fs/btrfs/free-space-cache.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Copyright (C) 2008 Red Hat.  All rights reserved.
   4  */
   5
   6 #include <linux/pagemap.h>
   7 #include <linux/sched.h>
   8 #include <linux/sched/signal.h>
   9 #include <linux/slab.h>
  10 #include <linux/math64.h>
  11 #include <linux/ratelimit.h>
  12 #include <linux/error-injection.h>
  13 #include <linux/sched/mm.h>
  14 #include "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 struct inode *__lookup_free_space_inode(struct btrfs_root *root,
  52                                                struct btrfs_path *path,
  53                                                u64 offset)
  54 {
  55         struct btrfs_fs_info *fs_info = root->fs_info;
  56         struct btrfs_key key;
  57         struct btrfs_key location;
  58         struct btrfs_disk_key disk_key;
  59         struct btrfs_free_space_header *header;
  60         struct extent_buffer *leaf;
  61         struct inode *inode = NULL;
  62         unsigned nofs_flag;
  63         int ret;
  64
  65         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
  66         key.offset = offset;
  67         key.type = 0;
  68
  69         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
  70         if (ret < 0)
  71                 return ERR_PTR(ret);
  72         if (ret > 0) {
  73                 btrfs_release_path(path);
  74                 return ERR_PTR(-ENOENT);
  75         }
  76
  77         leaf = path->nodes[0];
  78         header = btrfs_item_ptr(leaf, path->slots[0],
  79                                 struct btrfs_free_space_header);
  80         btrfs_free_space_key(leaf, header, &disk_key);
  81         btrfs_disk_key_to_cpu(&location, &disk_key);
  82         btrfs_release_path(path);
  83
  84         /*
  85          * We are often under a trans handle at this point, so we need to make
  86          * sure NOFS is set to keep us from deadlocking.
  87          */
  88         nofs_flag = memalloc_nofs_save();
  89         inode = btrfs_iget_path(fs_info->sb, location.objectid, root, path);
  90         btrfs_release_path(path);
  91         memalloc_nofs_restore(nofs_flag);
  92         if (IS_ERR(inode))
  93                 return inode;
  94
  95         mapping_set_gfp_mask(inode->i_mapping,
  96                         mapping_gfp_constraint(inode->i_mapping,
  97                         ~(__GFP_FS | __GFP_HIGHMEM)));
  98
  99         return inode;
 100 }
 101
 102 struct inode *lookup_free_space_inode(struct btrfs_block_group *block_group,
 103                 struct btrfs_path *path)
 104 {
 105         struct btrfs_fs_info *fs_info = block_group->fs_info;
 106         struct inode *inode = NULL;
 107         u32 flags = BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 108
 109         spin_lock(&block_group->lock);
 110         if (block_group->inode)
 111                 inode = igrab(block_group->inode);
 112         spin_unlock(&block_group->lock);
 113         if (inode)
 114                 return inode;
 115
 116         inode = __lookup_free_space_inode(fs_info->tree_root, path,
 117                                           block_group->start);
 118         if (IS_ERR(inode))
 119                 return inode;
 120
 121         spin_lock(&block_group->lock);
 122         if (!((BTRFS_I(inode)->flags & flags) == flags)) {
 123                 btrfs_info(fs_info, "Old style space inode found, converting.");
 124                 BTRFS_I(inode)->flags |= BTRFS_INODE_NODATASUM |
 125                         BTRFS_INODE_NODATACOW;
 126                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 127         }
 128
 129         if (!block_group->iref) {
 130                 block_group->inode = igrab(inode);
 131                 block_group->iref = 1;
 132         }
 133         spin_unlock(&block_group->lock);
 134
 135         return inode;
 136 }
 137
 138 static int __create_free_space_inode(struct btrfs_root *root,
 139                                      struct btrfs_trans_handle *trans,
 140                                      struct btrfs_path *path,
 141                                      u64 ino, u64 offset)
 142 {
 143         struct btrfs_key key;
 144         struct btrfs_disk_key disk_key;
 145         struct btrfs_free_space_header *header;
 146         struct btrfs_inode_item *inode_item;
 147         struct extent_buffer *leaf;
 148         /* We inline CRCs for the free disk space cache */
 149         const u64 flags = BTRFS_INODE_NOCOMPRESS | BTRFS_INODE_PREALLOC |
 150                           BTRFS_INODE_NODATASUM | BTRFS_INODE_NODATACOW;
 151         int ret;
 152
 153         ret = btrfs_insert_empty_inode(trans, root, path, ino);
 154         if (ret)
 155                 return ret;
 156
 157         leaf = path->nodes[0];
 158         inode_item = btrfs_item_ptr(leaf, path->slots[0],
 159                                     struct btrfs_inode_item);
 160         btrfs_item_key(leaf, &disk_key, path->slots[0]);
 161         memzero_extent_buffer(leaf, (unsigned long)inode_item,
 162                              sizeof(*inode_item));
 163         btrfs_set_inode_generation(leaf, inode_item, trans->transid);
 164         btrfs_set_inode_size(leaf, inode_item, 0);
 165         btrfs_set_inode_nbytes(leaf, inode_item, 0);
 166         btrfs_set_inode_uid(leaf, inode_item, 0);
 167         btrfs_set_inode_gid(leaf, inode_item, 0);
 168         btrfs_set_inode_mode(leaf, inode_item, S_IFREG | 0600);
 169         btrfs_set_inode_flags(leaf, inode_item, flags);
 170         btrfs_set_inode_nlink(leaf, inode_item, 1);
 171         btrfs_set_inode_transid(leaf, inode_item, trans->transid);
 172         btrfs_set_inode_block_group(leaf, inode_item, offset);
 173         btrfs_mark_buffer_dirty(leaf);
 174         btrfs_release_path(path);
 175
 176         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 177         key.offset = offset;
 178         key.type = 0;
 179         ret = btrfs_insert_empty_item(trans, root, path, &key,
 180                                       sizeof(struct btrfs_free_space_header));
 181         if (ret < 0) {
 182                 btrfs_release_path(path);
 183                 return ret;
 184         }
 185
 186         leaf = path->nodes[0];
 187         header = btrfs_item_ptr(leaf, path->slots[0],
 188                                 struct btrfs_free_space_header);
 189         memzero_extent_buffer(leaf, (unsigned long)header, sizeof(*header));
 190         btrfs_set_free_space_key(leaf, header, &disk_key);
 191         btrfs_mark_buffer_dirty(leaf);
 192         btrfs_release_path(path);
 193
 194         return 0;
 195 }
 196
 197 int create_free_space_inode(struct btrfs_trans_handle *trans,
 198                             struct btrfs_block_group *block_group,
 199                             struct btrfs_path *path)
 200 {
 201         int ret;
 202         u64 ino;
 203
 204         ret = btrfs_get_free_objectid(trans->fs_info->tree_root, &ino);
 205         if (ret < 0)
 206                 return ret;
 207
 208         return __create_free_space_inode(trans->fs_info->tree_root, trans, path,
 209                                          ino, block_group->start);
 210 }
 211
 212 /*
 213  * inode is an optional sink: if it is NULL, btrfs_remove_free_space_inode
 214  * handles lookup, otherwise it takes ownership and iputs the inode.
 215  * Don't reuse an inode pointer after passing it into this function.
 216  */
 217 int btrfs_remove_free_space_inode(struct btrfs_trans_handle *trans,
 218                                   struct inode *inode,
 219                                   struct btrfs_block_group *block_group)
 220 {
 221         struct btrfs_path *path;
 222         struct btrfs_key key;
 223         int ret = 0;
 224
 225         path = btrfs_alloc_path();
 226         if (!path)
 227                 return -ENOMEM;
 228
 229         if (!inode)
 230                 inode = lookup_free_space_inode(block_group, path);
 231         if (IS_ERR(inode)) {
 232                 if (PTR_ERR(inode) != -ENOENT)
 233                         ret = PTR_ERR(inode);
 234                 goto out;
 235         }
 236         ret = btrfs_orphan_add(trans, BTRFS_I(inode));
 237         if (ret) {
 238                 btrfs_add_delayed_iput(inode);
 239                 goto out;
 240         }
 241         clear_nlink(inode);
 242         /* One for the block groups ref */
 243         spin_lock(&block_group->lock);
 244         if (block_group->iref) {
 245                 block_group->iref = 0;
 246                 block_group->inode = NULL;
 247                 spin_unlock(&block_group->lock);
 248                 iput(inode);
 249         } else {
 250                 spin_unlock(&block_group->lock);
 251         }
 252         /* One for the lookup ref */
 253         btrfs_add_delayed_iput(inode);
 254
 255         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 256         key.type = 0;
 257         key.offset = block_group->start;
 258         ret = btrfs_search_slot(trans, trans->fs_info->tree_root, &key, path,
 259                                 -1, 1);
 260         if (ret) {
 261                 if (ret > 0)
 262                         ret = 0;
 263                 goto out;
 264         }
 265         ret = btrfs_del_item(trans, trans->fs_info->tree_root, path);
 266 out:
 267         btrfs_free_path(path);
 268         return ret;
 269 }
 270
 271 int btrfs_check_trunc_cache_free_space(struct btrfs_fs_info *fs_info,
 272                                        struct btrfs_block_rsv *rsv)
 273 {
 274         u64 needed_bytes;
 275         int ret;
 276
 277         /* 1 for slack space, 1 for updating the inode */
 278         needed_bytes = btrfs_calc_insert_metadata_size(fs_info, 1) +
 279                 btrfs_calc_metadata_size(fs_info, 1);
 280
 281         spin_lock(&rsv->lock);
 282         if (rsv->reserved < needed_bytes)
 283                 ret = -ENOSPC;
 284         else
 285                 ret = 0;
 286         spin_unlock(&rsv->lock);
 287         return ret;
 288 }
 289
 290 int btrfs_truncate_free_space_cache(struct btrfs_trans_handle *trans,
 291                                     struct btrfs_block_group *block_group,
 292                                     struct inode *vfs_inode)
 293 {
 294         struct btrfs_truncate_control control = {
 295                 .inode = BTRFS_I(vfs_inode),
 296                 .new_size = 0,
 297                 .ino = btrfs_ino(BTRFS_I(vfs_inode)),
 298                 .min_type = BTRFS_EXTENT_DATA_KEY,
 299                 .clear_extent_range = true,
 300         };
 301         struct btrfs_inode *inode = BTRFS_I(vfs_inode);
 302         struct btrfs_root *root = inode->root;
 303         struct extent_state *cached_state = NULL;
 304         int ret = 0;
 305         bool locked = false;
 306
 307         if (block_group) {
 308                 struct btrfs_path *path = btrfs_alloc_path();
 309
 310                 if (!path) {
 311                         ret = -ENOMEM;
 312                         goto fail;
 313                 }
 314                 locked = true;
 315                 mutex_lock(&trans->transaction->cache_write_mutex);
 316                 if (!list_empty(&block_group->io_list)) {
 317                         list_del_init(&block_group->io_list);
 318
 319                         btrfs_wait_cache_io(trans, block_group, path);
 320                         btrfs_put_block_group(block_group);
 321                 }
 322
 323                 /*
 324                  * now that we've truncated the cache away, its no longer
 325                  * setup or written
 326                  */
 327                 spin_lock(&block_group->lock);
 328                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
 329                 spin_unlock(&block_group->lock);
 330                 btrfs_free_path(path);
 331         }
 332
 333         btrfs_i_size_write(inode, 0);
 334         truncate_pagecache(vfs_inode, 0);
 335
 336         lock_extent_bits(&inode->io_tree, 0, (u64)-1, &cached_state);
 337         btrfs_drop_extent_cache(inode, 0, (u64)-1, 0);
 338
 339         /*
 340          * We skip the throttling logic for free space cache inodes, so we don't
 341          * need to check for -EAGAIN.
 342          */
 343         ret = btrfs_truncate_inode_items(trans, root, &control);
 344
 345         inode_sub_bytes(&inode->vfs_inode, control.sub_bytes);
 346         btrfs_inode_safe_disk_i_size_write(inode, control.last_size);
 347
 348         unlock_extent_cached(&inode->io_tree, 0, (u64)-1, &cached_state);
 349         if (ret)
 350                 goto fail;
 351
 352         ret = btrfs_update_inode(trans, root, inode);
 353
 354 fail:
 355         if (locked)
 356                 mutex_unlock(&trans->transaction->cache_write_mutex);
 357         if (ret)
 358                 btrfs_abort_transaction(trans, ret);
 359
 360         return ret;
 361 }
 362
 363 static void readahead_cache(struct inode *inode)
 364 {
 365         struct file_ra_state ra;
 366         unsigned long last_index;
 367
 368         file_ra_state_init(&ra, inode->i_mapping);
 369         last_index = (i_size_read(inode) - 1) >> PAGE_SHIFT;
 370
 371         page_cache_sync_readahead(inode->i_mapping, &ra, NULL, 0, last_index);
 372 }
 373
 374 static int io_ctl_init(struct btrfs_io_ctl *io_ctl, struct inode *inode,
 375                        int write)
 376 {
 377         int num_pages;
 378
 379         num_pages = DIV_ROUND_UP(i_size_read(inode), PAGE_SIZE);
 380
 381         /* Make sure we can fit our crcs and generation into the first page */
 382         if (write && (num_pages * sizeof(u32) + sizeof(u64)) > PAGE_SIZE)
 383                 return -ENOSPC;
 384
 385         memset(io_ctl, 0, sizeof(struct btrfs_io_ctl));
 386
 387         io_ctl->pages = kcalloc(num_pages, sizeof(struct page *), GFP_NOFS);
 388         if (!io_ctl->pages)
 389                 return -ENOMEM;
 390
 391         io_ctl->num_pages = num_pages;
 392         io_ctl->fs_info = btrfs_sb(inode->i_sb);
 393         io_ctl->inode = inode;
 394
 395         return 0;
 396 }
 397 ALLOW_ERROR_INJECTION(io_ctl_init, ERRNO);
 398
 399 static void io_ctl_free(struct btrfs_io_ctl *io_ctl)
 400 {
 401         kfree(io_ctl->pages);
 402         io_ctl->pages = NULL;
 403 }
 404
 405 static void io_ctl_unmap_page(struct btrfs_io_ctl *io_ctl)
 406 {
 407         if (io_ctl->cur) {
 408                 io_ctl->cur = NULL;
 409                 io_ctl->orig = NULL;
 410         }
 411 }
 412
 413 static void io_ctl_map_page(struct btrfs_io_ctl *io_ctl, int clear)
 414 {
 415         ASSERT(io_ctl->index < io_ctl->num_pages);
 416         io_ctl->page = io_ctl->pages[io_ctl->index++];
 417         io_ctl->cur = page_address(io_ctl->page);
 418         io_ctl->orig = io_ctl->cur;
 419         io_ctl->size = PAGE_SIZE;
 420         if (clear)
 421                 clear_page(io_ctl->cur);
 422 }
 423
 424 static void io_ctl_drop_pages(struct btrfs_io_ctl *io_ctl)
 425 {
 426         int i;
 427
 428         io_ctl_unmap_page(io_ctl);
 429
 430         for (i = 0; i < io_ctl->num_pages; i++) {
 431                 if (io_ctl->pages[i]) {
 432                         btrfs_page_clear_checked(io_ctl->fs_info,
 433                                         io_ctl->pages[i],
 434                                         page_offset(io_ctl->pages[i]),
 435                                         PAGE_SIZE);
 436                         unlock_page(io_ctl->pages[i]);
 437                         put_page(io_ctl->pages[i]);
 438                 }
 439         }
 440 }
 441
 442 static int io_ctl_prepare_pages(struct btrfs_io_ctl *io_ctl, bool uptodate)
 443 {
 444         struct page *page;
 445         struct inode *inode = io_ctl->inode;
 446         gfp_t mask = btrfs_alloc_write_mask(inode->i_mapping);
 447         int i;
 448
 449         for (i = 0; i < io_ctl->num_pages; i++) {
 450                 int ret;
 451
 452                 page = find_or_create_page(inode->i_mapping, i, mask);
 453                 if (!page) {
 454                         io_ctl_drop_pages(io_ctl);
 455                         return -ENOMEM;
 456                 }
 457
 458                 ret = set_page_extent_mapped(page);
 459                 if (ret < 0) {
 460                         unlock_page(page);
 461                         put_page(page);
 462                         io_ctl_drop_pages(io_ctl);
 463                         return ret;
 464                 }
 465
 466                 io_ctl->pages[i] = page;
 467                 if (uptodate && !PageUptodate(page)) {
 468                         btrfs_read_folio(NULL, page_folio(page));
 469                         lock_page(page);
 470                         if (page->mapping != inode->i_mapping) {
 471                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 472                                           "free space cache page truncated");
 473                                 io_ctl_drop_pages(io_ctl);
 474                                 return -EIO;
 475                         }
 476                         if (!PageUptodate(page)) {
 477                                 btrfs_err(BTRFS_I(inode)->root->fs_info,
 478                                            "error reading free space cache");
 479                                 io_ctl_drop_pages(io_ctl);
 480                                 return -EIO;
 481                         }
 482                 }
 483         }
 484
 485         for (i = 0; i < io_ctl->num_pages; i++)
 486                 clear_page_dirty_for_io(io_ctl->pages[i]);
 487
 488         return 0;
 489 }
 490
 491 static void io_ctl_set_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 492 {
 493         io_ctl_map_page(io_ctl, 1);
 494
 495         /*
 496          * Skip the csum areas.  If we don't check crcs then we just have a
 497          * 64bit chunk at the front of the first page.
 498          */
 499         io_ctl->cur += (sizeof(u32) * io_ctl->num_pages);
 500         io_ctl->size -= sizeof(u64) + (sizeof(u32) * io_ctl->num_pages);
 501
 502         put_unaligned_le64(generation, io_ctl->cur);
 503         io_ctl->cur += sizeof(u64);
 504 }
 505
 506 static int io_ctl_check_generation(struct btrfs_io_ctl *io_ctl, u64 generation)
 507 {
 508         u64 cache_gen;
 509
 510         /*
 511          * Skip the crc area.  If we don't check crcs then we just have a 64bit
 512          * 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         cache_gen = get_unaligned_le64(io_ctl->cur);
 518         if (cache_gen != generation) {
 519                 btrfs_err_rl(io_ctl->fs_info,
 520                         "space cache generation (%llu) does not match inode (%llu)",
 521                                 cache_gen, generation);
 522                 io_ctl_unmap_page(io_ctl);
 523                 return -EIO;
 524         }
 525         io_ctl->cur += sizeof(u64);
 526         return 0;
 527 }
 528
 529 static void io_ctl_set_crc(struct btrfs_io_ctl *io_ctl, int index)
 530 {
 531         u32 *tmp;
 532         u32 crc = ~(u32)0;
 533         unsigned offset = 0;
 534
 535         if (index == 0)
 536                 offset = sizeof(u32) * io_ctl->num_pages;
 537
 538         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 539         btrfs_crc32c_final(crc, (u8 *)&crc);
 540         io_ctl_unmap_page(io_ctl);
 541         tmp = page_address(io_ctl->pages[0]);
 542         tmp += index;
 543         *tmp = crc;
 544 }
 545
 546 static int io_ctl_check_crc(struct btrfs_io_ctl *io_ctl, int index)
 547 {
 548         u32 *tmp, val;
 549         u32 crc = ~(u32)0;
 550         unsigned offset = 0;
 551
 552         if (index == 0)
 553                 offset = sizeof(u32) * io_ctl->num_pages;
 554
 555         tmp = page_address(io_ctl->pages[0]);
 556         tmp += index;
 557         val = *tmp;
 558
 559         io_ctl_map_page(io_ctl, 0);
 560         crc = btrfs_crc32c(crc, io_ctl->orig + offset, PAGE_SIZE - offset);
 561         btrfs_crc32c_final(crc, (u8 *)&crc);
 562         if (val != crc) {
 563                 btrfs_err_rl(io_ctl->fs_info,
 564                         "csum mismatch on free space cache");
 565                 io_ctl_unmap_page(io_ctl);
 566                 return -EIO;
 567         }
 568
 569         return 0;
 570 }
 571
 572 static int io_ctl_add_entry(struct btrfs_io_ctl *io_ctl, u64 offset, u64 bytes,
 573                             void *bitmap)
 574 {
 575         struct btrfs_free_space_entry *entry;
 576
 577         if (!io_ctl->cur)
 578                 return -ENOSPC;
 579
 580         entry = io_ctl->cur;
 581         put_unaligned_le64(offset, &entry->offset);
 582         put_unaligned_le64(bytes, &entry->bytes);
 583         entry->type = (bitmap) ? BTRFS_FREE_SPACE_BITMAP :
 584                 BTRFS_FREE_SPACE_EXTENT;
 585         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 586         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 587
 588         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 589                 return 0;
 590
 591         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 592
 593         /* No more pages to map */
 594         if (io_ctl->index >= io_ctl->num_pages)
 595                 return 0;
 596
 597         /* map the next page */
 598         io_ctl_map_page(io_ctl, 1);
 599         return 0;
 600 }
 601
 602 static int io_ctl_add_bitmap(struct btrfs_io_ctl *io_ctl, void *bitmap)
 603 {
 604         if (!io_ctl->cur)
 605                 return -ENOSPC;
 606
 607         /*
 608          * If we aren't at the start of the current page, unmap this one and
 609          * map the next one if there is any left.
 610          */
 611         if (io_ctl->cur != io_ctl->orig) {
 612                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 613                 if (io_ctl->index >= io_ctl->num_pages)
 614                         return -ENOSPC;
 615                 io_ctl_map_page(io_ctl, 0);
 616         }
 617
 618         copy_page(io_ctl->cur, bitmap);
 619         io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 620         if (io_ctl->index < io_ctl->num_pages)
 621                 io_ctl_map_page(io_ctl, 0);
 622         return 0;
 623 }
 624
 625 static void io_ctl_zero_remaining_pages(struct btrfs_io_ctl *io_ctl)
 626 {
 627         /*
 628          * If we're not on the boundary we know we've modified the page and we
 629          * need to crc the page.
 630          */
 631         if (io_ctl->cur != io_ctl->orig)
 632                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 633         else
 634                 io_ctl_unmap_page(io_ctl);
 635
 636         while (io_ctl->index < io_ctl->num_pages) {
 637                 io_ctl_map_page(io_ctl, 1);
 638                 io_ctl_set_crc(io_ctl, io_ctl->index - 1);
 639         }
 640 }
 641
 642 static int io_ctl_read_entry(struct btrfs_io_ctl *io_ctl,
 643                             struct btrfs_free_space *entry, u8 *type)
 644 {
 645         struct btrfs_free_space_entry *e;
 646         int ret;
 647
 648         if (!io_ctl->cur) {
 649                 ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 650                 if (ret)
 651                         return ret;
 652         }
 653
 654         e = io_ctl->cur;
 655         entry->offset = get_unaligned_le64(&e->offset);
 656         entry->bytes = get_unaligned_le64(&e->bytes);
 657         *type = e->type;
 658         io_ctl->cur += sizeof(struct btrfs_free_space_entry);
 659         io_ctl->size -= sizeof(struct btrfs_free_space_entry);
 660
 661         if (io_ctl->size >= sizeof(struct btrfs_free_space_entry))
 662                 return 0;
 663
 664         io_ctl_unmap_page(io_ctl);
 665
 666         return 0;
 667 }
 668
 669 static int io_ctl_read_bitmap(struct btrfs_io_ctl *io_ctl,
 670                               struct btrfs_free_space *entry)
 671 {
 672         int ret;
 673
 674         ret = io_ctl_check_crc(io_ctl, io_ctl->index);
 675         if (ret)
 676                 return ret;
 677
 678         copy_page(entry->bitmap, io_ctl->cur);
 679         io_ctl_unmap_page(io_ctl);
 680
 681         return 0;
 682 }
 683
 684 static void recalculate_thresholds(struct btrfs_free_space_ctl *ctl)
 685 {
 686         struct btrfs_block_group *block_group = ctl->block_group;
 687         u64 max_bytes;
 688         u64 bitmap_bytes;
 689         u64 extent_bytes;
 690         u64 size = block_group->length;
 691         u64 bytes_per_bg = BITS_PER_BITMAP * ctl->unit;
 692         u64 max_bitmaps = div64_u64(size + bytes_per_bg - 1, bytes_per_bg);
 693
 694         max_bitmaps = max_t(u64, max_bitmaps, 1);
 695
 696         ASSERT(ctl->total_bitmaps <= max_bitmaps);
 697
 698         /*
 699          * We are trying to keep the total amount of memory used per 1GiB of
 700          * space to be MAX_CACHE_BYTES_PER_GIG.  However, with a reclamation
 701          * mechanism of pulling extents >= FORCE_EXTENT_THRESHOLD out of
 702          * bitmaps, we may end up using more memory than this.
 703          */
 704         if (size < SZ_1G)
 705                 max_bytes = MAX_CACHE_BYTES_PER_GIG;
 706         else
 707                 max_bytes = MAX_CACHE_BYTES_PER_GIG * div_u64(size, SZ_1G);
 708
 709         bitmap_bytes = ctl->total_bitmaps * ctl->unit;
 710
 711         /*
 712          * we want the extent entry threshold to always be at most 1/2 the max
 713          * bytes we can have, or whatever is less than that.
 714          */
 715         extent_bytes = max_bytes - bitmap_bytes;
 716         extent_bytes = min_t(u64, extent_bytes, max_bytes >> 1);
 717
 718         ctl->extents_thresh =
 719                 div_u64(extent_bytes, sizeof(struct btrfs_free_space));
 720 }
 721
 722 static int __load_free_space_cache(struct btrfs_root *root, struct inode *inode,
 723                                    struct btrfs_free_space_ctl *ctl,
 724                                    struct btrfs_path *path, u64 offset)
 725 {
 726         struct btrfs_fs_info *fs_info = root->fs_info;
 727         struct btrfs_free_space_header *header;
 728         struct extent_buffer *leaf;
 729         struct btrfs_io_ctl io_ctl;
 730         struct btrfs_key key;
 731         struct btrfs_free_space *e, *n;
 732         LIST_HEAD(bitmaps);
 733         u64 num_entries;
 734         u64 num_bitmaps;
 735         u64 generation;
 736         u8 type;
 737         int ret = 0;
 738
 739         /* Nothing in the space cache, goodbye */
 740         if (!i_size_read(inode))
 741                 return 0;
 742
 743         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
 744         key.offset = offset;
 745         key.type = 0;
 746
 747         ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
 748         if (ret < 0)
 749                 return 0;
 750         else if (ret > 0) {
 751                 btrfs_release_path(path);
 752                 return 0;
 753         }
 754
 755         ret = -1;
 756
 757         leaf = path->nodes[0];
 758         header = btrfs_item_ptr(leaf, path->slots[0],
 759                                 struct btrfs_free_space_header);
 760         num_entries = btrfs_free_space_entries(leaf, header);
 761         num_bitmaps = btrfs_free_space_bitmaps(leaf, header);
 762         generation = btrfs_free_space_generation(leaf, header);
 763         btrfs_release_path(path);
 764
 765         if (!BTRFS_I(inode)->generation) {
 766                 btrfs_info(fs_info,
 767                            "the free space cache file (%llu) is invalid, skip it",
 768                            offset);
 769                 return 0;
 770         }
 771
 772         if (BTRFS_I(inode)->generation != generation) {
 773                 btrfs_err(fs_info,
 774                           "free space inode generation (%llu) did not match free space cache generation (%llu)",
 775                           BTRFS_I(inode)->generation, generation);
 776                 return 0;
 777         }
 778
 779         if (!num_entries)
 780                 return 0;
 781
 782         ret = io_ctl_init(&io_ctl, inode, 0);
 783         if (ret)
 784                 return ret;
 785
 786         readahead_cache(inode);
 787
 788         ret = io_ctl_prepare_pages(&io_ctl, true);
 789         if (ret)
 790                 goto out;
 791
 792         ret = io_ctl_check_crc(&io_ctl, 0);
 793         if (ret)
 794                 goto free_cache;
 795
 796         ret = io_ctl_check_generation(&io_ctl, generation);
 797         if (ret)
 798                 goto free_cache;
 799
 800         while (num_entries) {
 801                 e = kmem_cache_zalloc(btrfs_free_space_cachep,
 802                                       GFP_NOFS);
 803                 if (!e) {
 804                         ret = -ENOMEM;
 805                         goto free_cache;
 806                 }
 807
 808                 ret = io_ctl_read_entry(&io_ctl, e, &type);
 809                 if (ret) {
 810                         kmem_cache_free(btrfs_free_space_cachep, e);
 811                         goto free_cache;
 812                 }
 813
 814                 if (!e->bytes) {
 815                         ret = -1;
 816                         kmem_cache_free(btrfs_free_space_cachep, e);
 817                         goto free_cache;
 818                 }
 819
 820                 if (type == BTRFS_FREE_SPACE_EXTENT) {
 821                         spin_lock(&ctl->tree_lock);
 822                         ret = link_free_space(ctl, e);
 823                         spin_unlock(&ctl->tree_lock);
 824                         if (ret) {
 825                                 btrfs_err(fs_info,
 826                                         "Duplicate entries in free space cache, dumping");
 827                                 kmem_cache_free(btrfs_free_space_cachep, e);
 828                                 goto free_cache;
 829                         }
 830                 } else {
 831                         ASSERT(num_bitmaps);
 832                         num_bitmaps--;
 833                         e->bitmap = kmem_cache_zalloc(
 834                                         btrfs_free_space_bitmap_cachep, GFP_NOFS);
 835                         if (!e->bitmap) {
 836                                 ret = -ENOMEM;
 837                                 kmem_cache_free(
 838                                         btrfs_free_space_cachep, e);
 839                                 goto free_cache;
 840                         }
 841                         spin_lock(&ctl->tree_lock);
 842                         ret = link_free_space(ctl, e);
 843                         ctl->total_bitmaps++;
 844                         recalculate_thresholds(ctl);
 845                         spin_unlock(&ctl->tree_lock);
 846                         if (ret) {
 847                                 btrfs_err(fs_info,
 848                                         "Duplicate entries in free space cache, dumping");
 849                                 kmem_cache_free(btrfs_free_space_cachep, e);
 850                                 goto free_cache;
 851                         }
 852                         list_add_tail(&e->list, &bitmaps);
 853                 }
 854
 855                 num_entries--;
 856         }
 857
 858         io_ctl_unmap_page(&io_ctl);
 859
 860         /*
 861          * We add the bitmaps at the end of the entries in order that
 862          * the bitmap entries are added to the cache.
 863          */
 864         list_for_each_entry_safe(e, n, &bitmaps, list) {
 865                 list_del_init(&e->list);
 866                 ret = io_ctl_read_bitmap(&io_ctl, e);
 867                 if (ret)
 868                         goto free_cache;
 869         }
 870
 871         io_ctl_drop_pages(&io_ctl);
 872         ret = 1;
 873 out:
 874         io_ctl_free(&io_ctl);
 875         return ret;
 876 free_cache:
 877         io_ctl_drop_pages(&io_ctl);
 878         __btrfs_remove_free_space_cache(ctl);
 879         goto out;
 880 }
 881
 882 static int copy_free_space_cache(struct btrfs_block_group *block_group,
 883                                  struct btrfs_free_space_ctl *ctl)
 884 {
 885         struct btrfs_free_space *info;
 886         struct rb_node *n;
 887         int ret = 0;
 888
 889         while (!ret && (n = rb_first(&ctl->free_space_offset)) != NULL) {
 890                 info = rb_entry(n, struct btrfs_free_space, offset_index);
 891                 if (!info->bitmap) {
 892                         unlink_free_space(ctl, info, true);
 893                         ret = btrfs_add_free_space(block_group, info->offset,
 894                                                    info->bytes);
 895                         kmem_cache_free(btrfs_free_space_cachep, info);
 896                 } else {
 897                         u64 offset = info->offset;
 898                         u64 bytes = ctl->unit;
 899
 900                         while (search_bitmap(ctl, info, &offset, &bytes,
 901                                              false) == 0) {
 902                                 ret = btrfs_add_free_space(block_group, offset,
 903                                                            bytes);
 904                                 if (ret)
 905                                         break;
 906                                 bitmap_clear_bits(ctl, info, offset, bytes, true);
 907                                 offset = info->offset;
 908                                 bytes = ctl->unit;
 909                         }
 910                         free_bitmap(ctl, info);
 911                 }
 912                 cond_resched();
 913         }
 914         return ret;
 915 }
 916
 917 int load_free_space_cache(struct btrfs_block_group *block_group)
 918 {
 919         struct btrfs_fs_info *fs_info = block_group->fs_info;
 920         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
 921         struct btrfs_free_space_ctl tmp_ctl = {};
 922         struct inode *inode;
 923         struct btrfs_path *path;
 924         int ret = 0;
 925         bool matched;
 926         u64 used = block_group->used;
 927
 928         /*
 929          * Because we could potentially discard our loaded free space, we want
 930          * to load everything into a temporary structure first, and then if it's
 931          * valid copy it all into the actual free space ctl.
 932          */
 933         btrfs_init_free_space_ctl(block_group, &tmp_ctl);
 934
 935         /*
 936          * If this block group has been marked to be cleared for one reason or
 937          * another then we can't trust the on disk cache, so just return.
 938          */
 939         spin_lock(&block_group->lock);
 940         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 941                 spin_unlock(&block_group->lock);
 942                 return 0;
 943         }
 944         spin_unlock(&block_group->lock);
 945
 946         path = btrfs_alloc_path();
 947         if (!path)
 948                 return 0;
 949         path->search_commit_root = 1;
 950         path->skip_locking = 1;
 951
 952         /*
 953          * We must pass a path with search_commit_root set to btrfs_iget in
 954          * order to avoid a deadlock when allocating extents for the tree root.
 955          *
 956          * When we are COWing an extent buffer from the tree root, when looking
 957          * for a free extent, at extent-tree.c:find_free_extent(), we can find
 958          * block group without its free space cache loaded. When we find one
 959          * we must load its space cache which requires reading its free space
 960          * cache's inode item from the root tree. If this inode item is located
 961          * in the same leaf that we started COWing before, then we end up in
 962          * deadlock on the extent buffer (trying to read lock it when we
 963          * previously write locked it).
 964          *
 965          * It's safe to read the inode item using the commit root because
 966          * block groups, once loaded, stay in memory forever (until they are
 967          * removed) as well as their space caches once loaded. New block groups
 968          * once created get their ->cached field set to BTRFS_CACHE_FINISHED so
 969          * we will never try to read their inode item while the fs is mounted.
 970          */
 971         inode = lookup_free_space_inode(block_group, path);
 972         if (IS_ERR(inode)) {
 973                 btrfs_free_path(path);
 974                 return 0;
 975         }
 976
 977         /* We may have converted the inode and made the cache invalid. */
 978         spin_lock(&block_group->lock);
 979         if (block_group->disk_cache_state != BTRFS_DC_WRITTEN) {
 980                 spin_unlock(&block_group->lock);
 981                 btrfs_free_path(path);
 982                 goto out;
 983         }
 984         spin_unlock(&block_group->lock);
 985
 986         ret = __load_free_space_cache(fs_info->tree_root, inode, &tmp_ctl,
 987                                       path, block_group->start);
 988         btrfs_free_path(path);
 989         if (ret <= 0)
 990                 goto out;
 991
 992         matched = (tmp_ctl.free_space == (block_group->length - used -
 993                                           block_group->bytes_super));
 994
 995         if (matched) {
 996                 ret = copy_free_space_cache(block_group, &tmp_ctl);
 997                 /*
 998                  * ret == 1 means we successfully loaded the free space cache,
 999                  * so we need to re-set it here.
1000                  */
1001                 if (ret == 0)
1002                         ret = 1;
1003         } else {
1004                 __btrfs_remove_free_space_cache(&tmp_ctl);
1005                 btrfs_warn(fs_info,
1006                            "block group %llu has wrong amount of free space",
1007                            block_group->start);
1008                 ret = -1;
1009         }
1010 out:
1011         if (ret < 0) {
1012                 /* This cache is bogus, make sure it gets cleared */
1013                 spin_lock(&block_group->lock);
1014                 block_group->disk_cache_state = BTRFS_DC_CLEAR;
1015                 spin_unlock(&block_group->lock);
1016                 ret = 0;
1017
1018                 btrfs_warn(fs_info,
1019                            "failed to load free space cache for block group %llu, rebuilding it now",
1020                            block_group->start);
1021         }
1022
1023         spin_lock(&ctl->tree_lock);
1024         btrfs_discard_update_discardable(block_group);
1025         spin_unlock(&ctl->tree_lock);
1026         iput(inode);
1027         return ret;
1028 }
1029
1030 static noinline_for_stack
1031 int write_cache_extent_entries(struct btrfs_io_ctl *io_ctl,
1032                               struct btrfs_free_space_ctl *ctl,
1033                               struct btrfs_block_group *block_group,
1034                               int *entries, int *bitmaps,
1035                               struct list_head *bitmap_list)
1036 {
1037         int ret;
1038         struct btrfs_free_cluster *cluster = NULL;
1039         struct btrfs_free_cluster *cluster_locked = NULL;
1040         struct rb_node *node = rb_first(&ctl->free_space_offset);
1041         struct btrfs_trim_range *trim_entry;
1042
1043         /* Get the cluster for this block_group if it exists */
1044         if (block_group && !list_empty(&block_group->cluster_list)) {
1045                 cluster = list_entry(block_group->cluster_list.next,
1046                                      struct btrfs_free_cluster,
1047                                      block_group_list);
1048         }
1049
1050         if (!node && cluster) {
1051                 cluster_locked = cluster;
1052                 spin_lock(&cluster_locked->lock);
1053                 node = rb_first(&cluster->root);
1054                 cluster = NULL;
1055         }
1056
1057         /* Write out the extent entries */
1058         while (node) {
1059                 struct btrfs_free_space *e;
1060
1061                 e = rb_entry(node, struct btrfs_free_space, offset_index);
1062                 *entries += 1;
1063
1064                 ret = io_ctl_add_entry(io_ctl, e->offset, e->bytes,
1065                                        e->bitmap);
1066                 if (ret)
1067                         goto fail;
1068
1069                 if (e->bitmap) {
1070                         list_add_tail(&e->list, bitmap_list);
1071                         *bitmaps += 1;
1072                 }
1073                 node = rb_next(node);
1074                 if (!node && cluster) {
1075                         node = rb_first(&cluster->root);
1076                         cluster_locked = cluster;
1077                         spin_lock(&cluster_locked->lock);
1078                         cluster = NULL;
1079                 }
1080         }
1081         if (cluster_locked) {
1082                 spin_unlock(&cluster_locked->lock);
1083                 cluster_locked = NULL;
1084         }
1085
1086         /*
1087          * Make sure we don't miss any range that was removed from our rbtree
1088          * because trimming is running. Otherwise after a umount+mount (or crash
1089          * after committing the transaction) we would leak free space and get
1090          * an inconsistent free space cache report from fsck.
1091          */
1092         list_for_each_entry(trim_entry, &ctl->trimming_ranges, list) {
1093                 ret = io_ctl_add_entry(io_ctl, trim_entry->start,
1094                                        trim_entry->bytes, NULL);
1095                 if (ret)
1096                         goto fail;
1097                 *entries += 1;
1098         }
1099
1100         return 0;
1101 fail:
1102         if (cluster_locked)
1103                 spin_unlock(&cluster_locked->lock);
1104         return -ENOSPC;
1105 }
1106
1107 static noinline_for_stack int
1108 update_cache_item(struct btrfs_trans_handle *trans,
1109                   struct btrfs_root *root,
1110                   struct inode *inode,
1111                   struct btrfs_path *path, u64 offset,
1112                   int entries, int bitmaps)
1113 {
1114         struct btrfs_key key;
1115         struct btrfs_free_space_header *header;
1116         struct extent_buffer *leaf;
1117         int ret;
1118
1119         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
1120         key.offset = offset;
1121         key.type = 0;
1122
1123         ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
1124         if (ret < 0) {
1125                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1126                                  EXTENT_DELALLOC, 0, 0, NULL);
1127                 goto fail;
1128         }
1129         leaf = path->nodes[0];
1130         if (ret > 0) {
1131                 struct btrfs_key found_key;
1132                 ASSERT(path->slots[0]);
1133                 path->slots[0]--;
1134                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
1135                 if (found_key.objectid != BTRFS_FREE_SPACE_OBJECTID ||
1136                     found_key.offset != offset) {
1137                         clear_extent_bit(&BTRFS_I(inode)->io_tree, 0,
1138                                          inode->i_size - 1, EXTENT_DELALLOC, 0,
1139                                          0, NULL);
1140                         btrfs_release_path(path);
1141                         goto fail;
1142                 }
1143         }
1144
1145         BTRFS_I(inode)->generation = trans->transid;
1146         header = btrfs_item_ptr(leaf, path->slots[0],
1147                                 struct btrfs_free_space_header);
1148         btrfs_set_free_space_entries(leaf, header, entries);
1149         btrfs_set_free_space_bitmaps(leaf, header, bitmaps);
1150         btrfs_set_free_space_generation(leaf, header, trans->transid);
1151         btrfs_mark_buffer_dirty(leaf);
1152         btrfs_release_path(path);
1153
1154         return 0;
1155
1156 fail:
1157         return -1;
1158 }
1159
1160 static noinline_for_stack int write_pinned_extent_entries(
1161                             struct btrfs_trans_handle *trans,
1162                             struct btrfs_block_group *block_group,
1163                             struct btrfs_io_ctl *io_ctl,
1164                             int *entries)
1165 {
1166         u64 start, extent_start, extent_end, len;
1167         struct extent_io_tree *unpin = NULL;
1168         int ret;
1169
1170         if (!block_group)
1171                 return 0;
1172
1173         /*
1174          * We want to add any pinned extents to our free space cache
1175          * so we don't leak the space
1176          *
1177          * We shouldn't have switched the pinned extents yet so this is the
1178          * right one
1179          */
1180         unpin = &trans->transaction->pinned_extents;
1181
1182         start = block_group->start;
1183
1184         while (start < block_group->start + block_group->length) {
1185                 ret = find_first_extent_bit(unpin, start,
1186                                             &extent_start, &extent_end,
1187                                             EXTENT_DIRTY, NULL);
1188                 if (ret)
1189                         return 0;
1190
1191                 /* This pinned extent is out of our range */
1192                 if (extent_start >= block_group->start + block_group->length)
1193                         return 0;
1194
1195                 extent_start = max(extent_start, start);
1196                 extent_end = min(block_group->start + block_group->length,
1197                                  extent_end + 1);
1198                 len = extent_end - extent_start;
1199
1200                 *entries += 1;
1201                 ret = io_ctl_add_entry(io_ctl, extent_start, len, NULL);
1202                 if (ret)
1203                         return -ENOSPC;
1204
1205                 start = extent_end;
1206         }
1207
1208         return 0;
1209 }
1210
1211 static noinline_for_stack int
1212 write_bitmap_entries(struct btrfs_io_ctl *io_ctl, struct list_head *bitmap_list)
1213 {
1214         struct btrfs_free_space *entry, *next;
1215         int ret;
1216
1217         /* Write out the bitmaps */
1218         list_for_each_entry_safe(entry, next, bitmap_list, list) {
1219                 ret = io_ctl_add_bitmap(io_ctl, entry->bitmap);
1220                 if (ret)
1221                         return -ENOSPC;
1222                 list_del_init(&entry->list);
1223         }
1224
1225         return 0;
1226 }
1227
1228 static int flush_dirty_cache(struct inode *inode)
1229 {
1230         int ret;
1231
1232         ret = btrfs_wait_ordered_range(inode, 0, (u64)-1);
1233         if (ret)
1234                 clear_extent_bit(&BTRFS_I(inode)->io_tree, 0, inode->i_size - 1,
1235                                  EXTENT_DELALLOC, 0, 0, NULL);
1236
1237         return ret;
1238 }
1239
1240 static void noinline_for_stack
1241 cleanup_bitmap_list(struct list_head *bitmap_list)
1242 {
1243         struct btrfs_free_space *entry, *next;
1244
1245         list_for_each_entry_safe(entry, next, bitmap_list, list)
1246                 list_del_init(&entry->list);
1247 }
1248
1249 static void noinline_for_stack
1250 cleanup_write_cache_enospc(struct inode *inode,
1251                            struct btrfs_io_ctl *io_ctl,
1252                            struct extent_state **cached_state)
1253 {
1254         io_ctl_drop_pages(io_ctl);
1255         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1256                              i_size_read(inode) - 1, cached_state);
1257 }
1258
1259 static int __btrfs_wait_cache_io(struct btrfs_root *root,
1260                                  struct btrfs_trans_handle *trans,
1261                                  struct btrfs_block_group *block_group,
1262                                  struct btrfs_io_ctl *io_ctl,
1263                                  struct btrfs_path *path, u64 offset)
1264 {
1265         int ret;
1266         struct inode *inode = io_ctl->inode;
1267
1268         if (!inode)
1269                 return 0;
1270
1271         /* Flush the dirty pages in the cache file. */
1272         ret = flush_dirty_cache(inode);
1273         if (ret)
1274                 goto out;
1275
1276         /* Update the cache item to tell everyone this cache file is valid. */
1277         ret = update_cache_item(trans, root, inode, path, offset,
1278                                 io_ctl->entries, io_ctl->bitmaps);
1279 out:
1280         if (ret) {
1281                 invalidate_inode_pages2(inode->i_mapping);
1282                 BTRFS_I(inode)->generation = 0;
1283                 if (block_group)
1284                         btrfs_debug(root->fs_info,
1285           "failed to write free space cache for block group %llu error %d",
1286                                   block_group->start, ret);
1287         }
1288         btrfs_update_inode(trans, root, BTRFS_I(inode));
1289
1290         if (block_group) {
1291                 /* the dirty list is protected by the dirty_bgs_lock */
1292                 spin_lock(&trans->transaction->dirty_bgs_lock);
1293
1294                 /* the disk_cache_state is protected by the block group lock */
1295                 spin_lock(&block_group->lock);
1296
1297                 /*
1298                  * only mark this as written if we didn't get put back on
1299                  * the dirty list while waiting for IO.   Otherwise our
1300                  * cache state won't be right, and we won't get written again
1301                  */
1302                 if (!ret && list_empty(&block_group->dirty_list))
1303                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1304                 else if (ret)
1305                         block_group->disk_cache_state = BTRFS_DC_ERROR;
1306
1307                 spin_unlock(&block_group->lock);
1308                 spin_unlock(&trans->transaction->dirty_bgs_lock);
1309                 io_ctl->inode = NULL;
1310                 iput(inode);
1311         }
1312
1313         return ret;
1314
1315 }
1316
1317 int btrfs_wait_cache_io(struct btrfs_trans_handle *trans,
1318                         struct btrfs_block_group *block_group,
1319                         struct btrfs_path *path)
1320 {
1321         return __btrfs_wait_cache_io(block_group->fs_info->tree_root, trans,
1322                                      block_group, &block_group->io_ctl,
1323                                      path, block_group->start);
1324 }
1325
1326 /**
1327  * Write out cached info to an inode
1328  *
1329  * @root:        root the inode belongs to
1330  * @inode:       freespace inode we are writing out
1331  * @ctl:         free space cache we are going to write out
1332  * @block_group: block_group for this cache if it belongs to a block_group
1333  * @io_ctl:      holds context for the io
1334  * @trans:       the trans handle
1335  *
1336  * This function writes out a free space cache struct to disk for quick recovery
1337  * on mount.  This will return 0 if it was successful in writing the cache out,
1338  * or an errno if it was not.
1339  */
1340 static int __btrfs_write_out_cache(struct btrfs_root *root, struct inode *inode,
1341                                    struct btrfs_free_space_ctl *ctl,
1342                                    struct btrfs_block_group *block_group,
1343                                    struct btrfs_io_ctl *io_ctl,
1344                                    struct btrfs_trans_handle *trans)
1345 {
1346         struct extent_state *cached_state = NULL;
1347         LIST_HEAD(bitmap_list);
1348         int entries = 0;
1349         int bitmaps = 0;
1350         int ret;
1351         int must_iput = 0;
1352
1353         if (!i_size_read(inode))
1354                 return -EIO;
1355
1356         WARN_ON(io_ctl->pages);
1357         ret = io_ctl_init(io_ctl, inode, 1);
1358         if (ret)
1359                 return ret;
1360
1361         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA)) {
1362                 down_write(&block_group->data_rwsem);
1363                 spin_lock(&block_group->lock);
1364                 if (block_group->delalloc_bytes) {
1365                         block_group->disk_cache_state = BTRFS_DC_WRITTEN;
1366                         spin_unlock(&block_group->lock);
1367                         up_write(&block_group->data_rwsem);
1368                         BTRFS_I(inode)->generation = 0;
1369                         ret = 0;
1370                         must_iput = 1;
1371                         goto out;
1372                 }
1373                 spin_unlock(&block_group->lock);
1374         }
1375
1376         /* Lock all pages first so we can lock the extent safely. */
1377         ret = io_ctl_prepare_pages(io_ctl, false);
1378         if (ret)
1379                 goto out_unlock;
1380
1381         lock_extent_bits(&BTRFS_I(inode)->io_tree, 0, i_size_read(inode) - 1,
1382                          &cached_state);
1383
1384         io_ctl_set_generation(io_ctl, trans->transid);
1385
1386         mutex_lock(&ctl->cache_writeout_mutex);
1387         /* Write out the extent entries in the free space cache */
1388         spin_lock(&ctl->tree_lock);
1389         ret = write_cache_extent_entries(io_ctl, ctl,
1390                                          block_group, &entries, &bitmaps,
1391                                          &bitmap_list);
1392         if (ret)
1393                 goto out_nospc_locked;
1394
1395         /*
1396          * Some spaces that are freed in the current transaction are pinned,
1397          * they will be added into free space cache after the transaction is
1398          * committed, we shouldn't lose them.
1399          *
1400          * If this changes while we are working we'll get added back to
1401          * the dirty list and redo it.  No locking needed
1402          */
1403         ret = write_pinned_extent_entries(trans, block_group, io_ctl, &entries);
1404         if (ret)
1405                 goto out_nospc_locked;
1406
1407         /*
1408          * At last, we write out all the bitmaps and keep cache_writeout_mutex
1409          * locked while doing it because a concurrent trim can be manipulating
1410          * or freeing the bitmap.
1411          */
1412         ret = write_bitmap_entries(io_ctl, &bitmap_list);
1413         spin_unlock(&ctl->tree_lock);
1414         mutex_unlock(&ctl->cache_writeout_mutex);
1415         if (ret)
1416                 goto out_nospc;
1417
1418         /* Zero out the rest of the pages just to make sure */
1419         io_ctl_zero_remaining_pages(io_ctl);
1420
1421         /* Everything is written out, now we dirty the pages in the file. */
1422         ret = btrfs_dirty_pages(BTRFS_I(inode), io_ctl->pages,
1423                                 io_ctl->num_pages, 0, i_size_read(inode),
1424                                 &cached_state, false);
1425         if (ret)
1426                 goto out_nospc;
1427
1428         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1429                 up_write(&block_group->data_rwsem);
1430         /*
1431          * Release the pages and unlock the extent, we will flush
1432          * them out later
1433          */
1434         io_ctl_drop_pages(io_ctl);
1435         io_ctl_free(io_ctl);
1436
1437         unlock_extent_cached(&BTRFS_I(inode)->io_tree, 0,
1438                              i_size_read(inode) - 1, &cached_state);
1439
1440         /*
1441          * at this point the pages are under IO and we're happy,
1442          * The caller is responsible for waiting on them and updating
1443          * the cache and the inode
1444          */
1445         io_ctl->entries = entries;
1446         io_ctl->bitmaps = bitmaps;
1447
1448         ret = btrfs_fdatawrite_range(inode, 0, (u64)-1);
1449         if (ret)
1450                 goto out;
1451
1452         return 0;
1453
1454 out_nospc_locked:
1455         cleanup_bitmap_list(&bitmap_list);
1456         spin_unlock(&ctl->tree_lock);
1457         mutex_unlock(&ctl->cache_writeout_mutex);
1458
1459 out_nospc:
1460         cleanup_write_cache_enospc(inode, io_ctl, &cached_state);
1461
1462 out_unlock:
1463         if (block_group && (block_group->flags & BTRFS_BLOCK_GROUP_DATA))
1464                 up_write(&block_group->data_rwsem);
1465
1466 out:
1467         io_ctl->inode = NULL;
1468         io_ctl_free(io_ctl);
1469         if (ret) {
1470                 invalidate_inode_pages2(inode->i_mapping);
1471                 BTRFS_I(inode)->generation = 0;
1472         }
1473         btrfs_update_inode(trans, root, BTRFS_I(inode));
1474         if (must_iput)
1475                 iput(inode);
1476         return ret;
1477 }
1478
1479 int btrfs_write_out_cache(struct btrfs_trans_handle *trans,
1480                           struct btrfs_block_group *block_group,
1481                           struct btrfs_path *path)
1482 {
1483         struct btrfs_fs_info *fs_info = trans->fs_info;
1484         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
1485         struct inode *inode;
1486         int ret = 0;
1487
1488         spin_lock(&block_group->lock);
1489         if (block_group->disk_cache_state < BTRFS_DC_SETUP) {
1490                 spin_unlock(&block_group->lock);
1491                 return 0;
1492         }
1493         spin_unlock(&block_group->lock);
1494
1495         inode = lookup_free_space_inode(block_group, path);
1496         if (IS_ERR(inode))
1497                 return 0;
1498
1499         ret = __btrfs_write_out_cache(fs_info->tree_root, inode, ctl,
1500                                 block_group, &block_group->io_ctl, trans);
1501         if (ret) {
1502                 btrfs_debug(fs_info,
1503           "failed to write free space cache for block group %llu error %d",
1504                           block_group->start, ret);
1505                 spin_lock(&block_group->lock);
1506                 block_group->disk_cache_state = BTRFS_DC_ERROR;
1507                 spin_unlock(&block_group->lock);
1508
1509                 block_group->io_ctl.inode = NULL;
1510                 iput(inode);
1511         }
1512
1513         /*
1514          * if ret == 0 the caller is expected to call btrfs_wait_cache_io
1515          * to wait for IO and put the inode
1516          */
1517
1518         return ret;
1519 }
1520
1521 static inline unsigned long offset_to_bit(u64 bitmap_start, u32 unit,
1522                                           u64 offset)
1523 {
1524         ASSERT(offset >= bitmap_start);
1525         offset -= bitmap_start;
1526         return (unsigned long)(div_u64(offset, unit));
1527 }
1528
1529 static inline unsigned long bytes_to_bits(u64 bytes, u32 unit)
1530 {
1531         return (unsigned long)(div_u64(bytes, unit));
1532 }
1533
1534 static inline u64 offset_to_bitmap(struct btrfs_free_space_ctl *ctl,
1535                                    u64 offset)
1536 {
1537         u64 bitmap_start;
1538         u64 bytes_per_bitmap;
1539
1540         bytes_per_bitmap = BITS_PER_BITMAP * ctl->unit;
1541         bitmap_start = offset - ctl->start;
1542         bitmap_start = div64_u64(bitmap_start, bytes_per_bitmap);
1543         bitmap_start *= bytes_per_bitmap;
1544         bitmap_start += ctl->start;
1545
1546         return bitmap_start;
1547 }
1548
1549 static int tree_insert_offset(struct rb_root *root, u64 offset,
1550                               struct rb_node *node, int bitmap)
1551 {
1552         struct rb_node **p = &root->rb_node;
1553         struct rb_node *parent = NULL;
1554         struct btrfs_free_space *info;
1555
1556         while (*p) {
1557                 parent = *p;
1558                 info = rb_entry(parent, struct btrfs_free_space, offset_index);
1559
1560                 if (offset < info->offset) {
1561                         p = &(*p)->rb_left;
1562                 } else if (offset > info->offset) {
1563                         p = &(*p)->rb_right;
1564                 } else {
1565                         /*
1566                          * we could have a bitmap entry and an extent entry
1567                          * share the same offset.  If this is the case, we want
1568                          * the extent entry to always be found first if we do a
1569                          * linear search through the tree, since we want to have
1570                          * the quickest allocation time, and allocating from an
1571                          * extent is faster than allocating from a bitmap.  So
1572                          * if we're inserting a bitmap and we find an entry at
1573                          * this offset, we want to go right, or after this entry
1574                          * logically.  If we are inserting an extent and we've
1575                          * found a bitmap, we want to go left, or before
1576                          * logically.
1577                          */
1578                         if (bitmap) {
1579                                 if (info->bitmap) {
1580                                         WARN_ON_ONCE(1);
1581                                         return -EEXIST;
1582                                 }
1583                                 p = &(*p)->rb_right;
1584                         } else {
1585                                 if (!info->bitmap) {
1586                                         WARN_ON_ONCE(1);
1587                                         return -EEXIST;
1588                                 }
1589                                 p = &(*p)->rb_left;
1590                         }
1591                 }
1592         }
1593
1594         rb_link_node(node, parent, p);
1595         rb_insert_color(node, root);
1596
1597         return 0;
1598 }
1599
1600 /*
1601  * This is a little subtle.  We *only* have ->max_extent_size set if we actually
1602  * searched through the bitmap and figured out the largest ->max_extent_size,
1603  * otherwise it's 0.  In the case that it's 0 we don't want to tell the
1604  * allocator the wrong thing, we want to use the actual real max_extent_size
1605  * we've found already if it's larger, or we want to use ->bytes.
1606  *
1607  * This matters because find_free_space() will skip entries who's ->bytes is
1608  * less than the required bytes.  So if we didn't search down this bitmap, we
1609  * may pick some previous entry that has a smaller ->max_extent_size than we
1610  * have.  For example, assume we have two entries, one that has
1611  * ->max_extent_size set to 4K and ->bytes set to 1M.  A second entry hasn't set
1612  * ->max_extent_size yet, has ->bytes set to 8K and it's contiguous.  We will
1613  *  call into find_free_space(), and return with max_extent_size == 4K, because
1614  *  that first bitmap entry had ->max_extent_size set, but the second one did
1615  *  not.  If instead we returned 8K we'd come in searching for 8K, and find the
1616  *  8K contiguous range.
1617  *
1618  *  Consider the other case, we have 2 8K chunks in that second entry and still
1619  *  don't have ->max_extent_size set.  We'll return 16K, and the next time the
1620  *  allocator comes in it'll fully search our second bitmap, and this time it'll
1621  *  get an uptodate value of 8K as the maximum chunk size.  Then we'll get the
1622  *  right allocation the next loop through.
1623  */
1624 static inline u64 get_max_extent_size(const struct btrfs_free_space *entry)
1625 {
1626         if (entry->bitmap && entry->max_extent_size)
1627                 return entry->max_extent_size;
1628         return entry->bytes;
1629 }
1630
1631 /*
1632  * We want the largest entry to be leftmost, so this is inverted from what you'd
1633  * normally expect.
1634  */
1635 static bool entry_less(struct rb_node *node, const struct rb_node *parent)
1636 {
1637         const struct btrfs_free_space *entry, *exist;
1638
1639         entry = rb_entry(node, struct btrfs_free_space, bytes_index);
1640         exist = rb_entry(parent, struct btrfs_free_space, bytes_index);
1641         return get_max_extent_size(exist) < get_max_extent_size(entry);
1642 }
1643
1644 /*
1645  * searches the tree for the given offset.
1646  *
1647  * fuzzy - If this is set, then we are trying to make an allocation, and we just
1648  * want a section that has at least bytes size and comes at or after the given
1649  * offset.
1650  */
1651 static struct btrfs_free_space *
1652 tree_search_offset(struct btrfs_free_space_ctl *ctl,
1653                    u64 offset, int bitmap_only, int fuzzy)
1654 {
1655         struct rb_node *n = ctl->free_space_offset.rb_node;
1656         struct btrfs_free_space *entry = NULL, *prev = NULL;
1657
1658         /* find entry that is closest to the 'offset' */
1659         while (n) {
1660                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1661                 prev = entry;
1662
1663                 if (offset < entry->offset)
1664                         n = n->rb_left;
1665                 else if (offset > entry->offset)
1666                         n = n->rb_right;
1667                 else
1668                         break;
1669
1670                 entry = NULL;
1671         }
1672
1673         if (bitmap_only) {
1674                 if (!entry)
1675                         return NULL;
1676                 if (entry->bitmap)
1677                         return entry;
1678
1679                 /*
1680                  * bitmap entry and extent entry may share same offset,
1681                  * in that case, bitmap entry comes after extent entry.
1682                  */
1683                 n = rb_next(n);
1684                 if (!n)
1685                         return NULL;
1686                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1687                 if (entry->offset != offset)
1688                         return NULL;
1689
1690                 WARN_ON(!entry->bitmap);
1691                 return entry;
1692         } else if (entry) {
1693                 if (entry->bitmap) {
1694                         /*
1695                          * if previous extent entry covers the offset,
1696                          * we should return it instead of the bitmap entry
1697                          */
1698                         n = rb_prev(&entry->offset_index);
1699                         if (n) {
1700                                 prev = rb_entry(n, struct btrfs_free_space,
1701                                                 offset_index);
1702                                 if (!prev->bitmap &&
1703                                     prev->offset + prev->bytes > offset)
1704                                         entry = prev;
1705                         }
1706                 }
1707                 return entry;
1708         }
1709
1710         if (!prev)
1711                 return NULL;
1712
1713         /* find last entry before the 'offset' */
1714         entry = prev;
1715         if (entry->offset > offset) {
1716                 n = rb_prev(&entry->offset_index);
1717                 if (n) {
1718                         entry = rb_entry(n, struct btrfs_free_space,
1719                                         offset_index);
1720                         ASSERT(entry->offset <= offset);
1721                 } else {
1722                         if (fuzzy)
1723                                 return entry;
1724                         else
1725                                 return NULL;
1726                 }
1727         }
1728
1729         if (entry->bitmap) {
1730                 n = rb_prev(&entry->offset_index);
1731                 if (n) {
1732                         prev = rb_entry(n, struct btrfs_free_space,
1733                                         offset_index);
1734                         if (!prev->bitmap &&
1735                             prev->offset + prev->bytes > offset)
1736                                 return prev;
1737                 }
1738                 if (entry->offset + BITS_PER_BITMAP * ctl->unit > offset)
1739                         return entry;
1740         } else if (entry->offset + entry->bytes > offset)
1741                 return entry;
1742
1743         if (!fuzzy)
1744                 return NULL;
1745
1746         while (1) {
1747                 n = rb_next(&entry->offset_index);
1748                 if (!n)
1749                         return NULL;
1750                 entry = rb_entry(n, struct btrfs_free_space, offset_index);
1751                 if (entry->bitmap) {
1752                         if (entry->offset + BITS_PER_BITMAP *
1753                             ctl->unit > offset)
1754                                 break;
1755                 } else {
1756                         if (entry->offset + entry->bytes > offset)
1757                                 break;
1758                 }
1759         }
1760         return entry;
1761 }
1762
1763 static inline void unlink_free_space(struct btrfs_free_space_ctl *ctl,
1764                                      struct btrfs_free_space *info,
1765                                      bool update_stat)
1766 {
1767         rb_erase(&info->offset_index, &ctl->free_space_offset);
1768         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1769         ctl->free_extents--;
1770
1771         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1772                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
1773                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= info->bytes;
1774         }
1775
1776         if (update_stat)
1777                 ctl->free_space -= info->bytes;
1778 }
1779
1780 static int link_free_space(struct btrfs_free_space_ctl *ctl,
1781                            struct btrfs_free_space *info)
1782 {
1783         int ret = 0;
1784
1785         ASSERT(info->bytes || info->bitmap);
1786         ret = tree_insert_offset(&ctl->free_space_offset, info->offset,
1787                                  &info->offset_index, (info->bitmap != NULL));
1788         if (ret)
1789                 return ret;
1790
1791         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1792
1793         if (!info->bitmap && !btrfs_free_space_trimmed(info)) {
1794                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
1795                 ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
1796         }
1797
1798         ctl->free_space += info->bytes;
1799         ctl->free_extents++;
1800         return ret;
1801 }
1802
1803 static void relink_bitmap_entry(struct btrfs_free_space_ctl *ctl,
1804                                 struct btrfs_free_space *info)
1805 {
1806         ASSERT(info->bitmap);
1807
1808         /*
1809          * If our entry is empty it's because we're on a cluster and we don't
1810          * want to re-link it into our ctl bytes index.
1811          */
1812         if (RB_EMPTY_NODE(&info->bytes_index))
1813                 return;
1814
1815         rb_erase_cached(&info->bytes_index, &ctl->free_space_bytes);
1816         rb_add_cached(&info->bytes_index, &ctl->free_space_bytes, entry_less);
1817 }
1818
1819 static inline void bitmap_clear_bits(struct btrfs_free_space_ctl *ctl,
1820                                      struct btrfs_free_space *info,
1821                                      u64 offset, u64 bytes, bool update_stat)
1822 {
1823         unsigned long start, count, end;
1824         int extent_delta = -1;
1825
1826         start = offset_to_bit(info->offset, ctl->unit, offset);
1827         count = bytes_to_bits(bytes, ctl->unit);
1828         end = start + count;
1829         ASSERT(end <= BITS_PER_BITMAP);
1830
1831         bitmap_clear(info->bitmap, start, count);
1832
1833         info->bytes -= bytes;
1834         if (info->max_extent_size > ctl->unit)
1835                 info->max_extent_size = 0;
1836
1837         relink_bitmap_entry(ctl, info);
1838
1839         if (start && test_bit(start - 1, info->bitmap))
1840                 extent_delta++;
1841
1842         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1843                 extent_delta++;
1844
1845         info->bitmap_extents += extent_delta;
1846         if (!btrfs_free_space_trimmed(info)) {
1847                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1848                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
1849         }
1850
1851         if (update_stat)
1852                 ctl->free_space -= bytes;
1853 }
1854
1855 static void bitmap_set_bits(struct btrfs_free_space_ctl *ctl,
1856                             struct btrfs_free_space *info, u64 offset,
1857                             u64 bytes)
1858 {
1859         unsigned long start, count, end;
1860         int extent_delta = 1;
1861
1862         start = offset_to_bit(info->offset, ctl->unit, offset);
1863         count = bytes_to_bits(bytes, ctl->unit);
1864         end = start + count;
1865         ASSERT(end <= BITS_PER_BITMAP);
1866
1867         bitmap_set(info->bitmap, start, count);
1868
1869         /*
1870          * We set some bytes, we have no idea what the max extent size is
1871          * anymore.
1872          */
1873         info->max_extent_size = 0;
1874         info->bytes += bytes;
1875         ctl->free_space += bytes;
1876
1877         relink_bitmap_entry(ctl, info);
1878
1879         if (start && test_bit(start - 1, info->bitmap))
1880                 extent_delta--;
1881
1882         if (end < BITS_PER_BITMAP && test_bit(end, info->bitmap))
1883                 extent_delta--;
1884
1885         info->bitmap_extents += extent_delta;
1886         if (!btrfs_free_space_trimmed(info)) {
1887                 ctl->discardable_extents[BTRFS_STAT_CURR] += extent_delta;
1888                 ctl->discardable_bytes[BTRFS_STAT_CURR] += bytes;
1889         }
1890 }
1891
1892 /*
1893  * If we can not find suitable extent, we will use bytes to record
1894  * the size of the max extent.
1895  */
1896 static int search_bitmap(struct btrfs_free_space_ctl *ctl,
1897                          struct btrfs_free_space *bitmap_info, u64 *offset,
1898                          u64 *bytes, bool for_alloc)
1899 {
1900         unsigned long found_bits = 0;
1901         unsigned long max_bits = 0;
1902         unsigned long bits, i;
1903         unsigned long next_zero;
1904         unsigned long extent_bits;
1905
1906         /*
1907          * Skip searching the bitmap if we don't have a contiguous section that
1908          * is large enough for this allocation.
1909          */
1910         if (for_alloc &&
1911             bitmap_info->max_extent_size &&
1912             bitmap_info->max_extent_size < *bytes) {
1913                 *bytes = bitmap_info->max_extent_size;
1914                 return -1;
1915         }
1916
1917         i = offset_to_bit(bitmap_info->offset, ctl->unit,
1918                           max_t(u64, *offset, bitmap_info->offset));
1919         bits = bytes_to_bits(*bytes, ctl->unit);
1920
1921         for_each_set_bit_from(i, bitmap_info->bitmap, BITS_PER_BITMAP) {
1922                 if (for_alloc && bits == 1) {
1923                         found_bits = 1;
1924                         break;
1925                 }
1926                 next_zero = find_next_zero_bit(bitmap_info->bitmap,
1927                                                BITS_PER_BITMAP, i);
1928                 extent_bits = next_zero - i;
1929                 if (extent_bits >= bits) {
1930                         found_bits = extent_bits;
1931                         break;
1932                 } else if (extent_bits > max_bits) {
1933                         max_bits = extent_bits;
1934                 }
1935                 i = next_zero;
1936         }
1937
1938         if (found_bits) {
1939                 *offset = (u64)(i * ctl->unit) + bitmap_info->offset;
1940                 *bytes = (u64)(found_bits) * ctl->unit;
1941                 return 0;
1942         }
1943
1944         *bytes = (u64)(max_bits) * ctl->unit;
1945         bitmap_info->max_extent_size = *bytes;
1946         relink_bitmap_entry(ctl, bitmap_info);
1947         return -1;
1948 }
1949
1950 /* Cache the size of the max extent in bytes */
1951 static struct btrfs_free_space *
1952 find_free_space(struct btrfs_free_space_ctl *ctl, u64 *offset, u64 *bytes,
1953                 unsigned long align, u64 *max_extent_size, bool use_bytes_index)
1954 {
1955         struct btrfs_free_space *entry;
1956         struct rb_node *node;
1957         u64 tmp;
1958         u64 align_off;
1959         int ret;
1960
1961         if (!ctl->free_space_offset.rb_node)
1962                 goto out;
1963 again:
1964         if (use_bytes_index) {
1965                 node = rb_first_cached(&ctl->free_space_bytes);
1966         } else {
1967                 entry = tree_search_offset(ctl, offset_to_bitmap(ctl, *offset),
1968                                            0, 1);
1969                 if (!entry)
1970                         goto out;
1971                 node = &entry->offset_index;
1972         }
1973
1974         for (; node; node = rb_next(node)) {
1975                 if (use_bytes_index)
1976                         entry = rb_entry(node, struct btrfs_free_space,
1977                                          bytes_index);
1978                 else
1979                         entry = rb_entry(node, struct btrfs_free_space,
1980                                          offset_index);
1981
1982                 /*
1983                  * If we are using the bytes index then all subsequent entries
1984                  * in this tree are going to be < bytes, so simply set the max
1985                  * extent size and exit the loop.
1986                  *
1987                  * If we're using the offset index then we need to keep going
1988                  * through the rest of the tree.
1989                  */
1990                 if (entry->bytes < *bytes) {
1991                         *max_extent_size = max(get_max_extent_size(entry),
1992                                                *max_extent_size);
1993                         if (use_bytes_index)
1994                                 break;
1995                         continue;
1996                 }
1997
1998                 /* make sure the space returned is big enough
1999                  * to match our requested alignment
2000                  */
2001                 if (*bytes >= align) {
2002                         tmp = entry->offset - ctl->start + align - 1;
2003                         tmp = div64_u64(tmp, align);
2004                         tmp = tmp * align + ctl->start;
2005                         align_off = tmp - entry->offset;
2006                 } else {
2007                         align_off = 0;
2008                         tmp = entry->offset;
2009                 }
2010
2011                 /*
2012                  * We don't break here if we're using the bytes index because we
2013                  * may have another entry that has the correct alignment that is
2014                  * the right size, so we don't want to miss that possibility.
2015                  * At worst this adds another loop through the logic, but if we
2016                  * broke here we could prematurely ENOSPC.
2017                  */
2018                 if (entry->bytes < *bytes + align_off) {
2019                         *max_extent_size = max(get_max_extent_size(entry),
2020                                                *max_extent_size);
2021                         continue;
2022                 }
2023
2024                 if (entry->bitmap) {
2025                         struct rb_node *old_next = rb_next(node);
2026                         u64 size = *bytes;
2027
2028                         ret = search_bitmap(ctl, entry, &tmp, &size, true);
2029                         if (!ret) {
2030                                 *offset = tmp;
2031                                 *bytes = size;
2032                                 return entry;
2033                         } else {
2034                                 *max_extent_size =
2035                                         max(get_max_extent_size(entry),
2036                                             *max_extent_size);
2037                         }
2038
2039                         /*
2040                          * The bitmap may have gotten re-arranged in the space
2041                          * index here because the max_extent_size may have been
2042                          * updated.  Start from the beginning again if this
2043                          * happened.
2044                          */
2045                         if (use_bytes_index && old_next != rb_next(node))
2046                                 goto again;
2047                         continue;
2048                 }
2049
2050                 *offset = tmp;
2051                 *bytes = entry->bytes - align_off;
2052                 return entry;
2053         }
2054 out:
2055         return NULL;
2056 }
2057
2058 static void add_new_bitmap(struct btrfs_free_space_ctl *ctl,
2059                            struct btrfs_free_space *info, u64 offset)
2060 {
2061         info->offset = offset_to_bitmap(ctl, offset);
2062         info->bytes = 0;
2063         info->bitmap_extents = 0;
2064         INIT_LIST_HEAD(&info->list);
2065         link_free_space(ctl, info);
2066         ctl->total_bitmaps++;
2067         recalculate_thresholds(ctl);
2068 }
2069
2070 static void free_bitmap(struct btrfs_free_space_ctl *ctl,
2071                         struct btrfs_free_space *bitmap_info)
2072 {
2073         /*
2074          * Normally when this is called, the bitmap is completely empty. However,
2075          * if we are blowing up the free space cache for one reason or another
2076          * via __btrfs_remove_free_space_cache(), then it may not be freed and
2077          * we may leave stats on the table.
2078          */
2079         if (bitmap_info->bytes && !btrfs_free_space_trimmed(bitmap_info)) {
2080                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
2081                         bitmap_info->bitmap_extents;
2082                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bitmap_info->bytes;
2083
2084         }
2085         unlink_free_space(ctl, bitmap_info, true);
2086         kmem_cache_free(btrfs_free_space_bitmap_cachep, bitmap_info->bitmap);
2087         kmem_cache_free(btrfs_free_space_cachep, bitmap_info);
2088         ctl->total_bitmaps--;
2089         recalculate_thresholds(ctl);
2090 }
2091
2092 static noinline int remove_from_bitmap(struct btrfs_free_space_ctl *ctl,
2093                               struct btrfs_free_space *bitmap_info,
2094                               u64 *offset, u64 *bytes)
2095 {
2096         u64 end;
2097         u64 search_start, search_bytes;
2098         int ret;
2099
2100 again:
2101         end = bitmap_info->offset + (u64)(BITS_PER_BITMAP * ctl->unit) - 1;
2102
2103         /*
2104          * We need to search for bits in this bitmap.  We could only cover some
2105          * of the extent in this bitmap thanks to how we add space, so we need
2106          * to search for as much as it as we can and clear that amount, and then
2107          * go searching for the next bit.
2108          */
2109         search_start = *offset;
2110         search_bytes = ctl->unit;
2111         search_bytes = min(search_bytes, end - search_start + 1);
2112         ret = search_bitmap(ctl, bitmap_info, &search_start, &search_bytes,
2113                             false);
2114         if (ret < 0 || search_start != *offset)
2115                 return -EINVAL;
2116
2117         /* We may have found more bits than what we need */
2118         search_bytes = min(search_bytes, *bytes);
2119
2120         /* Cannot clear past the end of the bitmap */
2121         search_bytes = min(search_bytes, end - search_start + 1);
2122
2123         bitmap_clear_bits(ctl, bitmap_info, search_start, search_bytes, true);
2124         *offset += search_bytes;
2125         *bytes -= search_bytes;
2126
2127         if (*bytes) {
2128                 struct rb_node *next = rb_next(&bitmap_info->offset_index);
2129                 if (!bitmap_info->bytes)
2130                         free_bitmap(ctl, bitmap_info);
2131
2132                 /*
2133                  * no entry after this bitmap, but we still have bytes to
2134                  * remove, so something has gone wrong.
2135                  */
2136                 if (!next)
2137                         return -EINVAL;
2138
2139                 bitmap_info = rb_entry(next, struct btrfs_free_space,
2140                                        offset_index);
2141
2142                 /*
2143                  * if the next entry isn't a bitmap we need to return to let the
2144                  * extent stuff do its work.
2145                  */
2146                 if (!bitmap_info->bitmap)
2147                         return -EAGAIN;
2148
2149                 /*
2150                  * Ok the next item is a bitmap, but it may not actually hold
2151                  * the information for the rest of this free space stuff, so
2152                  * look for it, and if we don't find it return so we can try
2153                  * everything over again.
2154                  */
2155                 search_start = *offset;
2156                 search_bytes = ctl->unit;
2157                 ret = search_bitmap(ctl, bitmap_info, &search_start,
2158                                     &search_bytes, false);
2159                 if (ret < 0 || search_start != *offset)
2160                         return -EAGAIN;
2161
2162                 goto again;
2163         } else if (!bitmap_info->bytes)
2164                 free_bitmap(ctl, bitmap_info);
2165
2166         return 0;
2167 }
2168
2169 static u64 add_bytes_to_bitmap(struct btrfs_free_space_ctl *ctl,
2170                                struct btrfs_free_space *info, u64 offset,
2171                                u64 bytes, enum btrfs_trim_state trim_state)
2172 {
2173         u64 bytes_to_set = 0;
2174         u64 end;
2175
2176         /*
2177          * This is a tradeoff to make bitmap trim state minimal.  We mark the
2178          * whole bitmap untrimmed if at any point we add untrimmed regions.
2179          */
2180         if (trim_state == BTRFS_TRIM_STATE_UNTRIMMED) {
2181                 if (btrfs_free_space_trimmed(info)) {
2182                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
2183                                 info->bitmap_extents;
2184                         ctl->discardable_bytes[BTRFS_STAT_CURR] += info->bytes;
2185                 }
2186                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2187         }
2188
2189         end = info->offset + (u64)(BITS_PER_BITMAP * ctl->unit);
2190
2191         bytes_to_set = min(end - offset, bytes);
2192
2193         bitmap_set_bits(ctl, info, offset, bytes_to_set);
2194
2195         return bytes_to_set;
2196
2197 }
2198
2199 static bool use_bitmap(struct btrfs_free_space_ctl *ctl,
2200                       struct btrfs_free_space *info)
2201 {
2202         struct btrfs_block_group *block_group = ctl->block_group;
2203         struct btrfs_fs_info *fs_info = block_group->fs_info;
2204         bool forced = false;
2205
2206 #ifdef CONFIG_BTRFS_DEBUG
2207         if (btrfs_should_fragment_free_space(block_group))
2208                 forced = true;
2209 #endif
2210
2211         /* This is a way to reclaim large regions from the bitmaps. */
2212         if (!forced && info->bytes >= FORCE_EXTENT_THRESHOLD)
2213                 return false;
2214
2215         /*
2216          * If we are below the extents threshold then we can add this as an
2217          * extent, and don't have to deal with the bitmap
2218          */
2219         if (!forced && ctl->free_extents < ctl->extents_thresh) {
2220                 /*
2221                  * If this block group has some small extents we don't want to
2222                  * use up all of our free slots in the cache with them, we want
2223                  * to reserve them to larger extents, however if we have plenty
2224                  * of cache left then go ahead an dadd them, no sense in adding
2225                  * the overhead of a bitmap if we don't have to.
2226                  */
2227                 if (info->bytes <= fs_info->sectorsize * 8) {
2228                         if (ctl->free_extents * 3 <= ctl->extents_thresh)
2229                                 return false;
2230                 } else {
2231                         return false;
2232                 }
2233         }
2234
2235         /*
2236          * The original block groups from mkfs can be really small, like 8
2237          * megabytes, so don't bother with a bitmap for those entries.  However
2238          * some block groups can be smaller than what a bitmap would cover but
2239          * are still large enough that they could overflow the 32k memory limit,
2240          * so allow those block groups to still be allowed to have a bitmap
2241          * entry.
2242          */
2243         if (((BITS_PER_BITMAP * ctl->unit) >> 1) > block_group->length)
2244                 return false;
2245
2246         return true;
2247 }
2248
2249 static const struct btrfs_free_space_op free_space_op = {
2250         .use_bitmap             = use_bitmap,
2251 };
2252
2253 static int insert_into_bitmap(struct btrfs_free_space_ctl *ctl,
2254                               struct btrfs_free_space *info)
2255 {
2256         struct btrfs_free_space *bitmap_info;
2257         struct btrfs_block_group *block_group = NULL;
2258         int added = 0;
2259         u64 bytes, offset, bytes_added;
2260         enum btrfs_trim_state trim_state;
2261         int ret;
2262
2263         bytes = info->bytes;
2264         offset = info->offset;
2265         trim_state = info->trim_state;
2266
2267         if (!ctl->op->use_bitmap(ctl, info))
2268                 return 0;
2269
2270         if (ctl->op == &free_space_op)
2271                 block_group = ctl->block_group;
2272 again:
2273         /*
2274          * Since we link bitmaps right into the cluster we need to see if we
2275          * have a cluster here, and if so and it has our bitmap we need to add
2276          * the free space to that bitmap.
2277          */
2278         if (block_group && !list_empty(&block_group->cluster_list)) {
2279                 struct btrfs_free_cluster *cluster;
2280                 struct rb_node *node;
2281                 struct btrfs_free_space *entry;
2282
2283                 cluster = list_entry(block_group->cluster_list.next,
2284                                      struct btrfs_free_cluster,
2285                                      block_group_list);
2286                 spin_lock(&cluster->lock);
2287                 node = rb_first(&cluster->root);
2288                 if (!node) {
2289                         spin_unlock(&cluster->lock);
2290                         goto no_cluster_bitmap;
2291                 }
2292
2293                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2294                 if (!entry->bitmap) {
2295                         spin_unlock(&cluster->lock);
2296                         goto no_cluster_bitmap;
2297                 }
2298
2299                 if (entry->offset == offset_to_bitmap(ctl, offset)) {
2300                         bytes_added = add_bytes_to_bitmap(ctl, entry, offset,
2301                                                           bytes, trim_state);
2302                         bytes -= bytes_added;
2303                         offset += bytes_added;
2304                 }
2305                 spin_unlock(&cluster->lock);
2306                 if (!bytes) {
2307                         ret = 1;
2308                         goto out;
2309                 }
2310         }
2311
2312 no_cluster_bitmap:
2313         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2314                                          1, 0);
2315         if (!bitmap_info) {
2316                 ASSERT(added == 0);
2317                 goto new_bitmap;
2318         }
2319
2320         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
2321                                           trim_state);
2322         bytes -= bytes_added;
2323         offset += bytes_added;
2324         added = 0;
2325
2326         if (!bytes) {
2327                 ret = 1;
2328                 goto out;
2329         } else
2330                 goto again;
2331
2332 new_bitmap:
2333         if (info && info->bitmap) {
2334                 add_new_bitmap(ctl, info, offset);
2335                 added = 1;
2336                 info = NULL;
2337                 goto again;
2338         } else {
2339                 spin_unlock(&ctl->tree_lock);
2340
2341                 /* no pre-allocated info, allocate a new one */
2342                 if (!info) {
2343                         info = kmem_cache_zalloc(btrfs_free_space_cachep,
2344                                                  GFP_NOFS);
2345                         if (!info) {
2346                                 spin_lock(&ctl->tree_lock);
2347                                 ret = -ENOMEM;
2348                                 goto out;
2349                         }
2350                 }
2351
2352                 /* allocate the bitmap */
2353                 info->bitmap = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep,
2354                                                  GFP_NOFS);
2355                 info->trim_state = BTRFS_TRIM_STATE_TRIMMED;
2356                 spin_lock(&ctl->tree_lock);
2357                 if (!info->bitmap) {
2358                         ret = -ENOMEM;
2359                         goto out;
2360                 }
2361                 goto again;
2362         }
2363
2364 out:
2365         if (info) {
2366                 if (info->bitmap)
2367                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
2368                                         info->bitmap);
2369                 kmem_cache_free(btrfs_free_space_cachep, info);
2370         }
2371
2372         return ret;
2373 }
2374
2375 /*
2376  * Free space merging rules:
2377  *  1) Merge trimmed areas together
2378  *  2) Let untrimmed areas coalesce with trimmed areas
2379  *  3) Always pull neighboring regions from bitmaps
2380  *
2381  * The above rules are for when we merge free space based on btrfs_trim_state.
2382  * Rules 2 and 3 are subtle because they are suboptimal, but are done for the
2383  * same reason: to promote larger extent regions which makes life easier for
2384  * find_free_extent().  Rule 2 enables coalescing based on the common path
2385  * being returning free space from btrfs_finish_extent_commit().  So when free
2386  * space is trimmed, it will prevent aggregating trimmed new region and
2387  * untrimmed regions in the rb_tree.  Rule 3 is purely to obtain larger extents
2388  * and provide find_free_extent() with the largest extents possible hoping for
2389  * the reuse path.
2390  */
2391 static bool try_merge_free_space(struct btrfs_free_space_ctl *ctl,
2392                           struct btrfs_free_space *info, bool update_stat)
2393 {
2394         struct btrfs_free_space *left_info = NULL;
2395         struct btrfs_free_space *right_info;
2396         bool merged = false;
2397         u64 offset = info->offset;
2398         u64 bytes = info->bytes;
2399         const bool is_trimmed = btrfs_free_space_trimmed(info);
2400
2401         /*
2402          * first we want to see if there is free space adjacent to the range we
2403          * are adding, if there is remove that struct and add a new one to
2404          * cover the entire range
2405          */
2406         right_info = tree_search_offset(ctl, offset + bytes, 0, 0);
2407         if (right_info && rb_prev(&right_info->offset_index))
2408                 left_info = rb_entry(rb_prev(&right_info->offset_index),
2409                                      struct btrfs_free_space, offset_index);
2410         else if (!right_info)
2411                 left_info = tree_search_offset(ctl, offset - 1, 0, 0);
2412
2413         /* See try_merge_free_space() comment. */
2414         if (right_info && !right_info->bitmap &&
2415             (!is_trimmed || btrfs_free_space_trimmed(right_info))) {
2416                 unlink_free_space(ctl, right_info, update_stat);
2417                 info->bytes += right_info->bytes;
2418                 kmem_cache_free(btrfs_free_space_cachep, right_info);
2419                 merged = true;
2420         }
2421
2422         /* See try_merge_free_space() comment. */
2423         if (left_info && !left_info->bitmap &&
2424             left_info->offset + left_info->bytes == offset &&
2425             (!is_trimmed || btrfs_free_space_trimmed(left_info))) {
2426                 unlink_free_space(ctl, left_info, update_stat);
2427                 info->offset = left_info->offset;
2428                 info->bytes += left_info->bytes;
2429                 kmem_cache_free(btrfs_free_space_cachep, left_info);
2430                 merged = true;
2431         }
2432
2433         return merged;
2434 }
2435
2436 static bool steal_from_bitmap_to_end(struct btrfs_free_space_ctl *ctl,
2437                                      struct btrfs_free_space *info,
2438                                      bool update_stat)
2439 {
2440         struct btrfs_free_space *bitmap;
2441         unsigned long i;
2442         unsigned long j;
2443         const u64 end = info->offset + info->bytes;
2444         const u64 bitmap_offset = offset_to_bitmap(ctl, end);
2445         u64 bytes;
2446
2447         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2448         if (!bitmap)
2449                 return false;
2450
2451         i = offset_to_bit(bitmap->offset, ctl->unit, end);
2452         j = find_next_zero_bit(bitmap->bitmap, BITS_PER_BITMAP, i);
2453         if (j == i)
2454                 return false;
2455         bytes = (j - i) * ctl->unit;
2456         info->bytes += bytes;
2457
2458         /* See try_merge_free_space() comment. */
2459         if (!btrfs_free_space_trimmed(bitmap))
2460                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2461
2462         bitmap_clear_bits(ctl, bitmap, end, bytes, update_stat);
2463
2464         if (!bitmap->bytes)
2465                 free_bitmap(ctl, bitmap);
2466
2467         return true;
2468 }
2469
2470 static bool steal_from_bitmap_to_front(struct btrfs_free_space_ctl *ctl,
2471                                        struct btrfs_free_space *info,
2472                                        bool update_stat)
2473 {
2474         struct btrfs_free_space *bitmap;
2475         u64 bitmap_offset;
2476         unsigned long i;
2477         unsigned long j;
2478         unsigned long prev_j;
2479         u64 bytes;
2480
2481         bitmap_offset = offset_to_bitmap(ctl, info->offset);
2482         /* If we're on a boundary, try the previous logical bitmap. */
2483         if (bitmap_offset == info->offset) {
2484                 if (info->offset == 0)
2485                         return false;
2486                 bitmap_offset = offset_to_bitmap(ctl, info->offset - 1);
2487         }
2488
2489         bitmap = tree_search_offset(ctl, bitmap_offset, 1, 0);
2490         if (!bitmap)
2491                 return false;
2492
2493         i = offset_to_bit(bitmap->offset, ctl->unit, info->offset) - 1;
2494         j = 0;
2495         prev_j = (unsigned long)-1;
2496         for_each_clear_bit_from(j, bitmap->bitmap, BITS_PER_BITMAP) {
2497                 if (j > i)
2498                         break;
2499                 prev_j = j;
2500         }
2501         if (prev_j == i)
2502                 return false;
2503
2504         if (prev_j == (unsigned long)-1)
2505                 bytes = (i + 1) * ctl->unit;
2506         else
2507                 bytes = (i - prev_j) * ctl->unit;
2508
2509         info->offset -= bytes;
2510         info->bytes += bytes;
2511
2512         /* See try_merge_free_space() comment. */
2513         if (!btrfs_free_space_trimmed(bitmap))
2514                 info->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2515
2516         bitmap_clear_bits(ctl, bitmap, info->offset, bytes, update_stat);
2517
2518         if (!bitmap->bytes)
2519                 free_bitmap(ctl, bitmap);
2520
2521         return true;
2522 }
2523
2524 /*
2525  * We prefer always to allocate from extent entries, both for clustered and
2526  * non-clustered allocation requests. So when attempting to add a new extent
2527  * entry, try to see if there's adjacent free space in bitmap entries, and if
2528  * there is, migrate that space from the bitmaps to the extent.
2529  * Like this we get better chances of satisfying space allocation requests
2530  * because we attempt to satisfy them based on a single cache entry, and never
2531  * on 2 or more entries - even if the entries represent a contiguous free space
2532  * region (e.g. 1 extent entry + 1 bitmap entry starting where the extent entry
2533  * ends).
2534  */
2535 static void steal_from_bitmap(struct btrfs_free_space_ctl *ctl,
2536                               struct btrfs_free_space *info,
2537                               bool update_stat)
2538 {
2539         /*
2540          * Only work with disconnected entries, as we can change their offset,
2541          * and must be extent entries.
2542          */
2543         ASSERT(!info->bitmap);
2544         ASSERT(RB_EMPTY_NODE(&info->offset_index));
2545
2546         if (ctl->total_bitmaps > 0) {
2547                 bool stole_end;
2548                 bool stole_front = false;
2549
2550                 stole_end = steal_from_bitmap_to_end(ctl, info, update_stat);
2551                 if (ctl->total_bitmaps > 0)
2552                         stole_front = steal_from_bitmap_to_front(ctl, info,
2553                                                                  update_stat);
2554
2555                 if (stole_end || stole_front)
2556                         try_merge_free_space(ctl, info, update_stat);
2557         }
2558 }
2559
2560 int __btrfs_add_free_space(struct btrfs_block_group *block_group,
2561                            u64 offset, u64 bytes,
2562                            enum btrfs_trim_state trim_state)
2563 {
2564         struct btrfs_fs_info *fs_info = block_group->fs_info;
2565         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2566         struct btrfs_free_space *info;
2567         int ret = 0;
2568         u64 filter_bytes = bytes;
2569
2570         ASSERT(!btrfs_is_zoned(fs_info));
2571
2572         info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
2573         if (!info)
2574                 return -ENOMEM;
2575
2576         info->offset = offset;
2577         info->bytes = bytes;
2578         info->trim_state = trim_state;
2579         RB_CLEAR_NODE(&info->offset_index);
2580         RB_CLEAR_NODE(&info->bytes_index);
2581
2582         spin_lock(&ctl->tree_lock);
2583
2584         if (try_merge_free_space(ctl, info, true))
2585                 goto link;
2586
2587         /*
2588          * There was no extent directly to the left or right of this new
2589          * extent then we know we're going to have to allocate a new extent, so
2590          * before we do that see if we need to drop this into a bitmap
2591          */
2592         ret = insert_into_bitmap(ctl, info);
2593         if (ret < 0) {
2594                 goto out;
2595         } else if (ret) {
2596                 ret = 0;
2597                 goto out;
2598         }
2599 link:
2600         /*
2601          * Only steal free space from adjacent bitmaps if we're sure we're not
2602          * going to add the new free space to existing bitmap entries - because
2603          * that would mean unnecessary work that would be reverted. Therefore
2604          * attempt to steal space from bitmaps if we're adding an extent entry.
2605          */
2606         steal_from_bitmap(ctl, info, true);
2607
2608         filter_bytes = max(filter_bytes, info->bytes);
2609
2610         ret = link_free_space(ctl, info);
2611         if (ret)
2612                 kmem_cache_free(btrfs_free_space_cachep, info);
2613 out:
2614         btrfs_discard_update_discardable(block_group);
2615         spin_unlock(&ctl->tree_lock);
2616
2617         if (ret) {
2618                 btrfs_crit(fs_info, "unable to add free space :%d", ret);
2619                 ASSERT(ret != -EEXIST);
2620         }
2621
2622         if (trim_state != BTRFS_TRIM_STATE_TRIMMED) {
2623                 btrfs_discard_check_filter(block_group, filter_bytes);
2624                 btrfs_discard_queue_work(&fs_info->discard_ctl, block_group);
2625         }
2626
2627         return ret;
2628 }
2629
2630 static int __btrfs_add_free_space_zoned(struct btrfs_block_group *block_group,
2631                                         u64 bytenr, u64 size, bool used)
2632 {
2633         struct btrfs_space_info *sinfo = block_group->space_info;
2634         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2635         u64 offset = bytenr - block_group->start;
2636         u64 to_free, to_unusable;
2637         int bg_reclaim_threshold = 0;
2638         bool initial = (size == block_group->length);
2639         u64 reclaimable_unusable;
2640
2641         WARN_ON(!initial && offset + size > block_group->zone_capacity);
2642
2643         if (!initial)
2644                 bg_reclaim_threshold = READ_ONCE(sinfo->bg_reclaim_threshold);
2645
2646         spin_lock(&ctl->tree_lock);
2647         if (!used)
2648                 to_free = size;
2649         else if (initial)
2650                 to_free = block_group->zone_capacity;
2651         else if (offset >= block_group->alloc_offset)
2652                 to_free = size;
2653         else if (offset + size <= block_group->alloc_offset)
2654                 to_free = 0;
2655         else
2656                 to_free = offset + size - block_group->alloc_offset;
2657         to_unusable = size - to_free;
2658
2659         ctl->free_space += to_free;
2660         /*
2661          * If the block group is read-only, we should account freed space into
2662          * bytes_readonly.
2663          */
2664         if (!block_group->ro)
2665                 block_group->zone_unusable += to_unusable;
2666         spin_unlock(&ctl->tree_lock);
2667         if (!used) {
2668                 spin_lock(&block_group->lock);
2669                 block_group->alloc_offset -= size;
2670                 spin_unlock(&block_group->lock);
2671         }
2672
2673         reclaimable_unusable = block_group->zone_unusable -
2674                                (block_group->length - block_group->zone_capacity);
2675         /* All the region is now unusable. Mark it as unused and reclaim */
2676         if (block_group->zone_unusable == block_group->length) {
2677                 btrfs_mark_bg_unused(block_group);
2678         } else if (bg_reclaim_threshold &&
2679                    reclaimable_unusable >=
2680                    div_factor_fine(block_group->zone_capacity,
2681                                    bg_reclaim_threshold)) {
2682                 btrfs_mark_bg_to_reclaim(block_group);
2683         }
2684
2685         return 0;
2686 }
2687
2688 int btrfs_add_free_space(struct btrfs_block_group *block_group,
2689                          u64 bytenr, u64 size)
2690 {
2691         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2692
2693         if (btrfs_is_zoned(block_group->fs_info))
2694                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2695                                                     true);
2696
2697         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC))
2698                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2699
2700         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2701 }
2702
2703 int btrfs_add_free_space_unused(struct btrfs_block_group *block_group,
2704                                 u64 bytenr, u64 size)
2705 {
2706         if (btrfs_is_zoned(block_group->fs_info))
2707                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2708                                                     false);
2709
2710         return btrfs_add_free_space(block_group, bytenr, size);
2711 }
2712
2713 /*
2714  * This is a subtle distinction because when adding free space back in general,
2715  * we want it to be added as untrimmed for async. But in the case where we add
2716  * it on loading of a block group, we want to consider it trimmed.
2717  */
2718 int btrfs_add_free_space_async_trimmed(struct btrfs_block_group *block_group,
2719                                        u64 bytenr, u64 size)
2720 {
2721         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
2722
2723         if (btrfs_is_zoned(block_group->fs_info))
2724                 return __btrfs_add_free_space_zoned(block_group, bytenr, size,
2725                                                     true);
2726
2727         if (btrfs_test_opt(block_group->fs_info, DISCARD_SYNC) ||
2728             btrfs_test_opt(block_group->fs_info, DISCARD_ASYNC))
2729                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
2730
2731         return __btrfs_add_free_space(block_group, bytenr, size, trim_state);
2732 }
2733
2734 int btrfs_remove_free_space(struct btrfs_block_group *block_group,
2735                             u64 offset, u64 bytes)
2736 {
2737         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2738         struct btrfs_free_space *info;
2739         int ret;
2740         bool re_search = false;
2741
2742         if (btrfs_is_zoned(block_group->fs_info)) {
2743                 /*
2744                  * This can happen with conventional zones when replaying log.
2745                  * Since the allocation info of tree-log nodes are not recorded
2746                  * to the extent-tree, calculate_alloc_pointer() failed to
2747                  * advance the allocation pointer after last allocated tree log
2748                  * node blocks.
2749                  *
2750                  * This function is called from
2751                  * btrfs_pin_extent_for_log_replay() when replaying the log.
2752                  * Advance the pointer not to overwrite the tree-log nodes.
2753                  */
2754                 if (block_group->start + block_group->alloc_offset <
2755                     offset + bytes) {
2756                         block_group->alloc_offset =
2757                                 offset + bytes - block_group->start;
2758                 }
2759                 return 0;
2760         }
2761
2762         spin_lock(&ctl->tree_lock);
2763
2764 again:
2765         ret = 0;
2766         if (!bytes)
2767                 goto out_lock;
2768
2769         info = tree_search_offset(ctl, offset, 0, 0);
2770         if (!info) {
2771                 /*
2772                  * oops didn't find an extent that matched the space we wanted
2773                  * to remove, look for a bitmap instead
2774                  */
2775                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
2776                                           1, 0);
2777                 if (!info) {
2778                         /*
2779                          * If we found a partial bit of our free space in a
2780                          * bitmap but then couldn't find the other part this may
2781                          * be a problem, so WARN about it.
2782                          */
2783                         WARN_ON(re_search);
2784                         goto out_lock;
2785                 }
2786         }
2787
2788         re_search = false;
2789         if (!info->bitmap) {
2790                 unlink_free_space(ctl, info, true);
2791                 if (offset == info->offset) {
2792                         u64 to_free = min(bytes, info->bytes);
2793
2794                         info->bytes -= to_free;
2795                         info->offset += to_free;
2796                         if (info->bytes) {
2797                                 ret = link_free_space(ctl, info);
2798                                 WARN_ON(ret);
2799                         } else {
2800                                 kmem_cache_free(btrfs_free_space_cachep, info);
2801                         }
2802
2803                         offset += to_free;
2804                         bytes -= to_free;
2805                         goto again;
2806                 } else {
2807                         u64 old_end = info->bytes + info->offset;
2808
2809                         info->bytes = offset - info->offset;
2810                         ret = link_free_space(ctl, info);
2811                         WARN_ON(ret);
2812                         if (ret)
2813                                 goto out_lock;
2814
2815                         /* Not enough bytes in this entry to satisfy us */
2816                         if (old_end < offset + bytes) {
2817                                 bytes -= old_end - offset;
2818                                 offset = old_end;
2819                                 goto again;
2820                         } else if (old_end == offset + bytes) {
2821                                 /* all done */
2822                                 goto out_lock;
2823                         }
2824                         spin_unlock(&ctl->tree_lock);
2825
2826                         ret = __btrfs_add_free_space(block_group,
2827                                                      offset + bytes,
2828                                                      old_end - (offset + bytes),
2829                                                      info->trim_state);
2830                         WARN_ON(ret);
2831                         goto out;
2832                 }
2833         }
2834
2835         ret = remove_from_bitmap(ctl, info, &offset, &bytes);
2836         if (ret == -EAGAIN) {
2837                 re_search = true;
2838                 goto again;
2839         }
2840 out_lock:
2841         btrfs_discard_update_discardable(block_group);
2842         spin_unlock(&ctl->tree_lock);
2843 out:
2844         return ret;
2845 }
2846
2847 void btrfs_dump_free_space(struct btrfs_block_group *block_group,
2848                            u64 bytes)
2849 {
2850         struct btrfs_fs_info *fs_info = block_group->fs_info;
2851         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2852         struct btrfs_free_space *info;
2853         struct rb_node *n;
2854         int count = 0;
2855
2856         /*
2857          * Zoned btrfs does not use free space tree and cluster. Just print
2858          * out the free space after the allocation offset.
2859          */
2860         if (btrfs_is_zoned(fs_info)) {
2861                 btrfs_info(fs_info, "free space %llu active %d",
2862                            block_group->zone_capacity - block_group->alloc_offset,
2863                            block_group->zone_is_active);
2864                 return;
2865         }
2866
2867         spin_lock(&ctl->tree_lock);
2868         for (n = rb_first(&ctl->free_space_offset); n; n = rb_next(n)) {
2869                 info = rb_entry(n, struct btrfs_free_space, offset_index);
2870                 if (info->bytes >= bytes && !block_group->ro)
2871                         count++;
2872                 btrfs_crit(fs_info, "entry offset %llu, bytes %llu, bitmap %s",
2873                            info->offset, info->bytes,
2874                        (info->bitmap) ? "yes" : "no");
2875         }
2876         spin_unlock(&ctl->tree_lock);
2877         btrfs_info(fs_info, "block group has cluster?: %s",
2878                list_empty(&block_group->cluster_list) ? "no" : "yes");
2879         btrfs_info(fs_info,
2880                    "%d blocks of free space at or bigger than bytes is", count);
2881 }
2882
2883 void btrfs_init_free_space_ctl(struct btrfs_block_group *block_group,
2884                                struct btrfs_free_space_ctl *ctl)
2885 {
2886         struct btrfs_fs_info *fs_info = block_group->fs_info;
2887
2888         spin_lock_init(&ctl->tree_lock);
2889         ctl->unit = fs_info->sectorsize;
2890         ctl->start = block_group->start;
2891         ctl->block_group = block_group;
2892         ctl->op = &free_space_op;
2893         ctl->free_space_bytes = RB_ROOT_CACHED;
2894         INIT_LIST_HEAD(&ctl->trimming_ranges);
2895         mutex_init(&ctl->cache_writeout_mutex);
2896
2897         /*
2898          * we only want to have 32k of ram per block group for keeping
2899          * track of free space, and if we pass 1/2 of that we want to
2900          * start converting things over to using bitmaps
2901          */
2902         ctl->extents_thresh = (SZ_32K / 2) / sizeof(struct btrfs_free_space);
2903 }
2904
2905 /*
2906  * for a given cluster, put all of its extents back into the free
2907  * space cache.  If the block group passed doesn't match the block group
2908  * pointed to by the cluster, someone else raced in and freed the
2909  * cluster already.  In that case, we just return without changing anything
2910  */
2911 static void __btrfs_return_cluster_to_free_space(
2912                              struct btrfs_block_group *block_group,
2913                              struct btrfs_free_cluster *cluster)
2914 {
2915         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2916         struct btrfs_free_space *entry;
2917         struct rb_node *node;
2918
2919         spin_lock(&cluster->lock);
2920         if (cluster->block_group != block_group) {
2921                 spin_unlock(&cluster->lock);
2922                 return;
2923         }
2924
2925         cluster->block_group = NULL;
2926         cluster->window_start = 0;
2927         list_del_init(&cluster->block_group_list);
2928
2929         node = rb_first(&cluster->root);
2930         while (node) {
2931                 bool bitmap;
2932
2933                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
2934                 node = rb_next(&entry->offset_index);
2935                 rb_erase(&entry->offset_index, &cluster->root);
2936                 RB_CLEAR_NODE(&entry->offset_index);
2937
2938                 bitmap = (entry->bitmap != NULL);
2939                 if (!bitmap) {
2940                         /* Merging treats extents as if they were new */
2941                         if (!btrfs_free_space_trimmed(entry)) {
2942                                 ctl->discardable_extents[BTRFS_STAT_CURR]--;
2943                                 ctl->discardable_bytes[BTRFS_STAT_CURR] -=
2944                                         entry->bytes;
2945                         }
2946
2947                         try_merge_free_space(ctl, entry, false);
2948                         steal_from_bitmap(ctl, entry, false);
2949
2950                         /* As we insert directly, update these statistics */
2951                         if (!btrfs_free_space_trimmed(entry)) {
2952                                 ctl->discardable_extents[BTRFS_STAT_CURR]++;
2953                                 ctl->discardable_bytes[BTRFS_STAT_CURR] +=
2954                                         entry->bytes;
2955                         }
2956                 }
2957                 tree_insert_offset(&ctl->free_space_offset,
2958                                    entry->offset, &entry->offset_index, bitmap);
2959                 rb_add_cached(&entry->bytes_index, &ctl->free_space_bytes,
2960                               entry_less);
2961         }
2962         cluster->root = RB_ROOT;
2963         spin_unlock(&cluster->lock);
2964         btrfs_put_block_group(block_group);
2965 }
2966
2967 static void __btrfs_remove_free_space_cache_locked(
2968                                 struct btrfs_free_space_ctl *ctl)
2969 {
2970         struct btrfs_free_space *info;
2971         struct rb_node *node;
2972
2973         while ((node = rb_last(&ctl->free_space_offset)) != NULL) {
2974                 info = rb_entry(node, struct btrfs_free_space, offset_index);
2975                 if (!info->bitmap) {
2976                         unlink_free_space(ctl, info, true);
2977                         kmem_cache_free(btrfs_free_space_cachep, info);
2978                 } else {
2979                         free_bitmap(ctl, info);
2980                 }
2981
2982                 cond_resched_lock(&ctl->tree_lock);
2983         }
2984 }
2985
2986 void __btrfs_remove_free_space_cache(struct btrfs_free_space_ctl *ctl)
2987 {
2988         spin_lock(&ctl->tree_lock);
2989         __btrfs_remove_free_space_cache_locked(ctl);
2990         if (ctl->block_group)
2991                 btrfs_discard_update_discardable(ctl->block_group);
2992         spin_unlock(&ctl->tree_lock);
2993 }
2994
2995 void btrfs_remove_free_space_cache(struct btrfs_block_group *block_group)
2996 {
2997         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
2998         struct btrfs_free_cluster *cluster;
2999         struct list_head *head;
3000
3001         spin_lock(&ctl->tree_lock);
3002         while ((head = block_group->cluster_list.next) !=
3003                &block_group->cluster_list) {
3004                 cluster = list_entry(head, struct btrfs_free_cluster,
3005                                      block_group_list);
3006
3007                 WARN_ON(cluster->block_group != block_group);
3008                 __btrfs_return_cluster_to_free_space(block_group, cluster);
3009
3010                 cond_resched_lock(&ctl->tree_lock);
3011         }
3012         __btrfs_remove_free_space_cache_locked(ctl);
3013         btrfs_discard_update_discardable(block_group);
3014         spin_unlock(&ctl->tree_lock);
3015
3016 }
3017
3018 /**
3019  * btrfs_is_free_space_trimmed - see if everything is trimmed
3020  * @block_group: block_group of interest
3021  *
3022  * Walk @block_group's free space rb_tree to determine if everything is trimmed.
3023  */
3024 bool btrfs_is_free_space_trimmed(struct btrfs_block_group *block_group)
3025 {
3026         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3027         struct btrfs_free_space *info;
3028         struct rb_node *node;
3029         bool ret = true;
3030
3031         spin_lock(&ctl->tree_lock);
3032         node = rb_first(&ctl->free_space_offset);
3033
3034         while (node) {
3035                 info = rb_entry(node, struct btrfs_free_space, offset_index);
3036
3037                 if (!btrfs_free_space_trimmed(info)) {
3038                         ret = false;
3039                         break;
3040                 }
3041
3042                 node = rb_next(node);
3043         }
3044
3045         spin_unlock(&ctl->tree_lock);
3046         return ret;
3047 }
3048
3049 u64 btrfs_find_space_for_alloc(struct btrfs_block_group *block_group,
3050                                u64 offset, u64 bytes, u64 empty_size,
3051                                u64 *max_extent_size)
3052 {
3053         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3054         struct btrfs_discard_ctl *discard_ctl =
3055                                         &block_group->fs_info->discard_ctl;
3056         struct btrfs_free_space *entry = NULL;
3057         u64 bytes_search = bytes + empty_size;
3058         u64 ret = 0;
3059         u64 align_gap = 0;
3060         u64 align_gap_len = 0;
3061         enum btrfs_trim_state align_gap_trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3062         bool use_bytes_index = (offset == block_group->start);
3063
3064         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3065
3066         spin_lock(&ctl->tree_lock);
3067         entry = find_free_space(ctl, &offset, &bytes_search,
3068                                 block_group->full_stripe_len, max_extent_size,
3069                                 use_bytes_index);
3070         if (!entry)
3071                 goto out;
3072
3073         ret = offset;
3074         if (entry->bitmap) {
3075                 bitmap_clear_bits(ctl, entry, offset, bytes, true);
3076
3077                 if (!btrfs_free_space_trimmed(entry))
3078                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3079
3080                 if (!entry->bytes)
3081                         free_bitmap(ctl, entry);
3082         } else {
3083                 unlink_free_space(ctl, entry, true);
3084                 align_gap_len = offset - entry->offset;
3085                 align_gap = entry->offset;
3086                 align_gap_trim_state = entry->trim_state;
3087
3088                 if (!btrfs_free_space_trimmed(entry))
3089                         atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3090
3091                 entry->offset = offset + bytes;
3092                 WARN_ON(entry->bytes < bytes + align_gap_len);
3093
3094                 entry->bytes -= bytes + align_gap_len;
3095                 if (!entry->bytes)
3096                         kmem_cache_free(btrfs_free_space_cachep, entry);
3097                 else
3098                         link_free_space(ctl, entry);
3099         }
3100 out:
3101         btrfs_discard_update_discardable(block_group);
3102         spin_unlock(&ctl->tree_lock);
3103
3104         if (align_gap_len)
3105                 __btrfs_add_free_space(block_group, align_gap, align_gap_len,
3106                                        align_gap_trim_state);
3107         return ret;
3108 }
3109
3110 /*
3111  * given a cluster, put all of its extents back into the free space
3112  * cache.  If a block group is passed, this function will only free
3113  * a cluster that belongs to the passed block group.
3114  *
3115  * Otherwise, it'll get a reference on the block group pointed to by the
3116  * cluster and remove the cluster from it.
3117  */
3118 void btrfs_return_cluster_to_free_space(
3119                                struct btrfs_block_group *block_group,
3120                                struct btrfs_free_cluster *cluster)
3121 {
3122         struct btrfs_free_space_ctl *ctl;
3123
3124         /* first, get a safe pointer to the block group */
3125         spin_lock(&cluster->lock);
3126         if (!block_group) {
3127                 block_group = cluster->block_group;
3128                 if (!block_group) {
3129                         spin_unlock(&cluster->lock);
3130                         return;
3131                 }
3132         } else if (cluster->block_group != block_group) {
3133                 /* someone else has already freed it don't redo their work */
3134                 spin_unlock(&cluster->lock);
3135                 return;
3136         }
3137         btrfs_get_block_group(block_group);
3138         spin_unlock(&cluster->lock);
3139
3140         ctl = block_group->free_space_ctl;
3141
3142         /* now return any extents the cluster had on it */
3143         spin_lock(&ctl->tree_lock);
3144         __btrfs_return_cluster_to_free_space(block_group, cluster);
3145         spin_unlock(&ctl->tree_lock);
3146
3147         btrfs_discard_queue_work(&block_group->fs_info->discard_ctl, block_group);
3148
3149         /* finally drop our ref */
3150         btrfs_put_block_group(block_group);
3151 }
3152
3153 static u64 btrfs_alloc_from_bitmap(struct btrfs_block_group *block_group,
3154                                    struct btrfs_free_cluster *cluster,
3155                                    struct btrfs_free_space *entry,
3156                                    u64 bytes, u64 min_start,
3157                                    u64 *max_extent_size)
3158 {
3159         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3160         int err;
3161         u64 search_start = cluster->window_start;
3162         u64 search_bytes = bytes;
3163         u64 ret = 0;
3164
3165         search_start = min_start;
3166         search_bytes = bytes;
3167
3168         err = search_bitmap(ctl, entry, &search_start, &search_bytes, true);
3169         if (err) {
3170                 *max_extent_size = max(get_max_extent_size(entry),
3171                                        *max_extent_size);
3172                 return 0;
3173         }
3174
3175         ret = search_start;
3176         bitmap_clear_bits(ctl, entry, ret, bytes, false);
3177
3178         return ret;
3179 }
3180
3181 /*
3182  * given a cluster, try to allocate 'bytes' from it, returns 0
3183  * if it couldn't find anything suitably large, or a logical disk offset
3184  * if things worked out
3185  */
3186 u64 btrfs_alloc_from_cluster(struct btrfs_block_group *block_group,
3187                              struct btrfs_free_cluster *cluster, u64 bytes,
3188                              u64 min_start, u64 *max_extent_size)
3189 {
3190         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3191         struct btrfs_discard_ctl *discard_ctl =
3192                                         &block_group->fs_info->discard_ctl;
3193         struct btrfs_free_space *entry = NULL;
3194         struct rb_node *node;
3195         u64 ret = 0;
3196
3197         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3198
3199         spin_lock(&cluster->lock);
3200         if (bytes > cluster->max_size)
3201                 goto out;
3202
3203         if (cluster->block_group != block_group)
3204                 goto out;
3205
3206         node = rb_first(&cluster->root);
3207         if (!node)
3208                 goto out;
3209
3210         entry = rb_entry(node, struct btrfs_free_space, offset_index);
3211         while (1) {
3212                 if (entry->bytes < bytes)
3213                         *max_extent_size = max(get_max_extent_size(entry),
3214                                                *max_extent_size);
3215
3216                 if (entry->bytes < bytes ||
3217                     (!entry->bitmap && entry->offset < min_start)) {
3218                         node = rb_next(&entry->offset_index);
3219                         if (!node)
3220                                 break;
3221                         entry = rb_entry(node, struct btrfs_free_space,
3222                                          offset_index);
3223                         continue;
3224                 }
3225
3226                 if (entry->bitmap) {
3227                         ret = btrfs_alloc_from_bitmap(block_group,
3228                                                       cluster, entry, bytes,
3229                                                       cluster->window_start,
3230                                                       max_extent_size);
3231                         if (ret == 0) {
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                         cluster->window_start += bytes;
3240                 } else {
3241                         ret = entry->offset;
3242
3243                         entry->offset += bytes;
3244                         entry->bytes -= bytes;
3245                 }
3246
3247                 break;
3248         }
3249 out:
3250         spin_unlock(&cluster->lock);
3251
3252         if (!ret)
3253                 return 0;
3254
3255         spin_lock(&ctl->tree_lock);
3256
3257         if (!btrfs_free_space_trimmed(entry))
3258                 atomic64_add(bytes, &discard_ctl->discard_bytes_saved);
3259
3260         ctl->free_space -= bytes;
3261         if (!entry->bitmap && !btrfs_free_space_trimmed(entry))
3262                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= bytes;
3263
3264         spin_lock(&cluster->lock);
3265         if (entry->bytes == 0) {
3266                 rb_erase(&entry->offset_index, &cluster->root);
3267                 ctl->free_extents--;
3268                 if (entry->bitmap) {
3269                         kmem_cache_free(btrfs_free_space_bitmap_cachep,
3270                                         entry->bitmap);
3271                         ctl->total_bitmaps--;
3272                         recalculate_thresholds(ctl);
3273                 } else if (!btrfs_free_space_trimmed(entry)) {
3274                         ctl->discardable_extents[BTRFS_STAT_CURR]--;
3275                 }
3276                 kmem_cache_free(btrfs_free_space_cachep, entry);
3277         }
3278
3279         spin_unlock(&cluster->lock);
3280         spin_unlock(&ctl->tree_lock);
3281
3282         return ret;
3283 }
3284
3285 static int btrfs_bitmap_cluster(struct btrfs_block_group *block_group,
3286                                 struct btrfs_free_space *entry,
3287                                 struct btrfs_free_cluster *cluster,
3288                                 u64 offset, u64 bytes,
3289                                 u64 cont1_bytes, u64 min_bytes)
3290 {
3291         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3292         unsigned long next_zero;
3293         unsigned long i;
3294         unsigned long want_bits;
3295         unsigned long min_bits;
3296         unsigned long found_bits;
3297         unsigned long max_bits = 0;
3298         unsigned long start = 0;
3299         unsigned long total_found = 0;
3300         int ret;
3301
3302         i = offset_to_bit(entry->offset, ctl->unit,
3303                           max_t(u64, offset, entry->offset));
3304         want_bits = bytes_to_bits(bytes, ctl->unit);
3305         min_bits = bytes_to_bits(min_bytes, ctl->unit);
3306
3307         /*
3308          * Don't bother looking for a cluster in this bitmap if it's heavily
3309          * fragmented.
3310          */
3311         if (entry->max_extent_size &&
3312             entry->max_extent_size < cont1_bytes)
3313                 return -ENOSPC;
3314 again:
3315         found_bits = 0;
3316         for_each_set_bit_from(i, entry->bitmap, BITS_PER_BITMAP) {
3317                 next_zero = find_next_zero_bit(entry->bitmap,
3318                                                BITS_PER_BITMAP, i);
3319                 if (next_zero - i >= min_bits) {
3320                         found_bits = next_zero - i;
3321                         if (found_bits > max_bits)
3322                                 max_bits = found_bits;
3323                         break;
3324                 }
3325                 if (next_zero - i > max_bits)
3326                         max_bits = next_zero - i;
3327                 i = next_zero;
3328         }
3329
3330         if (!found_bits) {
3331                 entry->max_extent_size = (u64)max_bits * ctl->unit;
3332                 return -ENOSPC;
3333         }
3334
3335         if (!total_found) {
3336                 start = i;
3337                 cluster->max_size = 0;
3338         }
3339
3340         total_found += found_bits;
3341
3342         if (cluster->max_size < found_bits * ctl->unit)
3343                 cluster->max_size = found_bits * ctl->unit;
3344
3345         if (total_found < want_bits || cluster->max_size < cont1_bytes) {
3346                 i = next_zero + 1;
3347                 goto again;
3348         }
3349
3350         cluster->window_start = start * ctl->unit + entry->offset;
3351         rb_erase(&entry->offset_index, &ctl->free_space_offset);
3352         rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3353
3354         /*
3355          * We need to know if we're currently on the normal space index when we
3356          * manipulate the bitmap so that we know we need to remove and re-insert
3357          * it into the space_index tree.  Clear the bytes_index node here so the
3358          * bitmap manipulation helpers know not to mess with the space_index
3359          * until this bitmap entry is added back into the normal cache.
3360          */
3361         RB_CLEAR_NODE(&entry->bytes_index);
3362
3363         ret = tree_insert_offset(&cluster->root, entry->offset,
3364                                  &entry->offset_index, 1);
3365         ASSERT(!ret); /* -EEXIST; Logic error */
3366
3367         trace_btrfs_setup_cluster(block_group, cluster,
3368                                   total_found * ctl->unit, 1);
3369         return 0;
3370 }
3371
3372 /*
3373  * This searches the block group for just extents to fill the cluster with.
3374  * Try to find a cluster with at least bytes total bytes, at least one
3375  * extent of cont1_bytes, and other clusters of at least min_bytes.
3376  */
3377 static noinline int
3378 setup_cluster_no_bitmap(struct btrfs_block_group *block_group,
3379                         struct btrfs_free_cluster *cluster,
3380                         struct list_head *bitmaps, u64 offset, u64 bytes,
3381                         u64 cont1_bytes, u64 min_bytes)
3382 {
3383         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3384         struct btrfs_free_space *first = NULL;
3385         struct btrfs_free_space *entry = NULL;
3386         struct btrfs_free_space *last;
3387         struct rb_node *node;
3388         u64 window_free;
3389         u64 max_extent;
3390         u64 total_size = 0;
3391
3392         entry = tree_search_offset(ctl, offset, 0, 1);
3393         if (!entry)
3394                 return -ENOSPC;
3395
3396         /*
3397          * We don't want bitmaps, so just move along until we find a normal
3398          * extent entry.
3399          */
3400         while (entry->bitmap || entry->bytes < min_bytes) {
3401                 if (entry->bitmap && list_empty(&entry->list))
3402                         list_add_tail(&entry->list, bitmaps);
3403                 node = rb_next(&entry->offset_index);
3404                 if (!node)
3405                         return -ENOSPC;
3406                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3407         }
3408
3409         window_free = entry->bytes;
3410         max_extent = entry->bytes;
3411         first = entry;
3412         last = entry;
3413
3414         for (node = rb_next(&entry->offset_index); node;
3415              node = rb_next(&entry->offset_index)) {
3416                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3417
3418                 if (entry->bitmap) {
3419                         if (list_empty(&entry->list))
3420                                 list_add_tail(&entry->list, bitmaps);
3421                         continue;
3422                 }
3423
3424                 if (entry->bytes < min_bytes)
3425                         continue;
3426
3427                 last = entry;
3428                 window_free += entry->bytes;
3429                 if (entry->bytes > max_extent)
3430                         max_extent = entry->bytes;
3431         }
3432
3433         if (window_free < bytes || max_extent < cont1_bytes)
3434                 return -ENOSPC;
3435
3436         cluster->window_start = first->offset;
3437
3438         node = &first->offset_index;
3439
3440         /*
3441          * now we've found our entries, pull them out of the free space
3442          * cache and put them into the cluster rbtree
3443          */
3444         do {
3445                 int ret;
3446
3447                 entry = rb_entry(node, struct btrfs_free_space, offset_index);
3448                 node = rb_next(&entry->offset_index);
3449                 if (entry->bitmap || entry->bytes < min_bytes)
3450                         continue;
3451
3452                 rb_erase(&entry->offset_index, &ctl->free_space_offset);
3453                 rb_erase_cached(&entry->bytes_index, &ctl->free_space_bytes);
3454                 ret = tree_insert_offset(&cluster->root, entry->offset,
3455                                          &entry->offset_index, 0);
3456                 total_size += entry->bytes;
3457                 ASSERT(!ret); /* -EEXIST; Logic error */
3458         } while (node && entry != last);
3459
3460         cluster->max_size = max_extent;
3461         trace_btrfs_setup_cluster(block_group, cluster, total_size, 0);
3462         return 0;
3463 }
3464
3465 /*
3466  * This specifically looks for bitmaps that may work in the cluster, we assume
3467  * that we have already failed to find extents that will work.
3468  */
3469 static noinline int
3470 setup_cluster_bitmap(struct btrfs_block_group *block_group,
3471                      struct btrfs_free_cluster *cluster,
3472                      struct list_head *bitmaps, u64 offset, u64 bytes,
3473                      u64 cont1_bytes, u64 min_bytes)
3474 {
3475         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3476         struct btrfs_free_space *entry = NULL;
3477         int ret = -ENOSPC;
3478         u64 bitmap_offset = offset_to_bitmap(ctl, offset);
3479
3480         if (ctl->total_bitmaps == 0)
3481                 return -ENOSPC;
3482
3483         /*
3484          * The bitmap that covers offset won't be in the list unless offset
3485          * is just its start offset.
3486          */
3487         if (!list_empty(bitmaps))
3488                 entry = list_first_entry(bitmaps, struct btrfs_free_space, list);
3489
3490         if (!entry || entry->offset != bitmap_offset) {
3491                 entry = tree_search_offset(ctl, bitmap_offset, 1, 0);
3492                 if (entry && list_empty(&entry->list))
3493                         list_add(&entry->list, bitmaps);
3494         }
3495
3496         list_for_each_entry(entry, bitmaps, list) {
3497                 if (entry->bytes < bytes)
3498                         continue;
3499                 ret = btrfs_bitmap_cluster(block_group, entry, cluster, offset,
3500                                            bytes, cont1_bytes, min_bytes);
3501                 if (!ret)
3502                         return 0;
3503         }
3504
3505         /*
3506          * The bitmaps list has all the bitmaps that record free space
3507          * starting after offset, so no more search is required.
3508          */
3509         return -ENOSPC;
3510 }
3511
3512 /*
3513  * here we try to find a cluster of blocks in a block group.  The goal
3514  * is to find at least bytes+empty_size.
3515  * We might not find them all in one contiguous area.
3516  *
3517  * returns zero and sets up cluster if things worked out, otherwise
3518  * it returns -enospc
3519  */
3520 int btrfs_find_space_cluster(struct btrfs_block_group *block_group,
3521                              struct btrfs_free_cluster *cluster,
3522                              u64 offset, u64 bytes, u64 empty_size)
3523 {
3524         struct btrfs_fs_info *fs_info = block_group->fs_info;
3525         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3526         struct btrfs_free_space *entry, *tmp;
3527         LIST_HEAD(bitmaps);
3528         u64 min_bytes;
3529         u64 cont1_bytes;
3530         int ret;
3531
3532         /*
3533          * Choose the minimum extent size we'll require for this
3534          * cluster.  For SSD_SPREAD, don't allow any fragmentation.
3535          * For metadata, allow allocates with smaller extents.  For
3536          * data, keep it dense.
3537          */
3538         if (btrfs_test_opt(fs_info, SSD_SPREAD)) {
3539                 cont1_bytes = min_bytes = bytes + empty_size;
3540         } else if (block_group->flags & BTRFS_BLOCK_GROUP_METADATA) {
3541                 cont1_bytes = bytes;
3542                 min_bytes = fs_info->sectorsize;
3543         } else {
3544                 cont1_bytes = max(bytes, (bytes + empty_size) >> 2);
3545                 min_bytes = fs_info->sectorsize;
3546         }
3547
3548         spin_lock(&ctl->tree_lock);
3549
3550         /*
3551          * If we know we don't have enough space to make a cluster don't even
3552          * bother doing all the work to try and find one.
3553          */
3554         if (ctl->free_space < bytes) {
3555                 spin_unlock(&ctl->tree_lock);
3556                 return -ENOSPC;
3557         }
3558
3559         spin_lock(&cluster->lock);
3560
3561         /* someone already found a cluster, hooray */
3562         if (cluster->block_group) {
3563                 ret = 0;
3564                 goto out;
3565         }
3566
3567         trace_btrfs_find_cluster(block_group, offset, bytes, empty_size,
3568                                  min_bytes);
3569
3570         ret = setup_cluster_no_bitmap(block_group, cluster, &bitmaps, offset,
3571                                       bytes + empty_size,
3572                                       cont1_bytes, min_bytes);
3573         if (ret)
3574                 ret = setup_cluster_bitmap(block_group, cluster, &bitmaps,
3575                                            offset, bytes + empty_size,
3576                                            cont1_bytes, min_bytes);
3577
3578         /* Clear our temporary list */
3579         list_for_each_entry_safe(entry, tmp, &bitmaps, list)
3580                 list_del_init(&entry->list);
3581
3582         if (!ret) {
3583                 btrfs_get_block_group(block_group);
3584                 list_add_tail(&cluster->block_group_list,
3585                               &block_group->cluster_list);
3586                 cluster->block_group = block_group;
3587         } else {
3588                 trace_btrfs_failed_cluster_setup(block_group);
3589         }
3590 out:
3591         spin_unlock(&cluster->lock);
3592         spin_unlock(&ctl->tree_lock);
3593
3594         return ret;
3595 }
3596
3597 /*
3598  * simple code to zero out a cluster
3599  */
3600 void btrfs_init_free_cluster(struct btrfs_free_cluster *cluster)
3601 {
3602         spin_lock_init(&cluster->lock);
3603         spin_lock_init(&cluster->refill_lock);
3604         cluster->root = RB_ROOT;
3605         cluster->max_size = 0;
3606         cluster->fragmented = false;
3607         INIT_LIST_HEAD(&cluster->block_group_list);
3608         cluster->block_group = NULL;
3609 }
3610
3611 static int do_trimming(struct btrfs_block_group *block_group,
3612                        u64 *total_trimmed, u64 start, u64 bytes,
3613                        u64 reserved_start, u64 reserved_bytes,
3614                        enum btrfs_trim_state reserved_trim_state,
3615                        struct btrfs_trim_range *trim_entry)
3616 {
3617         struct btrfs_space_info *space_info = block_group->space_info;
3618         struct btrfs_fs_info *fs_info = block_group->fs_info;
3619         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3620         int ret;
3621         int update = 0;
3622         const u64 end = start + bytes;
3623         const u64 reserved_end = reserved_start + reserved_bytes;
3624         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3625         u64 trimmed = 0;
3626
3627         spin_lock(&space_info->lock);
3628         spin_lock(&block_group->lock);
3629         if (!block_group->ro) {
3630                 block_group->reserved += reserved_bytes;
3631                 space_info->bytes_reserved += reserved_bytes;
3632                 update = 1;
3633         }
3634         spin_unlock(&block_group->lock);
3635         spin_unlock(&space_info->lock);
3636
3637         ret = btrfs_discard_extent(fs_info, start, bytes, &trimmed);
3638         if (!ret) {
3639                 *total_trimmed += trimmed;
3640                 trim_state = BTRFS_TRIM_STATE_TRIMMED;
3641         }
3642
3643         mutex_lock(&ctl->cache_writeout_mutex);
3644         if (reserved_start < start)
3645                 __btrfs_add_free_space(block_group, reserved_start,
3646                                        start - reserved_start,
3647                                        reserved_trim_state);
3648         if (start + bytes < reserved_start + reserved_bytes)
3649                 __btrfs_add_free_space(block_group, end, reserved_end - end,
3650                                        reserved_trim_state);
3651         __btrfs_add_free_space(block_group, start, bytes, trim_state);
3652         list_del(&trim_entry->list);
3653         mutex_unlock(&ctl->cache_writeout_mutex);
3654
3655         if (update) {
3656                 spin_lock(&space_info->lock);
3657                 spin_lock(&block_group->lock);
3658                 if (block_group->ro)
3659                         space_info->bytes_readonly += reserved_bytes;
3660                 block_group->reserved -= reserved_bytes;
3661                 space_info->bytes_reserved -= reserved_bytes;
3662                 spin_unlock(&block_group->lock);
3663                 spin_unlock(&space_info->lock);
3664         }
3665
3666         return ret;
3667 }
3668
3669 /*
3670  * If @async is set, then we will trim 1 region and return.
3671  */
3672 static int trim_no_bitmap(struct btrfs_block_group *block_group,
3673                           u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3674                           bool async)
3675 {
3676         struct btrfs_discard_ctl *discard_ctl =
3677                                         &block_group->fs_info->discard_ctl;
3678         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3679         struct btrfs_free_space *entry;
3680         struct rb_node *node;
3681         int ret = 0;
3682         u64 extent_start;
3683         u64 extent_bytes;
3684         enum btrfs_trim_state extent_trim_state;
3685         u64 bytes;
3686         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3687
3688         while (start < end) {
3689                 struct btrfs_trim_range trim_entry;
3690
3691                 mutex_lock(&ctl->cache_writeout_mutex);
3692                 spin_lock(&ctl->tree_lock);
3693
3694                 if (ctl->free_space < minlen)
3695                         goto out_unlock;
3696
3697                 entry = tree_search_offset(ctl, start, 0, 1);
3698                 if (!entry)
3699                         goto out_unlock;
3700
3701                 /* Skip bitmaps and if async, already trimmed entries */
3702                 while (entry->bitmap ||
3703                        (async && btrfs_free_space_trimmed(entry))) {
3704                         node = rb_next(&entry->offset_index);
3705                         if (!node)
3706                                 goto out_unlock;
3707                         entry = rb_entry(node, struct btrfs_free_space,
3708                                          offset_index);
3709                 }
3710
3711                 if (entry->offset >= end)
3712                         goto out_unlock;
3713
3714                 extent_start = entry->offset;
3715                 extent_bytes = entry->bytes;
3716                 extent_trim_state = entry->trim_state;
3717                 if (async) {
3718                         start = entry->offset;
3719                         bytes = entry->bytes;
3720                         if (bytes < minlen) {
3721                                 spin_unlock(&ctl->tree_lock);
3722                                 mutex_unlock(&ctl->cache_writeout_mutex);
3723                                 goto next;
3724                         }
3725                         unlink_free_space(ctl, entry, true);
3726                         /*
3727                          * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3728                          * If X < BTRFS_ASYNC_DISCARD_MIN_FILTER, we won't trim
3729                          * X when we come back around.  So trim it now.
3730                          */
3731                         if (max_discard_size &&
3732                             bytes >= (max_discard_size +
3733                                       BTRFS_ASYNC_DISCARD_MIN_FILTER)) {
3734                                 bytes = max_discard_size;
3735                                 extent_bytes = max_discard_size;
3736                                 entry->offset += max_discard_size;
3737                                 entry->bytes -= max_discard_size;
3738                                 link_free_space(ctl, entry);
3739                         } else {
3740                                 kmem_cache_free(btrfs_free_space_cachep, entry);
3741                         }
3742                 } else {
3743                         start = max(start, extent_start);
3744                         bytes = min(extent_start + extent_bytes, end) - start;
3745                         if (bytes < minlen) {
3746                                 spin_unlock(&ctl->tree_lock);
3747                                 mutex_unlock(&ctl->cache_writeout_mutex);
3748                                 goto next;
3749                         }
3750
3751                         unlink_free_space(ctl, entry, true);
3752                         kmem_cache_free(btrfs_free_space_cachep, entry);
3753                 }
3754
3755                 spin_unlock(&ctl->tree_lock);
3756                 trim_entry.start = extent_start;
3757                 trim_entry.bytes = extent_bytes;
3758                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3759                 mutex_unlock(&ctl->cache_writeout_mutex);
3760
3761                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3762                                   extent_start, extent_bytes, extent_trim_state,
3763                                   &trim_entry);
3764                 if (ret) {
3765                         block_group->discard_cursor = start + bytes;
3766                         break;
3767                 }
3768 next:
3769                 start += bytes;
3770                 block_group->discard_cursor = start;
3771                 if (async && *total_trimmed)
3772                         break;
3773
3774                 if (fatal_signal_pending(current)) {
3775                         ret = -ERESTARTSYS;
3776                         break;
3777                 }
3778
3779                 cond_resched();
3780         }
3781
3782         return ret;
3783
3784 out_unlock:
3785         block_group->discard_cursor = btrfs_block_group_end(block_group);
3786         spin_unlock(&ctl->tree_lock);
3787         mutex_unlock(&ctl->cache_writeout_mutex);
3788
3789         return ret;
3790 }
3791
3792 /*
3793  * If we break out of trimming a bitmap prematurely, we should reset the
3794  * trimming bit.  In a rather contrieved case, it's possible to race here so
3795  * reset the state to BTRFS_TRIM_STATE_UNTRIMMED.
3796  *
3797  * start = start of bitmap
3798  * end = near end of bitmap
3799  *
3800  * Thread 1:                    Thread 2:
3801  * trim_bitmaps(start)
3802  *                              trim_bitmaps(end)
3803  *                              end_trimming_bitmap()
3804  * reset_trimming_bitmap()
3805  */
3806 static void reset_trimming_bitmap(struct btrfs_free_space_ctl *ctl, u64 offset)
3807 {
3808         struct btrfs_free_space *entry;
3809
3810         spin_lock(&ctl->tree_lock);
3811         entry = tree_search_offset(ctl, offset, 1, 0);
3812         if (entry) {
3813                 if (btrfs_free_space_trimmed(entry)) {
3814                         ctl->discardable_extents[BTRFS_STAT_CURR] +=
3815                                 entry->bitmap_extents;
3816                         ctl->discardable_bytes[BTRFS_STAT_CURR] += entry->bytes;
3817                 }
3818                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3819         }
3820
3821         spin_unlock(&ctl->tree_lock);
3822 }
3823
3824 static void end_trimming_bitmap(struct btrfs_free_space_ctl *ctl,
3825                                 struct btrfs_free_space *entry)
3826 {
3827         if (btrfs_free_space_trimming_bitmap(entry)) {
3828                 entry->trim_state = BTRFS_TRIM_STATE_TRIMMED;
3829                 ctl->discardable_extents[BTRFS_STAT_CURR] -=
3830                         entry->bitmap_extents;
3831                 ctl->discardable_bytes[BTRFS_STAT_CURR] -= entry->bytes;
3832         }
3833 }
3834
3835 /*
3836  * If @async is set, then we will trim 1 region and return.
3837  */
3838 static int trim_bitmaps(struct btrfs_block_group *block_group,
3839                         u64 *total_trimmed, u64 start, u64 end, u64 minlen,
3840                         u64 maxlen, bool async)
3841 {
3842         struct btrfs_discard_ctl *discard_ctl =
3843                                         &block_group->fs_info->discard_ctl;
3844         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3845         struct btrfs_free_space *entry;
3846         int ret = 0;
3847         int ret2;
3848         u64 bytes;
3849         u64 offset = offset_to_bitmap(ctl, start);
3850         const u64 max_discard_size = READ_ONCE(discard_ctl->max_discard_size);
3851
3852         while (offset < end) {
3853                 bool next_bitmap = false;
3854                 struct btrfs_trim_range trim_entry;
3855
3856                 mutex_lock(&ctl->cache_writeout_mutex);
3857                 spin_lock(&ctl->tree_lock);
3858
3859                 if (ctl->free_space < minlen) {
3860                         block_group->discard_cursor =
3861                                 btrfs_block_group_end(block_group);
3862                         spin_unlock(&ctl->tree_lock);
3863                         mutex_unlock(&ctl->cache_writeout_mutex);
3864                         break;
3865                 }
3866
3867                 entry = tree_search_offset(ctl, offset, 1, 0);
3868                 /*
3869                  * Bitmaps are marked trimmed lossily now to prevent constant
3870                  * discarding of the same bitmap (the reason why we are bound
3871                  * by the filters).  So, retrim the block group bitmaps when we
3872                  * are preparing to punt to the unused_bgs list.  This uses
3873                  * @minlen to determine if we are in BTRFS_DISCARD_INDEX_UNUSED
3874                  * which is the only discard index which sets minlen to 0.
3875                  */
3876                 if (!entry || (async && minlen && start == offset &&
3877                                btrfs_free_space_trimmed(entry))) {
3878                         spin_unlock(&ctl->tree_lock);
3879                         mutex_unlock(&ctl->cache_writeout_mutex);
3880                         next_bitmap = true;
3881                         goto next;
3882                 }
3883
3884                 /*
3885                  * Async discard bitmap trimming begins at by setting the start
3886                  * to be key.objectid and the offset_to_bitmap() aligns to the
3887                  * start of the bitmap.  This lets us know we are fully
3888                  * scanning the bitmap rather than only some portion of it.
3889                  */
3890                 if (start == offset)
3891                         entry->trim_state = BTRFS_TRIM_STATE_TRIMMING;
3892
3893                 bytes = minlen;
3894                 ret2 = search_bitmap(ctl, entry, &start, &bytes, false);
3895                 if (ret2 || start >= end) {
3896                         /*
3897                          * We lossily consider a bitmap trimmed if we only skip
3898                          * over regions <= BTRFS_ASYNC_DISCARD_MIN_FILTER.
3899                          */
3900                         if (ret2 && minlen <= BTRFS_ASYNC_DISCARD_MIN_FILTER)
3901                                 end_trimming_bitmap(ctl, entry);
3902                         else
3903                                 entry->trim_state = BTRFS_TRIM_STATE_UNTRIMMED;
3904                         spin_unlock(&ctl->tree_lock);
3905                         mutex_unlock(&ctl->cache_writeout_mutex);
3906                         next_bitmap = true;
3907                         goto next;
3908                 }
3909
3910                 /*
3911                  * We already trimmed a region, but are using the locking above
3912                  * to reset the trim_state.
3913                  */
3914                 if (async && *total_trimmed) {
3915                         spin_unlock(&ctl->tree_lock);
3916                         mutex_unlock(&ctl->cache_writeout_mutex);
3917                         goto out;
3918                 }
3919
3920                 bytes = min(bytes, end - start);
3921                 if (bytes < minlen || (async && maxlen && bytes > maxlen)) {
3922                         spin_unlock(&ctl->tree_lock);
3923                         mutex_unlock(&ctl->cache_writeout_mutex);
3924                         goto next;
3925                 }
3926
3927                 /*
3928                  * Let bytes = BTRFS_MAX_DISCARD_SIZE + X.
3929                  * If X < @minlen, we won't trim X when we come back around.
3930                  * So trim it now.  We differ here from trimming extents as we
3931                  * don't keep individual state per bit.
3932                  */
3933                 if (async &&
3934                     max_discard_size &&
3935                     bytes > (max_discard_size + minlen))
3936                         bytes = max_discard_size;
3937
3938                 bitmap_clear_bits(ctl, entry, start, bytes, true);
3939                 if (entry->bytes == 0)
3940                         free_bitmap(ctl, entry);
3941
3942                 spin_unlock(&ctl->tree_lock);
3943                 trim_entry.start = start;
3944                 trim_entry.bytes = bytes;
3945                 list_add_tail(&trim_entry.list, &ctl->trimming_ranges);
3946                 mutex_unlock(&ctl->cache_writeout_mutex);
3947
3948                 ret = do_trimming(block_group, total_trimmed, start, bytes,
3949                                   start, bytes, 0, &trim_entry);
3950                 if (ret) {
3951                         reset_trimming_bitmap(ctl, offset);
3952                         block_group->discard_cursor =
3953                                 btrfs_block_group_end(block_group);
3954                         break;
3955                 }
3956 next:
3957                 if (next_bitmap) {
3958                         offset += BITS_PER_BITMAP * ctl->unit;
3959                         start = offset;
3960                 } else {
3961                         start += bytes;
3962                 }
3963                 block_group->discard_cursor = start;
3964
3965                 if (fatal_signal_pending(current)) {
3966                         if (start != offset)
3967                                 reset_trimming_bitmap(ctl, offset);
3968                         ret = -ERESTARTSYS;
3969                         break;
3970                 }
3971
3972                 cond_resched();
3973         }
3974
3975         if (offset >= end)
3976                 block_group->discard_cursor = end;
3977
3978 out:
3979         return ret;
3980 }
3981
3982 int btrfs_trim_block_group(struct btrfs_block_group *block_group,
3983                            u64 *trimmed, u64 start, u64 end, u64 minlen)
3984 {
3985         struct btrfs_free_space_ctl *ctl = block_group->free_space_ctl;
3986         int ret;
3987         u64 rem = 0;
3988
3989         ASSERT(!btrfs_is_zoned(block_group->fs_info));
3990
3991         *trimmed = 0;
3992
3993         spin_lock(&block_group->lock);
3994         if (block_group->removed) {
3995                 spin_unlock(&block_group->lock);
3996                 return 0;
3997         }
3998         btrfs_freeze_block_group(block_group);
3999         spin_unlock(&block_group->lock);
4000
4001         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, false);
4002         if (ret)
4003                 goto out;
4004
4005         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, 0, false);
4006         div64_u64_rem(end, BITS_PER_BITMAP * ctl->unit, &rem);
4007         /* If we ended in the middle of a bitmap, reset the trimming flag */
4008         if (rem)
4009                 reset_trimming_bitmap(ctl, offset_to_bitmap(ctl, end));
4010 out:
4011         btrfs_unfreeze_block_group(block_group);
4012         return ret;
4013 }
4014
4015 int btrfs_trim_block_group_extents(struct btrfs_block_group *block_group,
4016                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4017                                    bool async)
4018 {
4019         int ret;
4020
4021         *trimmed = 0;
4022
4023         spin_lock(&block_group->lock);
4024         if (block_group->removed) {
4025                 spin_unlock(&block_group->lock);
4026                 return 0;
4027         }
4028         btrfs_freeze_block_group(block_group);
4029         spin_unlock(&block_group->lock);
4030
4031         ret = trim_no_bitmap(block_group, trimmed, start, end, minlen, async);
4032         btrfs_unfreeze_block_group(block_group);
4033
4034         return ret;
4035 }
4036
4037 int btrfs_trim_block_group_bitmaps(struct btrfs_block_group *block_group,
4038                                    u64 *trimmed, u64 start, u64 end, u64 minlen,
4039                                    u64 maxlen, bool async)
4040 {
4041         int ret;
4042
4043         *trimmed = 0;
4044
4045         spin_lock(&block_group->lock);
4046         if (block_group->removed) {
4047                 spin_unlock(&block_group->lock);
4048                 return 0;
4049         }
4050         btrfs_freeze_block_group(block_group);
4051         spin_unlock(&block_group->lock);
4052
4053         ret = trim_bitmaps(block_group, trimmed, start, end, minlen, maxlen,
4054                            async);
4055
4056         btrfs_unfreeze_block_group(block_group);
4057
4058         return ret;
4059 }
4060
4061 bool btrfs_free_space_cache_v1_active(struct btrfs_fs_info *fs_info)
4062 {
4063         return btrfs_super_cache_generation(fs_info->super_copy);
4064 }
4065
4066 static int cleanup_free_space_cache_v1(struct btrfs_fs_info *fs_info,
4067                                        struct btrfs_trans_handle *trans)
4068 {
4069         struct btrfs_block_group *block_group;
4070         struct rb_node *node;
4071         int ret = 0;
4072
4073         btrfs_info(fs_info, "cleaning free space cache v1");
4074
4075         node = rb_first_cached(&fs_info->block_group_cache_tree);
4076         while (node) {
4077                 block_group = rb_entry(node, struct btrfs_block_group, cache_node);
4078                 ret = btrfs_remove_free_space_inode(trans, NULL, block_group);
4079                 if (ret)
4080                         goto out;
4081                 node = rb_next(node);
4082         }
4083 out:
4084         return ret;
4085 }
4086
4087 int btrfs_set_free_space_cache_v1_active(struct btrfs_fs_info *fs_info, bool active)
4088 {
4089         struct btrfs_trans_handle *trans;
4090         int ret;
4091
4092         /*
4093          * update_super_roots will appropriately set or unset
4094          * super_copy->cache_generation based on SPACE_CACHE and
4095          * BTRFS_FS_CLEANUP_SPACE_CACHE_V1. For this reason, we need a
4096          * transaction commit whether we are enabling space cache v1 and don't
4097          * have any other work to do, or are disabling it and removing free
4098          * space inodes.
4099          */
4100         trans = btrfs_start_transaction(fs_info->tree_root, 0);
4101         if (IS_ERR(trans))
4102                 return PTR_ERR(trans);
4103
4104         if (!active) {
4105                 set_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4106                 ret = cleanup_free_space_cache_v1(fs_info, trans);
4107                 if (ret) {
4108                         btrfs_abort_transaction(trans, ret);
4109                         btrfs_end_transaction(trans);
4110                         goto out;
4111                 }
4112         }
4113
4114         ret = btrfs_commit_transaction(trans);
4115 out:
4116         clear_bit(BTRFS_FS_CLEANUP_SPACE_CACHE_V1, &fs_info->flags);
4117
4118         return ret;
4119 }
4120
4121 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
4122 /*
4123  * Use this if you need to make a bitmap or extent entry specifically, it
4124  * doesn't do any of the merging that add_free_space does, this acts a lot like
4125  * how the free space cache loading stuff works, so you can get really weird
4126  * configurations.
4127  */
4128 int test_add_free_space_entry(struct btrfs_block_group *cache,
4129                               u64 offset, u64 bytes, bool bitmap)
4130 {
4131         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4132         struct btrfs_free_space *info = NULL, *bitmap_info;
4133         void *map = NULL;
4134         enum btrfs_trim_state trim_state = BTRFS_TRIM_STATE_TRIMMED;
4135         u64 bytes_added;
4136         int ret;
4137
4138 again:
4139         if (!info) {
4140                 info = kmem_cache_zalloc(btrfs_free_space_cachep, GFP_NOFS);
4141                 if (!info)
4142                         return -ENOMEM;
4143         }
4144
4145         if (!bitmap) {
4146                 spin_lock(&ctl->tree_lock);
4147                 info->offset = offset;
4148                 info->bytes = bytes;
4149                 info->max_extent_size = 0;
4150                 ret = link_free_space(ctl, info);
4151                 spin_unlock(&ctl->tree_lock);
4152                 if (ret)
4153                         kmem_cache_free(btrfs_free_space_cachep, info);
4154                 return ret;
4155         }
4156
4157         if (!map) {
4158                 map = kmem_cache_zalloc(btrfs_free_space_bitmap_cachep, GFP_NOFS);
4159                 if (!map) {
4160                         kmem_cache_free(btrfs_free_space_cachep, info);
4161                         return -ENOMEM;
4162                 }
4163         }
4164
4165         spin_lock(&ctl->tree_lock);
4166         bitmap_info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4167                                          1, 0);
4168         if (!bitmap_info) {
4169                 info->bitmap = map;
4170                 map = NULL;
4171                 add_new_bitmap(ctl, info, offset);
4172                 bitmap_info = info;
4173                 info = NULL;
4174         }
4175
4176         bytes_added = add_bytes_to_bitmap(ctl, bitmap_info, offset, bytes,
4177                                           trim_state);
4178
4179         bytes -= bytes_added;
4180         offset += bytes_added;
4181         spin_unlock(&ctl->tree_lock);
4182
4183         if (bytes)
4184                 goto again;
4185
4186         if (info)
4187                 kmem_cache_free(btrfs_free_space_cachep, info);
4188         if (map)
4189                 kmem_cache_free(btrfs_free_space_bitmap_cachep, map);
4190         return 0;
4191 }
4192
4193 /*
4194  * Checks to see if the given range is in the free space cache.  This is really
4195  * just used to check the absence of space, so if there is free space in the
4196  * range at all we will return 1.
4197  */
4198 int test_check_exists(struct btrfs_block_group *cache,
4199                       u64 offset, u64 bytes)
4200 {
4201         struct btrfs_free_space_ctl *ctl = cache->free_space_ctl;
4202         struct btrfs_free_space *info;
4203         int ret = 0;
4204
4205         spin_lock(&ctl->tree_lock);
4206         info = tree_search_offset(ctl, offset, 0, 0);
4207         if (!info) {
4208                 info = tree_search_offset(ctl, offset_to_bitmap(ctl, offset),
4209                                           1, 0);
4210                 if (!info)
4211                         goto out;
4212         }
4213
4214 have_info:
4215         if (info->bitmap) {
4216                 u64 bit_off, bit_bytes;
4217                 struct rb_node *n;
4218                 struct btrfs_free_space *tmp;
4219
4220                 bit_off = offset;
4221                 bit_bytes = ctl->unit;
4222                 ret = search_bitmap(ctl, info, &bit_off, &bit_bytes, false);
4223                 if (!ret) {
4224                         if (bit_off == offset) {
4225                                 ret = 1;
4226                                 goto out;
4227                         } else if (bit_off > offset &&
4228                                    offset + bytes > bit_off) {
4229                                 ret = 1;
4230                                 goto out;
4231                         }
4232                 }
4233
4234                 n = rb_prev(&info->offset_index);
4235                 while (n) {
4236                         tmp = rb_entry(n, struct btrfs_free_space,
4237                                        offset_index);
4238                         if (tmp->offset + tmp->bytes < offset)
4239                                 break;
4240                         if (offset + bytes < tmp->offset) {
4241                                 n = rb_prev(&tmp->offset_index);
4242                                 continue;
4243                         }
4244                         info = tmp;
4245                         goto have_info;
4246                 }
4247
4248                 n = rb_next(&info->offset_index);
4249                 while (n) {
4250                         tmp = rb_entry(n, struct btrfs_free_space,
4251                                        offset_index);
4252                         if (offset + bytes < tmp->offset)
4253                                 break;
4254                         if (tmp->offset + tmp->bytes < offset) {
4255                                 n = rb_next(&tmp->offset_index);
4256                                 continue;
4257                         }
4258                         info = tmp;
4259                         goto have_info;
4260                 }
4261
4262                 ret = 0;
4263                 goto out;
4264         }
4265
4266         if (info->offset == offset) {
4267                 ret = 1;
4268                 goto out;
4269         }
4270
4271         if (offset > info->offset && offset < info->offset + info->bytes)
4272                 ret = 1;
4273 out:
4274         spin_unlock(&ctl->tree_lock);
4275         return ret;
4276 }
4277 #endif /* CONFIG_BTRFS_FS_RUN_SANITY_TESTS */