xref: /openbmc/linux/drivers/md/bcache/journal.c (revision 39b6f3aa)
1 /*
2  * bcache journalling code, for btree insertions
3  *
4  * Copyright 2012 Google, Inc.
5  */
6 
7 #include "bcache.h"
8 #include "btree.h"
9 #include "debug.h"
10 #include "request.h"
11 
12 /*
13  * Journal replay/recovery:
14  *
15  * This code is all driven from run_cache_set(); we first read the journal
16  * entries, do some other stuff, then we mark all the keys in the journal
17  * entries (same as garbage collection would), then we replay them - reinserting
18  * them into the cache in precisely the same order as they appear in the
19  * journal.
20  *
21  * We only journal keys that go in leaf nodes, which simplifies things quite a
22  * bit.
23  */
24 
25 static void journal_read_endio(struct bio *bio, int error)
26 {
27 	struct closure *cl = bio->bi_private;
28 	closure_put(cl);
29 }
30 
31 static int journal_read_bucket(struct cache *ca, struct list_head *list,
32 			       struct btree_op *op, unsigned bucket_index)
33 {
34 	struct journal_device *ja = &ca->journal;
35 	struct bio *bio = &ja->bio;
36 
37 	struct journal_replay *i;
38 	struct jset *j, *data = ca->set->journal.w[0].data;
39 	unsigned len, left, offset = 0;
40 	int ret = 0;
41 	sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
42 
43 	pr_debug("reading %llu", (uint64_t) bucket);
44 
45 	while (offset < ca->sb.bucket_size) {
46 reread:		left = ca->sb.bucket_size - offset;
47 		len = min_t(unsigned, left, PAGE_SECTORS * 8);
48 
49 		bio_reset(bio);
50 		bio->bi_sector	= bucket + offset;
51 		bio->bi_bdev	= ca->bdev;
52 		bio->bi_rw	= READ;
53 		bio->bi_size	= len << 9;
54 
55 		bio->bi_end_io	= journal_read_endio;
56 		bio->bi_private = &op->cl;
57 		bch_bio_map(bio, data);
58 
59 		closure_bio_submit(bio, &op->cl, ca);
60 		closure_sync(&op->cl);
61 
62 		/* This function could be simpler now since we no longer write
63 		 * journal entries that overlap bucket boundaries; this means
64 		 * the start of a bucket will always have a valid journal entry
65 		 * if it has any journal entries at all.
66 		 */
67 
68 		j = data;
69 		while (len) {
70 			struct list_head *where;
71 			size_t blocks, bytes = set_bytes(j);
72 
73 			if (j->magic != jset_magic(ca->set))
74 				return ret;
75 
76 			if (bytes > left << 9)
77 				return ret;
78 
79 			if (bytes > len << 9)
80 				goto reread;
81 
82 			if (j->csum != csum_set(j))
83 				return ret;
84 
85 			blocks = set_blocks(j, ca->set);
86 
87 			while (!list_empty(list)) {
88 				i = list_first_entry(list,
89 					struct journal_replay, list);
90 				if (i->j.seq >= j->last_seq)
91 					break;
92 				list_del(&i->list);
93 				kfree(i);
94 			}
95 
96 			list_for_each_entry_reverse(i, list, list) {
97 				if (j->seq == i->j.seq)
98 					goto next_set;
99 
100 				if (j->seq < i->j.last_seq)
101 					goto next_set;
102 
103 				if (j->seq > i->j.seq) {
104 					where = &i->list;
105 					goto add;
106 				}
107 			}
108 
109 			where = list;
110 add:
111 			i = kmalloc(offsetof(struct journal_replay, j) +
112 				    bytes, GFP_KERNEL);
113 			if (!i)
114 				return -ENOMEM;
115 			memcpy(&i->j, j, bytes);
116 			list_add(&i->list, where);
117 			ret = 1;
118 
119 			ja->seq[bucket_index] = j->seq;
120 next_set:
121 			offset	+= blocks * ca->sb.block_size;
122 			len	-= blocks * ca->sb.block_size;
123 			j = ((void *) j) + blocks * block_bytes(ca);
124 		}
125 	}
126 
127 	return ret;
128 }
129 
130 int bch_journal_read(struct cache_set *c, struct list_head *list,
131 			struct btree_op *op)
132 {
133 #define read_bucket(b)							\
134 	({								\
135 		int ret = journal_read_bucket(ca, list, op, b);		\
136 		__set_bit(b, bitmap);					\
137 		if (ret < 0)						\
138 			return ret;					\
139 		ret;							\
140 	})
141 
142 	struct cache *ca;
143 	unsigned iter;
144 
145 	for_each_cache(ca, c, iter) {
146 		struct journal_device *ja = &ca->journal;
147 		unsigned long bitmap[SB_JOURNAL_BUCKETS / BITS_PER_LONG];
148 		unsigned i, l, r, m;
149 		uint64_t seq;
150 
151 		bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
152 		pr_debug("%u journal buckets", ca->sb.njournal_buckets);
153 
154 		/* Read journal buckets ordered by golden ratio hash to quickly
155 		 * find a sequence of buckets with valid journal entries
156 		 */
157 		for (i = 0; i < ca->sb.njournal_buckets; i++) {
158 			l = (i * 2654435769U) % ca->sb.njournal_buckets;
159 
160 			if (test_bit(l, bitmap))
161 				break;
162 
163 			if (read_bucket(l))
164 				goto bsearch;
165 		}
166 
167 		/* If that fails, check all the buckets we haven't checked
168 		 * already
169 		 */
170 		pr_debug("falling back to linear search");
171 
172 		for (l = 0; l < ca->sb.njournal_buckets; l++) {
173 			if (test_bit(l, bitmap))
174 				continue;
175 
176 			if (read_bucket(l))
177 				goto bsearch;
178 		}
179 bsearch:
180 		/* Binary search */
181 		m = r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
182 		pr_debug("starting binary search, l %u r %u", l, r);
183 
184 		while (l + 1 < r) {
185 			m = (l + r) >> 1;
186 
187 			if (read_bucket(m))
188 				l = m;
189 			else
190 				r = m;
191 		}
192 
193 		/* Read buckets in reverse order until we stop finding more
194 		 * journal entries
195 		 */
196 		pr_debug("finishing up");
197 		l = m;
198 
199 		while (1) {
200 			if (!l--)
201 				l = ca->sb.njournal_buckets - 1;
202 
203 			if (l == m)
204 				break;
205 
206 			if (test_bit(l, bitmap))
207 				continue;
208 
209 			if (!read_bucket(l))
210 				break;
211 		}
212 
213 		seq = 0;
214 
215 		for (i = 0; i < ca->sb.njournal_buckets; i++)
216 			if (ja->seq[i] > seq) {
217 				seq = ja->seq[i];
218 				ja->cur_idx = ja->discard_idx =
219 					ja->last_idx = i;
220 
221 			}
222 	}
223 
224 	c->journal.seq = list_entry(list->prev,
225 				    struct journal_replay,
226 				    list)->j.seq;
227 
228 	return 0;
229 #undef read_bucket
230 }
231 
232 void bch_journal_mark(struct cache_set *c, struct list_head *list)
233 {
234 	atomic_t p = { 0 };
235 	struct bkey *k;
236 	struct journal_replay *i;
237 	struct journal *j = &c->journal;
238 	uint64_t last = j->seq;
239 
240 	/*
241 	 * journal.pin should never fill up - we never write a journal
242 	 * entry when it would fill up. But if for some reason it does, we
243 	 * iterate over the list in reverse order so that we can just skip that
244 	 * refcount instead of bugging.
245 	 */
246 
247 	list_for_each_entry_reverse(i, list, list) {
248 		BUG_ON(last < i->j.seq);
249 		i->pin = NULL;
250 
251 		while (last-- != i->j.seq)
252 			if (fifo_free(&j->pin) > 1) {
253 				fifo_push_front(&j->pin, p);
254 				atomic_set(&fifo_front(&j->pin), 0);
255 			}
256 
257 		if (fifo_free(&j->pin) > 1) {
258 			fifo_push_front(&j->pin, p);
259 			i->pin = &fifo_front(&j->pin);
260 			atomic_set(i->pin, 1);
261 		}
262 
263 		for (k = i->j.start;
264 		     k < end(&i->j);
265 		     k = bkey_next(k)) {
266 			unsigned j;
267 
268 			for (j = 0; j < KEY_PTRS(k); j++) {
269 				struct bucket *g = PTR_BUCKET(c, k, j);
270 				atomic_inc(&g->pin);
271 
272 				if (g->prio == BTREE_PRIO &&
273 				    !ptr_stale(c, k, j))
274 					g->prio = INITIAL_PRIO;
275 			}
276 
277 			__bch_btree_mark_key(c, 0, k);
278 		}
279 	}
280 }
281 
282 int bch_journal_replay(struct cache_set *s, struct list_head *list,
283 			  struct btree_op *op)
284 {
285 	int ret = 0, keys = 0, entries = 0;
286 	struct bkey *k;
287 	struct journal_replay *i =
288 		list_entry(list->prev, struct journal_replay, list);
289 
290 	uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
291 
292 	list_for_each_entry(i, list, list) {
293 		BUG_ON(i->pin && atomic_read(i->pin) != 1);
294 
295 		if (n != i->j.seq)
296 			pr_err(
297 		"journal entries %llu-%llu missing! (replaying %llu-%llu)\n",
298 		n, i->j.seq - 1, start, end);
299 
300 		for (k = i->j.start;
301 		     k < end(&i->j);
302 		     k = bkey_next(k)) {
303 			pr_debug("%s", pkey(k));
304 			bkey_copy(op->keys.top, k);
305 			bch_keylist_push(&op->keys);
306 
307 			op->journal = i->pin;
308 			atomic_inc(op->journal);
309 
310 			ret = bch_btree_insert(op, s);
311 			if (ret)
312 				goto err;
313 
314 			BUG_ON(!bch_keylist_empty(&op->keys));
315 			keys++;
316 
317 			cond_resched();
318 		}
319 
320 		if (i->pin)
321 			atomic_dec(i->pin);
322 		n = i->j.seq + 1;
323 		entries++;
324 	}
325 
326 	pr_info("journal replay done, %i keys in %i entries, seq %llu",
327 		keys, entries, end);
328 
329 	while (!list_empty(list)) {
330 		i = list_first_entry(list, struct journal_replay, list);
331 		list_del(&i->list);
332 		kfree(i);
333 	}
334 err:
335 	closure_sync(&op->cl);
336 	return ret;
337 }
338 
339 /* Journalling */
340 
341 static void btree_flush_write(struct cache_set *c)
342 {
343 	/*
344 	 * Try to find the btree node with that references the oldest journal
345 	 * entry, best is our current candidate and is locked if non NULL:
346 	 */
347 	struct btree *b, *best = NULL;
348 	unsigned iter;
349 
350 	for_each_cached_btree(b, c, iter) {
351 		if (!down_write_trylock(&b->lock))
352 			continue;
353 
354 		if (!btree_node_dirty(b) ||
355 		    !btree_current_write(b)->journal) {
356 			rw_unlock(true, b);
357 			continue;
358 		}
359 
360 		if (!best)
361 			best = b;
362 		else if (journal_pin_cmp(c,
363 					 btree_current_write(best),
364 					 btree_current_write(b))) {
365 			rw_unlock(true, best);
366 			best = b;
367 		} else
368 			rw_unlock(true, b);
369 	}
370 
371 	if (best)
372 		goto out;
373 
374 	/* We can't find the best btree node, just pick the first */
375 	list_for_each_entry(b, &c->btree_cache, list)
376 		if (!b->level && btree_node_dirty(b)) {
377 			best = b;
378 			rw_lock(true, best, best->level);
379 			goto found;
380 		}
381 
382 out:
383 	if (!best)
384 		return;
385 found:
386 	if (btree_node_dirty(best))
387 		bch_btree_write(best, true, NULL);
388 	rw_unlock(true, best);
389 }
390 
391 #define last_seq(j)	((j)->seq - fifo_used(&(j)->pin) + 1)
392 
393 static void journal_discard_endio(struct bio *bio, int error)
394 {
395 	struct journal_device *ja =
396 		container_of(bio, struct journal_device, discard_bio);
397 	struct cache *ca = container_of(ja, struct cache, journal);
398 
399 	atomic_set(&ja->discard_in_flight, DISCARD_DONE);
400 
401 	closure_wake_up(&ca->set->journal.wait);
402 	closure_put(&ca->set->cl);
403 }
404 
405 static void journal_discard_work(struct work_struct *work)
406 {
407 	struct journal_device *ja =
408 		container_of(work, struct journal_device, discard_work);
409 
410 	submit_bio(0, &ja->discard_bio);
411 }
412 
413 static void do_journal_discard(struct cache *ca)
414 {
415 	struct journal_device *ja = &ca->journal;
416 	struct bio *bio = &ja->discard_bio;
417 
418 	if (!ca->discard) {
419 		ja->discard_idx = ja->last_idx;
420 		return;
421 	}
422 
423 	switch (atomic_read(&ja->discard_in_flight) == DISCARD_IN_FLIGHT) {
424 	case DISCARD_IN_FLIGHT:
425 		return;
426 
427 	case DISCARD_DONE:
428 		ja->discard_idx = (ja->discard_idx + 1) %
429 			ca->sb.njournal_buckets;
430 
431 		atomic_set(&ja->discard_in_flight, DISCARD_READY);
432 		/* fallthrough */
433 
434 	case DISCARD_READY:
435 		if (ja->discard_idx == ja->last_idx)
436 			return;
437 
438 		atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
439 
440 		bio_init(bio);
441 		bio->bi_sector		= bucket_to_sector(ca->set,
442 						ca->sb.d[ja->discard_idx]);
443 		bio->bi_bdev		= ca->bdev;
444 		bio->bi_rw		= REQ_WRITE|REQ_DISCARD;
445 		bio->bi_max_vecs	= 1;
446 		bio->bi_io_vec		= bio->bi_inline_vecs;
447 		bio->bi_size		= bucket_bytes(ca);
448 		bio->bi_end_io		= journal_discard_endio;
449 
450 		closure_get(&ca->set->cl);
451 		INIT_WORK(&ja->discard_work, journal_discard_work);
452 		schedule_work(&ja->discard_work);
453 	}
454 }
455 
456 static void journal_reclaim(struct cache_set *c)
457 {
458 	struct bkey *k = &c->journal.key;
459 	struct cache *ca;
460 	uint64_t last_seq;
461 	unsigned iter, n = 0;
462 	atomic_t p;
463 
464 	while (!atomic_read(&fifo_front(&c->journal.pin)))
465 		fifo_pop(&c->journal.pin, p);
466 
467 	last_seq = last_seq(&c->journal);
468 
469 	/* Update last_idx */
470 
471 	for_each_cache(ca, c, iter) {
472 		struct journal_device *ja = &ca->journal;
473 
474 		while (ja->last_idx != ja->cur_idx &&
475 		       ja->seq[ja->last_idx] < last_seq)
476 			ja->last_idx = (ja->last_idx + 1) %
477 				ca->sb.njournal_buckets;
478 	}
479 
480 	for_each_cache(ca, c, iter)
481 		do_journal_discard(ca);
482 
483 	if (c->journal.blocks_free)
484 		return;
485 
486 	/*
487 	 * Allocate:
488 	 * XXX: Sort by free journal space
489 	 */
490 
491 	for_each_cache(ca, c, iter) {
492 		struct journal_device *ja = &ca->journal;
493 		unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
494 
495 		/* No space available on this device */
496 		if (next == ja->discard_idx)
497 			continue;
498 
499 		ja->cur_idx = next;
500 		k->ptr[n++] = PTR(0,
501 				  bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
502 				  ca->sb.nr_this_dev);
503 	}
504 
505 	bkey_init(k);
506 	SET_KEY_PTRS(k, n);
507 
508 	if (n)
509 		c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
510 
511 	if (!journal_full(&c->journal))
512 		__closure_wake_up(&c->journal.wait);
513 }
514 
515 void bch_journal_next(struct journal *j)
516 {
517 	atomic_t p = { 1 };
518 
519 	j->cur = (j->cur == j->w)
520 		? &j->w[1]
521 		: &j->w[0];
522 
523 	/*
524 	 * The fifo_push() needs to happen at the same time as j->seq is
525 	 * incremented for last_seq() to be calculated correctly
526 	 */
527 	BUG_ON(!fifo_push(&j->pin, p));
528 	atomic_set(&fifo_back(&j->pin), 1);
529 
530 	j->cur->data->seq	= ++j->seq;
531 	j->cur->need_write	= false;
532 	j->cur->data->keys	= 0;
533 
534 	if (fifo_full(&j->pin))
535 		pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
536 }
537 
538 static void journal_write_endio(struct bio *bio, int error)
539 {
540 	struct journal_write *w = bio->bi_private;
541 
542 	cache_set_err_on(error, w->c, "journal io error");
543 	closure_put(&w->c->journal.io.cl);
544 }
545 
546 static void journal_write(struct closure *);
547 
548 static void journal_write_done(struct closure *cl)
549 {
550 	struct journal *j = container_of(cl, struct journal, io.cl);
551 	struct cache_set *c = container_of(j, struct cache_set, journal);
552 
553 	struct journal_write *w = (j->cur == j->w)
554 		? &j->w[1]
555 		: &j->w[0];
556 
557 	__closure_wake_up(&w->wait);
558 
559 	if (c->journal_delay_ms)
560 		closure_delay(&j->io, msecs_to_jiffies(c->journal_delay_ms));
561 
562 	continue_at(cl, journal_write, system_wq);
563 }
564 
565 static void journal_write_unlocked(struct closure *cl)
566 	__releases(c->journal.lock)
567 {
568 	struct cache_set *c = container_of(cl, struct cache_set, journal.io.cl);
569 	struct cache *ca;
570 	struct journal_write *w = c->journal.cur;
571 	struct bkey *k = &c->journal.key;
572 	unsigned i, sectors = set_blocks(w->data, c) * c->sb.block_size;
573 
574 	struct bio *bio;
575 	struct bio_list list;
576 	bio_list_init(&list);
577 
578 	if (!w->need_write) {
579 		/*
580 		 * XXX: have to unlock closure before we unlock journal lock,
581 		 * else we race with bch_journal(). But this way we race
582 		 * against cache set unregister. Doh.
583 		 */
584 		set_closure_fn(cl, NULL, NULL);
585 		closure_sub(cl, CLOSURE_RUNNING + 1);
586 		spin_unlock(&c->journal.lock);
587 		return;
588 	} else if (journal_full(&c->journal)) {
589 		journal_reclaim(c);
590 		spin_unlock(&c->journal.lock);
591 
592 		btree_flush_write(c);
593 		continue_at(cl, journal_write, system_wq);
594 	}
595 
596 	c->journal.blocks_free -= set_blocks(w->data, c);
597 
598 	w->data->btree_level = c->root->level;
599 
600 	bkey_copy(&w->data->btree_root, &c->root->key);
601 	bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
602 
603 	for_each_cache(ca, c, i)
604 		w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
605 
606 	w->data->magic		= jset_magic(c);
607 	w->data->version	= BCACHE_JSET_VERSION;
608 	w->data->last_seq	= last_seq(&c->journal);
609 	w->data->csum		= csum_set(w->data);
610 
611 	for (i = 0; i < KEY_PTRS(k); i++) {
612 		ca = PTR_CACHE(c, k, i);
613 		bio = &ca->journal.bio;
614 
615 		atomic_long_add(sectors, &ca->meta_sectors_written);
616 
617 		bio_reset(bio);
618 		bio->bi_sector	= PTR_OFFSET(k, i);
619 		bio->bi_bdev	= ca->bdev;
620 		bio->bi_rw	= REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH;
621 		bio->bi_size	= sectors << 9;
622 
623 		bio->bi_end_io	= journal_write_endio;
624 		bio->bi_private = w;
625 		bch_bio_map(bio, w->data);
626 
627 		trace_bcache_journal_write(bio);
628 		bio_list_add(&list, bio);
629 
630 		SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
631 
632 		ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
633 	}
634 
635 	atomic_dec_bug(&fifo_back(&c->journal.pin));
636 	bch_journal_next(&c->journal);
637 	journal_reclaim(c);
638 
639 	spin_unlock(&c->journal.lock);
640 
641 	while ((bio = bio_list_pop(&list)))
642 		closure_bio_submit(bio, cl, c->cache[0]);
643 
644 	continue_at(cl, journal_write_done, NULL);
645 }
646 
647 static void journal_write(struct closure *cl)
648 {
649 	struct cache_set *c = container_of(cl, struct cache_set, journal.io.cl);
650 
651 	spin_lock(&c->journal.lock);
652 	journal_write_unlocked(cl);
653 }
654 
655 static void __journal_try_write(struct cache_set *c, bool noflush)
656 	__releases(c->journal.lock)
657 {
658 	struct closure *cl = &c->journal.io.cl;
659 
660 	if (!closure_trylock(cl, &c->cl))
661 		spin_unlock(&c->journal.lock);
662 	else if (noflush && journal_full(&c->journal)) {
663 		spin_unlock(&c->journal.lock);
664 		continue_at(cl, journal_write, system_wq);
665 	} else
666 		journal_write_unlocked(cl);
667 }
668 
669 #define journal_try_write(c)	__journal_try_write(c, false)
670 
671 void bch_journal_meta(struct cache_set *c, struct closure *cl)
672 {
673 	struct journal_write *w;
674 
675 	if (CACHE_SYNC(&c->sb)) {
676 		spin_lock(&c->journal.lock);
677 
678 		w = c->journal.cur;
679 		w->need_write = true;
680 
681 		if (cl)
682 			BUG_ON(!closure_wait(&w->wait, cl));
683 
684 		__journal_try_write(c, true);
685 	}
686 }
687 
688 /*
689  * Entry point to the journalling code - bio_insert() and btree_invalidate()
690  * pass bch_journal() a list of keys to be journalled, and then
691  * bch_journal() hands those same keys off to btree_insert_async()
692  */
693 
694 void bch_journal(struct closure *cl)
695 {
696 	struct btree_op *op = container_of(cl, struct btree_op, cl);
697 	struct cache_set *c = op->c;
698 	struct journal_write *w;
699 	size_t b, n = ((uint64_t *) op->keys.top) - op->keys.list;
700 
701 	if (op->type != BTREE_INSERT ||
702 	    !CACHE_SYNC(&c->sb))
703 		goto out;
704 
705 	/*
706 	 * If we're looping because we errored, might already be waiting on
707 	 * another journal write:
708 	 */
709 	while (atomic_read(&cl->parent->remaining) & CLOSURE_WAITING)
710 		closure_sync(cl->parent);
711 
712 	spin_lock(&c->journal.lock);
713 
714 	if (journal_full(&c->journal)) {
715 		/* XXX: tracepoint */
716 		closure_wait(&c->journal.wait, cl);
717 
718 		journal_reclaim(c);
719 		spin_unlock(&c->journal.lock);
720 
721 		btree_flush_write(c);
722 		continue_at(cl, bch_journal, bcache_wq);
723 	}
724 
725 	w = c->journal.cur;
726 	w->need_write = true;
727 	b = __set_blocks(w->data, w->data->keys + n, c);
728 
729 	if (b * c->sb.block_size > PAGE_SECTORS << JSET_BITS ||
730 	    b > c->journal.blocks_free) {
731 		/* XXX: If we were inserting so many keys that they won't fit in
732 		 * an _empty_ journal write, we'll deadlock. For now, handle
733 		 * this in bch_keylist_realloc() - but something to think about.
734 		 */
735 		BUG_ON(!w->data->keys);
736 
737 		/* XXX: tracepoint */
738 		BUG_ON(!closure_wait(&w->wait, cl));
739 
740 		closure_flush(&c->journal.io);
741 
742 		journal_try_write(c);
743 		continue_at(cl, bch_journal, bcache_wq);
744 	}
745 
746 	memcpy(end(w->data), op->keys.list, n * sizeof(uint64_t));
747 	w->data->keys += n;
748 
749 	op->journal = &fifo_back(&c->journal.pin);
750 	atomic_inc(op->journal);
751 
752 	if (op->flush_journal) {
753 		closure_flush(&c->journal.io);
754 		closure_wait(&w->wait, cl->parent);
755 	}
756 
757 	journal_try_write(c);
758 out:
759 	bch_btree_insert_async(cl);
760 }
761 
762 void bch_journal_free(struct cache_set *c)
763 {
764 	free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
765 	free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
766 	free_fifo(&c->journal.pin);
767 }
768 
769 int bch_journal_alloc(struct cache_set *c)
770 {
771 	struct journal *j = &c->journal;
772 
773 	closure_init_unlocked(&j->io);
774 	spin_lock_init(&j->lock);
775 
776 	c->journal_delay_ms = 100;
777 
778 	j->w[0].c = c;
779 	j->w[1].c = c;
780 
781 	if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
782 	    !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
783 	    !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
784 		return -ENOMEM;
785 
786 	return 0;
787 }
788