2 * bcache journalling code, for btree insertions
4 * Copyright 2012 Google, Inc.
11 #include <trace/events/bcache.h>
14 * Journal replay/recovery:
16 * This code is all driven from run_cache_set(); we first read the journal
17 * entries, do some other stuff, then we mark all the keys in the journal
18 * entries (same as garbage collection would), then we replay them - reinserting
19 * them into the cache in precisely the same order as they appear in the
22 * We only journal keys that go in leaf nodes, which simplifies things quite a
26 static void journal_read_endio(struct bio
*bio
, int error
)
28 struct closure
*cl
= bio
->bi_private
;
32 static int journal_read_bucket(struct cache
*ca
, struct list_head
*list
,
33 struct btree_op
*op
, unsigned bucket_index
)
35 struct journal_device
*ja
= &ca
->journal
;
36 struct bio
*bio
= &ja
->bio
;
38 struct journal_replay
*i
;
39 struct jset
*j
, *data
= ca
->set
->journal
.w
[0].data
;
40 unsigned len
, left
, offset
= 0;
42 sector_t bucket
= bucket_to_sector(ca
->set
, ca
->sb
.d
[bucket_index
]);
44 pr_debug("reading %llu", (uint64_t) bucket
);
46 while (offset
< ca
->sb
.bucket_size
) {
47 reread
: left
= ca
->sb
.bucket_size
- offset
;
48 len
= min_t(unsigned, left
, PAGE_SECTORS
* 8);
51 bio
->bi_sector
= bucket
+ offset
;
52 bio
->bi_bdev
= ca
->bdev
;
54 bio
->bi_size
= len
<< 9;
56 bio
->bi_end_io
= journal_read_endio
;
57 bio
->bi_private
= &op
->cl
;
58 bch_bio_map(bio
, data
);
60 closure_bio_submit(bio
, &op
->cl
, ca
);
61 closure_sync(&op
->cl
);
63 /* This function could be simpler now since we no longer write
64 * journal entries that overlap bucket boundaries; this means
65 * the start of a bucket will always have a valid journal entry
66 * if it has any journal entries at all.
71 struct list_head
*where
;
72 size_t blocks
, bytes
= set_bytes(j
);
74 if (j
->magic
!= jset_magic(ca
->set
))
77 if (bytes
> left
<< 9)
83 if (j
->csum
!= csum_set(j
))
86 blocks
= set_blocks(j
, ca
->set
);
88 while (!list_empty(list
)) {
89 i
= list_first_entry(list
,
90 struct journal_replay
, list
);
91 if (i
->j
.seq
>= j
->last_seq
)
97 list_for_each_entry_reverse(i
, list
, list
) {
98 if (j
->seq
== i
->j
.seq
)
101 if (j
->seq
< i
->j
.last_seq
)
104 if (j
->seq
> i
->j
.seq
) {
112 i
= kmalloc(offsetof(struct journal_replay
, j
) +
116 memcpy(&i
->j
, j
, bytes
);
117 list_add(&i
->list
, where
);
120 ja
->seq
[bucket_index
] = j
->seq
;
122 offset
+= blocks
* ca
->sb
.block_size
;
123 len
-= blocks
* ca
->sb
.block_size
;
124 j
= ((void *) j
) + blocks
* block_bytes(ca
);
131 int bch_journal_read(struct cache_set
*c
, struct list_head
*list
,
134 #define read_bucket(b) \
136 int ret = journal_read_bucket(ca, list, op, b); \
137 __set_bit(b, bitmap); \
146 for_each_cache(ca
, c
, iter
) {
147 struct journal_device
*ja
= &ca
->journal
;
148 unsigned long bitmap
[SB_JOURNAL_BUCKETS
/ BITS_PER_LONG
];
152 bitmap_zero(bitmap
, SB_JOURNAL_BUCKETS
);
153 pr_debug("%u journal buckets", ca
->sb
.njournal_buckets
);
156 * Read journal buckets ordered by golden ratio hash to quickly
157 * find a sequence of buckets with valid journal entries
159 for (i
= 0; i
< ca
->sb
.njournal_buckets
; i
++) {
160 l
= (i
* 2654435769U) % ca
->sb
.njournal_buckets
;
162 if (test_bit(l
, bitmap
))
170 * If that fails, check all the buckets we haven't checked
173 pr_debug("falling back to linear search");
175 for (l
= find_first_zero_bit(bitmap
, ca
->sb
.njournal_buckets
);
176 l
< ca
->sb
.njournal_buckets
;
177 l
= find_next_zero_bit(bitmap
, ca
->sb
.njournal_buckets
, l
+ 1))
181 if (list_empty(list
))
185 m
= r
= find_next_bit(bitmap
, ca
->sb
.njournal_buckets
, l
+ 1);
186 pr_debug("starting binary search, l %u r %u", l
, r
);
189 seq
= list_entry(list
->prev
, struct journal_replay
,
195 if (seq
!= list_entry(list
->prev
, struct journal_replay
,
203 * Read buckets in reverse order until we stop finding more
206 pr_debug("finishing up: m %u njournal_buckets %u",
207 m
, ca
->sb
.njournal_buckets
);
212 l
= ca
->sb
.njournal_buckets
- 1;
217 if (test_bit(l
, bitmap
))
226 for (i
= 0; i
< ca
->sb
.njournal_buckets
; i
++)
227 if (ja
->seq
[i
] > seq
) {
229 ja
->cur_idx
= ja
->discard_idx
=
235 if (!list_empty(list
))
236 c
->journal
.seq
= list_entry(list
->prev
,
237 struct journal_replay
,
244 void bch_journal_mark(struct cache_set
*c
, struct list_head
*list
)
248 struct journal_replay
*i
;
249 struct journal
*j
= &c
->journal
;
250 uint64_t last
= j
->seq
;
253 * journal.pin should never fill up - we never write a journal
254 * entry when it would fill up. But if for some reason it does, we
255 * iterate over the list in reverse order so that we can just skip that
256 * refcount instead of bugging.
259 list_for_each_entry_reverse(i
, list
, list
) {
260 BUG_ON(last
< i
->j
.seq
);
263 while (last
-- != i
->j
.seq
)
264 if (fifo_free(&j
->pin
) > 1) {
265 fifo_push_front(&j
->pin
, p
);
266 atomic_set(&fifo_front(&j
->pin
), 0);
269 if (fifo_free(&j
->pin
) > 1) {
270 fifo_push_front(&j
->pin
, p
);
271 i
->pin
= &fifo_front(&j
->pin
);
272 atomic_set(i
->pin
, 1);
280 for (j
= 0; j
< KEY_PTRS(k
); j
++) {
281 struct bucket
*g
= PTR_BUCKET(c
, k
, j
);
284 if (g
->prio
== BTREE_PRIO
&&
286 g
->prio
= INITIAL_PRIO
;
289 __bch_btree_mark_key(c
, 0, k
);
294 int bch_journal_replay(struct cache_set
*s
, struct list_head
*list
,
297 int ret
= 0, keys
= 0, entries
= 0;
299 struct journal_replay
*i
=
300 list_entry(list
->prev
, struct journal_replay
, list
);
302 uint64_t start
= i
->j
.last_seq
, end
= i
->j
.seq
, n
= start
;
303 struct keylist keylist
;
305 bch_keylist_init(&keylist
);
307 list_for_each_entry(i
, list
, list
) {
308 BUG_ON(i
->pin
&& atomic_read(i
->pin
) != 1);
310 cache_set_err_on(n
!= i
->j
.seq
, s
,
311 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
312 n
, i
->j
.seq
- 1, start
, end
);
317 trace_bcache_journal_replay_key(k
);
319 bkey_copy(keylist
.top
, k
);
320 bch_keylist_push(&keylist
);
322 op
->journal
= i
->pin
;
324 ret
= bch_btree_insert(op
, s
, &keylist
);
328 BUG_ON(!bch_keylist_empty(&keylist
));
340 pr_info("journal replay done, %i keys in %i entries, seq %llu",
343 while (!list_empty(list
)) {
344 i
= list_first_entry(list
, struct journal_replay
, list
);
349 closure_sync(&op
->cl
);
355 static void btree_flush_write(struct cache_set
*c
)
358 * Try to find the btree node with that references the oldest journal
359 * entry, best is our current candidate and is locked if non NULL:
361 struct btree
*b
, *best
;
366 for_each_cached_btree(b
, c
, i
)
367 if (btree_current_write(b
)->journal
) {
370 else if (journal_pin_cmp(c
,
371 btree_current_write(best
),
372 btree_current_write(b
))) {
379 rw_lock(true, b
, b
->level
);
381 if (!btree_current_write(b
)->journal
) {
387 bch_btree_node_write(b
, NULL
);
392 #define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1)
394 static void journal_discard_endio(struct bio
*bio
, int error
)
396 struct journal_device
*ja
=
397 container_of(bio
, struct journal_device
, discard_bio
);
398 struct cache
*ca
= container_of(ja
, struct cache
, journal
);
400 atomic_set(&ja
->discard_in_flight
, DISCARD_DONE
);
402 closure_wake_up(&ca
->set
->journal
.wait
);
403 closure_put(&ca
->set
->cl
);
406 static void journal_discard_work(struct work_struct
*work
)
408 struct journal_device
*ja
=
409 container_of(work
, struct journal_device
, discard_work
);
411 submit_bio(0, &ja
->discard_bio
);
414 static void do_journal_discard(struct cache
*ca
)
416 struct journal_device
*ja
= &ca
->journal
;
417 struct bio
*bio
= &ja
->discard_bio
;
420 ja
->discard_idx
= ja
->last_idx
;
424 switch (atomic_read(&ja
->discard_in_flight
)) {
425 case DISCARD_IN_FLIGHT
:
429 ja
->discard_idx
= (ja
->discard_idx
+ 1) %
430 ca
->sb
.njournal_buckets
;
432 atomic_set(&ja
->discard_in_flight
, DISCARD_READY
);
436 if (ja
->discard_idx
== ja
->last_idx
)
439 atomic_set(&ja
->discard_in_flight
, DISCARD_IN_FLIGHT
);
442 bio
->bi_sector
= bucket_to_sector(ca
->set
,
443 ca
->sb
.d
[ja
->discard_idx
]);
444 bio
->bi_bdev
= ca
->bdev
;
445 bio
->bi_rw
= REQ_WRITE
|REQ_DISCARD
;
446 bio
->bi_max_vecs
= 1;
447 bio
->bi_io_vec
= bio
->bi_inline_vecs
;
448 bio
->bi_size
= bucket_bytes(ca
);
449 bio
->bi_end_io
= journal_discard_endio
;
451 closure_get(&ca
->set
->cl
);
452 INIT_WORK(&ja
->discard_work
, journal_discard_work
);
453 schedule_work(&ja
->discard_work
);
457 static void journal_reclaim(struct cache_set
*c
)
459 struct bkey
*k
= &c
->journal
.key
;
462 unsigned iter
, n
= 0;
465 while (!atomic_read(&fifo_front(&c
->journal
.pin
)))
466 fifo_pop(&c
->journal
.pin
, p
);
468 last_seq
= last_seq(&c
->journal
);
470 /* Update last_idx */
472 for_each_cache(ca
, c
, iter
) {
473 struct journal_device
*ja
= &ca
->journal
;
475 while (ja
->last_idx
!= ja
->cur_idx
&&
476 ja
->seq
[ja
->last_idx
] < last_seq
)
477 ja
->last_idx
= (ja
->last_idx
+ 1) %
478 ca
->sb
.njournal_buckets
;
481 for_each_cache(ca
, c
, iter
)
482 do_journal_discard(ca
);
484 if (c
->journal
.blocks_free
)
489 * XXX: Sort by free journal space
492 for_each_cache(ca
, c
, iter
) {
493 struct journal_device
*ja
= &ca
->journal
;
494 unsigned next
= (ja
->cur_idx
+ 1) % ca
->sb
.njournal_buckets
;
496 /* No space available on this device */
497 if (next
== ja
->discard_idx
)
502 bucket_to_sector(c
, ca
->sb
.d
[ja
->cur_idx
]),
510 c
->journal
.blocks_free
= c
->sb
.bucket_size
>> c
->block_bits
;
512 if (!journal_full(&c
->journal
))
513 __closure_wake_up(&c
->journal
.wait
);
516 void bch_journal_next(struct journal
*j
)
520 j
->cur
= (j
->cur
== j
->w
)
525 * The fifo_push() needs to happen at the same time as j->seq is
526 * incremented for last_seq() to be calculated correctly
528 BUG_ON(!fifo_push(&j
->pin
, p
));
529 atomic_set(&fifo_back(&j
->pin
), 1);
531 j
->cur
->data
->seq
= ++j
->seq
;
532 j
->cur
->need_write
= false;
533 j
->cur
->data
->keys
= 0;
535 if (fifo_full(&j
->pin
))
536 pr_debug("journal_pin full (%zu)", fifo_used(&j
->pin
));
539 static void journal_write_endio(struct bio
*bio
, int error
)
541 struct journal_write
*w
= bio
->bi_private
;
543 cache_set_err_on(error
, w
->c
, "journal io error");
544 closure_put(&w
->c
->journal
.io
);
547 static void journal_write(struct closure
*);
549 static void journal_write_done(struct closure
*cl
)
551 struct journal
*j
= container_of(cl
, struct journal
, io
);
552 struct journal_write
*w
= (j
->cur
== j
->w
)
556 __closure_wake_up(&w
->wait
);
557 continue_at_nobarrier(cl
, journal_write
, system_wq
);
560 static void journal_write_unlocked(struct closure
*cl
)
561 __releases(c
->journal
.lock
)
563 struct cache_set
*c
= container_of(cl
, struct cache_set
, journal
.io
);
565 struct journal_write
*w
= c
->journal
.cur
;
566 struct bkey
*k
= &c
->journal
.key
;
567 unsigned i
, sectors
= set_blocks(w
->data
, c
) * c
->sb
.block_size
;
570 struct bio_list list
;
571 bio_list_init(&list
);
573 if (!w
->need_write
) {
575 * XXX: have to unlock closure before we unlock journal lock,
576 * else we race with bch_journal(). But this way we race
577 * against cache set unregister. Doh.
579 set_closure_fn(cl
, NULL
, NULL
);
580 closure_sub(cl
, CLOSURE_RUNNING
+ 1);
581 spin_unlock(&c
->journal
.lock
);
583 } else if (journal_full(&c
->journal
)) {
585 spin_unlock(&c
->journal
.lock
);
587 btree_flush_write(c
);
588 continue_at(cl
, journal_write
, system_wq
);
591 c
->journal
.blocks_free
-= set_blocks(w
->data
, c
);
593 w
->data
->btree_level
= c
->root
->level
;
595 bkey_copy(&w
->data
->btree_root
, &c
->root
->key
);
596 bkey_copy(&w
->data
->uuid_bucket
, &c
->uuid_bucket
);
598 for_each_cache(ca
, c
, i
)
599 w
->data
->prio_bucket
[ca
->sb
.nr_this_dev
] = ca
->prio_buckets
[0];
601 w
->data
->magic
= jset_magic(c
);
602 w
->data
->version
= BCACHE_JSET_VERSION
;
603 w
->data
->last_seq
= last_seq(&c
->journal
);
604 w
->data
->csum
= csum_set(w
->data
);
606 for (i
= 0; i
< KEY_PTRS(k
); i
++) {
607 ca
= PTR_CACHE(c
, k
, i
);
608 bio
= &ca
->journal
.bio
;
610 atomic_long_add(sectors
, &ca
->meta_sectors_written
);
613 bio
->bi_sector
= PTR_OFFSET(k
, i
);
614 bio
->bi_bdev
= ca
->bdev
;
615 bio
->bi_rw
= REQ_WRITE
|REQ_SYNC
|REQ_META
|REQ_FLUSH
|REQ_FUA
;
616 bio
->bi_size
= sectors
<< 9;
618 bio
->bi_end_io
= journal_write_endio
;
620 bch_bio_map(bio
, w
->data
);
622 trace_bcache_journal_write(bio
);
623 bio_list_add(&list
, bio
);
625 SET_PTR_OFFSET(k
, i
, PTR_OFFSET(k
, i
) + sectors
);
627 ca
->journal
.seq
[ca
->journal
.cur_idx
] = w
->data
->seq
;
630 atomic_dec_bug(&fifo_back(&c
->journal
.pin
));
631 bch_journal_next(&c
->journal
);
634 spin_unlock(&c
->journal
.lock
);
636 while ((bio
= bio_list_pop(&list
)))
637 closure_bio_submit(bio
, cl
, c
->cache
[0]);
639 continue_at(cl
, journal_write_done
, NULL
);
642 static void journal_write(struct closure
*cl
)
644 struct cache_set
*c
= container_of(cl
, struct cache_set
, journal
.io
);
646 spin_lock(&c
->journal
.lock
);
647 journal_write_unlocked(cl
);
650 static void journal_try_write(struct cache_set
*c
)
651 __releases(c
->journal
.lock
)
653 struct closure
*cl
= &c
->journal
.io
;
654 struct journal_write
*w
= c
->journal
.cur
;
656 w
->need_write
= true;
658 if (closure_trylock(cl
, &c
->cl
))
659 journal_write_unlocked(cl
);
661 spin_unlock(&c
->journal
.lock
);
664 static struct journal_write
*journal_wait_for_write(struct cache_set
*c
,
670 closure_init_stack(&cl
);
672 spin_lock(&c
->journal
.lock
);
675 struct journal_write
*w
= c
->journal
.cur
;
677 sectors
= __set_blocks(w
->data
, w
->data
->keys
+ nkeys
,
678 c
) * c
->sb
.block_size
;
680 if (sectors
<= min_t(size_t,
681 c
->journal
.blocks_free
* c
->sb
.block_size
,
682 PAGE_SECTORS
<< JSET_BITS
))
685 /* XXX: tracepoint */
686 if (!journal_full(&c
->journal
)) {
687 trace_bcache_journal_entry_full(c
);
690 * XXX: If we were inserting so many keys that they
691 * won't fit in an _empty_ journal write, we'll
692 * deadlock. For now, handle this in
693 * bch_keylist_realloc() - but something to think about.
695 BUG_ON(!w
->data
->keys
);
697 closure_wait(&w
->wait
, &cl
);
698 journal_try_write(c
); /* unlocks */
700 trace_bcache_journal_full(c
);
702 closure_wait(&c
->journal
.wait
, &cl
);
704 spin_unlock(&c
->journal
.lock
);
706 btree_flush_write(c
);
710 spin_lock(&c
->journal
.lock
);
714 static void journal_write_work(struct work_struct
*work
)
716 struct cache_set
*c
= container_of(to_delayed_work(work
),
719 spin_lock(&c
->journal
.lock
);
720 journal_try_write(c
);
724 * Entry point to the journalling code - bio_insert() and btree_invalidate()
725 * pass bch_journal() a list of keys to be journalled, and then
726 * bch_journal() hands those same keys off to btree_insert_async()
729 atomic_t
*bch_journal(struct cache_set
*c
,
730 struct keylist
*keys
,
731 struct closure
*parent
)
733 struct journal_write
*w
;
736 if (!CACHE_SYNC(&c
->sb
))
739 w
= journal_wait_for_write(c
, bch_keylist_nkeys(keys
));
741 memcpy(end(w
->data
), keys
->keys
, bch_keylist_bytes(keys
));
742 w
->data
->keys
+= bch_keylist_nkeys(keys
);
744 ret
= &fifo_back(&c
->journal
.pin
);
748 closure_wait(&w
->wait
, parent
);
749 journal_try_write(c
);
750 } else if (!w
->need_write
) {
751 schedule_delayed_work(&c
->journal
.work
,
752 msecs_to_jiffies(c
->journal_delay_ms
));
753 spin_unlock(&c
->journal
.lock
);
755 spin_unlock(&c
->journal
.lock
);
762 void bch_journal_meta(struct cache_set
*c
, struct closure
*cl
)
767 bch_keylist_init(&keys
);
769 ref
= bch_journal(c
, &keys
, cl
);
774 void bch_journal_free(struct cache_set
*c
)
776 free_pages((unsigned long) c
->journal
.w
[1].data
, JSET_BITS
);
777 free_pages((unsigned long) c
->journal
.w
[0].data
, JSET_BITS
);
778 free_fifo(&c
->journal
.pin
);
781 int bch_journal_alloc(struct cache_set
*c
)
783 struct journal
*j
= &c
->journal
;
785 closure_init_unlocked(&j
->io
);
786 spin_lock_init(&j
->lock
);
787 INIT_DELAYED_WORK(&j
->work
, journal_write_work
);
789 c
->journal_delay_ms
= 100;
794 if (!(init_fifo(&j
->pin
, JOURNAL_PIN
, GFP_KERNEL
)) ||
795 !(j
->w
[0].data
= (void *) __get_free_pages(GFP_KERNEL
, JSET_BITS
)) ||
796 !(j
->w
[1].data
= (void *) __get_free_pages(GFP_KERNEL
, JSET_BITS
)))
This page took 0.048441 seconds and 6 git commands to generate.