Commit | Line | Data |
---|---|---|
280c2908 JB |
1 | // SPDX-License-Identifier: GPL-2.0 |
2 | ||
784352fe | 3 | #include "misc.h" |
280c2908 JB |
4 | #include "ctree.h" |
5 | #include "space-info.h" | |
6 | #include "sysfs.h" | |
7 | #include "volumes.h" | |
5da6afeb | 8 | #include "free-space-cache.h" |
0d9764f6 JB |
9 | #include "ordered-data.h" |
10 | #include "transaction.h" | |
aac0023c | 11 | #include "block-group.h" |
b0931513 | 12 | #include "zoned.h" |
c7f13d42 | 13 | #include "fs.h" |
07e81dc9 | 14 | #include "accessors.h" |
a0231804 | 15 | #include "extent-tree.h" |
280c2908 | 16 | |
4b8b0528 JB |
17 | /* |
18 | * HOW DOES SPACE RESERVATION WORK | |
19 | * | |
20 | * If you want to know about delalloc specifically, there is a separate comment | |
21 | * for that with the delalloc code. This comment is about how the whole system | |
22 | * works generally. | |
23 | * | |
24 | * BASIC CONCEPTS | |
25 | * | |
26 | * 1) space_info. This is the ultimate arbiter of how much space we can use. | |
27 | * There's a description of the bytes_ fields with the struct declaration, | |
28 | * refer to that for specifics on each field. Suffice it to say that for | |
29 | * reservations we care about total_bytes - SUM(space_info->bytes_) when | |
30 | * determining if there is space to make an allocation. There is a space_info | |
31 | * for METADATA, SYSTEM, and DATA areas. | |
32 | * | |
33 | * 2) block_rsv's. These are basically buckets for every different type of | |
34 | * metadata reservation we have. You can see the comment in the block_rsv | |
35 | * code on the rules for each type, but generally block_rsv->reserved is how | |
36 | * much space is accounted for in space_info->bytes_may_use. | |
37 | * | |
38 | * 3) btrfs_calc*_size. These are the worst case calculations we used based | |
39 | * on the number of items we will want to modify. We have one for changing | |
40 | * items, and one for inserting new items. Generally we use these helpers to | |
41 | * determine the size of the block reserves, and then use the actual bytes | |
42 | * values to adjust the space_info counters. | |
43 | * | |
44 | * MAKING RESERVATIONS, THE NORMAL CASE | |
45 | * | |
46 | * We call into either btrfs_reserve_data_bytes() or | |
47 | * btrfs_reserve_metadata_bytes(), depending on which we're looking for, with | |
48 | * num_bytes we want to reserve. | |
49 | * | |
50 | * ->reserve | |
51 | * space_info->bytes_may_reserve += num_bytes | |
52 | * | |
53 | * ->extent allocation | |
54 | * Call btrfs_add_reserved_bytes() which does | |
55 | * space_info->bytes_may_reserve -= num_bytes | |
56 | * space_info->bytes_reserved += extent_bytes | |
57 | * | |
58 | * ->insert reference | |
59 | * Call btrfs_update_block_group() which does | |
60 | * space_info->bytes_reserved -= extent_bytes | |
61 | * space_info->bytes_used += extent_bytes | |
62 | * | |
63 | * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority) | |
64 | * | |
65 | * Assume we are unable to simply make the reservation because we do not have | |
66 | * enough space | |
67 | * | |
68 | * -> __reserve_bytes | |
69 | * create a reserve_ticket with ->bytes set to our reservation, add it to | |
70 | * the tail of space_info->tickets, kick async flush thread | |
71 | * | |
72 | * ->handle_reserve_ticket | |
73 | * wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set | |
74 | * on the ticket. | |
75 | * | |
76 | * -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space | |
77 | * Flushes various things attempting to free up space. | |
78 | * | |
79 | * -> btrfs_try_granting_tickets() | |
80 | * This is called by anything that either subtracts space from | |
81 | * space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the | |
82 | * space_info->total_bytes. This loops through the ->priority_tickets and | |
83 | * then the ->tickets list checking to see if the reservation can be | |
84 | * completed. If it can the space is added to space_info->bytes_may_use and | |
85 | * the ticket is woken up. | |
86 | * | |
87 | * -> ticket wakeup | |
88 | * Check if ->bytes == 0, if it does we got our reservation and we can carry | |
89 | * on, if not return the appropriate error (ENOSPC, but can be EINTR if we | |
90 | * were interrupted.) | |
91 | * | |
92 | * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY | |
93 | * | |
94 | * Same as the above, except we add ourselves to the | |
95 | * space_info->priority_tickets, and we do not use ticket->wait, we simply | |
96 | * call flush_space() ourselves for the states that are safe for us to call | |
97 | * without deadlocking and hope for the best. | |
98 | * | |
99 | * THE FLUSHING STATES | |
100 | * | |
101 | * Generally speaking we will have two cases for each state, a "nice" state | |
102 | * and a "ALL THE THINGS" state. In btrfs we delay a lot of work in order to | |
103 | * reduce the locking over head on the various trees, and even to keep from | |
104 | * doing any work at all in the case of delayed refs. Each of these delayed | |
105 | * things however hold reservations, and so letting them run allows us to | |
106 | * reclaim space so we can make new reservations. | |
107 | * | |
108 | * FLUSH_DELAYED_ITEMS | |
109 | * Every inode has a delayed item to update the inode. Take a simple write | |
110 | * for example, we would update the inode item at write time to update the | |
111 | * mtime, and then again at finish_ordered_io() time in order to update the | |
112 | * isize or bytes. We keep these delayed items to coalesce these operations | |
113 | * into a single operation done on demand. These are an easy way to reclaim | |
114 | * metadata space. | |
115 | * | |
116 | * FLUSH_DELALLOC | |
117 | * Look at the delalloc comment to get an idea of how much space is reserved | |
118 | * for delayed allocation. We can reclaim some of this space simply by | |
119 | * running delalloc, but usually we need to wait for ordered extents to | |
120 | * reclaim the bulk of this space. | |
121 | * | |
122 | * FLUSH_DELAYED_REFS | |
123 | * We have a block reserve for the outstanding delayed refs space, and every | |
124 | * delayed ref operation holds a reservation. Running these is a quick way | |
125 | * to reclaim space, but we want to hold this until the end because COW can | |
126 | * churn a lot and we can avoid making some extent tree modifications if we | |
127 | * are able to delay for as long as possible. | |
128 | * | |
129 | * ALLOC_CHUNK | |
130 | * We will skip this the first time through space reservation, because of | |
131 | * overcommit and we don't want to have a lot of useless metadata space when | |
132 | * our worst case reservations will likely never come true. | |
133 | * | |
134 | * RUN_DELAYED_IPUTS | |
135 | * If we're freeing inodes we're likely freeing checksums, file extent | |
136 | * items, and extent tree items. Loads of space could be freed up by these | |
137 | * operations, however they won't be usable until the transaction commits. | |
138 | * | |
139 | * COMMIT_TRANS | |
c416a30c JB |
140 | * This will commit the transaction. Historically we had a lot of logic |
141 | * surrounding whether or not we'd commit the transaction, but this waits born | |
142 | * out of a pre-tickets era where we could end up committing the transaction | |
143 | * thousands of times in a row without making progress. Now thanks to our | |
144 | * ticketing system we know if we're not making progress and can error | |
145 | * everybody out after a few commits rather than burning the disk hoping for | |
146 | * a different answer. | |
f00c42dd | 147 | * |
4b8b0528 JB |
148 | * OVERCOMMIT |
149 | * | |
150 | * Because we hold so many reservations for metadata we will allow you to | |
151 | * reserve more space than is currently free in the currently allocate | |
152 | * metadata space. This only happens with metadata, data does not allow | |
153 | * overcommitting. | |
154 | * | |
155 | * You can see the current logic for when we allow overcommit in | |
156 | * btrfs_can_overcommit(), but it only applies to unallocated space. If there | |
157 | * is no unallocated space to be had, all reservations are kept within the | |
158 | * free space in the allocated metadata chunks. | |
159 | * | |
160 | * Because of overcommitting, you generally want to use the | |
161 | * btrfs_can_overcommit() logic for metadata allocations, as it does the right | |
162 | * thing with or without extra unallocated space. | |
163 | */ | |
164 | ||
e1f60a65 | 165 | u64 __pure btrfs_space_info_used(struct btrfs_space_info *s_info, |
280c2908 JB |
166 | bool may_use_included) |
167 | { | |
168 | ASSERT(s_info); | |
169 | return s_info->bytes_used + s_info->bytes_reserved + | |
170 | s_info->bytes_pinned + s_info->bytes_readonly + | |
169e0da9 | 171 | s_info->bytes_zone_unusable + |
280c2908 JB |
172 | (may_use_included ? s_info->bytes_may_use : 0); |
173 | } | |
174 | ||
175 | /* | |
176 | * after adding space to the filesystem, we need to clear the full flags | |
177 | * on all the space infos. | |
178 | */ | |
179 | void btrfs_clear_space_info_full(struct btrfs_fs_info *info) | |
180 | { | |
181 | struct list_head *head = &info->space_info; | |
182 | struct btrfs_space_info *found; | |
183 | ||
72804905 | 184 | list_for_each_entry(found, head, list) |
280c2908 | 185 | found->full = 0; |
280c2908 JB |
186 | } |
187 | ||
bb5a098d JB |
188 | /* |
189 | * Block groups with more than this value (percents) of unusable space will be | |
190 | * scheduled for background reclaim. | |
191 | */ | |
192 | #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH (75) | |
193 | ||
f6fca391 SR |
194 | /* |
195 | * Calculate chunk size depending on volume type (regular or zoned). | |
196 | */ | |
197 | static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags) | |
198 | { | |
199 | if (btrfs_is_zoned(fs_info)) | |
200 | return fs_info->zone_size; | |
201 | ||
202 | ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK); | |
203 | ||
204 | if (flags & BTRFS_BLOCK_GROUP_DATA) | |
5da431b7 | 205 | return BTRFS_MAX_DATA_CHUNK_SIZE; |
f6fca391 SR |
206 | else if (flags & BTRFS_BLOCK_GROUP_SYSTEM) |
207 | return SZ_32M; | |
208 | ||
209 | /* Handle BTRFS_BLOCK_GROUP_METADATA */ | |
210 | if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G) | |
211 | return SZ_1G; | |
212 | ||
213 | return SZ_256M; | |
214 | } | |
215 | ||
216 | /* | |
217 | * Update default chunk size. | |
218 | */ | |
219 | void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info, | |
220 | u64 chunk_size) | |
221 | { | |
222 | WRITE_ONCE(space_info->chunk_size, chunk_size); | |
223 | } | |
224 | ||
280c2908 JB |
225 | static int create_space_info(struct btrfs_fs_info *info, u64 flags) |
226 | { | |
227 | ||
228 | struct btrfs_space_info *space_info; | |
229 | int i; | |
230 | int ret; | |
231 | ||
232 | space_info = kzalloc(sizeof(*space_info), GFP_NOFS); | |
233 | if (!space_info) | |
234 | return -ENOMEM; | |
235 | ||
280c2908 JB |
236 | for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) |
237 | INIT_LIST_HEAD(&space_info->block_groups[i]); | |
238 | init_rwsem(&space_info->groups_sem); | |
239 | spin_lock_init(&space_info->lock); | |
240 | space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK; | |
241 | space_info->force_alloc = CHUNK_ALLOC_NO_FORCE; | |
280c2908 JB |
242 | INIT_LIST_HEAD(&space_info->ro_bgs); |
243 | INIT_LIST_HEAD(&space_info->tickets); | |
244 | INIT_LIST_HEAD(&space_info->priority_tickets); | |
88a777a6 | 245 | space_info->clamp = 1; |
f6fca391 | 246 | btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags)); |
280c2908 | 247 | |
bb5a098d JB |
248 | if (btrfs_is_zoned(info)) |
249 | space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH; | |
250 | ||
b882327a DS |
251 | ret = btrfs_sysfs_add_space_info_type(info, space_info); |
252 | if (ret) | |
280c2908 | 253 | return ret; |
280c2908 | 254 | |
72804905 | 255 | list_add(&space_info->list, &info->space_info); |
280c2908 JB |
256 | if (flags & BTRFS_BLOCK_GROUP_DATA) |
257 | info->data_sinfo = space_info; | |
258 | ||
259 | return ret; | |
260 | } | |
261 | ||
262 | int btrfs_init_space_info(struct btrfs_fs_info *fs_info) | |
263 | { | |
264 | struct btrfs_super_block *disk_super; | |
265 | u64 features; | |
266 | u64 flags; | |
267 | int mixed = 0; | |
268 | int ret; | |
269 | ||
270 | disk_super = fs_info->super_copy; | |
271 | if (!btrfs_super_root(disk_super)) | |
272 | return -EINVAL; | |
273 | ||
274 | features = btrfs_super_incompat_flags(disk_super); | |
275 | if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS) | |
276 | mixed = 1; | |
277 | ||
278 | flags = BTRFS_BLOCK_GROUP_SYSTEM; | |
279 | ret = create_space_info(fs_info, flags); | |
280 | if (ret) | |
281 | goto out; | |
282 | ||
283 | if (mixed) { | |
284 | flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA; | |
285 | ret = create_space_info(fs_info, flags); | |
286 | } else { | |
287 | flags = BTRFS_BLOCK_GROUP_METADATA; | |
288 | ret = create_space_info(fs_info, flags); | |
289 | if (ret) | |
290 | goto out; | |
291 | ||
292 | flags = BTRFS_BLOCK_GROUP_DATA; | |
293 | ret = create_space_info(fs_info, flags); | |
294 | } | |
295 | out: | |
296 | return ret; | |
297 | } | |
298 | ||
9d4b0a12 | 299 | void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info, |
723de71d | 300 | struct btrfs_block_group *block_group) |
280c2908 JB |
301 | { |
302 | struct btrfs_space_info *found; | |
723de71d | 303 | int factor, index; |
280c2908 | 304 | |
9d4b0a12 | 305 | factor = btrfs_bg_type_to_factor(block_group->flags); |
280c2908 | 306 | |
9d4b0a12 | 307 | found = btrfs_find_space_info(info, block_group->flags); |
280c2908 JB |
308 | ASSERT(found); |
309 | spin_lock(&found->lock); | |
9d4b0a12 | 310 | found->total_bytes += block_group->length; |
3349b57f | 311 | if (test_bit(BLOCK_GROUP_FLAG_ZONE_IS_ACTIVE, &block_group->runtime_flags)) |
9d4b0a12 JB |
312 | found->active_total_bytes += block_group->length; |
313 | found->disk_total += block_group->length * factor; | |
314 | found->bytes_used += block_group->used; | |
315 | found->disk_used += block_group->used * factor; | |
316 | found->bytes_readonly += block_group->bytes_super; | |
317 | found->bytes_zone_unusable += block_group->zone_unusable; | |
318 | if (block_group->length > 0) | |
280c2908 | 319 | found->full = 0; |
18fa2284 | 320 | btrfs_try_granting_tickets(info, found); |
280c2908 | 321 | spin_unlock(&found->lock); |
723de71d JB |
322 | |
323 | block_group->space_info = found; | |
324 | ||
325 | index = btrfs_bg_flags_to_raid_index(block_group->flags); | |
326 | down_write(&found->groups_sem); | |
327 | list_add_tail(&block_group->list, &found->block_groups[index]); | |
328 | up_write(&found->groups_sem); | |
280c2908 JB |
329 | } |
330 | ||
331 | struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info, | |
332 | u64 flags) | |
333 | { | |
334 | struct list_head *head = &info->space_info; | |
335 | struct btrfs_space_info *found; | |
336 | ||
337 | flags &= BTRFS_BLOCK_GROUP_TYPE_MASK; | |
338 | ||
72804905 JB |
339 | list_for_each_entry(found, head, list) { |
340 | if (found->flags & flags) | |
280c2908 | 341 | return found; |
280c2908 | 342 | } |
280c2908 JB |
343 | return NULL; |
344 | } | |
41783ef2 | 345 | |
fa121a26 JB |
346 | static u64 calc_available_free_space(struct btrfs_fs_info *fs_info, |
347 | struct btrfs_space_info *space_info, | |
348 | enum btrfs_reserve_flush_enum flush) | |
41783ef2 | 349 | { |
41783ef2 | 350 | u64 profile; |
41783ef2 | 351 | u64 avail; |
41783ef2 JB |
352 | int factor; |
353 | ||
9f246926 | 354 | if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM) |
41783ef2 JB |
355 | profile = btrfs_system_alloc_profile(fs_info); |
356 | else | |
357 | profile = btrfs_metadata_alloc_profile(fs_info); | |
358 | ||
41783ef2 JB |
359 | avail = atomic64_read(&fs_info->free_chunk_space); |
360 | ||
361 | /* | |
362 | * If we have dup, raid1 or raid10 then only half of the free | |
363 | * space is actually usable. For raid56, the space info used | |
364 | * doesn't include the parity drive, so we don't have to | |
365 | * change the math | |
366 | */ | |
367 | factor = btrfs_bg_type_to_factor(profile); | |
368 | avail = div_u64(avail, factor); | |
369 | ||
370 | /* | |
371 | * If we aren't flushing all things, let us overcommit up to | |
372 | * 1/2th of the space. If we can flush, don't let us overcommit | |
373 | * too much, let it overcommit up to 1/8 of the space. | |
374 | */ | |
375 | if (flush == BTRFS_RESERVE_FLUSH_ALL) | |
376 | avail >>= 3; | |
377 | else | |
378 | avail >>= 1; | |
fa121a26 JB |
379 | return avail; |
380 | } | |
381 | ||
6a921de5 NA |
382 | static inline u64 writable_total_bytes(struct btrfs_fs_info *fs_info, |
383 | struct btrfs_space_info *space_info) | |
384 | { | |
385 | /* | |
386 | * On regular filesystem, all total_bytes are always writable. On zoned | |
387 | * filesystem, there may be a limitation imposed by max_active_zones. | |
388 | * For metadata allocation, we cannot finish an existing active block | |
389 | * group to avoid a deadlock. Thus, we need to consider only the active | |
390 | * groups to be writable for metadata space. | |
391 | */ | |
392 | if (!btrfs_is_zoned(fs_info) || (space_info->flags & BTRFS_BLOCK_GROUP_DATA)) | |
393 | return space_info->total_bytes; | |
394 | ||
395 | return space_info->active_total_bytes; | |
396 | } | |
397 | ||
fa121a26 JB |
398 | int btrfs_can_overcommit(struct btrfs_fs_info *fs_info, |
399 | struct btrfs_space_info *space_info, u64 bytes, | |
400 | enum btrfs_reserve_flush_enum flush) | |
401 | { | |
402 | u64 avail; | |
403 | u64 used; | |
404 | ||
405 | /* Don't overcommit when in mixed mode */ | |
406 | if (space_info->flags & BTRFS_BLOCK_GROUP_DATA) | |
407 | return 0; | |
408 | ||
409 | used = btrfs_space_info_used(space_info, true); | |
79417d04 NA |
410 | if (btrfs_is_zoned(fs_info) && (space_info->flags & BTRFS_BLOCK_GROUP_METADATA)) |
411 | avail = 0; | |
412 | else | |
413 | avail = calc_available_free_space(fs_info, space_info, flush); | |
41783ef2 | 414 | |
6a921de5 | 415 | if (used + bytes < writable_total_bytes(fs_info, space_info) + avail) |
41783ef2 JB |
416 | return 1; |
417 | return 0; | |
418 | } | |
b338b013 | 419 | |
d611add4 FM |
420 | static void remove_ticket(struct btrfs_space_info *space_info, |
421 | struct reserve_ticket *ticket) | |
422 | { | |
423 | if (!list_empty(&ticket->list)) { | |
424 | list_del_init(&ticket->list); | |
425 | ASSERT(space_info->reclaim_size >= ticket->bytes); | |
426 | space_info->reclaim_size -= ticket->bytes; | |
427 | } | |
428 | } | |
429 | ||
b338b013 JB |
430 | /* |
431 | * This is for space we already have accounted in space_info->bytes_may_use, so | |
432 | * basically when we're returning space from block_rsv's. | |
433 | */ | |
18fa2284 JB |
434 | void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info, |
435 | struct btrfs_space_info *space_info) | |
b338b013 | 436 | { |
b338b013 | 437 | struct list_head *head; |
b338b013 | 438 | enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH; |
b338b013 | 439 | |
18fa2284 | 440 | lockdep_assert_held(&space_info->lock); |
b338b013 | 441 | |
18fa2284 | 442 | head = &space_info->priority_tickets; |
b338b013 | 443 | again: |
91182645 JB |
444 | while (!list_empty(head)) { |
445 | struct reserve_ticket *ticket; | |
446 | u64 used = btrfs_space_info_used(space_info, true); | |
447 | ||
448 | ticket = list_first_entry(head, struct reserve_ticket, list); | |
449 | ||
1a9fd417 | 450 | /* Check and see if our ticket can be satisfied now. */ |
6a921de5 | 451 | if ((used + ticket->bytes <= writable_total_bytes(fs_info, space_info)) || |
a30a3d20 JB |
452 | btrfs_can_overcommit(fs_info, space_info, ticket->bytes, |
453 | flush)) { | |
91182645 JB |
454 | btrfs_space_info_update_bytes_may_use(fs_info, |
455 | space_info, | |
456 | ticket->bytes); | |
d611add4 | 457 | remove_ticket(space_info, ticket); |
b338b013 JB |
458 | ticket->bytes = 0; |
459 | space_info->tickets_id++; | |
460 | wake_up(&ticket->wait); | |
461 | } else { | |
91182645 | 462 | break; |
b338b013 JB |
463 | } |
464 | } | |
465 | ||
91182645 | 466 | if (head == &space_info->priority_tickets) { |
b338b013 JB |
467 | head = &space_info->tickets; |
468 | flush = BTRFS_RESERVE_FLUSH_ALL; | |
469 | goto again; | |
470 | } | |
b338b013 | 471 | } |
5da6afeb JB |
472 | |
473 | #define DUMP_BLOCK_RSV(fs_info, rsv_name) \ | |
474 | do { \ | |
475 | struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name; \ | |
476 | spin_lock(&__rsv->lock); \ | |
477 | btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu", \ | |
478 | __rsv->size, __rsv->reserved); \ | |
479 | spin_unlock(&__rsv->lock); \ | |
480 | } while (0) | |
481 | ||
25a860c4 QW |
482 | static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info) |
483 | { | |
484 | switch (space_info->flags) { | |
485 | case BTRFS_BLOCK_GROUP_SYSTEM: | |
486 | return "SYSTEM"; | |
487 | case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA: | |
488 | return "DATA+METADATA"; | |
489 | case BTRFS_BLOCK_GROUP_DATA: | |
490 | return "DATA"; | |
491 | case BTRFS_BLOCK_GROUP_METADATA: | |
492 | return "METADATA"; | |
493 | default: | |
494 | return "UNKNOWN"; | |
495 | } | |
496 | } | |
497 | ||
8e327b9c QW |
498 | static void dump_global_block_rsv(struct btrfs_fs_info *fs_info) |
499 | { | |
500 | DUMP_BLOCK_RSV(fs_info, global_block_rsv); | |
501 | DUMP_BLOCK_RSV(fs_info, trans_block_rsv); | |
502 | DUMP_BLOCK_RSV(fs_info, chunk_block_rsv); | |
503 | DUMP_BLOCK_RSV(fs_info, delayed_block_rsv); | |
504 | DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv); | |
505 | } | |
506 | ||
84fe47a4 JB |
507 | static void __btrfs_dump_space_info(struct btrfs_fs_info *fs_info, |
508 | struct btrfs_space_info *info) | |
5da6afeb | 509 | { |
25a860c4 | 510 | const char *flag_str = space_info_flag_to_str(info); |
84fe47a4 | 511 | lockdep_assert_held(&info->lock); |
5da6afeb | 512 | |
0619b790 | 513 | /* The free space could be negative in case of overcommit */ |
25a860c4 QW |
514 | btrfs_info(fs_info, "space_info %s has %lld free, is %sfull", |
515 | flag_str, | |
0619b790 | 516 | (s64)(info->total_bytes - btrfs_space_info_used(info, true)), |
5da6afeb JB |
517 | info->full ? "" : "not "); |
518 | btrfs_info(fs_info, | |
25a860c4 | 519 | "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu", |
5da6afeb JB |
520 | info->total_bytes, info->bytes_used, info->bytes_pinned, |
521 | info->bytes_reserved, info->bytes_may_use, | |
169e0da9 | 522 | info->bytes_readonly, info->bytes_zone_unusable); |
84fe47a4 JB |
523 | } |
524 | ||
525 | void btrfs_dump_space_info(struct btrfs_fs_info *fs_info, | |
526 | struct btrfs_space_info *info, u64 bytes, | |
527 | int dump_block_groups) | |
528 | { | |
32da5386 | 529 | struct btrfs_block_group *cache; |
84fe47a4 JB |
530 | int index = 0; |
531 | ||
532 | spin_lock(&info->lock); | |
533 | __btrfs_dump_space_info(fs_info, info); | |
8e327b9c | 534 | dump_global_block_rsv(fs_info); |
84fe47a4 JB |
535 | spin_unlock(&info->lock); |
536 | ||
5da6afeb JB |
537 | if (!dump_block_groups) |
538 | return; | |
539 | ||
540 | down_read(&info->groups_sem); | |
541 | again: | |
542 | list_for_each_entry(cache, &info->block_groups[index], list) { | |
543 | spin_lock(&cache->lock); | |
544 | btrfs_info(fs_info, | |
169e0da9 | 545 | "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu zone_unusable %s", |
b3470b5d | 546 | cache->start, cache->length, cache->used, cache->pinned, |
169e0da9 NA |
547 | cache->reserved, cache->zone_unusable, |
548 | cache->ro ? "[readonly]" : ""); | |
5da6afeb | 549 | spin_unlock(&cache->lock); |
ab0db043 | 550 | btrfs_dump_free_space(cache, bytes); |
5da6afeb JB |
551 | } |
552 | if (++index < BTRFS_NR_RAID_TYPES) | |
553 | goto again; | |
554 | up_read(&info->groups_sem); | |
555 | } | |
0d9764f6 | 556 | |
0d9764f6 JB |
557 | static inline u64 calc_reclaim_items_nr(struct btrfs_fs_info *fs_info, |
558 | u64 to_reclaim) | |
559 | { | |
560 | u64 bytes; | |
561 | u64 nr; | |
562 | ||
2bd36e7b | 563 | bytes = btrfs_calc_insert_metadata_size(fs_info, 1); |
0d9764f6 JB |
564 | nr = div64_u64(to_reclaim, bytes); |
565 | if (!nr) | |
566 | nr = 1; | |
567 | return nr; | |
568 | } | |
569 | ||
570 | #define EXTENT_SIZE_PER_ITEM SZ_256K | |
571 | ||
572 | /* | |
573 | * shrink metadata reservation for delalloc | |
574 | */ | |
920a9958 JB |
575 | static void shrink_delalloc(struct btrfs_fs_info *fs_info, |
576 | struct btrfs_space_info *space_info, | |
385f421f JB |
577 | u64 to_reclaim, bool wait_ordered, |
578 | bool for_preempt) | |
0d9764f6 | 579 | { |
0d9764f6 JB |
580 | struct btrfs_trans_handle *trans; |
581 | u64 delalloc_bytes; | |
5deb17e1 | 582 | u64 ordered_bytes; |
0d9764f6 JB |
583 | u64 items; |
584 | long time_left; | |
0d9764f6 JB |
585 | int loops; |
586 | ||
03fe78cc JB |
587 | delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes); |
588 | ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes); | |
589 | if (delalloc_bytes == 0 && ordered_bytes == 0) | |
590 | return; | |
591 | ||
0d9764f6 | 592 | /* Calc the number of the pages we need flush for space reservation */ |
d7f81fac JB |
593 | if (to_reclaim == U64_MAX) { |
594 | items = U64_MAX; | |
595 | } else { | |
596 | /* | |
597 | * to_reclaim is set to however much metadata we need to | |
598 | * reclaim, but reclaiming that much data doesn't really track | |
03fe78cc JB |
599 | * exactly. What we really want to do is reclaim full inode's |
600 | * worth of reservations, however that's not available to us | |
601 | * here. We will take a fraction of the delalloc bytes for our | |
602 | * flushing loops and hope for the best. Delalloc will expand | |
603 | * the amount we write to cover an entire dirty extent, which | |
604 | * will reclaim the metadata reservation for that range. If | |
605 | * it's not enough subsequent flush stages will be more | |
606 | * aggressive. | |
d7f81fac | 607 | */ |
03fe78cc | 608 | to_reclaim = max(to_reclaim, delalloc_bytes >> 3); |
d7f81fac | 609 | items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2; |
d7f81fac | 610 | } |
0d9764f6 | 611 | |
0d031dc4 | 612 | trans = current->journal_info; |
0d9764f6 | 613 | |
0d9764f6 JB |
614 | /* |
615 | * If we are doing more ordered than delalloc we need to just wait on | |
616 | * ordered extents, otherwise we'll waste time trying to flush delalloc | |
617 | * that likely won't give us the space back we need. | |
618 | */ | |
385f421f | 619 | if (ordered_bytes > delalloc_bytes && !for_preempt) |
0d9764f6 JB |
620 | wait_ordered = true; |
621 | ||
622 | loops = 0; | |
5deb17e1 | 623 | while ((delalloc_bytes || ordered_bytes) && loops < 3) { |
9db4dc24 NB |
624 | u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT; |
625 | long nr_pages = min_t(u64, temp, LONG_MAX); | |
e1646070 | 626 | int async_pages; |
e076ab2a JB |
627 | |
628 | btrfs_start_delalloc_roots(fs_info, nr_pages, true); | |
0d9764f6 | 629 | |
e1646070 JB |
630 | /* |
631 | * We need to make sure any outstanding async pages are now | |
632 | * processed before we continue. This is because things like | |
633 | * sync_inode() try to be smart and skip writing if the inode is | |
634 | * marked clean. We don't use filemap_fwrite for flushing | |
635 | * because we want to control how many pages we write out at a | |
636 | * time, thus this is the only safe way to make sure we've | |
637 | * waited for outstanding compressed workers to have started | |
638 | * their jobs and thus have ordered extents set up properly. | |
639 | * | |
640 | * This exists because we do not want to wait for each | |
641 | * individual inode to finish its async work, we simply want to | |
642 | * start the IO on everybody, and then come back here and wait | |
643 | * for all of the async work to catch up. Once we're done with | |
644 | * that we know we'll have ordered extents for everything and we | |
645 | * can decide if we wait for that or not. | |
646 | * | |
647 | * If we choose to replace this in the future, make absolutely | |
648 | * sure that the proper waiting is being done in the async case, | |
649 | * as there have been bugs in that area before. | |
650 | */ | |
651 | async_pages = atomic_read(&fs_info->async_delalloc_pages); | |
652 | if (!async_pages) | |
653 | goto skip_async; | |
654 | ||
655 | /* | |
656 | * We don't want to wait forever, if we wrote less pages in this | |
657 | * loop than we have outstanding, only wait for that number of | |
658 | * pages, otherwise we can wait for all async pages to finish | |
659 | * before continuing. | |
660 | */ | |
661 | if (async_pages > nr_pages) | |
662 | async_pages -= nr_pages; | |
663 | else | |
664 | async_pages = 0; | |
665 | wait_event(fs_info->async_submit_wait, | |
666 | atomic_read(&fs_info->async_delalloc_pages) <= | |
667 | async_pages); | |
668 | skip_async: | |
0d9764f6 JB |
669 | loops++; |
670 | if (wait_ordered && !trans) { | |
671 | btrfs_wait_ordered_roots(fs_info, items, 0, (u64)-1); | |
672 | } else { | |
673 | time_left = schedule_timeout_killable(1); | |
674 | if (time_left) | |
675 | break; | |
676 | } | |
448b966b | 677 | |
385f421f JB |
678 | /* |
679 | * If we are for preemption we just want a one-shot of delalloc | |
680 | * flushing so we can stop flushing if we decide we don't need | |
681 | * to anymore. | |
682 | */ | |
683 | if (for_preempt) | |
684 | break; | |
685 | ||
448b966b JB |
686 | spin_lock(&space_info->lock); |
687 | if (list_empty(&space_info->tickets) && | |
688 | list_empty(&space_info->priority_tickets)) { | |
689 | spin_unlock(&space_info->lock); | |
690 | break; | |
691 | } | |
692 | spin_unlock(&space_info->lock); | |
693 | ||
0d9764f6 JB |
694 | delalloc_bytes = percpu_counter_sum_positive( |
695 | &fs_info->delalloc_bytes); | |
5deb17e1 JB |
696 | ordered_bytes = percpu_counter_sum_positive( |
697 | &fs_info->ordered_bytes); | |
0d9764f6 JB |
698 | } |
699 | } | |
700 | ||
0d9764f6 JB |
701 | /* |
702 | * Try to flush some data based on policy set by @state. This is only advisory | |
703 | * and may fail for various reasons. The caller is supposed to examine the | |
704 | * state of @space_info to detect the outcome. | |
705 | */ | |
706 | static void flush_space(struct btrfs_fs_info *fs_info, | |
707 | struct btrfs_space_info *space_info, u64 num_bytes, | |
4b02b00f | 708 | enum btrfs_flush_state state, bool for_preempt) |
0d9764f6 | 709 | { |
ce5603d0 | 710 | struct btrfs_root *root = fs_info->tree_root; |
0d9764f6 JB |
711 | struct btrfs_trans_handle *trans; |
712 | int nr; | |
713 | int ret = 0; | |
714 | ||
715 | switch (state) { | |
716 | case FLUSH_DELAYED_ITEMS_NR: | |
717 | case FLUSH_DELAYED_ITEMS: | |
718 | if (state == FLUSH_DELAYED_ITEMS_NR) | |
719 | nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2; | |
720 | else | |
721 | nr = -1; | |
722 | ||
723 | trans = btrfs_join_transaction(root); | |
724 | if (IS_ERR(trans)) { | |
725 | ret = PTR_ERR(trans); | |
726 | break; | |
727 | } | |
728 | ret = btrfs_run_delayed_items_nr(trans, nr); | |
729 | btrfs_end_transaction(trans); | |
730 | break; | |
731 | case FLUSH_DELALLOC: | |
732 | case FLUSH_DELALLOC_WAIT: | |
03fe78cc JB |
733 | case FLUSH_DELALLOC_FULL: |
734 | if (state == FLUSH_DELALLOC_FULL) | |
735 | num_bytes = U64_MAX; | |
920a9958 | 736 | shrink_delalloc(fs_info, space_info, num_bytes, |
03fe78cc | 737 | state != FLUSH_DELALLOC, for_preempt); |
0d9764f6 JB |
738 | break; |
739 | case FLUSH_DELAYED_REFS_NR: | |
740 | case FLUSH_DELAYED_REFS: | |
741 | trans = btrfs_join_transaction(root); | |
742 | if (IS_ERR(trans)) { | |
743 | ret = PTR_ERR(trans); | |
744 | break; | |
745 | } | |
746 | if (state == FLUSH_DELAYED_REFS_NR) | |
747 | nr = calc_reclaim_items_nr(fs_info, num_bytes); | |
748 | else | |
749 | nr = 0; | |
750 | btrfs_run_delayed_refs(trans, nr); | |
751 | btrfs_end_transaction(trans); | |
752 | break; | |
753 | case ALLOC_CHUNK: | |
754 | case ALLOC_CHUNK_FORCE: | |
b0931513 NA |
755 | /* |
756 | * For metadata space on zoned filesystem, reaching here means we | |
757 | * don't have enough space left in active_total_bytes. Try to | |
758 | * activate a block group first, because we may have inactive | |
759 | * block group already allocated. | |
760 | */ | |
761 | ret = btrfs_zoned_activate_one_bg(fs_info, space_info, false); | |
762 | if (ret < 0) | |
763 | break; | |
764 | else if (ret == 1) | |
765 | break; | |
766 | ||
0d9764f6 JB |
767 | trans = btrfs_join_transaction(root); |
768 | if (IS_ERR(trans)) { | |
769 | ret = PTR_ERR(trans); | |
770 | break; | |
771 | } | |
772 | ret = btrfs_chunk_alloc(trans, | |
c6c45303 | 773 | btrfs_get_alloc_profile(fs_info, space_info->flags), |
0d9764f6 JB |
774 | (state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE : |
775 | CHUNK_ALLOC_FORCE); | |
776 | btrfs_end_transaction(trans); | |
b0931513 NA |
777 | |
778 | /* | |
779 | * For metadata space on zoned filesystem, allocating a new chunk | |
780 | * is not enough. We still need to activate the block * group. | |
781 | * Active the newly allocated block group by (maybe) finishing | |
782 | * a block group. | |
783 | */ | |
784 | if (ret == 1) { | |
785 | ret = btrfs_zoned_activate_one_bg(fs_info, space_info, true); | |
786 | /* | |
787 | * Revert to the original ret regardless we could finish | |
788 | * one block group or not. | |
789 | */ | |
790 | if (ret >= 0) | |
791 | ret = 1; | |
792 | } | |
793 | ||
0d9764f6 JB |
794 | if (ret > 0 || ret == -ENOSPC) |
795 | ret = 0; | |
796 | break; | |
844245b4 | 797 | case RUN_DELAYED_IPUTS: |
0d9764f6 JB |
798 | /* |
799 | * If we have pending delayed iputs then we could free up a | |
800 | * bunch of pinned space, so make sure we run the iputs before | |
801 | * we do our pinned bytes check below. | |
802 | */ | |
803 | btrfs_run_delayed_iputs(fs_info); | |
804 | btrfs_wait_on_delayed_iputs(fs_info); | |
844245b4 JB |
805 | break; |
806 | case COMMIT_TRANS: | |
c416a30c | 807 | ASSERT(current->journal_info == NULL); |
f00c42dd JB |
808 | trans = btrfs_join_transaction(root); |
809 | if (IS_ERR(trans)) { | |
810 | ret = PTR_ERR(trans); | |
811 | break; | |
812 | } | |
813 | ret = btrfs_commit_transaction(trans); | |
814 | break; | |
0d9764f6 JB |
815 | default: |
816 | ret = -ENOSPC; | |
817 | break; | |
818 | } | |
819 | ||
820 | trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state, | |
4b02b00f | 821 | ret, for_preempt); |
0d9764f6 JB |
822 | return; |
823 | } | |
824 | ||
825 | static inline u64 | |
826 | btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info, | |
9f246926 | 827 | struct btrfs_space_info *space_info) |
0d9764f6 | 828 | { |
0d9764f6 | 829 | u64 used; |
fa121a26 | 830 | u64 avail; |
6a921de5 | 831 | u64 total; |
db161806 | 832 | u64 to_reclaim = space_info->reclaim_size; |
0d9764f6 | 833 | |
db161806 | 834 | lockdep_assert_held(&space_info->lock); |
fa121a26 JB |
835 | |
836 | avail = calc_available_free_space(fs_info, space_info, | |
837 | BTRFS_RESERVE_FLUSH_ALL); | |
838 | used = btrfs_space_info_used(space_info, true); | |
839 | ||
840 | /* | |
841 | * We may be flushing because suddenly we have less space than we had | |
842 | * before, and now we're well over-committed based on our current free | |
843 | * space. If that's the case add in our overage so we make sure to put | |
844 | * appropriate pressure on the flushing state machine. | |
845 | */ | |
6a921de5 NA |
846 | total = writable_total_bytes(fs_info, space_info); |
847 | if (total + avail < used) | |
848 | to_reclaim += used - (total + avail); | |
fa121a26 | 849 | |
0d9764f6 JB |
850 | return to_reclaim; |
851 | } | |
852 | ||
ae7913ba | 853 | static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info, |
2e294c60 | 854 | struct btrfs_space_info *space_info) |
0d9764f6 | 855 | { |
610a6ef4 | 856 | u64 global_rsv_size = fs_info->global_block_rsv.reserved; |
2e294c60 | 857 | u64 ordered, delalloc; |
6a921de5 NA |
858 | u64 total = writable_total_bytes(fs_info, space_info); |
859 | u64 thresh; | |
2e294c60 | 860 | u64 used; |
0d9764f6 | 861 | |
6a921de5 NA |
862 | thresh = div_factor_fine(total, 90); |
863 | ||
bf7bd725 ND |
864 | lockdep_assert_held(&space_info->lock); |
865 | ||
0d9764f6 | 866 | /* If we're just plain full then async reclaim just slows us down. */ |
610a6ef4 JB |
867 | if ((space_info->bytes_used + space_info->bytes_reserved + |
868 | global_rsv_size) >= thresh) | |
ae7913ba | 869 | return false; |
0d9764f6 | 870 | |
11462397 JB |
871 | used = space_info->bytes_may_use + space_info->bytes_pinned; |
872 | ||
873 | /* The total flushable belongs to the global rsv, don't flush. */ | |
874 | if (global_rsv_size >= used) | |
875 | return false; | |
876 | ||
877 | /* | |
878 | * 128MiB is 1/4 of the maximum global rsv size. If we have less than | |
879 | * that devoted to other reservations then there's no sense in flushing, | |
880 | * we don't have a lot of things that need flushing. | |
881 | */ | |
882 | if (used - global_rsv_size <= SZ_128M) | |
883 | return false; | |
884 | ||
f205edf7 JB |
885 | /* |
886 | * We have tickets queued, bail so we don't compete with the async | |
887 | * flushers. | |
888 | */ | |
889 | if (space_info->reclaim_size) | |
890 | return false; | |
891 | ||
2e294c60 JB |
892 | /* |
893 | * If we have over half of the free space occupied by reservations or | |
894 | * pinned then we want to start flushing. | |
895 | * | |
896 | * We do not do the traditional thing here, which is to say | |
897 | * | |
898 | * if (used >= ((total_bytes + avail) / 2)) | |
899 | * return 1; | |
900 | * | |
901 | * because this doesn't quite work how we want. If we had more than 50% | |
902 | * of the space_info used by bytes_used and we had 0 available we'd just | |
903 | * constantly run the background flusher. Instead we want it to kick in | |
88a777a6 JB |
904 | * if our reclaimable space exceeds our clamped free space. |
905 | * | |
906 | * Our clamping range is 2^1 -> 2^8. Practically speaking that means | |
907 | * the following: | |
908 | * | |
909 | * Amount of RAM Minimum threshold Maximum threshold | |
910 | * | |
911 | * 256GiB 1GiB 128GiB | |
912 | * 128GiB 512MiB 64GiB | |
913 | * 64GiB 256MiB 32GiB | |
914 | * 32GiB 128MiB 16GiB | |
915 | * 16GiB 64MiB 8GiB | |
916 | * | |
917 | * These are the range our thresholds will fall in, corresponding to how | |
918 | * much delalloc we need for the background flusher to kick in. | |
2e294c60 | 919 | */ |
88a777a6 | 920 | |
2e294c60 JB |
921 | thresh = calc_available_free_space(fs_info, space_info, |
922 | BTRFS_RESERVE_FLUSH_ALL); | |
1239e2da JB |
923 | used = space_info->bytes_used + space_info->bytes_reserved + |
924 | space_info->bytes_readonly + global_rsv_size; | |
6a921de5 NA |
925 | if (used < total) |
926 | thresh += total - used; | |
88a777a6 | 927 | thresh >>= space_info->clamp; |
9f42d377 | 928 | |
2e294c60 | 929 | used = space_info->bytes_pinned; |
9f42d377 | 930 | |
2e294c60 JB |
931 | /* |
932 | * If we have more ordered bytes than delalloc bytes then we're either | |
933 | * doing a lot of DIO, or we simply don't have a lot of delalloc waiting | |
934 | * around. Preemptive flushing is only useful in that it can free up | |
935 | * space before tickets need to wait for things to finish. In the case | |
936 | * of ordered extents, preemptively waiting on ordered extents gets us | |
937 | * nothing, if our reservations are tied up in ordered extents we'll | |
938 | * simply have to slow down writers by forcing them to wait on ordered | |
939 | * extents. | |
940 | * | |
941 | * In the case that ordered is larger than delalloc, only include the | |
942 | * block reserves that we would actually be able to directly reclaim | |
943 | * from. In this case if we're heavy on metadata operations this will | |
944 | * clearly be heavy enough to warrant preemptive flushing. In the case | |
945 | * of heavy DIO or ordered reservations, preemptive flushing will just | |
946 | * waste time and cause us to slow down. | |
3e101569 JB |
947 | * |
948 | * We want to make sure we truly are maxed out on ordered however, so | |
949 | * cut ordered in half, and if it's still higher than delalloc then we | |
950 | * can keep flushing. This is to avoid the case where we start | |
951 | * flushing, and now delalloc == ordered and we stop preemptively | |
952 | * flushing when we could still have several gigs of delalloc to flush. | |
2e294c60 | 953 | */ |
3e101569 | 954 | ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1; |
2cdb3909 | 955 | delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes); |
2e294c60 JB |
956 | if (ordered >= delalloc) |
957 | used += fs_info->delayed_refs_rsv.reserved + | |
958 | fs_info->delayed_block_rsv.reserved; | |
9f42d377 | 959 | else |
30acce4e | 960 | used += space_info->bytes_may_use - global_rsv_size; |
0d9764f6 JB |
961 | |
962 | return (used >= thresh && !btrfs_fs_closing(fs_info) && | |
963 | !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state)); | |
964 | } | |
965 | ||
7f9fe614 JB |
966 | static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info, |
967 | struct btrfs_space_info *space_info, | |
968 | struct reserve_ticket *ticket) | |
969 | { | |
970 | struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv; | |
971 | u64 min_bytes; | |
972 | ||
1b0309ea JB |
973 | if (!ticket->steal) |
974 | return false; | |
975 | ||
7f9fe614 JB |
976 | if (global_rsv->space_info != space_info) |
977 | return false; | |
978 | ||
979 | spin_lock(&global_rsv->lock); | |
e6549c2a | 980 | min_bytes = div_factor(global_rsv->size, 1); |
7f9fe614 JB |
981 | if (global_rsv->reserved < min_bytes + ticket->bytes) { |
982 | spin_unlock(&global_rsv->lock); | |
983 | return false; | |
984 | } | |
985 | global_rsv->reserved -= ticket->bytes; | |
6d548b9e | 986 | remove_ticket(space_info, ticket); |
7f9fe614 | 987 | ticket->bytes = 0; |
7f9fe614 JB |
988 | wake_up(&ticket->wait); |
989 | space_info->tickets_id++; | |
990 | if (global_rsv->reserved < global_rsv->size) | |
991 | global_rsv->full = 0; | |
992 | spin_unlock(&global_rsv->lock); | |
993 | ||
994 | return true; | |
995 | } | |
996 | ||
2341ccd1 JB |
997 | /* |
998 | * maybe_fail_all_tickets - we've exhausted our flushing, start failing tickets | |
999 | * @fs_info - fs_info for this fs | |
1000 | * @space_info - the space info we were flushing | |
1001 | * | |
1002 | * We call this when we've exhausted our flushing ability and haven't made | |
1003 | * progress in satisfying tickets. The reservation code handles tickets in | |
1004 | * order, so if there is a large ticket first and then smaller ones we could | |
1005 | * very well satisfy the smaller tickets. This will attempt to wake up any | |
1006 | * tickets in the list to catch this case. | |
1007 | * | |
1008 | * This function returns true if it was able to make progress by clearing out | |
1009 | * other tickets, or if it stumbles across a ticket that was smaller than the | |
1010 | * first ticket. | |
1011 | */ | |
1012 | static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info, | |
1013 | struct btrfs_space_info *space_info) | |
0d9764f6 JB |
1014 | { |
1015 | struct reserve_ticket *ticket; | |
2341ccd1 | 1016 | u64 tickets_id = space_info->tickets_id; |
0e24f6d8 | 1017 | const bool aborted = BTRFS_FS_ERROR(fs_info); |
2341ccd1 | 1018 | |
fcdef39c JB |
1019 | trace_btrfs_fail_all_tickets(fs_info, space_info); |
1020 | ||
84fe47a4 JB |
1021 | if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) { |
1022 | btrfs_info(fs_info, "cannot satisfy tickets, dumping space info"); | |
1023 | __btrfs_dump_space_info(fs_info, space_info); | |
1024 | } | |
1025 | ||
2341ccd1 JB |
1026 | while (!list_empty(&space_info->tickets) && |
1027 | tickets_id == space_info->tickets_id) { | |
1028 | ticket = list_first_entry(&space_info->tickets, | |
1029 | struct reserve_ticket, list); | |
1030 | ||
1b0309ea | 1031 | if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket)) |
7f9fe614 JB |
1032 | return true; |
1033 | ||
0e24f6d8 | 1034 | if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG)) |
84fe47a4 JB |
1035 | btrfs_info(fs_info, "failing ticket with %llu bytes", |
1036 | ticket->bytes); | |
1037 | ||
d611add4 | 1038 | remove_ticket(space_info, ticket); |
0e24f6d8 JB |
1039 | if (aborted) |
1040 | ticket->error = -EIO; | |
1041 | else | |
1042 | ticket->error = -ENOSPC; | |
0d9764f6 | 1043 | wake_up(&ticket->wait); |
2341ccd1 JB |
1044 | |
1045 | /* | |
1046 | * We're just throwing tickets away, so more flushing may not | |
1047 | * trip over btrfs_try_granting_tickets, so we need to call it | |
1048 | * here to see if we can make progress with the next ticket in | |
1049 | * the list. | |
1050 | */ | |
0e24f6d8 JB |
1051 | if (!aborted) |
1052 | btrfs_try_granting_tickets(fs_info, space_info); | |
0d9764f6 | 1053 | } |
2341ccd1 | 1054 | return (tickets_id != space_info->tickets_id); |
0d9764f6 JB |
1055 | } |
1056 | ||
1057 | /* | |
1058 | * This is for normal flushers, we can wait all goddamned day if we want to. We | |
1059 | * will loop and continuously try to flush as long as we are making progress. | |
1060 | * We count progress as clearing off tickets each time we have to loop. | |
1061 | */ | |
1062 | static void btrfs_async_reclaim_metadata_space(struct work_struct *work) | |
1063 | { | |
1064 | struct btrfs_fs_info *fs_info; | |
1065 | struct btrfs_space_info *space_info; | |
1066 | u64 to_reclaim; | |
91e79a83 | 1067 | enum btrfs_flush_state flush_state; |
0d9764f6 JB |
1068 | int commit_cycles = 0; |
1069 | u64 last_tickets_id; | |
1070 | ||
1071 | fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work); | |
1072 | space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); | |
1073 | ||
1074 | spin_lock(&space_info->lock); | |
9f246926 | 1075 | to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); |
0d9764f6 JB |
1076 | if (!to_reclaim) { |
1077 | space_info->flush = 0; | |
1078 | spin_unlock(&space_info->lock); | |
1079 | return; | |
1080 | } | |
1081 | last_tickets_id = space_info->tickets_id; | |
1082 | spin_unlock(&space_info->lock); | |
1083 | ||
1084 | flush_state = FLUSH_DELAYED_ITEMS_NR; | |
1085 | do { | |
4b02b00f | 1086 | flush_space(fs_info, space_info, to_reclaim, flush_state, false); |
0d9764f6 JB |
1087 | spin_lock(&space_info->lock); |
1088 | if (list_empty(&space_info->tickets)) { | |
1089 | space_info->flush = 0; | |
1090 | spin_unlock(&space_info->lock); | |
1091 | return; | |
1092 | } | |
1093 | to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, | |
9f246926 | 1094 | space_info); |
0d9764f6 JB |
1095 | if (last_tickets_id == space_info->tickets_id) { |
1096 | flush_state++; | |
1097 | } else { | |
1098 | last_tickets_id = space_info->tickets_id; | |
1099 | flush_state = FLUSH_DELAYED_ITEMS_NR; | |
1100 | if (commit_cycles) | |
1101 | commit_cycles--; | |
1102 | } | |
1103 | ||
03fe78cc JB |
1104 | /* |
1105 | * We do not want to empty the system of delalloc unless we're | |
1106 | * under heavy pressure, so allow one trip through the flushing | |
1107 | * logic before we start doing a FLUSH_DELALLOC_FULL. | |
1108 | */ | |
1109 | if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles) | |
1110 | flush_state++; | |
1111 | ||
0d9764f6 JB |
1112 | /* |
1113 | * We don't want to force a chunk allocation until we've tried | |
1114 | * pretty hard to reclaim space. Think of the case where we | |
1115 | * freed up a bunch of space and so have a lot of pinned space | |
1116 | * to reclaim. We would rather use that than possibly create a | |
1117 | * underutilized metadata chunk. So if this is our first run | |
1118 | * through the flushing state machine skip ALLOC_CHUNK_FORCE and | |
1119 | * commit the transaction. If nothing has changed the next go | |
1120 | * around then we can force a chunk allocation. | |
1121 | */ | |
1122 | if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles) | |
1123 | flush_state++; | |
1124 | ||
1125 | if (flush_state > COMMIT_TRANS) { | |
1126 | commit_cycles++; | |
1127 | if (commit_cycles > 2) { | |
2341ccd1 | 1128 | if (maybe_fail_all_tickets(fs_info, space_info)) { |
0d9764f6 JB |
1129 | flush_state = FLUSH_DELAYED_ITEMS_NR; |
1130 | commit_cycles--; | |
1131 | } else { | |
1132 | space_info->flush = 0; | |
1133 | } | |
1134 | } else { | |
1135 | flush_state = FLUSH_DELAYED_ITEMS_NR; | |
1136 | } | |
1137 | } | |
1138 | spin_unlock(&space_info->lock); | |
1139 | } while (flush_state <= COMMIT_TRANS); | |
1140 | } | |
1141 | ||
576fa348 JB |
1142 | /* |
1143 | * This handles pre-flushing of metadata space before we get to the point that | |
1144 | * we need to start blocking threads on tickets. The logic here is different | |
1145 | * from the other flush paths because it doesn't rely on tickets to tell us how | |
1146 | * much we need to flush, instead it attempts to keep us below the 80% full | |
1147 | * watermark of space by flushing whichever reservation pool is currently the | |
1148 | * largest. | |
1149 | */ | |
1150 | static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work) | |
1151 | { | |
1152 | struct btrfs_fs_info *fs_info; | |
1153 | struct btrfs_space_info *space_info; | |
1154 | struct btrfs_block_rsv *delayed_block_rsv; | |
1155 | struct btrfs_block_rsv *delayed_refs_rsv; | |
1156 | struct btrfs_block_rsv *global_rsv; | |
1157 | struct btrfs_block_rsv *trans_rsv; | |
88a777a6 | 1158 | int loops = 0; |
576fa348 JB |
1159 | |
1160 | fs_info = container_of(work, struct btrfs_fs_info, | |
1161 | preempt_reclaim_work); | |
1162 | space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA); | |
1163 | delayed_block_rsv = &fs_info->delayed_block_rsv; | |
1164 | delayed_refs_rsv = &fs_info->delayed_refs_rsv; | |
1165 | global_rsv = &fs_info->global_block_rsv; | |
1166 | trans_rsv = &fs_info->trans_block_rsv; | |
1167 | ||
1168 | spin_lock(&space_info->lock); | |
2e294c60 | 1169 | while (need_preemptive_reclaim(fs_info, space_info)) { |
576fa348 JB |
1170 | enum btrfs_flush_state flush; |
1171 | u64 delalloc_size = 0; | |
1172 | u64 to_reclaim, block_rsv_size; | |
1173 | u64 global_rsv_size = global_rsv->reserved; | |
1174 | ||
88a777a6 JB |
1175 | loops++; |
1176 | ||
576fa348 JB |
1177 | /* |
1178 | * We don't have a precise counter for the metadata being | |
1179 | * reserved for delalloc, so we'll approximate it by subtracting | |
1180 | * out the block rsv's space from the bytes_may_use. If that | |
1181 | * amount is higher than the individual reserves, then we can | |
1182 | * assume it's tied up in delalloc reservations. | |
1183 | */ | |
1184 | block_rsv_size = global_rsv_size + | |
1185 | delayed_block_rsv->reserved + | |
1186 | delayed_refs_rsv->reserved + | |
1187 | trans_rsv->reserved; | |
1188 | if (block_rsv_size < space_info->bytes_may_use) | |
1189 | delalloc_size = space_info->bytes_may_use - block_rsv_size; | |
576fa348 JB |
1190 | |
1191 | /* | |
1192 | * We don't want to include the global_rsv in our calculation, | |
1193 | * because that's space we can't touch. Subtract it from the | |
1194 | * block_rsv_size for the next checks. | |
1195 | */ | |
1196 | block_rsv_size -= global_rsv_size; | |
1197 | ||
1198 | /* | |
1199 | * We really want to avoid flushing delalloc too much, as it | |
1200 | * could result in poor allocation patterns, so only flush it if | |
1201 | * it's larger than the rest of the pools combined. | |
1202 | */ | |
1203 | if (delalloc_size > block_rsv_size) { | |
1204 | to_reclaim = delalloc_size; | |
1205 | flush = FLUSH_DELALLOC; | |
1206 | } else if (space_info->bytes_pinned > | |
1207 | (delayed_block_rsv->reserved + | |
1208 | delayed_refs_rsv->reserved)) { | |
1209 | to_reclaim = space_info->bytes_pinned; | |
c416a30c | 1210 | flush = COMMIT_TRANS; |
576fa348 JB |
1211 | } else if (delayed_block_rsv->reserved > |
1212 | delayed_refs_rsv->reserved) { | |
1213 | to_reclaim = delayed_block_rsv->reserved; | |
1214 | flush = FLUSH_DELAYED_ITEMS_NR; | |
1215 | } else { | |
1216 | to_reclaim = delayed_refs_rsv->reserved; | |
1217 | flush = FLUSH_DELAYED_REFS_NR; | |
1218 | } | |
1219 | ||
06bae876 ND |
1220 | spin_unlock(&space_info->lock); |
1221 | ||
576fa348 JB |
1222 | /* |
1223 | * We don't want to reclaim everything, just a portion, so scale | |
1224 | * down the to_reclaim by 1/4. If it takes us down to 0, | |
1225 | * reclaim 1 items worth. | |
1226 | */ | |
1227 | to_reclaim >>= 2; | |
1228 | if (!to_reclaim) | |
1229 | to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1); | |
4b02b00f | 1230 | flush_space(fs_info, space_info, to_reclaim, flush, true); |
576fa348 JB |
1231 | cond_resched(); |
1232 | spin_lock(&space_info->lock); | |
576fa348 | 1233 | } |
88a777a6 JB |
1234 | |
1235 | /* We only went through once, back off our clamping. */ | |
1236 | if (loops == 1 && !space_info->reclaim_size) | |
1237 | space_info->clamp = max(1, space_info->clamp - 1); | |
e5ad49e2 | 1238 | trace_btrfs_done_preemptive_reclaim(fs_info, space_info); |
576fa348 JB |
1239 | spin_unlock(&space_info->lock); |
1240 | } | |
1241 | ||
1a7a92c8 JB |
1242 | /* |
1243 | * FLUSH_DELALLOC_WAIT: | |
1244 | * Space is freed from flushing delalloc in one of two ways. | |
1245 | * | |
1246 | * 1) compression is on and we allocate less space than we reserved | |
1247 | * 2) we are overwriting existing space | |
1248 | * | |
1249 | * For #1 that extra space is reclaimed as soon as the delalloc pages are | |
1250 | * COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent | |
1251 | * length to ->bytes_reserved, and subtracts the reserved space from | |
1252 | * ->bytes_may_use. | |
1253 | * | |
1254 | * For #2 this is trickier. Once the ordered extent runs we will drop the | |
1255 | * extent in the range we are overwriting, which creates a delayed ref for | |
1256 | * that freed extent. This however is not reclaimed until the transaction | |
1257 | * commits, thus the next stages. | |
1258 | * | |
1259 | * RUN_DELAYED_IPUTS | |
1260 | * If we are freeing inodes, we want to make sure all delayed iputs have | |
1261 | * completed, because they could have been on an inode with i_nlink == 0, and | |
1262 | * thus have been truncated and freed up space. But again this space is not | |
1263 | * immediately re-usable, it comes in the form of a delayed ref, which must be | |
1264 | * run and then the transaction must be committed. | |
1265 | * | |
1a7a92c8 | 1266 | * COMMIT_TRANS |
c416a30c JB |
1267 | * This is where we reclaim all of the pinned space generated by running the |
1268 | * iputs | |
c4923027 JB |
1269 | * |
1270 | * ALLOC_CHUNK_FORCE | |
1271 | * For data we start with alloc chunk force, however we could have been full | |
1272 | * before, and then the transaction commit could have freed new block groups, | |
1273 | * so if we now have space to allocate do the force chunk allocation. | |
1a7a92c8 | 1274 | */ |
57056740 | 1275 | static const enum btrfs_flush_state data_flush_states[] = { |
03fe78cc | 1276 | FLUSH_DELALLOC_FULL, |
57056740 | 1277 | RUN_DELAYED_IPUTS, |
57056740 | 1278 | COMMIT_TRANS, |
c4923027 | 1279 | ALLOC_CHUNK_FORCE, |
57056740 JB |
1280 | }; |
1281 | ||
1282 | static void btrfs_async_reclaim_data_space(struct work_struct *work) | |
0d9764f6 | 1283 | { |
57056740 JB |
1284 | struct btrfs_fs_info *fs_info; |
1285 | struct btrfs_space_info *space_info; | |
1286 | u64 last_tickets_id; | |
91e79a83 | 1287 | enum btrfs_flush_state flush_state = 0; |
57056740 JB |
1288 | |
1289 | fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work); | |
1290 | space_info = fs_info->data_sinfo; | |
1291 | ||
1292 | spin_lock(&space_info->lock); | |
1293 | if (list_empty(&space_info->tickets)) { | |
1294 | space_info->flush = 0; | |
1295 | spin_unlock(&space_info->lock); | |
1296 | return; | |
1297 | } | |
1298 | last_tickets_id = space_info->tickets_id; | |
1299 | spin_unlock(&space_info->lock); | |
1300 | ||
1301 | while (!space_info->full) { | |
4b02b00f | 1302 | flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); |
57056740 JB |
1303 | spin_lock(&space_info->lock); |
1304 | if (list_empty(&space_info->tickets)) { | |
1305 | space_info->flush = 0; | |
1306 | spin_unlock(&space_info->lock); | |
1307 | return; | |
1308 | } | |
0e24f6d8 JB |
1309 | |
1310 | /* Something happened, fail everything and bail. */ | |
1311 | if (BTRFS_FS_ERROR(fs_info)) | |
1312 | goto aborted_fs; | |
57056740 JB |
1313 | last_tickets_id = space_info->tickets_id; |
1314 | spin_unlock(&space_info->lock); | |
1315 | } | |
1316 | ||
1317 | while (flush_state < ARRAY_SIZE(data_flush_states)) { | |
1318 | flush_space(fs_info, space_info, U64_MAX, | |
4b02b00f | 1319 | data_flush_states[flush_state], false); |
57056740 JB |
1320 | spin_lock(&space_info->lock); |
1321 | if (list_empty(&space_info->tickets)) { | |
1322 | space_info->flush = 0; | |
1323 | spin_unlock(&space_info->lock); | |
1324 | return; | |
1325 | } | |
1326 | ||
1327 | if (last_tickets_id == space_info->tickets_id) { | |
1328 | flush_state++; | |
1329 | } else { | |
1330 | last_tickets_id = space_info->tickets_id; | |
1331 | flush_state = 0; | |
1332 | } | |
1333 | ||
1334 | if (flush_state >= ARRAY_SIZE(data_flush_states)) { | |
1335 | if (space_info->full) { | |
1336 | if (maybe_fail_all_tickets(fs_info, space_info)) | |
1337 | flush_state = 0; | |
1338 | else | |
1339 | space_info->flush = 0; | |
1340 | } else { | |
1341 | flush_state = 0; | |
1342 | } | |
0e24f6d8 JB |
1343 | |
1344 | /* Something happened, fail everything and bail. */ | |
1345 | if (BTRFS_FS_ERROR(fs_info)) | |
1346 | goto aborted_fs; | |
1347 | ||
57056740 JB |
1348 | } |
1349 | spin_unlock(&space_info->lock); | |
1350 | } | |
0e24f6d8 JB |
1351 | return; |
1352 | ||
1353 | aborted_fs: | |
1354 | maybe_fail_all_tickets(fs_info, space_info); | |
1355 | space_info->flush = 0; | |
1356 | spin_unlock(&space_info->lock); | |
57056740 JB |
1357 | } |
1358 | ||
1359 | void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info) | |
1360 | { | |
1361 | INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space); | |
1362 | INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space); | |
576fa348 JB |
1363 | INIT_WORK(&fs_info->preempt_reclaim_work, |
1364 | btrfs_preempt_reclaim_metadata_space); | |
0d9764f6 JB |
1365 | } |
1366 | ||
1367 | static const enum btrfs_flush_state priority_flush_states[] = { | |
1368 | FLUSH_DELAYED_ITEMS_NR, | |
1369 | FLUSH_DELAYED_ITEMS, | |
1370 | ALLOC_CHUNK, | |
1371 | }; | |
1372 | ||
d3984c90 JB |
1373 | static const enum btrfs_flush_state evict_flush_states[] = { |
1374 | FLUSH_DELAYED_ITEMS_NR, | |
1375 | FLUSH_DELAYED_ITEMS, | |
1376 | FLUSH_DELAYED_REFS_NR, | |
1377 | FLUSH_DELAYED_REFS, | |
1378 | FLUSH_DELALLOC, | |
1379 | FLUSH_DELALLOC_WAIT, | |
03fe78cc | 1380 | FLUSH_DELALLOC_FULL, |
d3984c90 JB |
1381 | ALLOC_CHUNK, |
1382 | COMMIT_TRANS, | |
1383 | }; | |
1384 | ||
0d9764f6 | 1385 | static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info, |
9ce2f423 JB |
1386 | struct btrfs_space_info *space_info, |
1387 | struct reserve_ticket *ticket, | |
1388 | const enum btrfs_flush_state *states, | |
1389 | int states_nr) | |
0d9764f6 JB |
1390 | { |
1391 | u64 to_reclaim; | |
9f35f76d | 1392 | int flush_state = 0; |
0d9764f6 JB |
1393 | |
1394 | spin_lock(&space_info->lock); | |
9f246926 | 1395 | to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info); |
9cd8dcdc JB |
1396 | /* |
1397 | * This is the priority reclaim path, so to_reclaim could be >0 still | |
143823cf | 1398 | * because we may have only satisfied the priority tickets and still |
9cd8dcdc JB |
1399 | * left non priority tickets on the list. We would then have |
1400 | * to_reclaim but ->bytes == 0. | |
1401 | */ | |
1402 | if (ticket->bytes == 0) { | |
0d9764f6 JB |
1403 | spin_unlock(&space_info->lock); |
1404 | return; | |
1405 | } | |
0d9764f6 | 1406 | |
9f35f76d JB |
1407 | while (flush_state < states_nr) { |
1408 | spin_unlock(&space_info->lock); | |
4b02b00f JB |
1409 | flush_space(fs_info, space_info, to_reclaim, states[flush_state], |
1410 | false); | |
0d9764f6 JB |
1411 | flush_state++; |
1412 | spin_lock(&space_info->lock); | |
1413 | if (ticket->bytes == 0) { | |
1414 | spin_unlock(&space_info->lock); | |
1415 | return; | |
1416 | } | |
9f35f76d JB |
1417 | } |
1418 | ||
ee6adbfd JB |
1419 | /* Attempt to steal from the global rsv if we can. */ |
1420 | if (!steal_from_global_rsv(fs_info, space_info, ticket)) { | |
1421 | ticket->error = -ENOSPC; | |
1422 | remove_ticket(space_info, ticket); | |
1423 | } | |
1424 | ||
9f35f76d JB |
1425 | /* |
1426 | * We must run try_granting_tickets here because we could be a large | |
1427 | * ticket in front of a smaller ticket that can now be satisfied with | |
1428 | * the available space. | |
1429 | */ | |
9f35f76d JB |
1430 | btrfs_try_granting_tickets(fs_info, space_info); |
1431 | spin_unlock(&space_info->lock); | |
0d9764f6 JB |
1432 | } |
1433 | ||
1004f686 JB |
1434 | static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info, |
1435 | struct btrfs_space_info *space_info, | |
57056740 | 1436 | struct reserve_ticket *ticket) |
1004f686 | 1437 | { |
9f35f76d | 1438 | spin_lock(&space_info->lock); |
9cd8dcdc JB |
1439 | |
1440 | /* We could have been granted before we got here. */ | |
1441 | if (ticket->bytes == 0) { | |
1442 | spin_unlock(&space_info->lock); | |
1443 | return; | |
1444 | } | |
1445 | ||
1004f686 | 1446 | while (!space_info->full) { |
9f35f76d | 1447 | spin_unlock(&space_info->lock); |
4b02b00f | 1448 | flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false); |
1004f686 JB |
1449 | spin_lock(&space_info->lock); |
1450 | if (ticket->bytes == 0) { | |
1451 | spin_unlock(&space_info->lock); | |
1452 | return; | |
1453 | } | |
1004f686 | 1454 | } |
9f35f76d JB |
1455 | |
1456 | ticket->error = -ENOSPC; | |
1457 | remove_ticket(space_info, ticket); | |
1458 | btrfs_try_granting_tickets(fs_info, space_info); | |
1459 | spin_unlock(&space_info->lock); | |
1004f686 JB |
1460 | } |
1461 | ||
374bf9c5 JB |
1462 | static void wait_reserve_ticket(struct btrfs_fs_info *fs_info, |
1463 | struct btrfs_space_info *space_info, | |
1464 | struct reserve_ticket *ticket) | |
0d9764f6 JB |
1465 | |
1466 | { | |
1467 | DEFINE_WAIT(wait); | |
0d9764f6 JB |
1468 | int ret = 0; |
1469 | ||
1470 | spin_lock(&space_info->lock); | |
1471 | while (ticket->bytes > 0 && ticket->error == 0) { | |
1472 | ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE); | |
1473 | if (ret) { | |
0cab7acc FM |
1474 | /* |
1475 | * Delete us from the list. After we unlock the space | |
1476 | * info, we don't want the async reclaim job to reserve | |
1477 | * space for this ticket. If that would happen, then the | |
1478 | * ticket's task would not known that space was reserved | |
1479 | * despite getting an error, resulting in a space leak | |
1480 | * (bytes_may_use counter of our space_info). | |
1481 | */ | |
d611add4 | 1482 | remove_ticket(space_info, ticket); |
374bf9c5 | 1483 | ticket->error = -EINTR; |
0d9764f6 JB |
1484 | break; |
1485 | } | |
1486 | spin_unlock(&space_info->lock); | |
1487 | ||
1488 | schedule(); | |
1489 | ||
1490 | finish_wait(&ticket->wait, &wait); | |
1491 | spin_lock(&space_info->lock); | |
1492 | } | |
0d9764f6 | 1493 | spin_unlock(&space_info->lock); |
0d9764f6 JB |
1494 | } |
1495 | ||
03235279 | 1496 | /** |
d98b188e NB |
1497 | * Do the appropriate flushing and waiting for a ticket |
1498 | * | |
1499 | * @fs_info: the filesystem | |
1500 | * @space_info: space info for the reservation | |
1501 | * @ticket: ticket for the reservation | |
ac1ea10e JB |
1502 | * @start_ns: timestamp when the reservation started |
1503 | * @orig_bytes: amount of bytes originally reserved | |
d98b188e | 1504 | * @flush: how much we can flush |
03235279 JB |
1505 | * |
1506 | * This does the work of figuring out how to flush for the ticket, waiting for | |
1507 | * the reservation, and returning the appropriate error if there is one. | |
1508 | */ | |
1509 | static int handle_reserve_ticket(struct btrfs_fs_info *fs_info, | |
1510 | struct btrfs_space_info *space_info, | |
1511 | struct reserve_ticket *ticket, | |
ac1ea10e | 1512 | u64 start_ns, u64 orig_bytes, |
03235279 JB |
1513 | enum btrfs_reserve_flush_enum flush) |
1514 | { | |
03235279 JB |
1515 | int ret; |
1516 | ||
d3984c90 | 1517 | switch (flush) { |
57056740 | 1518 | case BTRFS_RESERVE_FLUSH_DATA: |
d3984c90 | 1519 | case BTRFS_RESERVE_FLUSH_ALL: |
7f9fe614 | 1520 | case BTRFS_RESERVE_FLUSH_ALL_STEAL: |
03235279 | 1521 | wait_reserve_ticket(fs_info, space_info, ticket); |
d3984c90 JB |
1522 | break; |
1523 | case BTRFS_RESERVE_FLUSH_LIMIT: | |
9ce2f423 JB |
1524 | priority_reclaim_metadata_space(fs_info, space_info, ticket, |
1525 | priority_flush_states, | |
1526 | ARRAY_SIZE(priority_flush_states)); | |
d3984c90 JB |
1527 | break; |
1528 | case BTRFS_RESERVE_FLUSH_EVICT: | |
1529 | priority_reclaim_metadata_space(fs_info, space_info, ticket, | |
1530 | evict_flush_states, | |
1531 | ARRAY_SIZE(evict_flush_states)); | |
1532 | break; | |
1004f686 | 1533 | case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE: |
57056740 | 1534 | priority_reclaim_data_space(fs_info, space_info, ticket); |
1004f686 | 1535 | break; |
d3984c90 JB |
1536 | default: |
1537 | ASSERT(0); | |
1538 | break; | |
1539 | } | |
03235279 | 1540 | |
03235279 | 1541 | ret = ticket->error; |
03235279 | 1542 | ASSERT(list_empty(&ticket->list)); |
0cab7acc FM |
1543 | /* |
1544 | * Check that we can't have an error set if the reservation succeeded, | |
1545 | * as that would confuse tasks and lead them to error out without | |
1546 | * releasing reserved space (if an error happens the expectation is that | |
1547 | * space wasn't reserved at all). | |
1548 | */ | |
1549 | ASSERT(!(ticket->bytes == 0 && ticket->error)); | |
ac1ea10e JB |
1550 | trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes, |
1551 | start_ns, flush, ticket->error); | |
03235279 JB |
1552 | return ret; |
1553 | } | |
1554 | ||
666daa9f JB |
1555 | /* |
1556 | * This returns true if this flush state will go through the ordinary flushing | |
1557 | * code. | |
1558 | */ | |
1559 | static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush) | |
1560 | { | |
1561 | return (flush == BTRFS_RESERVE_FLUSH_ALL) || | |
1562 | (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL); | |
1563 | } | |
1564 | ||
88a777a6 JB |
1565 | static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info, |
1566 | struct btrfs_space_info *space_info) | |
1567 | { | |
1568 | u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes); | |
1569 | u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes); | |
1570 | ||
1571 | /* | |
1572 | * If we're heavy on ordered operations then clamping won't help us. We | |
1573 | * need to clamp specifically to keep up with dirty'ing buffered | |
1574 | * writers, because there's not a 1:1 correlation of writing delalloc | |
1575 | * and freeing space, like there is with flushing delayed refs or | |
1576 | * delayed nodes. If we're already more ordered than delalloc then | |
1577 | * we're keeping up, otherwise we aren't and should probably clamp. | |
1578 | */ | |
1579 | if (ordered < delalloc) | |
1580 | space_info->clamp = min(space_info->clamp + 1, 8); | |
1581 | } | |
1582 | ||
ee6adbfd JB |
1583 | static inline bool can_steal(enum btrfs_reserve_flush_enum flush) |
1584 | { | |
1585 | return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || | |
1586 | flush == BTRFS_RESERVE_FLUSH_EVICT); | |
1587 | } | |
1588 | ||
765c3fe9 JB |
1589 | /* |
1590 | * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to | |
1591 | * fail as quickly as possible. | |
1592 | */ | |
1593 | static inline bool can_ticket(enum btrfs_reserve_flush_enum flush) | |
1594 | { | |
1595 | return (flush != BTRFS_RESERVE_NO_FLUSH && | |
1596 | flush != BTRFS_RESERVE_FLUSH_EMERGENCY); | |
1597 | } | |
1598 | ||
0d9764f6 | 1599 | /** |
d98b188e NB |
1600 | * Try to reserve bytes from the block_rsv's space |
1601 | * | |
1602 | * @fs_info: the filesystem | |
1603 | * @space_info: space info we want to allocate from | |
1604 | * @orig_bytes: number of bytes we want | |
1605 | * @flush: whether or not we can flush to make our reservation | |
0d9764f6 JB |
1606 | * |
1607 | * This will reserve orig_bytes number of bytes from the space info associated | |
1608 | * with the block_rsv. If there is not enough space it will make an attempt to | |
1609 | * flush out space to make room. It will do this by flushing delalloc if | |
1610 | * possible or committing the transaction. If flush is 0 then no attempts to | |
1611 | * regain reservations will be made and this will fail if there is not enough | |
1612 | * space already. | |
1613 | */ | |
f3bda421 JB |
1614 | static int __reserve_bytes(struct btrfs_fs_info *fs_info, |
1615 | struct btrfs_space_info *space_info, u64 orig_bytes, | |
1616 | enum btrfs_reserve_flush_enum flush) | |
0d9764f6 | 1617 | { |
57056740 | 1618 | struct work_struct *async_work; |
0d9764f6 | 1619 | struct reserve_ticket ticket; |
ac1ea10e | 1620 | u64 start_ns = 0; |
0d9764f6 | 1621 | u64 used; |
0d9764f6 | 1622 | int ret = 0; |
ef1317a1 | 1623 | bool pending_tickets; |
0d9764f6 JB |
1624 | |
1625 | ASSERT(orig_bytes); | |
1626 | ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_ALL); | |
1627 | ||
57056740 JB |
1628 | if (flush == BTRFS_RESERVE_FLUSH_DATA) |
1629 | async_work = &fs_info->async_data_reclaim_work; | |
1630 | else | |
1631 | async_work = &fs_info->async_reclaim_work; | |
1632 | ||
0d9764f6 JB |
1633 | spin_lock(&space_info->lock); |
1634 | ret = -ENOSPC; | |
1635 | used = btrfs_space_info_used(space_info, true); | |
666daa9f JB |
1636 | |
1637 | /* | |
1638 | * We don't want NO_FLUSH allocations to jump everybody, they can | |
1639 | * generally handle ENOSPC in a different way, so treat them the same as | |
1640 | * normal flushers when it comes to skipping pending tickets. | |
1641 | */ | |
1642 | if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH)) | |
1643 | pending_tickets = !list_empty(&space_info->tickets) || | |
1644 | !list_empty(&space_info->priority_tickets); | |
1645 | else | |
1646 | pending_tickets = !list_empty(&space_info->priority_tickets); | |
0d9764f6 JB |
1647 | |
1648 | /* | |
9b4851bc GR |
1649 | * Carry on if we have enough space (short-circuit) OR call |
1650 | * can_overcommit() to ensure we can overcommit to continue. | |
0d9764f6 | 1651 | */ |
ef1317a1 | 1652 | if (!pending_tickets && |
6a921de5 | 1653 | ((used + orig_bytes <= writable_total_bytes(fs_info, space_info)) || |
a30a3d20 | 1654 | btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) { |
0d9764f6 JB |
1655 | btrfs_space_info_update_bytes_may_use(fs_info, space_info, |
1656 | orig_bytes); | |
0d9764f6 JB |
1657 | ret = 0; |
1658 | } | |
1659 | ||
765c3fe9 JB |
1660 | /* |
1661 | * Things are dire, we need to make a reservation so we don't abort. We | |
1662 | * will let this reservation go through as long as we have actual space | |
1663 | * left to allocate for the block. | |
1664 | */ | |
1665 | if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) { | |
1666 | used = btrfs_space_info_used(space_info, false); | |
1667 | if (used + orig_bytes <= | |
1668 | writable_total_bytes(fs_info, space_info)) { | |
1669 | btrfs_space_info_update_bytes_may_use(fs_info, space_info, | |
1670 | orig_bytes); | |
1671 | ret = 0; | |
1672 | } | |
1673 | } | |
1674 | ||
0d9764f6 JB |
1675 | /* |
1676 | * If we couldn't make a reservation then setup our reservation ticket | |
1677 | * and kick the async worker if it's not already running. | |
1678 | * | |
1679 | * If we are a priority flusher then we just need to add our ticket to | |
1680 | * the list and we will do our own flushing further down. | |
1681 | */ | |
765c3fe9 | 1682 | if (ret && can_ticket(flush)) { |
0d9764f6 JB |
1683 | ticket.bytes = orig_bytes; |
1684 | ticket.error = 0; | |
db161806 | 1685 | space_info->reclaim_size += ticket.bytes; |
0d9764f6 | 1686 | init_waitqueue_head(&ticket.wait); |
ee6adbfd | 1687 | ticket.steal = can_steal(flush); |
ac1ea10e JB |
1688 | if (trace_btrfs_reserve_ticket_enabled()) |
1689 | start_ns = ktime_get_ns(); | |
1690 | ||
7f9fe614 | 1691 | if (flush == BTRFS_RESERVE_FLUSH_ALL || |
57056740 JB |
1692 | flush == BTRFS_RESERVE_FLUSH_ALL_STEAL || |
1693 | flush == BTRFS_RESERVE_FLUSH_DATA) { | |
0d9764f6 JB |
1694 | list_add_tail(&ticket.list, &space_info->tickets); |
1695 | if (!space_info->flush) { | |
0aae4ca9 JB |
1696 | /* |
1697 | * We were forced to add a reserve ticket, so | |
1698 | * our preemptive flushing is unable to keep | |
1699 | * up. Clamp down on the threshold for the | |
1700 | * preemptive flushing in order to keep up with | |
1701 | * the workload. | |
1702 | */ | |
1703 | maybe_clamp_preempt(fs_info, space_info); | |
1704 | ||
0d9764f6 JB |
1705 | space_info->flush = 1; |
1706 | trace_btrfs_trigger_flush(fs_info, | |
1707 | space_info->flags, | |
1708 | orig_bytes, flush, | |
1709 | "enospc"); | |
57056740 | 1710 | queue_work(system_unbound_wq, async_work); |
0d9764f6 JB |
1711 | } |
1712 | } else { | |
1713 | list_add_tail(&ticket.list, | |
1714 | &space_info->priority_tickets); | |
1715 | } | |
1716 | } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) { | |
0d9764f6 JB |
1717 | /* |
1718 | * We will do the space reservation dance during log replay, | |
1719 | * which means we won't have fs_info->fs_root set, so don't do | |
1720 | * the async reclaim as we will panic. | |
1721 | */ | |
1722 | if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) && | |
ed738ba7 JB |
1723 | !work_busy(&fs_info->preempt_reclaim_work) && |
1724 | need_preemptive_reclaim(fs_info, space_info)) { | |
0d9764f6 JB |
1725 | trace_btrfs_trigger_flush(fs_info, space_info->flags, |
1726 | orig_bytes, flush, "preempt"); | |
1727 | queue_work(system_unbound_wq, | |
576fa348 | 1728 | &fs_info->preempt_reclaim_work); |
0d9764f6 JB |
1729 | } |
1730 | } | |
1731 | spin_unlock(&space_info->lock); | |
765c3fe9 | 1732 | if (!ret || !can_ticket(flush)) |
0d9764f6 JB |
1733 | return ret; |
1734 | ||
ac1ea10e JB |
1735 | return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns, |
1736 | orig_bytes, flush); | |
0d9764f6 JB |
1737 | } |
1738 | ||
1739 | /** | |
d98b188e NB |
1740 | * Trye to reserve metadata bytes from the block_rsv's space |
1741 | * | |
be8d1a2a | 1742 | * @fs_info: the filesystem |
d98b188e NB |
1743 | * @block_rsv: block_rsv we're allocating for |
1744 | * @orig_bytes: number of bytes we want | |
1745 | * @flush: whether or not we can flush to make our reservation | |
0d9764f6 JB |
1746 | * |
1747 | * This will reserve orig_bytes number of bytes from the space info associated | |
1748 | * with the block_rsv. If there is not enough space it will make an attempt to | |
1749 | * flush out space to make room. It will do this by flushing delalloc if | |
1750 | * possible or committing the transaction. If flush is 0 then no attempts to | |
1751 | * regain reservations will be made and this will fail if there is not enough | |
1752 | * space already. | |
1753 | */ | |
9270501c | 1754 | int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info, |
0d9764f6 JB |
1755 | struct btrfs_block_rsv *block_rsv, |
1756 | u64 orig_bytes, | |
1757 | enum btrfs_reserve_flush_enum flush) | |
1758 | { | |
0d9764f6 | 1759 | int ret; |
0d9764f6 | 1760 | |
f3bda421 | 1761 | ret = __reserve_bytes(fs_info, block_rsv->space_info, orig_bytes, flush); |
0d9764f6 JB |
1762 | if (ret == -ENOSPC) { |
1763 | trace_btrfs_space_reservation(fs_info, "space_info:enospc", | |
1764 | block_rsv->space_info->flags, | |
1765 | orig_bytes, 1); | |
1766 | ||
1767 | if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) | |
1768 | btrfs_dump_space_info(fs_info, block_rsv->space_info, | |
1769 | orig_bytes, 0); | |
1770 | } | |
1771 | return ret; | |
1772 | } | |
8698fc4e JB |
1773 | |
1774 | /** | |
d98b188e NB |
1775 | * Try to reserve data bytes for an allocation |
1776 | * | |
1777 | * @fs_info: the filesystem | |
1778 | * @bytes: number of bytes we need | |
1779 | * @flush: how we are allowed to flush | |
8698fc4e JB |
1780 | * |
1781 | * This will reserve bytes from the data space info. If there is not enough | |
1782 | * space then we will attempt to flush space as specified by flush. | |
1783 | */ | |
1784 | int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes, | |
1785 | enum btrfs_reserve_flush_enum flush) | |
1786 | { | |
1787 | struct btrfs_space_info *data_sinfo = fs_info->data_sinfo; | |
f3bda421 | 1788 | int ret; |
8698fc4e | 1789 | |
f3bda421 | 1790 | ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA || |
1daedb1d JB |
1791 | flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE || |
1792 | flush == BTRFS_RESERVE_NO_FLUSH); | |
8698fc4e JB |
1793 | ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA); |
1794 | ||
f3bda421 JB |
1795 | ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush); |
1796 | if (ret == -ENOSPC) { | |
1797 | trace_btrfs_space_reservation(fs_info, "space_info:enospc", | |
8698fc4e | 1798 | data_sinfo->flags, bytes, 1); |
f3bda421 JB |
1799 | if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) |
1800 | btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0); | |
1801 | } | |
8698fc4e JB |
1802 | return ret; |
1803 | } | |
8e327b9c QW |
1804 | |
1805 | /* Dump all the space infos when we abort a transaction due to ENOSPC. */ | |
1806 | __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info) | |
1807 | { | |
1808 | struct btrfs_space_info *space_info; | |
1809 | ||
1810 | btrfs_info(fs_info, "dumping space info:"); | |
1811 | list_for_each_entry(space_info, &fs_info->space_info, list) { | |
1812 | spin_lock(&space_info->lock); | |
1813 | __btrfs_dump_space_info(fs_info, space_info); | |
1814 | spin_unlock(&space_info->lock); | |
1815 | } | |
1816 | dump_global_block_rsv(fs_info); | |
1817 | } | |
e2f13b34 JB |
1818 | |
1819 | /* | |
1820 | * Account the unused space of all the readonly block group in the space_info. | |
1821 | * takes mirrors into account. | |
1822 | */ | |
1823 | u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo) | |
1824 | { | |
1825 | struct btrfs_block_group *block_group; | |
1826 | u64 free_bytes = 0; | |
1827 | int factor; | |
1828 | ||
1829 | /* It's df, we don't care if it's racy */ | |
1830 | if (list_empty(&sinfo->ro_bgs)) | |
1831 | return 0; | |
1832 | ||
1833 | spin_lock(&sinfo->lock); | |
1834 | list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) { | |
1835 | spin_lock(&block_group->lock); | |
1836 | ||
1837 | if (!block_group->ro) { | |
1838 | spin_unlock(&block_group->lock); | |
1839 | continue; | |
1840 | } | |
1841 | ||
1842 | factor = btrfs_bg_type_to_factor(block_group->flags); | |
1843 | free_bytes += (block_group->length - | |
1844 | block_group->used) * factor; | |
1845 | ||
1846 | spin_unlock(&block_group->lock); | |
1847 | } | |
1848 | spin_unlock(&sinfo->lock); | |
1849 | ||
1850 | return free_bytes; | |
1851 | } |