xref: /linux/drivers/md/bcache/alloc.c (revision 320fefa9e2edc67011e235ea1d50f0d00ddfe004)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Primary bucket allocation code
4  *
5  * Copyright 2012 Google, Inc.
6  *
7  * Allocation in bcache is done in terms of buckets:
8  *
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.
11  *
12  * Thus (assuming a bucket has no dirty data or metadata in it) we can reuse a
13  * bucket simply by incrementing its gen.
14  *
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.
18  *
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
22  * been overwritten.
23  *
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.
28  *
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.
32  *
33  * If we've got discards enabled, that happens when a bucket moves from the
34  * free_inc list to the free list.
35  *
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.
42  *
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.
49  *
50  * bch_bucket_alloc() allocates a single bucket from a specific cache.
51  *
52  * bch_bucket_alloc_set() allocates one  bucket from different caches
53  * out of a cache set.
54  *
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
57  * buckets are ready.
58  *
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.
62  */
63 
64 #include "bcache.h"
65 #include "btree.h"
66 
67 #include <linux/blkdev.h>
68 #include <linux/kthread.h>
69 #include <linux/random.h>
70 #include <trace/events/bcache.h>
71 
72 #define MAX_OPEN_BUCKETS 128
73 
74 /* Bucket heap / gen */
75 
76 uint8_t bch_inc_gen(struct cache *ca, struct bucket *b)
77 {
78 	uint8_t ret = ++b->gen;
79 
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);
82 
83 	return ret;
84 }
85 
86 void bch_rescale_priorities(struct cache_set *c, int sectors)
87 {
88 	struct cache *ca;
89 	struct bucket *b;
90 	unsigned long next = c->nbuckets * c->cache->sb.bucket_size / 1024;
91 	int r;
92 
93 	atomic_sub(sectors, &c->rescale);
94 
95 	do {
96 		r = atomic_read(&c->rescale);
97 
98 		if (r >= 0)
99 			return;
100 	} while (atomic_cmpxchg(&c->rescale, r, r + next) != r);
101 
102 	mutex_lock(&c->bucket_lock);
103 
104 	c->min_prio = USHRT_MAX;
105 
106 	ca = c->cache;
107 	for_each_bucket(b, ca)
108 		if (b->prio &&
109 		    b->prio != BTREE_PRIO &&
110 		    !atomic_read(&b->pin)) {
111 			b->prio--;
112 			c->min_prio = min(c->min_prio, b->prio);
113 		}
114 
115 	mutex_unlock(&c->bucket_lock);
116 }
117 
118 /*
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.
123  */
124 
125 static inline bool can_inc_bucket_gen(struct bucket *b)
126 {
127 	return bucket_gc_gen(b) < BUCKET_GC_GEN_MAX;
128 }
129 
130 bool bch_can_invalidate_bucket(struct cache *ca, struct bucket *b)
131 {
132 	BUG_ON(!ca->set->gc_mark_valid);
133 
134 	return (!GC_MARK(b) ||
135 		GC_MARK(b) == GC_MARK_RECLAIMABLE) &&
136 		!atomic_read(&b->pin) &&
137 		can_inc_bucket_gen(b);
138 }
139 
140 void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
141 {
142 	lockdep_assert_held(&ca->set->bucket_lock);
143 	BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE);
144 
145 	if (GC_SECTORS_USED(b))
146 		trace_bcache_invalidate(ca, b - ca->buckets);
147 
148 	bch_inc_gen(ca, b);
149 	b->prio = INITIAL_PRIO;
150 	atomic_inc(&b->pin);
151 }
152 
153 static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
154 {
155 	__bch_invalidate_one_bucket(ca, b);
156 
157 	fifo_push(&ca->free_inc, b - ca->buckets);
158 }
159 
160 /*
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
164  *
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.
167  */
168 
169 #define bucket_prio(b)							\
170 ({									\
171 	unsigned int min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8;	\
172 									\
173 	(b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b);	\
174 })
175 
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))
178 
179 static void invalidate_buckets_lru(struct cache *ca)
180 {
181 	struct bucket *b;
182 	ssize_t i;
183 
184 	ca->heap.used = 0;
185 
186 	for_each_bucket(b, ca) {
187 		if (!bch_can_invalidate_bucket(ca, b))
188 			continue;
189 
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);
195 		}
196 	}
197 
198 	for (i = ca->heap.used / 2 - 1; i >= 0; --i)
199 		heap_sift(&ca->heap, i, bucket_min_cmp);
200 
201 	while (!fifo_full(&ca->free_inc)) {
202 		if (!heap_pop(&ca->heap, b, bucket_min_cmp)) {
203 			/*
204 			 * We don't want to be calling invalidate_buckets()
205 			 * multiple times when it can't do anything
206 			 */
207 			ca->invalidate_needs_gc = 1;
208 			wake_up_gc(ca->set);
209 			return;
210 		}
211 
212 		bch_invalidate_one_bucket(ca, b);
213 	}
214 }
215 
216 static void invalidate_buckets_fifo(struct cache *ca)
217 {
218 	struct bucket *b;
219 	size_t checked = 0;
220 
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;
225 
226 		b = ca->buckets + ca->fifo_last_bucket++;
227 
228 		if (bch_can_invalidate_bucket(ca, b))
229 			bch_invalidate_one_bucket(ca, b);
230 
231 		if (++checked >= ca->sb.nbuckets) {
232 			ca->invalidate_needs_gc = 1;
233 			wake_up_gc(ca->set);
234 			return;
235 		}
236 	}
237 }
238 
239 static void invalidate_buckets_random(struct cache *ca)
240 {
241 	struct bucket *b;
242 	size_t checked = 0;
243 
244 	while (!fifo_full(&ca->free_inc)) {
245 		size_t n;
246 
247 		get_random_bytes(&n, sizeof(n));
248 
249 		n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket);
250 		n += ca->sb.first_bucket;
251 
252 		b = ca->buckets + n;
253 
254 		if (bch_can_invalidate_bucket(ca, b))
255 			bch_invalidate_one_bucket(ca, b);
256 
257 		if (++checked >= ca->sb.nbuckets / 2) {
258 			ca->invalidate_needs_gc = 1;
259 			wake_up_gc(ca->set);
260 			return;
261 		}
262 	}
263 }
264 
265 static void invalidate_buckets(struct cache *ca)
266 {
267 	BUG_ON(ca->invalidate_needs_gc);
268 
269 	switch (CACHE_REPLACEMENT(&ca->sb)) {
270 	case CACHE_REPLACEMENT_LRU:
271 		invalidate_buckets_lru(ca);
272 		break;
273 	case CACHE_REPLACEMENT_FIFO:
274 		invalidate_buckets_fifo(ca);
275 		break;
276 	case CACHE_REPLACEMENT_RANDOM:
277 		invalidate_buckets_random(ca);
278 		break;
279 	}
280 }
281 
282 #define allocator_wait(ca, cond)					\
283 do {									\
284 	while (1) {							\
285 		set_current_state(TASK_INTERRUPTIBLE);			\
286 		if (cond)						\
287 			break;						\
288 									\
289 		mutex_unlock(&(ca)->set->bucket_lock);			\
290 		if (kthread_should_stop() ||				\
291 		    test_bit(CACHE_SET_IO_DISABLE, &ca->set->flags)) {	\
292 			set_current_state(TASK_RUNNING);		\
293 			goto out;					\
294 		}							\
295 									\
296 		schedule();						\
297 		mutex_lock(&(ca)->set->bucket_lock);			\
298 	}								\
299 	__set_current_state(TASK_RUNNING);				\
300 } while (0)
301 
302 static int bch_allocator_push(struct cache *ca, long bucket)
303 {
304 	unsigned int i;
305 
306 	/* Prios/gens are actually the most important reserve */
307 	if (fifo_push(&ca->free[RESERVE_PRIO], bucket))
308 		return true;
309 
310 	for (i = 0; i < RESERVE_NR; i++)
311 		if (fifo_push(&ca->free[i], bucket))
312 			return true;
313 
314 	return false;
315 }
316 
317 static int bch_allocator_thread(void *arg)
318 {
319 	struct cache *ca = arg;
320 
321 	mutex_lock(&ca->set->bucket_lock);
322 
323 	while (1) {
324 		/*
325 		 * First, we pull buckets off of the unused and free_inc lists,
326 		 * possibly issue discards to them, then we add the bucket to
327 		 * the free list:
328 		 */
329 		while (1) {
330 			long bucket;
331 
332 			if (!fifo_pop(&ca->free_inc, bucket))
333 				break;
334 
335 			if (ca->discard) {
336 				mutex_unlock(&ca->set->bucket_lock);
337 				blkdev_issue_discard(ca->bdev,
338 					bucket_to_sector(ca->set, bucket),
339 					ca->sb.bucket_size, GFP_KERNEL);
340 				mutex_lock(&ca->set->bucket_lock);
341 			}
342 
343 			allocator_wait(ca, bch_allocator_push(ca, bucket));
344 			wake_up(&ca->set->btree_cache_wait);
345 			wake_up(&ca->set->bucket_wait);
346 		}
347 
348 		/*
349 		 * We've run out of free buckets, we need to find some buckets
350 		 * we can invalidate. First, invalidate them in memory and add
351 		 * them to the free_inc list:
352 		 */
353 
354 retry_invalidate:
355 		allocator_wait(ca, ca->set->gc_mark_valid &&
356 			       !ca->invalidate_needs_gc);
357 		invalidate_buckets(ca);
358 
359 		/*
360 		 * Now, we write their new gens to disk so we can start writing
361 		 * new stuff to them:
362 		 */
363 		allocator_wait(ca, !atomic_read(&ca->set->prio_blocked));
364 		if (CACHE_SYNC(&ca->sb)) {
365 			/*
366 			 * This could deadlock if an allocation with a btree
367 			 * node locked ever blocked - having the btree node
368 			 * locked would block garbage collection, but here we're
369 			 * waiting on garbage collection before we invalidate
370 			 * and free anything.
371 			 *
372 			 * But this should be safe since the btree code always
373 			 * uses btree_check_reserve() before allocating now, and
374 			 * if it fails it blocks without btree nodes locked.
375 			 */
376 			if (!fifo_full(&ca->free_inc))
377 				goto retry_invalidate;
378 
379 			if (bch_prio_write(ca, false) < 0) {
380 				ca->invalidate_needs_gc = 1;
381 				wake_up_gc(ca->set);
382 			}
383 		}
384 	}
385 out:
386 	wait_for_kthread_stop();
387 	return 0;
388 }
389 
390 /* Allocation */
391 
392 long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait)
393 {
394 	DEFINE_WAIT(w);
395 	struct bucket *b;
396 	long r;
397 
398 
399 	/* No allocation if CACHE_SET_IO_DISABLE bit is set */
400 	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &ca->set->flags)))
401 		return -1;
402 
403 	/* fastpath */
404 	if (fifo_pop(&ca->free[RESERVE_NONE], r) ||
405 	    fifo_pop(&ca->free[reserve], r))
406 		goto out;
407 
408 	if (!wait) {
409 		trace_bcache_alloc_fail(ca, reserve);
410 		return -1;
411 	}
412 
413 	do {
414 		prepare_to_wait(&ca->set->bucket_wait, &w,
415 				TASK_UNINTERRUPTIBLE);
416 
417 		mutex_unlock(&ca->set->bucket_lock);
418 		schedule();
419 		mutex_lock(&ca->set->bucket_lock);
420 	} while (!fifo_pop(&ca->free[RESERVE_NONE], r) &&
421 		 !fifo_pop(&ca->free[reserve], r));
422 
423 	finish_wait(&ca->set->bucket_wait, &w);
424 out:
425 	if (ca->alloc_thread)
426 		wake_up_process(ca->alloc_thread);
427 
428 	trace_bcache_alloc(ca, reserve);
429 
430 	if (expensive_debug_checks(ca->set)) {
431 		size_t iter;
432 		long i;
433 		unsigned int j;
434 
435 		for (iter = 0; iter < prio_buckets(ca) * 2; iter++)
436 			BUG_ON(ca->prio_buckets[iter] == (uint64_t) r);
437 
438 		for (j = 0; j < RESERVE_NR; j++)
439 			fifo_for_each(i, &ca->free[j], iter)
440 				BUG_ON(i == r);
441 		fifo_for_each(i, &ca->free_inc, iter)
442 			BUG_ON(i == r);
443 	}
444 
445 	b = ca->buckets + r;
446 
447 	BUG_ON(atomic_read(&b->pin) != 1);
448 
449 	SET_GC_SECTORS_USED(b, ca->sb.bucket_size);
450 
451 	if (reserve <= RESERVE_PRIO) {
452 		SET_GC_MARK(b, GC_MARK_METADATA);
453 		SET_GC_MOVE(b, 0);
454 		b->prio = BTREE_PRIO;
455 	} else {
456 		SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
457 		SET_GC_MOVE(b, 0);
458 		b->prio = INITIAL_PRIO;
459 	}
460 
461 	if (ca->set->avail_nbuckets > 0) {
462 		ca->set->avail_nbuckets--;
463 		bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
464 	}
465 
466 	return r;
467 }
468 
469 void __bch_bucket_free(struct cache *ca, struct bucket *b)
470 {
471 	SET_GC_MARK(b, 0);
472 	SET_GC_SECTORS_USED(b, 0);
473 
474 	if (ca->set->avail_nbuckets < ca->set->nbuckets) {
475 		ca->set->avail_nbuckets++;
476 		bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
477 	}
478 }
479 
480 void bch_bucket_free(struct cache_set *c, struct bkey *k)
481 {
482 	unsigned int i;
483 
484 	for (i = 0; i < KEY_PTRS(k); i++)
485 		__bch_bucket_free(c->cache, PTR_BUCKET(c, k, i));
486 }
487 
488 int __bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
489 			   struct bkey *k, bool wait)
490 {
491 	struct cache *ca;
492 	long b;
493 
494 	/* No allocation if CACHE_SET_IO_DISABLE bit is set */
495 	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
496 		return -1;
497 
498 	lockdep_assert_held(&c->bucket_lock);
499 
500 	bkey_init(k);
501 
502 	ca = c->cache;
503 	b = bch_bucket_alloc(ca, reserve, wait);
504 	if (b == -1)
505 		goto err;
506 
507 	k->ptr[0] = MAKE_PTR(ca->buckets[b].gen,
508 			     bucket_to_sector(c, b),
509 			     ca->sb.nr_this_dev);
510 
511 	SET_KEY_PTRS(k, 1);
512 
513 	return 0;
514 err:
515 	bch_bucket_free(c, k);
516 	bkey_put(c, k);
517 	return -1;
518 }
519 
520 int bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
521 			 struct bkey *k, bool wait)
522 {
523 	int ret;
524 
525 	mutex_lock(&c->bucket_lock);
526 	ret = __bch_bucket_alloc_set(c, reserve, k, wait);
527 	mutex_unlock(&c->bucket_lock);
528 	return ret;
529 }
530 
531 /* Sector allocator */
532 
533 struct open_bucket {
534 	struct list_head	list;
535 	unsigned int		last_write_point;
536 	unsigned int		sectors_free;
537 	BKEY_PADDED(key);
538 };
539 
540 /*
541  * We keep multiple buckets open for writes, and try to segregate different
542  * write streams for better cache utilization: first we try to segregate flash
543  * only volume write streams from cached devices, secondly we look for a bucket
544  * where the last write to it was sequential with the current write, and
545  * failing that we look for a bucket that was last used by the same task.
546  *
547  * The ideas is if you've got multiple tasks pulling data into the cache at the
548  * same time, you'll get better cache utilization if you try to segregate their
549  * data and preserve locality.
550  *
551  * For example, dirty sectors of flash only volume is not reclaimable, if their
552  * dirty sectors mixed with dirty sectors of cached device, such buckets will
553  * be marked as dirty and won't be reclaimed, though the dirty data of cached
554  * device have been written back to backend device.
555  *
556  * And say you've starting Firefox at the same time you're copying a
557  * bunch of files. Firefox will likely end up being fairly hot and stay in the
558  * cache awhile, but the data you copied might not be; if you wrote all that
559  * data to the same buckets it'd get invalidated at the same time.
560  *
561  * Both of those tasks will be doing fairly random IO so we can't rely on
562  * detecting sequential IO to segregate their data, but going off of the task
563  * should be a sane heuristic.
564  */
565 static struct open_bucket *pick_data_bucket(struct cache_set *c,
566 					    const struct bkey *search,
567 					    unsigned int write_point,
568 					    struct bkey *alloc)
569 {
570 	struct open_bucket *ret, *ret_task = NULL;
571 
572 	list_for_each_entry_reverse(ret, &c->data_buckets, list)
573 		if (UUID_FLASH_ONLY(&c->uuids[KEY_INODE(&ret->key)]) !=
574 		    UUID_FLASH_ONLY(&c->uuids[KEY_INODE(search)]))
575 			continue;
576 		else if (!bkey_cmp(&ret->key, search))
577 			goto found;
578 		else if (ret->last_write_point == write_point)
579 			ret_task = ret;
580 
581 	ret = ret_task ?: list_first_entry(&c->data_buckets,
582 					   struct open_bucket, list);
583 found:
584 	if (!ret->sectors_free && KEY_PTRS(alloc)) {
585 		ret->sectors_free = c->cache->sb.bucket_size;
586 		bkey_copy(&ret->key, alloc);
587 		bkey_init(alloc);
588 	}
589 
590 	if (!ret->sectors_free)
591 		ret = NULL;
592 
593 	return ret;
594 }
595 
596 /*
597  * Allocates some space in the cache to write to, and k to point to the newly
598  * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the
599  * end of the newly allocated space).
600  *
601  * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many
602  * sectors were actually allocated.
603  *
604  * If s->writeback is true, will not fail.
605  */
606 bool bch_alloc_sectors(struct cache_set *c,
607 		       struct bkey *k,
608 		       unsigned int sectors,
609 		       unsigned int write_point,
610 		       unsigned int write_prio,
611 		       bool wait)
612 {
613 	struct open_bucket *b;
614 	BKEY_PADDED(key) alloc;
615 	unsigned int i;
616 
617 	/*
618 	 * We might have to allocate a new bucket, which we can't do with a
619 	 * spinlock held. So if we have to allocate, we drop the lock, allocate
620 	 * and then retry. KEY_PTRS() indicates whether alloc points to
621 	 * allocated bucket(s).
622 	 */
623 
624 	bkey_init(&alloc.key);
625 	spin_lock(&c->data_bucket_lock);
626 
627 	while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) {
628 		unsigned int watermark = write_prio
629 			? RESERVE_MOVINGGC
630 			: RESERVE_NONE;
631 
632 		spin_unlock(&c->data_bucket_lock);
633 
634 		if (bch_bucket_alloc_set(c, watermark, &alloc.key, wait))
635 			return false;
636 
637 		spin_lock(&c->data_bucket_lock);
638 	}
639 
640 	/*
641 	 * If we had to allocate, we might race and not need to allocate the
642 	 * second time we call pick_data_bucket(). If we allocated a bucket but
643 	 * didn't use it, drop the refcount bch_bucket_alloc_set() took:
644 	 */
645 	if (KEY_PTRS(&alloc.key))
646 		bkey_put(c, &alloc.key);
647 
648 	for (i = 0; i < KEY_PTRS(&b->key); i++)
649 		EBUG_ON(ptr_stale(c, &b->key, i));
650 
651 	/* Set up the pointer to the space we're allocating: */
652 
653 	for (i = 0; i < KEY_PTRS(&b->key); i++)
654 		k->ptr[i] = b->key.ptr[i];
655 
656 	sectors = min(sectors, b->sectors_free);
657 
658 	SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors);
659 	SET_KEY_SIZE(k, sectors);
660 	SET_KEY_PTRS(k, KEY_PTRS(&b->key));
661 
662 	/*
663 	 * Move b to the end of the lru, and keep track of what this bucket was
664 	 * last used for:
665 	 */
666 	list_move_tail(&b->list, &c->data_buckets);
667 	bkey_copy_key(&b->key, k);
668 	b->last_write_point = write_point;
669 
670 	b->sectors_free	-= sectors;
671 
672 	for (i = 0; i < KEY_PTRS(&b->key); i++) {
673 		SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors);
674 
675 		atomic_long_add(sectors,
676 				&c->cache->sectors_written);
677 	}
678 
679 	if (b->sectors_free < c->cache->sb.block_size)
680 		b->sectors_free = 0;
681 
682 	/*
683 	 * k takes refcounts on the buckets it points to until it's inserted
684 	 * into the btree, but if we're done with this bucket we just transfer
685 	 * get_data_bucket()'s refcount.
686 	 */
687 	if (b->sectors_free)
688 		for (i = 0; i < KEY_PTRS(&b->key); i++)
689 			atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin);
690 
691 	spin_unlock(&c->data_bucket_lock);
692 	return true;
693 }
694 
695 /* Init */
696 
697 void bch_open_buckets_free(struct cache_set *c)
698 {
699 	struct open_bucket *b;
700 
701 	while (!list_empty(&c->data_buckets)) {
702 		b = list_first_entry(&c->data_buckets,
703 				     struct open_bucket, list);
704 		list_del(&b->list);
705 		kfree(b);
706 	}
707 }
708 
709 int bch_open_buckets_alloc(struct cache_set *c)
710 {
711 	int i;
712 
713 	spin_lock_init(&c->data_bucket_lock);
714 
715 	for (i = 0; i < MAX_OPEN_BUCKETS; i++) {
716 		struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL);
717 
718 		if (!b)
719 			return -ENOMEM;
720 
721 		list_add(&b->list, &c->data_buckets);
722 	}
723 
724 	return 0;
725 }
726 
727 int bch_cache_allocator_start(struct cache *ca)
728 {
729 	struct task_struct *k = kthread_run(bch_allocator_thread,
730 					    ca, "bcache_allocator");
731 	if (IS_ERR(k))
732 		return PTR_ERR(k);
733 
734 	ca->alloc_thread = k;
735 	return 0;
736 }
737