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/kthread.h>
68 #include <linux/random.h>
69 #include <trace/events/bcache.h>
71 #define MAX_OPEN_BUCKETS 128
73 /* Bucket heap / gen */
75 uint8_t bch_inc_gen(struct cache *ca, struct bucket *b)
77 uint8_t ret = ++b->gen;
79 ca->set->need_gc = max(ca->set->need_gc, bucket_gc_gen(b));
80 WARN_ON_ONCE(ca->set->need_gc > BUCKET_GC_GEN_MAX);
85 void bch_rescale_priorities(struct cache_set *c, int sectors)
89 unsigned next = c->nbuckets * c->sb.bucket_size / 1024;
93 atomic_sub(sectors, &c->rescale);
96 r = atomic_read(&c->rescale);
100 } while (atomic_cmpxchg(&c->rescale, r, r + next) != r);
102 mutex_lock(&c->bucket_lock);
104 c->min_prio = USHRT_MAX;
106 for_each_cache(ca, c, i)
107 for_each_bucket(b, ca)
109 b->prio != BTREE_PRIO &&
110 !atomic_read(&b->pin)) {
112 c->min_prio = min(c->min_prio, b->prio);
115 mutex_unlock(&c->bucket_lock);
119 * Background allocation thread: scans for buckets to be invalidated,
120 * invalidates them, rewrites prios/gens (marking them as invalidated on disk),
121 * then optionally issues discard commands to the newly free buckets, then puts
122 * them on the various freelists.
125 static inline bool can_inc_bucket_gen(struct bucket *b)
127 return bucket_gc_gen(b) < BUCKET_GC_GEN_MAX;
130 bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b)
132 BUG_ON(!ca->set->gc_mark_valid);
134 return (!GC_MARK(b) ||
135 GC_MARK(b) == GC_MARK_RECLAIMABLE) &&
136 !atomic_read(&b->pin) &&
137 can_inc_bucket_gen(b);
140 void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
142 lockdep_assert_held(&ca->set->bucket_lock);
143 BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE);
145 if (GC_SECTORS_USED(b))
146 trace_bcache_invalidate(ca, b - ca->buckets);
149 b->prio = INITIAL_PRIO;
153 static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
155 __bch_invalidate_one_bucket(ca, b);
157 fifo_push(&ca->free_inc, b - ca->buckets);
161 * Determines what order we're going to reuse buckets, smallest bucket_prio()
162 * first: we also take into account the number of sectors of live data in that
163 * bucket, and in order for that multiply to make sense we have to scale bucket
165 * Thus, we scale the bucket priorities so that the bucket with the smallest
166 * prio is worth 1/8th of what INITIAL_PRIO is worth.
169 #define bucket_prio(b) \
171 unsigned min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8; \
173 (b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b); \
176 #define bucket_max_cmp(l, r) (bucket_prio(l) < bucket_prio(r))
177 #define bucket_min_cmp(l, r) (bucket_prio(l) > bucket_prio(r))
179 static void invalidate_buckets_lru(struct cache *ca)
186 for_each_bucket(b, ca) {
187 if (!bch_can_invalidate_bucket(ca, b))
190 if (!heap_full(&ca->heap))
191 heap_add(&ca->heap, b, bucket_max_cmp);
192 else if (bucket_max_cmp(b, heap_peek(&ca->heap))) {
193 ca->heap.data[0] = b;
194 heap_sift(&ca->heap, 0, bucket_max_cmp);
198 for (i = ca->heap.used / 2 - 1; i >= 0; --i)
199 heap_sift(&ca->heap, i, bucket_min_cmp);
201 while (!fifo_full(&ca->free_inc)) {
202 if (!heap_pop(&ca->heap, b, bucket_min_cmp)) {
204 * We don't want to be calling invalidate_buckets()
205 * multiple times when it can't do anything
207 ca->invalidate_needs_gc = 1;
212 bch_invalidate_one_bucket(ca, b);
216 static void invalidate_buckets_fifo(struct cache *ca)
221 while (!fifo_full(&ca->free_inc)) {
222 if (ca->fifo_last_bucket < ca->sb.first_bucket ||
223 ca->fifo_last_bucket >= ca->sb.nbuckets)
224 ca->fifo_last_bucket = ca->sb.first_bucket;
226 b = ca->buckets + ca->fifo_last_bucket++;
228 if (bch_can_invalidate_bucket(ca, b))
229 bch_invalidate_one_bucket(ca, b);
231 if (++checked >= ca->sb.nbuckets) {
232 ca->invalidate_needs_gc = 1;
239 static void invalidate_buckets_random(struct cache *ca)
244 while (!fifo_full(&ca->free_inc)) {
246 get_random_bytes(&n, sizeof(n));
248 n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket);
249 n += ca->sb.first_bucket;
253 if (bch_can_invalidate_bucket(ca, b))
254 bch_invalidate_one_bucket(ca, b);
256 if (++checked >= ca->sb.nbuckets / 2) {
257 ca->invalidate_needs_gc = 1;
264 static void invalidate_buckets(struct cache *ca)
266 BUG_ON(ca->invalidate_needs_gc);
268 switch (CACHE_REPLACEMENT(&ca->sb)) {
269 case CACHE_REPLACEMENT_LRU:
270 invalidate_buckets_lru(ca);
272 case CACHE_REPLACEMENT_FIFO:
273 invalidate_buckets_fifo(ca);
275 case CACHE_REPLACEMENT_RANDOM:
276 invalidate_buckets_random(ca);
281 #define allocator_wait(ca, cond) \
284 set_current_state(TASK_INTERRUPTIBLE); \
288 mutex_unlock(&(ca)->set->bucket_lock); \
289 if (kthread_should_stop()) \
293 mutex_lock(&(ca)->set->bucket_lock); \
295 __set_current_state(TASK_RUNNING); \
298 static int bch_allocator_push(struct cache *ca, long bucket)
302 /* Prios/gens are actually the most important reserve */
303 if (fifo_push(&ca->free[RESERVE_PRIO], bucket))
306 for (i = 0; i < RESERVE_NR; i++)
307 if (fifo_push(&ca->free[i], bucket))
313 static int bch_allocator_thread(void *arg)
315 struct cache *ca = arg;
317 mutex_lock(&ca->set->bucket_lock);
321 * First, we pull buckets off of the unused and free_inc lists,
322 * possibly issue discards to them, then we add the bucket to
325 while (!fifo_empty(&ca->free_inc)) {
328 fifo_pop(&ca->free_inc, bucket);
331 mutex_unlock(&ca->set->bucket_lock);
332 blkdev_issue_discard(ca->bdev,
333 bucket_to_sector(ca->set, bucket),
334 ca->sb.bucket_size, GFP_KERNEL, 0);
335 mutex_lock(&ca->set->bucket_lock);
338 allocator_wait(ca, bch_allocator_push(ca, bucket));
339 wake_up(&ca->set->btree_cache_wait);
340 wake_up(&ca->set->bucket_wait);
344 * We've run out of free buckets, we need to find some buckets
345 * we can invalidate. First, invalidate them in memory and add
346 * them to the free_inc list:
350 allocator_wait(ca, ca->set->gc_mark_valid &&
351 !ca->invalidate_needs_gc);
352 invalidate_buckets(ca);
355 * Now, we write their new gens to disk so we can start writing
358 allocator_wait(ca, !atomic_read(&ca->set->prio_blocked));
359 if (CACHE_SYNC(&ca->set->sb)) {
361 * This could deadlock if an allocation with a btree
362 * node locked ever blocked - having the btree node
363 * locked would block garbage collection, but here we're
364 * waiting on garbage collection before we invalidate
367 * But this should be safe since the btree code always
368 * uses btree_check_reserve() before allocating now, and
369 * if it fails it blocks without btree nodes locked.
371 if (!fifo_full(&ca->free_inc))
372 goto retry_invalidate;
381 long bch_bucket_alloc(struct cache *ca, unsigned reserve, bool wait)
388 if (fifo_pop(&ca->free[RESERVE_NONE], r) ||
389 fifo_pop(&ca->free[reserve], r))
393 trace_bcache_alloc_fail(ca, reserve);
398 prepare_to_wait(&ca->set->bucket_wait, &w,
399 TASK_UNINTERRUPTIBLE);
401 mutex_unlock(&ca->set->bucket_lock);
403 mutex_lock(&ca->set->bucket_lock);
404 } while (!fifo_pop(&ca->free[RESERVE_NONE], r) &&
405 !fifo_pop(&ca->free[reserve], r));
407 finish_wait(&ca->set->bucket_wait, &w);
409 if (ca->alloc_thread)
410 wake_up_process(ca->alloc_thread);
412 trace_bcache_alloc(ca, reserve);
414 if (expensive_debug_checks(ca->set)) {
419 for (iter = 0; iter < prio_buckets(ca) * 2; iter++)
420 BUG_ON(ca->prio_buckets[iter] == (uint64_t) r);
422 for (j = 0; j < RESERVE_NR; j++)
423 fifo_for_each(i, &ca->free[j], iter)
425 fifo_for_each(i, &ca->free_inc, iter)
431 BUG_ON(atomic_read(&b->pin) != 1);
433 SET_GC_SECTORS_USED(b, ca->sb.bucket_size);
435 if (reserve <= RESERVE_PRIO) {
436 SET_GC_MARK(b, GC_MARK_METADATA);
438 b->prio = BTREE_PRIO;
440 SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
442 b->prio = INITIAL_PRIO;
445 if (ca->set->avail_nbuckets > 0) {
446 ca->set->avail_nbuckets--;
447 bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
453 void __bch_bucket_free(struct cache *ca, struct bucket *b)
456 SET_GC_SECTORS_USED(b, 0);
458 if (ca->set->avail_nbuckets < ca->set->nbuckets) {
459 ca->set->avail_nbuckets++;
460 bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
464 void bch_bucket_free(struct cache_set *c, struct bkey *k)
468 for (i = 0; i < KEY_PTRS(k); i++)
469 __bch_bucket_free(PTR_CACHE(c, k, i),
470 PTR_BUCKET(c, k, i));
473 int __bch_bucket_alloc_set(struct cache_set *c, unsigned reserve,
474 struct bkey *k, int n, bool wait)
478 lockdep_assert_held(&c->bucket_lock);
479 BUG_ON(!n || n > c->caches_loaded || n > 8);
483 /* sort by free space/prio of oldest data in caches */
485 for (i = 0; i < n; i++) {
486 struct cache *ca = c->cache_by_alloc[i];
487 long b = bch_bucket_alloc(ca, reserve, wait);
492 k->ptr[i] = PTR(ca->buckets[b].gen,
493 bucket_to_sector(c, b),
496 SET_KEY_PTRS(k, i + 1);
501 bch_bucket_free(c, k);
506 int bch_bucket_alloc_set(struct cache_set *c, unsigned reserve,
507 struct bkey *k, int n, bool wait)
510 mutex_lock(&c->bucket_lock);
511 ret = __bch_bucket_alloc_set(c, reserve, k, n, wait);
512 mutex_unlock(&c->bucket_lock);
516 /* Sector allocator */
519 struct list_head list;
520 unsigned last_write_point;
521 unsigned sectors_free;
526 * We keep multiple buckets open for writes, and try to segregate different
527 * write streams for better cache utilization: first we look for a bucket where
528 * the last write to it was sequential with the current write, and failing that
529 * we look for a bucket that was last used by the same task.
531 * The ideas is if you've got multiple tasks pulling data into the cache at the
532 * same time, you'll get better cache utilization if you try to segregate their
533 * data and preserve locality.
535 * For example, say you've starting Firefox at the same time you're copying a
536 * bunch of files. Firefox will likely end up being fairly hot and stay in the
537 * cache awhile, but the data you copied might not be; if you wrote all that
538 * data to the same buckets it'd get invalidated at the same time.
540 * Both of those tasks will be doing fairly random IO so we can't rely on
541 * detecting sequential IO to segregate their data, but going off of the task
542 * should be a sane heuristic.
544 static struct open_bucket *pick_data_bucket(struct cache_set *c,
545 const struct bkey *search,
546 unsigned write_point,
549 struct open_bucket *ret, *ret_task = NULL;
551 list_for_each_entry_reverse(ret, &c->data_buckets, list)
552 if (!bkey_cmp(&ret->key, search))
554 else if (ret->last_write_point == write_point)
557 ret = ret_task ?: list_first_entry(&c->data_buckets,
558 struct open_bucket, list);
560 if (!ret->sectors_free && KEY_PTRS(alloc)) {
561 ret->sectors_free = c->sb.bucket_size;
562 bkey_copy(&ret->key, alloc);
566 if (!ret->sectors_free)
573 * Allocates some space in the cache to write to, and k to point to the newly
574 * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the
575 * end of the newly allocated space).
577 * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many
578 * sectors were actually allocated.
580 * If s->writeback is true, will not fail.
582 bool bch_alloc_sectors(struct cache_set *c, struct bkey *k, unsigned sectors,
583 unsigned write_point, unsigned write_prio, bool wait)
585 struct open_bucket *b;
586 BKEY_PADDED(key) alloc;
590 * We might have to allocate a new bucket, which we can't do with a
591 * spinlock held. So if we have to allocate, we drop the lock, allocate
592 * and then retry. KEY_PTRS() indicates whether alloc points to
593 * allocated bucket(s).
596 bkey_init(&alloc.key);
597 spin_lock(&c->data_bucket_lock);
599 while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) {
600 unsigned watermark = write_prio
604 spin_unlock(&c->data_bucket_lock);
606 if (bch_bucket_alloc_set(c, watermark, &alloc.key, 1, wait))
609 spin_lock(&c->data_bucket_lock);
613 * If we had to allocate, we might race and not need to allocate the
614 * second time we call pick_data_bucket(). If we allocated a bucket but
615 * didn't use it, drop the refcount bch_bucket_alloc_set() took:
617 if (KEY_PTRS(&alloc.key))
618 bkey_put(c, &alloc.key);
620 for (i = 0; i < KEY_PTRS(&b->key); i++)
621 EBUG_ON(ptr_stale(c, &b->key, i));
623 /* Set up the pointer to the space we're allocating: */
625 for (i = 0; i < KEY_PTRS(&b->key); i++)
626 k->ptr[i] = b->key.ptr[i];
628 sectors = min(sectors, b->sectors_free);
630 SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors);
631 SET_KEY_SIZE(k, sectors);
632 SET_KEY_PTRS(k, KEY_PTRS(&b->key));
635 * Move b to the end of the lru, and keep track of what this bucket was
638 list_move_tail(&b->list, &c->data_buckets);
639 bkey_copy_key(&b->key, k);
640 b->last_write_point = write_point;
642 b->sectors_free -= sectors;
644 for (i = 0; i < KEY_PTRS(&b->key); i++) {
645 SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors);
647 atomic_long_add(sectors,
648 &PTR_CACHE(c, &b->key, i)->sectors_written);
651 if (b->sectors_free < c->sb.block_size)
655 * k takes refcounts on the buckets it points to until it's inserted
656 * into the btree, but if we're done with this bucket we just transfer
657 * get_data_bucket()'s refcount.
660 for (i = 0; i < KEY_PTRS(&b->key); i++)
661 atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin);
663 spin_unlock(&c->data_bucket_lock);
669 void bch_open_buckets_free(struct cache_set *c)
671 struct open_bucket *b;
673 while (!list_empty(&c->data_buckets)) {
674 b = list_first_entry(&c->data_buckets,
675 struct open_bucket, list);
681 int bch_open_buckets_alloc(struct cache_set *c)
685 spin_lock_init(&c->data_bucket_lock);
687 for (i = 0; i < MAX_OPEN_BUCKETS; i++) {
688 struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL);
692 list_add(&b->list, &c->data_buckets);
698 int bch_cache_allocator_start(struct cache *ca)
700 struct task_struct *k = kthread_run(bch_allocator_thread,
701 ca, "bcache_allocator");
705 ca->alloc_thread = k;