xref: /linux/fs/btrfs/space-info.c (revision ca283ea9)
1 // SPDX-License-Identifier: GPL-2.0
2 
3 #include "linux/spinlock.h"
4 #include <linux/minmax.h>
5 #include "misc.h"
6 #include "ctree.h"
7 #include "space-info.h"
8 #include "sysfs.h"
9 #include "volumes.h"
10 #include "free-space-cache.h"
11 #include "ordered-data.h"
12 #include "transaction.h"
13 #include "block-group.h"
14 #include "fs.h"
15 #include "accessors.h"
16 #include "extent-tree.h"
17 
18 /*
19  * HOW DOES SPACE RESERVATION WORK
20  *
21  * If you want to know about delalloc specifically, there is a separate comment
22  * for that with the delalloc code.  This comment is about how the whole system
23  * works generally.
24  *
25  * BASIC CONCEPTS
26  *
27  *   1) space_info.  This is the ultimate arbiter of how much space we can use.
28  *   There's a description of the bytes_ fields with the struct declaration,
29  *   refer to that for specifics on each field.  Suffice it to say that for
30  *   reservations we care about total_bytes - SUM(space_info->bytes_) when
31  *   determining if there is space to make an allocation.  There is a space_info
32  *   for METADATA, SYSTEM, and DATA areas.
33  *
34  *   2) block_rsv's.  These are basically buckets for every different type of
35  *   metadata reservation we have.  You can see the comment in the block_rsv
36  *   code on the rules for each type, but generally block_rsv->reserved is how
37  *   much space is accounted for in space_info->bytes_may_use.
38  *
39  *   3) btrfs_calc*_size.  These are the worst case calculations we used based
40  *   on the number of items we will want to modify.  We have one for changing
41  *   items, and one for inserting new items.  Generally we use these helpers to
42  *   determine the size of the block reserves, and then use the actual bytes
43  *   values to adjust the space_info counters.
44  *
45  * MAKING RESERVATIONS, THE NORMAL CASE
46  *
47  *   We call into either btrfs_reserve_data_bytes() or
48  *   btrfs_reserve_metadata_bytes(), depending on which we're looking for, with
49  *   num_bytes we want to reserve.
50  *
51  *   ->reserve
52  *     space_info->bytes_may_reserve += num_bytes
53  *
54  *   ->extent allocation
55  *     Call btrfs_add_reserved_bytes() which does
56  *     space_info->bytes_may_reserve -= num_bytes
57  *     space_info->bytes_reserved += extent_bytes
58  *
59  *   ->insert reference
60  *     Call btrfs_update_block_group() which does
61  *     space_info->bytes_reserved -= extent_bytes
62  *     space_info->bytes_used += extent_bytes
63  *
64  * MAKING RESERVATIONS, FLUSHING NORMALLY (non-priority)
65  *
66  *   Assume we are unable to simply make the reservation because we do not have
67  *   enough space
68  *
69  *   -> __reserve_bytes
70  *     create a reserve_ticket with ->bytes set to our reservation, add it to
71  *     the tail of space_info->tickets, kick async flush thread
72  *
73  *   ->handle_reserve_ticket
74  *     wait on ticket->wait for ->bytes to be reduced to 0, or ->error to be set
75  *     on the ticket.
76  *
77  *   -> btrfs_async_reclaim_metadata_space/btrfs_async_reclaim_data_space
78  *     Flushes various things attempting to free up space.
79  *
80  *   -> btrfs_try_granting_tickets()
81  *     This is called by anything that either subtracts space from
82  *     space_info->bytes_may_use, ->bytes_pinned, etc, or adds to the
83  *     space_info->total_bytes.  This loops through the ->priority_tickets and
84  *     then the ->tickets list checking to see if the reservation can be
85  *     completed.  If it can the space is added to space_info->bytes_may_use and
86  *     the ticket is woken up.
87  *
88  *   -> ticket wakeup
89  *     Check if ->bytes == 0, if it does we got our reservation and we can carry
90  *     on, if not return the appropriate error (ENOSPC, but can be EINTR if we
91  *     were interrupted.)
92  *
93  * MAKING RESERVATIONS, FLUSHING HIGH PRIORITY
94  *
95  *   Same as the above, except we add ourselves to the
96  *   space_info->priority_tickets, and we do not use ticket->wait, we simply
97  *   call flush_space() ourselves for the states that are safe for us to call
98  *   without deadlocking and hope for the best.
99  *
100  * THE FLUSHING STATES
101  *
102  *   Generally speaking we will have two cases for each state, a "nice" state
103  *   and a "ALL THE THINGS" state.  In btrfs we delay a lot of work in order to
104  *   reduce the locking over head on the various trees, and even to keep from
105  *   doing any work at all in the case of delayed refs.  Each of these delayed
106  *   things however hold reservations, and so letting them run allows us to
107  *   reclaim space so we can make new reservations.
108  *
109  *   FLUSH_DELAYED_ITEMS
110  *     Every inode has a delayed item to update the inode.  Take a simple write
111  *     for example, we would update the inode item at write time to update the
112  *     mtime, and then again at finish_ordered_io() time in order to update the
113  *     isize or bytes.  We keep these delayed items to coalesce these operations
114  *     into a single operation done on demand.  These are an easy way to reclaim
115  *     metadata space.
116  *
117  *   FLUSH_DELALLOC
118  *     Look at the delalloc comment to get an idea of how much space is reserved
119  *     for delayed allocation.  We can reclaim some of this space simply by
120  *     running delalloc, but usually we need to wait for ordered extents to
121  *     reclaim the bulk of this space.
122  *
123  *   FLUSH_DELAYED_REFS
124  *     We have a block reserve for the outstanding delayed refs space, and every
125  *     delayed ref operation holds a reservation.  Running these is a quick way
126  *     to reclaim space, but we want to hold this until the end because COW can
127  *     churn a lot and we can avoid making some extent tree modifications if we
128  *     are able to delay for as long as possible.
129  *
130  *   ALLOC_CHUNK
131  *     We will skip this the first time through space reservation, because of
132  *     overcommit and we don't want to have a lot of useless metadata space when
133  *     our worst case reservations will likely never come true.
134  *
135  *   RUN_DELAYED_IPUTS
136  *     If we're freeing inodes we're likely freeing checksums, file extent
137  *     items, and extent tree items.  Loads of space could be freed up by these
138  *     operations, however they won't be usable until the transaction commits.
139  *
140  *   COMMIT_TRANS
141  *     This will commit the transaction.  Historically we had a lot of logic
142  *     surrounding whether or not we'd commit the transaction, but this waits born
143  *     out of a pre-tickets era where we could end up committing the transaction
144  *     thousands of times in a row without making progress.  Now thanks to our
145  *     ticketing system we know if we're not making progress and can error
146  *     everybody out after a few commits rather than burning the disk hoping for
147  *     a different answer.
148  *
149  * OVERCOMMIT
150  *
151  *   Because we hold so many reservations for metadata we will allow you to
152  *   reserve more space than is currently free in the currently allocate
153  *   metadata space.  This only happens with metadata, data does not allow
154  *   overcommitting.
155  *
156  *   You can see the current logic for when we allow overcommit in
157  *   btrfs_can_overcommit(), but it only applies to unallocated space.  If there
158  *   is no unallocated space to be had, all reservations are kept within the
159  *   free space in the allocated metadata chunks.
160  *
161  *   Because of overcommitting, you generally want to use the
162  *   btrfs_can_overcommit() logic for metadata allocations, as it does the right
163  *   thing with or without extra unallocated space.
164  */
165 
btrfs_space_info_used(const struct btrfs_space_info * s_info,bool may_use_included)166 u64 __pure btrfs_space_info_used(const struct btrfs_space_info *s_info,
167 			  bool may_use_included)
168 {
169 	ASSERT(s_info);
170 	return s_info->bytes_used + s_info->bytes_reserved +
171 		s_info->bytes_pinned + s_info->bytes_readonly +
172 		s_info->bytes_zone_unusable +
173 		(may_use_included ? s_info->bytes_may_use : 0);
174 }
175 
176 /*
177  * after adding space to the filesystem, we need to clear the full flags
178  * on all the space infos.
179  */
btrfs_clear_space_info_full(struct btrfs_fs_info * info)180 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
181 {
182 	struct list_head *head = &info->space_info;
183 	struct btrfs_space_info *found;
184 
185 	list_for_each_entry(found, head, list)
186 		found->full = 0;
187 }
188 
189 /*
190  * Block groups with more than this value (percents) of unusable space will be
191  * scheduled for background reclaim.
192  */
193 #define BTRFS_DEFAULT_ZONED_RECLAIM_THRESH			(75)
194 
195 #define BTRFS_UNALLOC_BLOCK_GROUP_TARGET			(10ULL)
196 
197 /*
198  * Calculate chunk size depending on volume type (regular or zoned).
199  */
calc_chunk_size(const struct btrfs_fs_info * fs_info,u64 flags)200 static u64 calc_chunk_size(const struct btrfs_fs_info *fs_info, u64 flags)
201 {
202 	if (btrfs_is_zoned(fs_info))
203 		return fs_info->zone_size;
204 
205 	ASSERT(flags & BTRFS_BLOCK_GROUP_TYPE_MASK);
206 
207 	if (flags & BTRFS_BLOCK_GROUP_DATA)
208 		return BTRFS_MAX_DATA_CHUNK_SIZE;
209 	else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
210 		return SZ_32M;
211 
212 	/* Handle BTRFS_BLOCK_GROUP_METADATA */
213 	if (fs_info->fs_devices->total_rw_bytes > 50ULL * SZ_1G)
214 		return SZ_1G;
215 
216 	return SZ_256M;
217 }
218 
219 /*
220  * Update default chunk size.
221  */
btrfs_update_space_info_chunk_size(struct btrfs_space_info * space_info,u64 chunk_size)222 void btrfs_update_space_info_chunk_size(struct btrfs_space_info *space_info,
223 					u64 chunk_size)
224 {
225 	WRITE_ONCE(space_info->chunk_size, chunk_size);
226 }
227 
create_space_info(struct btrfs_fs_info * info,u64 flags)228 static int create_space_info(struct btrfs_fs_info *info, u64 flags)
229 {
230 
231 	struct btrfs_space_info *space_info;
232 	int i;
233 	int ret;
234 
235 	space_info = kzalloc(sizeof(*space_info), GFP_NOFS);
236 	if (!space_info)
237 		return -ENOMEM;
238 
239 	space_info->fs_info = info;
240 	for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
241 		INIT_LIST_HEAD(&space_info->block_groups[i]);
242 	init_rwsem(&space_info->groups_sem);
243 	spin_lock_init(&space_info->lock);
244 	space_info->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
245 	space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
246 	INIT_LIST_HEAD(&space_info->ro_bgs);
247 	INIT_LIST_HEAD(&space_info->tickets);
248 	INIT_LIST_HEAD(&space_info->priority_tickets);
249 	space_info->clamp = 1;
250 	btrfs_update_space_info_chunk_size(space_info, calc_chunk_size(info, flags));
251 
252 	if (btrfs_is_zoned(info))
253 		space_info->bg_reclaim_threshold = BTRFS_DEFAULT_ZONED_RECLAIM_THRESH;
254 
255 	ret = btrfs_sysfs_add_space_info_type(info, space_info);
256 	if (ret)
257 		return ret;
258 
259 	list_add(&space_info->list, &info->space_info);
260 	if (flags & BTRFS_BLOCK_GROUP_DATA)
261 		info->data_sinfo = space_info;
262 
263 	return ret;
264 }
265 
btrfs_init_space_info(struct btrfs_fs_info * fs_info)266 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
267 {
268 	struct btrfs_super_block *disk_super;
269 	u64 features;
270 	u64 flags;
271 	int mixed = 0;
272 	int ret;
273 
274 	disk_super = fs_info->super_copy;
275 	if (!btrfs_super_root(disk_super))
276 		return -EINVAL;
277 
278 	features = btrfs_super_incompat_flags(disk_super);
279 	if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
280 		mixed = 1;
281 
282 	flags = BTRFS_BLOCK_GROUP_SYSTEM;
283 	ret = create_space_info(fs_info, flags);
284 	if (ret)
285 		goto out;
286 
287 	if (mixed) {
288 		flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
289 		ret = create_space_info(fs_info, flags);
290 	} else {
291 		flags = BTRFS_BLOCK_GROUP_METADATA;
292 		ret = create_space_info(fs_info, flags);
293 		if (ret)
294 			goto out;
295 
296 		flags = BTRFS_BLOCK_GROUP_DATA;
297 		ret = create_space_info(fs_info, flags);
298 	}
299 out:
300 	return ret;
301 }
302 
btrfs_add_bg_to_space_info(struct btrfs_fs_info * info,struct btrfs_block_group * block_group)303 void btrfs_add_bg_to_space_info(struct btrfs_fs_info *info,
304 				struct btrfs_block_group *block_group)
305 {
306 	struct btrfs_space_info *found;
307 	int factor, index;
308 
309 	factor = btrfs_bg_type_to_factor(block_group->flags);
310 
311 	found = btrfs_find_space_info(info, block_group->flags);
312 	ASSERT(found);
313 	spin_lock(&found->lock);
314 	found->total_bytes += block_group->length;
315 	found->disk_total += block_group->length * factor;
316 	found->bytes_used += block_group->used;
317 	found->disk_used += block_group->used * factor;
318 	found->bytes_readonly += block_group->bytes_super;
319 	btrfs_space_info_update_bytes_zone_unusable(info, found, block_group->zone_unusable);
320 	if (block_group->length > 0)
321 		found->full = 0;
322 	btrfs_try_granting_tickets(info, found);
323 	spin_unlock(&found->lock);
324 
325 	block_group->space_info = found;
326 
327 	index = btrfs_bg_flags_to_raid_index(block_group->flags);
328 	down_write(&found->groups_sem);
329 	list_add_tail(&block_group->list, &found->block_groups[index]);
330 	up_write(&found->groups_sem);
331 }
332 
btrfs_find_space_info(struct btrfs_fs_info * info,u64 flags)333 struct btrfs_space_info *btrfs_find_space_info(struct btrfs_fs_info *info,
334 					       u64 flags)
335 {
336 	struct list_head *head = &info->space_info;
337 	struct btrfs_space_info *found;
338 
339 	flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
340 
341 	list_for_each_entry(found, head, list) {
342 		if (found->flags & flags)
343 			return found;
344 	}
345 	return NULL;
346 }
347 
calc_effective_data_chunk_size(struct btrfs_fs_info * fs_info)348 static u64 calc_effective_data_chunk_size(struct btrfs_fs_info *fs_info)
349 {
350 	struct btrfs_space_info *data_sinfo;
351 	u64 data_chunk_size;
352 
353 	/*
354 	 * Calculate the data_chunk_size, space_info->chunk_size is the
355 	 * "optimal" chunk size based on the fs size.  However when we actually
356 	 * allocate the chunk we will strip this down further, making it no
357 	 * more than 10% of the disk or 1G, whichever is smaller.
358 	 *
359 	 * On the zoned mode, we need to use zone_size (= data_sinfo->chunk_size)
360 	 * as it is.
361 	 */
362 	data_sinfo = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
363 	if (btrfs_is_zoned(fs_info))
364 		return data_sinfo->chunk_size;
365 	data_chunk_size = min(data_sinfo->chunk_size,
366 			      mult_perc(fs_info->fs_devices->total_rw_bytes, 10));
367 	return min_t(u64, data_chunk_size, SZ_1G);
368 }
369 
calc_available_free_space(struct btrfs_fs_info * fs_info,const struct btrfs_space_info * space_info,enum btrfs_reserve_flush_enum flush)370 static u64 calc_available_free_space(struct btrfs_fs_info *fs_info,
371 			  const struct btrfs_space_info *space_info,
372 			  enum btrfs_reserve_flush_enum flush)
373 {
374 	u64 profile;
375 	u64 avail;
376 	u64 data_chunk_size;
377 	int factor;
378 
379 	if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
380 		profile = btrfs_system_alloc_profile(fs_info);
381 	else
382 		profile = btrfs_metadata_alloc_profile(fs_info);
383 
384 	avail = atomic64_read(&fs_info->free_chunk_space);
385 
386 	/*
387 	 * If we have dup, raid1 or raid10 then only half of the free
388 	 * space is actually usable.  For raid56, the space info used
389 	 * doesn't include the parity drive, so we don't have to
390 	 * change the math
391 	 */
392 	factor = btrfs_bg_type_to_factor(profile);
393 	avail = div_u64(avail, factor);
394 	if (avail == 0)
395 		return 0;
396 
397 	data_chunk_size = calc_effective_data_chunk_size(fs_info);
398 
399 	/*
400 	 * Since data allocations immediately use block groups as part of the
401 	 * reservation, because we assume that data reservations will == actual
402 	 * usage, we could potentially overcommit and then immediately have that
403 	 * available space used by a data allocation, which could put us in a
404 	 * bind when we get close to filling the file system.
405 	 *
406 	 * To handle this simply remove the data_chunk_size from the available
407 	 * space.  If we are relatively empty this won't affect our ability to
408 	 * overcommit much, and if we're very close to full it'll keep us from
409 	 * getting into a position where we've given ourselves very little
410 	 * metadata wiggle room.
411 	 */
412 	if (avail <= data_chunk_size)
413 		return 0;
414 	avail -= data_chunk_size;
415 
416 	/*
417 	 * If we aren't flushing all things, let us overcommit up to
418 	 * 1/2th of the space. If we can flush, don't let us overcommit
419 	 * too much, let it overcommit up to 1/8 of the space.
420 	 */
421 	if (flush == BTRFS_RESERVE_FLUSH_ALL)
422 		avail >>= 3;
423 	else
424 		avail >>= 1;
425 
426 	/*
427 	 * On the zoned mode, we always allocate one zone as one chunk.
428 	 * Returning non-zone size alingned bytes here will result in
429 	 * less pressure for the async metadata reclaim process, and it
430 	 * will over-commit too much leading to ENOSPC. Align down to the
431 	 * zone size to avoid that.
432 	 */
433 	if (btrfs_is_zoned(fs_info))
434 		avail = ALIGN_DOWN(avail, fs_info->zone_size);
435 
436 	return avail;
437 }
438 
btrfs_can_overcommit(struct btrfs_fs_info * fs_info,const struct btrfs_space_info * space_info,u64 bytes,enum btrfs_reserve_flush_enum flush)439 int btrfs_can_overcommit(struct btrfs_fs_info *fs_info,
440 			 const struct btrfs_space_info *space_info, u64 bytes,
441 			 enum btrfs_reserve_flush_enum flush)
442 {
443 	u64 avail;
444 	u64 used;
445 
446 	/* Don't overcommit when in mixed mode */
447 	if (space_info->flags & BTRFS_BLOCK_GROUP_DATA)
448 		return 0;
449 
450 	used = btrfs_space_info_used(space_info, true);
451 	avail = calc_available_free_space(fs_info, space_info, flush);
452 
453 	if (used + bytes < space_info->total_bytes + avail)
454 		return 1;
455 	return 0;
456 }
457 
remove_ticket(struct btrfs_space_info * space_info,struct reserve_ticket * ticket)458 static void remove_ticket(struct btrfs_space_info *space_info,
459 			  struct reserve_ticket *ticket)
460 {
461 	if (!list_empty(&ticket->list)) {
462 		list_del_init(&ticket->list);
463 		ASSERT(space_info->reclaim_size >= ticket->bytes);
464 		space_info->reclaim_size -= ticket->bytes;
465 	}
466 }
467 
468 /*
469  * This is for space we already have accounted in space_info->bytes_may_use, so
470  * basically when we're returning space from block_rsv's.
471  */
btrfs_try_granting_tickets(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info)472 void btrfs_try_granting_tickets(struct btrfs_fs_info *fs_info,
473 				struct btrfs_space_info *space_info)
474 {
475 	struct list_head *head;
476 	enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_NO_FLUSH;
477 
478 	lockdep_assert_held(&space_info->lock);
479 
480 	head = &space_info->priority_tickets;
481 again:
482 	while (!list_empty(head)) {
483 		struct reserve_ticket *ticket;
484 		u64 used = btrfs_space_info_used(space_info, true);
485 
486 		ticket = list_first_entry(head, struct reserve_ticket, list);
487 
488 		/* Check and see if our ticket can be satisfied now. */
489 		if ((used + ticket->bytes <= space_info->total_bytes) ||
490 		    btrfs_can_overcommit(fs_info, space_info, ticket->bytes,
491 					 flush)) {
492 			btrfs_space_info_update_bytes_may_use(fs_info,
493 							      space_info,
494 							      ticket->bytes);
495 			remove_ticket(space_info, ticket);
496 			ticket->bytes = 0;
497 			space_info->tickets_id++;
498 			wake_up(&ticket->wait);
499 		} else {
500 			break;
501 		}
502 	}
503 
504 	if (head == &space_info->priority_tickets) {
505 		head = &space_info->tickets;
506 		flush = BTRFS_RESERVE_FLUSH_ALL;
507 		goto again;
508 	}
509 }
510 
511 #define DUMP_BLOCK_RSV(fs_info, rsv_name)				\
512 do {									\
513 	struct btrfs_block_rsv *__rsv = &(fs_info)->rsv_name;		\
514 	spin_lock(&__rsv->lock);					\
515 	btrfs_info(fs_info, #rsv_name ": size %llu reserved %llu",	\
516 		   __rsv->size, __rsv->reserved);			\
517 	spin_unlock(&__rsv->lock);					\
518 } while (0)
519 
space_info_flag_to_str(const struct btrfs_space_info * space_info)520 static const char *space_info_flag_to_str(const struct btrfs_space_info *space_info)
521 {
522 	switch (space_info->flags) {
523 	case BTRFS_BLOCK_GROUP_SYSTEM:
524 		return "SYSTEM";
525 	case BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA:
526 		return "DATA+METADATA";
527 	case BTRFS_BLOCK_GROUP_DATA:
528 		return "DATA";
529 	case BTRFS_BLOCK_GROUP_METADATA:
530 		return "METADATA";
531 	default:
532 		return "UNKNOWN";
533 	}
534 }
535 
dump_global_block_rsv(struct btrfs_fs_info * fs_info)536 static void dump_global_block_rsv(struct btrfs_fs_info *fs_info)
537 {
538 	DUMP_BLOCK_RSV(fs_info, global_block_rsv);
539 	DUMP_BLOCK_RSV(fs_info, trans_block_rsv);
540 	DUMP_BLOCK_RSV(fs_info, chunk_block_rsv);
541 	DUMP_BLOCK_RSV(fs_info, delayed_block_rsv);
542 	DUMP_BLOCK_RSV(fs_info, delayed_refs_rsv);
543 }
544 
__btrfs_dump_space_info(const struct btrfs_fs_info * fs_info,const struct btrfs_space_info * info)545 static void __btrfs_dump_space_info(const struct btrfs_fs_info *fs_info,
546 				    const struct btrfs_space_info *info)
547 {
548 	const char *flag_str = space_info_flag_to_str(info);
549 	lockdep_assert_held(&info->lock);
550 
551 	/* The free space could be negative in case of overcommit */
552 	btrfs_info(fs_info, "space_info %s has %lld free, is %sfull",
553 		   flag_str,
554 		   (s64)(info->total_bytes - btrfs_space_info_used(info, true)),
555 		   info->full ? "" : "not ");
556 	btrfs_info(fs_info,
557 "space_info total=%llu, used=%llu, pinned=%llu, reserved=%llu, may_use=%llu, readonly=%llu zone_unusable=%llu",
558 		info->total_bytes, info->bytes_used, info->bytes_pinned,
559 		info->bytes_reserved, info->bytes_may_use,
560 		info->bytes_readonly, info->bytes_zone_unusable);
561 }
562 
btrfs_dump_space_info(struct btrfs_fs_info * fs_info,struct btrfs_space_info * info,u64 bytes,int dump_block_groups)563 void btrfs_dump_space_info(struct btrfs_fs_info *fs_info,
564 			   struct btrfs_space_info *info, u64 bytes,
565 			   int dump_block_groups)
566 {
567 	struct btrfs_block_group *cache;
568 	u64 total_avail = 0;
569 	int index = 0;
570 
571 	spin_lock(&info->lock);
572 	__btrfs_dump_space_info(fs_info, info);
573 	dump_global_block_rsv(fs_info);
574 	spin_unlock(&info->lock);
575 
576 	if (!dump_block_groups)
577 		return;
578 
579 	down_read(&info->groups_sem);
580 again:
581 	list_for_each_entry(cache, &info->block_groups[index], list) {
582 		u64 avail;
583 
584 		spin_lock(&cache->lock);
585 		avail = cache->length - cache->used - cache->pinned -
586 			cache->reserved - cache->bytes_super - cache->zone_unusable;
587 		btrfs_info(fs_info,
588 "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %llu delalloc %llu super %llu zone_unusable (%llu bytes available) %s",
589 			   cache->start, cache->length, cache->used, cache->pinned,
590 			   cache->reserved, cache->delalloc_bytes,
591 			   cache->bytes_super, cache->zone_unusable,
592 			   avail, cache->ro ? "[readonly]" : "");
593 		spin_unlock(&cache->lock);
594 		btrfs_dump_free_space(cache, bytes);
595 		total_avail += avail;
596 	}
597 	if (++index < BTRFS_NR_RAID_TYPES)
598 		goto again;
599 	up_read(&info->groups_sem);
600 
601 	btrfs_info(fs_info, "%llu bytes available across all block groups", total_avail);
602 }
603 
calc_reclaim_items_nr(const struct btrfs_fs_info * fs_info,u64 to_reclaim)604 static inline u64 calc_reclaim_items_nr(const struct btrfs_fs_info *fs_info,
605 					u64 to_reclaim)
606 {
607 	u64 bytes;
608 	u64 nr;
609 
610 	bytes = btrfs_calc_insert_metadata_size(fs_info, 1);
611 	nr = div64_u64(to_reclaim, bytes);
612 	if (!nr)
613 		nr = 1;
614 	return nr;
615 }
616 
617 /*
618  * shrink metadata reservation for delalloc
619  */
shrink_delalloc(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,u64 to_reclaim,bool wait_ordered,bool for_preempt)620 static void shrink_delalloc(struct btrfs_fs_info *fs_info,
621 			    struct btrfs_space_info *space_info,
622 			    u64 to_reclaim, bool wait_ordered,
623 			    bool for_preempt)
624 {
625 	struct btrfs_trans_handle *trans;
626 	u64 delalloc_bytes;
627 	u64 ordered_bytes;
628 	u64 items;
629 	long time_left;
630 	int loops;
631 
632 	delalloc_bytes = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
633 	ordered_bytes = percpu_counter_sum_positive(&fs_info->ordered_bytes);
634 	if (delalloc_bytes == 0 && ordered_bytes == 0)
635 		return;
636 
637 	/* Calc the number of the pages we need flush for space reservation */
638 	if (to_reclaim == U64_MAX) {
639 		items = U64_MAX;
640 	} else {
641 		/*
642 		 * to_reclaim is set to however much metadata we need to
643 		 * reclaim, but reclaiming that much data doesn't really track
644 		 * exactly.  What we really want to do is reclaim full inode's
645 		 * worth of reservations, however that's not available to us
646 		 * here.  We will take a fraction of the delalloc bytes for our
647 		 * flushing loops and hope for the best.  Delalloc will expand
648 		 * the amount we write to cover an entire dirty extent, which
649 		 * will reclaim the metadata reservation for that range.  If
650 		 * it's not enough subsequent flush stages will be more
651 		 * aggressive.
652 		 */
653 		to_reclaim = max(to_reclaim, delalloc_bytes >> 3);
654 		items = calc_reclaim_items_nr(fs_info, to_reclaim) * 2;
655 	}
656 
657 	trans = current->journal_info;
658 
659 	/*
660 	 * If we are doing more ordered than delalloc we need to just wait on
661 	 * ordered extents, otherwise we'll waste time trying to flush delalloc
662 	 * that likely won't give us the space back we need.
663 	 */
664 	if (ordered_bytes > delalloc_bytes && !for_preempt)
665 		wait_ordered = true;
666 
667 	loops = 0;
668 	while ((delalloc_bytes || ordered_bytes) && loops < 3) {
669 		u64 temp = min(delalloc_bytes, to_reclaim) >> PAGE_SHIFT;
670 		long nr_pages = min_t(u64, temp, LONG_MAX);
671 		int async_pages;
672 
673 		btrfs_start_delalloc_roots(fs_info, nr_pages, true);
674 
675 		/*
676 		 * We need to make sure any outstanding async pages are now
677 		 * processed before we continue.  This is because things like
678 		 * sync_inode() try to be smart and skip writing if the inode is
679 		 * marked clean.  We don't use filemap_fwrite for flushing
680 		 * because we want to control how many pages we write out at a
681 		 * time, thus this is the only safe way to make sure we've
682 		 * waited for outstanding compressed workers to have started
683 		 * their jobs and thus have ordered extents set up properly.
684 		 *
685 		 * This exists because we do not want to wait for each
686 		 * individual inode to finish its async work, we simply want to
687 		 * start the IO on everybody, and then come back here and wait
688 		 * for all of the async work to catch up.  Once we're done with
689 		 * that we know we'll have ordered extents for everything and we
690 		 * can decide if we wait for that or not.
691 		 *
692 		 * If we choose to replace this in the future, make absolutely
693 		 * sure that the proper waiting is being done in the async case,
694 		 * as there have been bugs in that area before.
695 		 */
696 		async_pages = atomic_read(&fs_info->async_delalloc_pages);
697 		if (!async_pages)
698 			goto skip_async;
699 
700 		/*
701 		 * We don't want to wait forever, if we wrote less pages in this
702 		 * loop than we have outstanding, only wait for that number of
703 		 * pages, otherwise we can wait for all async pages to finish
704 		 * before continuing.
705 		 */
706 		if (async_pages > nr_pages)
707 			async_pages -= nr_pages;
708 		else
709 			async_pages = 0;
710 		wait_event(fs_info->async_submit_wait,
711 			   atomic_read(&fs_info->async_delalloc_pages) <=
712 			   async_pages);
713 skip_async:
714 		loops++;
715 		if (wait_ordered && !trans) {
716 			btrfs_wait_ordered_roots(fs_info, items, NULL);
717 		} else {
718 			time_left = schedule_timeout_killable(1);
719 			if (time_left)
720 				break;
721 		}
722 
723 		/*
724 		 * If we are for preemption we just want a one-shot of delalloc
725 		 * flushing so we can stop flushing if we decide we don't need
726 		 * to anymore.
727 		 */
728 		if (for_preempt)
729 			break;
730 
731 		spin_lock(&space_info->lock);
732 		if (list_empty(&space_info->tickets) &&
733 		    list_empty(&space_info->priority_tickets)) {
734 			spin_unlock(&space_info->lock);
735 			break;
736 		}
737 		spin_unlock(&space_info->lock);
738 
739 		delalloc_bytes = percpu_counter_sum_positive(
740 						&fs_info->delalloc_bytes);
741 		ordered_bytes = percpu_counter_sum_positive(
742 						&fs_info->ordered_bytes);
743 	}
744 }
745 
746 /*
747  * Try to flush some data based on policy set by @state. This is only advisory
748  * and may fail for various reasons. The caller is supposed to examine the
749  * state of @space_info to detect the outcome.
750  */
flush_space(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,u64 num_bytes,enum btrfs_flush_state state,bool for_preempt)751 static void flush_space(struct btrfs_fs_info *fs_info,
752 		       struct btrfs_space_info *space_info, u64 num_bytes,
753 		       enum btrfs_flush_state state, bool for_preempt)
754 {
755 	struct btrfs_root *root = fs_info->tree_root;
756 	struct btrfs_trans_handle *trans;
757 	int nr;
758 	int ret = 0;
759 
760 	switch (state) {
761 	case FLUSH_DELAYED_ITEMS_NR:
762 	case FLUSH_DELAYED_ITEMS:
763 		if (state == FLUSH_DELAYED_ITEMS_NR)
764 			nr = calc_reclaim_items_nr(fs_info, num_bytes) * 2;
765 		else
766 			nr = -1;
767 
768 		trans = btrfs_join_transaction_nostart(root);
769 		if (IS_ERR(trans)) {
770 			ret = PTR_ERR(trans);
771 			if (ret == -ENOENT)
772 				ret = 0;
773 			break;
774 		}
775 		ret = btrfs_run_delayed_items_nr(trans, nr);
776 		btrfs_end_transaction(trans);
777 		break;
778 	case FLUSH_DELALLOC:
779 	case FLUSH_DELALLOC_WAIT:
780 	case FLUSH_DELALLOC_FULL:
781 		if (state == FLUSH_DELALLOC_FULL)
782 			num_bytes = U64_MAX;
783 		shrink_delalloc(fs_info, space_info, num_bytes,
784 				state != FLUSH_DELALLOC, for_preempt);
785 		break;
786 	case FLUSH_DELAYED_REFS_NR:
787 	case FLUSH_DELAYED_REFS:
788 		trans = btrfs_join_transaction_nostart(root);
789 		if (IS_ERR(trans)) {
790 			ret = PTR_ERR(trans);
791 			if (ret == -ENOENT)
792 				ret = 0;
793 			break;
794 		}
795 		if (state == FLUSH_DELAYED_REFS_NR)
796 			btrfs_run_delayed_refs(trans, num_bytes);
797 		else
798 			btrfs_run_delayed_refs(trans, 0);
799 		btrfs_end_transaction(trans);
800 		break;
801 	case ALLOC_CHUNK:
802 	case ALLOC_CHUNK_FORCE:
803 		trans = btrfs_join_transaction(root);
804 		if (IS_ERR(trans)) {
805 			ret = PTR_ERR(trans);
806 			break;
807 		}
808 		ret = btrfs_chunk_alloc(trans,
809 				btrfs_get_alloc_profile(fs_info, space_info->flags),
810 				(state == ALLOC_CHUNK) ? CHUNK_ALLOC_NO_FORCE :
811 					CHUNK_ALLOC_FORCE);
812 		btrfs_end_transaction(trans);
813 
814 		if (ret > 0 || ret == -ENOSPC)
815 			ret = 0;
816 		break;
817 	case RUN_DELAYED_IPUTS:
818 		/*
819 		 * If we have pending delayed iputs then we could free up a
820 		 * bunch of pinned space, so make sure we run the iputs before
821 		 * we do our pinned bytes check below.
822 		 */
823 		btrfs_run_delayed_iputs(fs_info);
824 		btrfs_wait_on_delayed_iputs(fs_info);
825 		break;
826 	case COMMIT_TRANS:
827 		ASSERT(current->journal_info == NULL);
828 		/*
829 		 * We don't want to start a new transaction, just attach to the
830 		 * current one or wait it fully commits in case its commit is
831 		 * happening at the moment. Note: we don't use a nostart join
832 		 * because that does not wait for a transaction to fully commit
833 		 * (only for it to be unblocked, state TRANS_STATE_UNBLOCKED).
834 		 */
835 		ret = btrfs_commit_current_transaction(root);
836 		break;
837 	default:
838 		ret = -ENOSPC;
839 		break;
840 	}
841 
842 	trace_btrfs_flush_space(fs_info, space_info->flags, num_bytes, state,
843 				ret, for_preempt);
844 	return;
845 }
846 
btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info * fs_info,const struct btrfs_space_info * space_info)847 static u64 btrfs_calc_reclaim_metadata_size(struct btrfs_fs_info *fs_info,
848 					    const struct btrfs_space_info *space_info)
849 {
850 	u64 used;
851 	u64 avail;
852 	u64 to_reclaim = space_info->reclaim_size;
853 
854 	lockdep_assert_held(&space_info->lock);
855 
856 	avail = calc_available_free_space(fs_info, space_info,
857 					  BTRFS_RESERVE_FLUSH_ALL);
858 	used = btrfs_space_info_used(space_info, true);
859 
860 	/*
861 	 * We may be flushing because suddenly we have less space than we had
862 	 * before, and now we're well over-committed based on our current free
863 	 * space.  If that's the case add in our overage so we make sure to put
864 	 * appropriate pressure on the flushing state machine.
865 	 */
866 	if (space_info->total_bytes + avail < used)
867 		to_reclaim += used - (space_info->total_bytes + avail);
868 
869 	return to_reclaim;
870 }
871 
need_preemptive_reclaim(struct btrfs_fs_info * fs_info,const struct btrfs_space_info * space_info)872 static bool need_preemptive_reclaim(struct btrfs_fs_info *fs_info,
873 				    const struct btrfs_space_info *space_info)
874 {
875 	const u64 global_rsv_size = btrfs_block_rsv_reserved(&fs_info->global_block_rsv);
876 	u64 ordered, delalloc;
877 	u64 thresh;
878 	u64 used;
879 
880 	thresh = mult_perc(space_info->total_bytes, 90);
881 
882 	lockdep_assert_held(&space_info->lock);
883 
884 	/* If we're just plain full then async reclaim just slows us down. */
885 	if ((space_info->bytes_used + space_info->bytes_reserved +
886 	     global_rsv_size) >= thresh)
887 		return false;
888 
889 	used = space_info->bytes_may_use + space_info->bytes_pinned;
890 
891 	/* The total flushable belongs to the global rsv, don't flush. */
892 	if (global_rsv_size >= used)
893 		return false;
894 
895 	/*
896 	 * 128MiB is 1/4 of the maximum global rsv size.  If we have less than
897 	 * that devoted to other reservations then there's no sense in flushing,
898 	 * we don't have a lot of things that need flushing.
899 	 */
900 	if (used - global_rsv_size <= SZ_128M)
901 		return false;
902 
903 	/*
904 	 * We have tickets queued, bail so we don't compete with the async
905 	 * flushers.
906 	 */
907 	if (space_info->reclaim_size)
908 		return false;
909 
910 	/*
911 	 * If we have over half of the free space occupied by reservations or
912 	 * pinned then we want to start flushing.
913 	 *
914 	 * We do not do the traditional thing here, which is to say
915 	 *
916 	 *   if (used >= ((total_bytes + avail) / 2))
917 	 *     return 1;
918 	 *
919 	 * because this doesn't quite work how we want.  If we had more than 50%
920 	 * of the space_info used by bytes_used and we had 0 available we'd just
921 	 * constantly run the background flusher.  Instead we want it to kick in
922 	 * if our reclaimable space exceeds our clamped free space.
923 	 *
924 	 * Our clamping range is 2^1 -> 2^8.  Practically speaking that means
925 	 * the following:
926 	 *
927 	 * Amount of RAM        Minimum threshold       Maximum threshold
928 	 *
929 	 *        256GiB                     1GiB                  128GiB
930 	 *        128GiB                   512MiB                   64GiB
931 	 *         64GiB                   256MiB                   32GiB
932 	 *         32GiB                   128MiB                   16GiB
933 	 *         16GiB                    64MiB                    8GiB
934 	 *
935 	 * These are the range our thresholds will fall in, corresponding to how
936 	 * much delalloc we need for the background flusher to kick in.
937 	 */
938 
939 	thresh = calc_available_free_space(fs_info, space_info,
940 					   BTRFS_RESERVE_FLUSH_ALL);
941 	used = space_info->bytes_used + space_info->bytes_reserved +
942 	       space_info->bytes_readonly + global_rsv_size;
943 	if (used < space_info->total_bytes)
944 		thresh += space_info->total_bytes - used;
945 	thresh >>= space_info->clamp;
946 
947 	used = space_info->bytes_pinned;
948 
949 	/*
950 	 * If we have more ordered bytes than delalloc bytes then we're either
951 	 * doing a lot of DIO, or we simply don't have a lot of delalloc waiting
952 	 * around.  Preemptive flushing is only useful in that it can free up
953 	 * space before tickets need to wait for things to finish.  In the case
954 	 * of ordered extents, preemptively waiting on ordered extents gets us
955 	 * nothing, if our reservations are tied up in ordered extents we'll
956 	 * simply have to slow down writers by forcing them to wait on ordered
957 	 * extents.
958 	 *
959 	 * In the case that ordered is larger than delalloc, only include the
960 	 * block reserves that we would actually be able to directly reclaim
961 	 * from.  In this case if we're heavy on metadata operations this will
962 	 * clearly be heavy enough to warrant preemptive flushing.  In the case
963 	 * of heavy DIO or ordered reservations, preemptive flushing will just
964 	 * waste time and cause us to slow down.
965 	 *
966 	 * We want to make sure we truly are maxed out on ordered however, so
967 	 * cut ordered in half, and if it's still higher than delalloc then we
968 	 * can keep flushing.  This is to avoid the case where we start
969 	 * flushing, and now delalloc == ordered and we stop preemptively
970 	 * flushing when we could still have several gigs of delalloc to flush.
971 	 */
972 	ordered = percpu_counter_read_positive(&fs_info->ordered_bytes) >> 1;
973 	delalloc = percpu_counter_read_positive(&fs_info->delalloc_bytes);
974 	if (ordered >= delalloc)
975 		used += btrfs_block_rsv_reserved(&fs_info->delayed_refs_rsv) +
976 			btrfs_block_rsv_reserved(&fs_info->delayed_block_rsv);
977 	else
978 		used += space_info->bytes_may_use - global_rsv_size;
979 
980 	return (used >= thresh && !btrfs_fs_closing(fs_info) &&
981 		!test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
982 }
983 
steal_from_global_rsv(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,struct reserve_ticket * ticket)984 static bool steal_from_global_rsv(struct btrfs_fs_info *fs_info,
985 				  struct btrfs_space_info *space_info,
986 				  struct reserve_ticket *ticket)
987 {
988 	struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
989 	u64 min_bytes;
990 
991 	if (!ticket->steal)
992 		return false;
993 
994 	if (global_rsv->space_info != space_info)
995 		return false;
996 
997 	spin_lock(&global_rsv->lock);
998 	min_bytes = mult_perc(global_rsv->size, 10);
999 	if (global_rsv->reserved < min_bytes + ticket->bytes) {
1000 		spin_unlock(&global_rsv->lock);
1001 		return false;
1002 	}
1003 	global_rsv->reserved -= ticket->bytes;
1004 	remove_ticket(space_info, ticket);
1005 	ticket->bytes = 0;
1006 	wake_up(&ticket->wait);
1007 	space_info->tickets_id++;
1008 	if (global_rsv->reserved < global_rsv->size)
1009 		global_rsv->full = 0;
1010 	spin_unlock(&global_rsv->lock);
1011 
1012 	return true;
1013 }
1014 
1015 /*
1016  * We've exhausted our flushing, start failing tickets.
1017  *
1018  * @fs_info - fs_info for this fs
1019  * @space_info - the space info we were flushing
1020  *
1021  * We call this when we've exhausted our flushing ability and haven't made
1022  * progress in satisfying tickets.  The reservation code handles tickets in
1023  * order, so if there is a large ticket first and then smaller ones we could
1024  * very well satisfy the smaller tickets.  This will attempt to wake up any
1025  * tickets in the list to catch this case.
1026  *
1027  * This function returns true if it was able to make progress by clearing out
1028  * other tickets, or if it stumbles across a ticket that was smaller than the
1029  * first ticket.
1030  */
maybe_fail_all_tickets(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info)1031 static bool maybe_fail_all_tickets(struct btrfs_fs_info *fs_info,
1032 				   struct btrfs_space_info *space_info)
1033 {
1034 	struct reserve_ticket *ticket;
1035 	u64 tickets_id = space_info->tickets_id;
1036 	const bool aborted = BTRFS_FS_ERROR(fs_info);
1037 
1038 	trace_btrfs_fail_all_tickets(fs_info, space_info);
1039 
1040 	if (btrfs_test_opt(fs_info, ENOSPC_DEBUG)) {
1041 		btrfs_info(fs_info, "cannot satisfy tickets, dumping space info");
1042 		__btrfs_dump_space_info(fs_info, space_info);
1043 	}
1044 
1045 	while (!list_empty(&space_info->tickets) &&
1046 	       tickets_id == space_info->tickets_id) {
1047 		ticket = list_first_entry(&space_info->tickets,
1048 					  struct reserve_ticket, list);
1049 
1050 		if (!aborted && steal_from_global_rsv(fs_info, space_info, ticket))
1051 			return true;
1052 
1053 		if (!aborted && btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1054 			btrfs_info(fs_info, "failing ticket with %llu bytes",
1055 				   ticket->bytes);
1056 
1057 		remove_ticket(space_info, ticket);
1058 		if (aborted)
1059 			ticket->error = -EIO;
1060 		else
1061 			ticket->error = -ENOSPC;
1062 		wake_up(&ticket->wait);
1063 
1064 		/*
1065 		 * We're just throwing tickets away, so more flushing may not
1066 		 * trip over btrfs_try_granting_tickets, so we need to call it
1067 		 * here to see if we can make progress with the next ticket in
1068 		 * the list.
1069 		 */
1070 		if (!aborted)
1071 			btrfs_try_granting_tickets(fs_info, space_info);
1072 	}
1073 	return (tickets_id != space_info->tickets_id);
1074 }
1075 
1076 /*
1077  * This is for normal flushers, we can wait all goddamned day if we want to.  We
1078  * will loop and continuously try to flush as long as we are making progress.
1079  * We count progress as clearing off tickets each time we have to loop.
1080  */
btrfs_async_reclaim_metadata_space(struct work_struct * work)1081 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
1082 {
1083 	struct btrfs_fs_info *fs_info;
1084 	struct btrfs_space_info *space_info;
1085 	u64 to_reclaim;
1086 	enum btrfs_flush_state flush_state;
1087 	int commit_cycles = 0;
1088 	u64 last_tickets_id;
1089 
1090 	fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
1091 	space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1092 
1093 	spin_lock(&space_info->lock);
1094 	to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1095 	if (!to_reclaim) {
1096 		space_info->flush = 0;
1097 		spin_unlock(&space_info->lock);
1098 		return;
1099 	}
1100 	last_tickets_id = space_info->tickets_id;
1101 	spin_unlock(&space_info->lock);
1102 
1103 	flush_state = FLUSH_DELAYED_ITEMS_NR;
1104 	do {
1105 		flush_space(fs_info, space_info, to_reclaim, flush_state, false);
1106 		spin_lock(&space_info->lock);
1107 		if (list_empty(&space_info->tickets)) {
1108 			space_info->flush = 0;
1109 			spin_unlock(&space_info->lock);
1110 			return;
1111 		}
1112 		to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info,
1113 							      space_info);
1114 		if (last_tickets_id == space_info->tickets_id) {
1115 			flush_state++;
1116 		} else {
1117 			last_tickets_id = space_info->tickets_id;
1118 			flush_state = FLUSH_DELAYED_ITEMS_NR;
1119 			if (commit_cycles)
1120 				commit_cycles--;
1121 		}
1122 
1123 		/*
1124 		 * We do not want to empty the system of delalloc unless we're
1125 		 * under heavy pressure, so allow one trip through the flushing
1126 		 * logic before we start doing a FLUSH_DELALLOC_FULL.
1127 		 */
1128 		if (flush_state == FLUSH_DELALLOC_FULL && !commit_cycles)
1129 			flush_state++;
1130 
1131 		/*
1132 		 * We don't want to force a chunk allocation until we've tried
1133 		 * pretty hard to reclaim space.  Think of the case where we
1134 		 * freed up a bunch of space and so have a lot of pinned space
1135 		 * to reclaim.  We would rather use that than possibly create a
1136 		 * underutilized metadata chunk.  So if this is our first run
1137 		 * through the flushing state machine skip ALLOC_CHUNK_FORCE and
1138 		 * commit the transaction.  If nothing has changed the next go
1139 		 * around then we can force a chunk allocation.
1140 		 */
1141 		if (flush_state == ALLOC_CHUNK_FORCE && !commit_cycles)
1142 			flush_state++;
1143 
1144 		if (flush_state > COMMIT_TRANS) {
1145 			commit_cycles++;
1146 			if (commit_cycles > 2) {
1147 				if (maybe_fail_all_tickets(fs_info, space_info)) {
1148 					flush_state = FLUSH_DELAYED_ITEMS_NR;
1149 					commit_cycles--;
1150 				} else {
1151 					space_info->flush = 0;
1152 				}
1153 			} else {
1154 				flush_state = FLUSH_DELAYED_ITEMS_NR;
1155 			}
1156 		}
1157 		spin_unlock(&space_info->lock);
1158 	} while (flush_state <= COMMIT_TRANS);
1159 }
1160 
1161 /*
1162  * This handles pre-flushing of metadata space before we get to the point that
1163  * we need to start blocking threads on tickets.  The logic here is different
1164  * from the other flush paths because it doesn't rely on tickets to tell us how
1165  * much we need to flush, instead it attempts to keep us below the 80% full
1166  * watermark of space by flushing whichever reservation pool is currently the
1167  * largest.
1168  */
btrfs_preempt_reclaim_metadata_space(struct work_struct * work)1169 static void btrfs_preempt_reclaim_metadata_space(struct work_struct *work)
1170 {
1171 	struct btrfs_fs_info *fs_info;
1172 	struct btrfs_space_info *space_info;
1173 	struct btrfs_block_rsv *delayed_block_rsv;
1174 	struct btrfs_block_rsv *delayed_refs_rsv;
1175 	struct btrfs_block_rsv *global_rsv;
1176 	struct btrfs_block_rsv *trans_rsv;
1177 	int loops = 0;
1178 
1179 	fs_info = container_of(work, struct btrfs_fs_info,
1180 			       preempt_reclaim_work);
1181 	space_info = btrfs_find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
1182 	delayed_block_rsv = &fs_info->delayed_block_rsv;
1183 	delayed_refs_rsv = &fs_info->delayed_refs_rsv;
1184 	global_rsv = &fs_info->global_block_rsv;
1185 	trans_rsv = &fs_info->trans_block_rsv;
1186 
1187 	spin_lock(&space_info->lock);
1188 	while (need_preemptive_reclaim(fs_info, space_info)) {
1189 		enum btrfs_flush_state flush;
1190 		u64 delalloc_size = 0;
1191 		u64 to_reclaim, block_rsv_size;
1192 		const u64 global_rsv_size = btrfs_block_rsv_reserved(global_rsv);
1193 
1194 		loops++;
1195 
1196 		/*
1197 		 * We don't have a precise counter for the metadata being
1198 		 * reserved for delalloc, so we'll approximate it by subtracting
1199 		 * out the block rsv's space from the bytes_may_use.  If that
1200 		 * amount is higher than the individual reserves, then we can
1201 		 * assume it's tied up in delalloc reservations.
1202 		 */
1203 		block_rsv_size = global_rsv_size +
1204 			btrfs_block_rsv_reserved(delayed_block_rsv) +
1205 			btrfs_block_rsv_reserved(delayed_refs_rsv) +
1206 			btrfs_block_rsv_reserved(trans_rsv);
1207 		if (block_rsv_size < space_info->bytes_may_use)
1208 			delalloc_size = space_info->bytes_may_use - block_rsv_size;
1209 
1210 		/*
1211 		 * We don't want to include the global_rsv in our calculation,
1212 		 * because that's space we can't touch.  Subtract it from the
1213 		 * block_rsv_size for the next checks.
1214 		 */
1215 		block_rsv_size -= global_rsv_size;
1216 
1217 		/*
1218 		 * We really want to avoid flushing delalloc too much, as it
1219 		 * could result in poor allocation patterns, so only flush it if
1220 		 * it's larger than the rest of the pools combined.
1221 		 */
1222 		if (delalloc_size > block_rsv_size) {
1223 			to_reclaim = delalloc_size;
1224 			flush = FLUSH_DELALLOC;
1225 		} else if (space_info->bytes_pinned >
1226 			   (btrfs_block_rsv_reserved(delayed_block_rsv) +
1227 			    btrfs_block_rsv_reserved(delayed_refs_rsv))) {
1228 			to_reclaim = space_info->bytes_pinned;
1229 			flush = COMMIT_TRANS;
1230 		} else if (btrfs_block_rsv_reserved(delayed_block_rsv) >
1231 			   btrfs_block_rsv_reserved(delayed_refs_rsv)) {
1232 			to_reclaim = btrfs_block_rsv_reserved(delayed_block_rsv);
1233 			flush = FLUSH_DELAYED_ITEMS_NR;
1234 		} else {
1235 			to_reclaim = btrfs_block_rsv_reserved(delayed_refs_rsv);
1236 			flush = FLUSH_DELAYED_REFS_NR;
1237 		}
1238 
1239 		spin_unlock(&space_info->lock);
1240 
1241 		/*
1242 		 * We don't want to reclaim everything, just a portion, so scale
1243 		 * down the to_reclaim by 1/4.  If it takes us down to 0,
1244 		 * reclaim 1 items worth.
1245 		 */
1246 		to_reclaim >>= 2;
1247 		if (!to_reclaim)
1248 			to_reclaim = btrfs_calc_insert_metadata_size(fs_info, 1);
1249 		flush_space(fs_info, space_info, to_reclaim, flush, true);
1250 		cond_resched();
1251 		spin_lock(&space_info->lock);
1252 	}
1253 
1254 	/* We only went through once, back off our clamping. */
1255 	if (loops == 1 && !space_info->reclaim_size)
1256 		space_info->clamp = max(1, space_info->clamp - 1);
1257 	trace_btrfs_done_preemptive_reclaim(fs_info, space_info);
1258 	spin_unlock(&space_info->lock);
1259 }
1260 
1261 /*
1262  * FLUSH_DELALLOC_WAIT:
1263  *   Space is freed from flushing delalloc in one of two ways.
1264  *
1265  *   1) compression is on and we allocate less space than we reserved
1266  *   2) we are overwriting existing space
1267  *
1268  *   For #1 that extra space is reclaimed as soon as the delalloc pages are
1269  *   COWed, by way of btrfs_add_reserved_bytes() which adds the actual extent
1270  *   length to ->bytes_reserved, and subtracts the reserved space from
1271  *   ->bytes_may_use.
1272  *
1273  *   For #2 this is trickier.  Once the ordered extent runs we will drop the
1274  *   extent in the range we are overwriting, which creates a delayed ref for
1275  *   that freed extent.  This however is not reclaimed until the transaction
1276  *   commits, thus the next stages.
1277  *
1278  * RUN_DELAYED_IPUTS
1279  *   If we are freeing inodes, we want to make sure all delayed iputs have
1280  *   completed, because they could have been on an inode with i_nlink == 0, and
1281  *   thus have been truncated and freed up space.  But again this space is not
1282  *   immediately re-usable, it comes in the form of a delayed ref, which must be
1283  *   run and then the transaction must be committed.
1284  *
1285  * COMMIT_TRANS
1286  *   This is where we reclaim all of the pinned space generated by running the
1287  *   iputs
1288  *
1289  * ALLOC_CHUNK_FORCE
1290  *   For data we start with alloc chunk force, however we could have been full
1291  *   before, and then the transaction commit could have freed new block groups,
1292  *   so if we now have space to allocate do the force chunk allocation.
1293  */
1294 static const enum btrfs_flush_state data_flush_states[] = {
1295 	FLUSH_DELALLOC_FULL,
1296 	RUN_DELAYED_IPUTS,
1297 	COMMIT_TRANS,
1298 	ALLOC_CHUNK_FORCE,
1299 };
1300 
btrfs_async_reclaim_data_space(struct work_struct * work)1301 static void btrfs_async_reclaim_data_space(struct work_struct *work)
1302 {
1303 	struct btrfs_fs_info *fs_info;
1304 	struct btrfs_space_info *space_info;
1305 	u64 last_tickets_id;
1306 	enum btrfs_flush_state flush_state = 0;
1307 
1308 	fs_info = container_of(work, struct btrfs_fs_info, async_data_reclaim_work);
1309 	space_info = fs_info->data_sinfo;
1310 
1311 	spin_lock(&space_info->lock);
1312 	if (list_empty(&space_info->tickets)) {
1313 		space_info->flush = 0;
1314 		spin_unlock(&space_info->lock);
1315 		return;
1316 	}
1317 	last_tickets_id = space_info->tickets_id;
1318 	spin_unlock(&space_info->lock);
1319 
1320 	while (!space_info->full) {
1321 		flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1322 		spin_lock(&space_info->lock);
1323 		if (list_empty(&space_info->tickets)) {
1324 			space_info->flush = 0;
1325 			spin_unlock(&space_info->lock);
1326 			return;
1327 		}
1328 
1329 		/* Something happened, fail everything and bail. */
1330 		if (BTRFS_FS_ERROR(fs_info))
1331 			goto aborted_fs;
1332 		last_tickets_id = space_info->tickets_id;
1333 		spin_unlock(&space_info->lock);
1334 	}
1335 
1336 	while (flush_state < ARRAY_SIZE(data_flush_states)) {
1337 		flush_space(fs_info, space_info, U64_MAX,
1338 			    data_flush_states[flush_state], false);
1339 		spin_lock(&space_info->lock);
1340 		if (list_empty(&space_info->tickets)) {
1341 			space_info->flush = 0;
1342 			spin_unlock(&space_info->lock);
1343 			return;
1344 		}
1345 
1346 		if (last_tickets_id == space_info->tickets_id) {
1347 			flush_state++;
1348 		} else {
1349 			last_tickets_id = space_info->tickets_id;
1350 			flush_state = 0;
1351 		}
1352 
1353 		if (flush_state >= ARRAY_SIZE(data_flush_states)) {
1354 			if (space_info->full) {
1355 				if (maybe_fail_all_tickets(fs_info, space_info))
1356 					flush_state = 0;
1357 				else
1358 					space_info->flush = 0;
1359 			} else {
1360 				flush_state = 0;
1361 			}
1362 
1363 			/* Something happened, fail everything and bail. */
1364 			if (BTRFS_FS_ERROR(fs_info))
1365 				goto aborted_fs;
1366 
1367 		}
1368 		spin_unlock(&space_info->lock);
1369 	}
1370 	return;
1371 
1372 aborted_fs:
1373 	maybe_fail_all_tickets(fs_info, space_info);
1374 	space_info->flush = 0;
1375 	spin_unlock(&space_info->lock);
1376 }
1377 
btrfs_init_async_reclaim_work(struct btrfs_fs_info * fs_info)1378 void btrfs_init_async_reclaim_work(struct btrfs_fs_info *fs_info)
1379 {
1380 	INIT_WORK(&fs_info->async_reclaim_work, btrfs_async_reclaim_metadata_space);
1381 	INIT_WORK(&fs_info->async_data_reclaim_work, btrfs_async_reclaim_data_space);
1382 	INIT_WORK(&fs_info->preempt_reclaim_work,
1383 		  btrfs_preempt_reclaim_metadata_space);
1384 }
1385 
1386 static const enum btrfs_flush_state priority_flush_states[] = {
1387 	FLUSH_DELAYED_ITEMS_NR,
1388 	FLUSH_DELAYED_ITEMS,
1389 	ALLOC_CHUNK,
1390 };
1391 
1392 static const enum btrfs_flush_state evict_flush_states[] = {
1393 	FLUSH_DELAYED_ITEMS_NR,
1394 	FLUSH_DELAYED_ITEMS,
1395 	FLUSH_DELAYED_REFS_NR,
1396 	FLUSH_DELAYED_REFS,
1397 	FLUSH_DELALLOC,
1398 	FLUSH_DELALLOC_WAIT,
1399 	FLUSH_DELALLOC_FULL,
1400 	ALLOC_CHUNK,
1401 	COMMIT_TRANS,
1402 };
1403 
priority_reclaim_metadata_space(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,struct reserve_ticket * ticket,const enum btrfs_flush_state * states,int states_nr)1404 static void priority_reclaim_metadata_space(struct btrfs_fs_info *fs_info,
1405 				struct btrfs_space_info *space_info,
1406 				struct reserve_ticket *ticket,
1407 				const enum btrfs_flush_state *states,
1408 				int states_nr)
1409 {
1410 	u64 to_reclaim;
1411 	int flush_state = 0;
1412 
1413 	spin_lock(&space_info->lock);
1414 	to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info, space_info);
1415 	/*
1416 	 * This is the priority reclaim path, so to_reclaim could be >0 still
1417 	 * because we may have only satisfied the priority tickets and still
1418 	 * left non priority tickets on the list.  We would then have
1419 	 * to_reclaim but ->bytes == 0.
1420 	 */
1421 	if (ticket->bytes == 0) {
1422 		spin_unlock(&space_info->lock);
1423 		return;
1424 	}
1425 
1426 	while (flush_state < states_nr) {
1427 		spin_unlock(&space_info->lock);
1428 		flush_space(fs_info, space_info, to_reclaim, states[flush_state],
1429 			    false);
1430 		flush_state++;
1431 		spin_lock(&space_info->lock);
1432 		if (ticket->bytes == 0) {
1433 			spin_unlock(&space_info->lock);
1434 			return;
1435 		}
1436 	}
1437 
1438 	/*
1439 	 * Attempt to steal from the global rsv if we can, except if the fs was
1440 	 * turned into error mode due to a transaction abort when flushing space
1441 	 * above, in that case fail with the abort error instead of returning
1442 	 * success to the caller if we can steal from the global rsv - this is
1443 	 * just to have caller fail immeditelly instead of later when trying to
1444 	 * modify the fs, making it easier to debug -ENOSPC problems.
1445 	 */
1446 	if (BTRFS_FS_ERROR(fs_info)) {
1447 		ticket->error = BTRFS_FS_ERROR(fs_info);
1448 		remove_ticket(space_info, ticket);
1449 	} else if (!steal_from_global_rsv(fs_info, space_info, ticket)) {
1450 		ticket->error = -ENOSPC;
1451 		remove_ticket(space_info, ticket);
1452 	}
1453 
1454 	/*
1455 	 * We must run try_granting_tickets here because we could be a large
1456 	 * ticket in front of a smaller ticket that can now be satisfied with
1457 	 * the available space.
1458 	 */
1459 	btrfs_try_granting_tickets(fs_info, space_info);
1460 	spin_unlock(&space_info->lock);
1461 }
1462 
priority_reclaim_data_space(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,struct reserve_ticket * ticket)1463 static void priority_reclaim_data_space(struct btrfs_fs_info *fs_info,
1464 					struct btrfs_space_info *space_info,
1465 					struct reserve_ticket *ticket)
1466 {
1467 	spin_lock(&space_info->lock);
1468 
1469 	/* We could have been granted before we got here. */
1470 	if (ticket->bytes == 0) {
1471 		spin_unlock(&space_info->lock);
1472 		return;
1473 	}
1474 
1475 	while (!space_info->full) {
1476 		spin_unlock(&space_info->lock);
1477 		flush_space(fs_info, space_info, U64_MAX, ALLOC_CHUNK_FORCE, false);
1478 		spin_lock(&space_info->lock);
1479 		if (ticket->bytes == 0) {
1480 			spin_unlock(&space_info->lock);
1481 			return;
1482 		}
1483 	}
1484 
1485 	ticket->error = -ENOSPC;
1486 	remove_ticket(space_info, ticket);
1487 	btrfs_try_granting_tickets(fs_info, space_info);
1488 	spin_unlock(&space_info->lock);
1489 }
1490 
wait_reserve_ticket(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,struct reserve_ticket * ticket)1491 static void wait_reserve_ticket(struct btrfs_fs_info *fs_info,
1492 				struct btrfs_space_info *space_info,
1493 				struct reserve_ticket *ticket)
1494 
1495 {
1496 	DEFINE_WAIT(wait);
1497 	int ret = 0;
1498 
1499 	spin_lock(&space_info->lock);
1500 	while (ticket->bytes > 0 && ticket->error == 0) {
1501 		ret = prepare_to_wait_event(&ticket->wait, &wait, TASK_KILLABLE);
1502 		if (ret) {
1503 			/*
1504 			 * Delete us from the list. After we unlock the space
1505 			 * info, we don't want the async reclaim job to reserve
1506 			 * space for this ticket. If that would happen, then the
1507 			 * ticket's task would not known that space was reserved
1508 			 * despite getting an error, resulting in a space leak
1509 			 * (bytes_may_use counter of our space_info).
1510 			 */
1511 			remove_ticket(space_info, ticket);
1512 			ticket->error = -EINTR;
1513 			break;
1514 		}
1515 		spin_unlock(&space_info->lock);
1516 
1517 		schedule();
1518 
1519 		finish_wait(&ticket->wait, &wait);
1520 		spin_lock(&space_info->lock);
1521 	}
1522 	spin_unlock(&space_info->lock);
1523 }
1524 
1525 /*
1526  * Do the appropriate flushing and waiting for a ticket.
1527  *
1528  * @fs_info:    the filesystem
1529  * @space_info: space info for the reservation
1530  * @ticket:     ticket for the reservation
1531  * @start_ns:   timestamp when the reservation started
1532  * @orig_bytes: amount of bytes originally reserved
1533  * @flush:      how much we can flush
1534  *
1535  * This does the work of figuring out how to flush for the ticket, waiting for
1536  * the reservation, and returning the appropriate error if there is one.
1537  */
handle_reserve_ticket(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,struct reserve_ticket * ticket,u64 start_ns,u64 orig_bytes,enum btrfs_reserve_flush_enum flush)1538 static int handle_reserve_ticket(struct btrfs_fs_info *fs_info,
1539 				 struct btrfs_space_info *space_info,
1540 				 struct reserve_ticket *ticket,
1541 				 u64 start_ns, u64 orig_bytes,
1542 				 enum btrfs_reserve_flush_enum flush)
1543 {
1544 	int ret;
1545 
1546 	switch (flush) {
1547 	case BTRFS_RESERVE_FLUSH_DATA:
1548 	case BTRFS_RESERVE_FLUSH_ALL:
1549 	case BTRFS_RESERVE_FLUSH_ALL_STEAL:
1550 		wait_reserve_ticket(fs_info, space_info, ticket);
1551 		break;
1552 	case BTRFS_RESERVE_FLUSH_LIMIT:
1553 		priority_reclaim_metadata_space(fs_info, space_info, ticket,
1554 						priority_flush_states,
1555 						ARRAY_SIZE(priority_flush_states));
1556 		break;
1557 	case BTRFS_RESERVE_FLUSH_EVICT:
1558 		priority_reclaim_metadata_space(fs_info, space_info, ticket,
1559 						evict_flush_states,
1560 						ARRAY_SIZE(evict_flush_states));
1561 		break;
1562 	case BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE:
1563 		priority_reclaim_data_space(fs_info, space_info, ticket);
1564 		break;
1565 	default:
1566 		ASSERT(0);
1567 		break;
1568 	}
1569 
1570 	ret = ticket->error;
1571 	ASSERT(list_empty(&ticket->list));
1572 	/*
1573 	 * Check that we can't have an error set if the reservation succeeded,
1574 	 * as that would confuse tasks and lead them to error out without
1575 	 * releasing reserved space (if an error happens the expectation is that
1576 	 * space wasn't reserved at all).
1577 	 */
1578 	ASSERT(!(ticket->bytes == 0 && ticket->error));
1579 	trace_btrfs_reserve_ticket(fs_info, space_info->flags, orig_bytes,
1580 				   start_ns, flush, ticket->error);
1581 	return ret;
1582 }
1583 
1584 /*
1585  * This returns true if this flush state will go through the ordinary flushing
1586  * code.
1587  */
is_normal_flushing(enum btrfs_reserve_flush_enum flush)1588 static inline bool is_normal_flushing(enum btrfs_reserve_flush_enum flush)
1589 {
1590 	return	(flush == BTRFS_RESERVE_FLUSH_ALL) ||
1591 		(flush == BTRFS_RESERVE_FLUSH_ALL_STEAL);
1592 }
1593 
maybe_clamp_preempt(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info)1594 static inline void maybe_clamp_preempt(struct btrfs_fs_info *fs_info,
1595 				       struct btrfs_space_info *space_info)
1596 {
1597 	u64 ordered = percpu_counter_sum_positive(&fs_info->ordered_bytes);
1598 	u64 delalloc = percpu_counter_sum_positive(&fs_info->delalloc_bytes);
1599 
1600 	/*
1601 	 * If we're heavy on ordered operations then clamping won't help us.  We
1602 	 * need to clamp specifically to keep up with dirty'ing buffered
1603 	 * writers, because there's not a 1:1 correlation of writing delalloc
1604 	 * and freeing space, like there is with flushing delayed refs or
1605 	 * delayed nodes.  If we're already more ordered than delalloc then
1606 	 * we're keeping up, otherwise we aren't and should probably clamp.
1607 	 */
1608 	if (ordered < delalloc)
1609 		space_info->clamp = min(space_info->clamp + 1, 8);
1610 }
1611 
can_steal(enum btrfs_reserve_flush_enum flush)1612 static inline bool can_steal(enum btrfs_reserve_flush_enum flush)
1613 {
1614 	return (flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1615 		flush == BTRFS_RESERVE_FLUSH_EVICT);
1616 }
1617 
1618 /*
1619  * NO_FLUSH and FLUSH_EMERGENCY don't want to create a ticket, they just want to
1620  * fail as quickly as possible.
1621  */
can_ticket(enum btrfs_reserve_flush_enum flush)1622 static inline bool can_ticket(enum btrfs_reserve_flush_enum flush)
1623 {
1624 	return (flush != BTRFS_RESERVE_NO_FLUSH &&
1625 		flush != BTRFS_RESERVE_FLUSH_EMERGENCY);
1626 }
1627 
1628 /*
1629  * Try to reserve bytes from the block_rsv's space.
1630  *
1631  * @fs_info:    the filesystem
1632  * @space_info: space info we want to allocate from
1633  * @orig_bytes: number of bytes we want
1634  * @flush:      whether or not we can flush to make our reservation
1635  *
1636  * This will reserve orig_bytes number of bytes from the space info associated
1637  * with the block_rsv.  If there is not enough space it will make an attempt to
1638  * flush out space to make room.  It will do this by flushing delalloc if
1639  * possible or committing the transaction.  If flush is 0 then no attempts to
1640  * regain reservations will be made and this will fail if there is not enough
1641  * space already.
1642  */
__reserve_bytes(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,u64 orig_bytes,enum btrfs_reserve_flush_enum flush)1643 static int __reserve_bytes(struct btrfs_fs_info *fs_info,
1644 			   struct btrfs_space_info *space_info, u64 orig_bytes,
1645 			   enum btrfs_reserve_flush_enum flush)
1646 {
1647 	struct work_struct *async_work;
1648 	struct reserve_ticket ticket;
1649 	u64 start_ns = 0;
1650 	u64 used;
1651 	int ret = -ENOSPC;
1652 	bool pending_tickets;
1653 
1654 	ASSERT(orig_bytes);
1655 	/*
1656 	 * If have a transaction handle (current->journal_info != NULL), then
1657 	 * the flush method can not be neither BTRFS_RESERVE_FLUSH_ALL* nor
1658 	 * BTRFS_RESERVE_FLUSH_EVICT, as we could deadlock because those
1659 	 * flushing methods can trigger transaction commits.
1660 	 */
1661 	if (current->journal_info) {
1662 		/* One assert per line for easier debugging. */
1663 		ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL);
1664 		ASSERT(flush != BTRFS_RESERVE_FLUSH_ALL_STEAL);
1665 		ASSERT(flush != BTRFS_RESERVE_FLUSH_EVICT);
1666 	}
1667 
1668 	if (flush == BTRFS_RESERVE_FLUSH_DATA)
1669 		async_work = &fs_info->async_data_reclaim_work;
1670 	else
1671 		async_work = &fs_info->async_reclaim_work;
1672 
1673 	spin_lock(&space_info->lock);
1674 	used = btrfs_space_info_used(space_info, true);
1675 
1676 	/*
1677 	 * We don't want NO_FLUSH allocations to jump everybody, they can
1678 	 * generally handle ENOSPC in a different way, so treat them the same as
1679 	 * normal flushers when it comes to skipping pending tickets.
1680 	 */
1681 	if (is_normal_flushing(flush) || (flush == BTRFS_RESERVE_NO_FLUSH))
1682 		pending_tickets = !list_empty(&space_info->tickets) ||
1683 			!list_empty(&space_info->priority_tickets);
1684 	else
1685 		pending_tickets = !list_empty(&space_info->priority_tickets);
1686 
1687 	/*
1688 	 * Carry on if we have enough space (short-circuit) OR call
1689 	 * can_overcommit() to ensure we can overcommit to continue.
1690 	 */
1691 	if (!pending_tickets &&
1692 	    ((used + orig_bytes <= space_info->total_bytes) ||
1693 	     btrfs_can_overcommit(fs_info, space_info, orig_bytes, flush))) {
1694 		btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1695 						      orig_bytes);
1696 		ret = 0;
1697 	}
1698 
1699 	/*
1700 	 * Things are dire, we need to make a reservation so we don't abort.  We
1701 	 * will let this reservation go through as long as we have actual space
1702 	 * left to allocate for the block.
1703 	 */
1704 	if (ret && unlikely(flush == BTRFS_RESERVE_FLUSH_EMERGENCY)) {
1705 		used = btrfs_space_info_used(space_info, false);
1706 		if (used + orig_bytes <= space_info->total_bytes) {
1707 			btrfs_space_info_update_bytes_may_use(fs_info, space_info,
1708 							      orig_bytes);
1709 			ret = 0;
1710 		}
1711 	}
1712 
1713 	/*
1714 	 * If we couldn't make a reservation then setup our reservation ticket
1715 	 * and kick the async worker if it's not already running.
1716 	 *
1717 	 * If we are a priority flusher then we just need to add our ticket to
1718 	 * the list and we will do our own flushing further down.
1719 	 */
1720 	if (ret && can_ticket(flush)) {
1721 		ticket.bytes = orig_bytes;
1722 		ticket.error = 0;
1723 		space_info->reclaim_size += ticket.bytes;
1724 		init_waitqueue_head(&ticket.wait);
1725 		ticket.steal = can_steal(flush);
1726 		if (trace_btrfs_reserve_ticket_enabled())
1727 			start_ns = ktime_get_ns();
1728 
1729 		if (flush == BTRFS_RESERVE_FLUSH_ALL ||
1730 		    flush == BTRFS_RESERVE_FLUSH_ALL_STEAL ||
1731 		    flush == BTRFS_RESERVE_FLUSH_DATA) {
1732 			list_add_tail(&ticket.list, &space_info->tickets);
1733 			if (!space_info->flush) {
1734 				/*
1735 				 * We were forced to add a reserve ticket, so
1736 				 * our preemptive flushing is unable to keep
1737 				 * up.  Clamp down on the threshold for the
1738 				 * preemptive flushing in order to keep up with
1739 				 * the workload.
1740 				 */
1741 				maybe_clamp_preempt(fs_info, space_info);
1742 
1743 				space_info->flush = 1;
1744 				trace_btrfs_trigger_flush(fs_info,
1745 							  space_info->flags,
1746 							  orig_bytes, flush,
1747 							  "enospc");
1748 				queue_work(system_unbound_wq, async_work);
1749 			}
1750 		} else {
1751 			list_add_tail(&ticket.list,
1752 				      &space_info->priority_tickets);
1753 		}
1754 	} else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
1755 		/*
1756 		 * We will do the space reservation dance during log replay,
1757 		 * which means we won't have fs_info->fs_root set, so don't do
1758 		 * the async reclaim as we will panic.
1759 		 */
1760 		if (!test_bit(BTRFS_FS_LOG_RECOVERING, &fs_info->flags) &&
1761 		    !work_busy(&fs_info->preempt_reclaim_work) &&
1762 		    need_preemptive_reclaim(fs_info, space_info)) {
1763 			trace_btrfs_trigger_flush(fs_info, space_info->flags,
1764 						  orig_bytes, flush, "preempt");
1765 			queue_work(system_unbound_wq,
1766 				   &fs_info->preempt_reclaim_work);
1767 		}
1768 	}
1769 	spin_unlock(&space_info->lock);
1770 	if (!ret || !can_ticket(flush))
1771 		return ret;
1772 
1773 	return handle_reserve_ticket(fs_info, space_info, &ticket, start_ns,
1774 				     orig_bytes, flush);
1775 }
1776 
1777 /*
1778  * Try to reserve metadata bytes from the block_rsv's space.
1779  *
1780  * @fs_info:    the filesystem
1781  * @space_info: the space_info we're allocating for
1782  * @orig_bytes: number of bytes we want
1783  * @flush:      whether or not we can flush to make our reservation
1784  *
1785  * This will reserve orig_bytes number of bytes from the space info associated
1786  * with the block_rsv.  If there is not enough space it will make an attempt to
1787  * flush out space to make room.  It will do this by flushing delalloc if
1788  * possible or committing the transaction.  If flush is 0 then no attempts to
1789  * regain reservations will be made and this will fail if there is not enough
1790  * space already.
1791  */
btrfs_reserve_metadata_bytes(struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,u64 orig_bytes,enum btrfs_reserve_flush_enum flush)1792 int btrfs_reserve_metadata_bytes(struct btrfs_fs_info *fs_info,
1793 				 struct btrfs_space_info *space_info,
1794 				 u64 orig_bytes,
1795 				 enum btrfs_reserve_flush_enum flush)
1796 {
1797 	int ret;
1798 
1799 	ret = __reserve_bytes(fs_info, space_info, orig_bytes, flush);
1800 	if (ret == -ENOSPC) {
1801 		trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1802 					      space_info->flags, orig_bytes, 1);
1803 
1804 		if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1805 			btrfs_dump_space_info(fs_info, space_info, orig_bytes, 0);
1806 	}
1807 	return ret;
1808 }
1809 
1810 /*
1811  * Try to reserve data bytes for an allocation.
1812  *
1813  * @fs_info: the filesystem
1814  * @bytes:   number of bytes we need
1815  * @flush:   how we are allowed to flush
1816  *
1817  * This will reserve bytes from the data space info.  If there is not enough
1818  * space then we will attempt to flush space as specified by flush.
1819  */
btrfs_reserve_data_bytes(struct btrfs_fs_info * fs_info,u64 bytes,enum btrfs_reserve_flush_enum flush)1820 int btrfs_reserve_data_bytes(struct btrfs_fs_info *fs_info, u64 bytes,
1821 			     enum btrfs_reserve_flush_enum flush)
1822 {
1823 	struct btrfs_space_info *data_sinfo = fs_info->data_sinfo;
1824 	int ret;
1825 
1826 	ASSERT(flush == BTRFS_RESERVE_FLUSH_DATA ||
1827 	       flush == BTRFS_RESERVE_FLUSH_FREE_SPACE_INODE ||
1828 	       flush == BTRFS_RESERVE_NO_FLUSH);
1829 	ASSERT(!current->journal_info || flush != BTRFS_RESERVE_FLUSH_DATA);
1830 
1831 	ret = __reserve_bytes(fs_info, data_sinfo, bytes, flush);
1832 	if (ret == -ENOSPC) {
1833 		trace_btrfs_space_reservation(fs_info, "space_info:enospc",
1834 					      data_sinfo->flags, bytes, 1);
1835 		if (btrfs_test_opt(fs_info, ENOSPC_DEBUG))
1836 			btrfs_dump_space_info(fs_info, data_sinfo, bytes, 0);
1837 	}
1838 	return ret;
1839 }
1840 
1841 /* Dump all the space infos when we abort a transaction due to ENOSPC. */
btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info * fs_info)1842 __cold void btrfs_dump_space_info_for_trans_abort(struct btrfs_fs_info *fs_info)
1843 {
1844 	struct btrfs_space_info *space_info;
1845 
1846 	btrfs_info(fs_info, "dumping space info:");
1847 	list_for_each_entry(space_info, &fs_info->space_info, list) {
1848 		spin_lock(&space_info->lock);
1849 		__btrfs_dump_space_info(fs_info, space_info);
1850 		spin_unlock(&space_info->lock);
1851 	}
1852 	dump_global_block_rsv(fs_info);
1853 }
1854 
1855 /*
1856  * Account the unused space of all the readonly block group in the space_info.
1857  * takes mirrors into account.
1858  */
btrfs_account_ro_block_groups_free_space(struct btrfs_space_info * sinfo)1859 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
1860 {
1861 	struct btrfs_block_group *block_group;
1862 	u64 free_bytes = 0;
1863 	int factor;
1864 
1865 	/* It's df, we don't care if it's racy */
1866 	if (list_empty(&sinfo->ro_bgs))
1867 		return 0;
1868 
1869 	spin_lock(&sinfo->lock);
1870 	list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
1871 		spin_lock(&block_group->lock);
1872 
1873 		if (!block_group->ro) {
1874 			spin_unlock(&block_group->lock);
1875 			continue;
1876 		}
1877 
1878 		factor = btrfs_bg_type_to_factor(block_group->flags);
1879 		free_bytes += (block_group->length -
1880 			       block_group->used) * factor;
1881 
1882 		spin_unlock(&block_group->lock);
1883 	}
1884 	spin_unlock(&sinfo->lock);
1885 
1886 	return free_bytes;
1887 }
1888 
calc_pct_ratio(u64 x,u64 y)1889 static u64 calc_pct_ratio(u64 x, u64 y)
1890 {
1891 	int err;
1892 
1893 	if (!y)
1894 		return 0;
1895 again:
1896 	err = check_mul_overflow(100, x, &x);
1897 	if (err)
1898 		goto lose_precision;
1899 	return div64_u64(x, y);
1900 lose_precision:
1901 	x >>= 10;
1902 	y >>= 10;
1903 	if (!y)
1904 		y = 1;
1905 	goto again;
1906 }
1907 
1908 /*
1909  * A reasonable buffer for unallocated space is 10 data block_groups.
1910  * If we claw this back repeatedly, we can still achieve efficient
1911  * utilization when near full, and not do too much reclaim while
1912  * always maintaining a solid buffer for workloads that quickly
1913  * allocate and pressure the unallocated space.
1914  */
calc_unalloc_target(struct btrfs_fs_info * fs_info)1915 static u64 calc_unalloc_target(struct btrfs_fs_info *fs_info)
1916 {
1917 	u64 chunk_sz = calc_effective_data_chunk_size(fs_info);
1918 
1919 	return BTRFS_UNALLOC_BLOCK_GROUP_TARGET * chunk_sz;
1920 }
1921 
1922 /*
1923  * The fundamental goal of automatic reclaim is to protect the filesystem's
1924  * unallocated space and thus minimize the probability of the filesystem going
1925  * read only when a metadata allocation failure causes a transaction abort.
1926  *
1927  * However, relocations happen into the space_info's unused space, therefore
1928  * automatic reclaim must also back off as that space runs low. There is no
1929  * value in doing trivial "relocations" of re-writing the same block group
1930  * into a fresh one.
1931  *
1932  * Furthermore, we want to avoid doing too much reclaim even if there are good
1933  * candidates. This is because the allocator is pretty good at filling up the
1934  * holes with writes. So we want to do just enough reclaim to try and stay
1935  * safe from running out of unallocated space but not be wasteful about it.
1936  *
1937  * Therefore, the dynamic reclaim threshold is calculated as follows:
1938  * - calculate a target unallocated amount of 5 block group sized chunks
1939  * - ratchet up the intensity of reclaim depending on how far we are from
1940  *   that target by using a formula of unalloc / target to set the threshold.
1941  *
1942  * Typically with 10 block groups as the target, the discrete values this comes
1943  * out to are 0, 10, 20, ... , 80, 90, and 99.
1944  */
calc_dynamic_reclaim_threshold(const struct btrfs_space_info * space_info)1945 static int calc_dynamic_reclaim_threshold(const struct btrfs_space_info *space_info)
1946 {
1947 	struct btrfs_fs_info *fs_info = space_info->fs_info;
1948 	u64 unalloc = atomic64_read(&fs_info->free_chunk_space);
1949 	u64 target = calc_unalloc_target(fs_info);
1950 	u64 alloc = space_info->total_bytes;
1951 	u64 used = btrfs_space_info_used(space_info, false);
1952 	u64 unused = alloc - used;
1953 	u64 want = target > unalloc ? target - unalloc : 0;
1954 	u64 data_chunk_size = calc_effective_data_chunk_size(fs_info);
1955 
1956 	/* If we have no unused space, don't bother, it won't work anyway. */
1957 	if (unused < data_chunk_size)
1958 		return 0;
1959 
1960 	/* Cast to int is OK because want <= target. */
1961 	return calc_pct_ratio(want, target);
1962 }
1963 
btrfs_calc_reclaim_threshold(const struct btrfs_space_info * space_info)1964 int btrfs_calc_reclaim_threshold(const struct btrfs_space_info *space_info)
1965 {
1966 	lockdep_assert_held(&space_info->lock);
1967 
1968 	if (READ_ONCE(space_info->dynamic_reclaim))
1969 		return calc_dynamic_reclaim_threshold(space_info);
1970 	return READ_ONCE(space_info->bg_reclaim_threshold);
1971 }
1972 
1973 /*
1974  * Under "urgent" reclaim, we will reclaim even fresh block groups that have
1975  * recently seen successful allocations, as we are desperate to reclaim
1976  * whatever we can to avoid ENOSPC in a transaction leading to a readonly fs.
1977  */
is_reclaim_urgent(struct btrfs_space_info * space_info)1978 static bool is_reclaim_urgent(struct btrfs_space_info *space_info)
1979 {
1980 	struct btrfs_fs_info *fs_info = space_info->fs_info;
1981 	u64 unalloc = atomic64_read(&fs_info->free_chunk_space);
1982 	u64 data_chunk_size = calc_effective_data_chunk_size(fs_info);
1983 
1984 	return unalloc < data_chunk_size;
1985 }
1986 
do_reclaim_sweep(const struct btrfs_fs_info * fs_info,struct btrfs_space_info * space_info,int raid)1987 static void do_reclaim_sweep(const struct btrfs_fs_info *fs_info,
1988 			     struct btrfs_space_info *space_info, int raid)
1989 {
1990 	struct btrfs_block_group *bg;
1991 	int thresh_pct;
1992 	bool try_again = true;
1993 	bool urgent;
1994 
1995 	spin_lock(&space_info->lock);
1996 	urgent = is_reclaim_urgent(space_info);
1997 	thresh_pct = btrfs_calc_reclaim_threshold(space_info);
1998 	spin_unlock(&space_info->lock);
1999 
2000 	down_read(&space_info->groups_sem);
2001 again:
2002 	list_for_each_entry(bg, &space_info->block_groups[raid], list) {
2003 		u64 thresh;
2004 		bool reclaim = false;
2005 
2006 		btrfs_get_block_group(bg);
2007 		spin_lock(&bg->lock);
2008 		thresh = mult_perc(bg->length, thresh_pct);
2009 		if (bg->used < thresh && bg->reclaim_mark) {
2010 			try_again = false;
2011 			reclaim = true;
2012 		}
2013 		bg->reclaim_mark++;
2014 		spin_unlock(&bg->lock);
2015 		if (reclaim)
2016 			btrfs_mark_bg_to_reclaim(bg);
2017 		btrfs_put_block_group(bg);
2018 	}
2019 
2020 	/*
2021 	 * In situations where we are very motivated to reclaim (low unalloc)
2022 	 * use two passes to make the reclaim mark check best effort.
2023 	 *
2024 	 * If we have any staler groups, we don't touch the fresher ones, but if we
2025 	 * really need a block group, do take a fresh one.
2026 	 */
2027 	if (try_again && urgent) {
2028 		try_again = false;
2029 		goto again;
2030 	}
2031 
2032 	up_read(&space_info->groups_sem);
2033 }
2034 
btrfs_space_info_update_reclaimable(struct btrfs_space_info * space_info,s64 bytes)2035 void btrfs_space_info_update_reclaimable(struct btrfs_space_info *space_info, s64 bytes)
2036 {
2037 	u64 chunk_sz = calc_effective_data_chunk_size(space_info->fs_info);
2038 
2039 	lockdep_assert_held(&space_info->lock);
2040 	space_info->reclaimable_bytes += bytes;
2041 
2042 	if (space_info->reclaimable_bytes >= chunk_sz)
2043 		btrfs_set_periodic_reclaim_ready(space_info, true);
2044 }
2045 
btrfs_set_periodic_reclaim_ready(struct btrfs_space_info * space_info,bool ready)2046 void btrfs_set_periodic_reclaim_ready(struct btrfs_space_info *space_info, bool ready)
2047 {
2048 	lockdep_assert_held(&space_info->lock);
2049 	if (!READ_ONCE(space_info->periodic_reclaim))
2050 		return;
2051 	if (ready != space_info->periodic_reclaim_ready) {
2052 		space_info->periodic_reclaim_ready = ready;
2053 		if (!ready)
2054 			space_info->reclaimable_bytes = 0;
2055 	}
2056 }
2057 
btrfs_should_periodic_reclaim(struct btrfs_space_info * space_info)2058 bool btrfs_should_periodic_reclaim(struct btrfs_space_info *space_info)
2059 {
2060 	bool ret;
2061 
2062 	if (space_info->flags & BTRFS_BLOCK_GROUP_SYSTEM)
2063 		return false;
2064 	if (!READ_ONCE(space_info->periodic_reclaim))
2065 		return false;
2066 
2067 	spin_lock(&space_info->lock);
2068 	ret = space_info->periodic_reclaim_ready;
2069 	btrfs_set_periodic_reclaim_ready(space_info, false);
2070 	spin_unlock(&space_info->lock);
2071 
2072 	return ret;
2073 }
2074 
btrfs_reclaim_sweep(const struct btrfs_fs_info * fs_info)2075 void btrfs_reclaim_sweep(const struct btrfs_fs_info *fs_info)
2076 {
2077 	int raid;
2078 	struct btrfs_space_info *space_info;
2079 
2080 	list_for_each_entry(space_info, &fs_info->space_info, list) {
2081 		if (!btrfs_should_periodic_reclaim(space_info))
2082 			continue;
2083 		for (raid = 0; raid < BTRFS_NR_RAID_TYPES; raid++)
2084 			do_reclaim_sweep(fs_info, space_info, raid);
2085 	}
2086 }
2087