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 "misc.h"
  15 #include "ctree.h"
  16 #include "free-space-cache.h"
  17 #include "transaction.h"
  18 #include "disk-io.h"
  19 #include "extent_io.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 #include "subpage.h"
  26 #include "inode-item.h"
  27
  28 #define BITS_PER_BITMAP         (PAGE_SIZE * 8UL)
  29 #define MAX_CACHE_BYTES_PER_GIG SZ_64K
  30 #define FORCE_EXTENT_THRESHOLD  SZ_1M
  31
  32 struct btrfs_trim_range {
  33         u64 start;
  34         u64 bytes;
  35         struct list_head list;
  36 };
  37
  38 static int link_free_space(struct btrfs_free_space_ctl *ctl,
  39                            struct btrfs_free_space *info);
  40 static void unlink_free_space(struct btrfs_free_space_ctl *ctl,
  41                               struct btrfs_free_space *info, bool update_stat);
  42 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
  43                          struct btrfs_free_space *bitmap_info, u64 *offset,
  44                          u64 *bytes, bool for_alloc);
  45 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
  46                         struct btrfs_free_space *bitmap_info);
  47 static void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
  48                               struct btrfs_free_space *info, u64 offset,
  49                               u64 bytes, bool update_stats);
  50
  51 static void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
  52 {
  53         struct btrfs_free_space *info;
  54         struct rb_node *node;
  55
  56         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
  57                 info = rb_entry(node, struct btrfs_free_space, offset_index);
  58                 if (!info->bitmap) {
  59                         unlink_free_space(ctl, info, true);
  60                         kmem_cache_free(btrfs_free_space_cachep, info);
  61                 } else {
  62                         free_bitmap(ctl, info);
  63                 }
  64
  65                 cond_resched_lock(&ctl->tree_lock);
  66         }
  67 }
  68
  69 static struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  70                                                struct btrfs_path *path,
  71                                                u64 offset)
  72 {
  73         struct btrfs_fs_info *fs_info = root->fs_info;
  74         struct btrfs_key key;
  75         struct btrfs_key location;
  76         struct btrfs_disk_key disk_key;
  77         struct btrfs_free_space_header *header;
  78         struct extent_buffer *leaf;
  79         struct inode *inode = NULL;
  80         unsigned nofs_flag;
  81         int ret;
  82
  83         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  84         key.offset = offset;
  85         key.type = 0;
  86
  87         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  88         if (ret < 0)
  89                 return ERR_PTR(ret);
  90         if (ret > 0) {
  91                 btrfs_release_path(path);
  92                 return ERR_PTR(-ENOENT);
  93         }
  94
  95         leaf = path->nodes[0];
  96         header = btrfs_item_ptr(leaf, path->slots[0],
  97                                 struct btrfs_free_space_header);
  98         btrfs_free_space_key(leaf, header, &disk_key);
  99         btrfs_disk_key_to_cpu(&location, &disk_key);
 100         btrfs_release_path(path);
 101
 102         /*
 103          * We are often under a trans handle at this point, so we need to make
 104          * sure NOFS is set to keep us from deadlocking.
 105          */
 106         nofs_flag = memalloc_nofs_save();
 107         inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
 108         btrfs_release_path(path);
 109         memalloc_nofs_restore(nofs_flag);
 110         if (IS_ERR(inode))
 111                 return inode;
 112
 113         mapping_set_gfp_mask(inode->i_mapping,
 114                         mapping_gfp_constraint(inode->i_mapping,
 115                         ~(__GFP_FS | __GFP_HIGHMEM)));
 116
 117         return inode;
 118 }
 119
 120 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
 121                 struct btrfs_path *path)
 122 {
 123         struct btrfs_fs_info *fs_info = block_group->fs_info;
 124         struct inode *inode = NULL;
 125         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 126
 127         spin_lock(&block_group->lock);
 128         if (block_group->inode)
 129                 inode = igrab(block_group->inode);
 130         spin_unlock(&block_group->lock);
 131         if (inode)
 132                 return inode;
 133
 134         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 135                                           block_group->start);
 136         if (IS_ERR(inode))
 137                 return inode;
 138
 139         spin_lock(&block_group->lock);
 140         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 141                 btrfs_info(fs_info, "Old style space inode found, converting.");
 142                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 143                         BTRFS_INODE_NODATACOW;
 144                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 145         }
 146
 147         if (!test_and_set_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags))
 148                 block_group->inode = igrab(inode);
 149         spin_unlock(&block_group->lock);
 150
 151         return inode;
 152 }
 153
 154 static int __create_free_space_inode(struct btrfs_root *root,
 155                                      struct btrfs_trans_handle *trans,
 156                                      struct btrfs_path *path,
 157                                      u64 ino, u64 offset)
 158 {
 159         struct btrfs_key key;
 160         struct btrfs_disk_key disk_key;
 161         struct btrfs_free_space_header *header;
 162         struct btrfs_inode_item *inode_item;
 163         struct extent_buffer *leaf;
 164         /* We inline CRCs for the free disk space cache */
 165         const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
 166                           BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 167         int ret;
 168
 169         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 170         if (ret)
 171                 return ret;
 172
 173         leaf = path->nodes[0];
 174         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 175                                     struct btrfs_inode_item);
 176         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 177         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 178                              sizeof(*inode_item));
 179         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 180         btrfs_set_inode_size(leaf, inode_item, 0);
 181         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 182         btrfs_set_inode_uid(leaf, inode_item, 0);
 183         btrfs_set_inode_gid(leaf, inode_item, 0);
 184         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 185         btrfs_set_inode_flags(leaf, inode_item, flags);
 186         btrfs_set_inode_nlink(leaf, inode_item, 1);
 187         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 188         btrfs_set_inode_block_group(leaf, inode_item, offset);
 189         btrfs_mark_buffer_dirty(leaf);
 190         btrfs_release_path(path);
 191
 192         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 193         key.offset = offset;
 194         key.type = 0;
 195         ret = btrfs_insert_empty_item(trans, root, path, &key,
 196                                       sizeof(struct btrfs_free_space_header));
 197         if (ret < 0) {
 198                 btrfs_release_path(path);
 199                 return ret;
 200         }
 201
 202         leaf = path->nodes[0];
 203         header = btrfs_item_ptr(leaf, path->slots[0],
 204                                 struct btrfs_free_space_header);
 205         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 206         btrfs_set_free_space_key(leaf, header, &disk_key);
 207         btrfs_mark_buffer_dirty(leaf);
 208         btrfs_release_path(path);
 209
 210         return 0;
 211 }
 212
 213 int create_free_space_inode(struct btrfs_trans_handle *trans,
 214                             struct btrfs_block_group *block_group,
 215                             struct btrfs_path *path)
 216 {
 217         int ret;
 218         u64 ino;
 219
 220         ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
 221         if (ret < 0)
 222                 return ret;
 223
 224         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 225                                          ino, block_group->start);
 226 }
 227
 228 /*
 229  * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
 230  * handles lookup, otherwise it takes ownership and iputs the inode.
 231  * Don't reuse an inode pointer after passing it into this function.
 232  */
 233 int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
 234                                   struct inode *inode,
 235                                   struct btrfs_block_group *block_group)
 236 {
 237         struct btrfs_path *path;
 238         struct btrfs_key key;
 239         int ret = 0;
 240
 241         path = btrfs_alloc_path();
 242         if (!path)
 243                 return -ENOMEM;
 244
 245         if (!inode)
 246                 inode = lookup_free_space_inode(block_group, path);
 247         if (IS_ERR(inode)) {
 248                 if (PTR_ERR(inode) != -ENOENT)
 249                         ret = PTR_ERR(inode);
 250                 goto out;
 251         }
 252         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 253         if (ret) {
 254                 btrfs_add_delayed_iput(inode);
 255                 goto out;
 256         }
 257         clear_nlink(inode);
 258         /* One for the block groups ref */
 259         spin_lock(&block_group->lock);
 260         if (test_and_clear_bit(BLOCK_GROUP_FLAG_IREF, &block_group->runtime_flags)) {
 261                 block_group->inode = NULL;
 262                 spin_unlock(&block_group->lock);
 263                 iput(inode);
 264         } else {
 265                 spin_unlock(&block_group->lock);
 266         }
 267         /* One for the lookup ref */
 268         btrfs_add_delayed_iput(inode);
 269
 270         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 271         key.type = 0;
 272         key.offset = block_group->start;
 273         ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
 274                                 -1, 1);
 275         if (ret) {
 276                 if (ret > 0)
 277                         ret = 0;
 278                 goto out;
 279         }
 280         ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
 281 out:
 282         btrfs_free_path(path);
 283         return ret;
 284 }
 285
 286 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 287                                        struct btrfs_block_rsv *rsv)
 288 {
 289         u64 needed_bytes;
 290         int ret;
 291
 292         /* 1 for slack space, 1 for updating the inode */
 293         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 294                 btrfs_calc_metadata_size(fs_info, 1);
 295
 296         spin_lock(&rsv->lock);
 297         if (rsv->reserved < needed_bytes)
 298                 ret = -ENOSPC;
 299         else
 300                 ret = 0;
 301         spin_unlock(&rsv->lock);
 302         return ret;
 303 }
 304
 305 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 306                                     struct btrfs_block_group *block_group,
 307                                     struct inode *vfs_inode)
 308 {
 309         struct btrfs_truncate_control control = {
 310                 .inode = BTRFS_I(vfs_inode),
 311                 .new_size = 0,
 312                 .ino = btrfs_ino(BTRFS_I(vfs_inode)),
 313                 .min_type = BTRFS_EXTENT_DATA_KEY,
 314                 .clear_extent_range = true,
 315         };
 316         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
 317         struct btrfs_root *root = inode->root;
 318         struct extent_state *cached_state = NULL;
 319         int ret = 0;
 320         bool locked = false;
 321
 322         if (block_group) {
 323                 struct btrfs_path *path = btrfs_alloc_path();
 324
 325                 if (!path) {
 326                         ret = -ENOMEM;
 327                         goto fail;
 328                 }
 329                 locked = true;
 330                 mutex_lock(&trans->transaction->cache_write_mutex);
 331                 if (!list_empty(&block_group->io_list)) {
 332                         list_del_init(&block_group->io_list);
 333
 334                         btrfs_wait_cache_io(trans, block_group, path);
 335                         btrfs_put_block_group(block_group);
 336                 }
 337
 338                 /*
 339                  * now that we've truncated the cache away, its no longer
 340                  * setup or written
 341                  */
 342                 spin_lock(&block_group->lock);
 343                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 344                 spin_unlock(&block_group->lock);
 345                 btrfs_free_path(path);
 346         }
 347
 348         btrfs_i_size_write(inode, 0);
 349         truncate_pagecache(vfs_inode, 0);
 350
 351         lock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 352         btrfs_drop_extent_map_range(inode, 0, (u64)-1, false);
 353
 354         /*
 355          * We skip the throttling logic for free space cache inodes, so we don't
 356          * need to check for -EAGAIN.
 357          */
 358         ret = btrfs_truncate_inode_items(trans, root, &control);
 359
 360         inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
 361         btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
 362
 363         unlock_extent(&inode->io_tree, 0, (u64)-1, &cached_state);
 364         if (ret)
 365                 goto fail;
 366
 367         ret = btrfs_update_inode(trans, root, inode);
 368
 369 fail:
 370         if (locked)
 371                 mutex_unlock(&trans->transaction->cache_write_mutex);
 372         if (ret)
 373                 btrfs_abort_transaction(trans, ret);
 374
 375         return ret;
 376 }
 377
 378 static void readahead_cache(struct inode *inode)
 379 {
 380         struct file_ra_state ra;
 381         unsigned long last_index;
 382
 383         file_ra_state_init(&ra, inode->i_mapping);
 384         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 385
 386         page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
 387 }
 388
 389 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 390                        int write)
 391 {
 392         int num_pages;
 393
 394         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 395
 396         /* Make sure we can fit our crcs and generation into the first page */
 397         if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 398                 return -ENOSPC;
 399
 400         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 401
 402         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 403         if (!io_ctl->pages)
 404                 return -ENOMEM;
 405
 406         io_ctl->num_pages = num_pages;
 407         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 408         io_ctl->inode = inode;
 409
 410         return 0;
 411 }
 412 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 413
 414 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 415 {
 416         kfree(io_ctl->pages);
 417         io_ctl->pages = NULL;
 418 }
 419
 420 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 421 {
 422         if (io_ctl->cur) {
 423                 io_ctl->cur = NULL;
 424                 io_ctl->orig = NULL;
 425         }
 426 }
 427
 428 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 429 {
 430         ASSERT(io_ctl->index < io_ctl->num_pages);
 431         io_ctl->page = io_ctl->pages[io_ctl->index++];
 432         io_ctl->cur = page_address(io_ctl->page);
 433         io_ctl->orig = io_ctl->cur;
 434         io_ctl->size = PAGE_SIZE;
 435         if (clear)
 436                 clear_page(io_ctl->cur);
 437 }
 438
 439 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 440 {
 441         int i;
 442
 443         io_ctl_unmap_page(io_ctl);
 444
 445         for (i = 0; i < io_ctl->num_pages; i++) {
 446                 if (io_ctl->pages[i]) {
 447                         btrfs_page_clear_checked(io_ctl->fs_info,
 448                                         io_ctl->pages[i],
 449                                         page_offset(io_ctl->pages[i]),
 450                                         PAGE_SIZE);
 451                         unlock_page(io_ctl->pages[i]);
 452                         put_page(io_ctl->pages[i]);
 453                 }
 454         }
 455 }
 456
 457 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 458 {
 459         struct page *page;
 460         struct inode *inode = io_ctl->inode;
 461         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 462         int i;
 463
 464         for (i = 0; i < io_ctl->num_pages; i++) {
 465                 int ret;
 466
 467                 page = find_or_create_page(inode->i_mapping, i, mask);
 468                 if (!page) {
 469                         io_ctl_drop_pages(io_ctl);
 470                         return -ENOMEM;
 471                 }
 472
 473                 ret = set_page_extent_mapped(page);
 474                 if (ret < 0) {
 475                         unlock_page(page);
 476                         put_page(page);
 477                         io_ctl_drop_pages(io_ctl);
 478                         return ret;
 479                 }
 480
 481                 io_ctl->pages[i] = page;
 482                 if (uptodate && !PageUptodate(page)) {
 483                         btrfs_read_folio(NULL, page_folio(page));
 484                         lock_page(page);
 485                         if (page->mapping != inode->i_mapping) {
 486                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 487                                           "free space cache page truncated");
 488                                 io_ctl_drop_pages(io_ctl);
 489                                 return -EIO;
 490                         }
 491                         if (!PageUptodate(page)) {
 492                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 493                                            "error reading free space cache");
 494                                 io_ctl_drop_pages(io_ctl);
 495                                 return -EIO;
 496                         }
 497                 }
 498         }
 499
 500         for (i = 0; i < io_ctl->num_pages; i++)
 501                 clear_page_dirty_for_io(io_ctl->pages[i]);
 502
 503         return 0;
 504 }
 505
 506 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 507 {
 508         io_ctl_map_page(io_ctl, 1);
 509
 510         /*
 511          * Skip the csum areas.  If we don't check crcs then we just have a
 512          * 64bit chunk at the front of the first page.
 513          */
 514         io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 515         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 516
 517         put_unaligned_le64(generation, io_ctl->cur);
 518         io_ctl->cur += sizeof(u64);
 519 }
 520
 521 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 522 {
 523         u64 cache_gen;
 524
 525         /*
 526          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 527          * chunk at the front of the first page.
 528          */
 529         io_ctl->cur += sizeof(u32) * io_ctl->num_pages;
 530         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 531
 532         cache_gen = get_unaligned_le64(io_ctl->cur);
 533         if (cache_gen != generation) {
 534                 btrfs_err_rl(io_ctl->fs_info,
 535                         "space cache generation (%llu) does not match inode (%llu)",
 536                                 cache_gen, generation);
 537                 io_ctl_unmap_page(io_ctl);
 538                 return -EIO;
 539         }
 540         io_ctl->cur += sizeof(u64);
 541         return 0;
 542 }
 543
 544 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 545 {
 546         u32 *tmp;
 547         u32 crc = ~(u32)0;
 548         unsigned offset = 0;
 549
 550         if (index == 0)
 551                 offset = sizeof(u32) * io_ctl->num_pages;
 552
 553         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 554         btrfs_crc32c_final(crc, (u8 *)&crc);
 555         io_ctl_unmap_page(io_ctl);
 556         tmp = page_address(io_ctl->pages[0]);
 557         tmp += index;
 558         *tmp = crc;
 559 }
 560
 561 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 562 {
 563         u32 *tmp, val;
 564         u32 crc = ~(u32)0;
 565         unsigned offset = 0;
 566
 567         if (index == 0)
 568                 offset = sizeof(u32) * io_ctl->num_pages;
 569
 570         tmp = page_address(io_ctl->pages[0]);
 571         tmp += index;
 572         val = *tmp;
 573
 574         io_ctl_map_page(io_ctl, 0);
 575         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 576         btrfs_crc32c_final(crc, (u8 *)&crc);
 577         if (val != crc) {
 578                 btrfs_err_rl(io_ctl->fs_info,
 579                         "csum mismatch on free space cache");
 580                 io_ctl_unmap_page(io_ctl);
 581                 return -EIO;
 582         }
 583
 584         return 0;
 585 }
 586
 587 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 588                             void *bitmap)
 589 {
 590         struct btrfs_free_space_entry *entry;
 591
 592         if (!io_ctl->cur)
 593                 return -ENOSPC;
 594
 595         entry = io_ctl->cur;
 596         put_unaligned_le64(offset, &entry->offset);
 597         put_unaligned_le64(bytes, &entry->bytes);
 598         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 599                 BTRFS_FREE_SPACE_EXTENT;
 600         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 601         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 602
 603         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 604                 return 0;
 605
 606         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 607
 608         /* No more pages to map */
 609         if (io_ctl->index >= io_ctl->num_pages)
 610                 return 0;
 611
 612         /* map the next page */
 613         io_ctl_map_page(io_ctl, 1);
 614         return 0;
 615 }
 616
 617 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 618 {
 619         if (!io_ctl->cur)
 620                 return -ENOSPC;
 621
 622         /*
 623          * If we aren't at the start of the current page, unmap this one and
 624          * map the next one if there is any left.
 625          */
 626         if (io_ctl->cur != io_ctl->orig) {
 627                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 628                 if (io_ctl->index >= io_ctl->num_pages)
 629                         return -ENOSPC;
 630                 io_ctl_map_page(io_ctl, 0);
 631         }
 632
 633         copy_page(io_ctl->cur, bitmap);
 634         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 635         if (io_ctl->index < io_ctl->num_pages)
 636                 io_ctl_map_page(io_ctl, 0);
 637         return 0;
 638 }
 639
 640 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 641 {
 642         /*
 643          * If we're not on the boundary we know we've modified the page and we
 644          * need to crc the page.
 645          */
 646         if (io_ctl->cur != io_ctl->orig)
 647                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 648         else
 649                 io_ctl_unmap_page(io_ctl);
 650
 651         while (io_ctl->index < io_ctl->num_pages) {
 652                 io_ctl_map_page(io_ctl, 1);
 653                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 654         }
 655 }
 656
 657 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 658                             struct btrfs_free_space *entry, u8 *type)
 659 {
 660         struct btrfs_free_space_entry *e;
 661         int ret;
 662
 663         if (!io_ctl->cur) {
 664                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 665                 if (ret)
 666                         return ret;
 667         }
 668
 669         e = io_ctl->cur;
 670         entry->offset = get_unaligned_le64(&e->offset);
 671         entry->bytes = get_unaligned_le64(&e->bytes);
 672         *type = e->type;
 673         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 674         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 675
 676         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 677                 return 0;
 678
 679         io_ctl_unmap_page(io_ctl);
 680
 681         return 0;
 682 }
 683
 684 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 685                               struct btrfs_free_space *entry)
 686 {
 687         int ret;
 688
 689         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 690         if (ret)
 691                 return ret;
 692
 693         copy_page(entry->bitmap, io_ctl->cur);
 694         io_ctl_unmap_page(io_ctl);
 695
 696         return 0;
 697 }
 698
 699 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 700 {
 701         struct btrfs_block_group *block_group = ctl->block_group;
 702         u64 max_bytes;
 703         u64 bitmap_bytes;
 704         u64 extent_bytes;
 705         u64 size = block_group->length;
 706         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
 707         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
 708
 709         max_bitmaps = max_t(u64, max_bitmaps, 1);
 710
 711         if (ctl->total_bitmaps > max_bitmaps)
 712                 btrfs_err(block_group->fs_info,
 713 "invalid free space control: bg start=%llu len=%llu total_bitmaps=%u unit=%u max_bitmaps=%llu bytes_per_bg=%llu",
 714                           block_group->start, block_group->length,
 715                           ctl->total_bitmaps, ctl->unit, max_bitmaps,
 716                           bytes_per_bg);
 717         ASSERT(ctl->total_bitmaps <= max_bitmaps);
 718
 719         /*
 720          * We are trying to keep the total amount of memory used per 1GiB of
 721          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
 722          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
 723          * bitmaps, we may end up using more memory than this.
 724          */
 725         if (size < SZ_1G)
 726                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
 727         else
 728                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
 729
 730         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
 731
 732         /*
 733          * we want the extent entry threshold to always be at most 1/2 the max
 734          * bytes we can have, or whatever is less than that.
 735          */
 736         extent_bytes = max_bytes - bitmap_bytes;
 737         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
 738
 739         ctl->extents_thresh =
 740                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
 741 }
 742
 743 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 744                                    struct btrfs_free_space_ctl *ctl,
 745                                    struct btrfs_path *path, u64 offset)
 746 {
 747         struct btrfs_fs_info *fs_info = root->fs_info;
 748         struct btrfs_free_space_header *header;
 749         struct extent_buffer *leaf;
 750         struct btrfs_io_ctl io_ctl;
 751         struct btrfs_key key;
 752         struct btrfs_free_space *e, *n;
 753         LIST_HEAD(bitmaps);
 754         u64 num_entries;
 755         u64 num_bitmaps;
 756         u64 generation;
 757         u8 type;
 758         int ret = 0;
 759
 760         /* Nothing in the space cache, goodbye */
 761         if (!i_size_read(inode))
 762                 return 0;
 763
 764         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 765         key.offset = offset;
 766         key.type = 0;
 767
 768         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 769         if (ret < 0)
 770                 return 0;
 771         else if (ret > 0) {
 772                 btrfs_release_path(path);
 773                 return 0;
 774         }
 775
 776         ret = -1;
 777
 778         leaf = path->nodes[0];
 779         header = btrfs_item_ptr(leaf, path->slots[0],
 780                                 struct btrfs_free_space_header);
 781         num_entries = btrfs_free_space_entries(leaf, header);
 782         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 783         generation = btrfs_free_space_generation(leaf, header);
 784         btrfs_release_path(path);
 785
 786         if (!BTRFS_I(inode)->generation) {
 787                 btrfs_info(fs_info,
 788                            "the free space cache file (%llu) is invalid, skip it",
 789                            offset);
 790                 return 0;
 791         }
 792
 793         if (BTRFS_I(inode)->generation != generation) {
 794                 btrfs_err(fs_info,
 795                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 796                           BTRFS_I(inode)->generation, generation);
 797                 return 0;
 798         }
 799
 800         if (!num_entries)
 801                 return 0;
 802
 803         ret = io_ctl_init(&io_ctl, inode, 0);
 804         if (ret)
 805                 return ret;
 806
 807         readahead_cache(inode);
 808
 809         ret = io_ctl_prepare_pages(&io_ctl, true);
 810         if (ret)
 811                 goto out;
 812
 813         ret = io_ctl_check_crc(&io_ctl, 0);
 814         if (ret)
 815                 goto free_cache;
 816
 817         ret = io_ctl_check_generation(&io_ctl, generation);
 818         if (ret)
 819                 goto free_cache;
 820
 821         while (num_entries) {
 822                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 823                                       GFP_NOFS);
 824                 if (!e) {
 825                         ret = -ENOMEM;
 826                         goto free_cache;
 827                 }
 828
 829                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 830                 if (ret) {
 831                         kmem_cache_free(btrfs_free_space_cachep, e);
 832                         goto free_cache;
 833                 }
 834
 835                 if (!e->bytes) {
 836                         ret = -1;
 837                         kmem_cache_free(btrfs_free_space_cachep, e);
 838                         goto free_cache;
 839                 }
 840
 841                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 842                         spin_lock(&ctl->tree_lock);
 843                         ret = link_free_space(ctl, e);
 844                         spin_unlock(&ctl->tree_lock);
 845                         if (ret) {
 846                                 btrfs_err(fs_info,
 847                                         "Duplicate entries in free space cache, dumping");
 848                                 kmem_cache_free(btrfs_free_space_cachep, e);
 849                                 goto free_cache;
 850                         }
 851                 } else {
 852                         ASSERT(num_bitmaps);
 853                         num_bitmaps--;
 854                         e->bitmap = kmem_cache_zalloc(
 855                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 856                         if (!e->bitmap) {
 857                                 ret = -ENOMEM;
 858                                 kmem_cache_free(
 859                                         btrfs_free_space_cachep, e);
 860                                 goto free_cache;
 861                         }
 862                         spin_lock(&ctl->tree_lock);
 863                         ret = link_free_space(ctl, e);
 864                         if (ret) {
 865                                 spin_unlock(&ctl->tree_lock);
 866                                 btrfs_err(fs_info,
 867                                         "Duplicate entries in free space cache, dumping");
 868                                 kmem_cache_free(btrfs_free_space_cachep, e);
 869                                 goto free_cache;
 870                         }
 871                         ctl->total_bitmaps++;
 872                         recalculate_thresholds(ctl);
 873                         spin_unlock(&ctl->tree_lock);
 874                         list_add_tail(&e->list, &bitmaps);
 875                 }
 876
 877                 num_entries--;
 878         }
 879
 880         io_ctl_unmap_page(&io_ctl);
 881
 882         /*
 883          * We add the bitmaps at the end of the entries in order that
 884          * the bitmap entries are added to the cache.
 885          */
 886         list_for_each_entry_safe(e, n, &bitmaps, list) {
 887                 list_del_init(&e->list);
 888                 ret = io_ctl_read_bitmap(&io_ctl, e);
 889                 if (ret)
 890                         goto free_cache;
 891         }
 892
 893         io_ctl_drop_pages(&io_ctl);
 894         ret = 1;
 895 out:
 896         io_ctl_free(&io_ctl);
 897         return ret;
 898 free_cache:
 899         io_ctl_drop_pages(&io_ctl);
 900
 901         spin_lock(&ctl->tree_lock);
 902         __btrfs_remove_free_space_cache(ctl);
 903         spin_unlock(&ctl->tree_lock);
 904         goto out;
 905 }
 906
 907 static int copy_free_space_cache(struct btrfs_block_group *block_group,
 908                                  struct btrfs_free_space_ctl *ctl)
 909 {
 910         struct btrfs_free_space *info;
 911         struct rb_node *n;
 912         int ret = 0;
 913
 914         while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
 915                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 916                 if (!info->bitmap) {
 917                         unlink_free_space(ctl, info, true);
 918                         ret = btrfs_add_free_space(block_group, info->offset,
 919                                                    info->bytes);
 920                         kmem_cache_free(btrfs_free_space_cachep, info);
 921                 } else {
 922                         u64 offset = info->offset;
 923                         u64 bytes = ctl->unit;
 924
 925                         while (search_bitmap(ctl, info, &offset, &bytes,
 926                                              false) == 0) {
 927                                 ret = btrfs_add_free_space(block_group, offset,
 928                                                            bytes);
 929                                 if (ret)
 930                                         break;
 931                                 bitmap_clear_bits(ctl, info, offset, bytes, true);
 932                                 offset = info->offset;
 933                                 bytes = ctl->unit;
 934                         }
 935                         free_bitmap(ctl, info);
 936                 }
 937                 cond_resched();
 938         }
 939         return ret;
 940 }
 941
 942 static struct lock_class_key btrfs_free_space_inode_key;
 943
 944 int load_free_space_cache(struct btrfs_block_group *block_group)
 945 {
 946         struct btrfs_fs_info *fs_info = block_group->fs_info;
 947         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 948         struct btrfs_free_space_ctl tmp_ctl = {};
 949         struct inode *inode;
 950         struct btrfs_path *path;
 951         int ret = 0;
 952         bool matched;
 953         u64 used = block_group->used;
 954
 955         /*
 956          * Because we could potentially discard our loaded free space, we want
 957          * to load everything into a temporary structure first, and then if it's
 958          * valid copy it all into the actual free space ctl.
 959          */
 960         btrfs_init_free_space_ctl(block_group, &tmp_ctl);
 961
 962         /*
 963          * If this block group has been marked to be cleared for one reason or
 964          * another then we can't trust the on disk cache, so just return.
 965          */
 966         spin_lock(&block_group->lock);
 967         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 968                 spin_unlock(&block_group->lock);
 969                 return 0;
 970         }
 971         spin_unlock(&block_group->lock);
 972
 973         path = btrfs_alloc_path();
 974         if (!path)
 975                 return 0;
 976         path->search_commit_root = 1;
 977         path->skip_locking = 1;
 978
 979         /*
 980          * We must pass a path with search_commit_root set to btrfs_iget in
 981          * order to avoid a deadlock when allocating extents for the tree root.
 982          *
 983          * When we are COWing an extent buffer from the tree root, when looking
 984          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 985          * block group without its free space cache loaded. When we find one
 986          * we must load its space cache which requires reading its free space
 987          * cache's inode item from the root tree. If this inode item is located
 988          * in the same leaf that we started COWing before, then we end up in
 989          * deadlock on the extent buffer (trying to read lock it when we
 990          * previously write locked it).
 991          *
 992          * It's safe to read the inode item using the commit root because
 993          * block groups, once loaded, stay in memory forever (until they are
 994          * removed) as well as their space caches once loaded. New block groups
 995          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 996          * we will never try to read their inode item while the fs is mounted.
 997          */
 998         inode = lookup_free_space_inode(block_group, path);
 999         if (IS_ERR(inode)) {
1000                 btrfs_free_path(path);
1001                 return 0;
1002         }
1003
1004         /* We may have converted the inode and made the cache invalid. */
1005         spin_lock(&block_group->lock);
1006         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
1007                 spin_unlock(&block_group->lock);
1008                 btrfs_free_path(path);
1009                 goto out;
1010         }
1011         spin_unlock(&block_group->lock);
1012
1013         /*
1014          * Reinitialize the class of struct inode's mapping->invalidate_lock for
1015          * free space inodes to prevent false positives related to locks for normal
1016          * inodes.
1017          */
1018         lockdep_set_class(&(&inode->i_data)->invalidate_lock,
1019                           &btrfs_free_space_inode_key);
1020
1021         ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
1022                                       path, block_group->start);
1023         btrfs_free_path(path);
1024         if (ret <= 0)
1025                 goto out;
1026
1027         matched = (tmp_ctl.free_space == (block_group->length - used -
1028                                           block_group->bytes_super));
1029
1030         if (matched) {
1031                 ret = copy_free_space_cache(block_group, &tmp_ctl);
1032                 /*
1033                  * ret == 1 means we successfully loaded the free space cache,
1034                  * so we need to re-set it here.
1035                  */
1036                 if (ret == 0)
1037                         ret = 1;
1038         } else {
1039                 /*
1040                  * We need to call the _locked variant so we don't try to update
1041                  * the discard counters.
1042                  */
1043                 spin_lock(&tmp_ctl.tree_lock);
1044                 __btrfs_remove_free_space_cache(&tmp_ctl);
1045                 spin_unlock(&tmp_ctl.tree_lock);
1046                 btrfs_warn(fs_info,
1047                            "block group %llu has wrong amount of free space",
1048                            block_group->start);
1049                 ret = -1;
1050         }
1051 out:
1052         if (ret < 0) {
1053                 /* This cache is bogus, make sure it gets cleared */
1054                 spin_lock(&block_group->lock);
1055                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
1056                 spin_unlock(&block_group->lock);
1057                 ret = 0;
1058
1059                 btrfs_warn(fs_info,
1060                            "failed to load free space cache for block group %llu, rebuilding it now",
1061                            block_group->start);
1062         }
1063
1064         spin_lock(&ctl->tree_lock);
1065         btrfs_discard_update_discardable(block_group);
1066         spin_unlock(&ctl->tree_lock);
1067         iput(inode);
1068         return ret;
1069 }
1070
1071 static noinline_for_stack
1072 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1073                               struct btrfs_free_space_ctl *ctl,
1074                               struct btrfs_block_group *block_group,
1075                               int *entries, int *bitmaps,
1076                               struct list_head *bitmap_list)
1077 {
1078         int ret;
1079         struct btrfs_free_cluster *cluster = NULL;
1080         struct btrfs_free_cluster *cluster_locked = NULL;
1081         struct rb_node *node = rb_first(&ctl->free_space_offset);
1082         struct btrfs_trim_range *trim_entry;
1083
1084         /* Get the cluster for this block_group if it exists */
1085         if (block_group && !list_empty(&block_group->cluster_list)) {
1086                 cluster = list_entry(block_group->cluster_list.next,
1087                                      struct btrfs_free_cluster,
1088                                      block_group_list);
1089         }
1090
1091         if (!node && cluster) {
1092                 cluster_locked = cluster;
1093                 spin_lock(&cluster_locked->lock);
1094                 node = rb_first(&cluster->root);
1095                 cluster = NULL;
1096         }
1097
1098         /* Write out the extent entries */
1099         while (node) {
1100                 struct btrfs_free_space *e;
1101
1102                 e = rb_entry(node, struct btrfs_free_space, offset_index);
1103                 *entries += 1;
1104
1105                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1106                                        e->bitmap);
1107                 if (ret)
1108                         goto fail;
1109
1110                 if (e->bitmap) {
1111                         list_add_tail(&e->list, bitmap_list);
1112                         *bitmaps += 1;
1113                 }
1114                 node = rb_next(node);
1115                 if (!node && cluster) {
1116                         node = rb_first(&cluster->root);
1117                         cluster_locked = cluster;
1118                         spin_lock(&cluster_locked->lock);
1119                         cluster = NULL;
1120                 }
1121         }
1122         if (cluster_locked) {
1123                 spin_unlock(&cluster_locked->lock);
1124                 cluster_locked = NULL;
1125         }
1126
1127         /*
1128          * Make sure we don't miss any range that was removed from our rbtree
1129          * because trimming is running. Otherwise after a umount+mount (or crash
1130          * after committing the transaction) we would leak free space and get
1131          * an inconsistent free space cache report from fsck.
1132          */
1133         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1134                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1135                                        trim_entry->bytes, NULL);
1136                 if (ret)
1137                         goto fail;
1138                 *entries += 1;
1139         }
1140
1141         return 0;
1142 fail:
1143         if (cluster_locked)
1144                 spin_unlock(&cluster_locked->lock);
1145         return -ENOSPC;
1146 }
1147
1148 static noinline_for_stack int
1149 update_cache_item(struct btrfs_trans_handle *trans,
1150                   struct btrfs_root *root,
1151                   struct inode *inode,
1152                   struct btrfs_path *path, u64 offset,
1153                   int entries, int bitmaps)
1154 {
1155         struct btrfs_key key;
1156         struct btrfs_free_space_header *header;
1157         struct extent_buffer *leaf;
1158         int ret;
1159
1160         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1161         key.offset = offset;
1162         key.type = 0;
1163
1164         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1165         if (ret < 0) {
1166                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1167                                  EXTENT_DELALLOC, NULL);
1168                 goto fail;
1169         }
1170         leaf = path->nodes[0];
1171         if (ret > 0) {
1172                 struct btrfs_key found_key;
1173                 ASSERT(path->slots[0]);
1174                 path->slots[0]--;
1175                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1176                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1177                     found_key.offset != offset) {
1178                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1179                                          inode->i_size - 1, EXTENT_DELALLOC,
1180                                          NULL);
1181                         btrfs_release_path(path);
1182                         goto fail;
1183                 }
1184         }
1185
1186         BTRFS_I(inode)->generation = trans->transid;
1187         header = btrfs_item_ptr(leaf, path->slots[0],
1188                                 struct btrfs_free_space_header);
1189         btrfs_set_free_space_entries(leaf, header, entries);
1190         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1191         btrfs_set_free_space_generation(leaf, header, trans->transid);
1192         btrfs_mark_buffer_dirty(leaf);
1193         btrfs_release_path(path);
1194
1195         return 0;
1196
1197 fail:
1198         return -1;
1199 }
1200
1201 static noinline_for_stack int write_pinned_extent_entries(
1202                             struct btrfs_trans_handle *trans,
1203                             struct btrfs_block_group *block_group,
1204                             struct btrfs_io_ctl *io_ctl,
1205                             int *entries)
1206 {
1207         u64 start, extent_start, extent_end, len;
1208         struct extent_io_tree *unpin = NULL;
1209         int ret;
1210
1211         if (!block_group)
1212                 return 0;
1213
1214         /*
1215          * We want to add any pinned extents to our free space cache
1216          * so we don't leak the space
1217          *
1218          * We shouldn't have switched the pinned extents yet so this is the
1219          * right one
1220          */
1221         unpin = &trans->transaction->pinned_extents;
1222
1223         start = block_group->start;
1224
1225         while (start < block_group->start + block_group->length) {
1226                 ret = find_first_extent_bit(unpin, start,
1227                                             &extent_start, &extent_end,
1228                                             EXTENT_DIRTY, NULL);
1229                 if (ret)
1230                         return 0;
1231
1232                 /* This pinned extent is out of our range */
1233                 if (extent_start >= block_group->start + block_group->length)
1234                         return 0;
1235
1236                 extent_start = max(extent_start, start);
1237                 extent_end = min(block_group->start + block_group->length,
1238                                  extent_end + 1);
1239                 len = extent_end - extent_start;
1240
1241                 *entries += 1;
1242                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1243                 if (ret)
1244                         return -ENOSPC;
1245
1246                 start = extent_end;
1247         }
1248
1249         return 0;
1250 }
1251
1252 static noinline_for_stack int
1253 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1254 {
1255         struct btrfs_free_space *entry, *next;
1256         int ret;
1257
1258         /* Write out the bitmaps */
1259         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1260                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1261                 if (ret)
1262                         return -ENOSPC;
1263                 list_del_init(&entry->list);
1264         }
1265
1266         return 0;
1267 }
1268
1269 static int flush_dirty_cache(struct inode *inode)
1270 {
1271         int ret;
1272
1273         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1274         if (ret)
1275                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1276                                  EXTENT_DELALLOC, NULL);
1277
1278         return ret;
1279 }
1280
1281 static void noinline_for_stack
1282 cleanup_bitmap_list(struct list_head *bitmap_list)
1283 {
1284         struct btrfs_free_space *entry, *next;
1285
1286         list_for_each_entry_safe(entry, next, bitmap_list, list)
1287                 list_del_init(&entry->list);
1288 }
1289
1290 static void noinline_for_stack
1291 cleanup_write_cache_enospc(struct inode *inode,
1292                            struct btrfs_io_ctl *io_ctl,
1293                            struct extent_state **cached_state)
1294 {
1295         io_ctl_drop_pages(io_ctl);
1296         unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1297                       cached_state);
1298 }
1299
1300 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1301                                  struct btrfs_trans_handle *trans,
1302                                  struct btrfs_block_group *block_group,
1303                                  struct btrfs_io_ctl *io_ctl,
1304                                  struct btrfs_path *path, u64 offset)
1305 {
1306         int ret;
1307         struct inode *inode = io_ctl->inode;
1308
1309         if (!inode)
1310                 return 0;
1311
1312         /* Flush the dirty pages in the cache file. */
1313         ret = flush_dirty_cache(inode);
1314         if (ret)
1315                 goto out;
1316
1317         /* Update the cache item to tell everyone this cache file is valid. */
1318         ret = update_cache_item(trans, root, inode, path, offset,
1319                                 io_ctl->entries, io_ctl->bitmaps);
1320 out:
1321         if (ret) {
1322                 invalidate_inode_pages2(inode->i_mapping);
1323                 BTRFS_I(inode)->generation = 0;
1324                 if (block_group)
1325                         btrfs_debug(root->fs_info,
1326           "failed to write free space cache for block group %llu error %d",
1327                                   block_group->start, ret);
1328         }
1329         btrfs_update_inode(trans, root, BTRFS_I(inode));
1330
1331         if (block_group) {
1332                 /* the dirty list is protected by the dirty_bgs_lock */
1333                 spin_lock(&trans->transaction->dirty_bgs_lock);
1334
1335                 /* the disk_cache_state is protected by the block group lock */
1336                 spin_lock(&block_group->lock);
1337
1338                 /*
1339                  * only mark this as written if we didn't get put back on
1340                  * the dirty list while waiting for IO.   Otherwise our
1341                  * cache state won't be right, and we won't get written again
1342                  */
1343                 if (!ret && list_empty(&block_group->dirty_list))
1344                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1345                 else if (ret)
1346                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1347
1348                 spin_unlock(&block_group->lock);
1349                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1350                 io_ctl->inode = NULL;
1351                 iput(inode);
1352         }
1353
1354         return ret;
1355
1356 }
1357
1358 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1359                         struct btrfs_block_group *block_group,
1360                         struct btrfs_path *path)
1361 {
1362         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1363                                      block_group, &block_group->io_ctl,
1364                                      path, block_group->start);
1365 }
1366
1367 /**
1368  * Write out cached info to an inode
1369  *
1370  * @root:        root the inode belongs to
1371  * @inode:       freespace inode we are writing out
1372  * @ctl:         free space cache we are going to write out
1373  * @block_group: block_group for this cache if it belongs to a block_group
1374  * @io_ctl:      holds context for the io
1375  * @trans:       the trans handle
1376  *
1377  * This function writes out a free space cache struct to disk for quick recovery
1378  * on mount.  This will return 0 if it was successful in writing the cache out,
1379  * or an errno if it was not.
1380  */
1381 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1382                                    struct btrfs_free_space_ctl *ctl,
1383                                    struct btrfs_block_group *block_group,
1384                                    struct btrfs_io_ctl *io_ctl,
1385                                    struct btrfs_trans_handle *trans)
1386 {
1387         struct extent_state *cached_state = NULL;
1388         LIST_HEAD(bitmap_list);
1389         int entries = 0;
1390         int bitmaps = 0;
1391         int ret;
1392         int must_iput = 0;
1393
1394         if (!i_size_read(inode))
1395                 return -EIO;
1396
1397         WARN_ON(io_ctl->pages);
1398         ret = io_ctl_init(io_ctl, inode, 1);
1399         if (ret)
1400                 return ret;
1401
1402         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1403                 down_write(&block_group->data_rwsem);
1404                 spin_lock(&block_group->lock);
1405                 if (block_group->delalloc_bytes) {
1406                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1407                         spin_unlock(&block_group->lock);
1408                         up_write(&block_group->data_rwsem);
1409                         BTRFS_I(inode)->generation = 0;
1410                         ret = 0;
1411                         must_iput = 1;
1412                         goto out;
1413                 }
1414                 spin_unlock(&block_group->lock);
1415         }
1416
1417         /* Lock all pages first so we can lock the extent safely. */
1418         ret = io_ctl_prepare_pages(io_ctl, false);
1419         if (ret)
1420                 goto out_unlock;
1421
1422         lock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1423                     &cached_state);
1424
1425         io_ctl_set_generation(io_ctl, trans->transid);
1426
1427         mutex_lock(&ctl->cache_writeout_mutex);
1428         /* Write out the extent entries in the free space cache */
1429         spin_lock(&ctl->tree_lock);
1430         ret = write_cache_extent_entries(io_ctl, ctl,
1431                                          block_group, &entries, &bitmaps,
1432                                          &bitmap_list);
1433         if (ret)
1434                 goto out_nospc_locked;
1435
1436         /*
1437          * Some spaces that are freed in the current transaction are pinned,
1438          * they will be added into free space cache after the transaction is
1439          * committed, we shouldn't lose them.
1440          *
1441          * If this changes while we are working we'll get added back to
1442          * the dirty list and redo it.  No locking needed
1443          */
1444         ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1445         if (ret)
1446                 goto out_nospc_locked;
1447
1448         /*
1449          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1450          * locked while doing it because a concurrent trim can be manipulating
1451          * or freeing the bitmap.
1452          */
1453         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1454         spin_unlock(&ctl->tree_lock);
1455         mutex_unlock(&ctl->cache_writeout_mutex);
1456         if (ret)
1457                 goto out_nospc;
1458
1459         /* Zero out the rest of the pages just to make sure */
1460         io_ctl_zero_remaining_pages(io_ctl);
1461
1462         /* Everything is written out, now we dirty the pages in the file. */
1463         ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1464                                 io_ctl->num_pages, 0, i_size_read(inode),
1465                                 &cached_state, false);
1466         if (ret)
1467                 goto out_nospc;
1468
1469         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1470                 up_write(&block_group->data_rwsem);
1471         /*
1472          * Release the pages and unlock the extent, we will flush
1473          * them out later
1474          */
1475         io_ctl_drop_pages(io_ctl);
1476         io_ctl_free(io_ctl);
1477
1478         unlock_extent(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1479                       &cached_state);
1480
1481         /*
1482          * at this point the pages are under IO and we're happy,
1483          * The caller is responsible for waiting on them and updating
1484          * the cache and the inode
1485          */
1486         io_ctl->entries = entries;
1487         io_ctl->bitmaps = bitmaps;
1488
1489         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1490         if (ret)
1491                 goto out;
1492
1493         return 0;
1494
1495 out_nospc_locked:
1496         cleanup_bitmap_list(&bitmap_list);
1497         spin_unlock(&ctl->tree_lock);
1498         mutex_unlock(&ctl->cache_writeout_mutex);
1499
1500 out_nospc:
1501         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1502
1503 out_unlock:
1504         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1505                 up_write(&block_group->data_rwsem);
1506
1507 out:
1508         io_ctl->inode = NULL;
1509         io_ctl_free(io_ctl);
1510         if (ret) {
1511                 invalidate_inode_pages2(inode->i_mapping);
1512                 BTRFS_I(inode)->generation = 0;
1513         }
1514         btrfs_update_inode(trans, root, BTRFS_I(inode));
1515         if (must_iput)
1516                 iput(inode);
1517         return ret;
1518 }
1519
1520 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1521                           struct btrfs_block_group *block_group,
1522                           struct btrfs_path *path)
1523 {
1524         struct btrfs_fs_info *fs_info = trans->fs_info;
1525         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1526         struct inode *inode;
1527         int ret = 0;
1528
1529         spin_lock(&block_group->lock);
1530         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1531                 spin_unlock(&block_group->lock);
1532                 return 0;
1533         }
1534         spin_unlock(&block_group->lock);
1535
1536         inode = lookup_free_space_inode(block_group, path);
1537         if (IS_ERR(inode))
1538                 return 0;
1539
1540         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1541                                 block_group, &block_group->io_ctl, trans);
1542         if (ret) {
1543                 btrfs_debug(fs_info,
1544           "failed to write free space cache for block group %llu error %d",
1545                           block_group->start, ret);
1546                 spin_lock(&block_group->lock);
1547                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1548                 spin_unlock(&block_group->lock);
1549
1550                 block_group->io_ctl.inode = NULL;
1551                 iput(inode);
1552         }
1553
1554         /*
1555          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1556          * to wait for IO and put the inode
1557          */
1558
1559         return ret;
1560 }
1561
1562 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1563                                           u64 offset)
1564 {
1565         ASSERT(offset >= bitmap_start);
1566         offset -= bitmap_start;
1567         return (unsigned long)(div_u64(offset, unit));
1568 }
1569
1570 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1571 {
1572         return (unsigned long)(div_u64(bytes, unit));
1573 }
1574
1575 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1576                                    u64 offset)
1577 {
1578         u64 bitmap_start;
1579         u64 bytes_per_bitmap;
1580
1581         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1582         bitmap_start = offset - ctl->start;
1583         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1584         bitmap_start *= bytes_per_bitmap;
1585         bitmap_start += ctl->start;
1586
1587         return bitmap_start;
1588 }
1589
1590 static int tree_insert_offset(struct rb_root *root, u64 offset,
1591                               struct rb_node *node, int bitmap)
1592 {
1593         struct rb_node **p = &root->rb_node;
1594         struct rb_node *parent = NULL;
1595         struct btrfs_free_space *info;
1596
1597         while (*p) {
1598                 parent = *p;
1599                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1600
1601                 if (offset < info->offset) {
1602                         p = &(*p)->rb_left;
1603                 } else if (offset > info->offset) {
1604                         p = &(*p)->rb_right;
1605                 } else {
1606                         /*
1607                          * we could have a bitmap entry and an extent entry
1608                          * share the same offset.  If this is the case, we want
1609                          * the extent entry to always be found first if we do a
1610                          * linear search through the tree, since we want to have
1611                          * the quickest allocation time, and allocating from an
1612                          * extent is faster than allocating from a bitmap.  So
1613                          * if we're inserting a bitmap and we find an entry at
1614                          * this offset, we want to go right, or after this entry
1615                          * logically.  If we are inserting an extent and we've
1616                          * found a bitmap, we want to go left, or before
1617                          * logically.
1618                          */
1619                         if (bitmap) {
1620                                 if (info->bitmap) {
1621                                         WARN_ON_ONCE(1);
1622                                         return -EEXIST;
1623                                 }
1624                                 p = &(*p)->rb_right;
1625                         } else {
1626                                 if (!info->bitmap) {
1627                                         WARN_ON_ONCE(1);
1628                                         return -EEXIST;
1629                                 }
1630                                 p = &(*p)->rb_left;
1631                         }
1632                 }
1633         }
1634
1635         rb_link_node(node, parent, p);
1636         rb_insert_color(node, root);
1637
1638         return 0;
1639 }
1640
1641 /*
1642  * This is a little subtle.  We *only* have ->max_extent_size set if we actually
1643  * searched through the bitmap and figured out the largest ->max_extent_size,
1644  * otherwise it's 0.  In the case that it's 0 we don't want to tell the
1645  * allocator the wrong thing, we want to use the actual real max_extent_size
1646  * we've found already if it's larger, or we want to use ->bytes.
1647  *
1648  * This matters because find_free_space() will skip entries who's ->bytes is
1649  * less than the required bytes.  So if we didn't search down this bitmap, we
1650  * may pick some previous entry that has a smaller ->max_extent_size than we
1651  * have.  For example, assume we have two entries, one that has
1652  * ->max_extent_size set to 4K and ->bytes set to 1M.  A second entry hasn't set
1653  * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous.  We will
1654  *  call into find_free_space(), and return with max_extent_size == 4K, because
1655  *  that first bitmap entry had ->max_extent_size set, but the second one did
1656  *  not.  If instead we returned 8K we'd come in searching for 8K, and find the
1657  *  8K contiguous range.
1658  *
1659  *  Consider the other case, we have 2 8K chunks in that second entry and still
1660  *  don't have ->max_extent_size set.  We'll return 16K, and the next time the
1661  *  allocator comes in it'll fully search our second bitmap, and this time it'll
1662  *  get an uptodate value of 8K as the maximum chunk size.  Then we'll get the
1663  *  right allocation the next loop through.
1664  */
1665 static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1666 {
1667         if (entry->bitmap && entry->max_extent_size)
1668                 return entry->max_extent_size;
1669         return entry->bytes;
1670 }
1671
1672 /*
1673  * We want the largest entry to be leftmost, so this is inverted from what you'd
1674  * normally expect.
1675  */
1676 static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1677 {
1678         const struct btrfs_free_space *entry, *exist;
1679
1680         entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1681         exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1682         return get_max_extent_size(exist) < get_max_extent_size(entry);
1683 }
1684
1685 /*
1686  * searches the tree for the given offset.
1687  *
1688  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1689  * want a section that has at least bytes size and comes at or after the given
1690  * offset.
1691  */
1692 static struct btrfs_free_space *
1693 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1694                    u64 offset, int bitmap_only, int fuzzy)
1695 {
1696         struct rb_node *n = ctl->free_space_offset.rb_node;
1697         struct btrfs_free_space *entry = NULL, *prev = NULL;
1698
1699         /* find entry that is closest to the 'offset' */
1700         while (n) {
1701                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1702                 prev = entry;
1703
1704                 if (offset < entry->offset)
1705                         n = n->rb_left;
1706                 else if (offset > entry->offset)
1707                         n = n->rb_right;
1708                 else
1709                         break;
1710
1711                 entry = NULL;
1712         }
1713
1714         if (bitmap_only) {
1715                 if (!entry)
1716                         return NULL;
1717                 if (entry->bitmap)
1718                         return entry;
1719
1720                 /*
1721                  * bitmap entry and extent entry may share same offset,
1722                  * in that case, bitmap entry comes after extent entry.
1723                  */
1724                 n = rb_next(n);
1725                 if (!n)
1726                         return NULL;
1727                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1728                 if (entry->offset != offset)
1729                         return NULL;
1730
1731                 WARN_ON(!entry->bitmap);
1732                 return entry;
1733         } else if (entry) {
1734                 if (entry->bitmap) {
1735                         /*
1736                          * if previous extent entry covers the offset,
1737                          * we should return it instead of the bitmap entry
1738                          */
1739                         n = rb_prev(&entry->offset_index);
1740                         if (n) {
1741                                 prev = rb_entry(n, struct btrfs_free_space,
1742                                                 offset_index);
1743                                 if (!prev->bitmap &&
1744                                     prev->offset + prev->bytes > offset)
1745                                         entry = prev;
1746                         }
1747                 }
1748                 return entry;
1749         }
1750
1751         if (!prev)
1752                 return NULL;
1753
1754         /* find last entry before the 'offset' */
1755         entry = prev;
1756         if (entry->offset > offset) {
1757                 n = rb_prev(&entry->offset_index);
1758                 if (n) {
1759                         entry = rb_entry(n, struct btrfs_free_space,
1760                                         offset_index);
1761                         ASSERT(entry->offset <= offset);
1762                 } else {
1763                         if (fuzzy)
1764                                 return entry;
1765                         else
1766                                 return NULL;
1767                 }
1768         }
1769
1770         if (entry->bitmap) {
1771                 n = rb_prev(&entry->offset_index);
1772                 if (n) {
1773                         prev = rb_entry(n, struct btrfs_free_space,
1774                                         offset_index);
1775                         if (!prev->bitmap &&
1776                             prev->offset + prev->bytes > offset)
1777                                 return prev;
1778                 }
1779                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1780                         return entry;
1781         } else if (entry->offset + entry->bytes > offset)
1782                 return entry;
1783
1784         if (!fuzzy)
1785                 return NULL;
1786
1787         while (1) {
1788                 n = rb_next(&entry->offset_index);
1789                 if (!n)
1790                         return NULL;
1791                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1792                 if (entry->bitmap) {
1793                         if (entry->offset + BITS_PER_BITMAP *
1794                             ctl->unit > offset)
1795                                 break;
1796                 } else {
1797                         if (entry->offset + entry->bytes > offset)
1798                                 break;
1799                 }
1800         }
1801         return entry;
1802 }
1803
1804 static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1805                                      struct btrfs_free_space *info,
1806                                      bool update_stat)
1807 {
1808         rb_erase(&info->offset_index, &ctl->free_space_offset);
1809         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1810         ctl->free_extents--;
1811
1812         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1813                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1814                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1815         }
1816
1817         if (update_stat)
1818                 ctl->free_space -= info->bytes;
1819 }
1820
1821 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1822                            struct btrfs_free_space *info)
1823 {
1824         int ret = 0;
1825
1826         ASSERT(info->bytes || info->bitmap);
1827         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1828                                  &info->offset_index, (info->bitmap != NULL));
1829         if (ret)
1830                 return ret;
1831
1832         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1833
1834         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1835                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1836                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1837         }
1838
1839         ctl->free_space += info->bytes;
1840         ctl->free_extents++;
1841         return ret;
1842 }
1843
1844 static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1845                                 struct btrfs_free_space *info)
1846 {
1847         ASSERT(info->bitmap);
1848
1849         /*
1850          * If our entry is empty it's because we're on a cluster and we don't
1851          * want to re-link it into our ctl bytes index.
1852          */
1853         if (RB_EMPTY_NODE(&info->bytes_index))
1854                 return;
1855
1856         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1857         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1858 }
1859
1860 static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1861                                      struct btrfs_free_space *info,
1862                                      u64 offset, u64 bytes, bool update_stat)
1863 {
1864         unsigned long start, count, end;
1865         int extent_delta = -1;
1866
1867         start = offset_to_bit(info->offset, ctl->unit, offset);
1868         count = bytes_to_bits(bytes, ctl->unit);
1869         end = start + count;
1870         ASSERT(end <= BITS_PER_BITMAP);
1871
1872         bitmap_clear(info->bitmap, start, count);
1873
1874         info->bytes -= bytes;
1875         if (info->max_extent_size > ctl->unit)
1876                 info->max_extent_size = 0;
1877
1878         relink_bitmap_entry(ctl, info);
1879
1880         if (start && test_bit(start - 1, info->bitmap))
1881                 extent_delta++;
1882
1883         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1884                 extent_delta++;
1885
1886         info->bitmap_extents += extent_delta;
1887         if (!btrfs_free_space_trimmed(info)) {
1888                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1889                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1890         }
1891
1892         if (update_stat)
1893                 ctl->free_space -= bytes;
1894 }
1895
1896 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1897                             struct btrfs_free_space *info, u64 offset,
1898                             u64 bytes)
1899 {
1900         unsigned long start, count, end;
1901         int extent_delta = 1;
1902
1903         start = offset_to_bit(info->offset, ctl->unit, offset);
1904         count = bytes_to_bits(bytes, ctl->unit);
1905         end = start + count;
1906         ASSERT(end <= BITS_PER_BITMAP);
1907
1908         bitmap_set(info->bitmap, start, count);
1909
1910         /*
1911          * We set some bytes, we have no idea what the max extent size is
1912          * anymore.
1913          */
1914         info->max_extent_size = 0;
1915         info->bytes += bytes;
1916         ctl->free_space += bytes;
1917
1918         relink_bitmap_entry(ctl, info);
1919
1920         if (start && test_bit(start - 1, info->bitmap))
1921                 extent_delta--;
1922
1923         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1924                 extent_delta--;
1925
1926         info->bitmap_extents += extent_delta;
1927         if (!btrfs_free_space_trimmed(info)) {
1928                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1929                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1930         }
1931 }
1932
1933 /*
1934  * If we can not find suitable extent, we will use bytes to record
1935  * the size of the max extent.
1936  */
1937 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1938                          struct btrfs_free_space *bitmap_info, u64 *offset,
1939                          u64 *bytes, bool for_alloc)
1940 {
1941         unsigned long found_bits = 0;
1942         unsigned long max_bits = 0;
1943         unsigned long bits, i;
1944         unsigned long next_zero;
1945         unsigned long extent_bits;
1946
1947         /*
1948          * Skip searching the bitmap if we don't have a contiguous section that
1949          * is large enough for this allocation.
1950          */
1951         if (for_alloc &&
1952             bitmap_info->max_extent_size &&
1953             bitmap_info->max_extent_size < *bytes) {
1954                 *bytes = bitmap_info->max_extent_size;
1955                 return -1;
1956         }
1957
1958         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1959                           max_t(u64, *offset, bitmap_info->offset));
1960         bits = bytes_to_bits(*bytes, ctl->unit);
1961
1962         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1963                 if (for_alloc && bits == 1) {
1964                         found_bits = 1;
1965                         break;
1966                 }
1967                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1968                                                BITS_PER_BITMAP, i);
1969                 extent_bits = next_zero - i;
1970                 if (extent_bits >= bits) {
1971                         found_bits = extent_bits;
1972                         break;
1973                 } else if (extent_bits > max_bits) {
1974                         max_bits = extent_bits;
1975                 }
1976                 i = next_zero;
1977         }
1978
1979         if (found_bits) {
1980                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1981                 *bytes = (u64)(found_bits) * ctl->unit;
1982                 return 0;
1983         }
1984
1985         *bytes = (u64)(max_bits) * ctl->unit;
1986         bitmap_info->max_extent_size = *bytes;
1987         relink_bitmap_entry(ctl, bitmap_info);
1988         return -1;
1989 }
1990
1991 /* Cache the size of the max extent in bytes */
1992 static struct btrfs_free_space *
1993 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1994                 unsigned long align, u64 *max_extent_size, bool use_bytes_index)
1995 {
1996         struct btrfs_free_space *entry;
1997         struct rb_node *node;
1998         u64 tmp;
1999         u64 align_off;
2000         int ret;
2001
2002         if (!ctl->free_space_offset.rb_node)
2003                 goto out;
2004 again:
2005         if (use_bytes_index) {
2006                 node = rb_first_cached(&ctl->free_space_bytes);
2007         } else {
2008                 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
2009                                            0, 1);
2010                 if (!entry)
2011                         goto out;
2012                 node = &entry->offset_index;
2013         }
2014
2015         for (; node; node = rb_next(node)) {
2016                 if (use_bytes_index)
2017                         entry = rb_entry(node, struct btrfs_free_space,
2018                                          bytes_index);
2019                 else
2020                         entry = rb_entry(node, struct btrfs_free_space,
2021                                          offset_index);
2022
2023                 /*
2024                  * If we are using the bytes index then all subsequent entries
2025                  * in this tree are going to be < bytes, so simply set the max
2026                  * extent size and exit the loop.
2027                  *
2028                  * If we're using the offset index then we need to keep going
2029                  * through the rest of the tree.
2030                  */
2031                 if (entry->bytes < *bytes) {
2032                         *max_extent_size = max(get_max_extent_size(entry),
2033                                                *max_extent_size);
2034                         if (use_bytes_index)
2035                                 break;
2036                         continue;
2037                 }
2038
2039                 /* make sure the space returned is big enough
2040                  * to match our requested alignment
2041                  */
2042                 if (*bytes >= align) {
2043                         tmp = entry->offset - ctl->start + align - 1;
2044                         tmp = div64_u64(tmp, align);
2045                         tmp = tmp * align + ctl->start;
2046                         align_off = tmp - entry->offset;
2047                 } else {
2048                         align_off = 0;
2049                         tmp = entry->offset;
2050                 }
2051
2052                 /*
2053                  * We don't break here if we're using the bytes index because we
2054                  * may have another entry that has the correct alignment that is
2055                  * the right size, so we don't want to miss that possibility.
2056                  * At worst this adds another loop through the logic, but if we
2057                  * broke here we could prematurely ENOSPC.
2058                  */
2059                 if (entry->bytes < *bytes + align_off) {
2060                         *max_extent_size = max(get_max_extent_size(entry),
2061                                                *max_extent_size);
2062                         continue;
2063                 }
2064
2065                 if (entry->bitmap) {
2066                         struct rb_node *old_next = rb_next(node);
2067                         u64 size = *bytes;
2068
2069                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
2070                         if (!ret) {
2071                                 *offset = tmp;
2072                                 *bytes = size;
2073                                 return entry;
2074                         } else {
2075                                 *max_extent_size =
2076                                         max(get_max_extent_size(entry),
2077                                             *max_extent_size);
2078                         }
2079
2080                         /*
2081                          * The bitmap may have gotten re-arranged in the space
2082                          * index here because the max_extent_size may have been
2083                          * updated.  Start from the beginning again if this
2084                          * happened.
2085                          */
2086                         if (use_bytes_index && old_next != rb_next(node))
2087                                 goto again;
2088                         continue;
2089                 }
2090
2091                 *offset = tmp;
2092                 *bytes = entry->bytes - align_off;
2093                 return entry;
2094         }
2095 out:
2096         return NULL;
2097 }
2098
2099 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2100                            struct btrfs_free_space *info, u64 offset)
2101 {
2102         info->offset = offset_to_bitmap(ctl, offset);
2103         info->bytes = 0;
2104         info->bitmap_extents = 0;
2105         INIT_LIST_HEAD(&info->list);
2106         link_free_space(ctl, info);
2107         ctl->total_bitmaps++;
2108         recalculate_thresholds(ctl);
2109 }
2110
2111 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2112                         struct btrfs_free_space *bitmap_info)
2113 {
2114         /*
2115          * Normally when this is called, the bitmap is completely empty. However,
2116          * if we are blowing up the free space cache for one reason or another
2117          * via __btrfs_remove_free_space_cache(), then it may not be freed and
2118          * we may leave stats on the table.
2119          */
2120         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2121                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
2122                         bitmap_info->bitmap_extents;
2123                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2124
2125         }
2126         unlink_free_space(ctl, bitmap_info, true);
2127         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2128         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2129         ctl->total_bitmaps--;
2130         recalculate_thresholds(ctl);
2131 }
2132
2133 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2134                               struct btrfs_free_space *bitmap_info,
2135                               u64 *offset, u64 *bytes)
2136 {
2137         u64 end;
2138         u64 search_start, search_bytes;
2139         int ret;
2140
2141 again:
2142         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2143
2144         /*
2145          * We need to search for bits in this bitmap.  We could only cover some
2146          * of the extent in this bitmap thanks to how we add space, so we need
2147          * to search for as much as it as we can and clear that amount, and then
2148          * go searching for the next bit.
2149          */
2150         search_start = *offset;
2151         search_bytes = ctl->unit;
2152         search_bytes = min(search_bytes, end - search_start + 1);
2153         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2154                             false);
2155         if (ret < 0 || search_start != *offset)
2156                 return -EINVAL;
2157
2158         /* We may have found more bits than what we need */
2159         search_bytes = min(search_bytes, *bytes);
2160
2161         /* Cannot clear past the end of the bitmap */
2162         search_bytes = min(search_bytes, end - search_start + 1);
2163
2164         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2165         *offset += search_bytes;
2166         *bytes -= search_bytes;
2167
2168         if (*bytes) {
2169                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2170                 if (!bitmap_info->bytes)
2171                         free_bitmap(ctl, bitmap_info);
2172
2173                 /*
2174                  * no entry after this bitmap, but we still have bytes to
2175                  * remove, so something has gone wrong.
2176                  */
2177                 if (!next)
2178                         return -EINVAL;
2179
2180                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2181                                        offset_index);
2182
2183                 /*
2184                  * if the next entry isn't a bitmap we need to return to let the
2185                  * extent stuff do its work.
2186                  */
2187                 if (!bitmap_info->bitmap)
2188                         return -EAGAIN;
2189
2190                 /*
2191                  * Ok the next item is a bitmap, but it may not actually hold
2192                  * the information for the rest of this free space stuff, so
2193                  * look for it, and if we don't find it return so we can try
2194                  * everything over again.
2195                  */
2196                 search_start = *offset;
2197                 search_bytes = ctl->unit;
2198                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2199                                     &search_bytes, false);
2200                 if (ret < 0 || search_start != *offset)
2201                         return -EAGAIN;
2202
2203                 goto again;
2204         } else if (!bitmap_info->bytes)
2205                 free_bitmap(ctl, bitmap_info);
2206
2207         return 0;
2208 }
2209
2210 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2211                                struct btrfs_free_space *info, u64 offset,
2212                                u64 bytes, enum btrfs_trim_state trim_state)
2213 {
2214         u64 bytes_to_set = 0;
2215         u64 end;
2216
2217         /*
2218          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2219          * whole bitmap untrimmed if at any point we add untrimmed regions.
2220          */
2221         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2222                 if (btrfs_free_space_trimmed(info)) {
2223                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2224                                 info->bitmap_extents;
2225                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2226                 }
2227                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2228         }
2229
2230         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2231
2232         bytes_to_set = min(end - offset, bytes);
2233
2234         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2235
2236         return bytes_to_set;
2237
2238 }
2239
2240 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2241                       struct btrfs_free_space *info)
2242 {
2243         struct btrfs_block_group *block_group = ctl->block_group;
2244         struct btrfs_fs_info *fs_info = block_group->fs_info;
2245         bool forced = false;
2246
2247 #ifdef CONFIG_BTRFS_DEBUG
2248         if (btrfs_should_fragment_free_space(block_group))
2249                 forced = true;
2250 #endif
2251
2252         /* This is a way to reclaim large regions from the bitmaps. */
2253         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2254                 return false;
2255
2256         /*
2257          * If we are below the extents threshold then we can add this as an
2258          * extent, and don't have to deal with the bitmap
2259          */
2260         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2261                 /*
2262                  * If this block group has some small extents we don't want to
2263                  * use up all of our free slots in the cache with them, we want
2264                  * to reserve them to larger extents, however if we have plenty
2265                  * of cache left then go ahead an dadd them, no sense in adding
2266                  * the overhead of a bitmap if we don't have to.
2267                  */
2268                 if (info->bytes <= fs_info->sectorsize * 8) {
2269                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2270                                 return false;
2271                 } else {
2272                         return false;
2273                 }
2274         }
2275
2276         /*
2277          * The original block groups from mkfs can be really small, like 8
2278          * megabytes, so don't bother with a bitmap for those entries.  However
2279          * some block groups can be smaller than what a bitmap would cover but
2280          * are still large enough that they could overflow the 32k memory limit,
2281          * so allow those block groups to still be allowed to have a bitmap
2282          * entry.
2283          */
2284         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2285                 return false;
2286
2287         return true;
2288 }
2289
2290 static const struct btrfs_free_space_op free_space_op = {
2291         .use_bitmap             = use_bitmap,
2292 };
2293
2294 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2295                               struct btrfs_free_space *info)
2296 {
2297         struct btrfs_free_space *bitmap_info;
2298         struct btrfs_block_group *block_group = NULL;
2299         int added = 0;
2300         u64 bytes, offset, bytes_added;
2301         enum btrfs_trim_state trim_state;
2302         int ret;
2303
2304         bytes = info->bytes;
2305         offset = info->offset;
2306         trim_state = info->trim_state;
2307
2308         if (!ctl->op->use_bitmap(ctl, info))
2309                 return 0;
2310
2311         if (ctl->op == &free_space_op)
2312                 block_group = ctl->block_group;
2313 again:
2314         /*
2315          * Since we link bitmaps right into the cluster we need to see if we
2316          * have a cluster here, and if so and it has our bitmap we need to add
2317          * the free space to that bitmap.
2318          */
2319         if (block_group && !list_empty(&block_group->cluster_list)) {
2320                 struct btrfs_free_cluster *cluster;
2321                 struct rb_node *node;
2322                 struct btrfs_free_space *entry;
2323
2324                 cluster = list_entry(block_group->cluster_list.next,
2325                                      struct btrfs_free_cluster,
2326                                      block_group_list);
2327                 spin_lock(&cluster->lock);
2328                 node = rb_first(&cluster->root);
2329                 if (!node) {
2330                         spin_unlock(&cluster->lock);
2331                         goto no_cluster_bitmap;
2332                 }
2333
2334                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2335                 if (!entry->bitmap) {
2336                         spin_unlock(&cluster->lock);
2337                         goto no_cluster_bitmap;
2338                 }
2339
2340                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2341                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2342                                                           bytes, trim_state);
2343                         bytes -= bytes_added;
2344                         offset += bytes_added;
2345                 }
2346                 spin_unlock(&cluster->lock);
2347                 if (!bytes) {
2348                         ret = 1;
2349                         goto out;
2350                 }
2351         }
2352
2353 no_cluster_bitmap:
2354         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2355                                          1, 0);
2356         if (!bitmap_info) {
2357                 ASSERT(added == 0);
2358                 goto new_bitmap;
2359         }
2360
2361         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2362                                           trim_state);
2363         bytes -= bytes_added;
2364         offset += bytes_added;
2365         added = 0;
2366
2367         if (!bytes) {
2368                 ret = 1;
2369                 goto out;
2370         } else
2371                 goto again;
2372
2373 new_bitmap:
2374         if (info && info->bitmap) {
2375                 add_new_bitmap(ctl, info, offset);
2376                 added = 1;
2377                 info = NULL;
2378                 goto again;
2379         } else {
2380                 spin_unlock(&ctl->tree_lock);
2381
2382                 /* no pre-allocated info, allocate a new one */
2383                 if (!info) {
2384                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2385                                                  GFP_NOFS);
2386                         if (!info) {
2387                                 spin_lock(&ctl->tree_lock);
2388                                 ret = -ENOMEM;
2389                                 goto out;
2390                         }
2391                 }
2392
2393                 /* allocate the bitmap */
2394                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2395                                                  GFP_NOFS);
2396                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2397                 spin_lock(&ctl->tree_lock);
2398                 if (!info->bitmap) {
2399                         ret = -ENOMEM;
2400                         goto out;
2401                 }
2402                 goto again;
2403         }
2404
2405 out:
2406         if (info) {
2407                 if (info->bitmap)
2408                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2409                                         info->bitmap);
2410                 kmem_cache_free(btrfs_free_space_cachep, info);
2411         }
2412
2413         return ret;
2414 }
2415
2416 /*
2417  * Free space merging rules:
2418  *  1) Merge trimmed areas together
2419  *  2) Let untrimmed areas coalesce with trimmed areas
2420  *  3) Always pull neighboring regions from bitmaps
2421  *
2422  * The above rules are for when we merge free space based on btrfs_trim_state.
2423  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2424  * same reason: to promote larger extent regions which makes life easier for
2425  * find_free_extent().  Rule 2 enables coalescing based on the common path
2426  * being returning free space from btrfs_finish_extent_commit().  So when free
2427  * space is trimmed, it will prevent aggregating trimmed new region and
2428  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2429  * and provide find_free_extent() with the largest extents possible hoping for
2430  * the reuse path.
2431  */
2432 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2433                           struct btrfs_free_space *info, bool update_stat)
2434 {
2435         struct btrfs_free_space *left_info = NULL;
2436         struct btrfs_free_space *right_info;
2437         bool merged = false;
2438         u64 offset = info->offset;
2439         u64 bytes = info->bytes;
2440         const bool is_trimmed = btrfs_free_space_trimmed(info);
2441
2442         /*
2443          * first we want to see if there is free space adjacent to the range we
2444          * are adding, if there is remove that struct and add a new one to
2445          * cover the entire range
2446          */
2447         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2448         if (right_info && rb_prev(&right_info->offset_index))
2449                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2450                                      struct btrfs_free_space, offset_index);
2451         else if (!right_info)
2452                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2453
2454         /* See try_merge_free_space() comment. */
2455         if (right_info && !right_info->bitmap &&
2456             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2457                 unlink_free_space(ctl, right_info, update_stat);
2458                 info->bytes += right_info->bytes;
2459                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2460                 merged = true;
2461         }
2462
2463         /* See try_merge_free_space() comment. */
2464         if (left_info && !left_info->bitmap &&
2465             left_info->offset + left_info->bytes == offset &&
2466             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2467                 unlink_free_space(ctl, left_info, update_stat);
2468                 info->offset = left_info->offset;
2469                 info->bytes += left_info->bytes;
2470                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2471                 merged = true;
2472         }
2473
2474         return merged;
2475 }
2476
2477 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2478                                      struct btrfs_free_space *info,
2479                                      bool update_stat)
2480 {
2481         struct btrfs_free_space *bitmap;
2482         unsigned long i;
2483         unsigned long j;
2484         const u64 end = info->offset + info->bytes;
2485         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2486         u64 bytes;
2487
2488         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2489         if (!bitmap)
2490                 return false;
2491
2492         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2493         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2494         if (j == i)
2495                 return false;
2496         bytes = (j - i) * ctl->unit;
2497         info->bytes += bytes;
2498
2499         /* See try_merge_free_space() comment. */
2500         if (!btrfs_free_space_trimmed(bitmap))
2501                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2502
2503         bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2504
2505         if (!bitmap->bytes)
2506                 free_bitmap(ctl, bitmap);
2507
2508         return true;
2509 }
2510
2511 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2512                                        struct btrfs_free_space *info,
2513                                        bool update_stat)
2514 {
2515         struct btrfs_free_space *bitmap;
2516         u64 bitmap_offset;
2517         unsigned long i;
2518         unsigned long j;
2519         unsigned long prev_j;
2520         u64 bytes;
2521
2522         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2523         /* If we're on a boundary, try the previous logical bitmap. */
2524         if (bitmap_offset == info->offset) {
2525                 if (info->offset == 0)
2526                         return false;
2527                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2528         }
2529
2530         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2531         if (!bitmap)
2532                 return false;
2533
2534         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2535         j = 0;
2536         prev_j = (unsigned long)-1;
2537         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2538                 if (j > i)
2539                         break;
2540                 prev_j = j;
2541         }
2542         if (prev_j == i)
2543                 return false;
2544
2545         if (prev_j == (unsigned long)-1)
2546                 bytes = (i + 1) * ctl->unit;
2547         else
2548                 bytes = (i - prev_j) * ctl->unit;
2549
2550         info->offset -= bytes;
2551         info->bytes += bytes;
2552
2553         /* See try_merge_free_space() comment. */
2554         if (!btrfs_free_space_trimmed(bitmap))
2555                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2556
2557         bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2558
2559         if (!bitmap->bytes)
2560                 free_bitmap(ctl, bitmap);
2561
2562         return true;
2563 }
2564
2565 /*
2566  * We prefer always to allocate from extent entries, both for clustered and
2567  * non-clustered allocation requests. So when attempting to add a new extent
2568  * entry, try to see if there's adjacent free space in bitmap entries, and if
2569  * there is, migrate that space from the bitmaps to the extent.
2570  * Like this we get better chances of satisfying space allocation requests
2571  * because we attempt to satisfy them based on a single cache entry, and never
2572  * on 2 or more entries - even if the entries represent a contiguous free space
2573  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2574  * ends).
2575  */
2576 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2577                               struct btrfs_free_space *info,
2578                               bool update_stat)
2579 {
2580         /*
2581          * Only work with disconnected entries, as we can change their offset,
2582          * and must be extent entries.
2583          */
2584         ASSERT(!info->bitmap);
2585         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2586
2587         if (ctl->total_bitmaps > 0) {
2588                 bool stole_end;
2589                 bool stole_front = false;
2590
2591                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2592                 if (ctl->total_bitmaps > 0)
2593                         stole_front = steal_from_bitmap_to_front(ctl, info,
2594                                                                  update_stat);
2595
2596                 if (stole_end || stole_front)
2597                         try_merge_free_space(ctl, info, update_stat);
2598         }
2599 }
2600
2601 int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2602                            u64 offset, u64 bytes,
2603                            enum btrfs_trim_state trim_state)
2604 {
2605         struct btrfs_fs_info *fs_info = block_group->fs_info;
2606         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2607         struct btrfs_free_space *info;
2608         int ret = 0;
2609         u64 filter_bytes = bytes;
2610
2611         ASSERT(!btrfs_is_zoned(fs_info));
2612
2613         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2614         if (!info)
2615                 return -ENOMEM;
2616
2617         info->offset = offset;
2618         info->bytes = bytes;
2619         info->trim_state = trim_state;
2620         RB_CLEAR_NODE(&info->offset_index);
2621         RB_CLEAR_NODE(&info->bytes_index);
2622
2623         spin_lock(&ctl->tree_lock);
2624
2625         if (try_merge_free_space(ctl, info, true))
2626                 goto link;
2627
2628         /*
2629          * There was no extent directly to the left or right of this new
2630          * extent then we know we're going to have to allocate a new extent, so
2631          * before we do that see if we need to drop this into a bitmap
2632          */
2633         ret = insert_into_bitmap(ctl, info);
2634         if (ret < 0) {
2635                 goto out;
2636         } else if (ret) {
2637                 ret = 0;
2638                 goto out;
2639         }
2640 link:
2641         /*
2642          * Only steal free space from adjacent bitmaps if we're sure we're not
2643          * going to add the new free space to existing bitmap entries - because
2644          * that would mean unnecessary work that would be reverted. Therefore
2645          * attempt to steal space from bitmaps if we're adding an extent entry.
2646          */
2647         steal_from_bitmap(ctl, info, true);
2648
2649         filter_bytes = max(filter_bytes, info->bytes);
2650
2651         ret = link_free_space(ctl, info);
2652         if (ret)
2653                 kmem_cache_free(btrfs_free_space_cachep, info);
2654 out:
2655         btrfs_discard_update_discardable(block_group);
2656         spin_unlock(&ctl->tree_lock);
2657
2658         if (ret) {
2659                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2660                 ASSERT(ret != -EEXIST);
2661         }
2662
2663         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2664                 btrfs_discard_check_filter(block_group, filter_bytes);
2665                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2666         }
2667
2668         return ret;
2669 }
2670
2671 static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2672                                         u64 bytenr, u64 size, bool used)
2673 {
2674         struct btrfs_space_info *sinfo = block_group->space_info;
2675         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2676         u64 offset = bytenr - block_group->start;
2677         u64 to_free, to_unusable;
2678         int bg_reclaim_threshold = 0;
2679         bool initial = (size == block_group->length);
2680         u64 reclaimable_unusable;
2681
2682         WARN_ON(!initial && offset + size > block_group->zone_capacity);
2683
2684         if (!initial)
2685                 bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2686
2687         spin_lock(&ctl->tree_lock);
2688         if (!used)
2689                 to_free = size;
2690         else if (initial)
2691                 to_free = block_group->zone_capacity;
2692         else if (offset >= block_group->alloc_offset)
2693                 to_free = size;
2694         else if (offset + size <= block_group->alloc_offset)
2695                 to_free = 0;
2696         else
2697                 to_free = offset + size - block_group->alloc_offset;
2698         to_unusable = size - to_free;
2699
2700         ctl->free_space += to_free;
2701         /*
2702          * If the block group is read-only, we should account freed space into
2703          * bytes_readonly.
2704          */
2705         if (!block_group->ro)
2706                 block_group->zone_unusable += to_unusable;
2707         spin_unlock(&ctl->tree_lock);
2708         if (!used) {
2709                 spin_lock(&block_group->lock);
2710                 block_group->alloc_offset -= size;
2711                 spin_unlock(&block_group->lock);
2712         }
2713
2714         reclaimable_unusable = block_group->zone_unusable -
2715                                (block_group->length - block_group->zone_capacity);
2716         /* All the region is now unusable. Mark it as unused and reclaim */
2717         if (block_group->zone_unusable == block_group->length) {
2718                 btrfs_mark_bg_unused(block_group);
2719         } else if (bg_reclaim_threshold &&
2720                    reclaimable_unusable >=
2721                    div_factor_fine(block_group->zone_capacity,
2722                                    bg_reclaim_threshold)) {
2723                 btrfs_mark_bg_to_reclaim(block_group);
2724         }
2725
2726         return 0;
2727 }
2728
2729 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2730                          u64 bytenr, u64 size)
2731 {
2732         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2733
2734         if (btrfs_is_zoned(block_group->fs_info))
2735                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2736                                                     true);
2737
2738         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2739                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2740
2741         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2742 }
2743
2744 int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2745                                 u64 bytenr, u64 size)
2746 {
2747         if (btrfs_is_zoned(block_group->fs_info))
2748                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2749                                                     false);
2750
2751         return btrfs_add_free_space(block_group, bytenr, size);
2752 }
2753
2754 /*
2755  * This is a subtle distinction because when adding free space back in general,
2756  * we want it to be added as untrimmed for async. But in the case where we add
2757  * it on loading of a block group, we want to consider it trimmed.
2758  */
2759 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2760                                        u64 bytenr, u64 size)
2761 {
2762         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2763
2764         if (btrfs_is_zoned(block_group->fs_info))
2765                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2766                                                     true);
2767
2768         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2769             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2770                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2771
2772         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2773 }
2774
2775 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2776                             u64 offset, u64 bytes)
2777 {
2778         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2779         struct btrfs_free_space *info;
2780         int ret;
2781         bool re_search = false;
2782
2783         if (btrfs_is_zoned(block_group->fs_info)) {
2784                 /*
2785                  * This can happen with conventional zones when replaying log.
2786                  * Since the allocation info of tree-log nodes are not recorded
2787                  * to the extent-tree, calculate_alloc_pointer() failed to
2788                  * advance the allocation pointer after last allocated tree log
2789                  * node blocks.
2790                  *
2791                  * This function is called from
2792                  * btrfs_pin_extent_for_log_replay() when replaying the log.
2793                  * Advance the pointer not to overwrite the tree-log nodes.
2794                  */
2795                 if (block_group->start + block_group->alloc_offset <
2796                     offset + bytes) {
2797                         block_group->alloc_offset =
2798                                 offset + bytes - block_group->start;
2799                 }
2800                 return 0;
2801         }
2802
2803         spin_lock(&ctl->tree_lock);
2804
2805 again:
2806         ret = 0;
2807         if (!bytes)
2808                 goto out_lock;
2809
2810         info = tree_search_offset(ctl, offset, 0, 0);
2811         if (!info) {
2812                 /*
2813                  * oops didn't find an extent that matched the space we wanted
2814                  * to remove, look for a bitmap instead
2815                  */
2816                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2817                                           1, 0);
2818                 if (!info) {
2819                         /*
2820                          * If we found a partial bit of our free space in a
2821                          * bitmap but then couldn't find the other part this may
2822                          * be a problem, so WARN about it.
2823                          */
2824                         WARN_ON(re_search);
2825                         goto out_lock;
2826                 }
2827         }
2828
2829         re_search = false;
2830         if (!info->bitmap) {
2831                 unlink_free_space(ctl, info, true);
2832                 if (offset == info->offset) {
2833                         u64 to_free = min(bytes, info->bytes);
2834
2835                         info->bytes -= to_free;
2836                         info->offset += to_free;
2837                         if (info->bytes) {
2838                                 ret = link_free_space(ctl, info);
2839                                 WARN_ON(ret);
2840                         } else {
2841                                 kmem_cache_free(btrfs_free_space_cachep, info);
2842                         }
2843
2844                         offset += to_free;
2845                         bytes -= to_free;
2846                         goto again;
2847                 } else {
2848                         u64 old_end = info->bytes + info->offset;
2849
2850                         info->bytes = offset - info->offset;
2851                         ret = link_free_space(ctl, info);
2852                         WARN_ON(ret);
2853                         if (ret)
2854                                 goto out_lock;
2855
2856                         /* Not enough bytes in this entry to satisfy us */
2857                         if (old_end < offset + bytes) {
2858                                 bytes -= old_end - offset;
2859                                 offset = old_end;
2860                                 goto again;
2861                         } else if (old_end == offset + bytes) {
2862                                 /* all done */
2863                                 goto out_lock;
2864                         }
2865                         spin_unlock(&ctl->tree_lock);
2866
2867                         ret = __btrfs_add_free_space(block_group,
2868                                                      offset + bytes,
2869                                                      old_end - (offset + bytes),
2870                                                      info->trim_state);
2871                         WARN_ON(ret);
2872                         goto out;
2873                 }
2874         }
2875
2876         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2877         if (ret == -EAGAIN) {
2878                 re_search = true;
2879                 goto again;
2880         }
2881 out_lock:
2882         btrfs_discard_update_discardable(block_group);
2883         spin_unlock(&ctl->tree_lock);
2884 out:
2885         return ret;
2886 }
2887
2888 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2889                            u64 bytes)
2890 {
2891         struct btrfs_fs_info *fs_info = block_group->fs_info;
2892         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2893         struct btrfs_free_space *info;
2894         struct rb_node *n;
2895         int count = 0;
2896
2897         /*
2898          * Zoned btrfs does not use free space tree and cluster. Just print
2899          * out the free space after the allocation offset.
2900          */
2901         if (btrfs_is_zoned(fs_info)) {
2902                 btrfs_info(fs_info, "free space %llu active %d",
2903                            block_group->zone_capacity - block_group->alloc_offset,
2904                            test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE,
2905                                     &block_group->runtime_flags));
2906                 return;
2907         }
2908
2909         spin_lock(&ctl->tree_lock);
2910         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2911                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2912                 if (info->bytes >= bytes && !block_group->ro)
2913                         count++;
2914                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2915                            info->offset, info->bytes,
2916                        (info->bitmap) ? "yes" : "no");
2917         }
2918         spin_unlock(&ctl->tree_lock);
2919         btrfs_info(fs_info, "block group has cluster?: %s",
2920                list_empty(&block_group->cluster_list) ? "no" : "yes");
2921         btrfs_info(fs_info,
2922                    "%d blocks of free space at or bigger than bytes is", count);
2923 }
2924
2925 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2926                                struct btrfs_free_space_ctl *ctl)
2927 {
2928         struct btrfs_fs_info *fs_info = block_group->fs_info;
2929
2930         spin_lock_init(&ctl->tree_lock);
2931         ctl->unit = fs_info->sectorsize;
2932         ctl->start = block_group->start;
2933         ctl->block_group = block_group;
2934         ctl->op = &free_space_op;
2935         ctl->free_space_bytes = RB_ROOT_CACHED;
2936         INIT_LIST_HEAD(&ctl->trimming_ranges);
2937         mutex_init(&ctl->cache_writeout_mutex);
2938
2939         /*
2940          * we only want to have 32k of ram per block group for keeping
2941          * track of free space, and if we pass 1/2 of that we want to
2942          * start converting things over to using bitmaps
2943          */
2944         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2945 }
2946
2947 /*
2948  * for a given cluster, put all of its extents back into the free
2949  * space cache.  If the block group passed doesn't match the block group
2950  * pointed to by the cluster, someone else raced in and freed the
2951  * cluster already.  In that case, we just return without changing anything
2952  */
2953 static void __btrfs_return_cluster_to_free_space(
2954                              struct btrfs_block_group *block_group,
2955                              struct btrfs_free_cluster *cluster)
2956 {
2957         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2958         struct btrfs_free_space *entry;
2959         struct rb_node *node;
2960
2961         spin_lock(&cluster->lock);
2962         if (cluster->block_group != block_group) {
2963                 spin_unlock(&cluster->lock);
2964                 return;
2965         }
2966
2967         cluster->block_group = NULL;
2968         cluster->window_start = 0;
2969         list_del_init(&cluster->block_group_list);
2970
2971         node = rb_first(&cluster->root);
2972         while (node) {
2973                 bool bitmap;
2974
2975                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2976                 node = rb_next(&entry->offset_index);
2977                 rb_erase(&entry->offset_index, &cluster->root);
2978                 RB_CLEAR_NODE(&entry->offset_index);
2979
2980                 bitmap = (entry->bitmap != NULL);
2981                 if (!bitmap) {
2982                         /* Merging treats extents as if they were new */
2983                         if (!btrfs_free_space_trimmed(entry)) {
2984                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2985                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2986                                         entry->bytes;
2987                         }
2988
2989                         try_merge_free_space(ctl, entry, false);
2990                         steal_from_bitmap(ctl, entry, false);
2991
2992                         /* As we insert directly, update these statistics */
2993                         if (!btrfs_free_space_trimmed(entry)) {
2994                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2995                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2996                                         entry->bytes;
2997                         }
2998                 }
2999                 tree_insert_offset(&ctl->free_space_offset,
3000                                    entry->offset, &entry->offset_index, bitmap);
3001                 rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
3002                               entry_less);
3003         }
3004         cluster->root = RB_ROOT;
3005         spin_unlock(&cluster->lock);
3006         btrfs_put_block_group(block_group);
3007 }
3008
3009 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
3010 {
3011         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3012         struct btrfs_free_cluster *cluster;
3013         struct list_head *head;
3014
3015         spin_lock(&ctl->tree_lock);
3016         while ((head = block_group->cluster_list.next) !=
3017                &block_group->cluster_list) {
3018                 cluster = list_entry(head, struct btrfs_free_cluster,
3019                                      block_group_list);
3020
3021                 WARN_ON(cluster->block_group != block_group);
3022                 __btrfs_return_cluster_to_free_space(block_group, cluster);
3023
3024                 cond_resched_lock(&ctl->tree_lock);
3025         }
3026         __btrfs_remove_free_space_cache(ctl);
3027         btrfs_discard_update_discardable(block_group);
3028         spin_unlock(&ctl->tree_lock);
3029
3030 }
3031
3032 /**
3033  * btrfs_is_free_space_trimmed - see if everything is trimmed
3034  * @block_group: block_group of interest
3035  *
3036  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3037  */
3038 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3039 {
3040         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3041         struct btrfs_free_space *info;
3042         struct rb_node *node;
3043         bool ret = true;
3044
3045         spin_lock(&ctl->tree_lock);
3046         node = rb_first(&ctl->free_space_offset);
3047
3048         while (node) {
3049                 info = rb_entry(node, struct btrfs_free_space, offset_index);
3050
3051                 if (!btrfs_free_space_trimmed(info)) {
3052                         ret = false;
3053                         break;
3054                 }
3055
3056                 node = rb_next(node);
3057         }
3058
3059         spin_unlock(&ctl->tree_lock);
3060         return ret;
3061 }
3062
3063 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3064                                u64 offset, u64 bytes, u64 empty_size,
3065                                u64 *max_extent_size)
3066 {
3067         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3068         struct btrfs_discard_ctl *discard_ctl =
3069                                         &block_group->fs_info->discard_ctl;
3070         struct btrfs_free_space *entry = NULL;
3071         u64 bytes_search = bytes + empty_size;
3072         u64 ret = 0;
3073         u64 align_gap = 0;
3074         u64 align_gap_len = 0;
3075         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3076         bool use_bytes_index = (offset == block_group->start);
3077
3078         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3079
3080         spin_lock(&ctl->tree_lock);
3081         entry = find_free_space(ctl, &offset, &bytes_search,
3082                                 block_group->full_stripe_len, max_extent_size,
3083                                 use_bytes_index);
3084         if (!entry)
3085                 goto out;
3086
3087         ret = offset;
3088         if (entry->bitmap) {
3089                 bitmap_clear_bits(ctl, entry, offset, bytes, true);
3090
3091                 if (!btrfs_free_space_trimmed(entry))
3092                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3093
3094                 if (!entry->bytes)
3095                         free_bitmap(ctl, entry);
3096         } else {
3097                 unlink_free_space(ctl, entry, true);
3098                 align_gap_len = offset - entry->offset;
3099                 align_gap = entry->offset;
3100                 align_gap_trim_state = entry->trim_state;
3101
3102                 if (!btrfs_free_space_trimmed(entry))
3103                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3104
3105                 entry->offset = offset + bytes;
3106                 WARN_ON(entry->bytes < bytes + align_gap_len);
3107
3108                 entry->bytes -= bytes + align_gap_len;
3109                 if (!entry->bytes)
3110                         kmem_cache_free(btrfs_free_space_cachep, entry);
3111                 else
3112                         link_free_space(ctl, entry);
3113         }
3114 out:
3115         btrfs_discard_update_discardable(block_group);
3116         spin_unlock(&ctl->tree_lock);
3117
3118         if (align_gap_len)
3119                 __btrfs_add_free_space(block_group, align_gap, align_gap_len,
3120                                        align_gap_trim_state);
3121         return ret;
3122 }
3123
3124 /*
3125  * given a cluster, put all of its extents back into the free space
3126  * cache.  If a block group is passed, this function will only free
3127  * a cluster that belongs to the passed block group.
3128  *
3129  * Otherwise, it'll get a reference on the block group pointed to by the
3130  * cluster and remove the cluster from it.
3131  */
3132 void btrfs_return_cluster_to_free_space(
3133                                struct btrfs_block_group *block_group,
3134                                struct btrfs_free_cluster *cluster)
3135 {
3136         struct btrfs_free_space_ctl *ctl;
3137
3138         /* first, get a safe pointer to the block group */
3139         spin_lock(&cluster->lock);
3140         if (!block_group) {
3141                 block_group = cluster->block_group;
3142                 if (!block_group) {
3143                         spin_unlock(&cluster->lock);
3144                         return;
3145                 }
3146         } else if (cluster->block_group != block_group) {
3147                 /* someone else has already freed it don't redo their work */
3148                 spin_unlock(&cluster->lock);
3149                 return;
3150         }
3151         btrfs_get_block_group(block_group);
3152         spin_unlock(&cluster->lock);
3153
3154         ctl = block_group->free_space_ctl;
3155
3156         /* now return any extents the cluster had on it */
3157         spin_lock(&ctl->tree_lock);
3158         __btrfs_return_cluster_to_free_space(block_group, cluster);
3159         spin_unlock(&ctl->tree_lock);
3160
3161         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3162
3163         /* finally drop our ref */
3164         btrfs_put_block_group(block_group);
3165 }
3166
3167 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3168                                    struct btrfs_free_cluster *cluster,
3169                                    struct btrfs_free_space *entry,
3170                                    u64 bytes, u64 min_start,
3171                                    u64 *max_extent_size)
3172 {
3173         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3174         int err;
3175         u64 search_start = cluster->window_start;
3176         u64 search_bytes = bytes;
3177         u64 ret = 0;
3178
3179         search_start = min_start;
3180         search_bytes = bytes;
3181
3182         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3183         if (err) {
3184                 *max_extent_size = max(get_max_extent_size(entry),
3185                                        *max_extent_size);
3186                 return 0;
3187         }
3188
3189         ret = search_start;
3190         bitmap_clear_bits(ctl, entry, ret, bytes, false);
3191
3192         return ret;
3193 }
3194
3195 /*
3196  * given a cluster, try to allocate 'bytes' from it, returns 0
3197  * if it couldn't find anything suitably large, or a logical disk offset
3198  * if things worked out
3199  */
3200 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3201                              struct btrfs_free_cluster *cluster, u64 bytes,
3202                              u64 min_start, u64 *max_extent_size)
3203 {
3204         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3205         struct btrfs_discard_ctl *discard_ctl =
3206                                         &block_group->fs_info->discard_ctl;
3207         struct btrfs_free_space *entry = NULL;
3208         struct rb_node *node;
3209         u64 ret = 0;
3210
3211         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3212
3213         spin_lock(&cluster->lock);
3214         if (bytes > cluster->max_size)
3215                 goto out;
3216
3217         if (cluster->block_group != block_group)
3218                 goto out;
3219
3220         node = rb_first(&cluster->root);
3221         if (!node)
3222                 goto out;
3223
3224         entry = rb_entry(node, struct btrfs_free_space, offset_index);
3225         while (1) {
3226                 if (entry->bytes < bytes)
3227                         *max_extent_size = max(get_max_extent_size(entry),
3228                                                *max_extent_size);
3229
3230                 if (entry->bytes < bytes ||
3231                     (!entry->bitmap && entry->offset < min_start)) {
3232                         node = rb_next(&entry->offset_index);
3233                         if (!node)
3234                                 break;
3235                         entry = rb_entry(node, struct btrfs_free_space,
3236                                          offset_index);
3237                         continue;
3238                 }
3239
3240                 if (entry->bitmap) {
3241                         ret = btrfs_alloc_from_bitmap(block_group,
3242                                                       cluster, entry, bytes,
3243                                                       cluster->window_start,
3244                                                       max_extent_size);
3245                         if (ret == 0) {
3246                                 node = rb_next(&entry->offset_index);
3247                                 if (!node)
3248                                         break;
3249                                 entry = rb_entry(node, struct btrfs_free_space,
3250                                                  offset_index);
3251                                 continue;
3252                         }
3253                         cluster->window_start += bytes;
3254                 } else {
3255                         ret = entry->offset;
3256
3257                         entry->offset += bytes;
3258                         entry->bytes -= bytes;
3259                 }
3260
3261                 break;
3262         }
3263 out:
3264         spin_unlock(&cluster->lock);
3265
3266         if (!ret)
3267                 return 0;
3268
3269         spin_lock(&ctl->tree_lock);
3270
3271         if (!btrfs_free_space_trimmed(entry))
3272                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3273
3274         ctl->free_space -= bytes;
3275         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3276                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3277
3278         spin_lock(&cluster->lock);
3279         if (entry->bytes == 0) {
3280                 rb_erase(&entry->offset_index, &cluster->root);
3281                 ctl->free_extents--;
3282                 if (entry->bitmap) {
3283                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3284                                         entry->bitmap);
3285                         ctl->total_bitmaps--;
3286                         recalculate_thresholds(ctl);
3287                 } else if (!btrfs_free_space_trimmed(entry)) {
3288                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3289                 }
3290                 kmem_cache_free(btrfs_free_space_cachep, entry);
3291         }
3292
3293         spin_unlock(&cluster->lock);
3294         spin_unlock(&ctl->tree_lock);
3295
3296         return ret;
3297 }
3298
3299 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3300                                 struct btrfs_free_space *entry,
3301                                 struct btrfs_free_cluster *cluster,
3302                                 u64 offset, u64 bytes,
3303                                 u64 cont1_bytes, u64 min_bytes)
3304 {
3305         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3306         unsigned long next_zero;
3307         unsigned long i;
3308         unsigned long want_bits;
3309         unsigned long min_bits;
3310         unsigned long found_bits;
3311         unsigned long max_bits = 0;
3312         unsigned long start = 0;
3313         unsigned long total_found = 0;
3314         int ret;
3315
3316         i = offset_to_bit(entry->offset, ctl->unit,
3317                           max_t(u64, offset, entry->offset));
3318         want_bits = bytes_to_bits(bytes, ctl->unit);
3319         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3320
3321         /*
3322          * Don't bother looking for a cluster in this bitmap if it's heavily
3323          * fragmented.
3324          */
3325         if (entry->max_extent_size &&
3326             entry->max_extent_size < cont1_bytes)
3327                 return -ENOSPC;
3328 again:
3329         found_bits = 0;
3330         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3331                 next_zero = find_next_zero_bit(entry->bitmap,
3332                                                BITS_PER_BITMAP, i);
3333                 if (next_zero - i >= min_bits) {
3334                         found_bits = next_zero - i;
3335                         if (found_bits > max_bits)
3336                                 max_bits = found_bits;
3337                         break;
3338                 }
3339                 if (next_zero - i > max_bits)
3340                         max_bits = next_zero - i;
3341                 i = next_zero;
3342         }
3343
3344         if (!found_bits) {
3345                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3346                 return -ENOSPC;
3347         }
3348
3349         if (!total_found) {
3350                 start = i;
3351                 cluster->max_size = 0;
3352         }
3353
3354         total_found += found_bits;
3355
3356         if (cluster->max_size < found_bits * ctl->unit)
3357                 cluster->max_size = found_bits * ctl->unit;
3358
3359         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3360                 i = next_zero + 1;
3361                 goto again;
3362         }
3363
3364         cluster->window_start = start * ctl->unit + entry->offset;
3365         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3366         rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3367
3368         /*
3369          * We need to know if we're currently on the normal space index when we
3370          * manipulate the bitmap so that we know we need to remove and re-insert
3371          * it into the space_index tree.  Clear the bytes_index node here so the
3372          * bitmap manipulation helpers know not to mess with the space_index
3373          * until this bitmap entry is added back into the normal cache.
3374          */
3375         RB_CLEAR_NODE(&entry->bytes_index);
3376
3377         ret = tree_insert_offset(&cluster->root, entry->offset,
3378                                  &entry->offset_index, 1);
3379         ASSERT(!ret); /* -EEXIST; Logic error */
3380
3381         trace_btrfs_setup_cluster(block_group, cluster,
3382                                   total_found * ctl->unit, 1);
3383         return 0;
3384 }
3385
3386 /*
3387  * This searches the block group for just extents to fill the cluster with.
3388  * Try to find a cluster with at least bytes total bytes, at least one
3389  * extent of cont1_bytes, and other clusters of at least min_bytes.
3390  */
3391 static noinline int
3392 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3393                         struct btrfs_free_cluster *cluster,
3394                         struct list_head *bitmaps, u64 offset, u64 bytes,
3395                         u64 cont1_bytes, u64 min_bytes)
3396 {
3397         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3398         struct btrfs_free_space *first = NULL;
3399         struct btrfs_free_space *entry = NULL;
3400         struct btrfs_free_space *last;
3401         struct rb_node *node;
3402         u64 window_free;
3403         u64 max_extent;
3404         u64 total_size = 0;
3405
3406         entry = tree_search_offset(ctl, offset, 0, 1);
3407         if (!entry)
3408                 return -ENOSPC;
3409
3410         /*
3411          * We don't want bitmaps, so just move along until we find a normal
3412          * extent entry.
3413          */
3414         while (entry->bitmap || entry->bytes < min_bytes) {
3415                 if (entry->bitmap && list_empty(&entry->list))
3416                         list_add_tail(&entry->list, bitmaps);
3417                 node = rb_next(&entry->offset_index);
3418                 if (!node)
3419                         return -ENOSPC;
3420                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3421         }
3422
3423         window_free = entry->bytes;
3424         max_extent = entry->bytes;
3425         first = entry;
3426         last = entry;
3427
3428         for (node = rb_next(&entry->offset_index); node;
3429              node = rb_next(&entry->offset_index)) {
3430                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3431
3432                 if (entry->bitmap) {
3433                         if (list_empty(&entry->list))
3434                                 list_add_tail(&entry->list, bitmaps);
3435                         continue;
3436                 }
3437
3438                 if (entry->bytes < min_bytes)
3439                         continue;
3440
3441                 last = entry;
3442                 window_free += entry->bytes;
3443                 if (entry->bytes > max_extent)
3444                         max_extent = entry->bytes;
3445         }
3446
3447         if (window_free < bytes || max_extent < cont1_bytes)
3448                 return -ENOSPC;
3449
3450         cluster->window_start = first->offset;
3451
3452         node = &first->offset_index;
3453
3454         /*
3455          * now we've found our entries, pull them out of the free space
3456          * cache and put them into the cluster rbtree
3457          */
3458         do {
3459                 int ret;
3460
3461                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3462                 node = rb_next(&entry->offset_index);
3463                 if (entry->bitmap || entry->bytes < min_bytes)
3464                         continue;
3465
3466                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3467                 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3468                 ret = tree_insert_offset(&cluster->root, entry->offset,
3469                                          &entry->offset_index, 0);
3470                 total_size += entry->bytes;
3471                 ASSERT(!ret); /* -EEXIST; Logic error */
3472         } while (node && entry != last);
3473
3474         cluster->max_size = max_extent;
3475         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3476         return 0;
3477 }
3478
3479 /*
3480  * This specifically looks for bitmaps that may work in the cluster, we assume
3481  * that we have already failed to find extents that will work.
3482  */
3483 static noinline int
3484 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3485                      struct btrfs_free_cluster *cluster,
3486                      struct list_head *bitmaps, u64 offset, u64 bytes,
3487                      u64 cont1_bytes, u64 min_bytes)
3488 {
3489         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3490         struct btrfs_free_space *entry = NULL;
3491         int ret = -ENOSPC;
3492         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3493
3494         if (ctl->total_bitmaps == 0)
3495                 return -ENOSPC;
3496
3497         /*
3498          * The bitmap that covers offset won't be in the list unless offset
3499          * is just its start offset.
3500          */
3501         if (!list_empty(bitmaps))
3502                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3503
3504         if (!entry || entry->offset != bitmap_offset) {
3505                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3506                 if (entry && list_empty(&entry->list))
3507                         list_add(&entry->list, bitmaps);
3508         }
3509
3510         list_for_each_entry(entry, bitmaps, list) {
3511                 if (entry->bytes < bytes)
3512                         continue;
3513                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3514                                            bytes, cont1_bytes, min_bytes);
3515                 if (!ret)
3516                         return 0;
3517         }
3518
3519         /*
3520          * The bitmaps list has all the bitmaps that record free space
3521          * starting after offset, so no more search is required.
3522          */
3523         return -ENOSPC;
3524 }
3525
3526 /*
3527  * here we try to find a cluster of blocks in a block group.  The goal
3528  * is to find at least bytes+empty_size.
3529  * We might not find them all in one contiguous area.
3530  *
3531  * returns zero and sets up cluster if things worked out, otherwise
3532  * it returns -enospc
3533  */
3534 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3535                              struct btrfs_free_cluster *cluster,
3536                              u64 offset, u64 bytes, u64 empty_size)
3537 {
3538         struct btrfs_fs_info *fs_info = block_group->fs_info;
3539         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3540         struct btrfs_free_space *entry, *tmp;
3541         LIST_HEAD(bitmaps);
3542         u64 min_bytes;
3543         u64 cont1_bytes;
3544         int ret;
3545
3546         /*
3547          * Choose the minimum extent size we'll require for this
3548          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3549          * For metadata, allow allocates with smaller extents.  For
3550          * data, keep it dense.
3551          */
3552         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3553                 cont1_bytes = bytes + empty_size;
3554                 min_bytes = cont1_bytes;
3555         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3556                 cont1_bytes = bytes;
3557                 min_bytes = fs_info->sectorsize;
3558         } else {
3559                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3560                 min_bytes = fs_info->sectorsize;
3561         }
3562
3563         spin_lock(&ctl->tree_lock);
3564
3565         /*
3566          * If we know we don't have enough space to make a cluster don't even
3567          * bother doing all the work to try and find one.
3568          */
3569         if (ctl->free_space < bytes) {
3570                 spin_unlock(&ctl->tree_lock);
3571                 return -ENOSPC;
3572         }
3573
3574         spin_lock(&cluster->lock);
3575
3576         /* someone already found a cluster, hooray */
3577         if (cluster->block_group) {
3578                 ret = 0;
3579                 goto out;
3580         }
3581
3582         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3583                                  min_bytes);
3584
3585         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3586                                       bytes + empty_size,
3587                                       cont1_bytes, min_bytes);
3588         if (ret)
3589                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3590                                            offset, bytes + empty_size,
3591                                            cont1_bytes, min_bytes);
3592
3593         /* Clear our temporary list */
3594         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3595                 list_del_init(&entry->list);
3596
3597         if (!ret) {
3598                 btrfs_get_block_group(block_group);
3599                 list_add_tail(&cluster->block_group_list,
3600                               &block_group->cluster_list);
3601                 cluster->block_group = block_group;
3602         } else {
3603                 trace_btrfs_failed_cluster_setup(block_group);
3604         }
3605 out:
3606         spin_unlock(&cluster->lock);
3607         spin_unlock(&ctl->tree_lock);
3608
3609         return ret;
3610 }
3611
3612 /*
3613  * simple code to zero out a cluster
3614  */
3615 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3616 {
3617         spin_lock_init(&cluster->lock);
3618         spin_lock_init(&cluster->refill_lock);
3619         cluster->root = RB_ROOT;
3620         cluster->max_size = 0;
3621         cluster->fragmented = false;
3622         INIT_LIST_HEAD(&cluster->block_group_list);
3623         cluster->block_group = NULL;
3624 }
3625
3626 static int do_trimming(struct btrfs_block_group *block_group,
3627                        u64 *total_trimmed, u64 start, u64 bytes,
3628                        u64 reserved_start, u64 reserved_bytes,
3629                        enum btrfs_trim_state reserved_trim_state,
3630                        struct btrfs_trim_range *trim_entry)
3631 {
3632         struct btrfs_space_info *space_info = block_group->space_info;
3633         struct btrfs_fs_info *fs_info = block_group->fs_info;
3634         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3635         int ret;
3636         int update = 0;
3637         const u64 end = start + bytes;
3638         const u64 reserved_end = reserved_start + reserved_bytes;
3639         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3640         u64 trimmed = 0;
3641
3642         spin_lock(&space_info->lock);
3643         spin_lock(&block_group->lock);
3644         if (!block_group->ro) {
3645                 block_group->reserved += reserved_bytes;
3646                 space_info->bytes_reserved += reserved_bytes;
3647                 update = 1;
3648         }
3649         spin_unlock(&block_group->lock);
3650         spin_unlock(&space_info->lock);
3651
3652         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3653         if (!ret) {
3654                 *total_trimmed += trimmed;
3655                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3656         }
3657
3658         mutex_lock(&ctl->cache_writeout_mutex);
3659         if (reserved_start < start)
3660                 __btrfs_add_free_space(block_group, reserved_start,
3661                                        start - reserved_start,
3662                                        reserved_trim_state);
3663         if (start + bytes < reserved_start + reserved_bytes)
3664                 __btrfs_add_free_space(block_group, end, reserved_end - end,
3665                                        reserved_trim_state);
3666         __btrfs_add_free_space(block_group, start, bytes, trim_state);
3667         list_del(&trim_entry->list);
3668         mutex_unlock(&ctl->cache_writeout_mutex);
3669
3670         if (update) {
3671                 spin_lock(&space_info->lock);
3672                 spin_lock(&block_group->lock);
3673                 if (block_group->ro)
3674                         space_info->bytes_readonly += reserved_bytes;
3675                 block_group->reserved -= reserved_bytes;
3676                 space_info->bytes_reserved -= reserved_bytes;
3677                 spin_unlock(&block_group->lock);
3678                 spin_unlock(&space_info->lock);
3679         }
3680
3681         return ret;
3682 }
3683
3684 /*
3685  * If @async is set, then we will trim 1 region and return.
3686  */
3687 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3688                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3689                           bool async)
3690 {
3691         struct btrfs_discard_ctl *discard_ctl =
3692                                         &block_group->fs_info->discard_ctl;
3693         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3694         struct btrfs_free_space *entry;
3695         struct rb_node *node;
3696         int ret = 0;
3697         u64 extent_start;
3698         u64 extent_bytes;
3699         enum btrfs_trim_state extent_trim_state;
3700         u64 bytes;
3701         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3702
3703         while (start < end) {
3704                 struct btrfs_trim_range trim_entry;
3705
3706                 mutex_lock(&ctl->cache_writeout_mutex);
3707                 spin_lock(&ctl->tree_lock);
3708
3709                 if (ctl->free_space < minlen)
3710                         goto out_unlock;
3711
3712                 entry = tree_search_offset(ctl, start, 0, 1);
3713                 if (!entry)
3714                         goto out_unlock;
3715
3716                 /* Skip bitmaps and if async, already trimmed entries */
3717                 while (entry->bitmap ||
3718                        (async && btrfs_free_space_trimmed(entry))) {
3719                         node = rb_next(&entry->offset_index);
3720                         if (!node)
3721                                 goto out_unlock;
3722                         entry = rb_entry(node, struct btrfs_free_space,
3723                                          offset_index);
3724                 }
3725
3726                 if (entry->offset >= end)
3727                         goto out_unlock;
3728
3729                 extent_start = entry->offset;
3730                 extent_bytes = entry->bytes;
3731                 extent_trim_state = entry->trim_state;
3732                 if (async) {
3733                         start = entry->offset;
3734                         bytes = entry->bytes;
3735                         if (bytes < minlen) {
3736                                 spin_unlock(&ctl->tree_lock);
3737                                 mutex_unlock(&ctl->cache_writeout_mutex);
3738                                 goto next;
3739                         }
3740                         unlink_free_space(ctl, entry, true);
3741                         /*
3742                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3743                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3744                          * X when we come back around.  So trim it now.
3745                          */
3746                         if (max_discard_size &&
3747                             bytes >= (max_discard_size +
3748                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3749                                 bytes = max_discard_size;
3750                                 extent_bytes = max_discard_size;
3751                                 entry->offset += max_discard_size;
3752                                 entry->bytes -= max_discard_size;
3753                                 link_free_space(ctl, entry);
3754                         } else {
3755                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3756                         }
3757                 } else {
3758                         start = max(start, extent_start);
3759                         bytes = min(extent_start + extent_bytes, end) - start;
3760                         if (bytes < minlen) {
3761                                 spin_unlock(&ctl->tree_lock);
3762                                 mutex_unlock(&ctl->cache_writeout_mutex);
3763                                 goto next;
3764                         }
3765
3766                         unlink_free_space(ctl, entry, true);
3767                         kmem_cache_free(btrfs_free_space_cachep, entry);
3768                 }
3769
3770                 spin_unlock(&ctl->tree_lock);
3771                 trim_entry.start = extent_start;
3772                 trim_entry.bytes = extent_bytes;
3773                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3774                 mutex_unlock(&ctl->cache_writeout_mutex);
3775
3776                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3777                                   extent_start, extent_bytes, extent_trim_state,
3778                                   &trim_entry);
3779                 if (ret) {
3780                         block_group->discard_cursor = start + bytes;
3781                         break;
3782                 }
3783 next:
3784                 start += bytes;
3785                 block_group->discard_cursor = start;
3786                 if (async && *total_trimmed)
3787                         break;
3788
3789                 if (fatal_signal_pending(current)) {
3790                         ret = -ERESTARTSYS;
3791                         break;
3792                 }
3793
3794                 cond_resched();
3795         }
3796
3797         return ret;
3798
3799 out_unlock:
3800         block_group->discard_cursor = btrfs_block_group_end(block_group);
3801         spin_unlock(&ctl->tree_lock);
3802         mutex_unlock(&ctl->cache_writeout_mutex);
3803
3804         return ret;
3805 }
3806
3807 /*
3808  * If we break out of trimming a bitmap prematurely, we should reset the
3809  * trimming bit.  In a rather contrieved case, it's possible to race here so
3810  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3811  *
3812  * start = start of bitmap
3813  * end = near end of bitmap
3814  *
3815  * Thread 1:                    Thread 2:
3816  * trim_bitmaps(start)
3817  *                              trim_bitmaps(end)
3818  *                              end_trimming_bitmap()
3819  * reset_trimming_bitmap()
3820  */
3821 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3822 {
3823         struct btrfs_free_space *entry;
3824
3825         spin_lock(&ctl->tree_lock);
3826         entry = tree_search_offset(ctl, offset, 1, 0);
3827         if (entry) {
3828                 if (btrfs_free_space_trimmed(entry)) {
3829                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3830                                 entry->bitmap_extents;
3831                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3832                 }
3833                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3834         }
3835
3836         spin_unlock(&ctl->tree_lock);
3837 }
3838
3839 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3840                                 struct btrfs_free_space *entry)
3841 {
3842         if (btrfs_free_space_trimming_bitmap(entry)) {
3843                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3844                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3845                         entry->bitmap_extents;
3846                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3847         }
3848 }
3849
3850 /*
3851  * If @async is set, then we will trim 1 region and return.
3852  */
3853 static int trim_bitmaps(struct btrfs_block_group *block_group,
3854                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3855                         u64 maxlen, bool async)
3856 {
3857         struct btrfs_discard_ctl *discard_ctl =
3858                                         &block_group->fs_info->discard_ctl;
3859         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3860         struct btrfs_free_space *entry;
3861         int ret = 0;
3862         int ret2;
3863         u64 bytes;
3864         u64 offset = offset_to_bitmap(ctl, start);
3865         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3866
3867         while (offset < end) {
3868                 bool next_bitmap = false;
3869                 struct btrfs_trim_range trim_entry;
3870
3871                 mutex_lock(&ctl->cache_writeout_mutex);
3872                 spin_lock(&ctl->tree_lock);
3873
3874                 if (ctl->free_space < minlen) {
3875                         block_group->discard_cursor =
3876                                 btrfs_block_group_end(block_group);
3877                         spin_unlock(&ctl->tree_lock);
3878                         mutex_unlock(&ctl->cache_writeout_mutex);
3879                         break;
3880                 }
3881
3882                 entry = tree_search_offset(ctl, offset, 1, 0);
3883                 /*
3884                  * Bitmaps are marked trimmed lossily now to prevent constant
3885                  * discarding of the same bitmap (the reason why we are bound
3886                  * by the filters).  So, retrim the block group bitmaps when we
3887                  * are preparing to punt to the unused_bgs list.  This uses
3888                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3889                  * which is the only discard index which sets minlen to 0.
3890                  */
3891                 if (!entry || (async && minlen && start == offset &&
3892                                btrfs_free_space_trimmed(entry))) {
3893                         spin_unlock(&ctl->tree_lock);
3894                         mutex_unlock(&ctl->cache_writeout_mutex);
3895                         next_bitmap = true;
3896                         goto next;
3897                 }
3898
3899                 /*
3900                  * Async discard bitmap trimming begins at by setting the start
3901                  * to be key.objectid and the offset_to_bitmap() aligns to the
3902                  * start of the bitmap.  This lets us know we are fully
3903                  * scanning the bitmap rather than only some portion of it.
3904                  */
3905                 if (start == offset)
3906                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3907
3908                 bytes = minlen;
3909                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3910                 if (ret2 || start >= end) {
3911                         /*
3912                          * We lossily consider a bitmap trimmed if we only skip
3913                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3914                          */
3915                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3916                                 end_trimming_bitmap(ctl, entry);
3917                         else
3918                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3919                         spin_unlock(&ctl->tree_lock);
3920                         mutex_unlock(&ctl->cache_writeout_mutex);
3921                         next_bitmap = true;
3922                         goto next;
3923                 }
3924
3925                 /*
3926                  * We already trimmed a region, but are using the locking above
3927                  * to reset the trim_state.
3928                  */
3929                 if (async && *total_trimmed) {
3930                         spin_unlock(&ctl->tree_lock);
3931                         mutex_unlock(&ctl->cache_writeout_mutex);
3932                         goto out;
3933                 }
3934
3935                 bytes = min(bytes, end - start);
3936                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3937                         spin_unlock(&ctl->tree_lock);
3938                         mutex_unlock(&ctl->cache_writeout_mutex);
3939                         goto next;
3940                 }
3941
3942                 /*
3943                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3944                  * If X < @minlen, we won't trim X when we come back around.
3945                  * So trim it now.  We differ here from trimming extents as we
3946                  * don't keep individual state per bit.
3947                  */
3948                 if (async &&
3949                     max_discard_size &&
3950                     bytes > (max_discard_size + minlen))
3951                         bytes = max_discard_size;
3952
3953                 bitmap_clear_bits(ctl, entry, start, bytes, true);
3954                 if (entry->bytes == 0)
3955                         free_bitmap(ctl, entry);
3956
3957                 spin_unlock(&ctl->tree_lock);
3958                 trim_entry.start = start;
3959                 trim_entry.bytes = bytes;
3960                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3961                 mutex_unlock(&ctl->cache_writeout_mutex);
3962
3963                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3964                                   start, bytes, 0, &trim_entry);
3965                 if (ret) {
3966                         reset_trimming_bitmap(ctl, offset);
3967                         block_group->discard_cursor =
3968                                 btrfs_block_group_end(block_group);
3969                         break;
3970                 }
3971 next:
3972                 if (next_bitmap) {
3973                         offset += BITS_PER_BITMAP * ctl->unit;
3974                         start = offset;
3975                 } else {
3976                         start += bytes;
3977                 }
3978                 block_group->discard_cursor = start;
3979
3980                 if (fatal_signal_pending(current)) {
3981                         if (start != offset)
3982                                 reset_trimming_bitmap(ctl, offset);
3983                         ret = -ERESTARTSYS;
3984                         break;
3985                 }
3986
3987                 cond_resched();
3988         }
3989
3990         if (offset >= end)
3991                 block_group->discard_cursor = end;
3992
3993 out:
3994         return ret;
3995 }
3996
3997 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3998                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3999 {
4000         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
4001         int ret;
4002         u64 rem = 0;
4003
4004         ASSERT(!btrfs_is_zoned(block_group->fs_info));
4005
4006         *trimmed = 0;
4007
4008         spin_lock(&block_group->lock);
4009         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4010                 spin_unlock(&block_group->lock);
4011                 return 0;
4012         }
4013         btrfs_freeze_block_group(block_group);
4014         spin_unlock(&block_group->lock);
4015
4016         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4017         if (ret)
4018                 goto out;
4019
4020         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4021         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4022         /* If we ended in the middle of a bitmap, reset the trimming flag */
4023         if (rem)
4024                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4025 out:
4026         btrfs_unfreeze_block_group(block_group);
4027         return ret;
4028 }
4029
4030 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4031                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4032                                    bool async)
4033 {
4034         int ret;
4035
4036         *trimmed = 0;
4037
4038         spin_lock(&block_group->lock);
4039         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4040                 spin_unlock(&block_group->lock);
4041                 return 0;
4042         }
4043         btrfs_freeze_block_group(block_group);
4044         spin_unlock(&block_group->lock);
4045
4046         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4047         btrfs_unfreeze_block_group(block_group);
4048
4049         return ret;
4050 }
4051
4052 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4053                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4054                                    u64 maxlen, bool async)
4055 {
4056         int ret;
4057
4058         *trimmed = 0;
4059
4060         spin_lock(&block_group->lock);
4061         if (test_bit(BLOCK_GROUP_FLAG_REMOVED, &block_group->runtime_flags)) {
4062                 spin_unlock(&block_group->lock);
4063                 return 0;
4064         }
4065         btrfs_freeze_block_group(block_group);
4066         spin_unlock(&block_group->lock);
4067
4068         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4069                            async);
4070
4071         btrfs_unfreeze_block_group(block_group);
4072
4073         return ret;
4074 }
4075
4076 bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4077 {
4078         return btrfs_super_cache_generation(fs_info->super_copy);
4079 }
4080
4081 static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4082                                        struct btrfs_trans_handle *trans)
4083 {
4084         struct btrfs_block_group *block_group;
4085         struct rb_node *node;
4086         int ret = 0;
4087
4088         btrfs_info(fs_info, "cleaning free space cache v1");
4089
4090         node = rb_first_cached(&fs_info->block_group_cache_tree);
4091         while (node) {
4092                 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4093                 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4094                 if (ret)
4095                         goto out;
4096                 node = rb_next(node);
4097         }
4098 out:
4099         return ret;
4100 }
4101
4102 int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4103 {
4104         struct btrfs_trans_handle *trans;
4105         int ret;
4106
4107         /*
4108          * update_super_roots will appropriately set or unset
4109          * super_copy->cache_generation based on SPACE_CACHE and
4110          * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4111          * transaction commit whether we are enabling space cache v1 and don't
4112          * have any other work to do, or are disabling it and removing free
4113          * space inodes.
4114          */
4115         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4116         if (IS_ERR(trans))
4117                 return PTR_ERR(trans);
4118
4119         if (!active) {
4120                 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4121                 ret = cleanup_free_space_cache_v1(fs_info, trans);
4122                 if (ret) {
4123                         btrfs_abort_transaction(trans, ret);
4124                         btrfs_end_transaction(trans);
4125                         goto out;
4126                 }
4127         }
4128
4129         ret = btrfs_commit_transaction(trans);
4130 out:
4131         clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4132
4133         return ret;
4134 }
4135
4136 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4137 /*
4138  * Use this if you need to make a bitmap or extent entry specifically, it
4139  * doesn't do any of the merging that add_free_space does, this acts a lot like
4140  * how the free space cache loading stuff works, so you can get really weird
4141  * configurations.
4142  */
4143 int test_add_free_space_entry(struct btrfs_block_group *cache,
4144                               u64 offset, u64 bytes, bool bitmap)
4145 {
4146         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4147         struct btrfs_free_space *info = NULL, *bitmap_info;
4148         void *map = NULL;
4149         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4150         u64 bytes_added;
4151         int ret;
4152
4153 again:
4154         if (!info) {
4155                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4156                 if (!info)
4157                         return -ENOMEM;
4158         }
4159
4160         if (!bitmap) {
4161                 spin_lock(&ctl->tree_lock);
4162                 info->offset = offset;
4163                 info->bytes = bytes;
4164                 info->max_extent_size = 0;
4165                 ret = link_free_space(ctl, info);
4166                 spin_unlock(&ctl->tree_lock);
4167                 if (ret)
4168                         kmem_cache_free(btrfs_free_space_cachep, info);
4169                 return ret;
4170         }
4171
4172         if (!map) {
4173                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4174                 if (!map) {
4175                         kmem_cache_free(btrfs_free_space_cachep, info);
4176                         return -ENOMEM;
4177                 }
4178         }
4179
4180         spin_lock(&ctl->tree_lock);
4181         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4182                                          1, 0);
4183         if (!bitmap_info) {
4184                 info->bitmap = map;
4185                 map = NULL;
4186                 add_new_bitmap(ctl, info, offset);
4187                 bitmap_info = info;
4188                 info = NULL;
4189         }
4190
4191         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4192                                           trim_state);
4193
4194         bytes -= bytes_added;
4195         offset += bytes_added;
4196         spin_unlock(&ctl->tree_lock);
4197
4198         if (bytes)
4199                 goto again;
4200
4201         if (info)
4202                 kmem_cache_free(btrfs_free_space_cachep, info);
4203         if (map)
4204                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4205         return 0;
4206 }
4207
4208 /*
4209  * Checks to see if the given range is in the free space cache.  This is really
4210  * just used to check the absence of space, so if there is free space in the
4211  * range at all we will return 1.
4212  */
4213 int test_check_exists(struct btrfs_block_group *cache,
4214                       u64 offset, u64 bytes)
4215 {
4216         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4217         struct btrfs_free_space *info;
4218         int ret = 0;
4219
4220         spin_lock(&ctl->tree_lock);
4221         info = tree_search_offset(ctl, offset, 0, 0);
4222         if (!info) {
4223                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4224                                           1, 0);
4225                 if (!info)
4226                         goto out;
4227         }
4228
4229 have_info:
4230         if (info->bitmap) {
4231                 u64 bit_off, bit_bytes;
4232                 struct rb_node *n;
4233                 struct btrfs_free_space *tmp;
4234
4235                 bit_off = offset;
4236                 bit_bytes = ctl->unit;
4237                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4238                 if (!ret) {
4239                         if (bit_off == offset) {
4240                                 ret = 1;
4241                                 goto out;
4242                         } else if (bit_off > offset &&
4243                                    offset + bytes > bit_off) {
4244                                 ret = 1;
4245                                 goto out;
4246                         }
4247                 }
4248
4249                 n = rb_prev(&info->offset_index);
4250                 while (n) {
4251                         tmp = rb_entry(n, struct btrfs_free_space,
4252                                        offset_index);
4253                         if (tmp->offset + tmp->bytes < offset)
4254                                 break;
4255                         if (offset + bytes < tmp->offset) {
4256                                 n = rb_prev(&tmp->offset_index);
4257                                 continue;
4258                         }
4259                         info = tmp;
4260                         goto have_info;
4261                 }
4262
4263                 n = rb_next(&info->offset_index);
4264                 while (n) {
4265                         tmp = rb_entry(n, struct btrfs_free_space,
4266                                        offset_index);
4267                         if (offset + bytes < tmp->offset)
4268                                 break;
4269                         if (tmp->offset + tmp->bytes < offset) {
4270                                 n = rb_next(&tmp->offset_index);
4271                                 continue;
4272                         }
4273                         info = tmp;
4274                         goto have_info;
4275                 }
4276
4277                 ret = 0;
4278                 goto out;
4279         }
4280
4281         if (info->offset == offset) {
4282                 ret = 1;
4283                 goto out;
4284         }
4285
4286         if (offset > info->offset && offset < info->offset + info->bytes)
4287                 ret = 1;
4288 out:
4289         spin_unlock(&ctl->tree_lock);
4290         return ret;
4291 }
4292 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */