1 // SPDX-License-Identifier: GPL-2.0 2 3 #include "misc.h" 4 #include "ctree.h" 5 #include "space-info.h" 6 #include "sysfs.h" 7 #include "volumes.h" 8 #include "free-space-cache.h" 9 #include "ordered-data.h" 10 #include "transaction.h" 11 #include "block-group.h" 12 13 /* 14 * HOW DOES SPACE RESERVATION WORK 15 * 16 * If you want to know about delalloc specifically, there is a separate comment 17 * for that with the delalloc code. This comment is about how the whole system 18 * works generally. 19 * 20 * BASIC CONCEPTS 21 * 22 * 1) space_info. This is the ultimate arbiter of how much space we can use. 23 * There's a description of the bytes_ fields with the struct declaration, 24 * refer to that for specifics on each field. Suffice it to say that for 25 * reservations we care about total_bytes - SUM(space_info->bytes_) when 26 * determining if there is space to make an allocation. There is a space_info 27 * for METADATA, SYSTEM, and DATA areas. 28 * 29 * 2) block_rsv's. These are basically buckets for every different type of 30 * metadata reservation we have. You can see the comment in the block_rsv 31 * code on the rules for each type, but generally block_rsv->reserved is how 32 * much space is accounted for in space_info->bytes_may_use. 33 * 34 * 3) btrfs_calc*_size. These are the worst case calculations we used based 35 * on the number of items we will want to modify. We have one for changing 36 * items, and one for inserting new items. Generally we use these helpers to 37 * determine the size of the block reserves, and then use the actual bytes 38 * values to adjust the space_info counters. 39 * 40 * MAKING RESERVATIONS, THE NORMAL CASE 41 * 42 * We call into either btrfs_reserve_data_bytes() or 43 * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with 44 * num_bytes we want to reserve. 45 * 46 * ->reserve 47 * space_info->bytes_may_reserve += num_bytes 48 * 49 * ->extent allocation 50 * Call btrfs_add_reserved_bytes() which does 51 * space_info->bytes_may_reserve -= num_bytes 52 * space_info->bytes_reserved += extent_bytes 53 * 54 * ->insert reference 55 * Call btrfs_update_block_group() which does 56 * space_info->bytes_reserved -= extent_bytes 57 * space_info->bytes_used += extent_bytes 58 * 59 * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) 60 * 61 * Assume we are unable to simply make the reservation because we do not have 62 * enough space 63 * 64 * -> __reserve_bytes 65 * create a reserve_ticket with ->bytes set to our reservation, add it to 66 * the tail of space_info->tickets, kick async flush thread 67 * 68 * ->handle_reserve_ticket 69 * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set 70 * on the ticket. 71 * 72 * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space 73 * Flushes various things attempting to free up space. 74 * 75 * -> btrfs_try_granting_tickets() 76 * This is called by anything that either subtracts space from 77 * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the 78 * space_info->total_bytes. This loops through the ->priority_tickets and 79 * then the ->tickets list checking to see if the reservation can be 80 * completed. If it can the space is added to space_info->bytes_may_use and 81 * the ticket is woken up. 82 * 83 * -> ticket wakeup 84 * Check if ->bytes == 0, if it does we got our reservation and we can carry 85 * on, if not return the appropriate error (ENOSPC, but can be EINTR if we 86 * were interrupted.) 87 * 88 * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY 89 * 90 * Same as the above, except we add ourselves to the 91 * space_info->priority_tickets, and we do not use ticket->wait, we simply 92 * call flush_space() ourselves for the states that are safe for us to call 93 * without deadlocking and hope for the best. 94 * 95 * THE FLUSHING STATES 96 * 97 * Generally speaking we will have two cases for each state, a "nice" state 98 * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to 99 * reduce the locking over head on the various trees, and even to keep from 100 * doing any work at all in the case of delayed refs. Each of these delayed 101 * things however hold reservations, and so letting them run allows us to 102 * reclaim space so we can make new reservations. 103 * 104 * FLUSH_DELAYED_ITEMS 105 * Every inode has a delayed item to update the inode. Take a simple write 106 * for example, we would update the inode item at write time to update the 107 * mtime, and then again at finish_ordered_io() time in order to update the 108 * isize or bytes. We keep these delayed items to coalesce these operations 109 * into a single operation done on demand. These are an easy way to reclaim 110 * metadata space. 111 * 112 * FLUSH_DELALLOC 113 * Look at the delalloc comment to get an idea of how much space is reserved 114 * for delayed allocation. We can reclaim some of this space simply by 115 * running delalloc, but usually we need to wait for ordered extents to 116 * reclaim the bulk of this space. 117 * 118 * FLUSH_DELAYED_REFS 119 * We have a block reserve for the outstanding delayed refs space, and every 120 * delayed ref operation holds a reservation. Running these is a quick way 121 * to reclaim space, but we want to hold this until the end because COW can 122 * churn a lot and we can avoid making some extent tree modifications if we 123 * are able to delay for as long as possible. 124 * 125 * ALLOC_CHUNK 126 * We will skip this the first time through space reservation, because of 127 * overcommit and we don't want to have a lot of useless metadata space when 128 * our worst case reservations will likely never come true. 129 * 130 * RUN_DELAYED_IPUTS 131 * If we're freeing inodes we're likely freeing checksums, file extent 132 * items, and extent tree items. Loads of space could be freed up by these 133 * operations, however they won't be usable until the transaction commits. 134 * 135 * COMMIT_TRANS 136 * This will commit the transaction. Historically we had a lot of logic 137 * surrounding whether or not we'd commit the transaction, but this waits born 138 * out of a pre-tickets era where we could end up committing the transaction 139 * thousands of times in a row without making progress. Now thanks to our 140 * ticketing system we know if we're not making progress and can error 141 * everybody out after a few commits rather than burning the disk hoping for 142 * a different answer. 143 * 144 * OVERCOMMIT 145 * 146 * Because we hold so many reservations for metadata we will allow you to 147 * reserve more space than is currently free in the currently allocate 148 * metadata space. This only happens with metadata, data does not allow 149 * overcommitting. 150 * 151 * You can see the current logic for when we allow overcommit in 152 * btrfs_can_overcommit(), but it only applies to unallocated space. If there 153 * is no unallocated space to be had, all reservations are kept within the 154 * free space in the allocated metadata chunks. 155 * 156 * Because of overcommitting, you generally want to use the 157 * btrfs_can_overcommit() logic for metadata allocations, as it does the right 158 * thing with or without extra unallocated space. 159 */ 160 161 u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, 162 bool may_use_included) 163 { 164 ASSERT(s_info); 165 return s_info->bytes_used + s_info->bytes_reserved + 166 s_info->bytes_pinned + s_info->bytes_readonly + 167 s_info->bytes_zone_unusable + 168 (may_use_included ? s_info->bytes_may_use : 0); 169 } 170 171 /* 172 * after adding space to the filesystem, we need to clear the full flags 173 * on all the space infos. 174 */ 175 void btrfs_clear_space_info_full(struct btrfs_fs_info *info) 176 { 177 struct list_head *head = &info->space_info; 178 struct btrfs_space_info *found; 179 180 list_for_each_entry(found, head, list) 181 found->full = 0; 182 } 183 184 static int create_space_info(struct btrfs_fs_info *info, u64 flags) 185 { 186 187 struct btrfs_space_info *space_info; 188 int i; 189 int ret; 190 191 space_info = kzalloc(sizeof(*space_info), GFP_NOFS); 192 if (!space_info) 193 return -ENOMEM; 194 195 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) 196 INIT_LIST_HEAD(&space_info->block_groups[i]); 197 init_rwsem(&space_info->groups_sem); 198 spin_lock_init(&space_info->lock); 199 space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; 200 space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; 201 INIT_LIST_HEAD(&space_info->ro_bgs); 202 INIT_LIST_HEAD(&space_info->tickets); 203 INIT_LIST_HEAD(&space_info->priority_tickets); 204 space_info->clamp = 1; 205 206 ret = btrfs_sysfs_add_space_info_type(info, space_info); 207 if (ret) 208 return ret; 209 210 list_add(&space_info->list, &info->space_info); 211 if (flags & BTRFS_BLOCK_GROUP_DATA) 212 info->data_sinfo = space_info; 213 214 return ret; 215 } 216 217 int btrfs_init_space_info(struct btrfs_fs_info *fs_info) 218 { 219 struct btrfs_super_block *disk_super; 220 u64 features; 221 u64 flags; 222 int mixed = 0; 223 int ret; 224 225 disk_super = fs_info->super_copy; 226 if (!btrfs_super_root(disk_super)) 227 return -EINVAL; 228 229 features = btrfs_super_incompat_flags(disk_super); 230 if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) 231 mixed = 1; 232 233 flags = BTRFS_BLOCK_GROUP_SYSTEM; 234 ret = create_space_info(fs_info, flags); 235 if (ret) 236 goto out; 237 238 if (mixed) { 239 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; 240 ret = create_space_info(fs_info, flags); 241 } else { 242 flags = BTRFS_BLOCK_GROUP_METADATA; 243 ret = create_space_info(fs_info, flags); 244 if (ret) 245 goto out; 246 247 flags = BTRFS_BLOCK_GROUP_DATA; 248 ret = create_space_info(fs_info, flags); 249 } 250 out: 251 return ret; 252 } 253 254 void btrfs_update_space_info(struct btrfs_fs_info *info, u64 flags, 255 u64 total_bytes, u64 bytes_used, 256 u64 bytes_readonly, u64 bytes_zone_unusable, 257 struct btrfs_space_info **space_info) 258 { 259 struct btrfs_space_info *found; 260 int factor; 261 262 factor = btrfs_bg_type_to_factor(flags); 263 264 found = btrfs_find_space_info(info, flags); 265 ASSERT(found); 266 spin_lock(&found->lock); 267 found->total_bytes += total_bytes; 268 found->disk_total += total_bytes * factor; 269 found->bytes_used += bytes_used; 270 found->disk_used += bytes_used * factor; 271 found->bytes_readonly += bytes_readonly; 272 found->bytes_zone_unusable += bytes_zone_unusable; 273 if (total_bytes > 0) 274 found->full = 0; 275 btrfs_try_granting_tickets(info, found); 276 spin_unlock(&found->lock); 277 *space_info = found; 278 } 279 280 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, 281 u64 flags) 282 { 283 struct list_head *head = &info->space_info; 284 struct btrfs_space_info *found; 285 286 flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; 287 288 list_for_each_entry(found, head, list) { 289 if (found->flags & flags) 290 return found; 291 } 292 return NULL; 293 } 294 295 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, 296 struct btrfs_space_info *space_info, 297 enum btrfs_reserve_flush_enum flush) 298 { 299 u64 profile; 300 u64 avail; 301 int factor; 302 303 if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) 304 profile = btrfs_system_alloc_profile(fs_info); 305 else 306 profile = btrfs_metadata_alloc_profile(fs_info); 307 308 avail = atomic64_read(&fs_info->free_chunk_space); 309 310 /* 311 * If we have dup, raid1 or raid10 then only half of the free 312 * space is actually usable. For raid56, the space info used 313 * doesn't include the parity drive, so we don't have to 314 * change the math 315 */ 316 factor = btrfs_bg_type_to_factor(profile); 317 avail = div_u64(avail, factor); 318 319 /* 320 * If we aren't flushing all things, let us overcommit up to 321 * 1/2th of the space. If we can flush, don't let us overcommit 322 * too much, let it overcommit up to 1/8 of the space. 323 */ 324 if (flush == BTRFS_RESERVE_FLUSH_ALL) 325 avail >>= 3; 326 else 327 avail >>= 1; 328 return avail; 329 } 330 331 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, 332 struct btrfs_space_info *space_info, u64 bytes, 333 enum btrfs_reserve_flush_enum flush) 334 { 335 u64 avail; 336 u64 used; 337 338 /* Don't overcommit when in mixed mode */ 339 if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) 340 return 0; 341 342 used = btrfs_space_info_used(space_info, true); 343 avail = calc_available_free_space(fs_info, space_info, flush); 344 345 if (used + bytes < space_info->total_bytes + avail) 346 return 1; 347 return 0; 348 } 349 350 static void remove_ticket(struct btrfs_space_info *space_info, 351 struct reserve_ticket *ticket) 352 { 353 if (!list_empty(&ticket->list)) { 354 list_del_init(&ticket->list); 355 ASSERT(space_info->reclaim_size >= ticket->bytes); 356 space_info->reclaim_size -= ticket->bytes; 357 } 358 } 359 360 /* 361 * This is for space we already have accounted in space_info->bytes_may_use, so 362 * basically when we're returning space from block_rsv's. 363 */ 364 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, 365 struct btrfs_space_info *space_info) 366 { 367 struct list_head *head; 368 enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; 369 370 lockdep_assert_held(&space_info->lock); 371 372 head = &space_info->priority_tickets; 373 again: 374 while (!list_empty(head)) { 375 struct reserve_ticket *ticket; 376 u64 used = btrfs_space_info_used(space_info, true); 377 378 ticket = list_first_entry(head, struct reserve_ticket, list); 379 380 /* Check and see if our ticket can be satisfied now. */ 381 if ((used + ticket->bytes <= space_info->total_bytes) || 382 btrfs_can_overcommit(fs_info, space_info, ticket->bytes, 383 flush)) { 384 btrfs_space_info_update_bytes_may_use(fs_info, 385 space_info, 386 ticket->bytes); 387 remove_ticket(space_info, ticket); 388 ticket->bytes = 0; 389 space_info->tickets_id++; 390 wake_up(&ticket->wait); 391 } else { 392 break; 393 } 394 } 395 396 if (head == &space_info->priority_tickets) { 397 head = &space_info->tickets; 398 flush = BTRFS_RESERVE_FLUSH_ALL; 399 goto again; 400 } 401 } 402 403 #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ 404 do { \ 405 struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ 406 spin_lock(&__rsv->lock); \ 407 btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ 408 __rsv->size, __rsv->reserved); \ 409 spin_unlock(&__rsv->lock); \ 410 } while (0) 411 412 static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 413 struct btrfs_space_info *info) 414 { 415 lockdep_assert_held(&info->lock); 416 417 /* The free space could be negative in case of overcommit */ 418 btrfs_info(fs_info, "space_info %llu has %lld free, is %sfull", 419 info->flags, 420 (s64)(info->total_bytes - btrfs_space_info_used(info, true)), 421 info->full ? "" : "not "); 422 btrfs_info(fs_info, 423 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu", 424 info->total_bytes, info->bytes_used, info->bytes_pinned, 425 info->bytes_reserved, info->bytes_may_use, 426 info->bytes_readonly, info->bytes_zone_unusable); 427 428 DUMP_BLOCK_RSV(fs_info, global_block_rsv); 429 DUMP_BLOCK_RSV(fs_info, trans_block_rsv); 430 DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); 431 DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); 432 DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); 433 434 } 435 436 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, 437 struct btrfs_space_info *info, u64 bytes, 438 int dump_block_groups) 439 { 440 struct btrfs_block_group *cache; 441 int index = 0; 442 443 spin_lock(&info->lock); 444 __btrfs_dump_space_info(fs_info, info); 445 spin_unlock(&info->lock); 446 447 if (!dump_block_groups) 448 return; 449 450 down_read(&info->groups_sem); 451 again: 452 list_for_each_entry(cache, &info->block_groups[index], list) { 453 spin_lock(&cache->lock); 454 btrfs_info(fs_info, 455 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s", 456 cache->start, cache->length, cache->used, cache->pinned, 457 cache->reserved, cache->zone_unusable, 458 cache->ro ? "[readonly]" : ""); 459 spin_unlock(&cache->lock); 460 btrfs_dump_free_space(cache, bytes); 461 } 462 if (++index < BTRFS_NR_RAID_TYPES) 463 goto again; 464 up_read(&info->groups_sem); 465 } 466 467 static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, 468 u64 to_reclaim) 469 { 470 u64 bytes; 471 u64 nr; 472 473 bytes = btrfs_calc_insert_metadata_size(fs_info, 1); 474 nr = div64_u64(to_reclaim, bytes); 475 if (!nr) 476 nr = 1; 477 return nr; 478 } 479 480 #define EXTENT_SIZE_PER_ITEM SZ_256K 481 482 /* 483 * shrink metadata reservation for delalloc 484 */ 485 static void shrink_delalloc(struct btrfs_fs_info *fs_info, 486 struct btrfs_space_info *space_info, 487 u64 to_reclaim, bool wait_ordered, 488 bool for_preempt) 489 { 490 struct btrfs_trans_handle *trans; 491 u64 delalloc_bytes; 492 u64 ordered_bytes; 493 u64 items; 494 long time_left; 495 int loops; 496 497 delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 498 ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); 499 if (delalloc_bytes == 0 && ordered_bytes == 0) 500 return; 501 502 /* Calc the number of the pages we need flush for space reservation */ 503 if (to_reclaim == U64_MAX) { 504 items = U64_MAX; 505 } else { 506 /* 507 * to_reclaim is set to however much metadata we need to 508 * reclaim, but reclaiming that much data doesn't really track 509 * exactly. What we really want to do is reclaim full inode's 510 * worth of reservations, however that's not available to us 511 * here. We will take a fraction of the delalloc bytes for our 512 * flushing loops and hope for the best. Delalloc will expand 513 * the amount we write to cover an entire dirty extent, which 514 * will reclaim the metadata reservation for that range. If 515 * it's not enough subsequent flush stages will be more 516 * aggressive. 517 */ 518 to_reclaim = max(to_reclaim, delalloc_bytes >> 3); 519 items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; 520 } 521 522 trans = (struct btrfs_trans_handle *)current->journal_info; 523 524 /* 525 * If we are doing more ordered than delalloc we need to just wait on 526 * ordered extents, otherwise we'll waste time trying to flush delalloc 527 * that likely won't give us the space back we need. 528 */ 529 if (ordered_bytes > delalloc_bytes && !for_preempt) 530 wait_ordered = true; 531 532 loops = 0; 533 while ((delalloc_bytes || ordered_bytes) && loops < 3) { 534 u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; 535 long nr_pages = min_t(u64, temp, LONG_MAX); 536 int async_pages; 537 538 btrfs_start_delalloc_roots(fs_info, nr_pages, true); 539 540 /* 541 * We need to make sure any outstanding async pages are now 542 * processed before we continue. This is because things like 543 * sync_inode() try to be smart and skip writing if the inode is 544 * marked clean. We don't use filemap_fwrite for flushing 545 * because we want to control how many pages we write out at a 546 * time, thus this is the only safe way to make sure we've 547 * waited for outstanding compressed workers to have started 548 * their jobs and thus have ordered extents set up properly. 549 * 550 * This exists because we do not want to wait for each 551 * individual inode to finish its async work, we simply want to 552 * start the IO on everybody, and then come back here and wait 553 * for all of the async work to catch up. Once we're done with 554 * that we know we'll have ordered extents for everything and we 555 * can decide if we wait for that or not. 556 * 557 * If we choose to replace this in the future, make absolutely 558 * sure that the proper waiting is being done in the async case, 559 * as there have been bugs in that area before. 560 */ 561 async_pages = atomic_read(&fs_info->async_delalloc_pages); 562 if (!async_pages) 563 goto skip_async; 564 565 /* 566 * We don't want to wait forever, if we wrote less pages in this 567 * loop than we have outstanding, only wait for that number of 568 * pages, otherwise we can wait for all async pages to finish 569 * before continuing. 570 */ 571 if (async_pages > nr_pages) 572 async_pages -= nr_pages; 573 else 574 async_pages = 0; 575 wait_event(fs_info->async_submit_wait, 576 atomic_read(&fs_info->async_delalloc_pages) <= 577 async_pages); 578 skip_async: 579 loops++; 580 if (wait_ordered && !trans) { 581 btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); 582 } else { 583 time_left = schedule_timeout_killable(1); 584 if (time_left) 585 break; 586 } 587 588 /* 589 * If we are for preemption we just want a one-shot of delalloc 590 * flushing so we can stop flushing if we decide we don't need 591 * to anymore. 592 */ 593 if (for_preempt) 594 break; 595 596 spin_lock(&space_info->lock); 597 if (list_empty(&space_info->tickets) && 598 list_empty(&space_info->priority_tickets)) { 599 spin_unlock(&space_info->lock); 600 break; 601 } 602 spin_unlock(&space_info->lock); 603 604 delalloc_bytes = percpu_counter_sum_positive( 605 &fs_info->delalloc_bytes); 606 ordered_bytes = percpu_counter_sum_positive( 607 &fs_info->ordered_bytes); 608 } 609 } 610 611 /* 612 * Try to flush some data based on policy set by @state. This is only advisory 613 * and may fail for various reasons. The caller is supposed to examine the 614 * state of @space_info to detect the outcome. 615 */ 616 static void flush_space(struct btrfs_fs_info *fs_info, 617 struct btrfs_space_info *space_info, u64 num_bytes, 618 enum btrfs_flush_state state, bool for_preempt) 619 { 620 struct btrfs_root *root = fs_info->tree_root; 621 struct btrfs_trans_handle *trans; 622 int nr; 623 int ret = 0; 624 625 switch (state) { 626 case FLUSH_DELAYED_ITEMS_NR: 627 case FLUSH_DELAYED_ITEMS: 628 if (state == FLUSH_DELAYED_ITEMS_NR) 629 nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; 630 else 631 nr = -1; 632 633 trans = btrfs_join_transaction(root); 634 if (IS_ERR(trans)) { 635 ret = PTR_ERR(trans); 636 break; 637 } 638 ret = btrfs_run_delayed_items_nr(trans, nr); 639 btrfs_end_transaction(trans); 640 break; 641 case FLUSH_DELALLOC: 642 case FLUSH_DELALLOC_WAIT: 643 case FLUSH_DELALLOC_FULL: 644 if (state == FLUSH_DELALLOC_FULL) 645 num_bytes = U64_MAX; 646 shrink_delalloc(fs_info, space_info, num_bytes, 647 state != FLUSH_DELALLOC, for_preempt); 648 break; 649 case FLUSH_DELAYED_REFS_NR: 650 case FLUSH_DELAYED_REFS: 651 trans = btrfs_join_transaction(root); 652 if (IS_ERR(trans)) { 653 ret = PTR_ERR(trans); 654 break; 655 } 656 if (state == FLUSH_DELAYED_REFS_NR) 657 nr = calc_reclaim_items_nr(fs_info, num_bytes); 658 else 659 nr = 0; 660 btrfs_run_delayed_refs(trans, nr); 661 btrfs_end_transaction(trans); 662 break; 663 case ALLOC_CHUNK: 664 case ALLOC_CHUNK_FORCE: 665 trans = btrfs_join_transaction(root); 666 if (IS_ERR(trans)) { 667 ret = PTR_ERR(trans); 668 break; 669 } 670 ret = btrfs_chunk_alloc(trans, 671 btrfs_get_alloc_profile(fs_info, space_info->flags), 672 (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : 673 CHUNK_ALLOC_FORCE); 674 btrfs_end_transaction(trans); 675 if (ret > 0 || ret == -ENOSPC) 676 ret = 0; 677 break; 678 case RUN_DELAYED_IPUTS: 679 /* 680 * If we have pending delayed iputs then we could free up a 681 * bunch of pinned space, so make sure we run the iputs before 682 * we do our pinned bytes check below. 683 */ 684 btrfs_run_delayed_iputs(fs_info); 685 btrfs_wait_on_delayed_iputs(fs_info); 686 break; 687 case COMMIT_TRANS: 688 ASSERT(current->journal_info == NULL); 689 trans = btrfs_join_transaction(root); 690 if (IS_ERR(trans)) { 691 ret = PTR_ERR(trans); 692 break; 693 } 694 ret = btrfs_commit_transaction(trans); 695 break; 696 default: 697 ret = -ENOSPC; 698 break; 699 } 700 701 trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, 702 ret, for_preempt); 703 return; 704 } 705 706 static inline u64 707 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, 708 struct btrfs_space_info *space_info) 709 { 710 u64 used; 711 u64 avail; 712 u64 to_reclaim = space_info->reclaim_size; 713 714 lockdep_assert_held(&space_info->lock); 715 716 avail = calc_available_free_space(fs_info, space_info, 717 BTRFS_RESERVE_FLUSH_ALL); 718 used = btrfs_space_info_used(space_info, true); 719 720 /* 721 * We may be flushing because suddenly we have less space than we had 722 * before, and now we're well over-committed based on our current free 723 * space. If that's the case add in our overage so we make sure to put 724 * appropriate pressure on the flushing state machine. 725 */ 726 if (space_info->total_bytes + avail < used) 727 to_reclaim += used - (space_info->total_bytes + avail); 728 729 return to_reclaim; 730 } 731 732 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, 733 struct btrfs_space_info *space_info) 734 { 735 u64 global_rsv_size = fs_info->global_block_rsv.reserved; 736 u64 ordered, delalloc; 737 u64 thresh = div_factor_fine(space_info->total_bytes, 90); 738 u64 used; 739 740 /* If we're just plain full then async reclaim just slows us down. */ 741 if ((space_info->bytes_used + space_info->bytes_reserved + 742 global_rsv_size) >= thresh) 743 return false; 744 745 used = space_info->bytes_may_use + space_info->bytes_pinned; 746 747 /* The total flushable belongs to the global rsv, don't flush. */ 748 if (global_rsv_size >= used) 749 return false; 750 751 /* 752 * 128MiB is 1/4 of the maximum global rsv size. If we have less than 753 * that devoted to other reservations then there's no sense in flushing, 754 * we don't have a lot of things that need flushing. 755 */ 756 if (used - global_rsv_size <= SZ_128M) 757 return false; 758 759 /* 760 * We have tickets queued, bail so we don't compete with the async 761 * flushers. 762 */ 763 if (space_info->reclaim_size) 764 return false; 765 766 /* 767 * If we have over half of the free space occupied by reservations or 768 * pinned then we want to start flushing. 769 * 770 * We do not do the traditional thing here, which is to say 771 * 772 * if (used >= ((total_bytes + avail) / 2)) 773 * return 1; 774 * 775 * because this doesn't quite work how we want. If we had more than 50% 776 * of the space_info used by bytes_used and we had 0 available we'd just 777 * constantly run the background flusher. Instead we want it to kick in 778 * if our reclaimable space exceeds our clamped free space. 779 * 780 * Our clamping range is 2^1 -> 2^8. Practically speaking that means 781 * the following: 782 * 783 * Amount of RAM Minimum threshold Maximum threshold 784 * 785 * 256GiB 1GiB 128GiB 786 * 128GiB 512MiB 64GiB 787 * 64GiB 256MiB 32GiB 788 * 32GiB 128MiB 16GiB 789 * 16GiB 64MiB 8GiB 790 * 791 * These are the range our thresholds will fall in, corresponding to how 792 * much delalloc we need for the background flusher to kick in. 793 */ 794 795 thresh = calc_available_free_space(fs_info, space_info, 796 BTRFS_RESERVE_FLUSH_ALL); 797 used = space_info->bytes_used + space_info->bytes_reserved + 798 space_info->bytes_readonly + global_rsv_size; 799 if (used < space_info->total_bytes) 800 thresh += space_info->total_bytes - used; 801 thresh >>= space_info->clamp; 802 803 used = space_info->bytes_pinned; 804 805 /* 806 * If we have more ordered bytes than delalloc bytes then we're either 807 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting 808 * around. Preemptive flushing is only useful in that it can free up 809 * space before tickets need to wait for things to finish. In the case 810 * of ordered extents, preemptively waiting on ordered extents gets us 811 * nothing, if our reservations are tied up in ordered extents we'll 812 * simply have to slow down writers by forcing them to wait on ordered 813 * extents. 814 * 815 * In the case that ordered is larger than delalloc, only include the 816 * block reserves that we would actually be able to directly reclaim 817 * from. In this case if we're heavy on metadata operations this will 818 * clearly be heavy enough to warrant preemptive flushing. In the case 819 * of heavy DIO or ordered reservations, preemptive flushing will just 820 * waste time and cause us to slow down. 821 * 822 * We want to make sure we truly are maxed out on ordered however, so 823 * cut ordered in half, and if it's still higher than delalloc then we 824 * can keep flushing. This is to avoid the case where we start 825 * flushing, and now delalloc == ordered and we stop preemptively 826 * flushing when we could still have several gigs of delalloc to flush. 827 */ 828 ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; 829 delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); 830 if (ordered >= delalloc) 831 used += fs_info->delayed_refs_rsv.reserved + 832 fs_info->delayed_block_rsv.reserved; 833 else 834 used += space_info->bytes_may_use - global_rsv_size; 835 836 return (used >= thresh && !btrfs_fs_closing(fs_info) && 837 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); 838 } 839 840 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, 841 struct btrfs_space_info *space_info, 842 struct reserve_ticket *ticket) 843 { 844 struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; 845 u64 min_bytes; 846 847 if (!ticket->steal) 848 return false; 849 850 if (global_rsv->space_info != space_info) 851 return false; 852 853 spin_lock(&global_rsv->lock); 854 min_bytes = div_factor(global_rsv->size, 1); 855 if (global_rsv->reserved < min_bytes + ticket->bytes) { 856 spin_unlock(&global_rsv->lock); 857 return false; 858 } 859 global_rsv->reserved -= ticket->bytes; 860 remove_ticket(space_info, ticket); 861 ticket->bytes = 0; 862 wake_up(&ticket->wait); 863 space_info->tickets_id++; 864 if (global_rsv->reserved < global_rsv->size) 865 global_rsv->full = 0; 866 spin_unlock(&global_rsv->lock); 867 868 return true; 869 } 870 871 /* 872 * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets 873 * @fs_info - fs_info for this fs 874 * @space_info - the space info we were flushing 875 * 876 * We call this when we've exhausted our flushing ability and haven't made 877 * progress in satisfying tickets. The reservation code handles tickets in 878 * order, so if there is a large ticket first and then smaller ones we could 879 * very well satisfy the smaller tickets. This will attempt to wake up any 880 * tickets in the list to catch this case. 881 * 882 * This function returns true if it was able to make progress by clearing out 883 * other tickets, or if it stumbles across a ticket that was smaller than the 884 * first ticket. 885 */ 886 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, 887 struct btrfs_space_info *space_info) 888 { 889 struct reserve_ticket *ticket; 890 u64 tickets_id = space_info->tickets_id; 891 const bool aborted = BTRFS_FS_ERROR(fs_info); 892 893 trace_btrfs_fail_all_tickets(fs_info, space_info); 894 895 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { 896 btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); 897 __btrfs_dump_space_info(fs_info, space_info); 898 } 899 900 while (!list_empty(&space_info->tickets) && 901 tickets_id == space_info->tickets_id) { 902 ticket = list_first_entry(&space_info->tickets, 903 struct reserve_ticket, list); 904 905 if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket)) 906 return true; 907 908 if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 909 btrfs_info(fs_info, "failing ticket with %llu bytes", 910 ticket->bytes); 911 912 remove_ticket(space_info, ticket); 913 if (aborted) 914 ticket->error = -EIO; 915 else 916 ticket->error = -ENOSPC; 917 wake_up(&ticket->wait); 918 919 /* 920 * We're just throwing tickets away, so more flushing may not 921 * trip over btrfs_try_granting_tickets, so we need to call it 922 * here to see if we can make progress with the next ticket in 923 * the list. 924 */ 925 if (!aborted) 926 btrfs_try_granting_tickets(fs_info, space_info); 927 } 928 return (tickets_id != space_info->tickets_id); 929 } 930 931 /* 932 * This is for normal flushers, we can wait all goddamned day if we want to. We 933 * will loop and continuously try to flush as long as we are making progress. 934 * We count progress as clearing off tickets each time we have to loop. 935 */ 936 static void btrfs_async_reclaim_metadata_space(struct work_struct *work) 937 { 938 struct btrfs_fs_info *fs_info; 939 struct btrfs_space_info *space_info; 940 u64 to_reclaim; 941 enum btrfs_flush_state flush_state; 942 int commit_cycles = 0; 943 u64 last_tickets_id; 944 945 fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); 946 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 947 948 spin_lock(&space_info->lock); 949 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 950 if (!to_reclaim) { 951 space_info->flush = 0; 952 spin_unlock(&space_info->lock); 953 return; 954 } 955 last_tickets_id = space_info->tickets_id; 956 spin_unlock(&space_info->lock); 957 958 flush_state = FLUSH_DELAYED_ITEMS_NR; 959 do { 960 flush_space(fs_info, space_info, to_reclaim, flush_state, false); 961 spin_lock(&space_info->lock); 962 if (list_empty(&space_info->tickets)) { 963 space_info->flush = 0; 964 spin_unlock(&space_info->lock); 965 return; 966 } 967 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, 968 space_info); 969 if (last_tickets_id == space_info->tickets_id) { 970 flush_state++; 971 } else { 972 last_tickets_id = space_info->tickets_id; 973 flush_state = FLUSH_DELAYED_ITEMS_NR; 974 if (commit_cycles) 975 commit_cycles--; 976 } 977 978 /* 979 * We do not want to empty the system of delalloc unless we're 980 * under heavy pressure, so allow one trip through the flushing 981 * logic before we start doing a FLUSH_DELALLOC_FULL. 982 */ 983 if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) 984 flush_state++; 985 986 /* 987 * We don't want to force a chunk allocation until we've tried 988 * pretty hard to reclaim space. Think of the case where we 989 * freed up a bunch of space and so have a lot of pinned space 990 * to reclaim. We would rather use that than possibly create a 991 * underutilized metadata chunk. So if this is our first run 992 * through the flushing state machine skip ALLOC_CHUNK_FORCE and 993 * commit the transaction. If nothing has changed the next go 994 * around then we can force a chunk allocation. 995 */ 996 if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) 997 flush_state++; 998 999 if (flush_state > COMMIT_TRANS) { 1000 commit_cycles++; 1001 if (commit_cycles > 2) { 1002 if (maybe_fail_all_tickets(fs_info, space_info)) { 1003 flush_state = FLUSH_DELAYED_ITEMS_NR; 1004 commit_cycles--; 1005 } else { 1006 space_info->flush = 0; 1007 } 1008 } else { 1009 flush_state = FLUSH_DELAYED_ITEMS_NR; 1010 } 1011 } 1012 spin_unlock(&space_info->lock); 1013 } while (flush_state <= COMMIT_TRANS); 1014 } 1015 1016 /* 1017 * This handles pre-flushing of metadata space before we get to the point that 1018 * we need to start blocking threads on tickets. The logic here is different 1019 * from the other flush paths because it doesn't rely on tickets to tell us how 1020 * much we need to flush, instead it attempts to keep us below the 80% full 1021 * watermark of space by flushing whichever reservation pool is currently the 1022 * largest. 1023 */ 1024 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) 1025 { 1026 struct btrfs_fs_info *fs_info; 1027 struct btrfs_space_info *space_info; 1028 struct btrfs_block_rsv *delayed_block_rsv; 1029 struct btrfs_block_rsv *delayed_refs_rsv; 1030 struct btrfs_block_rsv *global_rsv; 1031 struct btrfs_block_rsv *trans_rsv; 1032 int loops = 0; 1033 1034 fs_info = container_of(work, struct btrfs_fs_info, 1035 preempt_reclaim_work); 1036 space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); 1037 delayed_block_rsv = &fs_info->delayed_block_rsv; 1038 delayed_refs_rsv = &fs_info->delayed_refs_rsv; 1039 global_rsv = &fs_info->global_block_rsv; 1040 trans_rsv = &fs_info->trans_block_rsv; 1041 1042 spin_lock(&space_info->lock); 1043 while (need_preemptive_reclaim(fs_info, space_info)) { 1044 enum btrfs_flush_state flush; 1045 u64 delalloc_size = 0; 1046 u64 to_reclaim, block_rsv_size; 1047 u64 global_rsv_size = global_rsv->reserved; 1048 1049 loops++; 1050 1051 /* 1052 * We don't have a precise counter for the metadata being 1053 * reserved for delalloc, so we'll approximate it by subtracting 1054 * out the block rsv's space from the bytes_may_use. If that 1055 * amount is higher than the individual reserves, then we can 1056 * assume it's tied up in delalloc reservations. 1057 */ 1058 block_rsv_size = global_rsv_size + 1059 delayed_block_rsv->reserved + 1060 delayed_refs_rsv->reserved + 1061 trans_rsv->reserved; 1062 if (block_rsv_size < space_info->bytes_may_use) 1063 delalloc_size = space_info->bytes_may_use - block_rsv_size; 1064 1065 /* 1066 * We don't want to include the global_rsv in our calculation, 1067 * because that's space we can't touch. Subtract it from the 1068 * block_rsv_size for the next checks. 1069 */ 1070 block_rsv_size -= global_rsv_size; 1071 1072 /* 1073 * We really want to avoid flushing delalloc too much, as it 1074 * could result in poor allocation patterns, so only flush it if 1075 * it's larger than the rest of the pools combined. 1076 */ 1077 if (delalloc_size > block_rsv_size) { 1078 to_reclaim = delalloc_size; 1079 flush = FLUSH_DELALLOC; 1080 } else if (space_info->bytes_pinned > 1081 (delayed_block_rsv->reserved + 1082 delayed_refs_rsv->reserved)) { 1083 to_reclaim = space_info->bytes_pinned; 1084 flush = COMMIT_TRANS; 1085 } else if (delayed_block_rsv->reserved > 1086 delayed_refs_rsv->reserved) { 1087 to_reclaim = delayed_block_rsv->reserved; 1088 flush = FLUSH_DELAYED_ITEMS_NR; 1089 } else { 1090 to_reclaim = delayed_refs_rsv->reserved; 1091 flush = FLUSH_DELAYED_REFS_NR; 1092 } 1093 1094 spin_unlock(&space_info->lock); 1095 1096 /* 1097 * We don't want to reclaim everything, just a portion, so scale 1098 * down the to_reclaim by 1/4. If it takes us down to 0, 1099 * reclaim 1 items worth. 1100 */ 1101 to_reclaim >>= 2; 1102 if (!to_reclaim) 1103 to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); 1104 flush_space(fs_info, space_info, to_reclaim, flush, true); 1105 cond_resched(); 1106 spin_lock(&space_info->lock); 1107 } 1108 1109 /* We only went through once, back off our clamping. */ 1110 if (loops == 1 && !space_info->reclaim_size) 1111 space_info->clamp = max(1, space_info->clamp - 1); 1112 trace_btrfs_done_preemptive_reclaim(fs_info, space_info); 1113 spin_unlock(&space_info->lock); 1114 } 1115 1116 /* 1117 * FLUSH_DELALLOC_WAIT: 1118 * Space is freed from flushing delalloc in one of two ways. 1119 * 1120 * 1) compression is on and we allocate less space than we reserved 1121 * 2) we are overwriting existing space 1122 * 1123 * For #1 that extra space is reclaimed as soon as the delalloc pages are 1124 * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent 1125 * length to ->bytes_reserved, and subtracts the reserved space from 1126 * ->bytes_may_use. 1127 * 1128 * For #2 this is trickier. Once the ordered extent runs we will drop the 1129 * extent in the range we are overwriting, which creates a delayed ref for 1130 * that freed extent. This however is not reclaimed until the transaction 1131 * commits, thus the next stages. 1132 * 1133 * RUN_DELAYED_IPUTS 1134 * If we are freeing inodes, we want to make sure all delayed iputs have 1135 * completed, because they could have been on an inode with i_nlink == 0, and 1136 * thus have been truncated and freed up space. But again this space is not 1137 * immediately re-usable, it comes in the form of a delayed ref, which must be 1138 * run and then the transaction must be committed. 1139 * 1140 * COMMIT_TRANS 1141 * This is where we reclaim all of the pinned space generated by running the 1142 * iputs 1143 * 1144 * ALLOC_CHUNK_FORCE 1145 * For data we start with alloc chunk force, however we could have been full 1146 * before, and then the transaction commit could have freed new block groups, 1147 * so if we now have space to allocate do the force chunk allocation. 1148 */ 1149 static const enum btrfs_flush_state data_flush_states[] = { 1150 FLUSH_DELALLOC_FULL, 1151 RUN_DELAYED_IPUTS, 1152 COMMIT_TRANS, 1153 ALLOC_CHUNK_FORCE, 1154 }; 1155 1156 static void btrfs_async_reclaim_data_space(struct work_struct *work) 1157 { 1158 struct btrfs_fs_info *fs_info; 1159 struct btrfs_space_info *space_info; 1160 u64 last_tickets_id; 1161 enum btrfs_flush_state flush_state = 0; 1162 1163 fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); 1164 space_info = fs_info->data_sinfo; 1165 1166 spin_lock(&space_info->lock); 1167 if (list_empty(&space_info->tickets)) { 1168 space_info->flush = 0; 1169 spin_unlock(&space_info->lock); 1170 return; 1171 } 1172 last_tickets_id = space_info->tickets_id; 1173 spin_unlock(&space_info->lock); 1174 1175 while (!space_info->full) { 1176 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1177 spin_lock(&space_info->lock); 1178 if (list_empty(&space_info->tickets)) { 1179 space_info->flush = 0; 1180 spin_unlock(&space_info->lock); 1181 return; 1182 } 1183 1184 /* Something happened, fail everything and bail. */ 1185 if (BTRFS_FS_ERROR(fs_info)) 1186 goto aborted_fs; 1187 last_tickets_id = space_info->tickets_id; 1188 spin_unlock(&space_info->lock); 1189 } 1190 1191 while (flush_state < ARRAY_SIZE(data_flush_states)) { 1192 flush_space(fs_info, space_info, U64_MAX, 1193 data_flush_states[flush_state], false); 1194 spin_lock(&space_info->lock); 1195 if (list_empty(&space_info->tickets)) { 1196 space_info->flush = 0; 1197 spin_unlock(&space_info->lock); 1198 return; 1199 } 1200 1201 if (last_tickets_id == space_info->tickets_id) { 1202 flush_state++; 1203 } else { 1204 last_tickets_id = space_info->tickets_id; 1205 flush_state = 0; 1206 } 1207 1208 if (flush_state >= ARRAY_SIZE(data_flush_states)) { 1209 if (space_info->full) { 1210 if (maybe_fail_all_tickets(fs_info, space_info)) 1211 flush_state = 0; 1212 else 1213 space_info->flush = 0; 1214 } else { 1215 flush_state = 0; 1216 } 1217 1218 /* Something happened, fail everything and bail. */ 1219 if (BTRFS_FS_ERROR(fs_info)) 1220 goto aborted_fs; 1221 1222 } 1223 spin_unlock(&space_info->lock); 1224 } 1225 return; 1226 1227 aborted_fs: 1228 maybe_fail_all_tickets(fs_info, space_info); 1229 space_info->flush = 0; 1230 spin_unlock(&space_info->lock); 1231 } 1232 1233 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) 1234 { 1235 INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); 1236 INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); 1237 INIT_WORK(&fs_info->preempt_reclaim_work, 1238 btrfs_preempt_reclaim_metadata_space); 1239 } 1240 1241 static const enum btrfs_flush_state priority_flush_states[] = { 1242 FLUSH_DELAYED_ITEMS_NR, 1243 FLUSH_DELAYED_ITEMS, 1244 ALLOC_CHUNK, 1245 }; 1246 1247 static const enum btrfs_flush_state evict_flush_states[] = { 1248 FLUSH_DELAYED_ITEMS_NR, 1249 FLUSH_DELAYED_ITEMS, 1250 FLUSH_DELAYED_REFS_NR, 1251 FLUSH_DELAYED_REFS, 1252 FLUSH_DELALLOC, 1253 FLUSH_DELALLOC_WAIT, 1254 FLUSH_DELALLOC_FULL, 1255 ALLOC_CHUNK, 1256 COMMIT_TRANS, 1257 }; 1258 1259 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, 1260 struct btrfs_space_info *space_info, 1261 struct reserve_ticket *ticket, 1262 const enum btrfs_flush_state *states, 1263 int states_nr) 1264 { 1265 u64 to_reclaim; 1266 int flush_state = 0; 1267 1268 spin_lock(&space_info->lock); 1269 to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); 1270 /* 1271 * This is the priority reclaim path, so to_reclaim could be >0 still 1272 * because we may have only satisified the priority tickets and still 1273 * left non priority tickets on the list. We would then have 1274 * to_reclaim but ->bytes == 0. 1275 */ 1276 if (ticket->bytes == 0) { 1277 spin_unlock(&space_info->lock); 1278 return; 1279 } 1280 1281 while (flush_state < states_nr) { 1282 spin_unlock(&space_info->lock); 1283 flush_space(fs_info, space_info, to_reclaim, states[flush_state], 1284 false); 1285 flush_state++; 1286 spin_lock(&space_info->lock); 1287 if (ticket->bytes == 0) { 1288 spin_unlock(&space_info->lock); 1289 return; 1290 } 1291 } 1292 1293 /* Attempt to steal from the global rsv if we can. */ 1294 if (!steal_from_global_rsv(fs_info, space_info, ticket)) { 1295 ticket->error = -ENOSPC; 1296 remove_ticket(space_info, ticket); 1297 } 1298 1299 /* 1300 * We must run try_granting_tickets here because we could be a large 1301 * ticket in front of a smaller ticket that can now be satisfied with 1302 * the available space. 1303 */ 1304 btrfs_try_granting_tickets(fs_info, space_info); 1305 spin_unlock(&space_info->lock); 1306 } 1307 1308 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, 1309 struct btrfs_space_info *space_info, 1310 struct reserve_ticket *ticket) 1311 { 1312 spin_lock(&space_info->lock); 1313 1314 /* We could have been granted before we got here. */ 1315 if (ticket->bytes == 0) { 1316 spin_unlock(&space_info->lock); 1317 return; 1318 } 1319 1320 while (!space_info->full) { 1321 spin_unlock(&space_info->lock); 1322 flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); 1323 spin_lock(&space_info->lock); 1324 if (ticket->bytes == 0) { 1325 spin_unlock(&space_info->lock); 1326 return; 1327 } 1328 } 1329 1330 ticket->error = -ENOSPC; 1331 remove_ticket(space_info, ticket); 1332 btrfs_try_granting_tickets(fs_info, space_info); 1333 spin_unlock(&space_info->lock); 1334 } 1335 1336 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, 1337 struct btrfs_space_info *space_info, 1338 struct reserve_ticket *ticket) 1339 1340 { 1341 DEFINE_WAIT(wait); 1342 int ret = 0; 1343 1344 spin_lock(&space_info->lock); 1345 while (ticket->bytes > 0 && ticket->error == 0) { 1346 ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); 1347 if (ret) { 1348 /* 1349 * Delete us from the list. After we unlock the space 1350 * info, we don't want the async reclaim job to reserve 1351 * space for this ticket. If that would happen, then the 1352 * ticket's task would not known that space was reserved 1353 * despite getting an error, resulting in a space leak 1354 * (bytes_may_use counter of our space_info). 1355 */ 1356 remove_ticket(space_info, ticket); 1357 ticket->error = -EINTR; 1358 break; 1359 } 1360 spin_unlock(&space_info->lock); 1361 1362 schedule(); 1363 1364 finish_wait(&ticket->wait, &wait); 1365 spin_lock(&space_info->lock); 1366 } 1367 spin_unlock(&space_info->lock); 1368 } 1369 1370 /** 1371 * Do the appropriate flushing and waiting for a ticket 1372 * 1373 * @fs_info: the filesystem 1374 * @space_info: space info for the reservation 1375 * @ticket: ticket for the reservation 1376 * @start_ns: timestamp when the reservation started 1377 * @orig_bytes: amount of bytes originally reserved 1378 * @flush: how much we can flush 1379 * 1380 * This does the work of figuring out how to flush for the ticket, waiting for 1381 * the reservation, and returning the appropriate error if there is one. 1382 */ 1383 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, 1384 struct btrfs_space_info *space_info, 1385 struct reserve_ticket *ticket, 1386 u64 start_ns, u64 orig_bytes, 1387 enum btrfs_reserve_flush_enum flush) 1388 { 1389 int ret; 1390 1391 switch (flush) { 1392 case BTRFS_RESERVE_FLUSH_DATA: 1393 case BTRFS_RESERVE_FLUSH_ALL: 1394 case BTRFS_RESERVE_FLUSH_ALL_STEAL: 1395 wait_reserve_ticket(fs_info, space_info, ticket); 1396 break; 1397 case BTRFS_RESERVE_FLUSH_LIMIT: 1398 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1399 priority_flush_states, 1400 ARRAY_SIZE(priority_flush_states)); 1401 break; 1402 case BTRFS_RESERVE_FLUSH_EVICT: 1403 priority_reclaim_metadata_space(fs_info, space_info, ticket, 1404 evict_flush_states, 1405 ARRAY_SIZE(evict_flush_states)); 1406 break; 1407 case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: 1408 priority_reclaim_data_space(fs_info, space_info, ticket); 1409 break; 1410 default: 1411 ASSERT(0); 1412 break; 1413 } 1414 1415 ret = ticket->error; 1416 ASSERT(list_empty(&ticket->list)); 1417 /* 1418 * Check that we can't have an error set if the reservation succeeded, 1419 * as that would confuse tasks and lead them to error out without 1420 * releasing reserved space (if an error happens the expectation is that 1421 * space wasn't reserved at all). 1422 */ 1423 ASSERT(!(ticket->bytes == 0 && ticket->error)); 1424 trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, 1425 start_ns, flush, ticket->error); 1426 return ret; 1427 } 1428 1429 /* 1430 * This returns true if this flush state will go through the ordinary flushing 1431 * code. 1432 */ 1433 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) 1434 { 1435 return (flush == BTRFS_RESERVE_FLUSH_ALL) || 1436 (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); 1437 } 1438 1439 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, 1440 struct btrfs_space_info *space_info) 1441 { 1442 u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); 1443 u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); 1444 1445 /* 1446 * If we're heavy on ordered operations then clamping won't help us. We 1447 * need to clamp specifically to keep up with dirty'ing buffered 1448 * writers, because there's not a 1:1 correlation of writing delalloc 1449 * and freeing space, like there is with flushing delayed refs or 1450 * delayed nodes. If we're already more ordered than delalloc then 1451 * we're keeping up, otherwise we aren't and should probably clamp. 1452 */ 1453 if (ordered < delalloc) 1454 space_info->clamp = min(space_info->clamp + 1, 8); 1455 } 1456 1457 static inline bool can_steal(enum btrfs_reserve_flush_enum flush) 1458 { 1459 return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1460 flush == BTRFS_RESERVE_FLUSH_EVICT); 1461 } 1462 1463 /** 1464 * Try to reserve bytes from the block_rsv's space 1465 * 1466 * @fs_info: the filesystem 1467 * @space_info: space info we want to allocate from 1468 * @orig_bytes: number of bytes we want 1469 * @flush: whether or not we can flush to make our reservation 1470 * 1471 * This will reserve orig_bytes number of bytes from the space info associated 1472 * with the block_rsv. If there is not enough space it will make an attempt to 1473 * flush out space to make room. It will do this by flushing delalloc if 1474 * possible or committing the transaction. If flush is 0 then no attempts to 1475 * regain reservations will be made and this will fail if there is not enough 1476 * space already. 1477 */ 1478 static int __reserve_bytes(struct btrfs_fs_info *fs_info, 1479 struct btrfs_space_info *space_info, u64 orig_bytes, 1480 enum btrfs_reserve_flush_enum flush) 1481 { 1482 struct work_struct *async_work; 1483 struct reserve_ticket ticket; 1484 u64 start_ns = 0; 1485 u64 used; 1486 int ret = 0; 1487 bool pending_tickets; 1488 1489 ASSERT(orig_bytes); 1490 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); 1491 1492 if (flush == BTRFS_RESERVE_FLUSH_DATA) 1493 async_work = &fs_info->async_data_reclaim_work; 1494 else 1495 async_work = &fs_info->async_reclaim_work; 1496 1497 spin_lock(&space_info->lock); 1498 ret = -ENOSPC; 1499 used = btrfs_space_info_used(space_info, true); 1500 1501 /* 1502 * We don't want NO_FLUSH allocations to jump everybody, they can 1503 * generally handle ENOSPC in a different way, so treat them the same as 1504 * normal flushers when it comes to skipping pending tickets. 1505 */ 1506 if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) 1507 pending_tickets = !list_empty(&space_info->tickets) || 1508 !list_empty(&space_info->priority_tickets); 1509 else 1510 pending_tickets = !list_empty(&space_info->priority_tickets); 1511 1512 /* 1513 * Carry on if we have enough space (short-circuit) OR call 1514 * can_overcommit() to ensure we can overcommit to continue. 1515 */ 1516 if (!pending_tickets && 1517 ((used + orig_bytes <= space_info->total_bytes) || 1518 btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { 1519 btrfs_space_info_update_bytes_may_use(fs_info, space_info, 1520 orig_bytes); 1521 ret = 0; 1522 } 1523 1524 /* 1525 * If we couldn't make a reservation then setup our reservation ticket 1526 * and kick the async worker if it's not already running. 1527 * 1528 * If we are a priority flusher then we just need to add our ticket to 1529 * the list and we will do our own flushing further down. 1530 */ 1531 if (ret && flush != BTRFS_RESERVE_NO_FLUSH) { 1532 ticket.bytes = orig_bytes; 1533 ticket.error = 0; 1534 space_info->reclaim_size += ticket.bytes; 1535 init_waitqueue_head(&ticket.wait); 1536 ticket.steal = can_steal(flush); 1537 if (trace_btrfs_reserve_ticket_enabled()) 1538 start_ns = ktime_get_ns(); 1539 1540 if (flush == BTRFS_RESERVE_FLUSH_ALL || 1541 flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || 1542 flush == BTRFS_RESERVE_FLUSH_DATA) { 1543 list_add_tail(&ticket.list, &space_info->tickets); 1544 if (!space_info->flush) { 1545 /* 1546 * We were forced to add a reserve ticket, so 1547 * our preemptive flushing is unable to keep 1548 * up. Clamp down on the threshold for the 1549 * preemptive flushing in order to keep up with 1550 * the workload. 1551 */ 1552 maybe_clamp_preempt(fs_info, space_info); 1553 1554 space_info->flush = 1; 1555 trace_btrfs_trigger_flush(fs_info, 1556 space_info->flags, 1557 orig_bytes, flush, 1558 "enospc"); 1559 queue_work(system_unbound_wq, async_work); 1560 } 1561 } else { 1562 list_add_tail(&ticket.list, 1563 &space_info->priority_tickets); 1564 } 1565 } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { 1566 used += orig_bytes; 1567 /* 1568 * We will do the space reservation dance during log replay, 1569 * which means we won't have fs_info->fs_root set, so don't do 1570 * the async reclaim as we will panic. 1571 */ 1572 if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && 1573 !work_busy(&fs_info->preempt_reclaim_work) && 1574 need_preemptive_reclaim(fs_info, space_info)) { 1575 trace_btrfs_trigger_flush(fs_info, space_info->flags, 1576 orig_bytes, flush, "preempt"); 1577 queue_work(system_unbound_wq, 1578 &fs_info->preempt_reclaim_work); 1579 } 1580 } 1581 spin_unlock(&space_info->lock); 1582 if (!ret || flush == BTRFS_RESERVE_NO_FLUSH) 1583 return ret; 1584 1585 return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, 1586 orig_bytes, flush); 1587 } 1588 1589 /** 1590 * Trye to reserve metadata bytes from the block_rsv's space 1591 * 1592 * @fs_info: the filesystem 1593 * @block_rsv: block_rsv we're allocating for 1594 * @orig_bytes: number of bytes we want 1595 * @flush: whether or not we can flush to make our reservation 1596 * 1597 * This will reserve orig_bytes number of bytes from the space info associated 1598 * with the block_rsv. If there is not enough space it will make an attempt to 1599 * flush out space to make room. It will do this by flushing delalloc if 1600 * possible or committing the transaction. If flush is 0 then no attempts to 1601 * regain reservations will be made and this will fail if there is not enough 1602 * space already. 1603 */ 1604 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info, 1605 struct btrfs_block_rsv *block_rsv, 1606 u64 orig_bytes, 1607 enum btrfs_reserve_flush_enum flush) 1608 { 1609 int ret; 1610 1611 ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); 1612 if (ret == -ENOSPC) { 1613 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1614 block_rsv->space_info->flags, 1615 orig_bytes, 1); 1616 1617 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1618 btrfs_dump_space_info(fs_info, block_rsv->space_info, 1619 orig_bytes, 0); 1620 } 1621 return ret; 1622 } 1623 1624 /** 1625 * Try to reserve data bytes for an allocation 1626 * 1627 * @fs_info: the filesystem 1628 * @bytes: number of bytes we need 1629 * @flush: how we are allowed to flush 1630 * 1631 * This will reserve bytes from the data space info. If there is not enough 1632 * space then we will attempt to flush space as specified by flush. 1633 */ 1634 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, 1635 enum btrfs_reserve_flush_enum flush) 1636 { 1637 struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; 1638 int ret; 1639 1640 ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || 1641 flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE); 1642 ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); 1643 1644 ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); 1645 if (ret == -ENOSPC) { 1646 trace_btrfs_space_reservation(fs_info, "space_info:enospc", 1647 data_sinfo->flags, bytes, 1); 1648 if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) 1649 btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); 1650 } 1651 return ret; 1652 } 1653