journal.c

  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * bcache journalling code, for btree insertions
  4 *
  5 * Copyright 2012 Google, Inc.
  6 */
  7
  8#include "bcache.h"
  9#include "btree.h"
 10#include "debug.h"
 11#include "extents.h"
 12
 13#include <trace/events/bcache.h>
 14
 15/*
 16 * Journal replay/recovery:
 17 *
 18 * This code is all driven from run_cache_set(); we first read the journal
 19 * entries, do some other stuff, then we mark all the keys in the journal
 20 * entries (same as garbage collection would), then we replay them - reinserting
 21 * them into the cache in precisely the same order as they appear in the
 22 * journal.
 23 *
 24 * We only journal keys that go in leaf nodes, which simplifies things quite a
 25 * bit.
 26 */
 27
 28static void journal_read_endio(struct bio *bio)
 29{
 30	struct closure *cl = bio->bi_private;
 31
 32	closure_put(cl);
 33}
 34
 35static int journal_read_bucket(struct cache *ca, struct list_head *list,
 36			       unsigned int bucket_index)
 37{
 38	struct journal_device *ja = &ca->journal;
 39	struct bio *bio = &ja->bio;
 40
 41	struct journal_replay *i;
 42	struct jset *j, *data = ca->set->journal.w[0].data;
 43	struct closure cl;
 44	unsigned int len, left, offset = 0;
 45	int ret = 0;
 46	sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
 47
 48	closure_init_stack(&cl);
 49
 50	pr_debug("reading %u\n", bucket_index);
 51
 52	while (offset < ca->sb.bucket_size) {
 53reread:		left = ca->sb.bucket_size - offset;
 54		len = min_t(unsigned int, left, PAGE_SECTORS << JSET_BITS);
 55
 56		bio_reset(bio, ca->bdev, REQ_OP_READ);
 57		bio->bi_iter.bi_sector	= bucket + offset;
 58		bio->bi_iter.bi_size	= len << 9;
 59
 60		bio->bi_end_io	= journal_read_endio;
 61		bio->bi_private = &cl;
 62		bch_bio_map(bio, data);
 63
 64		closure_bio_submit(ca->set, bio, &cl);
 65		closure_sync(&cl);
 66
 67		/* This function could be simpler now since we no longer write
 68		 * journal entries that overlap bucket boundaries; this means
 69		 * the start of a bucket will always have a valid journal entry
 70		 * if it has any journal entries at all.
 71		 */
 72
 73		j = data;
 74		while (len) {
 75			struct list_head *where;
 76			size_t blocks, bytes = set_bytes(j);
 77
 78			if (j->magic != jset_magic(&ca->sb)) {
 79				pr_debug("%u: bad magic\n", bucket_index);
 80				return ret;
 81			}
 82
 83			if (bytes > left << 9 ||
 84			    bytes > PAGE_SIZE << JSET_BITS) {
 85				pr_info("%u: too big, %zu bytes, offset %u\n",
 86					bucket_index, bytes, offset);
 87				return ret;
 88			}
 89
 90			if (bytes > len << 9)
 91				goto reread;
 92
 93			if (j->csum != csum_set(j)) {
 94				pr_info("%u: bad csum, %zu bytes, offset %u\n",
 95					bucket_index, bytes, offset);
 96				return ret;
 97			}
 98
 99			blocks = set_blocks(j, block_bytes(ca));
100
101			/*
102			 * Nodes in 'list' are in linear increasing order of
103			 * i->j.seq, the node on head has the smallest (oldest)
104			 * journal seq, the node on tail has the biggest
105			 * (latest) journal seq.
106			 */
107
108			/*
109			 * Check from the oldest jset for last_seq. If
110			 * i->j.seq < j->last_seq, it means the oldest jset
111			 * in list is expired and useless, remove it from
112			 * this list. Otherwise, j is a candidate jset for
113			 * further following checks.
114			 */
115			while (!list_empty(list)) {
116				i = list_first_entry(list,
117					struct journal_replay, list);
118				if (i->j.seq >= j->last_seq)
119					break;
120				list_del(&i->list);
121				kfree(i);
122			}
123
124			/* iterate list in reverse order (from latest jset) */
125			list_for_each_entry_reverse(i, list, list) {
126				if (j->seq == i->j.seq)
127					goto next_set;
128
129				/*
130				 * if j->seq is less than any i->j.last_seq
131				 * in list, j is an expired and useless jset.
132				 */
133				if (j->seq < i->j.last_seq)
134					goto next_set;
135
136				/*
137				 * 'where' points to first jset in list which
138				 * is elder then j.
139				 */
140				if (j->seq > i->j.seq) {
141					where = &i->list;
142					goto add;
143				}
144			}
145
146			where = list;
147add:
148			i = kmalloc(offsetof(struct journal_replay, j) +
149				    bytes, GFP_KERNEL);
150			if (!i)
151				return -ENOMEM;
152			unsafe_memcpy(&i->j, j, bytes,
153				/* "bytes" was calculated by set_bytes() above */);
154			/* Add to the location after 'where' points to */
155			list_add(&i->list, where);
156			ret = 1;
157
158			if (j->seq > ja->seq[bucket_index])
159				ja->seq[bucket_index] = j->seq;
160next_set:
161			offset	+= blocks * ca->sb.block_size;
162			len	-= blocks * ca->sb.block_size;
163			j = ((void *) j) + blocks * block_bytes(ca);
164		}
165	}
166
167	return ret;
168}
169
170int bch_journal_read(struct cache_set *c, struct list_head *list)
171{
172#define read_bucket(b)							\
173	({								\
174		ret = journal_read_bucket(ca, list, b);			\
175		__set_bit(b, bitmap);					\
176		if (ret < 0)						\
177			return ret;					\
178		ret;							\
179	})
180
181	struct cache *ca = c->cache;
182	int ret = 0;
183	struct journal_device *ja = &ca->journal;
184	DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
185	unsigned int i, l, r, m;
186	uint64_t seq;
187
188	bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
189	pr_debug("%u journal buckets\n", ca->sb.njournal_buckets);
190
191	/*
192	 * Read journal buckets ordered by golden ratio hash to quickly
193	 * find a sequence of buckets with valid journal entries
194	 */
195	for (i = 0; i < ca->sb.njournal_buckets; i++) {
196		/*
197		 * We must try the index l with ZERO first for
198		 * correctness due to the scenario that the journal
199		 * bucket is circular buffer which might have wrapped
200		 */
201		l = (i * 2654435769U) % ca->sb.njournal_buckets;
202
203		if (test_bit(l, bitmap))
204			break;
205
206		if (read_bucket(l))
207			goto bsearch;
208	}
209
210	/*
211	 * If that fails, check all the buckets we haven't checked
212	 * already
213	 */
214	pr_debug("falling back to linear search\n");
215
216	for_each_clear_bit(l, bitmap, ca->sb.njournal_buckets)
217		if (read_bucket(l))
218			goto bsearch;
219
220	/* no journal entries on this device? */
221	if (l == ca->sb.njournal_buckets)
222		goto out;
223bsearch:
224	BUG_ON(list_empty(list));
225
226	/* Binary search */
227	m = l;
228	r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
229	pr_debug("starting binary search, l %u r %u\n", l, r);
230
231	while (l + 1 < r) {
232		seq = list_entry(list->prev, struct journal_replay,
233				 list)->j.seq;
234
235		m = (l + r) >> 1;
236		read_bucket(m);
237
238		if (seq != list_entry(list->prev, struct journal_replay,
239				      list)->j.seq)
240			l = m;
241		else
242			r = m;
243	}
244
245	/*
246	 * Read buckets in reverse order until we stop finding more
247	 * journal entries
248	 */
249	pr_debug("finishing up: m %u njournal_buckets %u\n",
250		 m, ca->sb.njournal_buckets);
251	l = m;
252
253	while (1) {
254		if (!l--)
255			l = ca->sb.njournal_buckets - 1;
256
257		if (l == m)
258			break;
259
260		if (test_bit(l, bitmap))
261			continue;
262
263		if (!read_bucket(l))
264			break;
265	}
266
267	seq = 0;
268
269	for (i = 0; i < ca->sb.njournal_buckets; i++)
270		if (ja->seq[i] > seq) {
271			seq = ja->seq[i];
272			/*
273			 * When journal_reclaim() goes to allocate for
274			 * the first time, it'll use the bucket after
275			 * ja->cur_idx
276			 */
277			ja->cur_idx = i;
278			ja->last_idx = (i + 1) % ca->sb.njournal_buckets;
279
280		}
281
282out:
283	if (!list_empty(list))
284		c->journal.seq = list_entry(list->prev,
285					    struct journal_replay,
286					    list)->j.seq;
287
288	return 0;
289#undef read_bucket
290}
291
292void bch_journal_mark(struct cache_set *c, struct list_head *list)
293{
294	atomic_t p = { 0 };
295	struct bkey *k;
296	struct journal_replay *i;
297	struct journal *j = &c->journal;
298	uint64_t last = j->seq;
299
300	/*
301	 * journal.pin should never fill up - we never write a journal
302	 * entry when it would fill up. But if for some reason it does, we
303	 * iterate over the list in reverse order so that we can just skip that
304	 * refcount instead of bugging.
305	 */
306
307	list_for_each_entry_reverse(i, list, list) {
308		BUG_ON(last < i->j.seq);
309		i->pin = NULL;
310
311		while (last-- != i->j.seq)
312			if (fifo_free(&j->pin) > 1) {
313				fifo_push_front(&j->pin, p);
314				atomic_set(&fifo_front(&j->pin), 0);
315			}
316
317		if (fifo_free(&j->pin) > 1) {
318			fifo_push_front(&j->pin, p);
319			i->pin = &fifo_front(&j->pin);
320			atomic_set(i->pin, 1);
321		}
322
323		for (k = i->j.start;
324		     k < bset_bkey_last(&i->j);
325		     k = bkey_next(k))
326			if (!__bch_extent_invalid(c, k)) {
327				unsigned int j;
328
329				for (j = 0; j < KEY_PTRS(k); j++)
330					if (ptr_available(c, k, j))
331						atomic_inc(&PTR_BUCKET(c, k, j)->pin);
332
333				bch_initial_mark_key(c, 0, k);
334			}
335	}
336}
337
338int bch_journal_replay(struct cache_set *s, struct list_head *list)
339{
340	int ret = 0, keys = 0, entries = 0;
341	struct bkey *k;
342	struct journal_replay *i =
343		list_entry(list->prev, struct journal_replay, list);
344
345	uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
346	struct keylist keylist;
347
348	list_for_each_entry(i, list, list) {
349		BUG_ON(i->pin && atomic_read(i->pin) != 1);
350
351		if (n != i->j.seq) {
352			pr_err("journal entries %llu-%llu missing! (replaying %llu-%llu)\n",
353				n, i->j.seq - 1, start, end);
354			ret = -EIO;
355			goto err;
356		}
357
358		for (k = i->j.start;
359		     k < bset_bkey_last(&i->j);
360		     k = bkey_next(k)) {
361			trace_bcache_journal_replay_key(k);
362
363			bch_keylist_init_single(&keylist, k);
364
365			ret = bch_btree_insert(s, &keylist, i->pin, NULL);
366			if (ret)
367				goto err;
368
369			BUG_ON(!bch_keylist_empty(&keylist));
370			keys++;
371
372			cond_resched();
373		}
374
375		if (i->pin)
376			atomic_dec(i->pin);
377		n = i->j.seq + 1;
378		entries++;
379	}
380
381	pr_info("journal replay done, %i keys in %i entries, seq %llu\n",
382		keys, entries, end);
383err:
384	while (!list_empty(list)) {
385		i = list_first_entry(list, struct journal_replay, list);
386		list_del(&i->list);
387		kfree(i);
388	}
389
390	return ret;
391}
392
393void bch_journal_space_reserve(struct journal *j)
394{
395	j->do_reserve = true;
396}
397
398/* Journalling */
399
400static void btree_flush_write(struct cache_set *c)
401{
402	struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR];
403	unsigned int i, nr;
404	int ref_nr;
405	atomic_t *fifo_front_p, *now_fifo_front_p;
406	size_t mask;
407
408	if (c->journal.btree_flushing)
409		return;
410
411	spin_lock(&c->journal.flush_write_lock);
412	if (c->journal.btree_flushing) {
413		spin_unlock(&c->journal.flush_write_lock);
414		return;
415	}
416	c->journal.btree_flushing = true;
417	spin_unlock(&c->journal.flush_write_lock);
418
419	/* get the oldest journal entry and check its refcount */
420	spin_lock(&c->journal.lock);
421	fifo_front_p = &fifo_front(&c->journal.pin);
422	ref_nr = atomic_read(fifo_front_p);
423	if (ref_nr <= 0) {
424		/*
425		 * do nothing if no btree node references
426		 * the oldest journal entry
427		 */
428		spin_unlock(&c->journal.lock);
429		goto out;
430	}
431	spin_unlock(&c->journal.lock);
432
433	mask = c->journal.pin.mask;
434	nr = 0;
435	atomic_long_inc(&c->flush_write);
436	memset(btree_nodes, 0, sizeof(btree_nodes));
437
438	mutex_lock(&c->bucket_lock);
439	list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
440		/*
441		 * It is safe to get now_fifo_front_p without holding
442		 * c->journal.lock here, because we don't need to know
443		 * the exactly accurate value, just check whether the
444		 * front pointer of c->journal.pin is changed.
445		 */
446		now_fifo_front_p = &fifo_front(&c->journal.pin);
447		/*
448		 * If the oldest journal entry is reclaimed and front
449		 * pointer of c->journal.pin changes, it is unnecessary
450		 * to scan c->btree_cache anymore, just quit the loop and
451		 * flush out what we have already.
452		 */
453		if (now_fifo_front_p != fifo_front_p)
454			break;
455		/*
456		 * quit this loop if all matching btree nodes are
457		 * scanned and record in btree_nodes[] already.
458		 */
459		ref_nr = atomic_read(fifo_front_p);
460		if (nr >= ref_nr)
461			break;
462
463		if (btree_node_journal_flush(b))
464			pr_err("BUG: flush_write bit should not be set here!\n");
465
466		mutex_lock(&b->write_lock);
467
468		if (!btree_node_dirty(b)) {
469			mutex_unlock(&b->write_lock);
470			continue;
471		}
472
473		if (!btree_current_write(b)->journal) {
474			mutex_unlock(&b->write_lock);
475			continue;
476		}
477
478		/*
479		 * Only select the btree node which exactly references
480		 * the oldest journal entry.
481		 *
482		 * If the journal entry pointed by fifo_front_p is
483		 * reclaimed in parallel, don't worry:
484		 * - the list_for_each_xxx loop will quit when checking
485		 *   next now_fifo_front_p.
486		 * - If there are matched nodes recorded in btree_nodes[],
487		 *   they are clean now (this is why and how the oldest
488		 *   journal entry can be reclaimed). These selected nodes
489		 *   will be ignored and skipped in the following for-loop.
490		 */
491		if (((btree_current_write(b)->journal - fifo_front_p) &
492		     mask) != 0) {
493			mutex_unlock(&b->write_lock);
494			continue;
495		}
496
497		set_btree_node_journal_flush(b);
498
499		mutex_unlock(&b->write_lock);
500
501		btree_nodes[nr++] = b;
502		/*
503		 * To avoid holding c->bucket_lock too long time,
504		 * only scan for BTREE_FLUSH_NR matched btree nodes
505		 * at most. If there are more btree nodes reference
506		 * the oldest journal entry, try to flush them next
507		 * time when btree_flush_write() is called.
508		 */
509		if (nr == BTREE_FLUSH_NR)
510			break;
511	}
512	mutex_unlock(&c->bucket_lock);
513
514	for (i = 0; i < nr; i++) {
515		b = btree_nodes[i];
516		if (!b) {
517			pr_err("BUG: btree_nodes[%d] is NULL\n", i);
518			continue;
519		}
520
521		/* safe to check without holding b->write_lock */
522		if (!btree_node_journal_flush(b)) {
523			pr_err("BUG: bnode %p: journal_flush bit cleaned\n", b);
524			continue;
525		}
526
527		mutex_lock(&b->write_lock);
528		if (!btree_current_write(b)->journal) {
529			clear_bit(BTREE_NODE_journal_flush, &b->flags);
530			mutex_unlock(&b->write_lock);
531			pr_debug("bnode %p: written by others\n", b);
532			continue;
533		}
534
535		if (!btree_node_dirty(b)) {
536			clear_bit(BTREE_NODE_journal_flush, &b->flags);
537			mutex_unlock(&b->write_lock);
538			pr_debug("bnode %p: dirty bit cleaned by others\n", b);
539			continue;
540		}
541
542		__bch_btree_node_write(b, NULL);
543		clear_bit(BTREE_NODE_journal_flush, &b->flags);
544		mutex_unlock(&b->write_lock);
545	}
546
547out:
548	spin_lock(&c->journal.flush_write_lock);
549	c->journal.btree_flushing = false;
550	spin_unlock(&c->journal.flush_write_lock);
551}
552
553#define last_seq(j)	((j)->seq - fifo_used(&(j)->pin) + 1)
554
555static unsigned int free_journal_buckets(struct cache_set *c)
556{
557	struct journal *j = &c->journal;
558	struct cache *ca = c->cache;
559	struct journal_device *ja = &c->cache->journal;
560	unsigned int n;
561
562	/* In case njournal_buckets is not power of 2 */
563	if (ja->cur_idx >= ja->last_idx)
564		n = ca->sb.njournal_buckets + ja->last_idx - ja->cur_idx;
565	else
566		n = ja->last_idx - ja->cur_idx;
567
568	if (n > (1 + j->do_reserve))
569		return n - (1 + j->do_reserve);
570
571	return 0;
572}
573
574static void journal_reclaim(struct cache_set *c)
575{
576	struct bkey *k = &c->journal.key;
577	struct cache *ca = c->cache;
578	uint64_t last_seq;
579	struct journal_device *ja = &ca->journal;
580	atomic_t p __maybe_unused;
581
582	atomic_long_inc(&c->reclaim);
583
584	while (!atomic_read(&fifo_front(&c->journal.pin)))
585		fifo_pop(&c->journal.pin, p);
586
587	last_seq = last_seq(&c->journal);
588
589	/* Update last_idx */
590
591	while (ja->last_idx != ja->cur_idx &&
592	       ja->seq[ja->last_idx] < last_seq)
593		ja->last_idx = (ja->last_idx + 1) %
594			ca->sb.njournal_buckets;
595
596	if (c->journal.blocks_free)
597		goto out;
598
599	if (!free_journal_buckets(c))
600		goto out;
601
602	ja->cur_idx = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
603	k->ptr[0] = MAKE_PTR(0,
604			     bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
605			     ca->sb.nr_this_dev);
606	atomic_long_inc(&c->reclaimed_journal_buckets);
607
608	bkey_init(k);
609	SET_KEY_PTRS(k, 1);
610	c->journal.blocks_free = ca->sb.bucket_size >> c->block_bits;
611
612out:
613	if (!journal_full(&c->journal))
614		__closure_wake_up(&c->journal.wait);
615}
616
617void bch_journal_next(struct journal *j)
618{
619	atomic_t p = { 1 };
620
621	j->cur = (j->cur == j->w)
622		? &j->w[1]
623		: &j->w[0];
624
625	/*
626	 * The fifo_push() needs to happen at the same time as j->seq is
627	 * incremented for last_seq() to be calculated correctly
628	 */
629	BUG_ON(!fifo_push(&j->pin, p));
630	atomic_set(&fifo_back(&j->pin), 1);
631
632	j->cur->data->seq	= ++j->seq;
633	j->cur->dirty		= false;
634	j->cur->need_write	= false;
635	j->cur->data->keys	= 0;
636
637	if (fifo_full(&j->pin))
638		pr_debug("journal_pin full (%zu)\n", fifo_used(&j->pin));
639}
640
641static void journal_write_endio(struct bio *bio)
642{
643	struct journal_write *w = bio->bi_private;
644
645	cache_set_err_on(bio->bi_status, w->c, "journal io error");
646	closure_put(&w->c->journal.io);
647}
648
649static CLOSURE_CALLBACK(journal_write);
650
651static CLOSURE_CALLBACK(journal_write_done)
652{
653	closure_type(j, struct journal, io);
654	struct journal_write *w = (j->cur == j->w)
655		? &j->w[1]
656		: &j->w[0];
657
658	__closure_wake_up(&w->wait);
659	continue_at_nobarrier(cl, journal_write, bch_journal_wq);
660}
661
662static CLOSURE_CALLBACK(journal_write_unlock)
663	__releases(&c->journal.lock)
664{
665	closure_type(c, struct cache_set, journal.io);
666
667	c->journal.io_in_flight = 0;
668	spin_unlock(&c->journal.lock);
669}
670
671static CLOSURE_CALLBACK(journal_write_unlocked)
672	__releases(c->journal.lock)
673{
674	closure_type(c, struct cache_set, journal.io);
675	struct cache *ca = c->cache;
676	struct journal_write *w = c->journal.cur;
677	struct bkey *k = &c->journal.key;
678	unsigned int i, sectors = set_blocks(w->data, block_bytes(ca)) *
679		ca->sb.block_size;
680
681	struct bio *bio;
682	struct bio_list list;
683
684	bio_list_init(&list);
685
686	if (!w->need_write) {
687		closure_return_with_destructor(cl, journal_write_unlock);
688		return;
689	} else if (journal_full(&c->journal)) {
690		journal_reclaim(c);
691		spin_unlock(&c->journal.lock);
692
693		btree_flush_write(c);
694		continue_at(cl, journal_write, bch_journal_wq);
695		return;
696	}
697
698	c->journal.blocks_free -= set_blocks(w->data, block_bytes(ca));
699
700	w->data->btree_level = c->root->level;
701
702	bkey_copy(&w->data->btree_root, &c->root->key);
703	bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
704
705	w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
706	w->data->magic		= jset_magic(&ca->sb);
707	w->data->version	= BCACHE_JSET_VERSION;
708	w->data->last_seq	= last_seq(&c->journal);
709	w->data->csum		= csum_set(w->data);
710
711	for (i = 0; i < KEY_PTRS(k); i++) {
712		ca = c->cache;
713		bio = &ca->journal.bio;
714
715		atomic_long_add(sectors, &ca->meta_sectors_written);
716
717		bio_reset(bio, ca->bdev, REQ_OP_WRITE | 
718			  REQ_SYNC | REQ_META | REQ_PREFLUSH | REQ_FUA);
719		bio->bi_iter.bi_sector	= PTR_OFFSET(k, i);
720		bio->bi_iter.bi_size = sectors << 9;
721
722		bio->bi_end_io	= journal_write_endio;
723		bio->bi_private = w;
724		bch_bio_map(bio, w->data);
725
726		trace_bcache_journal_write(bio, w->data->keys);
727		bio_list_add(&list, bio);
728
729		SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
730
731		ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
732	}
733
734	/* If KEY_PTRS(k) == 0, this jset gets lost in air */
735	BUG_ON(i == 0);
736
737	atomic_dec_bug(&fifo_back(&c->journal.pin));
738	bch_journal_next(&c->journal);
739	journal_reclaim(c);
740
741	spin_unlock(&c->journal.lock);
742
743	while ((bio = bio_list_pop(&list)))
744		closure_bio_submit(c, bio, cl);
745
746	continue_at(cl, journal_write_done, NULL);
747}
748
749static CLOSURE_CALLBACK(journal_write)
750{
751	closure_type(c, struct cache_set, journal.io);
752
753	spin_lock(&c->journal.lock);
754	journal_write_unlocked(&cl->work);
755}
756
757static void journal_try_write(struct cache_set *c)
758	__releases(c->journal.lock)
759{
760	struct closure *cl = &c->journal.io;
761	struct journal_write *w = c->journal.cur;
762
763	w->need_write = true;
764
765	if (!c->journal.io_in_flight) {
766		c->journal.io_in_flight = 1;
767		closure_call(cl, journal_write_unlocked, NULL, &c->cl);
768	} else {
769		spin_unlock(&c->journal.lock);
770	}
771}
772
773static struct journal_write *journal_wait_for_write(struct cache_set *c,
774						    unsigned int nkeys)
775	__acquires(&c->journal.lock)
776{
777	size_t sectors;
778	struct closure cl;
779	bool wait = false;
780	struct cache *ca = c->cache;
781
782	closure_init_stack(&cl);
783
784	spin_lock(&c->journal.lock);
785
786	while (1) {
787		struct journal_write *w = c->journal.cur;
788
789		sectors = __set_blocks(w->data, w->data->keys + nkeys,
790				       block_bytes(ca)) * ca->sb.block_size;
791
792		if (sectors <= min_t(size_t,
793				     c->journal.blocks_free * ca->sb.block_size,
794				     PAGE_SECTORS << JSET_BITS))
795			return w;
796
797		if (wait)
798			closure_wait(&c->journal.wait, &cl);
799
800		if (!journal_full(&c->journal)) {
801			if (wait)
802				trace_bcache_journal_entry_full(c);
803
804			/*
805			 * XXX: If we were inserting so many keys that they
806			 * won't fit in an _empty_ journal write, we'll
807			 * deadlock. For now, handle this in
808			 * bch_keylist_realloc() - but something to think about.
809			 */
810			BUG_ON(!w->data->keys);
811
812			journal_try_write(c); /* unlocks */
813		} else {
814			if (wait)
815				trace_bcache_journal_full(c);
816
817			journal_reclaim(c);
818			spin_unlock(&c->journal.lock);
819
820			btree_flush_write(c);
821		}
822
823		closure_sync(&cl);
824		spin_lock(&c->journal.lock);
825		wait = true;
826	}
827}
828
829static void journal_write_work(struct work_struct *work)
830{
831	struct cache_set *c = container_of(to_delayed_work(work),
832					   struct cache_set,
833					   journal.work);
834	spin_lock(&c->journal.lock);
835	if (c->journal.cur->dirty)
836		journal_try_write(c);
837	else
838		spin_unlock(&c->journal.lock);
839}
840
841/*
842 * Entry point to the journalling code - bio_insert() and btree_invalidate()
843 * pass bch_journal() a list of keys to be journalled, and then
844 * bch_journal() hands those same keys off to btree_insert_async()
845 */
846
847atomic_t *bch_journal(struct cache_set *c,
848		      struct keylist *keys,
849		      struct closure *parent)
850{
851	struct journal_write *w;
852	atomic_t *ret;
853
854	/* No journaling if CACHE_SET_IO_DISABLE set already */
855	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
856		return NULL;
857
858	if (!CACHE_SYNC(&c->cache->sb))
859		return NULL;
860
861	w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
862
863	memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
864	w->data->keys += bch_keylist_nkeys(keys);
865
866	ret = &fifo_back(&c->journal.pin);
867	atomic_inc(ret);
868
869	if (parent) {
870		closure_wait(&w->wait, parent);
871		journal_try_write(c);
872	} else if (!w->dirty) {
873		w->dirty = true;
874		queue_delayed_work(bch_flush_wq, &c->journal.work,
875				   msecs_to_jiffies(c->journal_delay_ms));
876		spin_unlock(&c->journal.lock);
877	} else {
878		spin_unlock(&c->journal.lock);
879	}
880
881
882	return ret;
883}
884
885void bch_journal_meta(struct cache_set *c, struct closure *cl)
886{
887	struct keylist keys;
888	atomic_t *ref;
889
890	bch_keylist_init(&keys);
891
892	ref = bch_journal(c, &keys, cl);
893	if (ref)
894		atomic_dec_bug(ref);
895}
896
897void bch_journal_free(struct cache_set *c)
898{
899	free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
900	free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
901	free_fifo(&c->journal.pin);
902}
903
904int bch_journal_alloc(struct cache_set *c)
905{
906	struct journal *j = &c->journal;
907
908	spin_lock_init(&j->lock);
909	spin_lock_init(&j->flush_write_lock);
910	INIT_DELAYED_WORK(&j->work, journal_write_work);
911
912	c->journal_delay_ms = 100;
913
914	j->w[0].c = c;
915	j->w[1].c = c;
916
917	if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
918	    !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) ||
919	    !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)))
920		return -ENOMEM;
921
922	return 0;
923}