xref: /linux/drivers/md/bcache/alloc.c (revision 31392048f55f98cb01ca709d32d06d926ab9760a)
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 	return (ca->set->gc_mark_valid || b->reclaimable_in_gc) &&
133 	       ((!GC_MARK(b) || GC_MARK(b) == GC_MARK_RECLAIMABLE) &&
134 	       !atomic_read(&b->pin) && can_inc_bucket_gen(b));
135 }
136 
137 void __bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
138 {
139 	lockdep_assert_held(&ca->set->bucket_lock);
140 	BUG_ON(GC_MARK(b) && GC_MARK(b) != GC_MARK_RECLAIMABLE);
141 
142 	if (GC_SECTORS_USED(b))
143 		trace_bcache_invalidate(ca, b - ca->buckets);
144 
145 	bch_inc_gen(ca, b);
146 	b->prio = INITIAL_PRIO;
147 	atomic_inc(&b->pin);
148 	b->reclaimable_in_gc = 0;
149 }
150 
151 static void bch_invalidate_one_bucket(struct cache *ca, struct bucket *b)
152 {
153 	__bch_invalidate_one_bucket(ca, b);
154 
155 	fifo_push(&ca->free_inc, b - ca->buckets);
156 }
157 
158 /*
159  * Determines what order we're going to reuse buckets, smallest bucket_prio()
160  * first: we also take into account the number of sectors of live data in that
161  * bucket, and in order for that multiply to make sense we have to scale bucket
162  *
163  * Thus, we scale the bucket priorities so that the bucket with the smallest
164  * prio is worth 1/8th of what INITIAL_PRIO is worth.
165  */
166 
167 #define bucket_prio(b)							\
168 ({									\
169 	unsigned int min_prio = (INITIAL_PRIO - ca->set->min_prio) / 8;	\
170 									\
171 	(b->prio - ca->set->min_prio + min_prio) * GC_SECTORS_USED(b);	\
172 })
173 
174 #define bucket_max_cmp(l, r)	(bucket_prio(l) < bucket_prio(r))
175 #define bucket_min_cmp(l, r)	(bucket_prio(l) > bucket_prio(r))
176 
177 static void invalidate_buckets_lru(struct cache *ca)
178 {
179 	struct bucket *b;
180 	ssize_t i;
181 
182 	ca->heap.used = 0;
183 
184 	for_each_bucket(b, ca) {
185 		if (!bch_can_invalidate_bucket(ca, b))
186 			continue;
187 
188 		if (!heap_full(&ca->heap))
189 			heap_add(&ca->heap, b, bucket_max_cmp);
190 		else if (bucket_max_cmp(b, heap_peek(&ca->heap))) {
191 			ca->heap.data[0] = b;
192 			heap_sift(&ca->heap, 0, bucket_max_cmp);
193 		}
194 	}
195 
196 	for (i = ca->heap.used / 2 - 1; i >= 0; --i)
197 		heap_sift(&ca->heap, i, bucket_min_cmp);
198 
199 	while (!fifo_full(&ca->free_inc)) {
200 		if (!heap_pop(&ca->heap, b, bucket_min_cmp)) {
201 			/*
202 			 * We don't want to be calling invalidate_buckets()
203 			 * multiple times when it can't do anything
204 			 */
205 			ca->invalidate_needs_gc = 1;
206 			wake_up_gc(ca->set);
207 			return;
208 		}
209 
210 		bch_invalidate_one_bucket(ca, b);
211 	}
212 }
213 
214 static void invalidate_buckets_fifo(struct cache *ca)
215 {
216 	struct bucket *b;
217 	size_t checked = 0;
218 
219 	while (!fifo_full(&ca->free_inc)) {
220 		if (ca->fifo_last_bucket <  ca->sb.first_bucket ||
221 		    ca->fifo_last_bucket >= ca->sb.nbuckets)
222 			ca->fifo_last_bucket = ca->sb.first_bucket;
223 
224 		b = ca->buckets + ca->fifo_last_bucket++;
225 
226 		if (bch_can_invalidate_bucket(ca, b))
227 			bch_invalidate_one_bucket(ca, b);
228 
229 		if (++checked >= ca->sb.nbuckets) {
230 			ca->invalidate_needs_gc = 1;
231 			wake_up_gc(ca->set);
232 			return;
233 		}
234 	}
235 }
236 
237 static void invalidate_buckets_random(struct cache *ca)
238 {
239 	struct bucket *b;
240 	size_t checked = 0;
241 
242 	while (!fifo_full(&ca->free_inc)) {
243 		size_t n;
244 
245 		get_random_bytes(&n, sizeof(n));
246 
247 		n %= (size_t) (ca->sb.nbuckets - ca->sb.first_bucket);
248 		n += ca->sb.first_bucket;
249 
250 		b = ca->buckets + n;
251 
252 		if (bch_can_invalidate_bucket(ca, b))
253 			bch_invalidate_one_bucket(ca, b);
254 
255 		if (++checked >= ca->sb.nbuckets / 2) {
256 			ca->invalidate_needs_gc = 1;
257 			wake_up_gc(ca->set);
258 			return;
259 		}
260 	}
261 }
262 
263 static void invalidate_buckets(struct cache *ca)
264 {
265 	BUG_ON(ca->invalidate_needs_gc);
266 
267 	switch (CACHE_REPLACEMENT(&ca->sb)) {
268 	case CACHE_REPLACEMENT_LRU:
269 		invalidate_buckets_lru(ca);
270 		break;
271 	case CACHE_REPLACEMENT_FIFO:
272 		invalidate_buckets_fifo(ca);
273 		break;
274 	case CACHE_REPLACEMENT_RANDOM:
275 		invalidate_buckets_random(ca);
276 		break;
277 	}
278 }
279 
280 #define allocator_wait(ca, cond)					\
281 do {									\
282 	while (1) {							\
283 		set_current_state(TASK_INTERRUPTIBLE);			\
284 		if (cond)						\
285 			break;						\
286 									\
287 		mutex_unlock(&(ca)->set->bucket_lock);			\
288 		if (kthread_should_stop() ||				\
289 		    test_bit(CACHE_SET_IO_DISABLE, &ca->set->flags)) {	\
290 			set_current_state(TASK_RUNNING);		\
291 			goto out;					\
292 		}							\
293 									\
294 		schedule();						\
295 		mutex_lock(&(ca)->set->bucket_lock);			\
296 	}								\
297 	__set_current_state(TASK_RUNNING);				\
298 } while (0)
299 
300 static int bch_allocator_push(struct cache *ca, long bucket)
301 {
302 	unsigned int i;
303 
304 	/* Prios/gens are actually the most important reserve */
305 	if (fifo_push(&ca->free[RESERVE_PRIO], bucket))
306 		return true;
307 
308 	for (i = 0; i < RESERVE_NR; i++)
309 		if (fifo_push(&ca->free[i], bucket))
310 			return true;
311 
312 	return false;
313 }
314 
315 static int bch_allocator_thread(void *arg)
316 {
317 	struct cache *ca = arg;
318 
319 	mutex_lock(&ca->set->bucket_lock);
320 
321 	while (1) {
322 		/*
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
325 		 * the free list:
326 		 */
327 		while (1) {
328 			long bucket;
329 
330 			if (!fifo_pop(&ca->free_inc, bucket))
331 				break;
332 
333 			if (ca->discard) {
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);
338 				mutex_lock(&ca->set->bucket_lock);
339 			}
340 
341 			allocator_wait(ca, bch_allocator_push(ca, bucket));
342 			wake_up(&ca->set->btree_cache_wait);
343 			wake_up(&ca->set->bucket_wait);
344 		}
345 
346 		/*
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:
350 		 */
351 
352 retry_invalidate:
353 		allocator_wait(ca, !ca->invalidate_needs_gc);
354 		invalidate_buckets(ca);
355 
356 		/*
357 		 * Now, we write their new gens to disk so we can start writing
358 		 * new stuff to them:
359 		 */
360 		allocator_wait(ca, !atomic_read(&ca->set->prio_blocked));
361 		if (CACHE_SYNC(&ca->sb)) {
362 			/*
363 			 * This could deadlock if an allocation with a btree
364 			 * node locked ever blocked - having the btree node
365 			 * locked would block garbage collection, but here we're
366 			 * waiting on garbage collection before we invalidate
367 			 * and free anything.
368 			 *
369 			 * But this should be safe since the btree code always
370 			 * uses btree_check_reserve() before allocating now, and
371 			 * if it fails it blocks without btree nodes locked.
372 			 */
373 			if (!fifo_full(&ca->free_inc))
374 				goto retry_invalidate;
375 
376 			if (bch_prio_write(ca, false) < 0) {
377 				ca->invalidate_needs_gc = 1;
378 				wake_up_gc(ca->set);
379 			}
380 		}
381 	}
382 out:
383 	wait_for_kthread_stop();
384 	return 0;
385 }
386 
387 /* Allocation */
388 
389 long bch_bucket_alloc(struct cache *ca, unsigned int reserve, bool wait)
390 {
391 	DEFINE_WAIT(w);
392 	struct bucket *b;
393 	long r;
394 
395 
396 	/* No allocation if CACHE_SET_IO_DISABLE bit is set */
397 	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &ca->set->flags)))
398 		return -1;
399 
400 	/* fastpath */
401 	if (fifo_pop(&ca->free[RESERVE_NONE], r) ||
402 	    fifo_pop(&ca->free[reserve], r))
403 		goto out;
404 
405 	if (!wait) {
406 		trace_bcache_alloc_fail(ca, reserve);
407 		return -1;
408 	}
409 
410 	do {
411 		prepare_to_wait(&ca->set->bucket_wait, &w,
412 				TASK_UNINTERRUPTIBLE);
413 
414 		mutex_unlock(&ca->set->bucket_lock);
415 		schedule();
416 		mutex_lock(&ca->set->bucket_lock);
417 	} while (!fifo_pop(&ca->free[RESERVE_NONE], r) &&
418 		 !fifo_pop(&ca->free[reserve], r));
419 
420 	finish_wait(&ca->set->bucket_wait, &w);
421 out:
422 	if (ca->alloc_thread)
423 		wake_up_process(ca->alloc_thread);
424 
425 	trace_bcache_alloc(ca, reserve);
426 
427 	if (expensive_debug_checks(ca->set)) {
428 		size_t iter;
429 		long i;
430 		unsigned int j;
431 
432 		for (iter = 0; iter < prio_buckets(ca) * 2; iter++)
433 			BUG_ON(ca->prio_buckets[iter] == (uint64_t) r);
434 
435 		for (j = 0; j < RESERVE_NR; j++)
436 			fifo_for_each(i, &ca->free[j], iter)
437 				BUG_ON(i == r);
438 		fifo_for_each(i, &ca->free_inc, iter)
439 			BUG_ON(i == r);
440 	}
441 
442 	b = ca->buckets + r;
443 
444 	BUG_ON(atomic_read(&b->pin) != 1);
445 
446 	SET_GC_SECTORS_USED(b, ca->sb.bucket_size);
447 
448 	if (reserve <= RESERVE_PRIO) {
449 		SET_GC_MARK(b, GC_MARK_METADATA);
450 		SET_GC_MOVE(b, 0);
451 		b->prio = BTREE_PRIO;
452 	} else {
453 		SET_GC_MARK(b, GC_MARK_RECLAIMABLE);
454 		SET_GC_MOVE(b, 0);
455 		b->prio = INITIAL_PRIO;
456 	}
457 
458 	if (ca->set->avail_nbuckets > 0) {
459 		ca->set->avail_nbuckets--;
460 		bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
461 	}
462 
463 	return r;
464 }
465 
466 void __bch_bucket_free(struct cache *ca, struct bucket *b)
467 {
468 	SET_GC_MARK(b, 0);
469 	SET_GC_SECTORS_USED(b, 0);
470 
471 	if (ca->set->avail_nbuckets < ca->set->nbuckets) {
472 		ca->set->avail_nbuckets++;
473 		bch_update_bucket_in_use(ca->set, &ca->set->gc_stats);
474 	}
475 }
476 
477 void bch_bucket_free(struct cache_set *c, struct bkey *k)
478 {
479 	unsigned int i;
480 
481 	for (i = 0; i < KEY_PTRS(k); i++)
482 		__bch_bucket_free(c->cache, PTR_BUCKET(c, k, i));
483 }
484 
485 int __bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
486 			   struct bkey *k, bool wait)
487 {
488 	struct cache *ca;
489 	long b;
490 
491 	/* No allocation if CACHE_SET_IO_DISABLE bit is set */
492 	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
493 		return -1;
494 
495 	lockdep_assert_held(&c->bucket_lock);
496 
497 	bkey_init(k);
498 
499 	ca = c->cache;
500 	b = bch_bucket_alloc(ca, reserve, wait);
501 	if (b < 0)
502 		return -1;
503 
504 	k->ptr[0] = MAKE_PTR(ca->buckets[b].gen,
505 			     bucket_to_sector(c, b),
506 			     ca->sb.nr_this_dev);
507 
508 	SET_KEY_PTRS(k, 1);
509 
510 	return 0;
511 }
512 
513 int bch_bucket_alloc_set(struct cache_set *c, unsigned int reserve,
514 			 struct bkey *k, bool wait)
515 {
516 	int ret;
517 
518 	mutex_lock(&c->bucket_lock);
519 	ret = __bch_bucket_alloc_set(c, reserve, k, wait);
520 	mutex_unlock(&c->bucket_lock);
521 	return ret;
522 }
523 
524 /* Sector allocator */
525 
526 struct open_bucket {
527 	struct list_head	list;
528 	unsigned int		last_write_point;
529 	unsigned int		sectors_free;
530 	BKEY_PADDED(key);
531 };
532 
533 /*
534  * We keep multiple buckets open for writes, and try to segregate different
535  * write streams for better cache utilization: first we try to segregate flash
536  * only volume write streams from cached devices, secondly we look for a bucket
537  * where the last write to it was sequential with the current write, and
538  * failing that we look for a bucket that was last used by the same task.
539  *
540  * The ideas is if you've got multiple tasks pulling data into the cache at the
541  * same time, you'll get better cache utilization if you try to segregate their
542  * data and preserve locality.
543  *
544  * For example, dirty sectors of flash only volume is not reclaimable, if their
545  * dirty sectors mixed with dirty sectors of cached device, such buckets will
546  * be marked as dirty and won't be reclaimed, though the dirty data of cached
547  * device have been written back to backend device.
548  *
549  * And say you've starting Firefox at the same time you're copying a
550  * bunch of files. Firefox will likely end up being fairly hot and stay in the
551  * cache awhile, but the data you copied might not be; if you wrote all that
552  * data to the same buckets it'd get invalidated at the same time.
553  *
554  * Both of those tasks will be doing fairly random IO so we can't rely on
555  * detecting sequential IO to segregate their data, but going off of the task
556  * should be a sane heuristic.
557  */
558 static struct open_bucket *pick_data_bucket(struct cache_set *c,
559 					    const struct bkey *search,
560 					    unsigned int write_point,
561 					    struct bkey *alloc)
562 {
563 	struct open_bucket *ret, *ret_task = NULL;
564 
565 	list_for_each_entry_reverse(ret, &c->data_buckets, list)
566 		if (UUID_FLASH_ONLY(&c->uuids[KEY_INODE(&ret->key)]) !=
567 		    UUID_FLASH_ONLY(&c->uuids[KEY_INODE(search)]))
568 			continue;
569 		else if (!bkey_cmp(&ret->key, search))
570 			goto found;
571 		else if (ret->last_write_point == write_point)
572 			ret_task = ret;
573 
574 	ret = ret_task ?: list_first_entry(&c->data_buckets,
575 					   struct open_bucket, list);
576 found:
577 	if (!ret->sectors_free && KEY_PTRS(alloc)) {
578 		ret->sectors_free = c->cache->sb.bucket_size;
579 		bkey_copy(&ret->key, alloc);
580 		bkey_init(alloc);
581 	}
582 
583 	if (!ret->sectors_free)
584 		ret = NULL;
585 
586 	return ret;
587 }
588 
589 /*
590  * Allocates some space in the cache to write to, and k to point to the newly
591  * allocated space, and updates KEY_SIZE(k) and KEY_OFFSET(k) (to point to the
592  * end of the newly allocated space).
593  *
594  * May allocate fewer sectors than @sectors, KEY_SIZE(k) indicates how many
595  * sectors were actually allocated.
596  *
597  * If s->writeback is true, will not fail.
598  */
599 bool bch_alloc_sectors(struct cache_set *c,
600 		       struct bkey *k,
601 		       unsigned int sectors,
602 		       unsigned int write_point,
603 		       unsigned int write_prio,
604 		       bool wait)
605 {
606 	struct open_bucket *b;
607 	BKEY_PADDED(key) alloc;
608 	unsigned int i;
609 
610 	/*
611 	 * We might have to allocate a new bucket, which we can't do with a
612 	 * spinlock held. So if we have to allocate, we drop the lock, allocate
613 	 * and then retry. KEY_PTRS() indicates whether alloc points to
614 	 * allocated bucket(s).
615 	 */
616 
617 	bkey_init(&alloc.key);
618 	spin_lock(&c->data_bucket_lock);
619 
620 	while (!(b = pick_data_bucket(c, k, write_point, &alloc.key))) {
621 		unsigned int watermark = write_prio
622 			? RESERVE_MOVINGGC
623 			: RESERVE_NONE;
624 
625 		spin_unlock(&c->data_bucket_lock);
626 
627 		if (bch_bucket_alloc_set(c, watermark, &alloc.key, wait))
628 			return false;
629 
630 		spin_lock(&c->data_bucket_lock);
631 	}
632 
633 	/*
634 	 * If we had to allocate, we might race and not need to allocate the
635 	 * second time we call pick_data_bucket(). If we allocated a bucket but
636 	 * didn't use it, drop the refcount bch_bucket_alloc_set() took:
637 	 */
638 	if (KEY_PTRS(&alloc.key))
639 		bkey_put(c, &alloc.key);
640 
641 	for (i = 0; i < KEY_PTRS(&b->key); i++)
642 		EBUG_ON(ptr_stale(c, &b->key, i));
643 
644 	/* Set up the pointer to the space we're allocating: */
645 
646 	for (i = 0; i < KEY_PTRS(&b->key); i++)
647 		k->ptr[i] = b->key.ptr[i];
648 
649 	sectors = min(sectors, b->sectors_free);
650 
651 	SET_KEY_OFFSET(k, KEY_OFFSET(k) + sectors);
652 	SET_KEY_SIZE(k, sectors);
653 	SET_KEY_PTRS(k, KEY_PTRS(&b->key));
654 
655 	/*
656 	 * Move b to the end of the lru, and keep track of what this bucket was
657 	 * last used for:
658 	 */
659 	list_move_tail(&b->list, &c->data_buckets);
660 	bkey_copy_key(&b->key, k);
661 	b->last_write_point = write_point;
662 
663 	b->sectors_free	-= sectors;
664 
665 	for (i = 0; i < KEY_PTRS(&b->key); i++) {
666 		SET_PTR_OFFSET(&b->key, i, PTR_OFFSET(&b->key, i) + sectors);
667 
668 		atomic_long_add(sectors,
669 				&c->cache->sectors_written);
670 	}
671 
672 	if (b->sectors_free < c->cache->sb.block_size)
673 		b->sectors_free = 0;
674 
675 	/*
676 	 * k takes refcounts on the buckets it points to until it's inserted
677 	 * into the btree, but if we're done with this bucket we just transfer
678 	 * get_data_bucket()'s refcount.
679 	 */
680 	if (b->sectors_free)
681 		for (i = 0; i < KEY_PTRS(&b->key); i++)
682 			atomic_inc(&PTR_BUCKET(c, &b->key, i)->pin);
683 
684 	spin_unlock(&c->data_bucket_lock);
685 	return true;
686 }
687 
688 /* Init */
689 
690 void bch_open_buckets_free(struct cache_set *c)
691 {
692 	struct open_bucket *b;
693 
694 	while (!list_empty(&c->data_buckets)) {
695 		b = list_first_entry(&c->data_buckets,
696 				     struct open_bucket, list);
697 		list_del(&b->list);
698 		kfree(b);
699 	}
700 }
701 
702 int bch_open_buckets_alloc(struct cache_set *c)
703 {
704 	int i;
705 
706 	spin_lock_init(&c->data_bucket_lock);
707 
708 	for (i = 0; i < MAX_OPEN_BUCKETS; i++) {
709 		struct open_bucket *b = kzalloc(sizeof(*b), GFP_KERNEL);
710 
711 		if (!b)
712 			return -ENOMEM;
713 
714 		list_add(&b->list, &c->data_buckets);
715 	}
716 
717 	return 0;
718 }
719 
720 int bch_cache_allocator_start(struct cache *ca)
721 {
722 	struct task_struct *k = kthread_run(bch_allocator_thread,
723 					    ca, "bcache_allocator");
724 	if (IS_ERR(k))
725 		return PTR_ERR(k);
726 
727 	ca->alloc_thread = k;
728 	return 0;
729 }
730