xref: /linux/drivers/md/bcache/journal.c (revision e7d759f31ca295d589f7420719c311870bb3166f)
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 
28 static void journal_read_endio(struct bio *bio)
29 {
30 	struct closure *cl = bio->bi_private;
31 
32 	closure_put(cl);
33 }
34 
35 static 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) {
53 reread:		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;
147 add:
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;
160 next_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 
170 int 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;
223 bsearch:
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 = ja->discard_idx = (i + 1) %
279 				ca->sb.njournal_buckets;
280 
281 		}
282 
283 out:
284 	if (!list_empty(list))
285 		c->journal.seq = list_entry(list->prev,
286 					    struct journal_replay,
287 					    list)->j.seq;
288 
289 	return 0;
290 #undef read_bucket
291 }
292 
293 void bch_journal_mark(struct cache_set *c, struct list_head *list)
294 {
295 	atomic_t p = { 0 };
296 	struct bkey *k;
297 	struct journal_replay *i;
298 	struct journal *j = &c->journal;
299 	uint64_t last = j->seq;
300 
301 	/*
302 	 * journal.pin should never fill up - we never write a journal
303 	 * entry when it would fill up. But if for some reason it does, we
304 	 * iterate over the list in reverse order so that we can just skip that
305 	 * refcount instead of bugging.
306 	 */
307 
308 	list_for_each_entry_reverse(i, list, list) {
309 		BUG_ON(last < i->j.seq);
310 		i->pin = NULL;
311 
312 		while (last-- != i->j.seq)
313 			if (fifo_free(&j->pin) > 1) {
314 				fifo_push_front(&j->pin, p);
315 				atomic_set(&fifo_front(&j->pin), 0);
316 			}
317 
318 		if (fifo_free(&j->pin) > 1) {
319 			fifo_push_front(&j->pin, p);
320 			i->pin = &fifo_front(&j->pin);
321 			atomic_set(i->pin, 1);
322 		}
323 
324 		for (k = i->j.start;
325 		     k < bset_bkey_last(&i->j);
326 		     k = bkey_next(k))
327 			if (!__bch_extent_invalid(c, k)) {
328 				unsigned int j;
329 
330 				for (j = 0; j < KEY_PTRS(k); j++)
331 					if (ptr_available(c, k, j))
332 						atomic_inc(&PTR_BUCKET(c, k, j)->pin);
333 
334 				bch_initial_mark_key(c, 0, k);
335 			}
336 	}
337 }
338 
339 static bool is_discard_enabled(struct cache_set *s)
340 {
341 	struct cache *ca = s->cache;
342 
343 	if (ca->discard)
344 		return true;
345 
346 	return false;
347 }
348 
349 int bch_journal_replay(struct cache_set *s, struct list_head *list)
350 {
351 	int ret = 0, keys = 0, entries = 0;
352 	struct bkey *k;
353 	struct journal_replay *i =
354 		list_entry(list->prev, struct journal_replay, list);
355 
356 	uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
357 	struct keylist keylist;
358 
359 	list_for_each_entry(i, list, list) {
360 		BUG_ON(i->pin && atomic_read(i->pin) != 1);
361 
362 		if (n != i->j.seq) {
363 			if (n == start && is_discard_enabled(s))
364 				pr_info("journal entries %llu-%llu may be discarded! (replaying %llu-%llu)\n",
365 					n, i->j.seq - 1, start, end);
366 			else {
367 				pr_err("journal entries %llu-%llu missing! (replaying %llu-%llu)\n",
368 					n, i->j.seq - 1, start, end);
369 				ret = -EIO;
370 				goto err;
371 			}
372 		}
373 
374 		for (k = i->j.start;
375 		     k < bset_bkey_last(&i->j);
376 		     k = bkey_next(k)) {
377 			trace_bcache_journal_replay_key(k);
378 
379 			bch_keylist_init_single(&keylist, k);
380 
381 			ret = bch_btree_insert(s, &keylist, i->pin, NULL);
382 			if (ret)
383 				goto err;
384 
385 			BUG_ON(!bch_keylist_empty(&keylist));
386 			keys++;
387 
388 			cond_resched();
389 		}
390 
391 		if (i->pin)
392 			atomic_dec(i->pin);
393 		n = i->j.seq + 1;
394 		entries++;
395 	}
396 
397 	pr_info("journal replay done, %i keys in %i entries, seq %llu\n",
398 		keys, entries, end);
399 err:
400 	while (!list_empty(list)) {
401 		i = list_first_entry(list, struct journal_replay, list);
402 		list_del(&i->list);
403 		kfree(i);
404 	}
405 
406 	return ret;
407 }
408 
409 void bch_journal_space_reserve(struct journal *j)
410 {
411 	j->do_reserve = true;
412 }
413 
414 /* Journalling */
415 
416 static void btree_flush_write(struct cache_set *c)
417 {
418 	struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR];
419 	unsigned int i, nr;
420 	int ref_nr;
421 	atomic_t *fifo_front_p, *now_fifo_front_p;
422 	size_t mask;
423 
424 	if (c->journal.btree_flushing)
425 		return;
426 
427 	spin_lock(&c->journal.flush_write_lock);
428 	if (c->journal.btree_flushing) {
429 		spin_unlock(&c->journal.flush_write_lock);
430 		return;
431 	}
432 	c->journal.btree_flushing = true;
433 	spin_unlock(&c->journal.flush_write_lock);
434 
435 	/* get the oldest journal entry and check its refcount */
436 	spin_lock(&c->journal.lock);
437 	fifo_front_p = &fifo_front(&c->journal.pin);
438 	ref_nr = atomic_read(fifo_front_p);
439 	if (ref_nr <= 0) {
440 		/*
441 		 * do nothing if no btree node references
442 		 * the oldest journal entry
443 		 */
444 		spin_unlock(&c->journal.lock);
445 		goto out;
446 	}
447 	spin_unlock(&c->journal.lock);
448 
449 	mask = c->journal.pin.mask;
450 	nr = 0;
451 	atomic_long_inc(&c->flush_write);
452 	memset(btree_nodes, 0, sizeof(btree_nodes));
453 
454 	mutex_lock(&c->bucket_lock);
455 	list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
456 		/*
457 		 * It is safe to get now_fifo_front_p without holding
458 		 * c->journal.lock here, because we don't need to know
459 		 * the exactly accurate value, just check whether the
460 		 * front pointer of c->journal.pin is changed.
461 		 */
462 		now_fifo_front_p = &fifo_front(&c->journal.pin);
463 		/*
464 		 * If the oldest journal entry is reclaimed and front
465 		 * pointer of c->journal.pin changes, it is unnecessary
466 		 * to scan c->btree_cache anymore, just quit the loop and
467 		 * flush out what we have already.
468 		 */
469 		if (now_fifo_front_p != fifo_front_p)
470 			break;
471 		/*
472 		 * quit this loop if all matching btree nodes are
473 		 * scanned and record in btree_nodes[] already.
474 		 */
475 		ref_nr = atomic_read(fifo_front_p);
476 		if (nr >= ref_nr)
477 			break;
478 
479 		if (btree_node_journal_flush(b))
480 			pr_err("BUG: flush_write bit should not be set here!\n");
481 
482 		mutex_lock(&b->write_lock);
483 
484 		if (!btree_node_dirty(b)) {
485 			mutex_unlock(&b->write_lock);
486 			continue;
487 		}
488 
489 		if (!btree_current_write(b)->journal) {
490 			mutex_unlock(&b->write_lock);
491 			continue;
492 		}
493 
494 		/*
495 		 * Only select the btree node which exactly references
496 		 * the oldest journal entry.
497 		 *
498 		 * If the journal entry pointed by fifo_front_p is
499 		 * reclaimed in parallel, don't worry:
500 		 * - the list_for_each_xxx loop will quit when checking
501 		 *   next now_fifo_front_p.
502 		 * - If there are matched nodes recorded in btree_nodes[],
503 		 *   they are clean now (this is why and how the oldest
504 		 *   journal entry can be reclaimed). These selected nodes
505 		 *   will be ignored and skipped in the following for-loop.
506 		 */
507 		if (((btree_current_write(b)->journal - fifo_front_p) &
508 		     mask) != 0) {
509 			mutex_unlock(&b->write_lock);
510 			continue;
511 		}
512 
513 		set_btree_node_journal_flush(b);
514 
515 		mutex_unlock(&b->write_lock);
516 
517 		btree_nodes[nr++] = b;
518 		/*
519 		 * To avoid holding c->bucket_lock too long time,
520 		 * only scan for BTREE_FLUSH_NR matched btree nodes
521 		 * at most. If there are more btree nodes reference
522 		 * the oldest journal entry, try to flush them next
523 		 * time when btree_flush_write() is called.
524 		 */
525 		if (nr == BTREE_FLUSH_NR)
526 			break;
527 	}
528 	mutex_unlock(&c->bucket_lock);
529 
530 	for (i = 0; i < nr; i++) {
531 		b = btree_nodes[i];
532 		if (!b) {
533 			pr_err("BUG: btree_nodes[%d] is NULL\n", i);
534 			continue;
535 		}
536 
537 		/* safe to check without holding b->write_lock */
538 		if (!btree_node_journal_flush(b)) {
539 			pr_err("BUG: bnode %p: journal_flush bit cleaned\n", b);
540 			continue;
541 		}
542 
543 		mutex_lock(&b->write_lock);
544 		if (!btree_current_write(b)->journal) {
545 			clear_bit(BTREE_NODE_journal_flush, &b->flags);
546 			mutex_unlock(&b->write_lock);
547 			pr_debug("bnode %p: written by others\n", b);
548 			continue;
549 		}
550 
551 		if (!btree_node_dirty(b)) {
552 			clear_bit(BTREE_NODE_journal_flush, &b->flags);
553 			mutex_unlock(&b->write_lock);
554 			pr_debug("bnode %p: dirty bit cleaned by others\n", b);
555 			continue;
556 		}
557 
558 		__bch_btree_node_write(b, NULL);
559 		clear_bit(BTREE_NODE_journal_flush, &b->flags);
560 		mutex_unlock(&b->write_lock);
561 	}
562 
563 out:
564 	spin_lock(&c->journal.flush_write_lock);
565 	c->journal.btree_flushing = false;
566 	spin_unlock(&c->journal.flush_write_lock);
567 }
568 
569 #define last_seq(j)	((j)->seq - fifo_used(&(j)->pin) + 1)
570 
571 static void journal_discard_endio(struct bio *bio)
572 {
573 	struct journal_device *ja =
574 		container_of(bio, struct journal_device, discard_bio);
575 	struct cache *ca = container_of(ja, struct cache, journal);
576 
577 	atomic_set(&ja->discard_in_flight, DISCARD_DONE);
578 
579 	closure_wake_up(&ca->set->journal.wait);
580 	closure_put(&ca->set->cl);
581 }
582 
583 static void journal_discard_work(struct work_struct *work)
584 {
585 	struct journal_device *ja =
586 		container_of(work, struct journal_device, discard_work);
587 
588 	submit_bio(&ja->discard_bio);
589 }
590 
591 static void do_journal_discard(struct cache *ca)
592 {
593 	struct journal_device *ja = &ca->journal;
594 	struct bio *bio = &ja->discard_bio;
595 
596 	if (!ca->discard) {
597 		ja->discard_idx = ja->last_idx;
598 		return;
599 	}
600 
601 	switch (atomic_read(&ja->discard_in_flight)) {
602 	case DISCARD_IN_FLIGHT:
603 		return;
604 
605 	case DISCARD_DONE:
606 		ja->discard_idx = (ja->discard_idx + 1) %
607 			ca->sb.njournal_buckets;
608 
609 		atomic_set(&ja->discard_in_flight, DISCARD_READY);
610 		fallthrough;
611 
612 	case DISCARD_READY:
613 		if (ja->discard_idx == ja->last_idx)
614 			return;
615 
616 		atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
617 
618 		bio_init(bio, ca->bdev, bio->bi_inline_vecs, 1, REQ_OP_DISCARD);
619 		bio->bi_iter.bi_sector	= bucket_to_sector(ca->set,
620 						ca->sb.d[ja->discard_idx]);
621 		bio->bi_iter.bi_size	= bucket_bytes(ca);
622 		bio->bi_end_io		= journal_discard_endio;
623 
624 		closure_get(&ca->set->cl);
625 		INIT_WORK(&ja->discard_work, journal_discard_work);
626 		queue_work(bch_journal_wq, &ja->discard_work);
627 	}
628 }
629 
630 static unsigned int free_journal_buckets(struct cache_set *c)
631 {
632 	struct journal *j = &c->journal;
633 	struct cache *ca = c->cache;
634 	struct journal_device *ja = &c->cache->journal;
635 	unsigned int n;
636 
637 	/* In case njournal_buckets is not power of 2 */
638 	if (ja->cur_idx >= ja->discard_idx)
639 		n = ca->sb.njournal_buckets +  ja->discard_idx - ja->cur_idx;
640 	else
641 		n = ja->discard_idx - ja->cur_idx;
642 
643 	if (n > (1 + j->do_reserve))
644 		return n - (1 + j->do_reserve);
645 
646 	return 0;
647 }
648 
649 static void journal_reclaim(struct cache_set *c)
650 {
651 	struct bkey *k = &c->journal.key;
652 	struct cache *ca = c->cache;
653 	uint64_t last_seq;
654 	struct journal_device *ja = &ca->journal;
655 	atomic_t p __maybe_unused;
656 
657 	atomic_long_inc(&c->reclaim);
658 
659 	while (!atomic_read(&fifo_front(&c->journal.pin)))
660 		fifo_pop(&c->journal.pin, p);
661 
662 	last_seq = last_seq(&c->journal);
663 
664 	/* Update last_idx */
665 
666 	while (ja->last_idx != ja->cur_idx &&
667 	       ja->seq[ja->last_idx] < last_seq)
668 		ja->last_idx = (ja->last_idx + 1) %
669 			ca->sb.njournal_buckets;
670 
671 	do_journal_discard(ca);
672 
673 	if (c->journal.blocks_free)
674 		goto out;
675 
676 	if (!free_journal_buckets(c))
677 		goto out;
678 
679 	ja->cur_idx = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
680 	k->ptr[0] = MAKE_PTR(0,
681 			     bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
682 			     ca->sb.nr_this_dev);
683 	atomic_long_inc(&c->reclaimed_journal_buckets);
684 
685 	bkey_init(k);
686 	SET_KEY_PTRS(k, 1);
687 	c->journal.blocks_free = ca->sb.bucket_size >> c->block_bits;
688 
689 out:
690 	if (!journal_full(&c->journal))
691 		__closure_wake_up(&c->journal.wait);
692 }
693 
694 void bch_journal_next(struct journal *j)
695 {
696 	atomic_t p = { 1 };
697 
698 	j->cur = (j->cur == j->w)
699 		? &j->w[1]
700 		: &j->w[0];
701 
702 	/*
703 	 * The fifo_push() needs to happen at the same time as j->seq is
704 	 * incremented for last_seq() to be calculated correctly
705 	 */
706 	BUG_ON(!fifo_push(&j->pin, p));
707 	atomic_set(&fifo_back(&j->pin), 1);
708 
709 	j->cur->data->seq	= ++j->seq;
710 	j->cur->dirty		= false;
711 	j->cur->need_write	= false;
712 	j->cur->data->keys	= 0;
713 
714 	if (fifo_full(&j->pin))
715 		pr_debug("journal_pin full (%zu)\n", fifo_used(&j->pin));
716 }
717 
718 static void journal_write_endio(struct bio *bio)
719 {
720 	struct journal_write *w = bio->bi_private;
721 
722 	cache_set_err_on(bio->bi_status, w->c, "journal io error");
723 	closure_put(&w->c->journal.io);
724 }
725 
726 static CLOSURE_CALLBACK(journal_write);
727 
728 static CLOSURE_CALLBACK(journal_write_done)
729 {
730 	closure_type(j, struct journal, io);
731 	struct journal_write *w = (j->cur == j->w)
732 		? &j->w[1]
733 		: &j->w[0];
734 
735 	__closure_wake_up(&w->wait);
736 	continue_at_nobarrier(cl, journal_write, bch_journal_wq);
737 }
738 
739 static CLOSURE_CALLBACK(journal_write_unlock)
740 	__releases(&c->journal.lock)
741 {
742 	closure_type(c, struct cache_set, journal.io);
743 
744 	c->journal.io_in_flight = 0;
745 	spin_unlock(&c->journal.lock);
746 }
747 
748 static CLOSURE_CALLBACK(journal_write_unlocked)
749 	__releases(c->journal.lock)
750 {
751 	closure_type(c, struct cache_set, journal.io);
752 	struct cache *ca = c->cache;
753 	struct journal_write *w = c->journal.cur;
754 	struct bkey *k = &c->journal.key;
755 	unsigned int i, sectors = set_blocks(w->data, block_bytes(ca)) *
756 		ca->sb.block_size;
757 
758 	struct bio *bio;
759 	struct bio_list list;
760 
761 	bio_list_init(&list);
762 
763 	if (!w->need_write) {
764 		closure_return_with_destructor(cl, journal_write_unlock);
765 		return;
766 	} else if (journal_full(&c->journal)) {
767 		journal_reclaim(c);
768 		spin_unlock(&c->journal.lock);
769 
770 		btree_flush_write(c);
771 		continue_at(cl, journal_write, bch_journal_wq);
772 		return;
773 	}
774 
775 	c->journal.blocks_free -= set_blocks(w->data, block_bytes(ca));
776 
777 	w->data->btree_level = c->root->level;
778 
779 	bkey_copy(&w->data->btree_root, &c->root->key);
780 	bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
781 
782 	w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
783 	w->data->magic		= jset_magic(&ca->sb);
784 	w->data->version	= BCACHE_JSET_VERSION;
785 	w->data->last_seq	= last_seq(&c->journal);
786 	w->data->csum		= csum_set(w->data);
787 
788 	for (i = 0; i < KEY_PTRS(k); i++) {
789 		ca = c->cache;
790 		bio = &ca->journal.bio;
791 
792 		atomic_long_add(sectors, &ca->meta_sectors_written);
793 
794 		bio_reset(bio, ca->bdev, REQ_OP_WRITE |
795 			  REQ_SYNC | REQ_META | REQ_PREFLUSH | REQ_FUA);
796 		bio->bi_iter.bi_sector	= PTR_OFFSET(k, i);
797 		bio->bi_iter.bi_size = sectors << 9;
798 
799 		bio->bi_end_io	= journal_write_endio;
800 		bio->bi_private = w;
801 		bch_bio_map(bio, w->data);
802 
803 		trace_bcache_journal_write(bio, w->data->keys);
804 		bio_list_add(&list, bio);
805 
806 		SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
807 
808 		ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
809 	}
810 
811 	/* If KEY_PTRS(k) == 0, this jset gets lost in air */
812 	BUG_ON(i == 0);
813 
814 	atomic_dec_bug(&fifo_back(&c->journal.pin));
815 	bch_journal_next(&c->journal);
816 	journal_reclaim(c);
817 
818 	spin_unlock(&c->journal.lock);
819 
820 	while ((bio = bio_list_pop(&list)))
821 		closure_bio_submit(c, bio, cl);
822 
823 	continue_at(cl, journal_write_done, NULL);
824 }
825 
826 static CLOSURE_CALLBACK(journal_write)
827 {
828 	closure_type(c, struct cache_set, journal.io);
829 
830 	spin_lock(&c->journal.lock);
831 	journal_write_unlocked(&cl->work);
832 }
833 
834 static void journal_try_write(struct cache_set *c)
835 	__releases(c->journal.lock)
836 {
837 	struct closure *cl = &c->journal.io;
838 	struct journal_write *w = c->journal.cur;
839 
840 	w->need_write = true;
841 
842 	if (!c->journal.io_in_flight) {
843 		c->journal.io_in_flight = 1;
844 		closure_call(cl, journal_write_unlocked, NULL, &c->cl);
845 	} else {
846 		spin_unlock(&c->journal.lock);
847 	}
848 }
849 
850 static struct journal_write *journal_wait_for_write(struct cache_set *c,
851 						    unsigned int nkeys)
852 	__acquires(&c->journal.lock)
853 {
854 	size_t sectors;
855 	struct closure cl;
856 	bool wait = false;
857 	struct cache *ca = c->cache;
858 
859 	closure_init_stack(&cl);
860 
861 	spin_lock(&c->journal.lock);
862 
863 	while (1) {
864 		struct journal_write *w = c->journal.cur;
865 
866 		sectors = __set_blocks(w->data, w->data->keys + nkeys,
867 				       block_bytes(ca)) * ca->sb.block_size;
868 
869 		if (sectors <= min_t(size_t,
870 				     c->journal.blocks_free * ca->sb.block_size,
871 				     PAGE_SECTORS << JSET_BITS))
872 			return w;
873 
874 		if (wait)
875 			closure_wait(&c->journal.wait, &cl);
876 
877 		if (!journal_full(&c->journal)) {
878 			if (wait)
879 				trace_bcache_journal_entry_full(c);
880 
881 			/*
882 			 * XXX: If we were inserting so many keys that they
883 			 * won't fit in an _empty_ journal write, we'll
884 			 * deadlock. For now, handle this in
885 			 * bch_keylist_realloc() - but something to think about.
886 			 */
887 			BUG_ON(!w->data->keys);
888 
889 			journal_try_write(c); /* unlocks */
890 		} else {
891 			if (wait)
892 				trace_bcache_journal_full(c);
893 
894 			journal_reclaim(c);
895 			spin_unlock(&c->journal.lock);
896 
897 			btree_flush_write(c);
898 		}
899 
900 		closure_sync(&cl);
901 		spin_lock(&c->journal.lock);
902 		wait = true;
903 	}
904 }
905 
906 static void journal_write_work(struct work_struct *work)
907 {
908 	struct cache_set *c = container_of(to_delayed_work(work),
909 					   struct cache_set,
910 					   journal.work);
911 	spin_lock(&c->journal.lock);
912 	if (c->journal.cur->dirty)
913 		journal_try_write(c);
914 	else
915 		spin_unlock(&c->journal.lock);
916 }
917 
918 /*
919  * Entry point to the journalling code - bio_insert() and btree_invalidate()
920  * pass bch_journal() a list of keys to be journalled, and then
921  * bch_journal() hands those same keys off to btree_insert_async()
922  */
923 
924 atomic_t *bch_journal(struct cache_set *c,
925 		      struct keylist *keys,
926 		      struct closure *parent)
927 {
928 	struct journal_write *w;
929 	atomic_t *ret;
930 
931 	/* No journaling if CACHE_SET_IO_DISABLE set already */
932 	if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
933 		return NULL;
934 
935 	if (!CACHE_SYNC(&c->cache->sb))
936 		return NULL;
937 
938 	w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
939 
940 	memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
941 	w->data->keys += bch_keylist_nkeys(keys);
942 
943 	ret = &fifo_back(&c->journal.pin);
944 	atomic_inc(ret);
945 
946 	if (parent) {
947 		closure_wait(&w->wait, parent);
948 		journal_try_write(c);
949 	} else if (!w->dirty) {
950 		w->dirty = true;
951 		queue_delayed_work(bch_flush_wq, &c->journal.work,
952 				   msecs_to_jiffies(c->journal_delay_ms));
953 		spin_unlock(&c->journal.lock);
954 	} else {
955 		spin_unlock(&c->journal.lock);
956 	}
957 
958 
959 	return ret;
960 }
961 
962 void bch_journal_meta(struct cache_set *c, struct closure *cl)
963 {
964 	struct keylist keys;
965 	atomic_t *ref;
966 
967 	bch_keylist_init(&keys);
968 
969 	ref = bch_journal(c, &keys, cl);
970 	if (ref)
971 		atomic_dec_bug(ref);
972 }
973 
974 void bch_journal_free(struct cache_set *c)
975 {
976 	free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
977 	free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
978 	free_fifo(&c->journal.pin);
979 }
980 
981 int bch_journal_alloc(struct cache_set *c)
982 {
983 	struct journal *j = &c->journal;
984 
985 	spin_lock_init(&j->lock);
986 	spin_lock_init(&j->flush_write_lock);
987 	INIT_DELAYED_WORK(&j->work, journal_write_work);
988 
989 	c->journal_delay_ms = 100;
990 
991 	j->w[0].c = c;
992 	j->w[1].c = c;
993 
994 	if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
995 	    !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) ||
996 	    !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)))
997 		return -ENOMEM;
998 
999 	return 0;
1000 }
1001