1 // SPDX-License-Identifier: GPL-2.0
3 * Primary bucket allocation code
5 * Copyright 2012 Google, Inc.
7 * Allocation in bcache is done in terms of buckets:
9 * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
10 * btree pointers - they must match for the pointer to be considered valid.
12 * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
13 * bucket simply by incrementing its gen.
15 * The gens (along with the priorities; it's really the gens are important but
16 * the code is named as if it's the priorities) are written in an arbitrary list
17 * of buckets on disk, with a pointer to them in the journal header.
19 * When we invalidate a bucket, we have to write its new gen to disk and wait
20 * for that write to complete before we use it - otherwise after a crash we
21 * could have pointers that appeared to be good but pointed to data that had
24 * Since the gens and priorities are all stored contiguously on disk, we can
25 * batch this up: We fill up the free_inc list with freshly invalidated buckets,
26 * call prio_write(), and when prio_write() finishes we pull buckets off the
27 * free_inc list and optionally discard them.
29 * free_inc isn't the only freelist - if it was, we'd often to sleep while
30 * priorities and gens were being written before we could allocate. c->free is a
31 * smaller freelist, and buckets on that list are always ready to be used.
33 * If we've got discards enabled, that happens when a bucket moves from the
34 * free_inc list to the free list.
36 * There is another freelist, because sometimes we have buckets that we know
37 * have nothing pointing into them - these we can reuse without waiting for
38 * priorities to be rewritten. These come from freed btree nodes and buckets
39 * that garbage collection discovered no longer had valid keys pointing into
40 * them (because they were overwritten). That's the unused list - buckets on the
41 * unused list move to the free list, optionally being discarded in the process.
43 * It's also important to ensure that gens don't wrap around - with respect to
44 * either the oldest gen in the btree or the gen on disk. This is quite
45 * difficult to do in practice, but we explicitly guard against it anyways - if
46 * a bucket is in danger of wrapping around we simply skip invalidating it that
47 * time around, and we garbage collect or rewrite the priorities sooner than we
48 * would have otherwise.
50 * bch_bucket_alloc() allocates a single bucket from a specific cache.
52 * bch_bucket_alloc_set() allocates one or more buckets from different caches
55 * free_some_buckets() drives all the processes described above. It's called
56 * from bch_bucket_alloc() and a few other places that need to make sure free
59 * invalidate_buckets_(lru|fifo)() find buckets that are available to be
60 * invalidated, and then invalidate them and stick them on the free_inc list -
61 * in either lru or fifo order.
67 #include <linux/blkdev.h>
68 #include <linux/kthread.h>
69 #include <linux/random.h>
70 #include <trace/events/bcache.h>
72 #define MAX_OPEN_BUCKETS 128
74 /* Bucket heap / gen */
76 uint8_t bch_inc_gen(struct cache *ca, struct bucket *b)
78 uint8_t ret = ++b->gen;
80 ca->set->need_gc = max(ca->set->need_gc, bucket_gc_gen(b));
81 WARN_ON_ONCE(ca->set->need_gc > BUCKET_GC_GEN_MAX);
86 void bch_rescale_priorities(struct cache_set *c, int sectors)
90 unsigned int next = c->nbuckets * c->sb.bucket_size / 1024;
94 atomic_sub(sectors, &c->rescale);
97 r = atomic_read(&c->rescale);
101 } while (atomic_cmpxchg(&c->rescale, r, r + next) != r);
103 mutex_lock(&c->bucket_lock);
105 c->min_prio = USHRT_MAX;
107 for_each_cache(ca, c, i)
108 for_each_bucket(b, ca)
110 b->prio != BTREE_PRIO &&
111 !atomic_read(&b->pin)) {
113 c->min_prio = min(c->min_prio, b->prio);
116 mutex_unlock(&c->bucket_lock);
120 * Background allocation thread: scans for buckets to be invalidated,
121 * invalidates them, rewrites prios/gens (marking them as invalidated on disk),
122 * then optionally issues discard commands to the newly free buckets, then puts
123 * them on the various freelists.
126 static inline bool can_inc_bucket_gen(struct bucket *b)
128 return bucket_gc_gen(b) < BUCKET_GC_GEN_MAX;
131 bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b)
133 BUG_ON(!ca->set->gc_mark_valid);
135 return (!GC_MARK(b) ||
136 GC_MARK(b) == GC_MARK_RECLAIMABLE) &&
137 !atomic_read(&b->pin) &&
138 can_inc_bucket_gen(b);
141 void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
143 lockdep_assert_held(&ca->set->bucket_lock);
144 BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE);
146 if (GC_SECTORS_USED(b))
147 trace_bcache_invalidate(ca, b - ca->buckets);
150 b->prio = INITIAL_PRIO;
154 static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
156 __bch_invalidate_one_bucket(ca, b);
158 fifo_push(&ca->free_inc, b - ca->buckets);
162 * Determines what order we're going to reuse buckets, smallest bucket_prio()
163 * first: we also take into account the number of sectors of live data in that
164 * bucket, and in order for that multiply to make sense we have to scale bucket
166 * Thus, we scale the bucket priorities so that the bucket with the smallest
167 * prio is worth 1/8th of what INITIAL_PRIO is worth.
170 #define bucket_prio(b) \
172 unsigned int min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8; \
174 (b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b); \
177 #define bucket_max_cmp(l, r) (bucket_prio(l) < bucket_prio(r))
178 #define bucket_min_cmp(l, r) (bucket_prio(l) > bucket_prio(r))
180 static void invalidate_buckets_lru(struct cache *ca)
187 for_each_bucket(b, ca) {
188 if (!bch_can_invalidate_bucket(ca, b))
191 if (!heap_full(&ca->heap))
192 heap_add(&ca->heap, b, bucket_max_cmp);
193 else if (bucket_max_cmp(b, heap_peek(&ca->heap))) {
194 ca->heap.data[0] = b;
195 heap_sift(&ca->heap, 0, bucket_max_cmp);
199 for (i = ca->heap.used / 2 - 1; i >= 0; --i)
200 heap_sift(&ca->heap, i, bucket_min_cmp);
202 while (!fifo_full(&ca->free_inc)) {
203 if (!heap_pop(&ca->heap, b, bucket_min_cmp)) {
205 * We don't want to be calling invalidate_buckets()
206 * multiple times when it can't do anything
208 ca->invalidate_needs_gc = 1;
213 bch_invalidate_one_bucket(ca, b);
217 static void invalidate_buckets_fifo(struct cache *ca)
222 while (!fifo_full(&ca->free_inc)) {
223 if (ca->fifo_last_bucket < ca->sb.first_bucket ||
224 ca->fifo_last_bucket >= ca->sb.nbuckets)
225 ca->fifo_last_bucket = ca->sb.first_bucket;
227 b = ca->buckets + ca->fifo_last_bucket++;
229 if (bch_can_invalidate_bucket(ca, b))
230 bch_invalidate_one_bucket(ca, b);
232 if (++checked >= ca->sb.nbuckets) {
233 ca->invalidate_needs_gc = 1;
240 static void invalidate_buckets_random(struct cache *ca)
245 while (!fifo_full(&ca->free_inc)) {
248 get_random_bytes(&n, sizeof(n));
250 n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket);
251 n += ca->sb.first_bucket;
255 if (bch_can_invalidate_bucket(ca, b))
256 bch_invalidate_one_bucket(ca, b);
258 if (++checked >= ca->sb.nbuckets / 2) {
259 ca->invalidate_needs_gc = 1;
266 static void invalidate_buckets(struct cache *ca)
268 BUG_ON(ca->invalidate_needs_gc);
270 switch (CACHE_REPLACEMENT(&ca->sb)) {
271 case CACHE_REPLACEMENT_LRU:
272 invalidate_buckets_lru(ca);
274 case CACHE_REPLACEMENT_FIFO:
275 invalidate_buckets_fifo(ca);
277 case CACHE_REPLACEMENT_RANDOM:
278 invalidate_buckets_random(ca);
283 #define allocator_wait(ca, cond) \
286 set_current_state(TASK_INTERRUPTIBLE); \
290 mutex_unlock(&(ca)->set->bucket_lock); \
291 if (kthread_should_stop() || \
292 test_bit(CACHE_SET_IO_DISABLE, &ca->set->flags)) { \
293 set_current_state(TASK_RUNNING); \
298 mutex_lock(&(ca)->set->bucket_lock); \
300 __set_current_state(TASK_RUNNING); \
303 static int bch_allocator_push(struct cache *ca, long bucket)
307 /* Prios/gens are actually the most important reserve */
308 if (fifo_push(&ca->free[RESERVE_PRIO], bucket))
311 for (i = 0; i < RESERVE_NR; i++)
312 if (fifo_push(&ca->free[i], bucket))
318 static int bch_allocator_thread(void *arg)
320 struct cache *ca = arg;
322 mutex_lock(&ca->set->bucket_lock);
326 * First, we pull buckets off of the unused and free_inc lists,
327 * possibly issue discards to them, then we add the bucket to
333 if (!fifo_pop(&ca->free_inc, bucket))
337 mutex_unlock(&ca->set->bucket_lock);
338 blkdev_issue_discard(ca->bdev,
339 bucket_to_sector(ca->set, bucket),
340 ca->sb.bucket_size, GFP_KERNEL, 0);
341 mutex_lock(&ca->set->bucket_lock);
344 allocator_wait(ca, bch_allocator_push(ca, bucket));
345 wake_up(&ca->set->btree_cache_wait);
346 wake_up(&ca->set->bucket_wait);
350 * We've run out of free buckets, we need to find some buckets
351 * we can invalidate. First, invalidate them in memory and add
352 * them to the free_inc list:
356 allocator_wait(ca, ca->set->gc_mark_valid &&
357 !ca->invalidate_needs_gc);
358 invalidate_buckets(ca);
361 * Now, we write their new gens to disk so we can start writing
364 allocator_wait(ca, !atomic_read(&ca->set->prio_blocked));
365 if (CACHE_SYNC(&ca->set->sb)) {
367 * This could deadlock if an allocation with a btree
368 * node locked ever blocked - having the btree node
369 * locked would block garbage collection, but here we're
370 * waiting on garbage collection before we invalidate
373 * But this should be safe since the btree code always
374 * uses btree_check_reserve() before allocating now, and
375 * if it fails it blocks without btree nodes locked.
377 if (!fifo_full(&ca->free_inc))
378 goto retry_invalidate;
380 if (bch_prio_write(ca, false) < 0) {
381 ca->invalidate_needs_gc = 1;
387 wait_for_kthread_stop();
393 long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait)
400 /* No allocation if CACHE_SET_IO_DISABLE bit is set */
401 if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &ca->set->flags)))
405 if (fifo_pop(&ca->free[RESERVE_NONE], r) ||
406 fifo_pop(&ca->free[reserve], r))
410 trace_bcache_alloc_fail(ca, reserve);
415 prepare_to_wait(&ca->set->bucket_wait, &w,
416 TASK_UNINTERRUPTIBLE);
418 mutex_unlock(&ca->set->bucket_lock);
420 mutex_lock(&ca->set->bucket_lock);
421 } while (!fifo_pop(&ca->free[RESERVE_NONE], r) &&
422 !fifo_pop(&ca->free[reserve], r));
424 finish_wait(&ca->set->bucket_wait, &w);
426 if (ca->alloc_thread)
427 wake_up_process(ca->alloc_thread);
429 trace_bcache_alloc(ca, reserve);
431 if (expensive_debug_checks(ca->set)) {
436 for (iter = 0; iter < prio_buckets(ca) * 2; iter++)
437 BUG_ON(ca->prio_buckets[iter] == (uint64_t) r);
439 for (j = 0; j < RESERVE_NR; j++)
440 fifo_for_each(i, &ca->free[j], iter)
442 fifo_for_each(i, &ca->free_inc, iter)
448 BUG_ON(atomic_read(&b->pin) != 1);
450 SET_GC_SECTORS_USED(b, ca->sb.bucket_size);
452 if (reserve <= RESERVE_PRIO) {
453 SET_GC_MARK(b, GC_MARK_METADATA);
455 b->prio = BTREE_PRIO;
457 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
459 b->prio = INITIAL_PRIO;
462 if (ca->set->avail_nbuckets > 0) {
463 ca->set->avail_nbuckets--;
464 bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
470 void __bch_bucket_free(struct cache *ca, struct bucket *b)
473 SET_GC_SECTORS_USED(b, 0);
475 if (ca->set->avail_nbuckets < ca->set->nbuckets) {
476 ca->set->avail_nbuckets++;
477 bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
481 void bch_bucket_free(struct cache_set *c, struct bkey *k)
485 for (i = 0; i < KEY_PTRS(k); i++)
486 __bch_bucket_free(PTR_CACHE(c, k, i),
487 PTR_BUCKET(c, k, i));
490 int __bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
491 struct bkey *k, int n, bool wait)
495 /* No allocation if CACHE_SET_IO_DISABLE bit is set */
496 if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
499 lockdep_assert_held(&c->bucket_lock);
500 BUG_ON(!n || n > c->caches_loaded || n > MAX_CACHES_PER_SET);
504 /* sort by free space/prio of oldest data in caches */
506 for (i = 0; i < n; i++) {
507 struct cache *ca = c->cache_by_alloc[i];
508 long b = bch_bucket_alloc(ca, reserve, wait);
513 k->ptr[i] = MAKE_PTR(ca->buckets[b].gen,
514 bucket_to_sector(c, b),
517 SET_KEY_PTRS(k, i + 1);
522 bch_bucket_free(c, k);
527 int bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
528 struct bkey *k, int n, bool wait)
532 mutex_lock(&c->bucket_lock);
533 ret = __bch_bucket_alloc_set(c, reserve, k, n, wait);
534 mutex_unlock(&c->bucket_lock);
538 /* Sector allocator */
541 struct list_head list;
542 unsigned int last_write_point;
543 unsigned int sectors_free;
548 * We keep multiple buckets open for writes, and try to segregate different
549 * write streams for better cache utilization: first we try to segregate flash
550 * only volume write streams from cached devices, secondly we look for a bucket
551 * where the last write to it was sequential with the current write, and
552 * failing that we look for a bucket that was last used by the same task.
554 * The ideas is if you've got multiple tasks pulling data into the cache at the
555 * same time, you'll get better cache utilization if you try to segregate their
556 * data and preserve locality.
558 * For example, dirty sectors of flash only volume is not reclaimable, if their
559 * dirty sectors mixed with dirty sectors of cached device, such buckets will
560 * be marked as dirty and won't be reclaimed, though the dirty data of cached
561 * device have been written back to backend device.
563 * And say you've starting Firefox at the same time you're copying a
564 * bunch of files. Firefox will likely end up being fairly hot and stay in the
565 * cache awhile, but the data you copied might not be; if you wrote all that
566 * data to the same buckets it'd get invalidated at the same time.
568 * Both of those tasks will be doing fairly random IO so we can't rely on
569 * detecting sequential IO to segregate their data, but going off of the task
570 * should be a sane heuristic.
572 static struct open_bucket *pick_data_bucket(struct cache_set *c,
573 const struct bkey *search,
574 unsigned int write_point,
577 struct open_bucket *ret, *ret_task = NULL;
579 list_for_each_entry_reverse(ret, &c->data_buckets, list)
580 if (UUID_FLASH_ONLY(&c->uuids[KEY_INODE(&ret->key)]) !=
581 UUID_FLASH_ONLY(&c->uuids[KEY_INODE(search)]))
583 else if (!bkey_cmp(&ret->key, search))
585 else if (ret->last_write_point == write_point)
588 ret = ret_task ?: list_first_entry(&c->data_buckets,
589 struct open_bucket, list);
591 if (!ret->sectors_free && KEY_PTRS(alloc)) {
592 ret->sectors_free = c->sb.bucket_size;
593 bkey_copy(&ret->key, alloc);
597 if (!ret->sectors_free)
604 * Allocates some space in the cache to write to, and k to point to the newly
605 * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the
606 * end of the newly allocated space).
608 * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many
609 * sectors were actually allocated.
611 * If s->writeback is true, will not fail.
613 bool bch_alloc_sectors(struct cache_set *c,
615 unsigned int sectors,
616 unsigned int write_point,
617 unsigned int write_prio,
620 struct open_bucket *b;
621 BKEY_PADDED(key) alloc;
625 * We might have to allocate a new bucket, which we can't do with a
626 * spinlock held. So if we have to allocate, we drop the lock, allocate
627 * and then retry. KEY_PTRS() indicates whether alloc points to
628 * allocated bucket(s).
631 bkey_init(&alloc.key);
632 spin_lock(&c->data_bucket_lock);
634 while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) {
635 unsigned int watermark = write_prio
639 spin_unlock(&c->data_bucket_lock);
641 if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, wait))
644 spin_lock(&c->data_bucket_lock);
648 * If we had to allocate, we might race and not need to allocate the
649 * second time we call pick_data_bucket(). If we allocated a bucket but
650 * didn't use it, drop the refcount bch_bucket_alloc_set() took:
652 if (KEY_PTRS(&alloc.key))
653 bkey_put(c, &alloc.key);
655 for (i = 0; i < KEY_PTRS(&b->key); i++)
656 EBUG_ON(ptr_stale(c, &b->key, i));
658 /* Set up the pointer to the space we're allocating: */
660 for (i = 0; i < KEY_PTRS(&b->key); i++)
661 k->ptr[i] = b->key.ptr[i];
663 sectors = min(sectors, b->sectors_free);
665 SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors);
666 SET_KEY_SIZE(k, sectors);
667 SET_KEY_PTRS(k, KEY_PTRS(&b->key));
670 * Move b to the end of the lru, and keep track of what this bucket was
673 list_move_tail(&b->list, &c->data_buckets);
674 bkey_copy_key(&b->key, k);
675 b->last_write_point = write_point;
677 b->sectors_free -= sectors;
679 for (i = 0; i < KEY_PTRS(&b->key); i++) {
680 SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors);
682 atomic_long_add(sectors,
683 &PTR_CACHE(c, &b->key, i)->sectors_written);
686 if (b->sectors_free < c->sb.block_size)
690 * k takes refcounts on the buckets it points to until it's inserted
691 * into the btree, but if we're done with this bucket we just transfer
692 * get_data_bucket()'s refcount.
695 for (i = 0; i < KEY_PTRS(&b->key); i++)
696 atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin);
698 spin_unlock(&c->data_bucket_lock);
704 void bch_open_buckets_free(struct cache_set *c)
706 struct open_bucket *b;
708 while (!list_empty(&c->data_buckets)) {
709 b = list_first_entry(&c->data_buckets,
710 struct open_bucket, list);
716 int bch_open_buckets_alloc(struct cache_set *c)
720 spin_lock_init(&c->data_bucket_lock);
722 for (i = 0; i < MAX_OPEN_BUCKETS; i++) {
723 struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL);
728 list_add(&b->list, &c->data_buckets);
734 int bch_cache_allocator_start(struct cache *ca)
736 struct task_struct *k = kthread_run(bch_allocator_thread,
737 ca, "bcache_allocator");
741 ca->alloc_thread = k;