1 // SPDX-License-Identifier: GPL-2.0
3 * bcache setup/teardown code, and some metadata io - read a superblock and
4 * figure out what to do with it.
6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7 * Copyright 2012 Google, Inc.
15 #include "writeback.h"
17 #include <linux/blkdev.h>
18 #include <linux/buffer_head.h>
19 #include <linux/debugfs.h>
20 #include <linux/genhd.h>
21 #include <linux/idr.h>
22 #include <linux/kthread.h>
23 #include <linux/module.h>
24 #include <linux/random.h>
25 #include <linux/reboot.h>
26 #include <linux/sysfs.h>
28 MODULE_LICENSE("GPL");
29 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
31 static const char bcache_magic[] = {
32 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
33 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
36 static const char invalid_uuid[] = {
37 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
38 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
41 static struct kobject *bcache_kobj;
42 struct mutex bch_register_lock;
43 LIST_HEAD(bch_cache_sets);
44 static LIST_HEAD(uncached_devices);
46 static int bcache_major;
47 static DEFINE_IDA(bcache_device_idx);
48 static wait_queue_head_t unregister_wait;
49 struct workqueue_struct *bcache_wq;
50 struct workqueue_struct *bch_journal_wq;
52 #define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
53 /* limitation of partitions number on single bcache device */
54 #define BCACHE_MINORS 128
55 /* limitation of bcache devices number on single system */
56 #define BCACHE_DEVICE_IDX_MAX ((1U << MINORBITS)/BCACHE_MINORS)
60 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
65 struct buffer_head *bh = __bread(bdev, 1, SB_SIZE);
71 s = (struct cache_sb *) bh->b_data;
73 sb->offset = le64_to_cpu(s->offset);
74 sb->version = le64_to_cpu(s->version);
76 memcpy(sb->magic, s->magic, 16);
77 memcpy(sb->uuid, s->uuid, 16);
78 memcpy(sb->set_uuid, s->set_uuid, 16);
79 memcpy(sb->label, s->label, SB_LABEL_SIZE);
81 sb->flags = le64_to_cpu(s->flags);
82 sb->seq = le64_to_cpu(s->seq);
83 sb->last_mount = le32_to_cpu(s->last_mount);
84 sb->first_bucket = le16_to_cpu(s->first_bucket);
85 sb->keys = le16_to_cpu(s->keys);
87 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
88 sb->d[i] = le64_to_cpu(s->d[i]);
90 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u",
91 sb->version, sb->flags, sb->seq, sb->keys);
93 err = "Not a bcache superblock";
94 if (sb->offset != SB_SECTOR)
97 if (memcmp(sb->magic, bcache_magic, 16))
100 err = "Too many journal buckets";
101 if (sb->keys > SB_JOURNAL_BUCKETS)
104 err = "Bad checksum";
105 if (s->csum != csum_set(s))
109 if (bch_is_zero(sb->uuid, 16))
112 sb->block_size = le16_to_cpu(s->block_size);
114 err = "Superblock block size smaller than device block size";
115 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
118 switch (sb->version) {
119 case BCACHE_SB_VERSION_BDEV:
120 sb->data_offset = BDEV_DATA_START_DEFAULT;
122 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
123 sb->data_offset = le64_to_cpu(s->data_offset);
125 err = "Bad data offset";
126 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
130 case BCACHE_SB_VERSION_CDEV:
131 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
132 sb->nbuckets = le64_to_cpu(s->nbuckets);
133 sb->bucket_size = le16_to_cpu(s->bucket_size);
135 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
136 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
138 err = "Too many buckets";
139 if (sb->nbuckets > LONG_MAX)
142 err = "Not enough buckets";
143 if (sb->nbuckets < 1 << 7)
146 err = "Bad block/bucket size";
147 if (!is_power_of_2(sb->block_size) ||
148 sb->block_size > PAGE_SECTORS ||
149 !is_power_of_2(sb->bucket_size) ||
150 sb->bucket_size < PAGE_SECTORS)
153 err = "Invalid superblock: device too small";
154 if (get_capacity(bdev->bd_disk) <
155 sb->bucket_size * sb->nbuckets)
159 if (bch_is_zero(sb->set_uuid, 16))
162 err = "Bad cache device number in set";
163 if (!sb->nr_in_set ||
164 sb->nr_in_set <= sb->nr_this_dev ||
165 sb->nr_in_set > MAX_CACHES_PER_SET)
168 err = "Journal buckets not sequential";
169 for (i = 0; i < sb->keys; i++)
170 if (sb->d[i] != sb->first_bucket + i)
173 err = "Too many journal buckets";
174 if (sb->first_bucket + sb->keys > sb->nbuckets)
177 err = "Invalid superblock: first bucket comes before end of super";
178 if (sb->first_bucket * sb->bucket_size < 16)
183 err = "Unsupported superblock version";
187 sb->last_mount = (u32)ktime_get_real_seconds();
190 get_page(bh->b_page);
197 static void write_bdev_super_endio(struct bio *bio)
199 struct cached_dev *dc = bio->bi_private;
200 /* XXX: error checking */
202 closure_put(&dc->sb_write);
205 static void __write_super(struct cache_sb *sb, struct bio *bio)
207 struct cache_sb *out = page_address(bio_first_page_all(bio));
210 bio->bi_iter.bi_sector = SB_SECTOR;
211 bio->bi_iter.bi_size = SB_SIZE;
212 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC|REQ_META);
213 bch_bio_map(bio, NULL);
215 out->offset = cpu_to_le64(sb->offset);
216 out->version = cpu_to_le64(sb->version);
218 memcpy(out->uuid, sb->uuid, 16);
219 memcpy(out->set_uuid, sb->set_uuid, 16);
220 memcpy(out->label, sb->label, SB_LABEL_SIZE);
222 out->flags = cpu_to_le64(sb->flags);
223 out->seq = cpu_to_le64(sb->seq);
225 out->last_mount = cpu_to_le32(sb->last_mount);
226 out->first_bucket = cpu_to_le16(sb->first_bucket);
227 out->keys = cpu_to_le16(sb->keys);
229 for (i = 0; i < sb->keys; i++)
230 out->d[i] = cpu_to_le64(sb->d[i]);
232 out->csum = csum_set(out);
234 pr_debug("ver %llu, flags %llu, seq %llu",
235 sb->version, sb->flags, sb->seq);
240 static void bch_write_bdev_super_unlock(struct closure *cl)
242 struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
244 up(&dc->sb_write_mutex);
247 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
249 struct closure *cl = &dc->sb_write;
250 struct bio *bio = &dc->sb_bio;
252 down(&dc->sb_write_mutex);
253 closure_init(cl, parent);
256 bio_set_dev(bio, dc->bdev);
257 bio->bi_end_io = write_bdev_super_endio;
258 bio->bi_private = dc;
261 /* I/O request sent to backing device */
262 __write_super(&dc->sb, bio);
264 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
267 static void write_super_endio(struct bio *bio)
269 struct cache *ca = bio->bi_private;
272 bch_count_io_errors(ca, bio->bi_status, 0,
273 "writing superblock");
274 closure_put(&ca->set->sb_write);
277 static void bcache_write_super_unlock(struct closure *cl)
279 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
281 up(&c->sb_write_mutex);
284 void bcache_write_super(struct cache_set *c)
286 struct closure *cl = &c->sb_write;
290 down(&c->sb_write_mutex);
291 closure_init(cl, &c->cl);
295 for_each_cache(ca, c, i) {
296 struct bio *bio = &ca->sb_bio;
298 ca->sb.version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
299 ca->sb.seq = c->sb.seq;
300 ca->sb.last_mount = c->sb.last_mount;
302 SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
305 bio_set_dev(bio, ca->bdev);
306 bio->bi_end_io = write_super_endio;
307 bio->bi_private = ca;
310 __write_super(&ca->sb, bio);
313 closure_return_with_destructor(cl, bcache_write_super_unlock);
318 static void uuid_endio(struct bio *bio)
320 struct closure *cl = bio->bi_private;
321 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
323 cache_set_err_on(bio->bi_status, c, "accessing uuids");
324 bch_bbio_free(bio, c);
328 static void uuid_io_unlock(struct closure *cl)
330 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
332 up(&c->uuid_write_mutex);
335 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
336 struct bkey *k, struct closure *parent)
338 struct closure *cl = &c->uuid_write;
339 struct uuid_entry *u;
344 down(&c->uuid_write_mutex);
345 closure_init(cl, parent);
347 for (i = 0; i < KEY_PTRS(k); i++) {
348 struct bio *bio = bch_bbio_alloc(c);
350 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
351 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
353 bio->bi_end_io = uuid_endio;
354 bio->bi_private = cl;
355 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
356 bch_bio_map(bio, c->uuids);
358 bch_submit_bbio(bio, c, k, i);
360 if (op != REQ_OP_WRITE)
364 bch_extent_to_text(buf, sizeof(buf), k);
365 pr_debug("%s UUIDs at %s", op == REQ_OP_WRITE ? "wrote" : "read", buf);
367 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
368 if (!bch_is_zero(u->uuid, 16))
369 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
370 u - c->uuids, u->uuid, u->label,
371 u->first_reg, u->last_reg, u->invalidated);
373 closure_return_with_destructor(cl, uuid_io_unlock);
376 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
378 struct bkey *k = &j->uuid_bucket;
380 if (__bch_btree_ptr_invalid(c, k))
381 return "bad uuid pointer";
383 bkey_copy(&c->uuid_bucket, k);
384 uuid_io(c, REQ_OP_READ, 0, k, cl);
386 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
387 struct uuid_entry_v0 *u0 = (void *) c->uuids;
388 struct uuid_entry *u1 = (void *) c->uuids;
394 * Since the new uuid entry is bigger than the old, we have to
395 * convert starting at the highest memory address and work down
396 * in order to do it in place
399 for (i = c->nr_uuids - 1;
402 memcpy(u1[i].uuid, u0[i].uuid, 16);
403 memcpy(u1[i].label, u0[i].label, 32);
405 u1[i].first_reg = u0[i].first_reg;
406 u1[i].last_reg = u0[i].last_reg;
407 u1[i].invalidated = u0[i].invalidated;
417 static int __uuid_write(struct cache_set *c)
423 closure_init_stack(&cl);
424 lockdep_assert_held(&bch_register_lock);
426 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
429 SET_KEY_SIZE(&k.key, c->sb.bucket_size);
430 uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
433 /* Only one bucket used for uuid write */
434 ca = PTR_CACHE(c, &k.key, 0);
435 atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
437 bkey_copy(&c->uuid_bucket, &k.key);
442 int bch_uuid_write(struct cache_set *c)
444 int ret = __uuid_write(c);
447 bch_journal_meta(c, NULL);
452 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
454 struct uuid_entry *u;
457 u < c->uuids + c->nr_uuids; u++)
458 if (!memcmp(u->uuid, uuid, 16))
464 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
466 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
468 return uuid_find(c, zero_uuid);
472 * Bucket priorities/gens:
474 * For each bucket, we store on disk its
478 * See alloc.c for an explanation of the gen. The priority is used to implement
479 * lru (and in the future other) cache replacement policies; for most purposes
480 * it's just an opaque integer.
482 * The gens and the priorities don't have a whole lot to do with each other, and
483 * it's actually the gens that must be written out at specific times - it's no
484 * big deal if the priorities don't get written, if we lose them we just reuse
485 * buckets in suboptimal order.
487 * On disk they're stored in a packed array, and in as many buckets are required
488 * to fit them all. The buckets we use to store them form a list; the journal
489 * header points to the first bucket, the first bucket points to the second
492 * This code is used by the allocation code; periodically (whenever it runs out
493 * of buckets to allocate from) the allocation code will invalidate some
494 * buckets, but it can't use those buckets until their new gens are safely on
498 static void prio_endio(struct bio *bio)
500 struct cache *ca = bio->bi_private;
502 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
503 bch_bbio_free(bio, ca->set);
504 closure_put(&ca->prio);
507 static void prio_io(struct cache *ca, uint64_t bucket, int op,
508 unsigned long op_flags)
510 struct closure *cl = &ca->prio;
511 struct bio *bio = bch_bbio_alloc(ca->set);
513 closure_init_stack(cl);
515 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
516 bio_set_dev(bio, ca->bdev);
517 bio->bi_iter.bi_size = bucket_bytes(ca);
519 bio->bi_end_io = prio_endio;
520 bio->bi_private = ca;
521 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
522 bch_bio_map(bio, ca->disk_buckets);
524 closure_bio_submit(ca->set, bio, &ca->prio);
528 int bch_prio_write(struct cache *ca, bool wait)
534 pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu",
535 fifo_used(&ca->free[RESERVE_PRIO]),
536 fifo_used(&ca->free[RESERVE_NONE]),
537 fifo_used(&ca->free_inc));
540 * Pre-check if there are enough free buckets. In the non-blocking
541 * scenario it's better to fail early rather than starting to allocate
542 * buckets and do a cleanup later in case of failure.
545 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
546 fifo_used(&ca->free[RESERVE_NONE]);
547 if (prio_buckets(ca) > avail)
551 closure_init_stack(&cl);
553 lockdep_assert_held(&ca->set->bucket_lock);
555 ca->disk_buckets->seq++;
557 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
558 &ca->meta_sectors_written);
560 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
562 struct prio_set *p = ca->disk_buckets;
563 struct bucket_disk *d = p->data;
564 struct bucket_disk *end = d + prios_per_bucket(ca);
566 for (b = ca->buckets + i * prios_per_bucket(ca);
567 b < ca->buckets + ca->sb.nbuckets && d < end;
569 d->prio = cpu_to_le16(b->prio);
573 p->next_bucket = ca->prio_buckets[i + 1];
574 p->magic = pset_magic(&ca->sb);
575 p->csum = bch_crc64(&p->magic, bucket_bytes(ca) - 8);
577 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
578 BUG_ON(bucket == -1);
580 mutex_unlock(&ca->set->bucket_lock);
581 prio_io(ca, bucket, REQ_OP_WRITE, 0);
582 mutex_lock(&ca->set->bucket_lock);
584 ca->prio_buckets[i] = bucket;
585 atomic_dec_bug(&ca->buckets[bucket].pin);
588 mutex_unlock(&ca->set->bucket_lock);
590 bch_journal_meta(ca->set, &cl);
593 mutex_lock(&ca->set->bucket_lock);
596 * Don't want the old priorities to get garbage collected until after we
597 * finish writing the new ones, and they're journalled
599 for (i = 0; i < prio_buckets(ca); i++) {
600 if (ca->prio_last_buckets[i])
601 __bch_bucket_free(ca,
602 &ca->buckets[ca->prio_last_buckets[i]]);
604 ca->prio_last_buckets[i] = ca->prio_buckets[i];
609 static void prio_read(struct cache *ca, uint64_t bucket)
611 struct prio_set *p = ca->disk_buckets;
612 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
614 unsigned int bucket_nr = 0;
616 for (b = ca->buckets;
617 b < ca->buckets + ca->sb.nbuckets;
620 ca->prio_buckets[bucket_nr] = bucket;
621 ca->prio_last_buckets[bucket_nr] = bucket;
624 prio_io(ca, bucket, REQ_OP_READ, 0);
627 bch_crc64(&p->magic, bucket_bytes(ca) - 8))
628 pr_warn("bad csum reading priorities");
630 if (p->magic != pset_magic(&ca->sb))
631 pr_warn("bad magic reading priorities");
633 bucket = p->next_bucket;
637 b->prio = le16_to_cpu(d->prio);
638 b->gen = b->last_gc = d->gen;
644 static int open_dev(struct block_device *b, fmode_t mode)
646 struct bcache_device *d = b->bd_disk->private_data;
648 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
655 static void release_dev(struct gendisk *b, fmode_t mode)
657 struct bcache_device *d = b->private_data;
662 static int ioctl_dev(struct block_device *b, fmode_t mode,
663 unsigned int cmd, unsigned long arg)
665 struct bcache_device *d = b->bd_disk->private_data;
667 return d->ioctl(d, mode, cmd, arg);
670 static const struct block_device_operations bcache_ops = {
672 .release = release_dev,
674 .owner = THIS_MODULE,
677 void bcache_device_stop(struct bcache_device *d)
679 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
680 closure_queue(&d->cl);
683 static void bcache_device_unlink(struct bcache_device *d)
685 lockdep_assert_held(&bch_register_lock);
687 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
691 sysfs_remove_link(&d->c->kobj, d->name);
692 sysfs_remove_link(&d->kobj, "cache");
694 for_each_cache(ca, d->c, i)
695 bd_unlink_disk_holder(ca->bdev, d->disk);
699 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
705 for_each_cache(ca, d->c, i)
706 bd_link_disk_holder(ca->bdev, d->disk);
708 snprintf(d->name, BCACHEDEVNAME_SIZE,
709 "%s%u", name, d->id);
711 WARN(sysfs_create_link(&d->kobj, &c->kobj, "cache") ||
712 sysfs_create_link(&c->kobj, &d->kobj, d->name),
713 "Couldn't create device <-> cache set symlinks");
715 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
718 static void bcache_device_detach(struct bcache_device *d)
720 lockdep_assert_held(&bch_register_lock);
722 atomic_dec(&d->c->attached_dev_nr);
724 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
725 struct uuid_entry *u = d->c->uuids + d->id;
727 SET_UUID_FLASH_ONLY(u, 0);
728 memcpy(u->uuid, invalid_uuid, 16);
729 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
730 bch_uuid_write(d->c);
733 bcache_device_unlink(d);
735 d->c->devices[d->id] = NULL;
736 closure_put(&d->c->caching);
740 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
747 if (id >= c->devices_max_used)
748 c->devices_max_used = id + 1;
750 closure_get(&c->caching);
753 static inline int first_minor_to_idx(int first_minor)
755 return (first_minor/BCACHE_MINORS);
758 static inline int idx_to_first_minor(int idx)
760 return (idx * BCACHE_MINORS);
763 static void bcache_device_free(struct bcache_device *d)
765 struct gendisk *disk = d->disk;
767 lockdep_assert_held(&bch_register_lock);
770 pr_info("%s stopped", disk->disk_name);
772 pr_err("bcache device (NULL gendisk) stopped");
775 bcache_device_detach(d);
778 bool disk_added = (disk->flags & GENHD_FL_UP) != 0;
784 blk_cleanup_queue(disk->queue);
786 ida_simple_remove(&bcache_device_idx,
787 first_minor_to_idx(disk->first_minor));
792 bioset_exit(&d->bio_split);
793 kvfree(d->full_dirty_stripes);
794 kvfree(d->stripe_sectors_dirty);
796 closure_debug_destroy(&d->cl);
799 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
802 struct request_queue *q;
803 const size_t max_stripes = min_t(size_t, INT_MAX,
804 SIZE_MAX / sizeof(atomic_t));
809 d->stripe_size = 1 << 31;
811 d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
813 if (!d->nr_stripes || d->nr_stripes > max_stripes) {
814 pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
815 (unsigned int)d->nr_stripes);
819 n = d->nr_stripes * sizeof(atomic_t);
820 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
821 if (!d->stripe_sectors_dirty)
824 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
825 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
826 if (!d->full_dirty_stripes)
827 goto out_free_stripe_sectors_dirty;
829 idx = ida_simple_get(&bcache_device_idx, 0,
830 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
832 goto out_free_full_dirty_stripes;
834 if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
835 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
838 d->disk = alloc_disk(BCACHE_MINORS);
840 goto out_bioset_exit;
842 set_capacity(d->disk, sectors);
843 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
845 d->disk->major = bcache_major;
846 d->disk->first_minor = idx_to_first_minor(idx);
847 d->disk->fops = &bcache_ops;
848 d->disk->private_data = d;
850 q = blk_alloc_queue(GFP_KERNEL);
854 blk_queue_make_request(q, NULL);
857 q->backing_dev_info->congested_data = d;
858 q->limits.max_hw_sectors = UINT_MAX;
859 q->limits.max_sectors = UINT_MAX;
860 q->limits.max_segment_size = UINT_MAX;
861 q->limits.max_segments = BIO_MAX_PAGES;
862 blk_queue_max_discard_sectors(q, UINT_MAX);
863 q->limits.discard_granularity = 512;
864 q->limits.io_min = block_size;
865 q->limits.logical_block_size = block_size;
866 q->limits.physical_block_size = block_size;
867 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
868 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
869 blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
871 blk_queue_write_cache(q, true, true);
876 bioset_exit(&d->bio_split);
878 ida_simple_remove(&bcache_device_idx, idx);
879 out_free_full_dirty_stripes:
880 kvfree(d->full_dirty_stripes);
881 out_free_stripe_sectors_dirty:
882 kvfree(d->stripe_sectors_dirty);
889 static void calc_cached_dev_sectors(struct cache_set *c)
891 uint64_t sectors = 0;
892 struct cached_dev *dc;
894 list_for_each_entry(dc, &c->cached_devs, list)
895 sectors += bdev_sectors(dc->bdev);
897 c->cached_dev_sectors = sectors;
900 #define BACKING_DEV_OFFLINE_TIMEOUT 5
901 static int cached_dev_status_update(void *arg)
903 struct cached_dev *dc = arg;
904 struct request_queue *q;
907 * If this delayed worker is stopping outside, directly quit here.
908 * dc->io_disable might be set via sysfs interface, so check it
911 while (!kthread_should_stop() && !dc->io_disable) {
912 q = bdev_get_queue(dc->bdev);
913 if (blk_queue_dying(q))
914 dc->offline_seconds++;
916 dc->offline_seconds = 0;
918 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
919 pr_err("%s: device offline for %d seconds",
920 dc->backing_dev_name,
921 BACKING_DEV_OFFLINE_TIMEOUT);
922 pr_err("%s: disable I/O request due to backing "
923 "device offline", dc->disk.name);
924 dc->io_disable = true;
925 /* let others know earlier that io_disable is true */
927 bcache_device_stop(&dc->disk);
930 schedule_timeout_interruptible(HZ);
933 wait_for_kthread_stop();
938 void bch_cached_dev_run(struct cached_dev *dc)
940 struct bcache_device *d = &dc->disk;
941 char buf[SB_LABEL_SIZE + 1];
944 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
949 memcpy(buf, dc->sb.label, SB_LABEL_SIZE);
950 buf[SB_LABEL_SIZE] = '\0';
951 env[2] = kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf);
953 if (atomic_xchg(&dc->running, 1)) {
960 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
963 closure_init_stack(&cl);
965 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
966 bch_write_bdev_super(dc, &cl);
971 bd_link_disk_holder(dc->bdev, dc->disk.disk);
973 * won't show up in the uevent file, use udevadm monitor -e instead
974 * only class / kset properties are persistent
976 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
980 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
981 sysfs_create_link(&disk_to_dev(d->disk)->kobj, &d->kobj, "bcache"))
982 pr_debug("error creating sysfs link");
984 dc->status_update_thread = kthread_run(cached_dev_status_update,
985 dc, "bcache_status_update");
986 if (IS_ERR(dc->status_update_thread)) {
987 pr_warn("failed to create bcache_status_update kthread, "
988 "continue to run without monitoring backing "
994 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
995 * work dc->writeback_rate_update is running. Wait until the routine
996 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
997 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
998 * seconds, give up waiting here and continue to cancel it too.
1000 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1002 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1005 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1009 schedule_timeout_interruptible(1);
1010 } while (time_out > 0);
1013 pr_warn("give up waiting for dc->writeback_write_update to quit");
1015 cancel_delayed_work_sync(&dc->writeback_rate_update);
1018 static void cached_dev_detach_finish(struct work_struct *w)
1020 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1023 closure_init_stack(&cl);
1025 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1026 BUG_ON(refcount_read(&dc->count));
1028 mutex_lock(&bch_register_lock);
1030 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1031 cancel_writeback_rate_update_dwork(dc);
1033 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1034 kthread_stop(dc->writeback_thread);
1035 dc->writeback_thread = NULL;
1038 memset(&dc->sb.set_uuid, 0, 16);
1039 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1041 bch_write_bdev_super(dc, &cl);
1044 calc_cached_dev_sectors(dc->disk.c);
1045 bcache_device_detach(&dc->disk);
1046 list_move(&dc->list, &uncached_devices);
1048 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1049 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1051 mutex_unlock(&bch_register_lock);
1053 pr_info("Caching disabled for %s", dc->backing_dev_name);
1055 /* Drop ref we took in cached_dev_detach() */
1056 closure_put(&dc->disk.cl);
1059 void bch_cached_dev_detach(struct cached_dev *dc)
1061 lockdep_assert_held(&bch_register_lock);
1063 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1066 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1070 * Block the device from being closed and freed until we're finished
1073 closure_get(&dc->disk.cl);
1075 bch_writeback_queue(dc);
1080 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1083 uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1084 struct uuid_entry *u;
1085 struct cached_dev *exist_dc, *t;
1087 if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1088 (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1092 pr_err("Can't attach %s: already attached",
1093 dc->backing_dev_name);
1097 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1098 pr_err("Can't attach %s: shutting down",
1099 dc->backing_dev_name);
1103 if (dc->sb.block_size < c->sb.block_size) {
1105 pr_err("Couldn't attach %s: block size less than set's block size",
1106 dc->backing_dev_name);
1110 /* Check whether already attached */
1111 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1112 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1113 pr_err("Tried to attach %s but duplicate UUID already attached",
1114 dc->backing_dev_name);
1120 u = uuid_find(c, dc->sb.uuid);
1123 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1124 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1125 memcpy(u->uuid, invalid_uuid, 16);
1126 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1131 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1132 pr_err("Couldn't find uuid for %s in set",
1133 dc->backing_dev_name);
1137 u = uuid_find_empty(c);
1139 pr_err("Not caching %s, no room for UUID",
1140 dc->backing_dev_name);
1146 * Deadlocks since we're called via sysfs...
1147 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1150 if (bch_is_zero(u->uuid, 16)) {
1153 closure_init_stack(&cl);
1155 memcpy(u->uuid, dc->sb.uuid, 16);
1156 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1157 u->first_reg = u->last_reg = rtime;
1160 memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1161 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1163 bch_write_bdev_super(dc, &cl);
1166 u->last_reg = rtime;
1170 bcache_device_attach(&dc->disk, c, u - c->uuids);
1171 list_move(&dc->list, &c->cached_devs);
1172 calc_cached_dev_sectors(c);
1175 * dc->c must be set before dc->count != 0 - paired with the mb in
1179 refcount_set(&dc->count, 1);
1181 /* Block writeback thread, but spawn it */
1182 down_write(&dc->writeback_lock);
1183 if (bch_cached_dev_writeback_start(dc)) {
1184 up_write(&dc->writeback_lock);
1188 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1189 atomic_set(&dc->has_dirty, 1);
1190 bch_writeback_queue(dc);
1193 bch_sectors_dirty_init(&dc->disk);
1195 bch_cached_dev_run(dc);
1196 bcache_device_link(&dc->disk, c, "bdev");
1197 atomic_inc(&c->attached_dev_nr);
1199 /* Allow the writeback thread to proceed */
1200 up_write(&dc->writeback_lock);
1202 pr_info("Caching %s as %s on set %pU",
1203 dc->backing_dev_name,
1204 dc->disk.disk->disk_name,
1205 dc->disk.c->sb.set_uuid);
1209 void bch_cached_dev_release(struct kobject *kobj)
1211 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1214 module_put(THIS_MODULE);
1217 static void cached_dev_free(struct closure *cl)
1219 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1221 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1222 cancel_writeback_rate_update_dwork(dc);
1224 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1225 kthread_stop(dc->writeback_thread);
1226 if (!IS_ERR_OR_NULL(dc->status_update_thread))
1227 kthread_stop(dc->status_update_thread);
1229 mutex_lock(&bch_register_lock);
1231 if (atomic_read(&dc->running))
1232 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1233 bcache_device_free(&dc->disk);
1234 list_del(&dc->list);
1236 mutex_unlock(&bch_register_lock);
1238 if (dc->sb_bio.bi_inline_vecs[0].bv_page)
1239 put_page(bio_first_page_all(&dc->sb_bio));
1241 if (!IS_ERR_OR_NULL(dc->bdev))
1242 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1244 wake_up(&unregister_wait);
1246 kobject_put(&dc->disk.kobj);
1249 static void cached_dev_flush(struct closure *cl)
1251 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1252 struct bcache_device *d = &dc->disk;
1254 mutex_lock(&bch_register_lock);
1255 bcache_device_unlink(d);
1256 mutex_unlock(&bch_register_lock);
1258 bch_cache_accounting_destroy(&dc->accounting);
1259 kobject_del(&d->kobj);
1261 continue_at(cl, cached_dev_free, system_wq);
1264 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1268 struct request_queue *q = bdev_get_queue(dc->bdev);
1270 __module_get(THIS_MODULE);
1271 INIT_LIST_HEAD(&dc->list);
1272 closure_init(&dc->disk.cl, NULL);
1273 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1274 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1275 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1276 sema_init(&dc->sb_write_mutex, 1);
1277 INIT_LIST_HEAD(&dc->io_lru);
1278 spin_lock_init(&dc->io_lock);
1279 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1281 dc->sequential_cutoff = 4 << 20;
1283 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1284 list_add(&io->lru, &dc->io_lru);
1285 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1288 dc->disk.stripe_size = q->limits.io_opt >> 9;
1290 if (dc->disk.stripe_size)
1291 dc->partial_stripes_expensive =
1292 q->limits.raid_partial_stripes_expensive;
1294 ret = bcache_device_init(&dc->disk, block_size,
1295 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1299 dc->disk.disk->queue->backing_dev_info->ra_pages =
1300 max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1301 q->backing_dev_info->ra_pages);
1303 atomic_set(&dc->io_errors, 0);
1304 dc->io_disable = false;
1305 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1306 /* default to auto */
1307 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1309 bch_cached_dev_request_init(dc);
1310 bch_cached_dev_writeback_init(dc);
1314 /* Cached device - bcache superblock */
1316 static void register_bdev(struct cache_sb *sb, struct page *sb_page,
1317 struct block_device *bdev,
1318 struct cached_dev *dc)
1320 const char *err = "cannot allocate memory";
1321 struct cache_set *c;
1323 bdevname(bdev, dc->backing_dev_name);
1324 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1326 dc->bdev->bd_holder = dc;
1328 bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1329 bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1333 if (cached_dev_init(dc, sb->block_size << 9))
1336 err = "error creating kobject";
1337 if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1340 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1343 pr_info("registered backing device %s", dc->backing_dev_name);
1345 list_add(&dc->list, &uncached_devices);
1346 /* attach to a matched cache set if it exists */
1347 list_for_each_entry(c, &bch_cache_sets, list)
1348 bch_cached_dev_attach(dc, c, NULL);
1350 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1351 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE)
1352 bch_cached_dev_run(dc);
1356 pr_notice("error %s: %s", dc->backing_dev_name, err);
1357 bcache_device_stop(&dc->disk);
1360 /* Flash only volumes */
1362 void bch_flash_dev_release(struct kobject *kobj)
1364 struct bcache_device *d = container_of(kobj, struct bcache_device,
1369 static void flash_dev_free(struct closure *cl)
1371 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1373 mutex_lock(&bch_register_lock);
1374 atomic_long_sub(bcache_dev_sectors_dirty(d),
1375 &d->c->flash_dev_dirty_sectors);
1376 bcache_device_free(d);
1377 mutex_unlock(&bch_register_lock);
1378 kobject_put(&d->kobj);
1381 static void flash_dev_flush(struct closure *cl)
1383 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1385 mutex_lock(&bch_register_lock);
1386 bcache_device_unlink(d);
1387 mutex_unlock(&bch_register_lock);
1388 kobject_del(&d->kobj);
1389 continue_at(cl, flash_dev_free, system_wq);
1392 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1394 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1399 closure_init(&d->cl, NULL);
1400 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1402 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1404 if (bcache_device_init(d, block_bytes(c), u->sectors))
1407 bcache_device_attach(d, c, u - c->uuids);
1408 bch_sectors_dirty_init(d);
1409 bch_flash_dev_request_init(d);
1412 if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1415 bcache_device_link(d, c, "volume");
1419 kobject_put(&d->kobj);
1423 static int flash_devs_run(struct cache_set *c)
1426 struct uuid_entry *u;
1429 u < c->uuids + c->nr_uuids && !ret;
1431 if (UUID_FLASH_ONLY(u))
1432 ret = flash_dev_run(c, u);
1437 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1439 struct uuid_entry *u;
1441 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1444 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1447 u = uuid_find_empty(c);
1449 pr_err("Can't create volume, no room for UUID");
1453 get_random_bytes(u->uuid, 16);
1454 memset(u->label, 0, 32);
1455 u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1457 SET_UUID_FLASH_ONLY(u, 1);
1458 u->sectors = size >> 9;
1462 return flash_dev_run(c, u);
1465 bool bch_cached_dev_error(struct cached_dev *dc)
1467 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1470 dc->io_disable = true;
1471 /* make others know io_disable is true earlier */
1474 pr_err("stop %s: too many IO errors on backing device %s\n",
1475 dc->disk.disk->disk_name, dc->backing_dev_name);
1477 bcache_device_stop(&dc->disk);
1484 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1488 if (c->on_error != ON_ERROR_PANIC &&
1489 test_bit(CACHE_SET_STOPPING, &c->flags))
1492 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1493 pr_info("CACHE_SET_IO_DISABLE already set");
1496 * XXX: we can be called from atomic context
1497 * acquire_console_sem();
1500 pr_err("bcache: error on %pU: ", c->sb.set_uuid);
1502 va_start(args, fmt);
1506 pr_err(", disabling caching\n");
1508 if (c->on_error == ON_ERROR_PANIC)
1509 panic("panic forced after error\n");
1511 bch_cache_set_unregister(c);
1515 void bch_cache_set_release(struct kobject *kobj)
1517 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1520 module_put(THIS_MODULE);
1523 static void cache_set_free(struct closure *cl)
1525 struct cache_set *c = container_of(cl, struct cache_set, cl);
1529 debugfs_remove(c->debug);
1531 bch_open_buckets_free(c);
1532 bch_btree_cache_free(c);
1533 bch_journal_free(c);
1535 mutex_lock(&bch_register_lock);
1536 for_each_cache(ca, c, i)
1539 c->cache[ca->sb.nr_this_dev] = NULL;
1540 kobject_put(&ca->kobj);
1543 bch_bset_sort_state_free(&c->sort);
1544 free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1546 if (c->moving_gc_wq)
1547 destroy_workqueue(c->moving_gc_wq);
1548 bioset_exit(&c->bio_split);
1549 mempool_exit(&c->fill_iter);
1550 mempool_exit(&c->bio_meta);
1551 mempool_exit(&c->search);
1555 mutex_unlock(&bch_register_lock);
1557 pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1558 wake_up(&unregister_wait);
1560 closure_debug_destroy(&c->cl);
1561 kobject_put(&c->kobj);
1564 static void cache_set_flush(struct closure *cl)
1566 struct cache_set *c = container_of(cl, struct cache_set, caching);
1571 bch_cache_accounting_destroy(&c->accounting);
1573 kobject_put(&c->internal);
1574 kobject_del(&c->kobj);
1576 if (!IS_ERR_OR_NULL(c->gc_thread))
1577 kthread_stop(c->gc_thread);
1579 if (!IS_ERR_OR_NULL(c->root))
1580 list_add(&c->root->list, &c->btree_cache);
1582 /* Should skip this if we're unregistering because of an error */
1583 list_for_each_entry(b, &c->btree_cache, list) {
1584 mutex_lock(&b->write_lock);
1585 if (btree_node_dirty(b))
1586 __bch_btree_node_write(b, NULL);
1587 mutex_unlock(&b->write_lock);
1590 for_each_cache(ca, c, i)
1591 if (ca->alloc_thread)
1592 kthread_stop(ca->alloc_thread);
1594 if (c->journal.cur) {
1595 cancel_delayed_work_sync(&c->journal.work);
1596 /* flush last journal entry if needed */
1597 c->journal.work.work.func(&c->journal.work.work);
1604 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1605 * cache set is unregistering due to too many I/O errors. In this condition,
1606 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1607 * value and whether the broken cache has dirty data:
1609 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1610 * BCH_CACHED_STOP_AUTO 0 NO
1611 * BCH_CACHED_STOP_AUTO 1 YES
1612 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1613 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1615 * The expected behavior is, if stop_when_cache_set_failed is configured to
1616 * "auto" via sysfs interface, the bcache device will not be stopped if the
1617 * backing device is clean on the broken cache device.
1619 static void conditional_stop_bcache_device(struct cache_set *c,
1620 struct bcache_device *d,
1621 struct cached_dev *dc)
1623 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1624 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1625 d->disk->disk_name, c->sb.set_uuid);
1626 bcache_device_stop(d);
1627 } else if (atomic_read(&dc->has_dirty)) {
1629 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1630 * and dc->has_dirty == 1
1632 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1633 d->disk->disk_name);
1635 * There might be a small time gap that cache set is
1636 * released but bcache device is not. Inside this time
1637 * gap, regular I/O requests will directly go into
1638 * backing device as no cache set attached to. This
1639 * behavior may also introduce potential inconsistence
1640 * data in writeback mode while cache is dirty.
1641 * Therefore before calling bcache_device_stop() due
1642 * to a broken cache device, dc->io_disable should be
1643 * explicitly set to true.
1645 dc->io_disable = true;
1646 /* make others know io_disable is true earlier */
1648 bcache_device_stop(d);
1651 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1652 * and dc->has_dirty == 0
1654 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1655 d->disk->disk_name);
1659 static void __cache_set_unregister(struct closure *cl)
1661 struct cache_set *c = container_of(cl, struct cache_set, caching);
1662 struct cached_dev *dc;
1663 struct bcache_device *d;
1666 mutex_lock(&bch_register_lock);
1668 for (i = 0; i < c->devices_max_used; i++) {
1673 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1674 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1675 dc = container_of(d, struct cached_dev, disk);
1676 bch_cached_dev_detach(dc);
1677 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1678 conditional_stop_bcache_device(c, d, dc);
1680 bcache_device_stop(d);
1684 mutex_unlock(&bch_register_lock);
1686 continue_at(cl, cache_set_flush, system_wq);
1689 void bch_cache_set_stop(struct cache_set *c)
1691 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1692 closure_queue(&c->caching);
1695 void bch_cache_set_unregister(struct cache_set *c)
1697 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1698 bch_cache_set_stop(c);
1701 #define alloc_bucket_pages(gfp, c) \
1702 ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(bucket_pages(c))))
1704 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1707 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1712 __module_get(THIS_MODULE);
1713 closure_init(&c->cl, NULL);
1714 set_closure_fn(&c->cl, cache_set_free, system_wq);
1716 closure_init(&c->caching, &c->cl);
1717 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1719 /* Maybe create continue_at_noreturn() and use it here? */
1720 closure_set_stopped(&c->cl);
1721 closure_put(&c->cl);
1723 kobject_init(&c->kobj, &bch_cache_set_ktype);
1724 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1726 bch_cache_accounting_init(&c->accounting, &c->cl);
1728 memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1729 c->sb.block_size = sb->block_size;
1730 c->sb.bucket_size = sb->bucket_size;
1731 c->sb.nr_in_set = sb->nr_in_set;
1732 c->sb.last_mount = sb->last_mount;
1733 c->bucket_bits = ilog2(sb->bucket_size);
1734 c->block_bits = ilog2(sb->block_size);
1735 c->nr_uuids = bucket_bytes(c) / sizeof(struct uuid_entry);
1736 c->devices_max_used = 0;
1737 atomic_set(&c->attached_dev_nr, 0);
1738 c->btree_pages = bucket_pages(c);
1739 if (c->btree_pages > BTREE_MAX_PAGES)
1740 c->btree_pages = max_t(int, c->btree_pages / 4,
1743 sema_init(&c->sb_write_mutex, 1);
1744 mutex_init(&c->bucket_lock);
1745 init_waitqueue_head(&c->btree_cache_wait);
1746 spin_lock_init(&c->btree_cannibalize_lock);
1747 init_waitqueue_head(&c->bucket_wait);
1748 init_waitqueue_head(&c->gc_wait);
1749 sema_init(&c->uuid_write_mutex, 1);
1751 spin_lock_init(&c->btree_gc_time.lock);
1752 spin_lock_init(&c->btree_split_time.lock);
1753 spin_lock_init(&c->btree_read_time.lock);
1755 bch_moving_init_cache_set(c);
1757 INIT_LIST_HEAD(&c->list);
1758 INIT_LIST_HEAD(&c->cached_devs);
1759 INIT_LIST_HEAD(&c->btree_cache);
1760 INIT_LIST_HEAD(&c->btree_cache_freeable);
1761 INIT_LIST_HEAD(&c->btree_cache_freed);
1762 INIT_LIST_HEAD(&c->data_buckets);
1764 iter_size = (sb->bucket_size / sb->block_size + 1) *
1765 sizeof(struct btree_iter_set);
1767 if (!(c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL)) ||
1768 mempool_init_slab_pool(&c->search, 32, bch_search_cache) ||
1769 mempool_init_kmalloc_pool(&c->bio_meta, 2,
1770 sizeof(struct bbio) + sizeof(struct bio_vec) *
1772 mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
1773 bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1774 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER) ||
1775 !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1776 !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1777 WQ_MEM_RECLAIM, 0)) ||
1778 bch_journal_alloc(c) ||
1779 bch_btree_cache_alloc(c) ||
1780 bch_open_buckets_alloc(c) ||
1781 bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1784 c->congested_read_threshold_us = 2000;
1785 c->congested_write_threshold_us = 20000;
1786 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1787 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1791 bch_cache_set_unregister(c);
1795 static int run_cache_set(struct cache_set *c)
1797 const char *err = "cannot allocate memory";
1798 struct cached_dev *dc, *t;
1803 struct journal_replay *l;
1805 closure_init_stack(&cl);
1807 for_each_cache(ca, c, i)
1808 c->nbuckets += ca->sb.nbuckets;
1811 if (CACHE_SYNC(&c->sb)) {
1815 err = "cannot allocate memory for journal";
1816 if (bch_journal_read(c, &journal))
1819 pr_debug("btree_journal_read() done");
1821 err = "no journal entries found";
1822 if (list_empty(&journal))
1825 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1827 err = "IO error reading priorities";
1828 for_each_cache(ca, c, i)
1829 prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1832 * If prio_read() fails it'll call cache_set_error and we'll
1833 * tear everything down right away, but if we perhaps checked
1834 * sooner we could avoid journal replay.
1839 err = "bad btree root";
1840 if (__bch_btree_ptr_invalid(c, k))
1843 err = "error reading btree root";
1844 c->root = bch_btree_node_get(c, NULL, k,
1847 if (IS_ERR_OR_NULL(c->root))
1850 list_del_init(&c->root->list);
1851 rw_unlock(true, c->root);
1853 err = uuid_read(c, j, &cl);
1857 err = "error in recovery";
1858 if (bch_btree_check(c))
1861 bch_journal_mark(c, &journal);
1862 bch_initial_gc_finish(c);
1863 pr_debug("btree_check() done");
1866 * bcache_journal_next() can't happen sooner, or
1867 * btree_gc_finish() will give spurious errors about last_gc >
1868 * gc_gen - this is a hack but oh well.
1870 bch_journal_next(&c->journal);
1872 err = "error starting allocator thread";
1873 for_each_cache(ca, c, i)
1874 if (bch_cache_allocator_start(ca))
1878 * First place it's safe to allocate: btree_check() and
1879 * btree_gc_finish() have to run before we have buckets to
1880 * allocate, and bch_bucket_alloc_set() might cause a journal
1881 * entry to be written so bcache_journal_next() has to be called
1884 * If the uuids were in the old format we have to rewrite them
1885 * before the next journal entry is written:
1887 if (j->version < BCACHE_JSET_VERSION_UUID)
1890 err = "bcache: replay journal failed";
1891 if (bch_journal_replay(c, &journal))
1894 pr_notice("invalidating existing data");
1896 for_each_cache(ca, c, i) {
1899 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1900 2, SB_JOURNAL_BUCKETS);
1902 for (j = 0; j < ca->sb.keys; j++)
1903 ca->sb.d[j] = ca->sb.first_bucket + j;
1906 bch_initial_gc_finish(c);
1908 err = "error starting allocator thread";
1909 for_each_cache(ca, c, i)
1910 if (bch_cache_allocator_start(ca))
1913 mutex_lock(&c->bucket_lock);
1914 for_each_cache(ca, c, i)
1915 bch_prio_write(ca, true);
1916 mutex_unlock(&c->bucket_lock);
1918 err = "cannot allocate new UUID bucket";
1919 if (__uuid_write(c))
1922 err = "cannot allocate new btree root";
1923 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1924 if (IS_ERR_OR_NULL(c->root))
1927 mutex_lock(&c->root->write_lock);
1928 bkey_copy_key(&c->root->key, &MAX_KEY);
1929 bch_btree_node_write(c->root, &cl);
1930 mutex_unlock(&c->root->write_lock);
1932 bch_btree_set_root(c->root);
1933 rw_unlock(true, c->root);
1936 * We don't want to write the first journal entry until
1937 * everything is set up - fortunately journal entries won't be
1938 * written until the SET_CACHE_SYNC() here:
1940 SET_CACHE_SYNC(&c->sb, true);
1942 bch_journal_next(&c->journal);
1943 bch_journal_meta(c, &cl);
1946 err = "error starting gc thread";
1947 if (bch_gc_thread_start(c))
1951 c->sb.last_mount = (u32)ktime_get_real_seconds();
1952 bcache_write_super(c);
1954 list_for_each_entry_safe(dc, t, &uncached_devices, list)
1955 bch_cached_dev_attach(dc, c, NULL);
1959 set_bit(CACHE_SET_RUNNING, &c->flags);
1962 while (!list_empty(&journal)) {
1963 l = list_first_entry(&journal, struct journal_replay, list);
1969 /* XXX: test this, it's broken */
1970 bch_cache_set_error(c, "%s", err);
1975 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
1977 return ca->sb.block_size == c->sb.block_size &&
1978 ca->sb.bucket_size == c->sb.bucket_size &&
1979 ca->sb.nr_in_set == c->sb.nr_in_set;
1982 static const char *register_cache_set(struct cache *ca)
1985 const char *err = "cannot allocate memory";
1986 struct cache_set *c;
1988 list_for_each_entry(c, &bch_cache_sets, list)
1989 if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
1990 if (c->cache[ca->sb.nr_this_dev])
1991 return "duplicate cache set member";
1993 if (!can_attach_cache(ca, c))
1994 return "cache sb does not match set";
1996 if (!CACHE_SYNC(&ca->sb))
1997 SET_CACHE_SYNC(&c->sb, false);
2002 c = bch_cache_set_alloc(&ca->sb);
2006 err = "error creating kobject";
2007 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
2008 kobject_add(&c->internal, &c->kobj, "internal"))
2011 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2014 bch_debug_init_cache_set(c);
2016 list_add(&c->list, &bch_cache_sets);
2018 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2019 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2020 sysfs_create_link(&c->kobj, &ca->kobj, buf))
2024 * A special case is both ca->sb.seq and c->sb.seq are 0,
2025 * such condition happens on a new created cache device whose
2026 * super block is never flushed yet. In this case c->sb.version
2027 * and other members should be updated too, otherwise we will
2028 * have a mistaken super block version in cache set.
2030 if (ca->sb.seq > c->sb.seq || c->sb.seq == 0) {
2031 c->sb.version = ca->sb.version;
2032 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2033 c->sb.flags = ca->sb.flags;
2034 c->sb.seq = ca->sb.seq;
2035 pr_debug("set version = %llu", c->sb.version);
2038 kobject_get(&ca->kobj);
2040 ca->set->cache[ca->sb.nr_this_dev] = ca;
2041 c->cache_by_alloc[c->caches_loaded++] = ca;
2043 if (c->caches_loaded == c->sb.nr_in_set) {
2044 err = "failed to run cache set";
2045 if (run_cache_set(c) < 0)
2051 bch_cache_set_unregister(c);
2057 void bch_cache_release(struct kobject *kobj)
2059 struct cache *ca = container_of(kobj, struct cache, kobj);
2063 BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2064 ca->set->cache[ca->sb.nr_this_dev] = NULL;
2067 free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2068 kfree(ca->prio_buckets);
2071 free_heap(&ca->heap);
2072 free_fifo(&ca->free_inc);
2074 for (i = 0; i < RESERVE_NR; i++)
2075 free_fifo(&ca->free[i]);
2077 if (ca->sb_bio.bi_inline_vecs[0].bv_page)
2078 put_page(bio_first_page_all(&ca->sb_bio));
2080 if (!IS_ERR_OR_NULL(ca->bdev))
2081 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2084 module_put(THIS_MODULE);
2087 static int cache_alloc(struct cache *ca)
2090 size_t btree_buckets;
2093 __module_get(THIS_MODULE);
2094 kobject_init(&ca->kobj, &bch_cache_ktype);
2096 bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2099 * when ca->sb.njournal_buckets is not zero, journal exists,
2100 * and in bch_journal_replay(), tree node may split,
2101 * so bucket of RESERVE_BTREE type is needed,
2102 * the worst situation is all journal buckets are valid journal,
2103 * and all the keys need to replay,
2104 * so the number of RESERVE_BTREE type buckets should be as much
2105 * as journal buckets
2107 btree_buckets = ca->sb.njournal_buckets ?: 8;
2108 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2110 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets, GFP_KERNEL) ||
2111 !init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca), GFP_KERNEL) ||
2112 !init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL) ||
2113 !init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL) ||
2114 !init_fifo(&ca->free_inc, free << 2, GFP_KERNEL) ||
2115 !init_heap(&ca->heap, free << 3, GFP_KERNEL) ||
2116 !(ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2117 ca->sb.nbuckets))) ||
2118 !(ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2119 prio_buckets(ca), 2),
2121 !(ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca)))
2124 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2126 for_each_bucket(b, ca)
2127 atomic_set(&b->pin, 0);
2132 static int register_cache(struct cache_sb *sb, struct page *sb_page,
2133 struct block_device *bdev, struct cache *ca)
2135 const char *err = NULL; /* must be set for any error case */
2138 bdevname(bdev, ca->cache_dev_name);
2139 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2141 ca->bdev->bd_holder = ca;
2143 bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2144 bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2147 if (blk_queue_discard(bdev_get_queue(bdev)))
2148 ca->discard = CACHE_DISCARD(&ca->sb);
2150 ret = cache_alloc(ca);
2152 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2154 err = "cache_alloc(): -ENOMEM";
2156 err = "cache_alloc(): unknown error";
2160 if (kobject_add(&ca->kobj,
2161 &part_to_dev(bdev->bd_part)->kobj,
2163 err = "error calling kobject_add";
2168 mutex_lock(&bch_register_lock);
2169 err = register_cache_set(ca);
2170 mutex_unlock(&bch_register_lock);
2177 pr_info("registered cache device %s", ca->cache_dev_name);
2180 kobject_put(&ca->kobj);
2184 pr_notice("error %s: %s", ca->cache_dev_name, err);
2189 /* Global interfaces/init */
2191 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2192 const char *buffer, size_t size);
2194 kobj_attribute_write(register, register_bcache);
2195 kobj_attribute_write(register_quiet, register_bcache);
2197 static bool bch_is_open_backing(struct block_device *bdev)
2199 struct cache_set *c, *tc;
2200 struct cached_dev *dc, *t;
2202 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2203 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2204 if (dc->bdev == bdev)
2206 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2207 if (dc->bdev == bdev)
2212 static bool bch_is_open_cache(struct block_device *bdev)
2214 struct cache_set *c, *tc;
2218 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2219 for_each_cache(ca, c, i)
2220 if (ca->bdev == bdev)
2225 static bool bch_is_open(struct block_device *bdev)
2227 return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2230 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2231 const char *buffer, size_t size)
2234 const char *err = "cannot allocate memory";
2236 struct cache_sb *sb = NULL;
2237 struct block_device *bdev = NULL;
2238 struct page *sb_page = NULL;
2240 if (!try_module_get(THIS_MODULE))
2243 path = kstrndup(buffer, size, GFP_KERNEL);
2247 sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2251 err = "failed to open device";
2252 bdev = blkdev_get_by_path(strim(path),
2253 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2256 if (bdev == ERR_PTR(-EBUSY)) {
2257 bdev = lookup_bdev(strim(path));
2258 mutex_lock(&bch_register_lock);
2259 if (!IS_ERR(bdev) && bch_is_open(bdev))
2260 err = "device already registered";
2262 err = "device busy";
2263 mutex_unlock(&bch_register_lock);
2266 if (attr == &ksysfs_register_quiet)
2272 err = "failed to set blocksize";
2273 if (set_blocksize(bdev, 4096))
2276 err = read_super(sb, bdev, &sb_page);
2280 err = "failed to register device";
2281 if (SB_IS_BDEV(sb)) {
2282 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2287 mutex_lock(&bch_register_lock);
2288 register_bdev(sb, sb_page, bdev, dc);
2289 mutex_unlock(&bch_register_lock);
2291 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2296 if (register_cache(sb, sb_page, bdev, ca) != 0)
2304 module_put(THIS_MODULE);
2308 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2310 pr_info("error %s: %s", path, err);
2315 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2317 if (code == SYS_DOWN ||
2319 code == SYS_POWER_OFF) {
2321 unsigned long start = jiffies;
2322 bool stopped = false;
2324 struct cache_set *c, *tc;
2325 struct cached_dev *dc, *tdc;
2327 mutex_lock(&bch_register_lock);
2329 if (list_empty(&bch_cache_sets) &&
2330 list_empty(&uncached_devices))
2333 pr_info("Stopping all devices:");
2335 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2336 bch_cache_set_stop(c);
2338 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2339 bcache_device_stop(&dc->disk);
2341 /* What's a condition variable? */
2343 long timeout = start + 2 * HZ - jiffies;
2345 stopped = list_empty(&bch_cache_sets) &&
2346 list_empty(&uncached_devices);
2348 if (timeout < 0 || stopped)
2351 prepare_to_wait(&unregister_wait, &wait,
2352 TASK_UNINTERRUPTIBLE);
2354 mutex_unlock(&bch_register_lock);
2355 schedule_timeout(timeout);
2356 mutex_lock(&bch_register_lock);
2359 finish_wait(&unregister_wait, &wait);
2362 pr_info("All devices stopped");
2364 pr_notice("Timeout waiting for devices to be closed");
2366 mutex_unlock(&bch_register_lock);
2372 static struct notifier_block reboot = {
2373 .notifier_call = bcache_reboot,
2374 .priority = INT_MAX, /* before any real devices */
2377 static void bcache_exit(void)
2382 kobject_put(bcache_kobj);
2384 destroy_workqueue(bcache_wq);
2386 destroy_workqueue(bch_journal_wq);
2389 unregister_blkdev(bcache_major, "bcache");
2390 unregister_reboot_notifier(&reboot);
2391 mutex_destroy(&bch_register_lock);
2394 static int __init bcache_init(void)
2396 static const struct attribute *files[] = {
2397 &ksysfs_register.attr,
2398 &ksysfs_register_quiet.attr,
2402 mutex_init(&bch_register_lock);
2403 init_waitqueue_head(&unregister_wait);
2404 register_reboot_notifier(&reboot);
2406 bcache_major = register_blkdev(0, "bcache");
2407 if (bcache_major < 0) {
2408 unregister_reboot_notifier(&reboot);
2409 mutex_destroy(&bch_register_lock);
2410 return bcache_major;
2413 bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2417 bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2418 if (!bch_journal_wq)
2421 bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2425 if (bch_request_init() ||
2426 sysfs_create_files(bcache_kobj, files))
2429 bch_debug_init(bcache_kobj);
2430 closure_debug_init();
2438 module_exit(bcache_exit);
2439 module_init(bcache_init);