2 * Main bcache entry point - handle a read or a write request and decide what to
3 * do with it; the make_request functions are called by the block layer.
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
13 #include "writeback.h"
15 #include <linux/module.h>
16 #include <linux/hash.h>
17 #include <linux/random.h>
18 #include <linux/backing-dev.h>
20 #include <trace/events/bcache.h>
22 #define CUTOFF_CACHE_ADD 95
23 #define CUTOFF_CACHE_READA 90
25 struct kmem_cache *bch_search_cache;
27 static void bch_data_insert_start(struct closure *);
29 static unsigned cache_mode(struct cached_dev *dc, struct bio *bio)
31 return BDEV_CACHE_MODE(&dc->sb);
34 static bool verify(struct cached_dev *dc, struct bio *bio)
39 static void bio_csum(struct bio *bio, struct bkey *k)
42 struct bvec_iter iter;
45 bio_for_each_segment(bv, bio, iter) {
46 void *d = kmap(bv.bv_page) + bv.bv_offset;
47 csum = bch_crc64_update(csum, d, bv.bv_len);
51 k->ptr[KEY_PTRS(k)] = csum & (~0ULL >> 1);
54 /* Insert data into cache */
56 static void bch_data_insert_keys(struct closure *cl)
58 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
59 atomic_t *journal_ref = NULL;
60 struct bkey *replace_key = op->replace ? &op->replace_key : NULL;
64 * If we're looping, might already be waiting on
65 * another journal write - can't wait on more than one journal write at
68 * XXX: this looks wrong
71 while (atomic_read(&s->cl.remaining) & CLOSURE_WAITING)
76 journal_ref = bch_journal(op->c, &op->insert_keys,
77 op->flush_journal ? cl : NULL);
79 ret = bch_btree_insert(op->c, &op->insert_keys,
80 journal_ref, replace_key);
82 op->replace_collision = true;
85 op->insert_data_done = true;
89 atomic_dec_bug(journal_ref);
91 if (!op->insert_data_done) {
92 continue_at(cl, bch_data_insert_start, op->wq);
96 bch_keylist_free(&op->insert_keys);
100 static int bch_keylist_realloc(struct keylist *l, unsigned u64s,
103 size_t oldsize = bch_keylist_nkeys(l);
104 size_t newsize = oldsize + u64s;
107 * The journalling code doesn't handle the case where the keys to insert
108 * is bigger than an empty write: If we just return -ENOMEM here,
109 * bio_insert() and bio_invalidate() will insert the keys created so far
110 * and finish the rest when the keylist is empty.
112 if (newsize * sizeof(uint64_t) > block_bytes(c) - sizeof(struct jset))
115 return __bch_keylist_realloc(l, u64s);
118 static void bch_data_invalidate(struct closure *cl)
120 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
121 struct bio *bio = op->bio;
123 pr_debug("invalidating %i sectors from %llu",
124 bio_sectors(bio), (uint64_t) bio->bi_iter.bi_sector);
126 while (bio_sectors(bio)) {
127 unsigned sectors = min(bio_sectors(bio),
128 1U << (KEY_SIZE_BITS - 1));
130 if (bch_keylist_realloc(&op->insert_keys, 2, op->c))
133 bio->bi_iter.bi_sector += sectors;
134 bio->bi_iter.bi_size -= sectors << 9;
136 bch_keylist_add(&op->insert_keys,
137 &KEY(op->inode, bio->bi_iter.bi_sector, sectors));
140 op->insert_data_done = true;
143 continue_at(cl, bch_data_insert_keys, op->wq);
146 static void bch_data_insert_error(struct closure *cl)
148 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
151 * Our data write just errored, which means we've got a bunch of keys to
152 * insert that point to data that wasn't succesfully written.
154 * We don't have to insert those keys but we still have to invalidate
155 * that region of the cache - so, if we just strip off all the pointers
156 * from the keys we'll accomplish just that.
159 struct bkey *src = op->insert_keys.keys, *dst = op->insert_keys.keys;
161 while (src != op->insert_keys.top) {
162 struct bkey *n = bkey_next(src);
164 SET_KEY_PTRS(src, 0);
165 memmove(dst, src, bkey_bytes(src));
167 dst = bkey_next(dst);
171 op->insert_keys.top = dst;
173 bch_data_insert_keys(cl);
176 static void bch_data_insert_endio(struct bio *bio)
178 struct closure *cl = bio->bi_private;
179 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
182 /* TODO: We could try to recover from this. */
184 op->error = bio->bi_error;
185 else if (!op->replace)
186 set_closure_fn(cl, bch_data_insert_error, op->wq);
188 set_closure_fn(cl, NULL, NULL);
191 bch_bbio_endio(op->c, bio, bio->bi_error, "writing data to cache");
194 static void bch_data_insert_start(struct closure *cl)
196 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
197 struct bio *bio = op->bio, *n;
199 if (atomic_sub_return(bio_sectors(bio), &op->c->sectors_to_gc) < 0) {
200 set_gc_sectors(op->c);
205 return bch_data_invalidate(cl);
208 * Journal writes are marked REQ_FLUSH; if the original write was a
209 * flush, it'll wait on the journal write.
211 bio->bi_rw &= ~(REQ_FLUSH|REQ_FUA);
216 struct bio_set *split = op->c->bio_split;
218 /* 1 for the device pointer and 1 for the chksum */
219 if (bch_keylist_realloc(&op->insert_keys,
220 3 + (op->csum ? 1 : 0),
222 continue_at(cl, bch_data_insert_keys, op->wq);
226 k = op->insert_keys.top;
228 SET_KEY_INODE(k, op->inode);
229 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
231 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
232 op->write_point, op->write_prio,
236 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
238 n->bi_end_io = bch_data_insert_endio;
242 SET_KEY_DIRTY(k, true);
244 for (i = 0; i < KEY_PTRS(k); i++)
245 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
249 SET_KEY_CSUM(k, op->csum);
253 trace_bcache_cache_insert(k);
254 bch_keylist_push(&op->insert_keys);
256 n->bi_rw |= REQ_WRITE;
257 bch_submit_bbio(n, op->c, k, 0);
260 op->insert_data_done = true;
261 continue_at(cl, bch_data_insert_keys, op->wq);
264 /* bch_alloc_sectors() blocks if s->writeback = true */
265 BUG_ON(op->writeback);
268 * But if it's not a writeback write we'd rather just bail out if
269 * there aren't any buckets ready to write to - it might take awhile and
270 * we might be starving btree writes for gc or something.
275 * Writethrough write: We can't complete the write until we've
276 * updated the index. But we don't want to delay the write while
277 * we wait for buckets to be freed up, so just invalidate the
281 return bch_data_invalidate(cl);
284 * From a cache miss, we can just insert the keys for the data
285 * we have written or bail out if we didn't do anything.
287 op->insert_data_done = true;
290 if (!bch_keylist_empty(&op->insert_keys))
291 continue_at(cl, bch_data_insert_keys, op->wq);
298 * bch_data_insert - stick some data in the cache
300 * This is the starting point for any data to end up in a cache device; it could
301 * be from a normal write, or a writeback write, or a write to a flash only
302 * volume - it's also used by the moving garbage collector to compact data in
303 * mostly empty buckets.
305 * It first writes the data to the cache, creating a list of keys to be inserted
306 * (if the data had to be fragmented there will be multiple keys); after the
307 * data is written it calls bch_journal, and after the keys have been added to
308 * the next journal write they're inserted into the btree.
310 * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
311 * and op->inode is used for the key inode.
313 * If s->bypass is true, instead of inserting the data it invalidates the
314 * region of the cache represented by s->cache_bio and op->inode.
316 void bch_data_insert(struct closure *cl)
318 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
320 trace_bcache_write(op->c, op->inode, op->bio,
321 op->writeback, op->bypass);
323 bch_keylist_init(&op->insert_keys);
325 bch_data_insert_start(cl);
330 unsigned bch_get_congested(struct cache_set *c)
335 if (!c->congested_read_threshold_us &&
336 !c->congested_write_threshold_us)
339 i = (local_clock_us() - c->congested_last_us) / 1024;
343 i += atomic_read(&c->congested);
350 i = fract_exp_two(i, 6);
352 rand = get_random_int();
353 i -= bitmap_weight(&rand, BITS_PER_LONG);
355 return i > 0 ? i : 1;
358 static void add_sequential(struct task_struct *t)
360 ewma_add(t->sequential_io_avg,
361 t->sequential_io, 8, 0);
363 t->sequential_io = 0;
366 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
368 return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
371 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
373 struct cache_set *c = dc->disk.c;
374 unsigned mode = cache_mode(dc, bio);
375 unsigned sectors, congested = bch_get_congested(c);
376 struct task_struct *task = current;
379 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
380 c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
381 (bio->bi_rw & REQ_DISCARD))
384 if (mode == CACHE_MODE_NONE ||
385 (mode == CACHE_MODE_WRITEAROUND &&
386 (bio->bi_rw & REQ_WRITE)))
389 if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
390 bio_sectors(bio) & (c->sb.block_size - 1)) {
391 pr_debug("skipping unaligned io");
395 if (bypass_torture_test(dc)) {
396 if ((get_random_int() & 3) == 3)
402 if (!congested && !dc->sequential_cutoff)
406 mode == CACHE_MODE_WRITEBACK &&
407 (bio->bi_rw & REQ_WRITE) &&
408 (bio->bi_rw & REQ_SYNC))
411 spin_lock(&dc->io_lock);
413 hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
414 if (i->last == bio->bi_iter.bi_sector &&
415 time_before(jiffies, i->jiffies))
418 i = list_first_entry(&dc->io_lru, struct io, lru);
420 add_sequential(task);
423 if (i->sequential + bio->bi_iter.bi_size > i->sequential)
424 i->sequential += bio->bi_iter.bi_size;
426 i->last = bio_end_sector(bio);
427 i->jiffies = jiffies + msecs_to_jiffies(5000);
428 task->sequential_io = i->sequential;
431 hlist_add_head(&i->hash, iohash(dc, i->last));
432 list_move_tail(&i->lru, &dc->io_lru);
434 spin_unlock(&dc->io_lock);
436 sectors = max(task->sequential_io,
437 task->sequential_io_avg) >> 9;
439 if (dc->sequential_cutoff &&
440 sectors >= dc->sequential_cutoff >> 9) {
441 trace_bcache_bypass_sequential(bio);
445 if (congested && sectors >= congested) {
446 trace_bcache_bypass_congested(bio);
451 bch_rescale_priorities(c, bio_sectors(bio));
454 bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
461 /* Stack frame for bio_complete */
465 struct bio *orig_bio;
466 struct bio *cache_miss;
467 struct bcache_device *d;
469 unsigned insert_bio_sectors;
470 unsigned recoverable:1;
472 unsigned read_dirty_data:1;
474 unsigned long start_time;
477 struct data_insert_op iop;
480 static void bch_cache_read_endio(struct bio *bio)
482 struct bbio *b = container_of(bio, struct bbio, bio);
483 struct closure *cl = bio->bi_private;
484 struct search *s = container_of(cl, struct search, cl);
487 * If the bucket was reused while our bio was in flight, we might have
488 * read the wrong data. Set s->error but not error so it doesn't get
489 * counted against the cache device, but we'll still reread the data
490 * from the backing device.
494 s->iop.error = bio->bi_error;
495 else if (!KEY_DIRTY(&b->key) &&
496 ptr_stale(s->iop.c, &b->key, 0)) {
497 atomic_long_inc(&s->iop.c->cache_read_races);
498 s->iop.error = -EINTR;
501 bch_bbio_endio(s->iop.c, bio, bio->bi_error, "reading from cache");
505 * Read from a single key, handling the initial cache miss if the key starts in
506 * the middle of the bio
508 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
510 struct search *s = container_of(op, struct search, op);
511 struct bio *n, *bio = &s->bio.bio;
512 struct bkey *bio_key;
515 if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
518 if (KEY_INODE(k) != s->iop.inode ||
519 KEY_START(k) > bio->bi_iter.bi_sector) {
520 unsigned bio_sectors = bio_sectors(bio);
521 unsigned sectors = KEY_INODE(k) == s->iop.inode
522 ? min_t(uint64_t, INT_MAX,
523 KEY_START(k) - bio->bi_iter.bi_sector)
526 int ret = s->d->cache_miss(b, s, bio, sectors);
527 if (ret != MAP_CONTINUE)
530 /* if this was a complete miss we shouldn't get here */
531 BUG_ON(bio_sectors <= sectors);
537 /* XXX: figure out best pointer - for multiple cache devices */
540 PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
543 s->read_dirty_data = true;
545 n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
546 KEY_OFFSET(k) - bio->bi_iter.bi_sector),
547 GFP_NOIO, s->d->bio_split);
549 bio_key = &container_of(n, struct bbio, bio)->key;
550 bch_bkey_copy_single_ptr(bio_key, k, ptr);
552 bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
553 bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
555 n->bi_end_io = bch_cache_read_endio;
556 n->bi_private = &s->cl;
559 * The bucket we're reading from might be reused while our bio
560 * is in flight, and we could then end up reading the wrong
563 * We guard against this by checking (in cache_read_endio()) if
564 * the pointer is stale again; if so, we treat it as an error
565 * and reread from the backing device (but we don't pass that
566 * error up anywhere).
569 __bch_submit_bbio(n, b->c);
570 return n == bio ? MAP_DONE : MAP_CONTINUE;
573 static void cache_lookup(struct closure *cl)
575 struct search *s = container_of(cl, struct search, iop.cl);
576 struct bio *bio = &s->bio.bio;
579 bch_btree_op_init(&s->op, -1);
581 ret = bch_btree_map_keys(&s->op, s->iop.c,
582 &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
583 cache_lookup_fn, MAP_END_KEY);
584 if (ret == -EAGAIN) {
585 continue_at(cl, cache_lookup, bcache_wq);
592 /* Common code for the make_request functions */
594 static void request_endio(struct bio *bio)
596 struct closure *cl = bio->bi_private;
599 struct search *s = container_of(cl, struct search, cl);
600 s->iop.error = bio->bi_error;
601 /* Only cache read errors are recoverable */
602 s->recoverable = false;
609 static void bio_complete(struct search *s)
612 generic_end_io_acct(bio_data_dir(s->orig_bio),
613 &s->d->disk->part0, s->start_time);
615 trace_bcache_request_end(s->d, s->orig_bio);
616 s->orig_bio->bi_error = s->iop.error;
617 bio_endio(s->orig_bio);
622 static void do_bio_hook(struct search *s, struct bio *orig_bio)
624 struct bio *bio = &s->bio.bio;
627 __bio_clone_fast(bio, orig_bio);
628 bio->bi_end_io = request_endio;
629 bio->bi_private = &s->cl;
634 static void search_free(struct closure *cl)
636 struct search *s = container_of(cl, struct search, cl);
642 closure_debug_destroy(cl);
643 mempool_free(s, s->d->c->search);
646 static inline struct search *search_alloc(struct bio *bio,
647 struct bcache_device *d)
651 s = mempool_alloc(d->c->search, GFP_NOIO);
653 closure_init(&s->cl, NULL);
657 s->cache_miss = NULL;
660 s->write = (bio->bi_rw & REQ_WRITE) != 0;
661 s->read_dirty_data = 0;
662 s->start_time = jiffies;
666 s->iop.inode = d->id;
667 s->iop.write_point = hash_long((unsigned long) current, 16);
668 s->iop.write_prio = 0;
671 s->iop.flush_journal = (bio->bi_rw & (REQ_FLUSH|REQ_FUA)) != 0;
672 s->iop.wq = bcache_wq;
679 static void cached_dev_bio_complete(struct closure *cl)
681 struct search *s = container_of(cl, struct search, cl);
682 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
690 static void cached_dev_cache_miss_done(struct closure *cl)
692 struct search *s = container_of(cl, struct search, cl);
694 if (s->iop.replace_collision)
695 bch_mark_cache_miss_collision(s->iop.c, s->d);
701 bio_for_each_segment_all(bv, s->iop.bio, i)
702 __free_page(bv->bv_page);
705 cached_dev_bio_complete(cl);
708 static void cached_dev_read_error(struct closure *cl)
710 struct search *s = container_of(cl, struct search, cl);
711 struct bio *bio = &s->bio.bio;
713 if (s->recoverable) {
714 /* Retry from the backing device: */
715 trace_bcache_read_retry(s->orig_bio);
718 do_bio_hook(s, s->orig_bio);
720 /* XXX: invalidate cache */
722 closure_bio_submit(bio, cl);
725 continue_at(cl, cached_dev_cache_miss_done, NULL);
728 static void cached_dev_read_done(struct closure *cl)
730 struct search *s = container_of(cl, struct search, cl);
731 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
734 * We had a cache miss; cache_bio now contains data ready to be inserted
737 * First, we copy the data we just read from cache_bio's bounce buffers
738 * to the buffers the original bio pointed to:
742 bio_reset(s->iop.bio);
743 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
744 s->iop.bio->bi_bdev = s->cache_miss->bi_bdev;
745 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
746 bch_bio_map(s->iop.bio, NULL);
748 bio_copy_data(s->cache_miss, s->iop.bio);
750 bio_put(s->cache_miss);
751 s->cache_miss = NULL;
754 if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
755 bch_data_verify(dc, s->orig_bio);
760 !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
761 BUG_ON(!s->iop.replace);
762 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
765 continue_at(cl, cached_dev_cache_miss_done, NULL);
768 static void cached_dev_read_done_bh(struct closure *cl)
770 struct search *s = container_of(cl, struct search, cl);
771 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
773 bch_mark_cache_accounting(s->iop.c, s->d,
774 !s->cache_miss, s->iop.bypass);
775 trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
778 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
779 else if (s->iop.bio || verify(dc, &s->bio.bio))
780 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
782 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
785 static int cached_dev_cache_miss(struct btree *b, struct search *s,
786 struct bio *bio, unsigned sectors)
788 int ret = MAP_CONTINUE;
790 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
791 struct bio *miss, *cache_bio;
793 if (s->cache_miss || s->iop.bypass) {
794 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
795 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
799 if (!(bio->bi_rw & REQ_RAHEAD) &&
800 !(bio->bi_rw & REQ_META) &&
801 s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
802 reada = min_t(sector_t, dc->readahead >> 9,
803 bdev_sectors(bio->bi_bdev) - bio_end_sector(bio));
805 s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
807 s->iop.replace_key = KEY(s->iop.inode,
808 bio->bi_iter.bi_sector + s->insert_bio_sectors,
809 s->insert_bio_sectors);
811 ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
815 s->iop.replace = true;
817 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
819 /* btree_search_recurse()'s btree iterator is no good anymore */
820 ret = miss == bio ? MAP_DONE : -EINTR;
822 cache_bio = bio_alloc_bioset(GFP_NOWAIT,
823 DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
828 cache_bio->bi_iter.bi_sector = miss->bi_iter.bi_sector;
829 cache_bio->bi_bdev = miss->bi_bdev;
830 cache_bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
832 cache_bio->bi_end_io = request_endio;
833 cache_bio->bi_private = &s->cl;
835 bch_bio_map(cache_bio, NULL);
836 if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
840 bch_mark_cache_readahead(s->iop.c, s->d);
842 s->cache_miss = miss;
843 s->iop.bio = cache_bio;
845 closure_bio_submit(cache_bio, &s->cl);
851 miss->bi_end_io = request_endio;
852 miss->bi_private = &s->cl;
853 closure_bio_submit(miss, &s->cl);
857 static void cached_dev_read(struct cached_dev *dc, struct search *s)
859 struct closure *cl = &s->cl;
861 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
862 continue_at(cl, cached_dev_read_done_bh, NULL);
867 static void cached_dev_write_complete(struct closure *cl)
869 struct search *s = container_of(cl, struct search, cl);
870 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
872 up_read_non_owner(&dc->writeback_lock);
873 cached_dev_bio_complete(cl);
876 static void cached_dev_write(struct cached_dev *dc, struct search *s)
878 struct closure *cl = &s->cl;
879 struct bio *bio = &s->bio.bio;
880 struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
881 struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
883 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
885 down_read_non_owner(&dc->writeback_lock);
886 if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
888 * We overlap with some dirty data undergoing background
889 * writeback, force this write to writeback
891 s->iop.bypass = false;
892 s->iop.writeback = true;
896 * Discards aren't _required_ to do anything, so skipping if
897 * check_overlapping returned true is ok
899 * But check_overlapping drops dirty keys for which io hasn't started,
900 * so we still want to call it.
902 if (bio->bi_rw & REQ_DISCARD)
903 s->iop.bypass = true;
905 if (should_writeback(dc, s->orig_bio,
908 s->iop.bypass = false;
909 s->iop.writeback = true;
913 s->iop.bio = s->orig_bio;
916 if (!(bio->bi_rw & REQ_DISCARD) ||
917 blk_queue_discard(bdev_get_queue(dc->bdev)))
918 closure_bio_submit(bio, cl);
919 } else if (s->iop.writeback) {
920 bch_writeback_add(dc);
923 if (bio->bi_rw & REQ_FLUSH) {
924 /* Also need to send a flush to the backing device */
925 struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
928 flush->bi_rw = WRITE_FLUSH;
929 flush->bi_bdev = bio->bi_bdev;
930 flush->bi_end_io = request_endio;
931 flush->bi_private = cl;
933 closure_bio_submit(flush, cl);
936 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
938 closure_bio_submit(bio, cl);
941 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
942 continue_at(cl, cached_dev_write_complete, NULL);
945 static void cached_dev_nodata(struct closure *cl)
947 struct search *s = container_of(cl, struct search, cl);
948 struct bio *bio = &s->bio.bio;
950 if (s->iop.flush_journal)
951 bch_journal_meta(s->iop.c, cl);
953 /* If it's a flush, we send the flush to the backing device too */
954 closure_bio_submit(bio, cl);
956 continue_at(cl, cached_dev_bio_complete, NULL);
959 /* Cached devices - read & write stuff */
961 static blk_qc_t cached_dev_make_request(struct request_queue *q,
965 struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
966 struct cached_dev *dc = container_of(d, struct cached_dev, disk);
967 int rw = bio_data_dir(bio);
969 generic_start_io_acct(rw, bio_sectors(bio), &d->disk->part0);
971 bio->bi_bdev = dc->bdev;
972 bio->bi_iter.bi_sector += dc->sb.data_offset;
974 if (cached_dev_get(dc)) {
975 s = search_alloc(bio, d);
976 trace_bcache_request_start(s->d, bio);
978 if (!bio->bi_iter.bi_size) {
980 * can't call bch_journal_meta from under
981 * generic_make_request
983 continue_at_nobarrier(&s->cl,
987 s->iop.bypass = check_should_bypass(dc, bio);
990 cached_dev_write(dc, s);
992 cached_dev_read(dc, s);
995 if ((bio->bi_rw & REQ_DISCARD) &&
996 !blk_queue_discard(bdev_get_queue(dc->bdev)))
999 generic_make_request(bio);
1002 return BLK_QC_T_NONE;
1005 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1006 unsigned int cmd, unsigned long arg)
1008 struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1009 return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1012 static int cached_dev_congested(void *data, int bits)
1014 struct bcache_device *d = data;
1015 struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1016 struct request_queue *q = bdev_get_queue(dc->bdev);
1019 if (bdi_congested(&q->backing_dev_info, bits))
1022 if (cached_dev_get(dc)) {
1026 for_each_cache(ca, d->c, i) {
1027 q = bdev_get_queue(ca->bdev);
1028 ret |= bdi_congested(&q->backing_dev_info, bits);
1037 void bch_cached_dev_request_init(struct cached_dev *dc)
1039 struct gendisk *g = dc->disk.disk;
1041 g->queue->make_request_fn = cached_dev_make_request;
1042 g->queue->backing_dev_info.congested_fn = cached_dev_congested;
1043 dc->disk.cache_miss = cached_dev_cache_miss;
1044 dc->disk.ioctl = cached_dev_ioctl;
1047 /* Flash backed devices */
1049 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1050 struct bio *bio, unsigned sectors)
1052 unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1054 swap(bio->bi_iter.bi_size, bytes);
1056 swap(bio->bi_iter.bi_size, bytes);
1058 bio_advance(bio, bytes);
1060 if (!bio->bi_iter.bi_size)
1063 return MAP_CONTINUE;
1066 static void flash_dev_nodata(struct closure *cl)
1068 struct search *s = container_of(cl, struct search, cl);
1070 if (s->iop.flush_journal)
1071 bch_journal_meta(s->iop.c, cl);
1073 continue_at(cl, search_free, NULL);
1076 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1081 struct bcache_device *d = bio->bi_bdev->bd_disk->private_data;
1082 int rw = bio_data_dir(bio);
1084 generic_start_io_acct(rw, bio_sectors(bio), &d->disk->part0);
1086 s = search_alloc(bio, d);
1090 trace_bcache_request_start(s->d, bio);
1092 if (!bio->bi_iter.bi_size) {
1094 * can't call bch_journal_meta from under
1095 * generic_make_request
1097 continue_at_nobarrier(&s->cl,
1100 return BLK_QC_T_NONE;
1102 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1103 &KEY(d->id, bio->bi_iter.bi_sector, 0),
1104 &KEY(d->id, bio_end_sector(bio), 0));
1106 s->iop.bypass = (bio->bi_rw & REQ_DISCARD) != 0;
1107 s->iop.writeback = true;
1110 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1112 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1115 continue_at(cl, search_free, NULL);
1116 return BLK_QC_T_NONE;
1119 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1120 unsigned int cmd, unsigned long arg)
1125 static int flash_dev_congested(void *data, int bits)
1127 struct bcache_device *d = data;
1128 struct request_queue *q;
1133 for_each_cache(ca, d->c, i) {
1134 q = bdev_get_queue(ca->bdev);
1135 ret |= bdi_congested(&q->backing_dev_info, bits);
1141 void bch_flash_dev_request_init(struct bcache_device *d)
1143 struct gendisk *g = d->disk;
1145 g->queue->make_request_fn = flash_dev_make_request;
1146 g->queue->backing_dev_info.congested_fn = flash_dev_congested;
1147 d->cache_miss = flash_dev_cache_miss;
1148 d->ioctl = flash_dev_ioctl;
1151 void bch_request_exit(void)
1153 if (bch_search_cache)
1154 kmem_cache_destroy(bch_search_cache);
1157 int __init bch_request_init(void)
1159 bch_search_cache = KMEM_CACHE(search, 0);
1160 if (!bch_search_cache)