2 * Primary bucket allocation code
4 * Copyright 2012 Google, Inc.
6 * Allocation in bcache is done in terms of buckets:
8 * Each bucket has associated an 8 bit gen; this gen corresponds to the gen in
9 * btree pointers - they must match for the pointer to be considered valid.
11 * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
12 * bucket simply by incrementing its gen.
14 * The gens (along with the priorities; it's really the gens are important but
15 * the code is named as if it's the priorities) are written in an arbitrary list
16 * of buckets on disk, with a pointer to them in the journal header.
18 * When we invalidate a bucket, we have to write its new gen to disk and wait
19 * for that write to complete before we use it - otherwise after a crash we
20 * could have pointers that appeared to be good but pointed to data that had
23 * Since the gens and priorities are all stored contiguously on disk, we can
24 * batch this up: We fill up the free_inc list with freshly invalidated buckets,
25 * call prio_write(), and when prio_write() finishes we pull buckets off the
26 * free_inc list and optionally discard them.
28 * free_inc isn't the only freelist - if it was, we'd often to sleep while
29 * priorities and gens were being written before we could allocate. c->free is a
30 * smaller freelist, and buckets on that list are always ready to be used.
32 * If we've got discards enabled, that happens when a bucket moves from the
33 * free_inc list to the free list.
35 * There is another freelist, because sometimes we have buckets that we know
36 * have nothing pointing into them - these we can reuse without waiting for
37 * priorities to be rewritten. These come from freed btree nodes and buckets
38 * that garbage collection discovered no longer had valid keys pointing into
39 * them (because they were overwritten). That's the unused list - buckets on the
40 * unused list move to the free list, optionally being discarded in the process.
42 * It's also important to ensure that gens don't wrap around - with respect to
43 * either the oldest gen in the btree or the gen on disk. This is quite
44 * difficult to do in practice, but we explicitly guard against it anyways - if
45 * a bucket is in danger of wrapping around we simply skip invalidating it that
46 * time around, and we garbage collect or rewrite the priorities sooner than we
47 * would have otherwise.
49 * bch_bucket_alloc() allocates a single bucket from a specific cache.
51 * bch_bucket_alloc_set() allocates one or more buckets from different caches
54 * free_some_buckets() drives all the processes described above. It's called
55 * from bch_bucket_alloc() and a few other places that need to make sure free
58 * invalidate_buckets_(lru|fifo)() find buckets that are available to be
59 * invalidated, and then invalidate them and stick them on the free_inc list -
60 * in either lru or fifo order.
66 #include <linux/blkdev.h>
67 #include <linux/freezer.h>
68 #include <linux/kthread.h>
69 #include <linux/random.h>
70 #include <trace/events/bcache.h>
72 /* Bucket heap / gen */
74 uint8_t bch_inc_gen(struct cache *ca, struct bucket *b)
76 uint8_t ret = ++b->gen;
78 ca->set->need_gc = max(ca->set->need_gc, bucket_gc_gen(b));
79 WARN_ON_ONCE(ca->set->need_gc > BUCKET_GC_GEN_MAX);
84 void bch_rescale_priorities(struct cache_set *c, int sectors)
88 unsigned next = c->nbuckets * c->sb.bucket_size / 1024;
92 atomic_sub(sectors, &c->rescale);
95 r = atomic_read(&c->rescale);
99 } while (atomic_cmpxchg(&c->rescale, r, r + next) != r);
101 mutex_lock(&c->bucket_lock);
103 c->min_prio = USHRT_MAX;
105 for_each_cache(ca, c, i)
106 for_each_bucket(b, ca)
108 b->prio != BTREE_PRIO &&
109 !atomic_read(&b->pin)) {
111 c->min_prio = min(c->min_prio, b->prio);
114 mutex_unlock(&c->bucket_lock);
118 * Background allocation thread: scans for buckets to be invalidated,
119 * invalidates them, rewrites prios/gens (marking them as invalidated on disk),
120 * then optionally issues discard commands to the newly free buckets, then puts
121 * them on the various freelists.
124 static inline bool can_inc_bucket_gen(struct bucket *b)
126 return bucket_gc_gen(b) < BUCKET_GC_GEN_MAX;
129 bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b)
131 BUG_ON(!ca->set->gc_mark_valid);
133 return (!GC_MARK(b) ||
134 GC_MARK(b) == GC_MARK_RECLAIMABLE) &&
135 !atomic_read(&b->pin) &&
136 can_inc_bucket_gen(b);
139 void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
141 lockdep_assert_held(&ca->set->bucket_lock);
142 BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE);
144 if (GC_SECTORS_USED(b))
145 trace_bcache_invalidate(ca, b - ca->buckets);
148 b->prio = INITIAL_PRIO;
152 static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
154 __bch_invalidate_one_bucket(ca, b);
156 fifo_push(&ca->free_inc, b - ca->buckets);
160 * Determines what order we're going to reuse buckets, smallest bucket_prio()
161 * first: we also take into account the number of sectors of live data in that
162 * bucket, and in order for that multiply to make sense we have to scale bucket
164 * Thus, we scale the bucket priorities so that the bucket with the smallest
165 * prio is worth 1/8th of what INITIAL_PRIO is worth.
168 #define bucket_prio(b) \
170 unsigned min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8; \
172 (b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b); \
175 #define bucket_max_cmp(l, r) (bucket_prio(l) < bucket_prio(r))
176 #define bucket_min_cmp(l, r) (bucket_prio(l) > bucket_prio(r))
178 static void invalidate_buckets_lru(struct cache *ca)
185 for_each_bucket(b, ca) {
186 if (!bch_can_invalidate_bucket(ca, b))
189 if (!heap_full(&ca->heap))
190 heap_add(&ca->heap, b, bucket_max_cmp);
191 else if (bucket_max_cmp(b, heap_peek(&ca->heap))) {
192 ca->heap.data[0] = b;
193 heap_sift(&ca->heap, 0, bucket_max_cmp);
197 for (i = ca->heap.used / 2 - 1; i >= 0; --i)
198 heap_sift(&ca->heap, i, bucket_min_cmp);
200 while (!fifo_full(&ca->free_inc)) {
201 if (!heap_pop(&ca->heap, b, bucket_min_cmp)) {
203 * We don't want to be calling invalidate_buckets()
204 * multiple times when it can't do anything
206 ca->invalidate_needs_gc = 1;
211 bch_invalidate_one_bucket(ca, b);
215 static void invalidate_buckets_fifo(struct cache *ca)
220 while (!fifo_full(&ca->free_inc)) {
221 if (ca->fifo_last_bucket < ca->sb.first_bucket ||
222 ca->fifo_last_bucket >= ca->sb.nbuckets)
223 ca->fifo_last_bucket = ca->sb.first_bucket;
225 b = ca->buckets + ca->fifo_last_bucket++;
227 if (bch_can_invalidate_bucket(ca, b))
228 bch_invalidate_one_bucket(ca, b);
230 if (++checked >= ca->sb.nbuckets) {
231 ca->invalidate_needs_gc = 1;
238 static void invalidate_buckets_random(struct cache *ca)
243 while (!fifo_full(&ca->free_inc)) {
245 get_random_bytes(&n, sizeof(n));
247 n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket);
248 n += ca->sb.first_bucket;
252 if (bch_can_invalidate_bucket(ca, b))
253 bch_invalidate_one_bucket(ca, b);
255 if (++checked >= ca->sb.nbuckets / 2) {
256 ca->invalidate_needs_gc = 1;
263 static void invalidate_buckets(struct cache *ca)
265 BUG_ON(ca->invalidate_needs_gc);
267 switch (CACHE_REPLACEMENT(&ca->sb)) {
268 case CACHE_REPLACEMENT_LRU:
269 invalidate_buckets_lru(ca);
271 case CACHE_REPLACEMENT_FIFO:
272 invalidate_buckets_fifo(ca);
274 case CACHE_REPLACEMENT_RANDOM:
275 invalidate_buckets_random(ca);
280 #define allocator_wait(ca, cond) \
283 set_current_state(TASK_INTERRUPTIBLE); \
287 mutex_unlock(&(ca)->set->bucket_lock); \
288 if (kthread_should_stop()) { \
289 set_current_state(TASK_RUNNING); \
295 mutex_lock(&(ca)->set->bucket_lock); \
297 __set_current_state(TASK_RUNNING); \
300 static int bch_allocator_push(struct cache *ca, long bucket)
304 /* Prios/gens are actually the most important reserve */
305 if (fifo_push(&ca->free[RESERVE_PRIO], bucket))
308 for (i = 0; i < RESERVE_NR; i++)
309 if (fifo_push(&ca->free[i], bucket))
315 static int bch_allocator_thread(void *arg)
317 struct cache *ca = arg;
319 mutex_lock(&ca->set->bucket_lock);
323 * First, we pull buckets off of the unused and free_inc lists,
324 * possibly issue discards to them, then we add the bucket to
330 if (!fifo_pop(&ca->free_inc, bucket))
334 mutex_unlock(&ca->set->bucket_lock);
335 blkdev_issue_discard(ca->bdev,
336 bucket_to_sector(ca->set, bucket),
337 ca->sb.bucket_size, GFP_KERNEL, 0);
338 mutex_lock(&ca->set->bucket_lock);
341 allocator_wait(ca, bch_allocator_push(ca, bucket));
342 wake_up(&ca->set->btree_cache_wait);
343 wake_up(&ca->set->bucket_wait);
347 * We've run out of free buckets, we need to find some buckets
348 * we can invalidate. First, invalidate them in memory and add
349 * them to the free_inc list:
353 allocator_wait(ca, ca->set->gc_mark_valid &&
354 !ca->invalidate_needs_gc);
355 invalidate_buckets(ca);
358 * Now, we write their new gens to disk so we can start writing
361 allocator_wait(ca, !atomic_read(&ca->set->prio_blocked));
362 if (CACHE_SYNC(&ca->set->sb)) {
364 * This could deadlock if an allocation with a btree
365 * node locked ever blocked - having the btree node
366 * locked would block garbage collection, but here we're
367 * waiting on garbage collection before we invalidate
370 * But this should be safe since the btree code always
371 * uses btree_check_reserve() before allocating now, and
372 * if it fails it blocks without btree nodes locked.
374 if (!fifo_full(&ca->free_inc))
375 goto retry_invalidate;
384 long bch_bucket_alloc(struct cache *ca, unsigned reserve, bool wait)
391 if (fifo_pop(&ca->free[RESERVE_NONE], r) ||
392 fifo_pop(&ca->free[reserve], r))
396 trace_bcache_alloc_fail(ca, reserve);
401 prepare_to_wait(&ca->set->bucket_wait, &w,
402 TASK_UNINTERRUPTIBLE);
404 mutex_unlock(&ca->set->bucket_lock);
406 mutex_lock(&ca->set->bucket_lock);
407 } while (!fifo_pop(&ca->free[RESERVE_NONE], r) &&
408 !fifo_pop(&ca->free[reserve], r));
410 finish_wait(&ca->set->bucket_wait, &w);
412 if (ca->alloc_thread)
413 wake_up_process(ca->alloc_thread);
415 trace_bcache_alloc(ca, reserve);
417 if (expensive_debug_checks(ca->set)) {
422 for (iter = 0; iter < prio_buckets(ca) * 2; iter++)
423 BUG_ON(ca->prio_buckets[iter] == (uint64_t) r);
425 for (j = 0; j < RESERVE_NR; j++)
426 fifo_for_each(i, &ca->free[j], iter)
428 fifo_for_each(i, &ca->free_inc, iter)
434 BUG_ON(atomic_read(&b->pin) != 1);
436 SET_GC_SECTORS_USED(b, ca->sb.bucket_size);
438 if (reserve <= RESERVE_PRIO) {
439 SET_GC_MARK(b, GC_MARK_METADATA);
441 b->prio = BTREE_PRIO;
443 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
445 b->prio = INITIAL_PRIO;
451 void __bch_bucket_free(struct cache *ca, struct bucket *b)
454 SET_GC_SECTORS_USED(b, 0);
457 void bch_bucket_free(struct cache_set *c, struct bkey *k)
461 for (i = 0; i < KEY_PTRS(k); i++)
462 __bch_bucket_free(PTR_CACHE(c, k, i),
463 PTR_BUCKET(c, k, i));
466 int __bch_bucket_alloc_set(struct cache_set *c, unsigned reserve,
467 struct bkey *k, int n, bool wait)
471 lockdep_assert_held(&c->bucket_lock);
472 BUG_ON(!n || n > c->caches_loaded || n > 8);
476 /* sort by free space/prio of oldest data in caches */
478 for (i = 0; i < n; i++) {
479 struct cache *ca = c->cache_by_alloc[i];
480 long b = bch_bucket_alloc(ca, reserve, wait);
485 k->ptr[i] = MAKE_PTR(ca->buckets[b].gen,
486 bucket_to_sector(c, b),
489 SET_KEY_PTRS(k, i + 1);
494 bch_bucket_free(c, k);
499 int bch_bucket_alloc_set(struct cache_set *c, unsigned reserve,
500 struct bkey *k, int n, bool wait)
503 mutex_lock(&c->bucket_lock);
504 ret = __bch_bucket_alloc_set(c, reserve, k, n, wait);
505 mutex_unlock(&c->bucket_lock);
509 /* Sector allocator */
512 struct list_head list;
513 unsigned last_write_point;
514 unsigned sectors_free;
519 * We keep multiple buckets open for writes, and try to segregate different
520 * write streams for better cache utilization: first we try to segregate flash
521 * only volume write streams from cached devices, secondly we look for a bucket
522 * where the last write to it was sequential with the current write, and
523 * failing that we look for a bucket that was last used by the same task.
525 * The ideas is if you've got multiple tasks pulling data into the cache at the
526 * same time, you'll get better cache utilization if you try to segregate their
527 * data and preserve locality.
529 * For example, dirty sectors of flash only volume is not reclaimable, if their
530 * dirty sectors mixed with dirty sectors of cached device, such buckets will
531 * be marked as dirty and won't be reclaimed, though the dirty data of cached
532 * device have been written back to backend device.
534 * And say you've starting Firefox at the same time you're copying a
535 * bunch of files. Firefox will likely end up being fairly hot and stay in the
536 * cache awhile, but the data you copied might not be; if you wrote all that
537 * data to the same buckets it'd get invalidated at the same time.
539 * Both of those tasks will be doing fairly random IO so we can't rely on
540 * detecting sequential IO to segregate their data, but going off of the task
541 * should be a sane heuristic.
543 static struct open_bucket *pick_data_bucket(struct cache_set *c,
544 const struct bkey *search,
545 unsigned write_point,
548 struct open_bucket *ret, *ret_task = NULL;
550 list_for_each_entry_reverse(ret, &c->data_buckets, list)
551 if (UUID_FLASH_ONLY(&c->uuids[KEY_INODE(&ret->key)]) !=
552 UUID_FLASH_ONLY(&c->uuids[KEY_INODE(search)]))
554 else if (!bkey_cmp(&ret->key, search))
556 else if (ret->last_write_point == write_point)
559 ret = ret_task ?: list_first_entry(&c->data_buckets,
560 struct open_bucket, list);
562 if (!ret->sectors_free && KEY_PTRS(alloc)) {
563 ret->sectors_free = c->sb.bucket_size;
564 bkey_copy(&ret->key, alloc);
568 if (!ret->sectors_free)
575 * Allocates some space in the cache to write to, and k to point to the newly
576 * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the
577 * end of the newly allocated space).
579 * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many
580 * sectors were actually allocated.
582 * If s->writeback is true, will not fail.
584 bool bch_alloc_sectors(struct cache_set *c, struct bkey *k, unsigned sectors,
585 unsigned write_point, unsigned write_prio, bool wait)
587 struct open_bucket *b;
588 BKEY_PADDED(key) alloc;
592 * We might have to allocate a new bucket, which we can't do with a
593 * spinlock held. So if we have to allocate, we drop the lock, allocate
594 * and then retry. KEY_PTRS() indicates whether alloc points to
595 * allocated bucket(s).
598 bkey_init(&alloc.key);
599 spin_lock(&c->data_bucket_lock);
601 while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) {
602 unsigned watermark = write_prio
606 spin_unlock(&c->data_bucket_lock);
608 if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, wait))
611 spin_lock(&c->data_bucket_lock);
615 * If we had to allocate, we might race and not need to allocate the
616 * second time we call find_data_bucket(). If we allocated a bucket but
617 * didn't use it, drop the refcount bch_bucket_alloc_set() took:
619 if (KEY_PTRS(&alloc.key))
620 bkey_put(c, &alloc.key);
622 for (i = 0; i < KEY_PTRS(&b->key); i++)
623 EBUG_ON(ptr_stale(c, &b->key, i));
625 /* Set up the pointer to the space we're allocating: */
627 for (i = 0; i < KEY_PTRS(&b->key); i++)
628 k->ptr[i] = b->key.ptr[i];
630 sectors = min(sectors, b->sectors_free);
632 SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors);
633 SET_KEY_SIZE(k, sectors);
634 SET_KEY_PTRS(k, KEY_PTRS(&b->key));
637 * Move b to the end of the lru, and keep track of what this bucket was
640 list_move_tail(&b->list, &c->data_buckets);
641 bkey_copy_key(&b->key, k);
642 b->last_write_point = write_point;
644 b->sectors_free -= sectors;
646 for (i = 0; i < KEY_PTRS(&b->key); i++) {
647 SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors);
649 atomic_long_add(sectors,
650 &PTR_CACHE(c, &b->key, i)->sectors_written);
653 if (b->sectors_free < c->sb.block_size)
657 * k takes refcounts on the buckets it points to until it's inserted
658 * into the btree, but if we're done with this bucket we just transfer
659 * get_data_bucket()'s refcount.
662 for (i = 0; i < KEY_PTRS(&b->key); i++)
663 atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin);
665 spin_unlock(&c->data_bucket_lock);
671 void bch_open_buckets_free(struct cache_set *c)
673 struct open_bucket *b;
675 while (!list_empty(&c->data_buckets)) {
676 b = list_first_entry(&c->data_buckets,
677 struct open_bucket, list);
683 int bch_open_buckets_alloc(struct cache_set *c)
687 spin_lock_init(&c->data_bucket_lock);
689 for (i = 0; i < 6; i++) {
690 struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL);
694 list_add(&b->list, &c->data_buckets);
700 int bch_cache_allocator_start(struct cache *ca)
702 struct task_struct *k = kthread_run(bch_allocator_thread,
703 ca, "bcache_allocator");
707 ca->alloc_thread = k;