Loading...
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 = ja->discard_idx = (i + 1) %
279 ca->sb.njournal_buckets;
280
281 }
282
283out:
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
293void 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
339static 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
349int 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);
399err:
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
409void bch_journal_space_reserve(struct journal *j)
410{
411 j->do_reserve = true;
412}
413
414/* Journalling */
415
416static 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
563out:
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
571static 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
583static 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
591static 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
630static 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
649static 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
689out:
690 if (!journal_full(&c->journal))
691 __closure_wake_up(&c->journal.wait);
692}
693
694void 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
718static 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
726static void journal_write(struct closure *cl);
727
728static void journal_write_done(struct closure *cl)
729{
730 struct journal *j = container_of(cl, 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
739static void journal_write_unlock(struct closure *cl)
740 __releases(&c->journal.lock)
741{
742 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
743
744 c->journal.io_in_flight = 0;
745 spin_unlock(&c->journal.lock);
746}
747
748static void journal_write_unlocked(struct closure *cl)
749 __releases(c->journal.lock)
750{
751 struct cache_set *c = container_of(cl, 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
826static void journal_write(struct closure *cl)
827{
828 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
829
830 spin_lock(&c->journal.lock);
831 journal_write_unlocked(cl);
832}
833
834static 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
850static 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
906static 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
924atomic_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
962void 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
974void 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
981int 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}
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 closure_put(cl);
32}
33
34static int journal_read_bucket(struct cache *ca, struct list_head *list,
35 unsigned bucket_index)
36{
37 struct journal_device *ja = &ca->journal;
38 struct bio *bio = &ja->bio;
39
40 struct journal_replay *i;
41 struct jset *j, *data = ca->set->journal.w[0].data;
42 struct closure cl;
43 unsigned len, left, offset = 0;
44 int ret = 0;
45 sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
46
47 closure_init_stack(&cl);
48
49 pr_debug("reading %u", bucket_index);
50
51 while (offset < ca->sb.bucket_size) {
52reread: left = ca->sb.bucket_size - offset;
53 len = min_t(unsigned, left, PAGE_SECTORS << JSET_BITS);
54
55 bio_reset(bio);
56 bio->bi_iter.bi_sector = bucket + offset;
57 bio_set_dev(bio, ca->bdev);
58 bio->bi_iter.bi_size = len << 9;
59
60 bio->bi_end_io = journal_read_endio;
61 bio->bi_private = &cl;
62 bio_set_op_attrs(bio, REQ_OP_READ, 0);
63 bch_bio_map(bio, data);
64
65 closure_bio_submit(ca->set, bio, &cl);
66 closure_sync(&cl);
67
68 /* This function could be simpler now since we no longer write
69 * journal entries that overlap bucket boundaries; this means
70 * the start of a bucket will always have a valid journal entry
71 * if it has any journal entries at all.
72 */
73
74 j = data;
75 while (len) {
76 struct list_head *where;
77 size_t blocks, bytes = set_bytes(j);
78
79 if (j->magic != jset_magic(&ca->sb)) {
80 pr_debug("%u: bad magic", bucket_index);
81 return ret;
82 }
83
84 if (bytes > left << 9 ||
85 bytes > PAGE_SIZE << JSET_BITS) {
86 pr_info("%u: too big, %zu bytes, offset %u",
87 bucket_index, bytes, offset);
88 return ret;
89 }
90
91 if (bytes > len << 9)
92 goto reread;
93
94 if (j->csum != csum_set(j)) {
95 pr_info("%u: bad csum, %zu bytes, offset %u",
96 bucket_index, bytes, offset);
97 return ret;
98 }
99
100 blocks = set_blocks(j, block_bytes(ca->set));
101
102 while (!list_empty(list)) {
103 i = list_first_entry(list,
104 struct journal_replay, list);
105 if (i->j.seq >= j->last_seq)
106 break;
107 list_del(&i->list);
108 kfree(i);
109 }
110
111 list_for_each_entry_reverse(i, list, list) {
112 if (j->seq == i->j.seq)
113 goto next_set;
114
115 if (j->seq < i->j.last_seq)
116 goto next_set;
117
118 if (j->seq > i->j.seq) {
119 where = &i->list;
120 goto add;
121 }
122 }
123
124 where = list;
125add:
126 i = kmalloc(offsetof(struct journal_replay, j) +
127 bytes, GFP_KERNEL);
128 if (!i)
129 return -ENOMEM;
130 memcpy(&i->j, j, bytes);
131 list_add(&i->list, where);
132 ret = 1;
133
134 ja->seq[bucket_index] = j->seq;
135next_set:
136 offset += blocks * ca->sb.block_size;
137 len -= blocks * ca->sb.block_size;
138 j = ((void *) j) + blocks * block_bytes(ca);
139 }
140 }
141
142 return ret;
143}
144
145int bch_journal_read(struct cache_set *c, struct list_head *list)
146{
147#define read_bucket(b) \
148 ({ \
149 int ret = journal_read_bucket(ca, list, b); \
150 __set_bit(b, bitmap); \
151 if (ret < 0) \
152 return ret; \
153 ret; \
154 })
155
156 struct cache *ca;
157 unsigned iter;
158
159 for_each_cache(ca, c, iter) {
160 struct journal_device *ja = &ca->journal;
161 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
162 unsigned i, l, r, m;
163 uint64_t seq;
164
165 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
166 pr_debug("%u journal buckets", ca->sb.njournal_buckets);
167
168 /*
169 * Read journal buckets ordered by golden ratio hash to quickly
170 * find a sequence of buckets with valid journal entries
171 */
172 for (i = 0; i < ca->sb.njournal_buckets; i++) {
173 /*
174 * We must try the index l with ZERO first for
175 * correctness due to the scenario that the journal
176 * bucket is circular buffer which might have wrapped
177 */
178 l = (i * 2654435769U) % ca->sb.njournal_buckets;
179
180 if (test_bit(l, bitmap))
181 break;
182
183 if (read_bucket(l))
184 goto bsearch;
185 }
186
187 /*
188 * If that fails, check all the buckets we haven't checked
189 * already
190 */
191 pr_debug("falling back to linear search");
192
193 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
194 l < ca->sb.njournal_buckets;
195 l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1))
196 if (read_bucket(l))
197 goto bsearch;
198
199 /* no journal entries on this device? */
200 if (l == ca->sb.njournal_buckets)
201 continue;
202bsearch:
203 BUG_ON(list_empty(list));
204
205 /* Binary search */
206 m = l;
207 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
208 pr_debug("starting binary search, l %u r %u", l, r);
209
210 while (l + 1 < r) {
211 seq = list_entry(list->prev, struct journal_replay,
212 list)->j.seq;
213
214 m = (l + r) >> 1;
215 read_bucket(m);
216
217 if (seq != list_entry(list->prev, struct journal_replay,
218 list)->j.seq)
219 l = m;
220 else
221 r = m;
222 }
223
224 /*
225 * Read buckets in reverse order until we stop finding more
226 * journal entries
227 */
228 pr_debug("finishing up: m %u njournal_buckets %u",
229 m, ca->sb.njournal_buckets);
230 l = m;
231
232 while (1) {
233 if (!l--)
234 l = ca->sb.njournal_buckets - 1;
235
236 if (l == m)
237 break;
238
239 if (test_bit(l, bitmap))
240 continue;
241
242 if (!read_bucket(l))
243 break;
244 }
245
246 seq = 0;
247
248 for (i = 0; i < ca->sb.njournal_buckets; i++)
249 if (ja->seq[i] > seq) {
250 seq = ja->seq[i];
251 /*
252 * When journal_reclaim() goes to allocate for
253 * the first time, it'll use the bucket after
254 * ja->cur_idx
255 */
256 ja->cur_idx = i;
257 ja->last_idx = ja->discard_idx = (i + 1) %
258 ca->sb.njournal_buckets;
259
260 }
261 }
262
263 if (!list_empty(list))
264 c->journal.seq = list_entry(list->prev,
265 struct journal_replay,
266 list)->j.seq;
267
268 return 0;
269#undef read_bucket
270}
271
272void bch_journal_mark(struct cache_set *c, struct list_head *list)
273{
274 atomic_t p = { 0 };
275 struct bkey *k;
276 struct journal_replay *i;
277 struct journal *j = &c->journal;
278 uint64_t last = j->seq;
279
280 /*
281 * journal.pin should never fill up - we never write a journal
282 * entry when it would fill up. But if for some reason it does, we
283 * iterate over the list in reverse order so that we can just skip that
284 * refcount instead of bugging.
285 */
286
287 list_for_each_entry_reverse(i, list, list) {
288 BUG_ON(last < i->j.seq);
289 i->pin = NULL;
290
291 while (last-- != i->j.seq)
292 if (fifo_free(&j->pin) > 1) {
293 fifo_push_front(&j->pin, p);
294 atomic_set(&fifo_front(&j->pin), 0);
295 }
296
297 if (fifo_free(&j->pin) > 1) {
298 fifo_push_front(&j->pin, p);
299 i->pin = &fifo_front(&j->pin);
300 atomic_set(i->pin, 1);
301 }
302
303 for (k = i->j.start;
304 k < bset_bkey_last(&i->j);
305 k = bkey_next(k))
306 if (!__bch_extent_invalid(c, k)) {
307 unsigned j;
308
309 for (j = 0; j < KEY_PTRS(k); j++)
310 if (ptr_available(c, k, j))
311 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
312
313 bch_initial_mark_key(c, 0, k);
314 }
315 }
316}
317
318int bch_journal_replay(struct cache_set *s, struct list_head *list)
319{
320 int ret = 0, keys = 0, entries = 0;
321 struct bkey *k;
322 struct journal_replay *i =
323 list_entry(list->prev, struct journal_replay, list);
324
325 uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
326 struct keylist keylist;
327
328 list_for_each_entry(i, list, list) {
329 BUG_ON(i->pin && atomic_read(i->pin) != 1);
330
331 cache_set_err_on(n != i->j.seq, s,
332"bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
333 n, i->j.seq - 1, start, end);
334
335 for (k = i->j.start;
336 k < bset_bkey_last(&i->j);
337 k = bkey_next(k)) {
338 trace_bcache_journal_replay_key(k);
339
340 bch_keylist_init_single(&keylist, k);
341
342 ret = bch_btree_insert(s, &keylist, i->pin, NULL);
343 if (ret)
344 goto err;
345
346 BUG_ON(!bch_keylist_empty(&keylist));
347 keys++;
348
349 cond_resched();
350 }
351
352 if (i->pin)
353 atomic_dec(i->pin);
354 n = i->j.seq + 1;
355 entries++;
356 }
357
358 pr_info("journal replay done, %i keys in %i entries, seq %llu",
359 keys, entries, end);
360err:
361 while (!list_empty(list)) {
362 i = list_first_entry(list, struct journal_replay, list);
363 list_del(&i->list);
364 kfree(i);
365 }
366
367 return ret;
368}
369
370/* Journalling */
371#define journal_max_cmp(l, r) \
372 (fifo_idx(&c->journal.pin, btree_current_write(l)->journal) < \
373 fifo_idx(&(c)->journal.pin, btree_current_write(r)->journal))
374#define journal_min_cmp(l, r) \
375 (fifo_idx(&c->journal.pin, btree_current_write(l)->journal) > \
376 fifo_idx(&(c)->journal.pin, btree_current_write(r)->journal))
377
378static void btree_flush_write(struct cache_set *c)
379{
380 /*
381 * Try to find the btree node with that references the oldest journal
382 * entry, best is our current candidate and is locked if non NULL:
383 */
384 struct btree *b;
385 int i;
386
387 atomic_long_inc(&c->flush_write);
388
389retry:
390 spin_lock(&c->journal.lock);
391 if (heap_empty(&c->flush_btree)) {
392 for_each_cached_btree(b, c, i)
393 if (btree_current_write(b)->journal) {
394 if (!heap_full(&c->flush_btree))
395 heap_add(&c->flush_btree, b,
396 journal_max_cmp);
397 else if (journal_max_cmp(b,
398 heap_peek(&c->flush_btree))) {
399 c->flush_btree.data[0] = b;
400 heap_sift(&c->flush_btree, 0,
401 journal_max_cmp);
402 }
403 }
404
405 for (i = c->flush_btree.used / 2 - 1; i >= 0; --i)
406 heap_sift(&c->flush_btree, i, journal_min_cmp);
407 }
408
409 b = NULL;
410 heap_pop(&c->flush_btree, b, journal_min_cmp);
411 spin_unlock(&c->journal.lock);
412
413 if (b) {
414 mutex_lock(&b->write_lock);
415 if (!btree_current_write(b)->journal) {
416 mutex_unlock(&b->write_lock);
417 /* We raced */
418 atomic_long_inc(&c->retry_flush_write);
419 goto retry;
420 }
421
422 __bch_btree_node_write(b, NULL);
423 mutex_unlock(&b->write_lock);
424 }
425}
426
427#define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1)
428
429static void journal_discard_endio(struct bio *bio)
430{
431 struct journal_device *ja =
432 container_of(bio, struct journal_device, discard_bio);
433 struct cache *ca = container_of(ja, struct cache, journal);
434
435 atomic_set(&ja->discard_in_flight, DISCARD_DONE);
436
437 closure_wake_up(&ca->set->journal.wait);
438 closure_put(&ca->set->cl);
439}
440
441static void journal_discard_work(struct work_struct *work)
442{
443 struct journal_device *ja =
444 container_of(work, struct journal_device, discard_work);
445
446 submit_bio(&ja->discard_bio);
447}
448
449static void do_journal_discard(struct cache *ca)
450{
451 struct journal_device *ja = &ca->journal;
452 struct bio *bio = &ja->discard_bio;
453
454 if (!ca->discard) {
455 ja->discard_idx = ja->last_idx;
456 return;
457 }
458
459 switch (atomic_read(&ja->discard_in_flight)) {
460 case DISCARD_IN_FLIGHT:
461 return;
462
463 case DISCARD_DONE:
464 ja->discard_idx = (ja->discard_idx + 1) %
465 ca->sb.njournal_buckets;
466
467 atomic_set(&ja->discard_in_flight, DISCARD_READY);
468 /* fallthrough */
469
470 case DISCARD_READY:
471 if (ja->discard_idx == ja->last_idx)
472 return;
473
474 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
475
476 bio_init(bio, bio->bi_inline_vecs, 1);
477 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
478 bio->bi_iter.bi_sector = bucket_to_sector(ca->set,
479 ca->sb.d[ja->discard_idx]);
480 bio_set_dev(bio, ca->bdev);
481 bio->bi_iter.bi_size = bucket_bytes(ca);
482 bio->bi_end_io = journal_discard_endio;
483
484 closure_get(&ca->set->cl);
485 INIT_WORK(&ja->discard_work, journal_discard_work);
486 schedule_work(&ja->discard_work);
487 }
488}
489
490static void journal_reclaim(struct cache_set *c)
491{
492 struct bkey *k = &c->journal.key;
493 struct cache *ca;
494 uint64_t last_seq;
495 unsigned iter, n = 0;
496 atomic_t p __maybe_unused;
497
498 atomic_long_inc(&c->reclaim);
499
500 while (!atomic_read(&fifo_front(&c->journal.pin)))
501 fifo_pop(&c->journal.pin, p);
502
503 last_seq = last_seq(&c->journal);
504
505 /* Update last_idx */
506
507 for_each_cache(ca, c, iter) {
508 struct journal_device *ja = &ca->journal;
509
510 while (ja->last_idx != ja->cur_idx &&
511 ja->seq[ja->last_idx] < last_seq)
512 ja->last_idx = (ja->last_idx + 1) %
513 ca->sb.njournal_buckets;
514 }
515
516 for_each_cache(ca, c, iter)
517 do_journal_discard(ca);
518
519 if (c->journal.blocks_free)
520 goto out;
521
522 /*
523 * Allocate:
524 * XXX: Sort by free journal space
525 */
526
527 for_each_cache(ca, c, iter) {
528 struct journal_device *ja = &ca->journal;
529 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
530
531 /* No space available on this device */
532 if (next == ja->discard_idx)
533 continue;
534
535 ja->cur_idx = next;
536 k->ptr[n++] = MAKE_PTR(0,
537 bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
538 ca->sb.nr_this_dev);
539 }
540
541 bkey_init(k);
542 SET_KEY_PTRS(k, n);
543
544 if (n)
545 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
546out:
547 if (!journal_full(&c->journal))
548 __closure_wake_up(&c->journal.wait);
549}
550
551void bch_journal_next(struct journal *j)
552{
553 atomic_t p = { 1 };
554
555 j->cur = (j->cur == j->w)
556 ? &j->w[1]
557 : &j->w[0];
558
559 /*
560 * The fifo_push() needs to happen at the same time as j->seq is
561 * incremented for last_seq() to be calculated correctly
562 */
563 BUG_ON(!fifo_push(&j->pin, p));
564 atomic_set(&fifo_back(&j->pin), 1);
565
566 j->cur->data->seq = ++j->seq;
567 j->cur->dirty = false;
568 j->cur->need_write = false;
569 j->cur->data->keys = 0;
570
571 if (fifo_full(&j->pin))
572 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
573}
574
575static void journal_write_endio(struct bio *bio)
576{
577 struct journal_write *w = bio->bi_private;
578
579 cache_set_err_on(bio->bi_status, w->c, "journal io error");
580 closure_put(&w->c->journal.io);
581}
582
583static void journal_write(struct closure *);
584
585static void journal_write_done(struct closure *cl)
586{
587 struct journal *j = container_of(cl, struct journal, io);
588 struct journal_write *w = (j->cur == j->w)
589 ? &j->w[1]
590 : &j->w[0];
591
592 __closure_wake_up(&w->wait);
593 continue_at_nobarrier(cl, journal_write, system_wq);
594}
595
596static void journal_write_unlock(struct closure *cl)
597 __releases(&c->journal.lock)
598{
599 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
600
601 c->journal.io_in_flight = 0;
602 spin_unlock(&c->journal.lock);
603}
604
605static void journal_write_unlocked(struct closure *cl)
606 __releases(c->journal.lock)
607{
608 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
609 struct cache *ca;
610 struct journal_write *w = c->journal.cur;
611 struct bkey *k = &c->journal.key;
612 unsigned i, sectors = set_blocks(w->data, block_bytes(c)) *
613 c->sb.block_size;
614
615 struct bio *bio;
616 struct bio_list list;
617 bio_list_init(&list);
618
619 if (!w->need_write) {
620 closure_return_with_destructor(cl, journal_write_unlock);
621 return;
622 } else if (journal_full(&c->journal)) {
623 journal_reclaim(c);
624 spin_unlock(&c->journal.lock);
625
626 btree_flush_write(c);
627 continue_at(cl, journal_write, system_wq);
628 return;
629 }
630
631 c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
632
633 w->data->btree_level = c->root->level;
634
635 bkey_copy(&w->data->btree_root, &c->root->key);
636 bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
637
638 for_each_cache(ca, c, i)
639 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
640
641 w->data->magic = jset_magic(&c->sb);
642 w->data->version = BCACHE_JSET_VERSION;
643 w->data->last_seq = last_seq(&c->journal);
644 w->data->csum = csum_set(w->data);
645
646 for (i = 0; i < KEY_PTRS(k); i++) {
647 ca = PTR_CACHE(c, k, i);
648 bio = &ca->journal.bio;
649
650 atomic_long_add(sectors, &ca->meta_sectors_written);
651
652 bio_reset(bio);
653 bio->bi_iter.bi_sector = PTR_OFFSET(k, i);
654 bio_set_dev(bio, ca->bdev);
655 bio->bi_iter.bi_size = sectors << 9;
656
657 bio->bi_end_io = journal_write_endio;
658 bio->bi_private = w;
659 bio_set_op_attrs(bio, REQ_OP_WRITE,
660 REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
661 bch_bio_map(bio, w->data);
662
663 trace_bcache_journal_write(bio);
664 bio_list_add(&list, bio);
665
666 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
667
668 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
669 }
670
671 atomic_dec_bug(&fifo_back(&c->journal.pin));
672 bch_journal_next(&c->journal);
673 journal_reclaim(c);
674
675 spin_unlock(&c->journal.lock);
676
677 while ((bio = bio_list_pop(&list)))
678 closure_bio_submit(c, bio, cl);
679
680 continue_at(cl, journal_write_done, NULL);
681}
682
683static void journal_write(struct closure *cl)
684{
685 struct cache_set *c = container_of(cl, struct cache_set, journal.io);
686
687 spin_lock(&c->journal.lock);
688 journal_write_unlocked(cl);
689}
690
691static void journal_try_write(struct cache_set *c)
692 __releases(c->journal.lock)
693{
694 struct closure *cl = &c->journal.io;
695 struct journal_write *w = c->journal.cur;
696
697 w->need_write = true;
698
699 if (!c->journal.io_in_flight) {
700 c->journal.io_in_flight = 1;
701 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
702 } else {
703 spin_unlock(&c->journal.lock);
704 }
705}
706
707static struct journal_write *journal_wait_for_write(struct cache_set *c,
708 unsigned nkeys)
709 __acquires(&c->journal.lock)
710{
711 size_t sectors;
712 struct closure cl;
713 bool wait = false;
714
715 closure_init_stack(&cl);
716
717 spin_lock(&c->journal.lock);
718
719 while (1) {
720 struct journal_write *w = c->journal.cur;
721
722 sectors = __set_blocks(w->data, w->data->keys + nkeys,
723 block_bytes(c)) * c->sb.block_size;
724
725 if (sectors <= min_t(size_t,
726 c->journal.blocks_free * c->sb.block_size,
727 PAGE_SECTORS << JSET_BITS))
728 return w;
729
730 if (wait)
731 closure_wait(&c->journal.wait, &cl);
732
733 if (!journal_full(&c->journal)) {
734 if (wait)
735 trace_bcache_journal_entry_full(c);
736
737 /*
738 * XXX: If we were inserting so many keys that they
739 * won't fit in an _empty_ journal write, we'll
740 * deadlock. For now, handle this in
741 * bch_keylist_realloc() - but something to think about.
742 */
743 BUG_ON(!w->data->keys);
744
745 journal_try_write(c); /* unlocks */
746 } else {
747 if (wait)
748 trace_bcache_journal_full(c);
749
750 journal_reclaim(c);
751 spin_unlock(&c->journal.lock);
752
753 btree_flush_write(c);
754 }
755
756 closure_sync(&cl);
757 spin_lock(&c->journal.lock);
758 wait = true;
759 }
760}
761
762static void journal_write_work(struct work_struct *work)
763{
764 struct cache_set *c = container_of(to_delayed_work(work),
765 struct cache_set,
766 journal.work);
767 spin_lock(&c->journal.lock);
768 if (c->journal.cur->dirty)
769 journal_try_write(c);
770 else
771 spin_unlock(&c->journal.lock);
772}
773
774/*
775 * Entry point to the journalling code - bio_insert() and btree_invalidate()
776 * pass bch_journal() a list of keys to be journalled, and then
777 * bch_journal() hands those same keys off to btree_insert_async()
778 */
779
780atomic_t *bch_journal(struct cache_set *c,
781 struct keylist *keys,
782 struct closure *parent)
783{
784 struct journal_write *w;
785 atomic_t *ret;
786
787 if (!CACHE_SYNC(&c->sb))
788 return NULL;
789
790 w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
791
792 memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
793 w->data->keys += bch_keylist_nkeys(keys);
794
795 ret = &fifo_back(&c->journal.pin);
796 atomic_inc(ret);
797
798 if (parent) {
799 closure_wait(&w->wait, parent);
800 journal_try_write(c);
801 } else if (!w->dirty) {
802 w->dirty = true;
803 schedule_delayed_work(&c->journal.work,
804 msecs_to_jiffies(c->journal_delay_ms));
805 spin_unlock(&c->journal.lock);
806 } else {
807 spin_unlock(&c->journal.lock);
808 }
809
810
811 return ret;
812}
813
814void bch_journal_meta(struct cache_set *c, struct closure *cl)
815{
816 struct keylist keys;
817 atomic_t *ref;
818
819 bch_keylist_init(&keys);
820
821 ref = bch_journal(c, &keys, cl);
822 if (ref)
823 atomic_dec_bug(ref);
824}
825
826void bch_journal_free(struct cache_set *c)
827{
828 free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
829 free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
830 free_fifo(&c->journal.pin);
831}
832
833int bch_journal_alloc(struct cache_set *c)
834{
835 struct journal *j = &c->journal;
836
837 spin_lock_init(&j->lock);
838 INIT_DELAYED_WORK(&j->work, journal_write_work);
839
840 c->journal_delay_ms = 100;
841
842 j->w[0].c = c;
843 j->w[1].c = c;
844
845 if (!(init_heap(&c->flush_btree, 128, GFP_KERNEL)) ||
846 !(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
847 !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
848 !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
849 return -ENOMEM;
850
851 return 0;
852}