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 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
;
41 unsigned len
, left
, offset
= 0;
43 sector_t bucket
= bucket_to_sector(ca
->set
, ca
->sb
.d
[bucket_index
]);
45 closure_init_stack(&cl
);
47 pr_debug("reading %u", bucket_index
);
49 while (offset
< ca
->sb
.bucket_size
) {
50 reread
: left
= ca
->sb
.bucket_size
- offset
;
51 len
= min_t(unsigned, left
, PAGE_SECTORS
<< JSET_BITS
);
54 bio
->bi_iter
.bi_sector
= bucket
+ offset
;
55 bio
->bi_bdev
= ca
->bdev
;
57 bio
->bi_iter
.bi_size
= len
<< 9;
59 bio
->bi_end_io
= journal_read_endio
;
60 bio
->bi_private
= &cl
;
61 bch_bio_map(bio
, data
);
63 closure_bio_submit(bio
, &cl
, ca
);
66 /* This function could be simpler now since we no longer write
67 * journal entries that overlap bucket boundaries; this means
68 * the start of a bucket will always have a valid journal entry
69 * if it has any journal entries at all.
74 struct list_head
*where
;
75 size_t blocks
, bytes
= set_bytes(j
);
77 if (j
->magic
!= jset_magic(&ca
->sb
)) {
78 pr_debug("%u: bad magic", bucket_index
);
82 if (bytes
> left
<< 9 ||
83 bytes
> PAGE_SIZE
<< JSET_BITS
) {
84 pr_info("%u: too big, %zu bytes, offset %u",
85 bucket_index
, bytes
, offset
);
92 if (j
->csum
!= csum_set(j
)) {
93 pr_info("%u: bad csum, %zu bytes, offset %u",
94 bucket_index
, bytes
, offset
);
98 blocks
= set_blocks(j
, block_bytes(ca
->set
));
100 while (!list_empty(list
)) {
101 i
= list_first_entry(list
,
102 struct journal_replay
, list
);
103 if (i
->j
.seq
>= j
->last_seq
)
109 list_for_each_entry_reverse(i
, list
, list
) {
110 if (j
->seq
== i
->j
.seq
)
113 if (j
->seq
< i
->j
.last_seq
)
116 if (j
->seq
> i
->j
.seq
) {
124 i
= kmalloc(offsetof(struct journal_replay
, j
) +
128 memcpy(&i
->j
, j
, bytes
);
129 list_add(&i
->list
, where
);
132 ja
->seq
[bucket_index
] = j
->seq
;
134 offset
+= blocks
* ca
->sb
.block_size
;
135 len
-= blocks
* ca
->sb
.block_size
;
136 j
= ((void *) j
) + blocks
* block_bytes(ca
);
143 int bch_journal_read(struct cache_set
*c
, struct list_head
*list
)
145 #define read_bucket(b) \
147 int ret = journal_read_bucket(ca, list, b); \
148 __set_bit(b, bitmap); \
157 for_each_cache(ca
, c
, iter
) {
158 struct journal_device
*ja
= &ca
->journal
;
159 unsigned long bitmap
[SB_JOURNAL_BUCKETS
/ BITS_PER_LONG
];
163 bitmap_zero(bitmap
, SB_JOURNAL_BUCKETS
);
164 pr_debug("%u journal buckets", ca
->sb
.njournal_buckets
);
167 * Read journal buckets ordered by golden ratio hash to quickly
168 * find a sequence of buckets with valid journal entries
170 for (i
= 0; i
< ca
->sb
.njournal_buckets
; i
++) {
171 l
= (i
* 2654435769U) % ca
->sb
.njournal_buckets
;
173 if (test_bit(l
, bitmap
))
181 * If that fails, check all the buckets we haven't checked
184 pr_debug("falling back to linear search");
186 for (l
= find_first_zero_bit(bitmap
, ca
->sb
.njournal_buckets
);
187 l
< ca
->sb
.njournal_buckets
;
188 l
= find_next_zero_bit(bitmap
, ca
->sb
.njournal_buckets
, l
+ 1))
192 if (list_empty(list
))
196 m
= r
= find_next_bit(bitmap
, ca
->sb
.njournal_buckets
, l
+ 1);
197 pr_debug("starting binary search, l %u r %u", l
, r
);
200 seq
= list_entry(list
->prev
, struct journal_replay
,
206 if (seq
!= list_entry(list
->prev
, struct journal_replay
,
214 * Read buckets in reverse order until we stop finding more
217 pr_debug("finishing up: m %u njournal_buckets %u",
218 m
, ca
->sb
.njournal_buckets
);
223 l
= ca
->sb
.njournal_buckets
- 1;
228 if (test_bit(l
, bitmap
))
237 for (i
= 0; i
< ca
->sb
.njournal_buckets
; i
++)
238 if (ja
->seq
[i
] > seq
) {
240 ja
->cur_idx
= ja
->discard_idx
=
246 if (!list_empty(list
))
247 c
->journal
.seq
= list_entry(list
->prev
,
248 struct journal_replay
,
255 void bch_journal_mark(struct cache_set
*c
, struct list_head
*list
)
259 struct journal_replay
*i
;
260 struct journal
*j
= &c
->journal
;
261 uint64_t last
= j
->seq
;
264 * journal.pin should never fill up - we never write a journal
265 * entry when it would fill up. But if for some reason it does, we
266 * iterate over the list in reverse order so that we can just skip that
267 * refcount instead of bugging.
270 list_for_each_entry_reverse(i
, list
, list
) {
271 BUG_ON(last
< i
->j
.seq
);
274 while (last
-- != i
->j
.seq
)
275 if (fifo_free(&j
->pin
) > 1) {
276 fifo_push_front(&j
->pin
, p
);
277 atomic_set(&fifo_front(&j
->pin
), 0);
280 if (fifo_free(&j
->pin
) > 1) {
281 fifo_push_front(&j
->pin
, p
);
282 i
->pin
= &fifo_front(&j
->pin
);
283 atomic_set(i
->pin
, 1);
287 k
< bset_bkey_last(&i
->j
);
291 for (j
= 0; j
< KEY_PTRS(k
); j
++) {
292 struct bucket
*g
= PTR_BUCKET(c
, k
, j
);
295 if (g
->prio
== BTREE_PRIO
&&
297 g
->prio
= INITIAL_PRIO
;
300 __bch_btree_mark_key(c
, 0, k
);
305 int bch_journal_replay(struct cache_set
*s
, struct list_head
*list
)
307 int ret
= 0, keys
= 0, entries
= 0;
309 struct journal_replay
*i
=
310 list_entry(list
->prev
, struct journal_replay
, list
);
312 uint64_t start
= i
->j
.last_seq
, end
= i
->j
.seq
, n
= start
;
313 struct keylist keylist
;
315 bch_keylist_init(&keylist
);
317 list_for_each_entry(i
, list
, list
) {
318 BUG_ON(i
->pin
&& atomic_read(i
->pin
) != 1);
320 cache_set_err_on(n
!= i
->j
.seq
, s
,
321 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
322 n
, i
->j
.seq
- 1, start
, end
);
325 k
< bset_bkey_last(&i
->j
);
327 trace_bcache_journal_replay_key(k
);
329 bkey_copy(keylist
.top
, k
);
330 bch_keylist_push(&keylist
);
332 ret
= bch_btree_insert(s
, &keylist
, i
->pin
, NULL
);
336 BUG_ON(!bch_keylist_empty(&keylist
));
348 pr_info("journal replay done, %i keys in %i entries, seq %llu",
351 while (!list_empty(list
)) {
352 i
= list_first_entry(list
, struct journal_replay
, list
);
362 static void btree_flush_write(struct cache_set
*c
)
365 * Try to find the btree node with that references the oldest journal
366 * entry, best is our current candidate and is locked if non NULL:
368 struct btree
*b
, *best
;
373 for_each_cached_btree(b
, c
, i
)
374 if (btree_current_write(b
)->journal
) {
377 else if (journal_pin_cmp(c
,
378 btree_current_write(best
)->journal
,
379 btree_current_write(b
)->journal
)) {
386 rw_lock(true, b
, b
->level
);
388 if (!btree_current_write(b
)->journal
) {
394 bch_btree_node_write(b
, NULL
);
399 #define last_seq(j) ((j)->seq - fifo_used(&(j)->pin) + 1)
401 static void journal_discard_endio(struct bio
*bio
, int error
)
403 struct journal_device
*ja
=
404 container_of(bio
, struct journal_device
, discard_bio
);
405 struct cache
*ca
= container_of(ja
, struct cache
, journal
);
407 atomic_set(&ja
->discard_in_flight
, DISCARD_DONE
);
409 closure_wake_up(&ca
->set
->journal
.wait
);
410 closure_put(&ca
->set
->cl
);
413 static void journal_discard_work(struct work_struct
*work
)
415 struct journal_device
*ja
=
416 container_of(work
, struct journal_device
, discard_work
);
418 submit_bio(0, &ja
->discard_bio
);
421 static void do_journal_discard(struct cache
*ca
)
423 struct journal_device
*ja
= &ca
->journal
;
424 struct bio
*bio
= &ja
->discard_bio
;
427 ja
->discard_idx
= ja
->last_idx
;
431 switch (atomic_read(&ja
->discard_in_flight
)) {
432 case DISCARD_IN_FLIGHT
:
436 ja
->discard_idx
= (ja
->discard_idx
+ 1) %
437 ca
->sb
.njournal_buckets
;
439 atomic_set(&ja
->discard_in_flight
, DISCARD_READY
);
443 if (ja
->discard_idx
== ja
->last_idx
)
446 atomic_set(&ja
->discard_in_flight
, DISCARD_IN_FLIGHT
);
449 bio
->bi_iter
.bi_sector
= bucket_to_sector(ca
->set
,
450 ca
->sb
.d
[ja
->discard_idx
]);
451 bio
->bi_bdev
= ca
->bdev
;
452 bio
->bi_rw
= REQ_WRITE
|REQ_DISCARD
;
453 bio
->bi_max_vecs
= 1;
454 bio
->bi_io_vec
= bio
->bi_inline_vecs
;
455 bio
->bi_iter
.bi_size
= bucket_bytes(ca
);
456 bio
->bi_end_io
= journal_discard_endio
;
458 closure_get(&ca
->set
->cl
);
459 INIT_WORK(&ja
->discard_work
, journal_discard_work
);
460 schedule_work(&ja
->discard_work
);
464 static void journal_reclaim(struct cache_set
*c
)
466 struct bkey
*k
= &c
->journal
.key
;
469 unsigned iter
, n
= 0;
472 while (!atomic_read(&fifo_front(&c
->journal
.pin
)))
473 fifo_pop(&c
->journal
.pin
, p
);
475 last_seq
= last_seq(&c
->journal
);
477 /* Update last_idx */
479 for_each_cache(ca
, c
, iter
) {
480 struct journal_device
*ja
= &ca
->journal
;
482 while (ja
->last_idx
!= ja
->cur_idx
&&
483 ja
->seq
[ja
->last_idx
] < last_seq
)
484 ja
->last_idx
= (ja
->last_idx
+ 1) %
485 ca
->sb
.njournal_buckets
;
488 for_each_cache(ca
, c
, iter
)
489 do_journal_discard(ca
);
491 if (c
->journal
.blocks_free
)
496 * XXX: Sort by free journal space
499 for_each_cache(ca
, c
, iter
) {
500 struct journal_device
*ja
= &ca
->journal
;
501 unsigned next
= (ja
->cur_idx
+ 1) % ca
->sb
.njournal_buckets
;
503 /* No space available on this device */
504 if (next
== ja
->discard_idx
)
509 bucket_to_sector(c
, ca
->sb
.d
[ja
->cur_idx
]),
517 c
->journal
.blocks_free
= c
->sb
.bucket_size
>> c
->block_bits
;
519 if (!journal_full(&c
->journal
))
520 __closure_wake_up(&c
->journal
.wait
);
523 void bch_journal_next(struct journal
*j
)
527 j
->cur
= (j
->cur
== j
->w
)
532 * The fifo_push() needs to happen at the same time as j->seq is
533 * incremented for last_seq() to be calculated correctly
535 BUG_ON(!fifo_push(&j
->pin
, p
));
536 atomic_set(&fifo_back(&j
->pin
), 1);
538 j
->cur
->data
->seq
= ++j
->seq
;
539 j
->cur
->need_write
= false;
540 j
->cur
->data
->keys
= 0;
542 if (fifo_full(&j
->pin
))
543 pr_debug("journal_pin full (%zu)", fifo_used(&j
->pin
));
546 static void journal_write_endio(struct bio
*bio
, int error
)
548 struct journal_write
*w
= bio
->bi_private
;
550 cache_set_err_on(error
, w
->c
, "journal io error");
551 closure_put(&w
->c
->journal
.io
);
554 static void journal_write(struct closure
*);
556 static void journal_write_done(struct closure
*cl
)
558 struct journal
*j
= container_of(cl
, struct journal
, io
);
559 struct journal_write
*w
= (j
->cur
== j
->w
)
563 __closure_wake_up(&w
->wait
);
564 continue_at_nobarrier(cl
, journal_write
, system_wq
);
567 static void journal_write_unlock(struct closure
*cl
)
569 struct cache_set
*c
= container_of(cl
, struct cache_set
, journal
.io
);
571 c
->journal
.io_in_flight
= 0;
572 spin_unlock(&c
->journal
.lock
);
575 static void journal_write_unlocked(struct closure
*cl
)
576 __releases(c
->journal
.lock
)
578 struct cache_set
*c
= container_of(cl
, struct cache_set
, journal
.io
);
580 struct journal_write
*w
= c
->journal
.cur
;
581 struct bkey
*k
= &c
->journal
.key
;
582 unsigned i
, sectors
= set_blocks(w
->data
, block_bytes(c
)) *
586 struct bio_list list
;
587 bio_list_init(&list
);
589 if (!w
->need_write
) {
590 closure_return_with_destructor(cl
, journal_write_unlock
);
591 } else if (journal_full(&c
->journal
)) {
593 spin_unlock(&c
->journal
.lock
);
595 btree_flush_write(c
);
596 continue_at(cl
, journal_write
, system_wq
);
599 c
->journal
.blocks_free
-= set_blocks(w
->data
, block_bytes(c
));
601 w
->data
->btree_level
= c
->root
->level
;
603 bkey_copy(&w
->data
->btree_root
, &c
->root
->key
);
604 bkey_copy(&w
->data
->uuid_bucket
, &c
->uuid_bucket
);
606 for_each_cache(ca
, c
, i
)
607 w
->data
->prio_bucket
[ca
->sb
.nr_this_dev
] = ca
->prio_buckets
[0];
609 w
->data
->magic
= jset_magic(&c
->sb
);
610 w
->data
->version
= BCACHE_JSET_VERSION
;
611 w
->data
->last_seq
= last_seq(&c
->journal
);
612 w
->data
->csum
= csum_set(w
->data
);
614 for (i
= 0; i
< KEY_PTRS(k
); i
++) {
615 ca
= PTR_CACHE(c
, k
, i
);
616 bio
= &ca
->journal
.bio
;
618 atomic_long_add(sectors
, &ca
->meta_sectors_written
);
621 bio
->bi_iter
.bi_sector
= PTR_OFFSET(k
, i
);
622 bio
->bi_bdev
= ca
->bdev
;
623 bio
->bi_rw
= REQ_WRITE
|REQ_SYNC
|REQ_META
|REQ_FLUSH
|REQ_FUA
;
624 bio
->bi_iter
.bi_size
= sectors
<< 9;
626 bio
->bi_end_io
= journal_write_endio
;
628 bch_bio_map(bio
, w
->data
);
630 trace_bcache_journal_write(bio
);
631 bio_list_add(&list
, bio
);
633 SET_PTR_OFFSET(k
, i
, PTR_OFFSET(k
, i
) + sectors
);
635 ca
->journal
.seq
[ca
->journal
.cur_idx
] = w
->data
->seq
;
638 atomic_dec_bug(&fifo_back(&c
->journal
.pin
));
639 bch_journal_next(&c
->journal
);
642 spin_unlock(&c
->journal
.lock
);
644 while ((bio
= bio_list_pop(&list
)))
645 closure_bio_submit(bio
, cl
, c
->cache
[0]);
647 continue_at(cl
, journal_write_done
, NULL
);
650 static void journal_write(struct closure
*cl
)
652 struct cache_set
*c
= container_of(cl
, struct cache_set
, journal
.io
);
654 spin_lock(&c
->journal
.lock
);
655 journal_write_unlocked(cl
);
658 static void journal_try_write(struct cache_set
*c
)
659 __releases(c
->journal
.lock
)
661 struct closure
*cl
= &c
->journal
.io
;
662 struct journal_write
*w
= c
->journal
.cur
;
664 w
->need_write
= true;
666 if (!c
->journal
.io_in_flight
) {
667 c
->journal
.io_in_flight
= 1;
668 closure_call(cl
, journal_write_unlocked
, NULL
, &c
->cl
);
670 spin_unlock(&c
->journal
.lock
);
674 static struct journal_write
*journal_wait_for_write(struct cache_set
*c
,
681 closure_init_stack(&cl
);
683 spin_lock(&c
->journal
.lock
);
686 struct journal_write
*w
= c
->journal
.cur
;
688 sectors
= __set_blocks(w
->data
, w
->data
->keys
+ nkeys
,
689 block_bytes(c
)) * c
->sb
.block_size
;
691 if (sectors
<= min_t(size_t,
692 c
->journal
.blocks_free
* c
->sb
.block_size
,
693 PAGE_SECTORS
<< JSET_BITS
))
697 closure_wait(&c
->journal
.wait
, &cl
);
699 if (!journal_full(&c
->journal
)) {
701 trace_bcache_journal_entry_full(c
);
704 * XXX: If we were inserting so many keys that they
705 * won't fit in an _empty_ journal write, we'll
706 * deadlock. For now, handle this in
707 * bch_keylist_realloc() - but something to think about.
709 BUG_ON(!w
->data
->keys
);
711 journal_try_write(c
); /* unlocks */
714 trace_bcache_journal_full(c
);
717 spin_unlock(&c
->journal
.lock
);
719 btree_flush_write(c
);
723 spin_lock(&c
->journal
.lock
);
728 static void journal_write_work(struct work_struct
*work
)
730 struct cache_set
*c
= container_of(to_delayed_work(work
),
733 spin_lock(&c
->journal
.lock
);
734 journal_try_write(c
);
738 * Entry point to the journalling code - bio_insert() and btree_invalidate()
739 * pass bch_journal() a list of keys to be journalled, and then
740 * bch_journal() hands those same keys off to btree_insert_async()
743 atomic_t
*bch_journal(struct cache_set
*c
,
744 struct keylist
*keys
,
745 struct closure
*parent
)
747 struct journal_write
*w
;
750 if (!CACHE_SYNC(&c
->sb
))
753 w
= journal_wait_for_write(c
, bch_keylist_nkeys(keys
));
755 memcpy(bset_bkey_last(w
->data
), keys
->keys
, bch_keylist_bytes(keys
));
756 w
->data
->keys
+= bch_keylist_nkeys(keys
);
758 ret
= &fifo_back(&c
->journal
.pin
);
762 closure_wait(&w
->wait
, parent
);
763 journal_try_write(c
);
764 } else if (!w
->need_write
) {
765 schedule_delayed_work(&c
->journal
.work
,
766 msecs_to_jiffies(c
->journal_delay_ms
));
767 spin_unlock(&c
->journal
.lock
);
769 spin_unlock(&c
->journal
.lock
);
776 void bch_journal_meta(struct cache_set
*c
, struct closure
*cl
)
781 bch_keylist_init(&keys
);
783 ref
= bch_journal(c
, &keys
, cl
);
788 void bch_journal_free(struct cache_set
*c
)
790 free_pages((unsigned long) c
->journal
.w
[1].data
, JSET_BITS
);
791 free_pages((unsigned long) c
->journal
.w
[0].data
, JSET_BITS
);
792 free_fifo(&c
->journal
.pin
);
795 int bch_journal_alloc(struct cache_set
*c
)
797 struct journal
*j
= &c
->journal
;
799 spin_lock_init(&j
->lock
);
800 INIT_DELAYED_WORK(&j
->work
, journal_write_work
);
802 c
->journal_delay_ms
= 100;
807 if (!(init_fifo(&j
->pin
, JOURNAL_PIN
, GFP_KERNEL
)) ||
808 !(j
->w
[0].data
= (void *) __get_free_pages(GFP_KERNEL
, JSET_BITS
)) ||
809 !(j
->w
[1].data
= (void *) __get_free_pages(GFP_KERNEL
, JSET_BITS
)))