2 * bcache setup/teardown code, and some metadata io - read a superblock and
3 * figure out what to do with it.
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
14 #include "writeback.h"
16 #include <linux/blkdev.h>
17 #include <linux/buffer_head.h>
18 #include <linux/debugfs.h>
19 #include <linux/genhd.h>
20 #include <linux/idr.h>
21 #include <linux/kthread.h>
22 #include <linux/module.h>
23 #include <linux/random.h>
24 #include <linux/reboot.h>
25 #include <linux/sysfs.h>
27 MODULE_LICENSE("GPL");
28 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
30 static const char bcache_magic[] = {
31 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
32 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
35 static const char invalid_uuid[] = {
36 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
37 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
40 /* Default is -1; we skip past it for struct cached_dev's cache mode */
41 const char * const bch_cache_modes[] = {
50 /* Default is -1; we skip past it for stop_when_cache_set_failed */
51 const char * const bch_stop_on_failure_modes[] = {
58 static struct kobject *bcache_kobj;
59 struct mutex bch_register_lock;
60 LIST_HEAD(bch_cache_sets);
61 static LIST_HEAD(uncached_devices);
63 static int bcache_major;
64 static DEFINE_IDA(bcache_device_idx);
65 static wait_queue_head_t unregister_wait;
66 struct workqueue_struct *bcache_wq;
68 #define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
69 /* limitation of partitions number on single bcache device */
70 #define BCACHE_MINORS 128
71 /* limitation of bcache devices number on single system */
72 #define BCACHE_DEVICE_IDX_MAX ((1U << MINORBITS)/BCACHE_MINORS)
76 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
81 struct buffer_head *bh = __bread(bdev, 1, SB_SIZE);
87 s = (struct cache_sb *) bh->b_data;
89 sb->offset = le64_to_cpu(s->offset);
90 sb->version = le64_to_cpu(s->version);
92 memcpy(sb->magic, s->magic, 16);
93 memcpy(sb->uuid, s->uuid, 16);
94 memcpy(sb->set_uuid, s->set_uuid, 16);
95 memcpy(sb->label, s->label, SB_LABEL_SIZE);
97 sb->flags = le64_to_cpu(s->flags);
98 sb->seq = le64_to_cpu(s->seq);
99 sb->last_mount = le32_to_cpu(s->last_mount);
100 sb->first_bucket = le16_to_cpu(s->first_bucket);
101 sb->keys = le16_to_cpu(s->keys);
103 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
104 sb->d[i] = le64_to_cpu(s->d[i]);
106 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u",
107 sb->version, sb->flags, sb->seq, sb->keys);
109 err = "Not a bcache superblock";
110 if (sb->offset != SB_SECTOR)
113 if (memcmp(sb->magic, bcache_magic, 16))
116 err = "Too many journal buckets";
117 if (sb->keys > SB_JOURNAL_BUCKETS)
120 err = "Bad checksum";
121 if (s->csum != csum_set(s))
125 if (bch_is_zero(sb->uuid, 16))
128 sb->block_size = le16_to_cpu(s->block_size);
130 err = "Superblock block size smaller than device block size";
131 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
134 switch (sb->version) {
135 case BCACHE_SB_VERSION_BDEV:
136 sb->data_offset = BDEV_DATA_START_DEFAULT;
138 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
139 sb->data_offset = le64_to_cpu(s->data_offset);
141 err = "Bad data offset";
142 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
146 case BCACHE_SB_VERSION_CDEV:
147 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
148 sb->nbuckets = le64_to_cpu(s->nbuckets);
149 sb->bucket_size = le16_to_cpu(s->bucket_size);
151 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
152 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
154 err = "Too many buckets";
155 if (sb->nbuckets > LONG_MAX)
158 err = "Not enough buckets";
159 if (sb->nbuckets < 1 << 7)
162 err = "Bad block/bucket size";
163 if (!is_power_of_2(sb->block_size) ||
164 sb->block_size > PAGE_SECTORS ||
165 !is_power_of_2(sb->bucket_size) ||
166 sb->bucket_size < PAGE_SECTORS)
169 err = "Invalid superblock: device too small";
170 if (get_capacity(bdev->bd_disk) < sb->bucket_size * sb->nbuckets)
174 if (bch_is_zero(sb->set_uuid, 16))
177 err = "Bad cache device number in set";
178 if (!sb->nr_in_set ||
179 sb->nr_in_set <= sb->nr_this_dev ||
180 sb->nr_in_set > MAX_CACHES_PER_SET)
183 err = "Journal buckets not sequential";
184 for (i = 0; i < sb->keys; i++)
185 if (sb->d[i] != sb->first_bucket + i)
188 err = "Too many journal buckets";
189 if (sb->first_bucket + sb->keys > sb->nbuckets)
192 err = "Invalid superblock: first bucket comes before end of super";
193 if (sb->first_bucket * sb->bucket_size < 16)
198 err = "Unsupported superblock version";
202 sb->last_mount = get_seconds();
205 get_page(bh->b_page);
212 static void write_bdev_super_endio(struct bio *bio)
214 struct cached_dev *dc = bio->bi_private;
215 /* XXX: error checking */
217 closure_put(&dc->sb_write);
220 static void __write_super(struct cache_sb *sb, struct bio *bio)
222 struct cache_sb *out = page_address(bio_first_page_all(bio));
225 bio->bi_iter.bi_sector = SB_SECTOR;
226 bio->bi_iter.bi_size = SB_SIZE;
227 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC|REQ_META);
228 bch_bio_map(bio, NULL);
230 out->offset = cpu_to_le64(sb->offset);
231 out->version = cpu_to_le64(sb->version);
233 memcpy(out->uuid, sb->uuid, 16);
234 memcpy(out->set_uuid, sb->set_uuid, 16);
235 memcpy(out->label, sb->label, SB_LABEL_SIZE);
237 out->flags = cpu_to_le64(sb->flags);
238 out->seq = cpu_to_le64(sb->seq);
240 out->last_mount = cpu_to_le32(sb->last_mount);
241 out->first_bucket = cpu_to_le16(sb->first_bucket);
242 out->keys = cpu_to_le16(sb->keys);
244 for (i = 0; i < sb->keys; i++)
245 out->d[i] = cpu_to_le64(sb->d[i]);
247 out->csum = csum_set(out);
249 pr_debug("ver %llu, flags %llu, seq %llu",
250 sb->version, sb->flags, sb->seq);
255 static void bch_write_bdev_super_unlock(struct closure *cl)
257 struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
259 up(&dc->sb_write_mutex);
262 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
264 struct closure *cl = &dc->sb_write;
265 struct bio *bio = &dc->sb_bio;
267 down(&dc->sb_write_mutex);
268 closure_init(cl, parent);
271 bio_set_dev(bio, dc->bdev);
272 bio->bi_end_io = write_bdev_super_endio;
273 bio->bi_private = dc;
276 /* I/O request sent to backing device */
277 __write_super(&dc->sb, bio);
279 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
282 static void write_super_endio(struct bio *bio)
284 struct cache *ca = bio->bi_private;
287 bch_count_io_errors(ca, bio->bi_status, 0,
288 "writing superblock");
289 closure_put(&ca->set->sb_write);
292 static void bcache_write_super_unlock(struct closure *cl)
294 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
296 up(&c->sb_write_mutex);
299 void bcache_write_super(struct cache_set *c)
301 struct closure *cl = &c->sb_write;
305 down(&c->sb_write_mutex);
306 closure_init(cl, &c->cl);
310 for_each_cache(ca, c, i) {
311 struct bio *bio = &ca->sb_bio;
313 ca->sb.version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
314 ca->sb.seq = c->sb.seq;
315 ca->sb.last_mount = c->sb.last_mount;
317 SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
320 bio_set_dev(bio, ca->bdev);
321 bio->bi_end_io = write_super_endio;
322 bio->bi_private = ca;
325 __write_super(&ca->sb, bio);
328 closure_return_with_destructor(cl, bcache_write_super_unlock);
333 static void uuid_endio(struct bio *bio)
335 struct closure *cl = bio->bi_private;
336 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
338 cache_set_err_on(bio->bi_status, c, "accessing uuids");
339 bch_bbio_free(bio, c);
343 static void uuid_io_unlock(struct closure *cl)
345 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
347 up(&c->uuid_write_mutex);
350 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
351 struct bkey *k, struct closure *parent)
353 struct closure *cl = &c->uuid_write;
354 struct uuid_entry *u;
359 down(&c->uuid_write_mutex);
360 closure_init(cl, parent);
362 for (i = 0; i < KEY_PTRS(k); i++) {
363 struct bio *bio = bch_bbio_alloc(c);
365 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
366 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
368 bio->bi_end_io = uuid_endio;
369 bio->bi_private = cl;
370 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
371 bch_bio_map(bio, c->uuids);
373 bch_submit_bbio(bio, c, k, i);
375 if (op != REQ_OP_WRITE)
379 bch_extent_to_text(buf, sizeof(buf), k);
380 pr_debug("%s UUIDs at %s", op == REQ_OP_WRITE ? "wrote" : "read", buf);
382 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
383 if (!bch_is_zero(u->uuid, 16))
384 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
385 u - c->uuids, u->uuid, u->label,
386 u->first_reg, u->last_reg, u->invalidated);
388 closure_return_with_destructor(cl, uuid_io_unlock);
391 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
393 struct bkey *k = &j->uuid_bucket;
395 if (__bch_btree_ptr_invalid(c, k))
396 return "bad uuid pointer";
398 bkey_copy(&c->uuid_bucket, k);
399 uuid_io(c, REQ_OP_READ, 0, k, cl);
401 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
402 struct uuid_entry_v0 *u0 = (void *) c->uuids;
403 struct uuid_entry *u1 = (void *) c->uuids;
409 * Since the new uuid entry is bigger than the old, we have to
410 * convert starting at the highest memory address and work down
411 * in order to do it in place
414 for (i = c->nr_uuids - 1;
417 memcpy(u1[i].uuid, u0[i].uuid, 16);
418 memcpy(u1[i].label, u0[i].label, 32);
420 u1[i].first_reg = u0[i].first_reg;
421 u1[i].last_reg = u0[i].last_reg;
422 u1[i].invalidated = u0[i].invalidated;
432 static int __uuid_write(struct cache_set *c)
436 closure_init_stack(&cl);
438 lockdep_assert_held(&bch_register_lock);
440 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
443 SET_KEY_SIZE(&k.key, c->sb.bucket_size);
444 uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
447 bkey_copy(&c->uuid_bucket, &k.key);
452 int bch_uuid_write(struct cache_set *c)
454 int ret = __uuid_write(c);
457 bch_journal_meta(c, NULL);
462 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
464 struct uuid_entry *u;
467 u < c->uuids + c->nr_uuids; u++)
468 if (!memcmp(u->uuid, uuid, 16))
474 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
476 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
477 return uuid_find(c, zero_uuid);
481 * Bucket priorities/gens:
483 * For each bucket, we store on disk its
487 * See alloc.c for an explanation of the gen. The priority is used to implement
488 * lru (and in the future other) cache replacement policies; for most purposes
489 * it's just an opaque integer.
491 * The gens and the priorities don't have a whole lot to do with each other, and
492 * it's actually the gens that must be written out at specific times - it's no
493 * big deal if the priorities don't get written, if we lose them we just reuse
494 * buckets in suboptimal order.
496 * On disk they're stored in a packed array, and in as many buckets are required
497 * to fit them all. The buckets we use to store them form a list; the journal
498 * header points to the first bucket, the first bucket points to the second
501 * This code is used by the allocation code; periodically (whenever it runs out
502 * of buckets to allocate from) the allocation code will invalidate some
503 * buckets, but it can't use those buckets until their new gens are safely on
507 static void prio_endio(struct bio *bio)
509 struct cache *ca = bio->bi_private;
511 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
512 bch_bbio_free(bio, ca->set);
513 closure_put(&ca->prio);
516 static void prio_io(struct cache *ca, uint64_t bucket, int op,
517 unsigned long op_flags)
519 struct closure *cl = &ca->prio;
520 struct bio *bio = bch_bbio_alloc(ca->set);
522 closure_init_stack(cl);
524 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
525 bio_set_dev(bio, ca->bdev);
526 bio->bi_iter.bi_size = bucket_bytes(ca);
528 bio->bi_end_io = prio_endio;
529 bio->bi_private = ca;
530 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
531 bch_bio_map(bio, ca->disk_buckets);
533 closure_bio_submit(ca->set, bio, &ca->prio);
537 void bch_prio_write(struct cache *ca)
543 closure_init_stack(&cl);
545 lockdep_assert_held(&ca->set->bucket_lock);
547 ca->disk_buckets->seq++;
549 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
550 &ca->meta_sectors_written);
552 //pr_debug("free %zu, free_inc %zu, unused %zu", fifo_used(&ca->free),
553 // fifo_used(&ca->free_inc), fifo_used(&ca->unused));
555 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
557 struct prio_set *p = ca->disk_buckets;
558 struct bucket_disk *d = p->data;
559 struct bucket_disk *end = d + prios_per_bucket(ca);
561 for (b = ca->buckets + i * prios_per_bucket(ca);
562 b < ca->buckets + ca->sb.nbuckets && d < end;
564 d->prio = cpu_to_le16(b->prio);
568 p->next_bucket = ca->prio_buckets[i + 1];
569 p->magic = pset_magic(&ca->sb);
570 p->csum = bch_crc64(&p->magic, bucket_bytes(ca) - 8);
572 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, true);
573 BUG_ON(bucket == -1);
575 mutex_unlock(&ca->set->bucket_lock);
576 prio_io(ca, bucket, REQ_OP_WRITE, 0);
577 mutex_lock(&ca->set->bucket_lock);
579 ca->prio_buckets[i] = bucket;
580 atomic_dec_bug(&ca->buckets[bucket].pin);
583 mutex_unlock(&ca->set->bucket_lock);
585 bch_journal_meta(ca->set, &cl);
588 mutex_lock(&ca->set->bucket_lock);
591 * Don't want the old priorities to get garbage collected until after we
592 * finish writing the new ones, and they're journalled
594 for (i = 0; i < prio_buckets(ca); i++) {
595 if (ca->prio_last_buckets[i])
596 __bch_bucket_free(ca,
597 &ca->buckets[ca->prio_last_buckets[i]]);
599 ca->prio_last_buckets[i] = ca->prio_buckets[i];
603 static void prio_read(struct cache *ca, uint64_t bucket)
605 struct prio_set *p = ca->disk_buckets;
606 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
608 unsigned bucket_nr = 0;
610 for (b = ca->buckets;
611 b < ca->buckets + ca->sb.nbuckets;
614 ca->prio_buckets[bucket_nr] = bucket;
615 ca->prio_last_buckets[bucket_nr] = bucket;
618 prio_io(ca, bucket, REQ_OP_READ, 0);
620 if (p->csum != bch_crc64(&p->magic, bucket_bytes(ca) - 8))
621 pr_warn("bad csum reading priorities");
623 if (p->magic != pset_magic(&ca->sb))
624 pr_warn("bad magic reading priorities");
626 bucket = p->next_bucket;
630 b->prio = le16_to_cpu(d->prio);
631 b->gen = b->last_gc = d->gen;
637 static int open_dev(struct block_device *b, fmode_t mode)
639 struct bcache_device *d = b->bd_disk->private_data;
640 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
647 static void release_dev(struct gendisk *b, fmode_t mode)
649 struct bcache_device *d = b->private_data;
653 static int ioctl_dev(struct block_device *b, fmode_t mode,
654 unsigned int cmd, unsigned long arg)
656 struct bcache_device *d = b->bd_disk->private_data;
657 return d->ioctl(d, mode, cmd, arg);
660 static const struct block_device_operations bcache_ops = {
662 .release = release_dev,
664 .owner = THIS_MODULE,
667 void bcache_device_stop(struct bcache_device *d)
669 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
670 closure_queue(&d->cl);
673 static void bcache_device_unlink(struct bcache_device *d)
675 lockdep_assert_held(&bch_register_lock);
677 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
681 sysfs_remove_link(&d->c->kobj, d->name);
682 sysfs_remove_link(&d->kobj, "cache");
684 for_each_cache(ca, d->c, i)
685 bd_unlink_disk_holder(ca->bdev, d->disk);
689 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
695 for_each_cache(ca, d->c, i)
696 bd_link_disk_holder(ca->bdev, d->disk);
698 snprintf(d->name, BCACHEDEVNAME_SIZE,
699 "%s%u", name, d->id);
701 WARN(sysfs_create_link(&d->kobj, &c->kobj, "cache") ||
702 sysfs_create_link(&c->kobj, &d->kobj, d->name),
703 "Couldn't create device <-> cache set symlinks");
705 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
708 static void bcache_device_detach(struct bcache_device *d)
710 lockdep_assert_held(&bch_register_lock);
712 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
713 struct uuid_entry *u = d->c->uuids + d->id;
715 SET_UUID_FLASH_ONLY(u, 0);
716 memcpy(u->uuid, invalid_uuid, 16);
717 u->invalidated = cpu_to_le32(get_seconds());
718 bch_uuid_write(d->c);
721 bcache_device_unlink(d);
723 d->c->devices[d->id] = NULL;
724 closure_put(&d->c->caching);
728 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
735 if (id >= c->devices_max_used)
736 c->devices_max_used = id + 1;
738 closure_get(&c->caching);
741 static inline int first_minor_to_idx(int first_minor)
743 return (first_minor/BCACHE_MINORS);
746 static inline int idx_to_first_minor(int idx)
748 return (idx * BCACHE_MINORS);
751 static void bcache_device_free(struct bcache_device *d)
753 lockdep_assert_held(&bch_register_lock);
755 pr_info("%s stopped", d->disk->disk_name);
758 bcache_device_detach(d);
759 if (d->disk && d->disk->flags & GENHD_FL_UP)
760 del_gendisk(d->disk);
761 if (d->disk && d->disk->queue)
762 blk_cleanup_queue(d->disk->queue);
764 ida_simple_remove(&bcache_device_idx,
765 first_minor_to_idx(d->disk->first_minor));
770 bioset_free(d->bio_split);
771 kvfree(d->full_dirty_stripes);
772 kvfree(d->stripe_sectors_dirty);
774 closure_debug_destroy(&d->cl);
777 static int bcache_device_init(struct bcache_device *d, unsigned block_size,
780 struct request_queue *q;
781 const size_t max_stripes = min_t(size_t, INT_MAX,
782 SIZE_MAX / sizeof(atomic_t));
787 d->stripe_size = 1 << 31;
789 d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
791 if (!d->nr_stripes || d->nr_stripes > max_stripes) {
792 pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
793 (unsigned)d->nr_stripes);
797 n = d->nr_stripes * sizeof(atomic_t);
798 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
799 if (!d->stripe_sectors_dirty)
802 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
803 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
804 if (!d->full_dirty_stripes)
807 idx = ida_simple_get(&bcache_device_idx, 0,
808 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
812 if (!(d->bio_split = bioset_create(4, offsetof(struct bbio, bio),
814 BIOSET_NEED_RESCUER)) ||
815 !(d->disk = alloc_disk(BCACHE_MINORS))) {
816 ida_simple_remove(&bcache_device_idx, idx);
820 set_capacity(d->disk, sectors);
821 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
823 d->disk->major = bcache_major;
824 d->disk->first_minor = idx_to_first_minor(idx);
825 d->disk->fops = &bcache_ops;
826 d->disk->private_data = d;
828 q = blk_alloc_queue(GFP_KERNEL);
832 blk_queue_make_request(q, NULL);
835 q->backing_dev_info->congested_data = d;
836 q->limits.max_hw_sectors = UINT_MAX;
837 q->limits.max_sectors = UINT_MAX;
838 q->limits.max_segment_size = UINT_MAX;
839 q->limits.max_segments = BIO_MAX_PAGES;
840 blk_queue_max_discard_sectors(q, UINT_MAX);
841 q->limits.discard_granularity = 512;
842 q->limits.io_min = block_size;
843 q->limits.logical_block_size = block_size;
844 q->limits.physical_block_size = block_size;
845 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
846 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
847 blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
849 blk_queue_write_cache(q, true, true);
856 static void calc_cached_dev_sectors(struct cache_set *c)
858 uint64_t sectors = 0;
859 struct cached_dev *dc;
861 list_for_each_entry(dc, &c->cached_devs, list)
862 sectors += bdev_sectors(dc->bdev);
864 c->cached_dev_sectors = sectors;
867 void bch_cached_dev_run(struct cached_dev *dc)
869 struct bcache_device *d = &dc->disk;
870 char buf[SB_LABEL_SIZE + 1];
873 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
878 memcpy(buf, dc->sb.label, SB_LABEL_SIZE);
879 buf[SB_LABEL_SIZE] = '\0';
880 env[2] = kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf);
882 if (atomic_xchg(&dc->running, 1)) {
889 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
891 closure_init_stack(&cl);
893 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
894 bch_write_bdev_super(dc, &cl);
899 bd_link_disk_holder(dc->bdev, dc->disk.disk);
900 /* won't show up in the uevent file, use udevadm monitor -e instead
901 * only class / kset properties are persistent */
902 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
906 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
907 sysfs_create_link(&disk_to_dev(d->disk)->kobj, &d->kobj, "bcache"))
908 pr_debug("error creating sysfs link");
912 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
913 * work dc->writeback_rate_update is running. Wait until the routine
914 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
915 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
916 * seconds, give up waiting here and continue to cancel it too.
918 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
920 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
923 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
927 schedule_timeout_interruptible(1);
928 } while (time_out > 0);
931 pr_warn("give up waiting for dc->writeback_write_update to quit");
933 cancel_delayed_work_sync(&dc->writeback_rate_update);
936 static void cached_dev_detach_finish(struct work_struct *w)
938 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
940 closure_init_stack(&cl);
942 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
943 BUG_ON(refcount_read(&dc->count));
945 mutex_lock(&bch_register_lock);
947 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
948 cancel_writeback_rate_update_dwork(dc);
950 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
951 kthread_stop(dc->writeback_thread);
952 dc->writeback_thread = NULL;
955 memset(&dc->sb.set_uuid, 0, 16);
956 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
958 bch_write_bdev_super(dc, &cl);
961 bcache_device_detach(&dc->disk);
962 list_move(&dc->list, &uncached_devices);
964 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
965 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
967 mutex_unlock(&bch_register_lock);
969 pr_info("Caching disabled for %s", dc->backing_dev_name);
971 /* Drop ref we took in cached_dev_detach() */
972 closure_put(&dc->disk.cl);
975 void bch_cached_dev_detach(struct cached_dev *dc)
977 lockdep_assert_held(&bch_register_lock);
979 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
982 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
986 * Block the device from being closed and freed until we're finished
989 closure_get(&dc->disk.cl);
991 bch_writeback_queue(dc);
996 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
999 uint32_t rtime = cpu_to_le32(get_seconds());
1000 struct uuid_entry *u;
1001 struct cached_dev *exist_dc, *t;
1003 if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1004 (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1008 pr_err("Can't attach %s: already attached",
1009 dc->backing_dev_name);
1013 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1014 pr_err("Can't attach %s: shutting down",
1015 dc->backing_dev_name);
1019 if (dc->sb.block_size < c->sb.block_size) {
1021 pr_err("Couldn't attach %s: block size less than set's block size",
1022 dc->backing_dev_name);
1026 /* Check whether already attached */
1027 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1028 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1029 pr_err("Tried to attach %s but duplicate UUID already attached",
1030 dc->backing_dev_name);
1036 u = uuid_find(c, dc->sb.uuid);
1039 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1040 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1041 memcpy(u->uuid, invalid_uuid, 16);
1042 u->invalidated = cpu_to_le32(get_seconds());
1047 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1048 pr_err("Couldn't find uuid for %s in set",
1049 dc->backing_dev_name);
1053 u = uuid_find_empty(c);
1055 pr_err("Not caching %s, no room for UUID",
1056 dc->backing_dev_name);
1061 /* Deadlocks since we're called via sysfs...
1062 sysfs_remove_file(&dc->kobj, &sysfs_attach);
1065 if (bch_is_zero(u->uuid, 16)) {
1067 closure_init_stack(&cl);
1069 memcpy(u->uuid, dc->sb.uuid, 16);
1070 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1071 u->first_reg = u->last_reg = rtime;
1074 memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1075 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1077 bch_write_bdev_super(dc, &cl);
1080 u->last_reg = rtime;
1084 bcache_device_attach(&dc->disk, c, u - c->uuids);
1085 list_move(&dc->list, &c->cached_devs);
1086 calc_cached_dev_sectors(c);
1090 * dc->c must be set before dc->count != 0 - paired with the mb in
1093 refcount_set(&dc->count, 1);
1095 /* Block writeback thread, but spawn it */
1096 down_write(&dc->writeback_lock);
1097 if (bch_cached_dev_writeback_start(dc)) {
1098 up_write(&dc->writeback_lock);
1102 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1103 bch_sectors_dirty_init(&dc->disk);
1104 atomic_set(&dc->has_dirty, 1);
1105 bch_writeback_queue(dc);
1108 bch_cached_dev_run(dc);
1109 bcache_device_link(&dc->disk, c, "bdev");
1111 /* Allow the writeback thread to proceed */
1112 up_write(&dc->writeback_lock);
1114 pr_info("Caching %s as %s on set %pU",
1115 dc->backing_dev_name,
1116 dc->disk.disk->disk_name,
1117 dc->disk.c->sb.set_uuid);
1121 void bch_cached_dev_release(struct kobject *kobj)
1123 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1126 module_put(THIS_MODULE);
1129 static void cached_dev_free(struct closure *cl)
1131 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1133 mutex_lock(&bch_register_lock);
1135 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1136 cancel_writeback_rate_update_dwork(dc);
1138 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1139 kthread_stop(dc->writeback_thread);
1140 if (dc->writeback_write_wq)
1141 destroy_workqueue(dc->writeback_write_wq);
1143 if (atomic_read(&dc->running))
1144 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1145 bcache_device_free(&dc->disk);
1146 list_del(&dc->list);
1148 mutex_unlock(&bch_register_lock);
1150 if (!IS_ERR_OR_NULL(dc->bdev))
1151 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1153 wake_up(&unregister_wait);
1155 kobject_put(&dc->disk.kobj);
1158 static void cached_dev_flush(struct closure *cl)
1160 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1161 struct bcache_device *d = &dc->disk;
1163 mutex_lock(&bch_register_lock);
1164 bcache_device_unlink(d);
1165 mutex_unlock(&bch_register_lock);
1167 bch_cache_accounting_destroy(&dc->accounting);
1168 kobject_del(&d->kobj);
1170 continue_at(cl, cached_dev_free, system_wq);
1173 static int cached_dev_init(struct cached_dev *dc, unsigned block_size)
1177 struct request_queue *q = bdev_get_queue(dc->bdev);
1179 __module_get(THIS_MODULE);
1180 INIT_LIST_HEAD(&dc->list);
1181 closure_init(&dc->disk.cl, NULL);
1182 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1183 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1184 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1185 sema_init(&dc->sb_write_mutex, 1);
1186 INIT_LIST_HEAD(&dc->io_lru);
1187 spin_lock_init(&dc->io_lock);
1188 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1190 dc->sequential_cutoff = 4 << 20;
1192 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1193 list_add(&io->lru, &dc->io_lru);
1194 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1197 dc->disk.stripe_size = q->limits.io_opt >> 9;
1199 if (dc->disk.stripe_size)
1200 dc->partial_stripes_expensive =
1201 q->limits.raid_partial_stripes_expensive;
1203 ret = bcache_device_init(&dc->disk, block_size,
1204 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1208 dc->disk.disk->queue->backing_dev_info->ra_pages =
1209 max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1210 q->backing_dev_info->ra_pages);
1212 atomic_set(&dc->io_errors, 0);
1213 dc->io_disable = false;
1214 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1215 /* default to auto */
1216 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1218 bch_cached_dev_request_init(dc);
1219 bch_cached_dev_writeback_init(dc);
1223 /* Cached device - bcache superblock */
1225 static void register_bdev(struct cache_sb *sb, struct page *sb_page,
1226 struct block_device *bdev,
1227 struct cached_dev *dc)
1229 const char *err = "cannot allocate memory";
1230 struct cache_set *c;
1232 bdevname(bdev, dc->backing_dev_name);
1233 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1235 dc->bdev->bd_holder = dc;
1237 bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1238 bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1242 if (cached_dev_init(dc, sb->block_size << 9))
1245 err = "error creating kobject";
1246 if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1249 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1252 pr_info("registered backing device %s", dc->backing_dev_name);
1254 list_add(&dc->list, &uncached_devices);
1255 list_for_each_entry(c, &bch_cache_sets, list)
1256 bch_cached_dev_attach(dc, c, NULL);
1258 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1259 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE)
1260 bch_cached_dev_run(dc);
1264 pr_notice("error %s: %s", dc->backing_dev_name, err);
1265 bcache_device_stop(&dc->disk);
1268 /* Flash only volumes */
1270 void bch_flash_dev_release(struct kobject *kobj)
1272 struct bcache_device *d = container_of(kobj, struct bcache_device,
1277 static void flash_dev_free(struct closure *cl)
1279 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1280 mutex_lock(&bch_register_lock);
1281 bcache_device_free(d);
1282 mutex_unlock(&bch_register_lock);
1283 kobject_put(&d->kobj);
1286 static void flash_dev_flush(struct closure *cl)
1288 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1290 mutex_lock(&bch_register_lock);
1291 bcache_device_unlink(d);
1292 mutex_unlock(&bch_register_lock);
1293 kobject_del(&d->kobj);
1294 continue_at(cl, flash_dev_free, system_wq);
1297 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1299 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1304 closure_init(&d->cl, NULL);
1305 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1307 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1309 if (bcache_device_init(d, block_bytes(c), u->sectors))
1312 bcache_device_attach(d, c, u - c->uuids);
1313 bch_sectors_dirty_init(d);
1314 bch_flash_dev_request_init(d);
1317 if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1320 bcache_device_link(d, c, "volume");
1324 kobject_put(&d->kobj);
1328 static int flash_devs_run(struct cache_set *c)
1331 struct uuid_entry *u;
1334 u < c->uuids + c->nr_uuids && !ret;
1336 if (UUID_FLASH_ONLY(u))
1337 ret = flash_dev_run(c, u);
1342 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1344 struct uuid_entry *u;
1346 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1349 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1352 u = uuid_find_empty(c);
1354 pr_err("Can't create volume, no room for UUID");
1358 get_random_bytes(u->uuid, 16);
1359 memset(u->label, 0, 32);
1360 u->first_reg = u->last_reg = cpu_to_le32(get_seconds());
1362 SET_UUID_FLASH_ONLY(u, 1);
1363 u->sectors = size >> 9;
1367 return flash_dev_run(c, u);
1370 bool bch_cached_dev_error(struct cached_dev *dc)
1372 struct cache_set *c;
1374 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1377 dc->io_disable = true;
1378 /* make others know io_disable is true earlier */
1381 pr_err("stop %s: too many IO errors on backing device %s\n",
1382 dc->disk.disk->disk_name, dc->backing_dev_name);
1385 * If the cached device is still attached to a cache set,
1386 * even dc->io_disable is true and no more I/O requests
1387 * accepted, cache device internal I/O (writeback scan or
1388 * garbage collection) may still prevent bcache device from
1389 * being stopped. So here CACHE_SET_IO_DISABLE should be
1390 * set to c->flags too, to make the internal I/O to cache
1391 * device rejected and stopped immediately.
1392 * If c is NULL, that means the bcache device is not attached
1393 * to any cache set, then no CACHE_SET_IO_DISABLE bit to set.
1396 if (c && test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1397 pr_info("CACHE_SET_IO_DISABLE already set");
1399 bcache_device_stop(&dc->disk);
1406 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1410 if (c->on_error != ON_ERROR_PANIC &&
1411 test_bit(CACHE_SET_STOPPING, &c->flags))
1414 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1415 pr_info("CACHE_SET_IO_DISABLE already set");
1417 /* XXX: we can be called from atomic context
1418 acquire_console_sem();
1421 printk(KERN_ERR "bcache: error on %pU: ", c->sb.set_uuid);
1423 va_start(args, fmt);
1427 printk(", disabling caching\n");
1429 if (c->on_error == ON_ERROR_PANIC)
1430 panic("panic forced after error\n");
1432 bch_cache_set_unregister(c);
1436 void bch_cache_set_release(struct kobject *kobj)
1438 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1440 module_put(THIS_MODULE);
1443 static void cache_set_free(struct closure *cl)
1445 struct cache_set *c = container_of(cl, struct cache_set, cl);
1449 if (!IS_ERR_OR_NULL(c->debug))
1450 debugfs_remove(c->debug);
1452 bch_open_buckets_free(c);
1453 bch_btree_cache_free(c);
1454 bch_journal_free(c);
1456 for_each_cache(ca, c, i)
1459 c->cache[ca->sb.nr_this_dev] = NULL;
1460 kobject_put(&ca->kobj);
1463 bch_bset_sort_state_free(&c->sort);
1464 free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1466 if (c->moving_gc_wq)
1467 destroy_workqueue(c->moving_gc_wq);
1469 bioset_free(c->bio_split);
1471 mempool_destroy(c->fill_iter);
1473 mempool_destroy(c->bio_meta);
1475 mempool_destroy(c->search);
1478 mutex_lock(&bch_register_lock);
1480 mutex_unlock(&bch_register_lock);
1482 pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1483 wake_up(&unregister_wait);
1485 closure_debug_destroy(&c->cl);
1486 kobject_put(&c->kobj);
1489 static void cache_set_flush(struct closure *cl)
1491 struct cache_set *c = container_of(cl, struct cache_set, caching);
1496 bch_cache_accounting_destroy(&c->accounting);
1498 kobject_put(&c->internal);
1499 kobject_del(&c->kobj);
1502 kthread_stop(c->gc_thread);
1504 if (!IS_ERR_OR_NULL(c->root))
1505 list_add(&c->root->list, &c->btree_cache);
1507 /* Should skip this if we're unregistering because of an error */
1508 list_for_each_entry(b, &c->btree_cache, list) {
1509 mutex_lock(&b->write_lock);
1510 if (btree_node_dirty(b))
1511 __bch_btree_node_write(b, NULL);
1512 mutex_unlock(&b->write_lock);
1515 for_each_cache(ca, c, i)
1516 if (ca->alloc_thread)
1517 kthread_stop(ca->alloc_thread);
1519 if (c->journal.cur) {
1520 cancel_delayed_work_sync(&c->journal.work);
1521 /* flush last journal entry if needed */
1522 c->journal.work.work.func(&c->journal.work.work);
1529 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1530 * cache set is unregistering due to too many I/O errors. In this condition,
1531 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1532 * value and whether the broken cache has dirty data:
1534 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1535 * BCH_CACHED_STOP_AUTO 0 NO
1536 * BCH_CACHED_STOP_AUTO 1 YES
1537 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1538 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1540 * The expected behavior is, if stop_when_cache_set_failed is configured to
1541 * "auto" via sysfs interface, the bcache device will not be stopped if the
1542 * backing device is clean on the broken cache device.
1544 static void conditional_stop_bcache_device(struct cache_set *c,
1545 struct bcache_device *d,
1546 struct cached_dev *dc)
1548 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1549 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1550 d->disk->disk_name, c->sb.set_uuid);
1551 bcache_device_stop(d);
1552 } else if (atomic_read(&dc->has_dirty)) {
1554 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1555 * and dc->has_dirty == 1
1557 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1558 d->disk->disk_name);
1560 * There might be a small time gap that cache set is
1561 * released but bcache device is not. Inside this time
1562 * gap, regular I/O requests will directly go into
1563 * backing device as no cache set attached to. This
1564 * behavior may also introduce potential inconsistence
1565 * data in writeback mode while cache is dirty.
1566 * Therefore before calling bcache_device_stop() due
1567 * to a broken cache device, dc->io_disable should be
1568 * explicitly set to true.
1570 dc->io_disable = true;
1571 /* make others know io_disable is true earlier */
1573 bcache_device_stop(d);
1576 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1577 * and dc->has_dirty == 0
1579 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1580 d->disk->disk_name);
1584 static void __cache_set_unregister(struct closure *cl)
1586 struct cache_set *c = container_of(cl, struct cache_set, caching);
1587 struct cached_dev *dc;
1588 struct bcache_device *d;
1591 mutex_lock(&bch_register_lock);
1593 for (i = 0; i < c->devices_max_used; i++) {
1598 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1599 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1600 dc = container_of(d, struct cached_dev, disk);
1601 bch_cached_dev_detach(dc);
1602 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1603 conditional_stop_bcache_device(c, d, dc);
1605 bcache_device_stop(d);
1609 mutex_unlock(&bch_register_lock);
1611 continue_at(cl, cache_set_flush, system_wq);
1614 void bch_cache_set_stop(struct cache_set *c)
1616 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1617 closure_queue(&c->caching);
1620 void bch_cache_set_unregister(struct cache_set *c)
1622 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1623 bch_cache_set_stop(c);
1626 #define alloc_bucket_pages(gfp, c) \
1627 ((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1629 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1632 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1636 __module_get(THIS_MODULE);
1637 closure_init(&c->cl, NULL);
1638 set_closure_fn(&c->cl, cache_set_free, system_wq);
1640 closure_init(&c->caching, &c->cl);
1641 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1643 /* Maybe create continue_at_noreturn() and use it here? */
1644 closure_set_stopped(&c->cl);
1645 closure_put(&c->cl);
1647 kobject_init(&c->kobj, &bch_cache_set_ktype);
1648 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1650 bch_cache_accounting_init(&c->accounting, &c->cl);
1652 memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1653 c->sb.block_size = sb->block_size;
1654 c->sb.bucket_size = sb->bucket_size;
1655 c->sb.nr_in_set = sb->nr_in_set;
1656 c->sb.last_mount = sb->last_mount;
1657 c->bucket_bits = ilog2(sb->bucket_size);
1658 c->block_bits = ilog2(sb->block_size);
1659 c->nr_uuids = bucket_bytes(c) / sizeof(struct uuid_entry);
1660 c->devices_max_used = 0;
1661 c->btree_pages = bucket_pages(c);
1662 if (c->btree_pages > BTREE_MAX_PAGES)
1663 c->btree_pages = max_t(int, c->btree_pages / 4,
1666 sema_init(&c->sb_write_mutex, 1);
1667 mutex_init(&c->bucket_lock);
1668 init_waitqueue_head(&c->btree_cache_wait);
1669 init_waitqueue_head(&c->bucket_wait);
1670 init_waitqueue_head(&c->gc_wait);
1671 sema_init(&c->uuid_write_mutex, 1);
1673 spin_lock_init(&c->btree_gc_time.lock);
1674 spin_lock_init(&c->btree_split_time.lock);
1675 spin_lock_init(&c->btree_read_time.lock);
1677 bch_moving_init_cache_set(c);
1679 INIT_LIST_HEAD(&c->list);
1680 INIT_LIST_HEAD(&c->cached_devs);
1681 INIT_LIST_HEAD(&c->btree_cache);
1682 INIT_LIST_HEAD(&c->btree_cache_freeable);
1683 INIT_LIST_HEAD(&c->btree_cache_freed);
1684 INIT_LIST_HEAD(&c->data_buckets);
1686 c->search = mempool_create_slab_pool(32, bch_search_cache);
1690 iter_size = (sb->bucket_size / sb->block_size + 1) *
1691 sizeof(struct btree_iter_set);
1693 if (!(c->devices = kzalloc(c->nr_uuids * sizeof(void *), GFP_KERNEL)) ||
1694 !(c->bio_meta = mempool_create_kmalloc_pool(2,
1695 sizeof(struct bbio) + sizeof(struct bio_vec) *
1696 bucket_pages(c))) ||
1697 !(c->fill_iter = mempool_create_kmalloc_pool(1, iter_size)) ||
1698 !(c->bio_split = bioset_create(4, offsetof(struct bbio, bio),
1700 BIOSET_NEED_RESCUER)) ||
1701 !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1702 !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1703 WQ_MEM_RECLAIM, 0)) ||
1704 bch_journal_alloc(c) ||
1705 bch_btree_cache_alloc(c) ||
1706 bch_open_buckets_alloc(c) ||
1707 bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1710 c->congested_read_threshold_us = 2000;
1711 c->congested_write_threshold_us = 20000;
1712 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1713 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1717 bch_cache_set_unregister(c);
1721 static void run_cache_set(struct cache_set *c)
1723 const char *err = "cannot allocate memory";
1724 struct cached_dev *dc, *t;
1729 closure_init_stack(&cl);
1731 for_each_cache(ca, c, i)
1732 c->nbuckets += ca->sb.nbuckets;
1735 if (CACHE_SYNC(&c->sb)) {
1740 err = "cannot allocate memory for journal";
1741 if (bch_journal_read(c, &journal))
1744 pr_debug("btree_journal_read() done");
1746 err = "no journal entries found";
1747 if (list_empty(&journal))
1750 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1752 err = "IO error reading priorities";
1753 for_each_cache(ca, c, i)
1754 prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1757 * If prio_read() fails it'll call cache_set_error and we'll
1758 * tear everything down right away, but if we perhaps checked
1759 * sooner we could avoid journal replay.
1764 err = "bad btree root";
1765 if (__bch_btree_ptr_invalid(c, k))
1768 err = "error reading btree root";
1769 c->root = bch_btree_node_get(c, NULL, k, j->btree_level, true, NULL);
1770 if (IS_ERR_OR_NULL(c->root))
1773 list_del_init(&c->root->list);
1774 rw_unlock(true, c->root);
1776 err = uuid_read(c, j, &cl);
1780 err = "error in recovery";
1781 if (bch_btree_check(c))
1784 bch_journal_mark(c, &journal);
1785 bch_initial_gc_finish(c);
1786 pr_debug("btree_check() done");
1789 * bcache_journal_next() can't happen sooner, or
1790 * btree_gc_finish() will give spurious errors about last_gc >
1791 * gc_gen - this is a hack but oh well.
1793 bch_journal_next(&c->journal);
1795 err = "error starting allocator thread";
1796 for_each_cache(ca, c, i)
1797 if (bch_cache_allocator_start(ca))
1801 * First place it's safe to allocate: btree_check() and
1802 * btree_gc_finish() have to run before we have buckets to
1803 * allocate, and bch_bucket_alloc_set() might cause a journal
1804 * entry to be written so bcache_journal_next() has to be called
1807 * If the uuids were in the old format we have to rewrite them
1808 * before the next journal entry is written:
1810 if (j->version < BCACHE_JSET_VERSION_UUID)
1813 bch_journal_replay(c, &journal);
1815 pr_notice("invalidating existing data");
1817 for_each_cache(ca, c, i) {
1820 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1821 2, SB_JOURNAL_BUCKETS);
1823 for (j = 0; j < ca->sb.keys; j++)
1824 ca->sb.d[j] = ca->sb.first_bucket + j;
1827 bch_initial_gc_finish(c);
1829 err = "error starting allocator thread";
1830 for_each_cache(ca, c, i)
1831 if (bch_cache_allocator_start(ca))
1834 mutex_lock(&c->bucket_lock);
1835 for_each_cache(ca, c, i)
1837 mutex_unlock(&c->bucket_lock);
1839 err = "cannot allocate new UUID bucket";
1840 if (__uuid_write(c))
1843 err = "cannot allocate new btree root";
1844 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1845 if (IS_ERR_OR_NULL(c->root))
1848 mutex_lock(&c->root->write_lock);
1849 bkey_copy_key(&c->root->key, &MAX_KEY);
1850 bch_btree_node_write(c->root, &cl);
1851 mutex_unlock(&c->root->write_lock);
1853 bch_btree_set_root(c->root);
1854 rw_unlock(true, c->root);
1857 * We don't want to write the first journal entry until
1858 * everything is set up - fortunately journal entries won't be
1859 * written until the SET_CACHE_SYNC() here:
1861 SET_CACHE_SYNC(&c->sb, true);
1863 bch_journal_next(&c->journal);
1864 bch_journal_meta(c, &cl);
1867 err = "error starting gc thread";
1868 if (bch_gc_thread_start(c))
1872 c->sb.last_mount = get_seconds();
1873 bcache_write_super(c);
1875 list_for_each_entry_safe(dc, t, &uncached_devices, list)
1876 bch_cached_dev_attach(dc, c, NULL);
1880 set_bit(CACHE_SET_RUNNING, &c->flags);
1884 /* XXX: test this, it's broken */
1885 bch_cache_set_error(c, "%s", err);
1888 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
1890 return ca->sb.block_size == c->sb.block_size &&
1891 ca->sb.bucket_size == c->sb.bucket_size &&
1892 ca->sb.nr_in_set == c->sb.nr_in_set;
1895 static const char *register_cache_set(struct cache *ca)
1898 const char *err = "cannot allocate memory";
1899 struct cache_set *c;
1901 list_for_each_entry(c, &bch_cache_sets, list)
1902 if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
1903 if (c->cache[ca->sb.nr_this_dev])
1904 return "duplicate cache set member";
1906 if (!can_attach_cache(ca, c))
1907 return "cache sb does not match set";
1909 if (!CACHE_SYNC(&ca->sb))
1910 SET_CACHE_SYNC(&c->sb, false);
1915 c = bch_cache_set_alloc(&ca->sb);
1919 err = "error creating kobject";
1920 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
1921 kobject_add(&c->internal, &c->kobj, "internal"))
1924 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
1927 bch_debug_init_cache_set(c);
1929 list_add(&c->list, &bch_cache_sets);
1931 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
1932 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
1933 sysfs_create_link(&c->kobj, &ca->kobj, buf))
1936 if (ca->sb.seq > c->sb.seq) {
1937 c->sb.version = ca->sb.version;
1938 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
1939 c->sb.flags = ca->sb.flags;
1940 c->sb.seq = ca->sb.seq;
1941 pr_debug("set version = %llu", c->sb.version);
1944 kobject_get(&ca->kobj);
1946 ca->set->cache[ca->sb.nr_this_dev] = ca;
1947 c->cache_by_alloc[c->caches_loaded++] = ca;
1949 if (c->caches_loaded == c->sb.nr_in_set)
1954 bch_cache_set_unregister(c);
1960 void bch_cache_release(struct kobject *kobj)
1962 struct cache *ca = container_of(kobj, struct cache, kobj);
1966 BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
1967 ca->set->cache[ca->sb.nr_this_dev] = NULL;
1970 free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
1971 kfree(ca->prio_buckets);
1974 free_heap(&ca->heap);
1975 free_fifo(&ca->free_inc);
1977 for (i = 0; i < RESERVE_NR; i++)
1978 free_fifo(&ca->free[i]);
1980 if (ca->sb_bio.bi_inline_vecs[0].bv_page)
1981 put_page(bio_first_page_all(&ca->sb_bio));
1983 if (!IS_ERR_OR_NULL(ca->bdev))
1984 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1987 module_put(THIS_MODULE);
1990 static int cache_alloc(struct cache *ca)
1993 size_t btree_buckets;
1996 __module_get(THIS_MODULE);
1997 kobject_init(&ca->kobj, &bch_cache_ktype);
1999 bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2002 * when ca->sb.njournal_buckets is not zero, journal exists,
2003 * and in bch_journal_replay(), tree node may split,
2004 * so bucket of RESERVE_BTREE type is needed,
2005 * the worst situation is all journal buckets are valid journal,
2006 * and all the keys need to replay,
2007 * so the number of RESERVE_BTREE type buckets should be as much
2008 * as journal buckets
2010 btree_buckets = ca->sb.njournal_buckets ?: 8;
2011 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2013 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets, GFP_KERNEL) ||
2014 !init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca), GFP_KERNEL) ||
2015 !init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL) ||
2016 !init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL) ||
2017 !init_fifo(&ca->free_inc, free << 2, GFP_KERNEL) ||
2018 !init_heap(&ca->heap, free << 3, GFP_KERNEL) ||
2019 !(ca->buckets = vzalloc(sizeof(struct bucket) *
2020 ca->sb.nbuckets)) ||
2021 !(ca->prio_buckets = kzalloc(sizeof(uint64_t) * prio_buckets(ca) *
2023 !(ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca)))
2026 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2028 for_each_bucket(b, ca)
2029 atomic_set(&b->pin, 0);
2034 static int register_cache(struct cache_sb *sb, struct page *sb_page,
2035 struct block_device *bdev, struct cache *ca)
2037 const char *err = NULL; /* must be set for any error case */
2040 bdevname(bdev, ca->cache_dev_name);
2041 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2043 ca->bdev->bd_holder = ca;
2045 bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2046 bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2049 if (blk_queue_discard(bdev_get_queue(bdev)))
2050 ca->discard = CACHE_DISCARD(&ca->sb);
2052 ret = cache_alloc(ca);
2054 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2056 err = "cache_alloc(): -ENOMEM";
2058 err = "cache_alloc(): unknown error";
2062 if (kobject_add(&ca->kobj, &part_to_dev(bdev->bd_part)->kobj, "bcache")) {
2063 err = "error calling kobject_add";
2068 mutex_lock(&bch_register_lock);
2069 err = register_cache_set(ca);
2070 mutex_unlock(&bch_register_lock);
2077 pr_info("registered cache device %s", ca->cache_dev_name);
2080 kobject_put(&ca->kobj);
2084 pr_notice("error %s: %s", ca->cache_dev_name, err);
2089 /* Global interfaces/init */
2091 static ssize_t register_bcache(struct kobject *, struct kobj_attribute *,
2092 const char *, size_t);
2094 kobj_attribute_write(register, register_bcache);
2095 kobj_attribute_write(register_quiet, register_bcache);
2097 static bool bch_is_open_backing(struct block_device *bdev) {
2098 struct cache_set *c, *tc;
2099 struct cached_dev *dc, *t;
2101 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2102 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2103 if (dc->bdev == bdev)
2105 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2106 if (dc->bdev == bdev)
2111 static bool bch_is_open_cache(struct block_device *bdev) {
2112 struct cache_set *c, *tc;
2116 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2117 for_each_cache(ca, c, i)
2118 if (ca->bdev == bdev)
2123 static bool bch_is_open(struct block_device *bdev) {
2124 return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2127 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2128 const char *buffer, size_t size)
2131 const char *err = "cannot allocate memory";
2133 struct cache_sb *sb = NULL;
2134 struct block_device *bdev = NULL;
2135 struct page *sb_page = NULL;
2137 if (!try_module_get(THIS_MODULE))
2140 if (!(path = kstrndup(buffer, size, GFP_KERNEL)) ||
2141 !(sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL)))
2144 err = "failed to open device";
2145 bdev = blkdev_get_by_path(strim(path),
2146 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2149 if (bdev == ERR_PTR(-EBUSY)) {
2150 bdev = lookup_bdev(strim(path));
2151 mutex_lock(&bch_register_lock);
2152 if (!IS_ERR(bdev) && bch_is_open(bdev))
2153 err = "device already registered";
2155 err = "device busy";
2156 mutex_unlock(&bch_register_lock);
2159 if (attr == &ksysfs_register_quiet)
2165 err = "failed to set blocksize";
2166 if (set_blocksize(bdev, 4096))
2169 err = read_super(sb, bdev, &sb_page);
2173 err = "failed to register device";
2174 if (SB_IS_BDEV(sb)) {
2175 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2179 mutex_lock(&bch_register_lock);
2180 register_bdev(sb, sb_page, bdev, dc);
2181 mutex_unlock(&bch_register_lock);
2183 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2187 if (register_cache(sb, sb_page, bdev, ca) != 0)
2195 module_put(THIS_MODULE);
2199 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2201 pr_info("error %s: %s", path, err);
2206 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2208 if (code == SYS_DOWN ||
2210 code == SYS_POWER_OFF) {
2212 unsigned long start = jiffies;
2213 bool stopped = false;
2215 struct cache_set *c, *tc;
2216 struct cached_dev *dc, *tdc;
2218 mutex_lock(&bch_register_lock);
2220 if (list_empty(&bch_cache_sets) &&
2221 list_empty(&uncached_devices))
2224 pr_info("Stopping all devices:");
2226 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2227 bch_cache_set_stop(c);
2229 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2230 bcache_device_stop(&dc->disk);
2232 /* What's a condition variable? */
2234 long timeout = start + 2 * HZ - jiffies;
2236 stopped = list_empty(&bch_cache_sets) &&
2237 list_empty(&uncached_devices);
2239 if (timeout < 0 || stopped)
2242 prepare_to_wait(&unregister_wait, &wait,
2243 TASK_UNINTERRUPTIBLE);
2245 mutex_unlock(&bch_register_lock);
2246 schedule_timeout(timeout);
2247 mutex_lock(&bch_register_lock);
2250 finish_wait(&unregister_wait, &wait);
2253 pr_info("All devices stopped");
2255 pr_notice("Timeout waiting for devices to be closed");
2257 mutex_unlock(&bch_register_lock);
2263 static struct notifier_block reboot = {
2264 .notifier_call = bcache_reboot,
2265 .priority = INT_MAX, /* before any real devices */
2268 static void bcache_exit(void)
2273 kobject_put(bcache_kobj);
2275 destroy_workqueue(bcache_wq);
2277 unregister_blkdev(bcache_major, "bcache");
2278 unregister_reboot_notifier(&reboot);
2279 mutex_destroy(&bch_register_lock);
2282 static int __init bcache_init(void)
2284 static const struct attribute *files[] = {
2285 &ksysfs_register.attr,
2286 &ksysfs_register_quiet.attr,
2290 mutex_init(&bch_register_lock);
2291 init_waitqueue_head(&unregister_wait);
2292 register_reboot_notifier(&reboot);
2294 bcache_major = register_blkdev(0, "bcache");
2295 if (bcache_major < 0) {
2296 unregister_reboot_notifier(&reboot);
2297 mutex_destroy(&bch_register_lock);
2298 return bcache_major;
2301 if (!(bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0)) ||
2302 !(bcache_kobj = kobject_create_and_add("bcache", fs_kobj)) ||
2303 bch_request_init() ||
2304 bch_debug_init(bcache_kobj) || closure_debug_init() ||
2305 sysfs_create_files(bcache_kobj, files))
2314 module_exit(bcache_exit);
2315 module_init(bcache_init);