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"
18 #include <linux/blkdev.h>
19 #include <linux/debugfs.h>
20 #include <linux/genhd.h>
21 #include <linux/idr.h>
22 #include <linux/kthread.h>
23 #include <linux/workqueue.h>
24 #include <linux/module.h>
25 #include <linux/random.h>
26 #include <linux/reboot.h>
27 #include <linux/sysfs.h>
29 unsigned int bch_cutoff_writeback;
30 unsigned int bch_cutoff_writeback_sync;
32 static const char bcache_magic[] = {
33 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
34 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
37 static const char invalid_uuid[] = {
38 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
39 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
42 static struct kobject *bcache_kobj;
43 struct mutex bch_register_lock;
44 bool bcache_is_reboot;
45 LIST_HEAD(bch_cache_sets);
46 static LIST_HEAD(uncached_devices);
48 static int bcache_major;
49 static DEFINE_IDA(bcache_device_idx);
50 static wait_queue_head_t unregister_wait;
51 struct workqueue_struct *bcache_wq;
52 struct workqueue_struct *bch_journal_wq;
55 #define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
56 /* limitation of partitions number on single bcache device */
57 #define BCACHE_MINORS 128
58 /* limitation of bcache devices number on single system */
59 #define BCACHE_DEVICE_IDX_MAX ((1U << MINORBITS)/BCACHE_MINORS)
63 static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
65 unsigned int bucket_size = le16_to_cpu(s->bucket_size);
67 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES &&
68 bch_has_feature_large_bucket(sb))
69 bucket_size |= le16_to_cpu(s->bucket_size_hi) << 16;
74 static const char *read_super_common(struct cache_sb *sb, struct block_device *bdev,
75 struct cache_sb_disk *s)
80 sb->first_bucket= le16_to_cpu(s->first_bucket);
81 sb->nbuckets = le64_to_cpu(s->nbuckets);
82 sb->bucket_size = get_bucket_size(sb, s);
84 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
85 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
87 err = "Too many journal buckets";
88 if (sb->keys > SB_JOURNAL_BUCKETS)
91 err = "Too many buckets";
92 if (sb->nbuckets > LONG_MAX)
95 err = "Not enough buckets";
96 if (sb->nbuckets < 1 << 7)
99 err = "Bad block size (not power of 2)";
100 if (!is_power_of_2(sb->block_size))
103 err = "Bad block size (larger than page size)";
104 if (sb->block_size > PAGE_SECTORS)
107 err = "Bad bucket size (not power of 2)";
108 if (!is_power_of_2(sb->bucket_size))
111 err = "Bad bucket size (smaller than page size)";
112 if (sb->bucket_size < PAGE_SECTORS)
115 err = "Invalid superblock: device too small";
116 if (get_capacity(bdev->bd_disk) <
117 sb->bucket_size * sb->nbuckets)
121 if (bch_is_zero(sb->set_uuid, 16))
124 err = "Bad cache device number in set";
125 if (!sb->nr_in_set ||
126 sb->nr_in_set <= sb->nr_this_dev ||
127 sb->nr_in_set > MAX_CACHES_PER_SET)
130 err = "Journal buckets not sequential";
131 for (i = 0; i < sb->keys; i++)
132 if (sb->d[i] != sb->first_bucket + i)
135 err = "Too many journal buckets";
136 if (sb->first_bucket + sb->keys > sb->nbuckets)
139 err = "Invalid superblock: first bucket comes before end of super";
140 if (sb->first_bucket * sb->bucket_size < 16)
149 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
150 struct cache_sb_disk **res)
153 struct cache_sb_disk *s;
157 page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
158 SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
161 s = page_address(page) + offset_in_page(SB_OFFSET);
163 sb->offset = le64_to_cpu(s->offset);
164 sb->version = le64_to_cpu(s->version);
166 memcpy(sb->magic, s->magic, 16);
167 memcpy(sb->uuid, s->uuid, 16);
168 memcpy(sb->set_uuid, s->set_uuid, 16);
169 memcpy(sb->label, s->label, SB_LABEL_SIZE);
171 sb->flags = le64_to_cpu(s->flags);
172 sb->seq = le64_to_cpu(s->seq);
173 sb->last_mount = le32_to_cpu(s->last_mount);
174 sb->keys = le16_to_cpu(s->keys);
176 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
177 sb->d[i] = le64_to_cpu(s->d[i]);
179 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
180 sb->version, sb->flags, sb->seq, sb->keys);
182 err = "Not a bcache superblock (bad offset)";
183 if (sb->offset != SB_SECTOR)
186 err = "Not a bcache superblock (bad magic)";
187 if (memcmp(sb->magic, bcache_magic, 16))
190 err = "Bad checksum";
191 if (s->csum != csum_set(s))
195 if (bch_is_zero(sb->uuid, 16))
198 sb->block_size = le16_to_cpu(s->block_size);
200 err = "Superblock block size smaller than device block size";
201 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
204 switch (sb->version) {
205 case BCACHE_SB_VERSION_BDEV:
206 sb->data_offset = BDEV_DATA_START_DEFAULT;
208 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
209 case BCACHE_SB_VERSION_BDEV_WITH_FEATURES:
210 sb->data_offset = le64_to_cpu(s->data_offset);
212 err = "Bad data offset";
213 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
217 case BCACHE_SB_VERSION_CDEV:
218 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
219 err = read_super_common(sb, bdev, s);
223 case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
225 * Feature bits are needed in read_super_common(),
226 * convert them firstly.
228 sb->feature_compat = le64_to_cpu(s->feature_compat);
229 sb->feature_incompat = le64_to_cpu(s->feature_incompat);
230 sb->feature_ro_compat = le64_to_cpu(s->feature_ro_compat);
231 err = read_super_common(sb, bdev, s);
236 err = "Unsupported superblock version";
240 sb->last_mount = (u32)ktime_get_real_seconds();
248 static void write_bdev_super_endio(struct bio *bio)
250 struct cached_dev *dc = bio->bi_private;
253 bch_count_backing_io_errors(dc, bio);
255 closure_put(&dc->sb_write);
258 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
263 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
264 bio->bi_iter.bi_sector = SB_SECTOR;
265 __bio_add_page(bio, virt_to_page(out), SB_SIZE,
266 offset_in_page(out));
268 out->offset = cpu_to_le64(sb->offset);
270 memcpy(out->uuid, sb->uuid, 16);
271 memcpy(out->set_uuid, sb->set_uuid, 16);
272 memcpy(out->label, sb->label, SB_LABEL_SIZE);
274 out->flags = cpu_to_le64(sb->flags);
275 out->seq = cpu_to_le64(sb->seq);
277 out->last_mount = cpu_to_le32(sb->last_mount);
278 out->first_bucket = cpu_to_le16(sb->first_bucket);
279 out->keys = cpu_to_le16(sb->keys);
281 for (i = 0; i < sb->keys; i++)
282 out->d[i] = cpu_to_le64(sb->d[i]);
284 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
285 out->feature_compat = cpu_to_le64(sb->feature_compat);
286 out->feature_incompat = cpu_to_le64(sb->feature_incompat);
287 out->feature_ro_compat = cpu_to_le64(sb->feature_ro_compat);
290 out->version = cpu_to_le64(sb->version);
291 out->csum = csum_set(out);
293 pr_debug("ver %llu, flags %llu, seq %llu\n",
294 sb->version, sb->flags, sb->seq);
299 static void bch_write_bdev_super_unlock(struct closure *cl)
301 struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
303 up(&dc->sb_write_mutex);
306 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
308 struct closure *cl = &dc->sb_write;
309 struct bio *bio = &dc->sb_bio;
311 down(&dc->sb_write_mutex);
312 closure_init(cl, parent);
314 bio_init(bio, dc->sb_bv, 1);
315 bio_set_dev(bio, dc->bdev);
316 bio->bi_end_io = write_bdev_super_endio;
317 bio->bi_private = dc;
320 /* I/O request sent to backing device */
321 __write_super(&dc->sb, dc->sb_disk, bio);
323 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
326 static void write_super_endio(struct bio *bio)
328 struct cache *ca = bio->bi_private;
331 bch_count_io_errors(ca, bio->bi_status, 0,
332 "writing superblock");
333 closure_put(&ca->set->sb_write);
336 static void bcache_write_super_unlock(struct closure *cl)
338 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
340 up(&c->sb_write_mutex);
343 void bcache_write_super(struct cache_set *c)
345 struct closure *cl = &c->sb_write;
346 struct cache *ca = c->cache;
347 struct bio *bio = &ca->sb_bio;
348 unsigned int version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
350 down(&c->sb_write_mutex);
351 closure_init(cl, &c->cl);
355 if (c->sb.version > version)
356 version = c->sb.version;
358 ca->sb.version = version;
359 ca->sb.seq = c->sb.seq;
360 ca->sb.last_mount = c->sb.last_mount;
362 SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
364 bio_init(bio, ca->sb_bv, 1);
365 bio_set_dev(bio, ca->bdev);
366 bio->bi_end_io = write_super_endio;
367 bio->bi_private = ca;
370 __write_super(&ca->sb, ca->sb_disk, bio);
372 closure_return_with_destructor(cl, bcache_write_super_unlock);
377 static void uuid_endio(struct bio *bio)
379 struct closure *cl = bio->bi_private;
380 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
382 cache_set_err_on(bio->bi_status, c, "accessing uuids");
383 bch_bbio_free(bio, c);
387 static void uuid_io_unlock(struct closure *cl)
389 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
391 up(&c->uuid_write_mutex);
394 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
395 struct bkey *k, struct closure *parent)
397 struct closure *cl = &c->uuid_write;
398 struct uuid_entry *u;
403 down(&c->uuid_write_mutex);
404 closure_init(cl, parent);
406 for (i = 0; i < KEY_PTRS(k); i++) {
407 struct bio *bio = bch_bbio_alloc(c);
409 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
410 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
412 bio->bi_end_io = uuid_endio;
413 bio->bi_private = cl;
414 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
415 bch_bio_map(bio, c->uuids);
417 bch_submit_bbio(bio, c, k, i);
419 if (op != REQ_OP_WRITE)
423 bch_extent_to_text(buf, sizeof(buf), k);
424 pr_debug("%s UUIDs at %s\n", op == REQ_OP_WRITE ? "wrote" : "read", buf);
426 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
427 if (!bch_is_zero(u->uuid, 16))
428 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
429 u - c->uuids, u->uuid, u->label,
430 u->first_reg, u->last_reg, u->invalidated);
432 closure_return_with_destructor(cl, uuid_io_unlock);
435 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
437 struct bkey *k = &j->uuid_bucket;
439 if (__bch_btree_ptr_invalid(c, k))
440 return "bad uuid pointer";
442 bkey_copy(&c->uuid_bucket, k);
443 uuid_io(c, REQ_OP_READ, 0, k, cl);
445 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
446 struct uuid_entry_v0 *u0 = (void *) c->uuids;
447 struct uuid_entry *u1 = (void *) c->uuids;
453 * Since the new uuid entry is bigger than the old, we have to
454 * convert starting at the highest memory address and work down
455 * in order to do it in place
458 for (i = c->nr_uuids - 1;
461 memcpy(u1[i].uuid, u0[i].uuid, 16);
462 memcpy(u1[i].label, u0[i].label, 32);
464 u1[i].first_reg = u0[i].first_reg;
465 u1[i].last_reg = u0[i].last_reg;
466 u1[i].invalidated = u0[i].invalidated;
476 static int __uuid_write(struct cache_set *c)
483 closure_init_stack(&cl);
484 lockdep_assert_held(&bch_register_lock);
486 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
489 size = meta_bucket_pages(&c->sb) * PAGE_SECTORS;
490 SET_KEY_SIZE(&k.key, size);
491 uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
494 /* Only one bucket used for uuid write */
495 ca = PTR_CACHE(c, &k.key, 0);
496 atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
498 bkey_copy(&c->uuid_bucket, &k.key);
503 int bch_uuid_write(struct cache_set *c)
505 int ret = __uuid_write(c);
508 bch_journal_meta(c, NULL);
513 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
515 struct uuid_entry *u;
518 u < c->uuids + c->nr_uuids; u++)
519 if (!memcmp(u->uuid, uuid, 16))
525 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
527 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
529 return uuid_find(c, zero_uuid);
533 * Bucket priorities/gens:
535 * For each bucket, we store on disk its
539 * See alloc.c for an explanation of the gen. The priority is used to implement
540 * lru (and in the future other) cache replacement policies; for most purposes
541 * it's just an opaque integer.
543 * The gens and the priorities don't have a whole lot to do with each other, and
544 * it's actually the gens that must be written out at specific times - it's no
545 * big deal if the priorities don't get written, if we lose them we just reuse
546 * buckets in suboptimal order.
548 * On disk they're stored in a packed array, and in as many buckets are required
549 * to fit them all. The buckets we use to store them form a list; the journal
550 * header points to the first bucket, the first bucket points to the second
553 * This code is used by the allocation code; periodically (whenever it runs out
554 * of buckets to allocate from) the allocation code will invalidate some
555 * buckets, but it can't use those buckets until their new gens are safely on
559 static void prio_endio(struct bio *bio)
561 struct cache *ca = bio->bi_private;
563 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
564 bch_bbio_free(bio, ca->set);
565 closure_put(&ca->prio);
568 static void prio_io(struct cache *ca, uint64_t bucket, int op,
569 unsigned long op_flags)
571 struct closure *cl = &ca->prio;
572 struct bio *bio = bch_bbio_alloc(ca->set);
574 closure_init_stack(cl);
576 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
577 bio_set_dev(bio, ca->bdev);
578 bio->bi_iter.bi_size = meta_bucket_bytes(&ca->sb);
580 bio->bi_end_io = prio_endio;
581 bio->bi_private = ca;
582 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
583 bch_bio_map(bio, ca->disk_buckets);
585 closure_bio_submit(ca->set, bio, &ca->prio);
589 int bch_prio_write(struct cache *ca, bool wait)
595 pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
596 fifo_used(&ca->free[RESERVE_PRIO]),
597 fifo_used(&ca->free[RESERVE_NONE]),
598 fifo_used(&ca->free_inc));
601 * Pre-check if there are enough free buckets. In the non-blocking
602 * scenario it's better to fail early rather than starting to allocate
603 * buckets and do a cleanup later in case of failure.
606 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
607 fifo_used(&ca->free[RESERVE_NONE]);
608 if (prio_buckets(ca) > avail)
612 closure_init_stack(&cl);
614 lockdep_assert_held(&ca->set->bucket_lock);
616 ca->disk_buckets->seq++;
618 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
619 &ca->meta_sectors_written);
621 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
623 struct prio_set *p = ca->disk_buckets;
624 struct bucket_disk *d = p->data;
625 struct bucket_disk *end = d + prios_per_bucket(ca);
627 for (b = ca->buckets + i * prios_per_bucket(ca);
628 b < ca->buckets + ca->sb.nbuckets && d < end;
630 d->prio = cpu_to_le16(b->prio);
634 p->next_bucket = ca->prio_buckets[i + 1];
635 p->magic = pset_magic(&ca->sb);
636 p->csum = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
638 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
639 BUG_ON(bucket == -1);
641 mutex_unlock(&ca->set->bucket_lock);
642 prio_io(ca, bucket, REQ_OP_WRITE, 0);
643 mutex_lock(&ca->set->bucket_lock);
645 ca->prio_buckets[i] = bucket;
646 atomic_dec_bug(&ca->buckets[bucket].pin);
649 mutex_unlock(&ca->set->bucket_lock);
651 bch_journal_meta(ca->set, &cl);
654 mutex_lock(&ca->set->bucket_lock);
657 * Don't want the old priorities to get garbage collected until after we
658 * finish writing the new ones, and they're journalled
660 for (i = 0; i < prio_buckets(ca); i++) {
661 if (ca->prio_last_buckets[i])
662 __bch_bucket_free(ca,
663 &ca->buckets[ca->prio_last_buckets[i]]);
665 ca->prio_last_buckets[i] = ca->prio_buckets[i];
670 static int prio_read(struct cache *ca, uint64_t bucket)
672 struct prio_set *p = ca->disk_buckets;
673 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
675 unsigned int bucket_nr = 0;
678 for (b = ca->buckets;
679 b < ca->buckets + ca->sb.nbuckets;
682 ca->prio_buckets[bucket_nr] = bucket;
683 ca->prio_last_buckets[bucket_nr] = bucket;
686 prio_io(ca, bucket, REQ_OP_READ, 0);
689 bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
690 pr_warn("bad csum reading priorities\n");
694 if (p->magic != pset_magic(&ca->sb)) {
695 pr_warn("bad magic reading priorities\n");
699 bucket = p->next_bucket;
703 b->prio = le16_to_cpu(d->prio);
704 b->gen = b->last_gc = d->gen;
714 static int open_dev(struct block_device *b, fmode_t mode)
716 struct bcache_device *d = b->bd_disk->private_data;
718 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
725 static void release_dev(struct gendisk *b, fmode_t mode)
727 struct bcache_device *d = b->private_data;
732 static int ioctl_dev(struct block_device *b, fmode_t mode,
733 unsigned int cmd, unsigned long arg)
735 struct bcache_device *d = b->bd_disk->private_data;
737 return d->ioctl(d, mode, cmd, arg);
740 static const struct block_device_operations bcache_cached_ops = {
741 .submit_bio = cached_dev_submit_bio,
743 .release = release_dev,
745 .owner = THIS_MODULE,
748 static const struct block_device_operations bcache_flash_ops = {
749 .submit_bio = flash_dev_submit_bio,
751 .release = release_dev,
753 .owner = THIS_MODULE,
756 void bcache_device_stop(struct bcache_device *d)
758 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
761 * - cached device: cached_dev_flush()
762 * - flash dev: flash_dev_flush()
764 closure_queue(&d->cl);
767 static void bcache_device_unlink(struct bcache_device *d)
769 lockdep_assert_held(&bch_register_lock);
771 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
772 struct cache *ca = d->c->cache;
774 sysfs_remove_link(&d->c->kobj, d->name);
775 sysfs_remove_link(&d->kobj, "cache");
777 bd_unlink_disk_holder(ca->bdev, d->disk);
781 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
784 struct cache *ca = c->cache;
787 bd_link_disk_holder(ca->bdev, d->disk);
789 snprintf(d->name, BCACHEDEVNAME_SIZE,
790 "%s%u", name, d->id);
792 ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
794 pr_err("Couldn't create device -> cache set symlink\n");
796 ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
798 pr_err("Couldn't create cache set -> device symlink\n");
800 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
803 static void bcache_device_detach(struct bcache_device *d)
805 lockdep_assert_held(&bch_register_lock);
807 atomic_dec(&d->c->attached_dev_nr);
809 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
810 struct uuid_entry *u = d->c->uuids + d->id;
812 SET_UUID_FLASH_ONLY(u, 0);
813 memcpy(u->uuid, invalid_uuid, 16);
814 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
815 bch_uuid_write(d->c);
818 bcache_device_unlink(d);
820 d->c->devices[d->id] = NULL;
821 closure_put(&d->c->caching);
825 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
832 if (id >= c->devices_max_used)
833 c->devices_max_used = id + 1;
835 closure_get(&c->caching);
838 static inline int first_minor_to_idx(int first_minor)
840 return (first_minor/BCACHE_MINORS);
843 static inline int idx_to_first_minor(int idx)
845 return (idx * BCACHE_MINORS);
848 static void bcache_device_free(struct bcache_device *d)
850 struct gendisk *disk = d->disk;
852 lockdep_assert_held(&bch_register_lock);
855 pr_info("%s stopped\n", disk->disk_name);
857 pr_err("bcache device (NULL gendisk) stopped\n");
860 bcache_device_detach(d);
863 bool disk_added = (disk->flags & GENHD_FL_UP) != 0;
869 blk_cleanup_queue(disk->queue);
871 ida_simple_remove(&bcache_device_idx,
872 first_minor_to_idx(disk->first_minor));
877 bioset_exit(&d->bio_split);
878 kvfree(d->full_dirty_stripes);
879 kvfree(d->stripe_sectors_dirty);
881 closure_debug_destroy(&d->cl);
884 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
885 sector_t sectors, struct block_device *cached_bdev,
886 const struct block_device_operations *ops)
888 struct request_queue *q;
889 const size_t max_stripes = min_t(size_t, INT_MAX,
890 SIZE_MAX / sizeof(atomic_t));
895 d->stripe_size = 1 << 31;
897 n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
898 if (!n || n > max_stripes) {
899 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
905 n = d->nr_stripes * sizeof(atomic_t);
906 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
907 if (!d->stripe_sectors_dirty)
910 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
911 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
912 if (!d->full_dirty_stripes)
915 idx = ida_simple_get(&bcache_device_idx, 0,
916 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
920 if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
921 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
924 d->disk = alloc_disk(BCACHE_MINORS);
928 set_capacity(d->disk, sectors);
929 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
931 d->disk->major = bcache_major;
932 d->disk->first_minor = idx_to_first_minor(idx);
934 d->disk->private_data = d;
936 q = blk_alloc_queue(NUMA_NO_NODE);
941 q->limits.max_hw_sectors = UINT_MAX;
942 q->limits.max_sectors = UINT_MAX;
943 q->limits.max_segment_size = UINT_MAX;
944 q->limits.max_segments = BIO_MAX_PAGES;
945 blk_queue_max_discard_sectors(q, UINT_MAX);
946 q->limits.discard_granularity = 512;
947 q->limits.io_min = block_size;
948 q->limits.logical_block_size = block_size;
949 q->limits.physical_block_size = block_size;
951 if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
953 * This should only happen with BCACHE_SB_VERSION_BDEV.
954 * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
956 pr_info("%s: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
957 d->disk->disk_name, q->limits.logical_block_size,
958 PAGE_SIZE, bdev_logical_block_size(cached_bdev));
960 /* This also adjusts physical block size/min io size if needed */
961 blk_queue_logical_block_size(q, bdev_logical_block_size(cached_bdev));
964 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
965 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
966 blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
968 blk_queue_write_cache(q, true, true);
973 ida_simple_remove(&bcache_device_idx, idx);
980 static void calc_cached_dev_sectors(struct cache_set *c)
982 uint64_t sectors = 0;
983 struct cached_dev *dc;
985 list_for_each_entry(dc, &c->cached_devs, list)
986 sectors += bdev_sectors(dc->bdev);
988 c->cached_dev_sectors = sectors;
991 #define BACKING_DEV_OFFLINE_TIMEOUT 5
992 static int cached_dev_status_update(void *arg)
994 struct cached_dev *dc = arg;
995 struct request_queue *q;
998 * If this delayed worker is stopping outside, directly quit here.
999 * dc->io_disable might be set via sysfs interface, so check it
1002 while (!kthread_should_stop() && !dc->io_disable) {
1003 q = bdev_get_queue(dc->bdev);
1004 if (blk_queue_dying(q))
1005 dc->offline_seconds++;
1007 dc->offline_seconds = 0;
1009 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1010 pr_err("%s: device offline for %d seconds\n",
1011 dc->backing_dev_name,
1012 BACKING_DEV_OFFLINE_TIMEOUT);
1013 pr_err("%s: disable I/O request due to backing device offline\n",
1015 dc->io_disable = true;
1016 /* let others know earlier that io_disable is true */
1018 bcache_device_stop(&dc->disk);
1021 schedule_timeout_interruptible(HZ);
1024 wait_for_kthread_stop();
1029 int bch_cached_dev_run(struct cached_dev *dc)
1031 struct bcache_device *d = &dc->disk;
1032 char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1035 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1036 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1040 if (dc->io_disable) {
1041 pr_err("I/O disabled on cached dev %s\n",
1042 dc->backing_dev_name);
1049 if (atomic_xchg(&dc->running, 1)) {
1053 pr_info("cached dev %s is running already\n",
1054 dc->backing_dev_name);
1059 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1062 closure_init_stack(&cl);
1064 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1065 bch_write_bdev_super(dc, &cl);
1070 bd_link_disk_holder(dc->bdev, dc->disk.disk);
1072 * won't show up in the uevent file, use udevadm monitor -e instead
1073 * only class / kset properties are persistent
1075 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1080 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1081 sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1082 &d->kobj, "bcache")) {
1083 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1087 dc->status_update_thread = kthread_run(cached_dev_status_update,
1088 dc, "bcache_status_update");
1089 if (IS_ERR(dc->status_update_thread)) {
1090 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1097 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1098 * work dc->writeback_rate_update is running. Wait until the routine
1099 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1100 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1101 * seconds, give up waiting here and continue to cancel it too.
1103 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1105 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1108 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1112 schedule_timeout_interruptible(1);
1113 } while (time_out > 0);
1116 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1118 cancel_delayed_work_sync(&dc->writeback_rate_update);
1121 static void cached_dev_detach_finish(struct work_struct *w)
1123 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1126 closure_init_stack(&cl);
1128 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1129 BUG_ON(refcount_read(&dc->count));
1132 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1133 cancel_writeback_rate_update_dwork(dc);
1135 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1136 kthread_stop(dc->writeback_thread);
1137 dc->writeback_thread = NULL;
1140 memset(&dc->sb.set_uuid, 0, 16);
1141 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1143 bch_write_bdev_super(dc, &cl);
1146 mutex_lock(&bch_register_lock);
1148 calc_cached_dev_sectors(dc->disk.c);
1149 bcache_device_detach(&dc->disk);
1150 list_move(&dc->list, &uncached_devices);
1152 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1153 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1155 mutex_unlock(&bch_register_lock);
1157 pr_info("Caching disabled for %s\n", dc->backing_dev_name);
1159 /* Drop ref we took in cached_dev_detach() */
1160 closure_put(&dc->disk.cl);
1163 void bch_cached_dev_detach(struct cached_dev *dc)
1165 lockdep_assert_held(&bch_register_lock);
1167 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1170 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1174 * Block the device from being closed and freed until we're finished
1177 closure_get(&dc->disk.cl);
1179 bch_writeback_queue(dc);
1184 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1187 uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1188 struct uuid_entry *u;
1189 struct cached_dev *exist_dc, *t;
1192 if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1193 (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1197 pr_err("Can't attach %s: already attached\n",
1198 dc->backing_dev_name);
1202 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1203 pr_err("Can't attach %s: shutting down\n",
1204 dc->backing_dev_name);
1208 if (dc->sb.block_size < c->sb.block_size) {
1210 pr_err("Couldn't attach %s: block size less than set's block size\n",
1211 dc->backing_dev_name);
1215 /* Check whether already attached */
1216 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1217 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1218 pr_err("Tried to attach %s but duplicate UUID already attached\n",
1219 dc->backing_dev_name);
1225 u = uuid_find(c, dc->sb.uuid);
1228 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1229 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1230 memcpy(u->uuid, invalid_uuid, 16);
1231 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1236 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1237 pr_err("Couldn't find uuid for %s in set\n",
1238 dc->backing_dev_name);
1242 u = uuid_find_empty(c);
1244 pr_err("Not caching %s, no room for UUID\n",
1245 dc->backing_dev_name);
1251 * Deadlocks since we're called via sysfs...
1252 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1255 if (bch_is_zero(u->uuid, 16)) {
1258 closure_init_stack(&cl);
1260 memcpy(u->uuid, dc->sb.uuid, 16);
1261 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1262 u->first_reg = u->last_reg = rtime;
1265 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1266 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1268 bch_write_bdev_super(dc, &cl);
1271 u->last_reg = rtime;
1275 bcache_device_attach(&dc->disk, c, u - c->uuids);
1276 list_move(&dc->list, &c->cached_devs);
1277 calc_cached_dev_sectors(c);
1280 * dc->c must be set before dc->count != 0 - paired with the mb in
1284 refcount_set(&dc->count, 1);
1286 /* Block writeback thread, but spawn it */
1287 down_write(&dc->writeback_lock);
1288 if (bch_cached_dev_writeback_start(dc)) {
1289 up_write(&dc->writeback_lock);
1290 pr_err("Couldn't start writeback facilities for %s\n",
1291 dc->disk.disk->disk_name);
1295 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1296 atomic_set(&dc->has_dirty, 1);
1297 bch_writeback_queue(dc);
1300 bch_sectors_dirty_init(&dc->disk);
1302 ret = bch_cached_dev_run(dc);
1303 if (ret && (ret != -EBUSY)) {
1304 up_write(&dc->writeback_lock);
1306 * bch_register_lock is held, bcache_device_stop() is not
1307 * able to be directly called. The kthread and kworker
1308 * created previously in bch_cached_dev_writeback_start()
1309 * have to be stopped manually here.
1311 kthread_stop(dc->writeback_thread);
1312 cancel_writeback_rate_update_dwork(dc);
1313 pr_err("Couldn't run cached device %s\n",
1314 dc->backing_dev_name);
1318 bcache_device_link(&dc->disk, c, "bdev");
1319 atomic_inc(&c->attached_dev_nr);
1321 /* Allow the writeback thread to proceed */
1322 up_write(&dc->writeback_lock);
1324 pr_info("Caching %s as %s on set %pU\n",
1325 dc->backing_dev_name,
1326 dc->disk.disk->disk_name,
1327 dc->disk.c->set_uuid);
1331 /* when dc->disk.kobj released */
1332 void bch_cached_dev_release(struct kobject *kobj)
1334 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1337 module_put(THIS_MODULE);
1340 static void cached_dev_free(struct closure *cl)
1342 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1344 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1345 cancel_writeback_rate_update_dwork(dc);
1347 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1348 kthread_stop(dc->writeback_thread);
1349 if (!IS_ERR_OR_NULL(dc->status_update_thread))
1350 kthread_stop(dc->status_update_thread);
1352 mutex_lock(&bch_register_lock);
1354 if (atomic_read(&dc->running))
1355 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1356 bcache_device_free(&dc->disk);
1357 list_del(&dc->list);
1359 mutex_unlock(&bch_register_lock);
1362 put_page(virt_to_page(dc->sb_disk));
1364 if (!IS_ERR_OR_NULL(dc->bdev))
1365 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1367 wake_up(&unregister_wait);
1369 kobject_put(&dc->disk.kobj);
1372 static void cached_dev_flush(struct closure *cl)
1374 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1375 struct bcache_device *d = &dc->disk;
1377 mutex_lock(&bch_register_lock);
1378 bcache_device_unlink(d);
1379 mutex_unlock(&bch_register_lock);
1381 bch_cache_accounting_destroy(&dc->accounting);
1382 kobject_del(&d->kobj);
1384 continue_at(cl, cached_dev_free, system_wq);
1387 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1391 struct request_queue *q = bdev_get_queue(dc->bdev);
1393 __module_get(THIS_MODULE);
1394 INIT_LIST_HEAD(&dc->list);
1395 closure_init(&dc->disk.cl, NULL);
1396 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1397 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1398 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1399 sema_init(&dc->sb_write_mutex, 1);
1400 INIT_LIST_HEAD(&dc->io_lru);
1401 spin_lock_init(&dc->io_lock);
1402 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1404 dc->sequential_cutoff = 4 << 20;
1406 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1407 list_add(&io->lru, &dc->io_lru);
1408 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1411 dc->disk.stripe_size = q->limits.io_opt >> 9;
1413 if (dc->disk.stripe_size)
1414 dc->partial_stripes_expensive =
1415 q->limits.raid_partial_stripes_expensive;
1417 ret = bcache_device_init(&dc->disk, block_size,
1418 dc->bdev->bd_part->nr_sects - dc->sb.data_offset,
1419 dc->bdev, &bcache_cached_ops);
1423 blk_queue_io_opt(dc->disk.disk->queue,
1424 max(queue_io_opt(dc->disk.disk->queue), queue_io_opt(q)));
1426 atomic_set(&dc->io_errors, 0);
1427 dc->io_disable = false;
1428 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1429 /* default to auto */
1430 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1432 bch_cached_dev_request_init(dc);
1433 bch_cached_dev_writeback_init(dc);
1437 /* Cached device - bcache superblock */
1439 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1440 struct block_device *bdev,
1441 struct cached_dev *dc)
1443 const char *err = "cannot allocate memory";
1444 struct cache_set *c;
1447 bdevname(bdev, dc->backing_dev_name);
1448 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1450 dc->bdev->bd_holder = dc;
1451 dc->sb_disk = sb_disk;
1453 if (cached_dev_init(dc, sb->block_size << 9))
1456 err = "error creating kobject";
1457 if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1460 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1463 pr_info("registered backing device %s\n", dc->backing_dev_name);
1465 list_add(&dc->list, &uncached_devices);
1466 /* attach to a matched cache set if it exists */
1467 list_for_each_entry(c, &bch_cache_sets, list)
1468 bch_cached_dev_attach(dc, c, NULL);
1470 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1471 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1472 err = "failed to run cached device";
1473 ret = bch_cached_dev_run(dc);
1480 pr_notice("error %s: %s\n", dc->backing_dev_name, err);
1481 bcache_device_stop(&dc->disk);
1485 /* Flash only volumes */
1487 /* When d->kobj released */
1488 void bch_flash_dev_release(struct kobject *kobj)
1490 struct bcache_device *d = container_of(kobj, struct bcache_device,
1495 static void flash_dev_free(struct closure *cl)
1497 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1499 mutex_lock(&bch_register_lock);
1500 atomic_long_sub(bcache_dev_sectors_dirty(d),
1501 &d->c->flash_dev_dirty_sectors);
1502 bcache_device_free(d);
1503 mutex_unlock(&bch_register_lock);
1504 kobject_put(&d->kobj);
1507 static void flash_dev_flush(struct closure *cl)
1509 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1511 mutex_lock(&bch_register_lock);
1512 bcache_device_unlink(d);
1513 mutex_unlock(&bch_register_lock);
1514 kobject_del(&d->kobj);
1515 continue_at(cl, flash_dev_free, system_wq);
1518 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1520 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1525 closure_init(&d->cl, NULL);
1526 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1528 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1530 if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1531 NULL, &bcache_flash_ops))
1534 bcache_device_attach(d, c, u - c->uuids);
1535 bch_sectors_dirty_init(d);
1536 bch_flash_dev_request_init(d);
1539 if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1542 bcache_device_link(d, c, "volume");
1546 kobject_put(&d->kobj);
1550 static int flash_devs_run(struct cache_set *c)
1553 struct uuid_entry *u;
1556 u < c->uuids + c->nr_uuids && !ret;
1558 if (UUID_FLASH_ONLY(u))
1559 ret = flash_dev_run(c, u);
1564 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1566 struct uuid_entry *u;
1568 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1571 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1574 u = uuid_find_empty(c);
1576 pr_err("Can't create volume, no room for UUID\n");
1580 get_random_bytes(u->uuid, 16);
1581 memset(u->label, 0, 32);
1582 u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1584 SET_UUID_FLASH_ONLY(u, 1);
1585 u->sectors = size >> 9;
1589 return flash_dev_run(c, u);
1592 bool bch_cached_dev_error(struct cached_dev *dc)
1594 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1597 dc->io_disable = true;
1598 /* make others know io_disable is true earlier */
1601 pr_err("stop %s: too many IO errors on backing device %s\n",
1602 dc->disk.disk->disk_name, dc->backing_dev_name);
1604 bcache_device_stop(&dc->disk);
1611 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1613 struct va_format vaf;
1616 if (c->on_error != ON_ERROR_PANIC &&
1617 test_bit(CACHE_SET_STOPPING, &c->flags))
1620 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1621 pr_info("CACHE_SET_IO_DISABLE already set\n");
1624 * XXX: we can be called from atomic context
1625 * acquire_console_sem();
1628 va_start(args, fmt);
1633 pr_err("error on %pU: %pV, disabling caching\n",
1638 if (c->on_error == ON_ERROR_PANIC)
1639 panic("panic forced after error\n");
1641 bch_cache_set_unregister(c);
1645 /* When c->kobj released */
1646 void bch_cache_set_release(struct kobject *kobj)
1648 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1651 module_put(THIS_MODULE);
1654 static void cache_set_free(struct closure *cl)
1656 struct cache_set *c = container_of(cl, struct cache_set, cl);
1659 debugfs_remove(c->debug);
1661 bch_open_buckets_free(c);
1662 bch_btree_cache_free(c);
1663 bch_journal_free(c);
1665 mutex_lock(&bch_register_lock);
1670 kobject_put(&ca->kobj);
1673 bch_bset_sort_state_free(&c->sort);
1674 free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->sb)));
1676 if (c->moving_gc_wq)
1677 destroy_workqueue(c->moving_gc_wq);
1678 bioset_exit(&c->bio_split);
1679 mempool_exit(&c->fill_iter);
1680 mempool_exit(&c->bio_meta);
1681 mempool_exit(&c->search);
1685 mutex_unlock(&bch_register_lock);
1687 pr_info("Cache set %pU unregistered\n", c->set_uuid);
1688 wake_up(&unregister_wait);
1690 closure_debug_destroy(&c->cl);
1691 kobject_put(&c->kobj);
1694 static void cache_set_flush(struct closure *cl)
1696 struct cache_set *c = container_of(cl, struct cache_set, caching);
1697 struct cache *ca = c->cache;
1700 bch_cache_accounting_destroy(&c->accounting);
1702 kobject_put(&c->internal);
1703 kobject_del(&c->kobj);
1705 if (!IS_ERR_OR_NULL(c->gc_thread))
1706 kthread_stop(c->gc_thread);
1708 if (!IS_ERR_OR_NULL(c->root))
1709 list_add(&c->root->list, &c->btree_cache);
1712 * Avoid flushing cached nodes if cache set is retiring
1713 * due to too many I/O errors detected.
1715 if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1716 list_for_each_entry(b, &c->btree_cache, list) {
1717 mutex_lock(&b->write_lock);
1718 if (btree_node_dirty(b))
1719 __bch_btree_node_write(b, NULL);
1720 mutex_unlock(&b->write_lock);
1723 if (ca->alloc_thread)
1724 kthread_stop(ca->alloc_thread);
1726 if (c->journal.cur) {
1727 cancel_delayed_work_sync(&c->journal.work);
1728 /* flush last journal entry if needed */
1729 c->journal.work.work.func(&c->journal.work.work);
1736 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1737 * cache set is unregistering due to too many I/O errors. In this condition,
1738 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1739 * value and whether the broken cache has dirty data:
1741 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1742 * BCH_CACHED_STOP_AUTO 0 NO
1743 * BCH_CACHED_STOP_AUTO 1 YES
1744 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1745 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1747 * The expected behavior is, if stop_when_cache_set_failed is configured to
1748 * "auto" via sysfs interface, the bcache device will not be stopped if the
1749 * backing device is clean on the broken cache device.
1751 static void conditional_stop_bcache_device(struct cache_set *c,
1752 struct bcache_device *d,
1753 struct cached_dev *dc)
1755 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1756 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1757 d->disk->disk_name, c->set_uuid);
1758 bcache_device_stop(d);
1759 } else if (atomic_read(&dc->has_dirty)) {
1761 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1762 * and dc->has_dirty == 1
1764 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1765 d->disk->disk_name);
1767 * There might be a small time gap that cache set is
1768 * released but bcache device is not. Inside this time
1769 * gap, regular I/O requests will directly go into
1770 * backing device as no cache set attached to. This
1771 * behavior may also introduce potential inconsistence
1772 * data in writeback mode while cache is dirty.
1773 * Therefore before calling bcache_device_stop() due
1774 * to a broken cache device, dc->io_disable should be
1775 * explicitly set to true.
1777 dc->io_disable = true;
1778 /* make others know io_disable is true earlier */
1780 bcache_device_stop(d);
1783 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1784 * and dc->has_dirty == 0
1786 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1787 d->disk->disk_name);
1791 static void __cache_set_unregister(struct closure *cl)
1793 struct cache_set *c = container_of(cl, struct cache_set, caching);
1794 struct cached_dev *dc;
1795 struct bcache_device *d;
1798 mutex_lock(&bch_register_lock);
1800 for (i = 0; i < c->devices_max_used; i++) {
1805 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1806 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1807 dc = container_of(d, struct cached_dev, disk);
1808 bch_cached_dev_detach(dc);
1809 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1810 conditional_stop_bcache_device(c, d, dc);
1812 bcache_device_stop(d);
1816 mutex_unlock(&bch_register_lock);
1818 continue_at(cl, cache_set_flush, system_wq);
1821 void bch_cache_set_stop(struct cache_set *c)
1823 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1824 /* closure_fn set to __cache_set_unregister() */
1825 closure_queue(&c->caching);
1828 void bch_cache_set_unregister(struct cache_set *c)
1830 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1831 bch_cache_set_stop(c);
1834 #define alloc_meta_bucket_pages(gfp, sb) \
1835 ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1837 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1840 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1845 __module_get(THIS_MODULE);
1846 closure_init(&c->cl, NULL);
1847 set_closure_fn(&c->cl, cache_set_free, system_wq);
1849 closure_init(&c->caching, &c->cl);
1850 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1852 /* Maybe create continue_at_noreturn() and use it here? */
1853 closure_set_stopped(&c->cl);
1854 closure_put(&c->cl);
1856 kobject_init(&c->kobj, &bch_cache_set_ktype);
1857 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1859 bch_cache_accounting_init(&c->accounting, &c->cl);
1861 memcpy(c->set_uuid, sb->set_uuid, 16);
1862 c->sb.block_size = sb->block_size;
1863 c->sb.bucket_size = sb->bucket_size;
1864 c->sb.nr_in_set = sb->nr_in_set;
1865 c->sb.last_mount = sb->last_mount;
1866 c->sb.version = sb->version;
1867 if (c->sb.version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
1868 c->sb.feature_compat = sb->feature_compat;
1869 c->sb.feature_ro_compat = sb->feature_ro_compat;
1870 c->sb.feature_incompat = sb->feature_incompat;
1873 c->bucket_bits = ilog2(sb->bucket_size);
1874 c->block_bits = ilog2(sb->block_size);
1875 c->nr_uuids = meta_bucket_bytes(&c->sb) / sizeof(struct uuid_entry);
1876 c->devices_max_used = 0;
1877 atomic_set(&c->attached_dev_nr, 0);
1878 c->btree_pages = meta_bucket_pages(&c->sb);
1879 if (c->btree_pages > BTREE_MAX_PAGES)
1880 c->btree_pages = max_t(int, c->btree_pages / 4,
1883 sema_init(&c->sb_write_mutex, 1);
1884 mutex_init(&c->bucket_lock);
1885 init_waitqueue_head(&c->btree_cache_wait);
1886 spin_lock_init(&c->btree_cannibalize_lock);
1887 init_waitqueue_head(&c->bucket_wait);
1888 init_waitqueue_head(&c->gc_wait);
1889 sema_init(&c->uuid_write_mutex, 1);
1891 spin_lock_init(&c->btree_gc_time.lock);
1892 spin_lock_init(&c->btree_split_time.lock);
1893 spin_lock_init(&c->btree_read_time.lock);
1895 bch_moving_init_cache_set(c);
1897 INIT_LIST_HEAD(&c->list);
1898 INIT_LIST_HEAD(&c->cached_devs);
1899 INIT_LIST_HEAD(&c->btree_cache);
1900 INIT_LIST_HEAD(&c->btree_cache_freeable);
1901 INIT_LIST_HEAD(&c->btree_cache_freed);
1902 INIT_LIST_HEAD(&c->data_buckets);
1904 iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1905 sizeof(struct btree_iter_set);
1907 c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1911 if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1914 if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1915 sizeof(struct bbio) +
1916 sizeof(struct bio_vec) * meta_bucket_pages(&c->sb)))
1919 if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1922 if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1923 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
1926 c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, &c->sb);
1930 c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1931 if (!c->moving_gc_wq)
1934 if (bch_journal_alloc(c))
1937 if (bch_btree_cache_alloc(c))
1940 if (bch_open_buckets_alloc(c))
1943 if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1946 c->congested_read_threshold_us = 2000;
1947 c->congested_write_threshold_us = 20000;
1948 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1949 c->idle_max_writeback_rate_enabled = 1;
1950 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1954 bch_cache_set_unregister(c);
1958 static int run_cache_set(struct cache_set *c)
1960 const char *err = "cannot allocate memory";
1961 struct cached_dev *dc, *t;
1962 struct cache *ca = c->cache;
1965 struct journal_replay *l;
1967 closure_init_stack(&cl);
1969 c->nbuckets = ca->sb.nbuckets;
1972 if (CACHE_SYNC(&c->sb)) {
1976 err = "cannot allocate memory for journal";
1977 if (bch_journal_read(c, &journal))
1980 pr_debug("btree_journal_read() done\n");
1982 err = "no journal entries found";
1983 if (list_empty(&journal))
1986 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1988 err = "IO error reading priorities";
1989 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
1993 * If prio_read() fails it'll call cache_set_error and we'll
1994 * tear everything down right away, but if we perhaps checked
1995 * sooner we could avoid journal replay.
2000 err = "bad btree root";
2001 if (__bch_btree_ptr_invalid(c, k))
2004 err = "error reading btree root";
2005 c->root = bch_btree_node_get(c, NULL, k,
2008 if (IS_ERR_OR_NULL(c->root))
2011 list_del_init(&c->root->list);
2012 rw_unlock(true, c->root);
2014 err = uuid_read(c, j, &cl);
2018 err = "error in recovery";
2019 if (bch_btree_check(c))
2022 bch_journal_mark(c, &journal);
2023 bch_initial_gc_finish(c);
2024 pr_debug("btree_check() done\n");
2027 * bcache_journal_next() can't happen sooner, or
2028 * btree_gc_finish() will give spurious errors about last_gc >
2029 * gc_gen - this is a hack but oh well.
2031 bch_journal_next(&c->journal);
2033 err = "error starting allocator thread";
2034 if (bch_cache_allocator_start(ca))
2038 * First place it's safe to allocate: btree_check() and
2039 * btree_gc_finish() have to run before we have buckets to
2040 * allocate, and bch_bucket_alloc_set() might cause a journal
2041 * entry to be written so bcache_journal_next() has to be called
2044 * If the uuids were in the old format we have to rewrite them
2045 * before the next journal entry is written:
2047 if (j->version < BCACHE_JSET_VERSION_UUID)
2050 err = "bcache: replay journal failed";
2051 if (bch_journal_replay(c, &journal))
2056 pr_notice("invalidating existing data\n");
2057 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2058 2, SB_JOURNAL_BUCKETS);
2060 for (j = 0; j < ca->sb.keys; j++)
2061 ca->sb.d[j] = ca->sb.first_bucket + j;
2063 bch_initial_gc_finish(c);
2065 err = "error starting allocator thread";
2066 if (bch_cache_allocator_start(ca))
2069 mutex_lock(&c->bucket_lock);
2070 bch_prio_write(ca, true);
2071 mutex_unlock(&c->bucket_lock);
2073 err = "cannot allocate new UUID bucket";
2074 if (__uuid_write(c))
2077 err = "cannot allocate new btree root";
2078 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2079 if (IS_ERR_OR_NULL(c->root))
2082 mutex_lock(&c->root->write_lock);
2083 bkey_copy_key(&c->root->key, &MAX_KEY);
2084 bch_btree_node_write(c->root, &cl);
2085 mutex_unlock(&c->root->write_lock);
2087 bch_btree_set_root(c->root);
2088 rw_unlock(true, c->root);
2091 * We don't want to write the first journal entry until
2092 * everything is set up - fortunately journal entries won't be
2093 * written until the SET_CACHE_SYNC() here:
2095 SET_CACHE_SYNC(&c->sb, true);
2097 bch_journal_next(&c->journal);
2098 bch_journal_meta(c, &cl);
2101 err = "error starting gc thread";
2102 if (bch_gc_thread_start(c))
2106 c->sb.last_mount = (u32)ktime_get_real_seconds();
2107 bcache_write_super(c);
2109 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2110 bch_cached_dev_attach(dc, c, NULL);
2114 set_bit(CACHE_SET_RUNNING, &c->flags);
2117 while (!list_empty(&journal)) {
2118 l = list_first_entry(&journal, struct journal_replay, list);
2125 bch_cache_set_error(c, "%s", err);
2130 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
2132 return ca->sb.block_size == c->sb.block_size &&
2133 ca->sb.bucket_size == c->sb.bucket_size &&
2134 ca->sb.nr_in_set == c->sb.nr_in_set;
2137 static const char *register_cache_set(struct cache *ca)
2140 const char *err = "cannot allocate memory";
2141 struct cache_set *c;
2143 list_for_each_entry(c, &bch_cache_sets, list)
2144 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2146 return "duplicate cache set member";
2148 if (!can_attach_cache(ca, c))
2149 return "cache sb does not match set";
2151 if (!CACHE_SYNC(&ca->sb))
2152 SET_CACHE_SYNC(&c->sb, false);
2157 c = bch_cache_set_alloc(&ca->sb);
2161 err = "error creating kobject";
2162 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2163 kobject_add(&c->internal, &c->kobj, "internal"))
2166 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2169 bch_debug_init_cache_set(c);
2171 list_add(&c->list, &bch_cache_sets);
2173 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2174 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2175 sysfs_create_link(&c->kobj, &ca->kobj, buf))
2179 * A special case is both ca->sb.seq and c->sb.seq are 0,
2180 * such condition happens on a new created cache device whose
2181 * super block is never flushed yet. In this case c->sb.version
2182 * and other members should be updated too, otherwise we will
2183 * have a mistaken super block version in cache set.
2185 if (ca->sb.seq > c->sb.seq || c->sb.seq == 0) {
2186 c->sb.version = ca->sb.version;
2187 memcpy(c->set_uuid, ca->sb.set_uuid, 16);
2188 c->sb.flags = ca->sb.flags;
2189 c->sb.seq = ca->sb.seq;
2190 pr_debug("set version = %llu\n", c->sb.version);
2193 kobject_get(&ca->kobj);
2195 ca->set->cache = ca;
2197 err = "failed to run cache set";
2198 if (run_cache_set(c) < 0)
2203 bch_cache_set_unregister(c);
2209 /* When ca->kobj released */
2210 void bch_cache_release(struct kobject *kobj)
2212 struct cache *ca = container_of(kobj, struct cache, kobj);
2216 BUG_ON(ca->set->cache != ca);
2217 ca->set->cache = NULL;
2220 free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2221 kfree(ca->prio_buckets);
2224 free_heap(&ca->heap);
2225 free_fifo(&ca->free_inc);
2227 for (i = 0; i < RESERVE_NR; i++)
2228 free_fifo(&ca->free[i]);
2231 put_page(virt_to_page(ca->sb_disk));
2233 if (!IS_ERR_OR_NULL(ca->bdev))
2234 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2237 module_put(THIS_MODULE);
2240 static int cache_alloc(struct cache *ca)
2243 size_t btree_buckets;
2246 const char *err = NULL;
2248 __module_get(THIS_MODULE);
2249 kobject_init(&ca->kobj, &bch_cache_ktype);
2251 bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2254 * when ca->sb.njournal_buckets is not zero, journal exists,
2255 * and in bch_journal_replay(), tree node may split,
2256 * so bucket of RESERVE_BTREE type is needed,
2257 * the worst situation is all journal buckets are valid journal,
2258 * and all the keys need to replay,
2259 * so the number of RESERVE_BTREE type buckets should be as much
2260 * as journal buckets
2262 btree_buckets = ca->sb.njournal_buckets ?: 8;
2263 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2266 err = "ca->sb.nbuckets is too small";
2270 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2272 err = "ca->free[RESERVE_BTREE] alloc failed";
2273 goto err_btree_alloc;
2276 if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2278 err = "ca->free[RESERVE_PRIO] alloc failed";
2279 goto err_prio_alloc;
2282 if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2283 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2284 goto err_movinggc_alloc;
2287 if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2288 err = "ca->free[RESERVE_NONE] alloc failed";
2289 goto err_none_alloc;
2292 if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2293 err = "ca->free_inc alloc failed";
2294 goto err_free_inc_alloc;
2297 if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2298 err = "ca->heap alloc failed";
2299 goto err_heap_alloc;
2302 ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2305 err = "ca->buckets alloc failed";
2306 goto err_buckets_alloc;
2309 ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2310 prio_buckets(ca), 2),
2312 if (!ca->prio_buckets) {
2313 err = "ca->prio_buckets alloc failed";
2314 goto err_prio_buckets_alloc;
2317 ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2318 if (!ca->disk_buckets) {
2319 err = "ca->disk_buckets alloc failed";
2320 goto err_disk_buckets_alloc;
2323 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2325 for_each_bucket(b, ca)
2326 atomic_set(&b->pin, 0);
2329 err_disk_buckets_alloc:
2330 kfree(ca->prio_buckets);
2331 err_prio_buckets_alloc:
2334 free_heap(&ca->heap);
2336 free_fifo(&ca->free_inc);
2338 free_fifo(&ca->free[RESERVE_NONE]);
2340 free_fifo(&ca->free[RESERVE_MOVINGGC]);
2342 free_fifo(&ca->free[RESERVE_PRIO]);
2344 free_fifo(&ca->free[RESERVE_BTREE]);
2347 module_put(THIS_MODULE);
2349 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2353 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2354 struct block_device *bdev, struct cache *ca)
2356 const char *err = NULL; /* must be set for any error case */
2359 bdevname(bdev, ca->cache_dev_name);
2360 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2362 ca->bdev->bd_holder = ca;
2363 ca->sb_disk = sb_disk;
2365 if (blk_queue_discard(bdev_get_queue(bdev)))
2366 ca->discard = CACHE_DISCARD(&ca->sb);
2368 ret = cache_alloc(ca);
2371 * If we failed here, it means ca->kobj is not initialized yet,
2372 * kobject_put() won't be called and there is no chance to
2373 * call blkdev_put() to bdev in bch_cache_release(). So we
2374 * explicitly call blkdev_put() here.
2376 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2378 err = "cache_alloc(): -ENOMEM";
2379 else if (ret == -EPERM)
2380 err = "cache_alloc(): cache device is too small";
2382 err = "cache_alloc(): unknown error";
2386 if (kobject_add(&ca->kobj,
2387 &part_to_dev(bdev->bd_part)->kobj,
2389 err = "error calling kobject_add";
2394 mutex_lock(&bch_register_lock);
2395 err = register_cache_set(ca);
2396 mutex_unlock(&bch_register_lock);
2403 pr_info("registered cache device %s\n", ca->cache_dev_name);
2406 kobject_put(&ca->kobj);
2410 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2415 /* Global interfaces/init */
2417 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2418 const char *buffer, size_t size);
2419 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2420 struct kobj_attribute *attr,
2421 const char *buffer, size_t size);
2423 kobj_attribute_write(register, register_bcache);
2424 kobj_attribute_write(register_quiet, register_bcache);
2425 kobj_attribute_write(pendings_cleanup, bch_pending_bdevs_cleanup);
2427 static bool bch_is_open_backing(struct block_device *bdev)
2429 struct cache_set *c, *tc;
2430 struct cached_dev *dc, *t;
2432 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2433 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2434 if (dc->bdev == bdev)
2436 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2437 if (dc->bdev == bdev)
2442 static bool bch_is_open_cache(struct block_device *bdev)
2444 struct cache_set *c, *tc;
2446 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2447 struct cache *ca = c->cache;
2449 if (ca->bdev == bdev)
2456 static bool bch_is_open(struct block_device *bdev)
2458 return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2461 struct async_reg_args {
2462 struct delayed_work reg_work;
2464 struct cache_sb *sb;
2465 struct cache_sb_disk *sb_disk;
2466 struct block_device *bdev;
2469 static void register_bdev_worker(struct work_struct *work)
2472 struct async_reg_args *args =
2473 container_of(work, struct async_reg_args, reg_work.work);
2474 struct cached_dev *dc;
2476 dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2479 put_page(virt_to_page(args->sb_disk));
2480 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2484 mutex_lock(&bch_register_lock);
2485 if (register_bdev(args->sb, args->sb_disk, args->bdev, dc) < 0)
2487 mutex_unlock(&bch_register_lock);
2491 pr_info("error %s: fail to register backing device\n",
2496 module_put(THIS_MODULE);
2499 static void register_cache_worker(struct work_struct *work)
2502 struct async_reg_args *args =
2503 container_of(work, struct async_reg_args, reg_work.work);
2506 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2509 put_page(virt_to_page(args->sb_disk));
2510 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2514 /* blkdev_put() will be called in bch_cache_release() */
2515 if (register_cache(args->sb, args->sb_disk, args->bdev, ca) != 0)
2520 pr_info("error %s: fail to register cache device\n",
2525 module_put(THIS_MODULE);
2528 static void register_device_aync(struct async_reg_args *args)
2530 if (SB_IS_BDEV(args->sb))
2531 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2533 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2535 /* 10 jiffies is enough for a delay */
2536 queue_delayed_work(system_wq, &args->reg_work, 10);
2539 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2540 const char *buffer, size_t size)
2544 struct cache_sb *sb;
2545 struct cache_sb_disk *sb_disk;
2546 struct block_device *bdev;
2548 bool async_registration = false;
2550 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2551 async_registration = true;
2555 err = "failed to reference bcache module";
2556 if (!try_module_get(THIS_MODULE))
2559 /* For latest state of bcache_is_reboot */
2561 err = "bcache is in reboot";
2562 if (bcache_is_reboot)
2563 goto out_module_put;
2566 err = "cannot allocate memory";
2567 path = kstrndup(buffer, size, GFP_KERNEL);
2569 goto out_module_put;
2571 sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2576 err = "failed to open device";
2577 bdev = blkdev_get_by_path(strim(path),
2578 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2581 if (bdev == ERR_PTR(-EBUSY)) {
2582 bdev = lookup_bdev(strim(path));
2583 mutex_lock(&bch_register_lock);
2584 if (!IS_ERR(bdev) && bch_is_open(bdev))
2585 err = "device already registered";
2587 err = "device busy";
2588 mutex_unlock(&bch_register_lock);
2591 if (attr == &ksysfs_register_quiet)
2597 err = "failed to set blocksize";
2598 if (set_blocksize(bdev, 4096))
2599 goto out_blkdev_put;
2601 err = read_super(sb, bdev, &sb_disk);
2603 goto out_blkdev_put;
2605 err = "failed to register device";
2607 if (async_registration) {
2608 /* register in asynchronous way */
2609 struct async_reg_args *args =
2610 kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2614 err = "cannot allocate memory";
2615 goto out_put_sb_page;
2620 args->sb_disk = sb_disk;
2622 register_device_aync(args);
2623 /* No wait and returns to user space */
2627 if (SB_IS_BDEV(sb)) {
2628 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2631 goto out_put_sb_page;
2633 mutex_lock(&bch_register_lock);
2634 ret = register_bdev(sb, sb_disk, bdev, dc);
2635 mutex_unlock(&bch_register_lock);
2636 /* blkdev_put() will be called in cached_dev_free() */
2640 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2643 goto out_put_sb_page;
2645 /* blkdev_put() will be called in bch_cache_release() */
2646 if (register_cache(sb, sb_disk, bdev, ca) != 0)
2653 module_put(THIS_MODULE);
2658 put_page(virt_to_page(sb_disk));
2660 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2667 module_put(THIS_MODULE);
2669 pr_info("error %s: %s\n", path?path:"", err);
2675 struct list_head list;
2676 struct cached_dev *dc;
2679 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2680 struct kobj_attribute *attr,
2684 LIST_HEAD(pending_devs);
2686 struct cached_dev *dc, *tdc;
2687 struct pdev *pdev, *tpdev;
2688 struct cache_set *c, *tc;
2690 mutex_lock(&bch_register_lock);
2691 list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2692 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2696 list_add(&pdev->list, &pending_devs);
2699 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2700 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2701 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2702 char *set_uuid = c->set_uuid;
2704 if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2705 list_del(&pdev->list);
2711 mutex_unlock(&bch_register_lock);
2713 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2714 pr_info("delete pdev %p\n", pdev);
2715 list_del(&pdev->list);
2716 bcache_device_stop(&pdev->dc->disk);
2723 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2725 if (bcache_is_reboot)
2728 if (code == SYS_DOWN ||
2730 code == SYS_POWER_OFF) {
2732 unsigned long start = jiffies;
2733 bool stopped = false;
2735 struct cache_set *c, *tc;
2736 struct cached_dev *dc, *tdc;
2738 mutex_lock(&bch_register_lock);
2740 if (bcache_is_reboot)
2743 /* New registration is rejected since now */
2744 bcache_is_reboot = true;
2746 * Make registering caller (if there is) on other CPU
2747 * core know bcache_is_reboot set to true earlier
2751 if (list_empty(&bch_cache_sets) &&
2752 list_empty(&uncached_devices))
2755 mutex_unlock(&bch_register_lock);
2757 pr_info("Stopping all devices:\n");
2760 * The reason bch_register_lock is not held to call
2761 * bch_cache_set_stop() and bcache_device_stop() is to
2762 * avoid potential deadlock during reboot, because cache
2763 * set or bcache device stopping process will acqurie
2764 * bch_register_lock too.
2766 * We are safe here because bcache_is_reboot sets to
2767 * true already, register_bcache() will reject new
2768 * registration now. bcache_is_reboot also makes sure
2769 * bcache_reboot() won't be re-entered on by other thread,
2770 * so there is no race in following list iteration by
2771 * list_for_each_entry_safe().
2773 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2774 bch_cache_set_stop(c);
2776 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2777 bcache_device_stop(&dc->disk);
2781 * Give an early chance for other kthreads and
2782 * kworkers to stop themselves
2786 /* What's a condition variable? */
2788 long timeout = start + 10 * HZ - jiffies;
2790 mutex_lock(&bch_register_lock);
2791 stopped = list_empty(&bch_cache_sets) &&
2792 list_empty(&uncached_devices);
2794 if (timeout < 0 || stopped)
2797 prepare_to_wait(&unregister_wait, &wait,
2798 TASK_UNINTERRUPTIBLE);
2800 mutex_unlock(&bch_register_lock);
2801 schedule_timeout(timeout);
2804 finish_wait(&unregister_wait, &wait);
2807 pr_info("All devices stopped\n");
2809 pr_notice("Timeout waiting for devices to be closed\n");
2811 mutex_unlock(&bch_register_lock);
2817 static struct notifier_block reboot = {
2818 .notifier_call = bcache_reboot,
2819 .priority = INT_MAX, /* before any real devices */
2822 static void bcache_exit(void)
2827 kobject_put(bcache_kobj);
2829 destroy_workqueue(bcache_wq);
2831 destroy_workqueue(bch_journal_wq);
2834 unregister_blkdev(bcache_major, "bcache");
2835 unregister_reboot_notifier(&reboot);
2836 mutex_destroy(&bch_register_lock);
2839 /* Check and fixup module parameters */
2840 static void check_module_parameters(void)
2842 if (bch_cutoff_writeback_sync == 0)
2843 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2844 else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2845 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2846 bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2847 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2850 if (bch_cutoff_writeback == 0)
2851 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2852 else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2853 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2854 bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2855 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2858 if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2859 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2860 bch_cutoff_writeback, bch_cutoff_writeback_sync);
2861 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2865 static int __init bcache_init(void)
2867 static const struct attribute *files[] = {
2868 &ksysfs_register.attr,
2869 &ksysfs_register_quiet.attr,
2870 &ksysfs_pendings_cleanup.attr,
2874 check_module_parameters();
2876 mutex_init(&bch_register_lock);
2877 init_waitqueue_head(&unregister_wait);
2878 register_reboot_notifier(&reboot);
2880 bcache_major = register_blkdev(0, "bcache");
2881 if (bcache_major < 0) {
2882 unregister_reboot_notifier(&reboot);
2883 mutex_destroy(&bch_register_lock);
2884 return bcache_major;
2887 bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2891 bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2892 if (!bch_journal_wq)
2895 bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2899 if (bch_request_init() ||
2900 sysfs_create_files(bcache_kobj, files))
2904 closure_debug_init();
2906 bcache_is_reboot = false;
2917 module_exit(bcache_exit);
2918 module_init(bcache_init);
2920 module_param(bch_cutoff_writeback, uint, 0);
2921 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2923 module_param(bch_cutoff_writeback_sync, uint, 0);
2924 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2926 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2927 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2928 MODULE_LICENSE("GPL");