block_group_list);
3264         cluster->block_group = NULL;
3265 }
3266
3267 static int do_trimming(struct btrfs_block_group *block_group,
3268                        u64 *total_trimmed, u64 start, u64 bytes,
3269                        u64 reserved_start, u64 reserved_bytes,
3270                        enum btrfs_trim_state reserved_trim_state,
3271                        struct btrfs_trim_range *trim_entry)
3272 {
3273         struct btrfs_space_info *space_info = block_group->space_info;
3274         struct btrfs_fs_info *fs_info = block_group->fs_info;
3275         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3276         int ret;
3277         int update = 0;
3278         const u64 end = start + bytes;
3279         const u64 reserved_end = reserved_start + reserved_bytes;
3280         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3281         u64 trimmed = 0;
3282
3283         spin_lock(&space_info->lock);
3284         spin_lock(&block_group->lock);
3285         if (!block_group->ro) {
3286                 block_group->reserved += reserved_bytes;
3287                 space_info->bytes_reserved += reserved_bytes;
3288                 update = 1;
3289         }
3290         spin_unlock(&block_group->lock);
3291         spin_unlock(&space_info->lock);
3292
3293         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3294         if (!ret) {
3295                 *total_trimmed += trimmed;
3296                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3297         }
3298
3299         mutex_lock(&ctl->cache_writeout_mutex);
3300         if (reserved_start < start)
3301                 __btrfs_add_free_space(fs_info, ctl, reserved_start,
3302                                        start - reserved_start,
3303                                        reserved_trim_state);
3304         if (start + bytes < reserved_start + reserved_bytes)
3305                 __btrfs_add_free_space(fs_info, ctl, end, reserved_end - end,
3306                                        reserved_trim_state);
3307         __btrfs_add_free_space(fs_info, ctl, start, bytes, trim_state);
3308         list_del(&trim_entry->list);
3309         mutex_unlock(&ctl->cache_writeout_mutex);
3310
3311         if (update) {
3312                 spin_lock(&space_info->lock);
3313                 spin_lock(&block_group->lock);
3314                 if (block_group->ro)
3315                         space_info->bytes_readonly += reserved_bytes;
3316                 block_group->reserved -= reserved_bytes;
3317                 space_info->bytes_reserved -= reserved_bytes;
3318                 spin_unlock(&block_group->lock);
3319                 spin_unlock(&space_info->lock);
3320         }
3321
3322         return ret;
3323 }
3324
3325 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3326                           u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3327 {
3328         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3329         struct btrfs_free_space *entry;
3330         struct rb_node *node;
3331         int ret = 0;
3332         u64 extent_start;
3333         u64 extent_bytes;
3334         enum btrfs_trim_state extent_trim_state;
3335         u64 bytes;
3336
3337         while (start < end) {
3338                 struct btrfs_trim_range trim_entry;
3339
3340                 mutex_lock(&ctl->cache_writeout_mutex);
3341                 spin_lock(&ctl->tree_lock);
3342
3343                 if (ctl->free_space < minlen) {
3344                         spin_unlock(&ctl->tree_lock);
3345                         mutex_unlock(&ctl->cache_writeout_mutex);
3346                         break;
3347                 }
3348
3349                 entry = tree_search_offset(ctl, start, 0, 1);
3350                 if (!entry) {
3351                         spin_unlock(&ctl->tree_lock);
3352                         mutex_unlock(&ctl->cache_writeout_mutex);
3353                         break;
3354                 }
3355
3356                 /* skip bitmaps */
3357                 while (entry->bitmap) {
3358                         node = rb_next(&entry->offset_index);
3359                         if (!node) {
3360                                 spin_unlock(&ctl->tree_lock);
3361                                 mutex_unlock(&ctl->cache_writeout_mutex);
3362                                 goto out;
3363                         }
3364                         entry = rb_entry(node, struct btrfs_free_space,
3365                                          offset_index);
3366                 }
3367
3368                 if (entry->offset >= end) {
3369                         spin_unlock(&ctl->tree_lock);
3370                         mutex_unlock(&ctl->cache_writeout_mutex);
3371                         break;
3372                 }
3373
3374                 extent_start = entry->offset;
3375                 extent_bytes = entry->bytes;
3376                 extent_trim_state = entry->trim_state;
3377                 start = max(start, extent_start);
3378                 bytes = min(extent_start + extent_bytes, end) - start;
3379                 if (bytes < minlen) {
3380                         spin_unlock(&ctl->tree_lock);
3381                         mutex_unlock(&ctl->cache_writeout_mutex);
3382                         goto next;
3383                 }
3384
3385                 unlink_free_space(ctl, entry);
3386                 kmem_cache_free(btrfs_free_space_cachep, entry);
3387
3388                 spin_unlock(&ctl->tree_lock);
3389                 trim_entry.start = extent_start;
3390                 trim_entry.bytes = extent_bytes;
3391                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3392                 mutex_unlock(&ctl->cache_writeout_mutex);
3393
3394                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3395                                   extent_start, extent_bytes, extent_trim_state,
3396                                   &trim_entry);
3397                 if (ret)
3398                         break;
3399 next:
3400                 start += bytes;
3401
3402                 if (fatal_signal_pending(current)) {
3403                         ret = -ERESTARTSYS;
3404                         break;
3405                 }
3406
3407                 cond_resched();
3408         }
3409 out:
3410         return ret;
3411 }
3412
3413 /*
3414  * If we break out of trimming a bitmap prematurely, we should reset the
3415  * trimming bit.  In a rather contrieved case, it's possible to race here so
3416  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3417  *
3418  * start = start of bitmap
3419  * end = near end of bitmap
3420  *
3421  * Thread 1:                    Thread 2:
3422  * trim_bitmaps(start)
3423  *                              trim_bitmaps(end)
3424  *                              end_trimming_bitmap()
3425  * reset_trimming_bitmap()
3426  */
3427 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3428 {
3429         struct btrfs_free_space *entry;
3430
3431         spin_lock(&ctl->tree_lock);
3432         entry = tree_search_offset(ctl, offset, 1, 0);
3433         if (entry)
3434                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3435         spin_unlock(&ctl->tree_lock);
3436 }
3437
3438 static void end_trimming_bitmap(struct btrfs_free_space *entry)
3439 {
3440         if (btrfs_free_space_trimming_bitmap(entry))
3441                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3442 }
3443
3444 static int trim_bitmaps(struct btrfs_block_group *block_group,
3445                         u64 *total_trimmed, u64 start, u64 end, u64 minlen)
3446 {
3447         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3448         struct btrfs_free_space *entry;
3449         int ret = 0;
3450         int ret2;
3451         u64 bytes;
3452         u64 offset = offset_to_bitmap(ctl, start);
3453
3454         while (offset < end) {
3455                 bool next_bitmap = false;
3456                 struct btrfs_trim_range trim_entry;
3457
3458                 mutex_lock(&ctl->cache_writeout_mutex);
3459                 spin_lock(&ctl->tree_lock);
3460
3461                 if (ctl->free_space < minlen) {
3462                         spin_unlock(&ctl->tree_lock);
3463                         mutex_unlock(&ctl->cache_writeout_mutex);
3464                         break;
3465                 }
3466
3467                 entry = tree_search_offset(ctl, offset, 1, 0);
3468                 if (!entry || btrfs_free_space_trimmed(entry)) {
3469                         spin_unlock(&ctl->tree_lock);
3470                         mutex_unlock(&ctl->cache_writeout_mutex);
3471                         next_bitmap = true;
3472                         goto next;
3473                 }
3474
3475                 /*
3476                  * Async discard bitmap trimming begins at by setting the start
3477                  * to be key.objectid and the offset_to_bitmap() aligns to the
3478                  * start of the bitmap.  This lets us know we are fully
3479                  * scanning the bitmap rather than only some portion of it.
3480                  */
3481                 if (start == offset)
3482                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3483
3484                 bytes = minlen;
3485                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3486                 if (ret2 || start >= end) {
3487                         /*
3488                          * This keeps the invariant that all bytes are trimmed
3489                          * if BTRFS_TRIM_STATE_TRIMMED is set on a bitmap.
3490                          */
3491                         if (ret2 && !minlen)
3492                                 end_trimming_bitmap(entry);
3493                         else
3494                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3495                         spin_unlock(&ctl->tree_lock);
3496                         mutex_unlock(&ctl->cache_writeout_mutex);
3497                         next_bitmap = true;
3498                         goto next;
3499                 }
3500
3501                 bytes = min(bytes, end - start);
3502                 if (bytes < minlen) {
3503                         entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3504                         spin_unlock(&ctl->tree_lock);
3505                         mutex_unlock(&ctl->cache_writeout_mutex);
3506                         goto next;
3507                 }
3508
3509                 bitmap_clear_bits(ctl, entry, start, bytes);
3510                 if (entry->bytes == 0)
3511                         free_bitmap(ctl, entry);
3512
3513                 spin_unlock(&ctl->tree_lock);
3514                 trim_entry.start = start;
3515                 trim_entry.bytes = bytes;
3516                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3517                 mutex_unlock(&ctl->cache_writeout_mutex);
3518
3519                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3520                                   start, bytes, 0, &trim_entry);
3521                 if (ret) {
3522                         reset_trimming_bitmap(ctl, offset);
3523                         break;
3524                 }
3525 next:
3526                 if (next_bitmap) {
3527                         offset += BITS_PER_BITMAP * ctl->unit;
3528                         start = offset;
3529                 } else {
3530                         start += bytes;
3531                 }
3532
3533                 if (fatal_signal_pending(current)) {
3534                         if (start != offset)
3535                                 reset_trimming_bitmap(ctl, offset);
3536                         ret = -ERESTARTSYS;
3537                         break;
3538                 }
3539
3540                 cond_resched();
3541         }
3542
3543         return ret;
3544 }
3545
3546 void btrfs_get_block_group_trimming(struct btrfs_block_group *cache)
3547 {
3548         atomic_inc(&cache->trimming);
3549 }
3550
3551 void btrfs_put_block_group_trimming(struct btrfs_block_group *block_group)
3552 {
3553         struct btrfs_fs_info *fs_info = block_group->fs_info;
3554         struct extent_map_tree *em_tree;
3555         struct extent_map *em;
3556         bool cleanup;
3557
3558         spin_lock(&block_group->lock);
3559         cleanup = (atomic_dec_and_test(&block_group->trimming) &&
3560                    block_group->removed);
3561         spin_unlock(&block_group->lock);
3562
3563         if (cleanup) {
3564                 mutex_lock(&fs_info->chunk_mutex);
3565                 em_tree = &fs_info->mapping_tree;
3566                 write_lock(&em_tree->lock);
3567                 em = lookup_extent_mapping(em_tree, block_group->start,
3568                                            1);
3569                 BUG_ON(!em); /* logic error, can't happen */
3570                 remove_extent_mapping(em_tree, em);
3571                 write_unlock(&em_tree->lock);
3572                 mutex_unlock(&fs_info->chunk_mutex);
3573
3574                 /* once for us and once for the tree */
3575                 free_extent_map(em);
3576                 free_extent_map(em);
3577
3578                 /*
3579                  * We've left one free space entry and other tasks trimming
3580                  * this block group have left 1 entry each one. Free them.
3581                  */
3582                 __btrfs_remove_free_space_cache(block_group->free_space_ctl);
3583         }
3584 }
3585
3586 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3587                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3588 {
3589         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3590         int ret;
3591         u64 rem = 0;
3592
3593         *trimmed = 0;
3594
3595         spin_lock(&block_group->lock);
3596         if (block_group->removed) {
3597                 spin_unlock(&block_group->lock);
3598                 return 0;
3599         }
3600         btrfs_get_block_group_trimming(block_group);
3601         spin_unlock(&block_group->lock);
3602
3603         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen);
3604         if (ret)
3605                 goto out;
3606
3607         ret = trim_bitmaps(block_group, trimmed, start, end, minlen);
3608         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
3609         /* If we ended in the middle of a bitmap, reset the trimming flag */
3610         if (rem)
3611                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
3612 out:
3613         btrfs_put_block_group_trimming(block_group);
3614         return ret;
3615 }
3616
3617 /*
3618  * Find the left-most item in the cache tree, and then return the
3619  * smallest inode number in the item.
3620  *
3621  * Note: the returned inode number may not be the smallest one in
3622  * the tree, if the left-most item is a bitmap.
3623  */
3624 u64 btrfs_find_ino_for_alloc(struct btrfs_root *fs_root)
3625 {
3626         struct btrfs_free_space_ctl *ctl = fs_root->free_ino_ctl;
3627         struct btrfs_free_space *entry = NULL;
3628         u64 ino = 0;
3629
3630         spin_lock(&ctl->tree_lock);
3631
3632         if (RB_EMPTY_ROOT(&ctl->free_space_offset))
3633                 goto out;
3634
3635         entry = rb_entry(rb_first(&ctl->free_space_offset),
3636                          struct btrfs_free_space, offset_index);
3637
3638         if (!entry->bitmap) {
3639                 ino = entry->offset;
3640
3641                 unlink_free_space(ctl, entry);
3642                 entry->offset++;
3643                 entry->bytes--;
3644                 if (!entry->bytes)
3645                         kmem_cache_free(btrfs_free_space_cachep, entry);
3646                 else
3647                         link_free_space(ctl, entry);
3648         } else {
3649                 u64 offset = 0;
3650                 u64 count = 1;
3651                 int ret;
3652
3653                 ret = search_bitmap(ctl, entry, &offset, &count, true);
3654                 /* Logic error; Should be empty if it can't find anything */
3655                 ASSERT(!ret);
3656
3657                 ino = offset;
3658                 bitmap_clear_bits(ctl, entry, offset, 1);
3659                 if (entry->bytes == 0)
3660                         free_bitmap(ctl, entry);
3661         }
3662 out:
3663         spin_unlock(&ctl->tree_lock);
3664
3665         return ino;
3666 }
3667
3668 struct inode *lookup_free_ino_inode(struct btrfs_root *root,
3669                                     struct btrfs_path *path)
3670 {
3671         struct inode *inode = NULL;
3672
3673         spin_lock(&root->ino_cache_lock);
3674         if (root->ino_cache_inode)
3675                 inode = igrab(root->ino_cache_inode);
3676         spin_unlock(&root->ino_cache_lock);
3677         if (inode)
3678                 return inode;
3679
3680         inode = __lookup_free_space_inode(root, path, 0);
3681         if (IS_ERR(inode))
3682                 return inode;
3683
3684         spin_lock(&root->ino_cache_lock);
3685         if (!btrfs_fs_closing(root->fs_info))
3686                 root->ino_cache_inode = igrab(inode);
3687         spin_unlock(&root->ino_cache_lock);
3688
3689         return inode;
3690 }
3691
3692 int create_free_ino_inode(struct btrfs_root *root,
3693                           struct btrfs_trans_handle *trans,
3694                           struct btrfs_path *path)
3695 {
3696         return __create_free_space_inode(root, trans, path,
3697                                          BTRFS_FREE_INO_OBJECTID, 0);
3698 }
3699
3700 int load_free_ino_cache(struct btrfs_fs_info *fs_info, struct btrfs_root *root)
3701 {
3702         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3703         struct btrfs_path *path;
3704         struct inode *inode;
3705         int ret = 0;
3706         u64 root_gen = btrfs_root_generation(&root->root_item);
3707
3708         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3709                 return 0;
3710
3711         /*
3712          * If we're unmounting then just return, since this does a search on the
3713          * normal root and not the commit root and we could deadlock.
3714          */
3715         if (btrfs_fs_closing(fs_info))
3716                 return 0;
3717
3718         path = btrfs_alloc_path();
3719         if (!path)
3720                 return 0;
3721
3722         inode = lookup_free_ino_inode(root, path);
3723         if (IS_ERR(inode))
3724                 goto out;
3725
3726         if (root_gen != BTRFS_I(inode)->generation)
3727                 goto out_put;
3728
3729         ret = __load_free_space_cache(root, inode, ctl, path, 0);
3730
3731         if (ret < 0)
3732                 btrfs_err(fs_info,
3733                         "failed to load free ino cache for root %llu",
3734                         root->root_key.objectid);
3735 out_put:
3736         iput(inode);
3737 out:
3738         btrfs_free_path(path);
3739         return ret;
3740 }
3741
3742 int btrfs_write_out_ino_cache(struct btrfs_root *root,
3743                               struct btrfs_trans_handle *trans,
3744                               struct btrfs_path *path,
3745                               struct inode *inode)
3746 {
3747         struct btrfs_fs_info *fs_info = root->fs_info;
3748         struct btrfs_free_space_ctl *ctl = root->free_ino_ctl;
3749         int ret;
3750         struct btrfs_io_ctl io_ctl;
3751         bool release_metadata = true;
3752
3753         if (!btrfs_test_opt(fs_info, INODE_MAP_CACHE))
3754                 return 0;
3755
3756         memset(&io_ctl, 0, sizeof(io_ctl));
3757         ret = __btrfs_write_out_cache(root, inode, ctl, NULL, &io_ctl, trans);
3758         if (!ret) {
3759                 /*
3760                  * At this point writepages() didn't error out, so our metadata
3761                  * reservation is released when the writeback finishes, at
3762                  * inode.c:btrfs_finish_ordered_io(), regardless of it finishing
3763                  * with or without an error.
3764                  */
3765                 release_metadata = false;
3766                 ret = btrfs_wait_cache_io_root(root, trans, &io_ctl, path);
3767         }
3768
3769         if (ret) {
3770                 if (release_metadata)
3771                         btrfs_delalloc_release_metadata(BTRFS_I(inode),
3772                                         inode->i_size, true);
3773 #ifdef DEBUG
3774                 btrfs_err(fs_info,
3775                           "failed to write free ino cache for root %llu",
3776                           root->root_key.objectid);
3777 #endif
3778         }
3779
3780         return ret;
3781 }
3782
3783 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3784 /*
3785  * Use this if you need to make a bitmap or extent entry specifically, it
3786  * doesn't do any of the merging that add_free_space does, this acts a lot like
3787  * how the free space cache loading stuff works, so you can get really weird
3788  * configurations.
3789  */
3790 int test_add_free_space_entry(struct btrfs_block_group *cache,
3791                               u64 offset, u64 bytes, bool bitmap)
3792 {
3793         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3794         struct btrfs_free_space *info = NULL, *bitmap_info;
3795         void *map = NULL;
3796         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
3797         u64 bytes_added;
3798         int ret;
3799
3800 again:
3801         if (!info) {
3802                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
3803                 if (!info)
3804                         return -ENOMEM;
3805         }
3806
3807         if (!bitmap) {
3808                 spin_lock(&ctl->tree_lock);
3809                 info->offset = offset;
3810                 info->bytes = bytes;
3811                 info->max_extent_size = 0;
3812                 ret = link_free_space(ctl, info);
3813                 spin_unlock(&ctl->tree_lock);
3814                 if (ret)
3815                         kmem_cache_free(btrfs_free_space_cachep, info);
3816                 return ret;
3817         }
3818
3819         if (!map) {
3820                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
3821                 if (!map) {
3822                         kmem_cache_free(btrfs_free_space_cachep, info);
3823                         return -ENOMEM;
3824                 }
3825         }
3826
3827         spin_lock(&ctl->tree_lock);
3828         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3829                                          1, 0);
3830         if (!bitmap_info) {
3831                 info->bitmap = map;
3832                 map = NULL;
3833                 add_new_bitmap(ctl, info, offset);
3834                 bitmap_info = info;
3835                 info = NULL;
3836         }
3837
3838         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
3839                                           trim_state);
3840
3841         bytes -= bytes_added;
3842         offset += bytes_added;
3843         spin_unlock(&ctl->tree_lock);
3844
3845         if (bytes)
3846                 goto again;
3847
3848         if (info)
3849                 kmem_cache_free(btrfs_free_space_cachep, info);
3850         if (map)
3851                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
3852         return 0;
3853 }
3854
3855 /*
3856  * Checks to see if the given range is in the free space cache.  This is really
3857  * just used to check the absence of space, so if there is free space in the
3858  * range at all we will return 1.
3859  */
3860 int test_check_exists(struct btrfs_block_group *cache,
3861                       u64 offset, u64 bytes)
3862 {
3863         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
3864         struct btrfs_free_space *info;
3865         int ret = 0;
3866
3867         spin_lock(&ctl->tree_lock);
3868         info = tree_search_offset(ctl, offset, 0, 0);
3869         if (!info) {
3870                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
3871                                           1, 0);
3872                 if (!info)
3873                         goto out;
3874         }
3875
3876 have_info:
3877         if (info->bitmap) {
3878                 u64 bit_off, bit_bytes;
3879                 struct rb_node *n;
3880                 struct btrfs_free_space *tmp;
3881
3882                 bit_off = offset;
3883                 bit_bytes = ctl->unit;
3884                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
3885                 if (!ret) {
3886                         if (bit_off == offset) {
3887                                 ret = 1;
3888                                 goto out;
3889                         } else if (bit_off > offset &&
3890                                    offset + bytes > bit_off) {
3891                                 ret = 1;
3892                                 goto out;
3893                         }
3894                 }
3895
3896                 n = rb_prev(&info->offset_index);
3897                 while (n) {
3898                         tmp = rb_entry(n, struct btrfs_free_space,
3899                                        offset_index);
3900                         if (tmp->offset + tmp->bytes < offset)
3901                                 break;
3902                         if (offset + bytes < tmp->offset) {
3903                                 n = rb_prev(&tmp->offset_index);
3904                                 continue;
3905                         }
3906                         info = tmp;
3907                         goto have_info;
3908                 }
3909
3910                 n = rb_next(&info->offset_index);
3911                 while (n) {
3912                         tmp = rb_entry(n, struct btrfs_free_space,
3913                                        offset_index);
3914                         if (offset + bytes < tmp->offset)
3915                                 break;
3916                         if (tmp->offset + tmp->bytes < offset) {
3917                                 n = rb_next(&tmp->offset_index);
3918                                 continue;
3919                         }
3920                         info = tmp;
3921                         goto have_info;
3922                 }
3923
3924                 ret = 0;
3925                 goto out;
3926         }
3927
3928         if (info->offset == offset) {
3929                 ret = 1;
3930                 goto out;
3931         }
3932
3933         if (offset > info->offset && offset < info->offset + info->bytes)
3934                 ret = 1;
3935 out:
3936         spin_unlock(&ctl->tree_lock);
3937         return ret;
3938 }
3939 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */