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;
84 op->status = BLK_STS_RESOURCE;
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);
181 if (bio->bi_status) {
182 /* TODO: We could try to recover from this. */
184 op->status = bio->bi_status;
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_status, "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)
203 return bch_data_invalidate(cl);
206 * Journal writes are marked REQ_PREFLUSH; if the original write was a
207 * flush, it'll wait on the journal write.
209 bio->bi_opf &= ~(REQ_PREFLUSH|REQ_FUA);
214 struct bio_set *split = op->c->bio_split;
216 /* 1 for the device pointer and 1 for the chksum */
217 if (bch_keylist_realloc(&op->insert_keys,
218 3 + (op->csum ? 1 : 0),
220 continue_at(cl, bch_data_insert_keys, op->wq);
224 k = op->insert_keys.top;
226 SET_KEY_INODE(k, op->inode);
227 SET_KEY_OFFSET(k, bio->bi_iter.bi_sector);
229 if (!bch_alloc_sectors(op->c, k, bio_sectors(bio),
230 op->write_point, op->write_prio,
234 n = bio_next_split(bio, KEY_SIZE(k), GFP_NOIO, split);
236 n->bi_end_io = bch_data_insert_endio;
240 SET_KEY_DIRTY(k, true);
242 for (i = 0; i < KEY_PTRS(k); i++)
243 SET_GC_MARK(PTR_BUCKET(op->c, k, i),
247 SET_KEY_CSUM(k, op->csum);
251 trace_bcache_cache_insert(k);
252 bch_keylist_push(&op->insert_keys);
254 bio_set_op_attrs(n, REQ_OP_WRITE, 0);
255 bch_submit_bbio(n, op->c, k, 0);
258 op->insert_data_done = true;
259 continue_at(cl, bch_data_insert_keys, op->wq);
262 /* bch_alloc_sectors() blocks if s->writeback = true */
263 BUG_ON(op->writeback);
266 * But if it's not a writeback write we'd rather just bail out if
267 * there aren't any buckets ready to write to - it might take awhile and
268 * we might be starving btree writes for gc or something.
273 * Writethrough write: We can't complete the write until we've
274 * updated the index. But we don't want to delay the write while
275 * we wait for buckets to be freed up, so just invalidate the
279 return bch_data_invalidate(cl);
282 * From a cache miss, we can just insert the keys for the data
283 * we have written or bail out if we didn't do anything.
285 op->insert_data_done = true;
288 if (!bch_keylist_empty(&op->insert_keys))
289 continue_at(cl, bch_data_insert_keys, op->wq);
296 * bch_data_insert - stick some data in the cache
298 * This is the starting point for any data to end up in a cache device; it could
299 * be from a normal write, or a writeback write, or a write to a flash only
300 * volume - it's also used by the moving garbage collector to compact data in
301 * mostly empty buckets.
303 * It first writes the data to the cache, creating a list of keys to be inserted
304 * (if the data had to be fragmented there will be multiple keys); after the
305 * data is written it calls bch_journal, and after the keys have been added to
306 * the next journal write they're inserted into the btree.
308 * It inserts the data in s->cache_bio; bi_sector is used for the key offset,
309 * and op->inode is used for the key inode.
311 * If s->bypass is true, instead of inserting the data it invalidates the
312 * region of the cache represented by s->cache_bio and op->inode.
314 void bch_data_insert(struct closure *cl)
316 struct data_insert_op *op = container_of(cl, struct data_insert_op, cl);
318 trace_bcache_write(op->c, op->inode, op->bio,
319 op->writeback, op->bypass);
321 bch_keylist_init(&op->insert_keys);
323 bch_data_insert_start(cl);
328 unsigned bch_get_congested(struct cache_set *c)
333 if (!c->congested_read_threshold_us &&
334 !c->congested_write_threshold_us)
337 i = (local_clock_us() - c->congested_last_us) / 1024;
341 i += atomic_read(&c->congested);
348 i = fract_exp_two(i, 6);
350 rand = get_random_int();
351 i -= bitmap_weight(&rand, BITS_PER_LONG);
353 return i > 0 ? i : 1;
356 static void add_sequential(struct task_struct *t)
358 ewma_add(t->sequential_io_avg,
359 t->sequential_io, 8, 0);
361 t->sequential_io = 0;
364 static struct hlist_head *iohash(struct cached_dev *dc, uint64_t k)
366 return &dc->io_hash[hash_64(k, RECENT_IO_BITS)];
369 static bool check_should_bypass(struct cached_dev *dc, struct bio *bio)
371 struct cache_set *c = dc->disk.c;
372 unsigned mode = cache_mode(dc, bio);
373 unsigned sectors, congested = bch_get_congested(c);
374 struct task_struct *task = current;
377 if (test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags) ||
378 c->gc_stats.in_use > CUTOFF_CACHE_ADD ||
379 (bio_op(bio) == REQ_OP_DISCARD))
382 if (mode == CACHE_MODE_NONE ||
383 (mode == CACHE_MODE_WRITEAROUND &&
384 op_is_write(bio_op(bio))))
387 if (bio->bi_iter.bi_sector & (c->sb.block_size - 1) ||
388 bio_sectors(bio) & (c->sb.block_size - 1)) {
389 pr_debug("skipping unaligned io");
393 if (bypass_torture_test(dc)) {
394 if ((get_random_int() & 3) == 3)
400 if (!congested && !dc->sequential_cutoff)
404 mode == CACHE_MODE_WRITEBACK &&
405 op_is_write(bio->bi_opf) &&
406 op_is_sync(bio->bi_opf))
409 spin_lock(&dc->io_lock);
411 hlist_for_each_entry(i, iohash(dc, bio->bi_iter.bi_sector), hash)
412 if (i->last == bio->bi_iter.bi_sector &&
413 time_before(jiffies, i->jiffies))
416 i = list_first_entry(&dc->io_lru, struct io, lru);
418 add_sequential(task);
421 if (i->sequential + bio->bi_iter.bi_size > i->sequential)
422 i->sequential += bio->bi_iter.bi_size;
424 i->last = bio_end_sector(bio);
425 i->jiffies = jiffies + msecs_to_jiffies(5000);
426 task->sequential_io = i->sequential;
429 hlist_add_head(&i->hash, iohash(dc, i->last));
430 list_move_tail(&i->lru, &dc->io_lru);
432 spin_unlock(&dc->io_lock);
434 sectors = max(task->sequential_io,
435 task->sequential_io_avg) >> 9;
437 if (dc->sequential_cutoff &&
438 sectors >= dc->sequential_cutoff >> 9) {
439 trace_bcache_bypass_sequential(bio);
443 if (congested && sectors >= congested) {
444 trace_bcache_bypass_congested(bio);
449 bch_rescale_priorities(c, bio_sectors(bio));
452 bch_mark_sectors_bypassed(c, dc, bio_sectors(bio));
459 /* Stack frame for bio_complete */
463 struct bio *orig_bio;
464 struct bio *cache_miss;
465 struct bcache_device *d;
467 unsigned insert_bio_sectors;
468 unsigned recoverable:1;
470 unsigned read_dirty_data:1;
472 unsigned long start_time;
475 struct data_insert_op iop;
478 static void bch_cache_read_endio(struct bio *bio)
480 struct bbio *b = container_of(bio, struct bbio, bio);
481 struct closure *cl = bio->bi_private;
482 struct search *s = container_of(cl, struct search, cl);
485 * If the bucket was reused while our bio was in flight, we might have
486 * read the wrong data. Set s->error but not error so it doesn't get
487 * counted against the cache device, but we'll still reread the data
488 * from the backing device.
492 s->iop.status = bio->bi_status;
493 else if (!KEY_DIRTY(&b->key) &&
494 ptr_stale(s->iop.c, &b->key, 0)) {
495 atomic_long_inc(&s->iop.c->cache_read_races);
496 s->iop.status = BLK_STS_IOERR;
499 bch_bbio_endio(s->iop.c, bio, bio->bi_status, "reading from cache");
503 * Read from a single key, handling the initial cache miss if the key starts in
504 * the middle of the bio
506 static int cache_lookup_fn(struct btree_op *op, struct btree *b, struct bkey *k)
508 struct search *s = container_of(op, struct search, op);
509 struct bio *n, *bio = &s->bio.bio;
510 struct bkey *bio_key;
513 if (bkey_cmp(k, &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0)) <= 0)
516 if (KEY_INODE(k) != s->iop.inode ||
517 KEY_START(k) > bio->bi_iter.bi_sector) {
518 unsigned bio_sectors = bio_sectors(bio);
519 unsigned sectors = KEY_INODE(k) == s->iop.inode
520 ? min_t(uint64_t, INT_MAX,
521 KEY_START(k) - bio->bi_iter.bi_sector)
524 int ret = s->d->cache_miss(b, s, bio, sectors);
525 if (ret != MAP_CONTINUE)
528 /* if this was a complete miss we shouldn't get here */
529 BUG_ON(bio_sectors <= sectors);
535 /* XXX: figure out best pointer - for multiple cache devices */
538 PTR_BUCKET(b->c, k, ptr)->prio = INITIAL_PRIO;
541 s->read_dirty_data = true;
543 n = bio_next_split(bio, min_t(uint64_t, INT_MAX,
544 KEY_OFFSET(k) - bio->bi_iter.bi_sector),
545 GFP_NOIO, s->d->bio_split);
547 bio_key = &container_of(n, struct bbio, bio)->key;
548 bch_bkey_copy_single_ptr(bio_key, k, ptr);
550 bch_cut_front(&KEY(s->iop.inode, n->bi_iter.bi_sector, 0), bio_key);
551 bch_cut_back(&KEY(s->iop.inode, bio_end_sector(n), 0), bio_key);
553 n->bi_end_io = bch_cache_read_endio;
554 n->bi_private = &s->cl;
557 * The bucket we're reading from might be reused while our bio
558 * is in flight, and we could then end up reading the wrong
561 * We guard against this by checking (in cache_read_endio()) if
562 * the pointer is stale again; if so, we treat it as an error
563 * and reread from the backing device (but we don't pass that
564 * error up anywhere).
567 __bch_submit_bbio(n, b->c);
568 return n == bio ? MAP_DONE : MAP_CONTINUE;
571 static void cache_lookup(struct closure *cl)
573 struct search *s = container_of(cl, struct search, iop.cl);
574 struct bio *bio = &s->bio.bio;
577 bch_btree_op_init(&s->op, -1);
579 ret = bch_btree_map_keys(&s->op, s->iop.c,
580 &KEY(s->iop.inode, bio->bi_iter.bi_sector, 0),
581 cache_lookup_fn, MAP_END_KEY);
582 if (ret == -EAGAIN) {
583 continue_at(cl, cache_lookup, bcache_wq);
590 /* Common code for the make_request functions */
592 static void request_endio(struct bio *bio)
594 struct closure *cl = bio->bi_private;
596 if (bio->bi_status) {
597 struct search *s = container_of(cl, struct search, cl);
598 s->iop.status = bio->bi_status;
599 /* Only cache read errors are recoverable */
600 s->recoverable = false;
607 static void bio_complete(struct search *s)
610 struct request_queue *q = s->orig_bio->bi_disk->queue;
611 generic_end_io_acct(q, bio_data_dir(s->orig_bio),
612 &s->d->disk->part0, s->start_time);
614 trace_bcache_request_end(s->d, s->orig_bio);
615 s->orig_bio->bi_status = s->iop.status;
616 bio_endio(s->orig_bio);
621 static void do_bio_hook(struct search *s, struct bio *orig_bio)
623 struct bio *bio = &s->bio.bio;
625 bio_init(bio, NULL, 0);
626 __bio_clone_fast(bio, orig_bio);
627 bio->bi_end_io = request_endio;
628 bio->bi_private = &s->cl;
633 static void search_free(struct closure *cl)
635 struct search *s = container_of(cl, struct search, cl);
641 closure_debug_destroy(cl);
642 mempool_free(s, s->d->c->search);
645 static inline struct search *search_alloc(struct bio *bio,
646 struct bcache_device *d)
650 s = mempool_alloc(d->c->search, GFP_NOIO);
652 closure_init(&s->cl, NULL);
656 s->cache_miss = NULL;
659 s->write = op_is_write(bio_op(bio));
660 s->read_dirty_data = 0;
661 s->start_time = jiffies;
665 s->iop.inode = d->id;
666 s->iop.write_point = hash_long((unsigned long) current, 16);
667 s->iop.write_prio = 0;
670 s->iop.flush_journal = op_is_flush(bio->bi_opf);
671 s->iop.wq = bcache_wq;
678 static void cached_dev_bio_complete(struct closure *cl)
680 struct search *s = container_of(cl, struct search, cl);
681 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
689 static void cached_dev_cache_miss_done(struct closure *cl)
691 struct search *s = container_of(cl, struct search, cl);
693 if (s->iop.replace_collision)
694 bch_mark_cache_miss_collision(s->iop.c, s->d);
697 bio_free_pages(s->iop.bio);
699 cached_dev_bio_complete(cl);
702 static void cached_dev_read_error(struct closure *cl)
704 struct search *s = container_of(cl, struct search, cl);
705 struct bio *bio = &s->bio.bio;
707 if (s->recoverable) {
708 /* Retry from the backing device: */
709 trace_bcache_read_retry(s->orig_bio);
712 do_bio_hook(s, s->orig_bio);
714 /* XXX: invalidate cache */
716 closure_bio_submit(bio, cl);
719 continue_at(cl, cached_dev_cache_miss_done, NULL);
722 static void cached_dev_read_done(struct closure *cl)
724 struct search *s = container_of(cl, struct search, cl);
725 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
728 * We had a cache miss; cache_bio now contains data ready to be inserted
731 * First, we copy the data we just read from cache_bio's bounce buffers
732 * to the buffers the original bio pointed to:
736 bio_reset(s->iop.bio);
737 s->iop.bio->bi_iter.bi_sector = s->cache_miss->bi_iter.bi_sector;
738 bio_copy_dev(s->iop.bio, s->cache_miss);
739 s->iop.bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
740 bch_bio_map(s->iop.bio, NULL);
742 bio_copy_data(s->cache_miss, s->iop.bio);
744 bio_put(s->cache_miss);
745 s->cache_miss = NULL;
748 if (verify(dc, &s->bio.bio) && s->recoverable && !s->read_dirty_data)
749 bch_data_verify(dc, s->orig_bio);
754 !test_bit(CACHE_SET_STOPPING, &s->iop.c->flags)) {
755 BUG_ON(!s->iop.replace);
756 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
759 continue_at(cl, cached_dev_cache_miss_done, NULL);
762 static void cached_dev_read_done_bh(struct closure *cl)
764 struct search *s = container_of(cl, struct search, cl);
765 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
767 bch_mark_cache_accounting(s->iop.c, s->d,
768 !s->cache_miss, s->iop.bypass);
769 trace_bcache_read(s->orig_bio, !s->cache_miss, s->iop.bypass);
772 continue_at_nobarrier(cl, cached_dev_read_error, bcache_wq);
773 else if (s->iop.bio || verify(dc, &s->bio.bio))
774 continue_at_nobarrier(cl, cached_dev_read_done, bcache_wq);
776 continue_at_nobarrier(cl, cached_dev_bio_complete, NULL);
779 static int cached_dev_cache_miss(struct btree *b, struct search *s,
780 struct bio *bio, unsigned sectors)
782 int ret = MAP_CONTINUE;
784 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
785 struct bio *miss, *cache_bio;
787 if (s->cache_miss || s->iop.bypass) {
788 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
789 ret = miss == bio ? MAP_DONE : MAP_CONTINUE;
793 if (!(bio->bi_opf & REQ_RAHEAD) &&
794 !(bio->bi_opf & REQ_META) &&
795 s->iop.c->gc_stats.in_use < CUTOFF_CACHE_READA)
796 reada = min_t(sector_t, dc->readahead >> 9,
797 get_capacity(bio->bi_disk) - bio_end_sector(bio));
799 s->insert_bio_sectors = min(sectors, bio_sectors(bio) + reada);
801 s->iop.replace_key = KEY(s->iop.inode,
802 bio->bi_iter.bi_sector + s->insert_bio_sectors,
803 s->insert_bio_sectors);
805 ret = bch_btree_insert_check_key(b, &s->op, &s->iop.replace_key);
809 s->iop.replace = true;
811 miss = bio_next_split(bio, sectors, GFP_NOIO, s->d->bio_split);
813 /* btree_search_recurse()'s btree iterator is no good anymore */
814 ret = miss == bio ? MAP_DONE : -EINTR;
816 cache_bio = bio_alloc_bioset(GFP_NOWAIT,
817 DIV_ROUND_UP(s->insert_bio_sectors, PAGE_SECTORS),
822 cache_bio->bi_iter.bi_sector = miss->bi_iter.bi_sector;
823 bio_copy_dev(cache_bio, miss);
824 cache_bio->bi_iter.bi_size = s->insert_bio_sectors << 9;
826 cache_bio->bi_end_io = request_endio;
827 cache_bio->bi_private = &s->cl;
829 bch_bio_map(cache_bio, NULL);
830 if (bio_alloc_pages(cache_bio, __GFP_NOWARN|GFP_NOIO))
834 bch_mark_cache_readahead(s->iop.c, s->d);
836 s->cache_miss = miss;
837 s->iop.bio = cache_bio;
839 closure_bio_submit(cache_bio, &s->cl);
845 miss->bi_end_io = request_endio;
846 miss->bi_private = &s->cl;
847 closure_bio_submit(miss, &s->cl);
851 static void cached_dev_read(struct cached_dev *dc, struct search *s)
853 struct closure *cl = &s->cl;
855 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
856 continue_at(cl, cached_dev_read_done_bh, NULL);
861 static void cached_dev_write_complete(struct closure *cl)
863 struct search *s = container_of(cl, struct search, cl);
864 struct cached_dev *dc = container_of(s->d, struct cached_dev, disk);
866 up_read_non_owner(&dc->writeback_lock);
867 cached_dev_bio_complete(cl);
870 static void cached_dev_write(struct cached_dev *dc, struct search *s)
872 struct closure *cl = &s->cl;
873 struct bio *bio = &s->bio.bio;
874 struct bkey start = KEY(dc->disk.id, bio->bi_iter.bi_sector, 0);
875 struct bkey end = KEY(dc->disk.id, bio_end_sector(bio), 0);
877 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys, &start, &end);
879 down_read_non_owner(&dc->writeback_lock);
880 if (bch_keybuf_check_overlapping(&dc->writeback_keys, &start, &end)) {
882 * We overlap with some dirty data undergoing background
883 * writeback, force this write to writeback
885 s->iop.bypass = false;
886 s->iop.writeback = true;
890 * Discards aren't _required_ to do anything, so skipping if
891 * check_overlapping returned true is ok
893 * But check_overlapping drops dirty keys for which io hasn't started,
894 * so we still want to call it.
896 if (bio_op(bio) == REQ_OP_DISCARD)
897 s->iop.bypass = true;
899 if (should_writeback(dc, s->orig_bio,
902 s->iop.bypass = false;
903 s->iop.writeback = true;
907 s->iop.bio = s->orig_bio;
910 if ((bio_op(bio) != REQ_OP_DISCARD) ||
911 blk_queue_discard(bdev_get_queue(dc->bdev)))
912 closure_bio_submit(bio, cl);
913 } else if (s->iop.writeback) {
914 bch_writeback_add(dc);
917 if (bio->bi_opf & REQ_PREFLUSH) {
918 /* Also need to send a flush to the backing device */
919 struct bio *flush = bio_alloc_bioset(GFP_NOIO, 0,
922 bio_copy_dev(flush, bio);
923 flush->bi_end_io = request_endio;
924 flush->bi_private = cl;
925 flush->bi_opf = REQ_OP_WRITE | REQ_PREFLUSH;
927 closure_bio_submit(flush, cl);
930 s->iop.bio = bio_clone_fast(bio, GFP_NOIO, dc->disk.bio_split);
932 closure_bio_submit(bio, cl);
935 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
936 continue_at(cl, cached_dev_write_complete, NULL);
939 static void cached_dev_nodata(struct closure *cl)
941 struct search *s = container_of(cl, struct search, cl);
942 struct bio *bio = &s->bio.bio;
944 if (s->iop.flush_journal)
945 bch_journal_meta(s->iop.c, cl);
947 /* If it's a flush, we send the flush to the backing device too */
948 closure_bio_submit(bio, cl);
950 continue_at(cl, cached_dev_bio_complete, NULL);
953 /* Cached devices - read & write stuff */
955 static blk_qc_t cached_dev_make_request(struct request_queue *q,
959 struct bcache_device *d = bio->bi_disk->private_data;
960 struct cached_dev *dc = container_of(d, struct cached_dev, disk);
961 int rw = bio_data_dir(bio);
963 generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
965 bio_set_dev(bio, dc->bdev);
966 bio->bi_iter.bi_sector += dc->sb.data_offset;
968 if (cached_dev_get(dc)) {
969 s = search_alloc(bio, d);
970 trace_bcache_request_start(s->d, bio);
972 if (!bio->bi_iter.bi_size) {
974 * can't call bch_journal_meta from under
975 * generic_make_request
977 continue_at_nobarrier(&s->cl,
981 s->iop.bypass = check_should_bypass(dc, bio);
984 cached_dev_write(dc, s);
986 cached_dev_read(dc, s);
989 if ((bio_op(bio) == REQ_OP_DISCARD) &&
990 !blk_queue_discard(bdev_get_queue(dc->bdev)))
993 generic_make_request(bio);
996 return BLK_QC_T_NONE;
999 static int cached_dev_ioctl(struct bcache_device *d, fmode_t mode,
1000 unsigned int cmd, unsigned long arg)
1002 struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1003 return __blkdev_driver_ioctl(dc->bdev, mode, cmd, arg);
1006 static int cached_dev_congested(void *data, int bits)
1008 struct bcache_device *d = data;
1009 struct cached_dev *dc = container_of(d, struct cached_dev, disk);
1010 struct request_queue *q = bdev_get_queue(dc->bdev);
1013 if (bdi_congested(q->backing_dev_info, bits))
1016 if (cached_dev_get(dc)) {
1020 for_each_cache(ca, d->c, i) {
1021 q = bdev_get_queue(ca->bdev);
1022 ret |= bdi_congested(q->backing_dev_info, bits);
1031 void bch_cached_dev_request_init(struct cached_dev *dc)
1033 struct gendisk *g = dc->disk.disk;
1035 g->queue->make_request_fn = cached_dev_make_request;
1036 g->queue->backing_dev_info->congested_fn = cached_dev_congested;
1037 dc->disk.cache_miss = cached_dev_cache_miss;
1038 dc->disk.ioctl = cached_dev_ioctl;
1041 /* Flash backed devices */
1043 static int flash_dev_cache_miss(struct btree *b, struct search *s,
1044 struct bio *bio, unsigned sectors)
1046 unsigned bytes = min(sectors, bio_sectors(bio)) << 9;
1048 swap(bio->bi_iter.bi_size, bytes);
1050 swap(bio->bi_iter.bi_size, bytes);
1052 bio_advance(bio, bytes);
1054 if (!bio->bi_iter.bi_size)
1057 return MAP_CONTINUE;
1060 static void flash_dev_nodata(struct closure *cl)
1062 struct search *s = container_of(cl, struct search, cl);
1064 if (s->iop.flush_journal)
1065 bch_journal_meta(s->iop.c, cl);
1067 continue_at(cl, search_free, NULL);
1070 static blk_qc_t flash_dev_make_request(struct request_queue *q,
1075 struct bcache_device *d = bio->bi_disk->private_data;
1076 int rw = bio_data_dir(bio);
1078 generic_start_io_acct(q, rw, bio_sectors(bio), &d->disk->part0);
1080 s = search_alloc(bio, d);
1084 trace_bcache_request_start(s->d, bio);
1086 if (!bio->bi_iter.bi_size) {
1088 * can't call bch_journal_meta from under
1089 * generic_make_request
1091 continue_at_nobarrier(&s->cl,
1094 return BLK_QC_T_NONE;
1096 bch_keybuf_check_overlapping(&s->iop.c->moving_gc_keys,
1097 &KEY(d->id, bio->bi_iter.bi_sector, 0),
1098 &KEY(d->id, bio_end_sector(bio), 0));
1100 s->iop.bypass = (bio_op(bio) == REQ_OP_DISCARD) != 0;
1101 s->iop.writeback = true;
1104 closure_call(&s->iop.cl, bch_data_insert, NULL, cl);
1106 closure_call(&s->iop.cl, cache_lookup, NULL, cl);
1109 continue_at(cl, search_free, NULL);
1110 return BLK_QC_T_NONE;
1113 static int flash_dev_ioctl(struct bcache_device *d, fmode_t mode,
1114 unsigned int cmd, unsigned long arg)
1119 static int flash_dev_congested(void *data, int bits)
1121 struct bcache_device *d = data;
1122 struct request_queue *q;
1127 for_each_cache(ca, d->c, i) {
1128 q = bdev_get_queue(ca->bdev);
1129 ret |= bdi_congested(q->backing_dev_info, bits);
1135 void bch_flash_dev_request_init(struct bcache_device *d)
1137 struct gendisk *g = d->disk;
1139 g->queue->make_request_fn = flash_dev_make_request;
1140 g->queue->backing_dev_info->congested_fn = flash_dev_congested;
1141 d->cache_miss = flash_dev_cache_miss;
1142 d->ioctl = flash_dev_ioctl;
1145 void bch_request_exit(void)
1147 if (bch_search_cache)
1148 kmem_cache_destroy(bch_search_cache);
1151 int __init bch_request_init(void)
1153 bch_search_cache = KMEM_CACHE(search, 0);
1154 if (!bch_search_cache)