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/pagemap.h>
20 #include <linux/debugfs.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_flush_wq;
53 struct workqueue_struct *bch_journal_wq;
56 #define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
57 /* limitation of partitions number on single bcache device */
58 #define BCACHE_MINORS 128
59 /* limitation of bcache devices number on single system */
60 #define BCACHE_DEVICE_IDX_MAX ((1U << MINORBITS)/BCACHE_MINORS)
64 static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
66 unsigned int bucket_size = le16_to_cpu(s->bucket_size);
68 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
69 if (bch_has_feature_large_bucket(sb)) {
70 unsigned int max, order;
72 max = sizeof(unsigned int) * BITS_PER_BYTE - 1;
73 order = le16_to_cpu(s->bucket_size);
75 * bcache tool will make sure the overflow won't
76 * happen, an error message here is enough.
79 pr_err("Bucket size (1 << %u) overflows\n",
81 bucket_size = 1 << order;
82 } else if (bch_has_feature_obso_large_bucket(sb)) {
84 le16_to_cpu(s->obso_bucket_size_hi) << 16;
91 static const char *read_super_common(struct cache_sb *sb, struct block_device *bdev,
92 struct cache_sb_disk *s)
97 sb->first_bucket= le16_to_cpu(s->first_bucket);
98 sb->nbuckets = le64_to_cpu(s->nbuckets);
99 sb->bucket_size = get_bucket_size(sb, s);
101 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
102 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
104 err = "Too many journal buckets";
105 if (sb->keys > SB_JOURNAL_BUCKETS)
108 err = "Too many buckets";
109 if (sb->nbuckets > LONG_MAX)
112 err = "Not enough buckets";
113 if (sb->nbuckets < 1 << 7)
116 err = "Bad block size (not power of 2)";
117 if (!is_power_of_2(sb->block_size))
120 err = "Bad block size (larger than page size)";
121 if (sb->block_size > PAGE_SECTORS)
124 err = "Bad bucket size (not power of 2)";
125 if (!is_power_of_2(sb->bucket_size))
128 err = "Bad bucket size (smaller than page size)";
129 if (sb->bucket_size < PAGE_SECTORS)
132 err = "Invalid superblock: device too small";
133 if (get_capacity(bdev->bd_disk) <
134 sb->bucket_size * sb->nbuckets)
138 if (bch_is_zero(sb->set_uuid, 16))
141 err = "Bad cache device number in set";
142 if (!sb->nr_in_set ||
143 sb->nr_in_set <= sb->nr_this_dev ||
144 sb->nr_in_set > MAX_CACHES_PER_SET)
147 err = "Journal buckets not sequential";
148 for (i = 0; i < sb->keys; i++)
149 if (sb->d[i] != sb->first_bucket + i)
152 err = "Too many journal buckets";
153 if (sb->first_bucket + sb->keys > sb->nbuckets)
156 err = "Invalid superblock: first bucket comes before end of super";
157 if (sb->first_bucket * sb->bucket_size < 16)
166 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
167 struct cache_sb_disk **res)
170 struct cache_sb_disk *s;
174 page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
175 SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
178 s = page_address(page) + offset_in_page(SB_OFFSET);
180 sb->offset = le64_to_cpu(s->offset);
181 sb->version = le64_to_cpu(s->version);
183 memcpy(sb->magic, s->magic, 16);
184 memcpy(sb->uuid, s->uuid, 16);
185 memcpy(sb->set_uuid, s->set_uuid, 16);
186 memcpy(sb->label, s->label, SB_LABEL_SIZE);
188 sb->flags = le64_to_cpu(s->flags);
189 sb->seq = le64_to_cpu(s->seq);
190 sb->last_mount = le32_to_cpu(s->last_mount);
191 sb->keys = le16_to_cpu(s->keys);
193 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
194 sb->d[i] = le64_to_cpu(s->d[i]);
196 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
197 sb->version, sb->flags, sb->seq, sb->keys);
199 err = "Not a bcache superblock (bad offset)";
200 if (sb->offset != SB_SECTOR)
203 err = "Not a bcache superblock (bad magic)";
204 if (memcmp(sb->magic, bcache_magic, 16))
207 err = "Bad checksum";
208 if (s->csum != csum_set(s))
212 if (bch_is_zero(sb->uuid, 16))
215 sb->block_size = le16_to_cpu(s->block_size);
217 err = "Superblock block size smaller than device block size";
218 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
221 switch (sb->version) {
222 case BCACHE_SB_VERSION_BDEV:
223 sb->data_offset = BDEV_DATA_START_DEFAULT;
225 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
226 case BCACHE_SB_VERSION_BDEV_WITH_FEATURES:
227 sb->data_offset = le64_to_cpu(s->data_offset);
229 err = "Bad data offset";
230 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
234 case BCACHE_SB_VERSION_CDEV:
235 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
236 err = read_super_common(sb, bdev, s);
240 case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
242 * Feature bits are needed in read_super_common(),
243 * convert them firstly.
245 sb->feature_compat = le64_to_cpu(s->feature_compat);
246 sb->feature_incompat = le64_to_cpu(s->feature_incompat);
247 sb->feature_ro_compat = le64_to_cpu(s->feature_ro_compat);
249 /* Check incompatible features */
250 err = "Unsupported compatible feature found";
251 if (bch_has_unknown_compat_features(sb))
254 err = "Unsupported read-only compatible feature found";
255 if (bch_has_unknown_ro_compat_features(sb))
258 err = "Unsupported incompatible feature found";
259 if (bch_has_unknown_incompat_features(sb))
262 err = read_super_common(sb, bdev, s);
267 err = "Unsupported superblock version";
271 sb->last_mount = (u32)ktime_get_real_seconds();
279 static void write_bdev_super_endio(struct bio *bio)
281 struct cached_dev *dc = bio->bi_private;
284 bch_count_backing_io_errors(dc, bio);
286 closure_put(&dc->sb_write);
289 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
294 bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
295 bio->bi_iter.bi_sector = SB_SECTOR;
296 __bio_add_page(bio, virt_to_page(out), SB_SIZE,
297 offset_in_page(out));
299 out->offset = cpu_to_le64(sb->offset);
301 memcpy(out->uuid, sb->uuid, 16);
302 memcpy(out->set_uuid, sb->set_uuid, 16);
303 memcpy(out->label, sb->label, SB_LABEL_SIZE);
305 out->flags = cpu_to_le64(sb->flags);
306 out->seq = cpu_to_le64(sb->seq);
308 out->last_mount = cpu_to_le32(sb->last_mount);
309 out->first_bucket = cpu_to_le16(sb->first_bucket);
310 out->keys = cpu_to_le16(sb->keys);
312 for (i = 0; i < sb->keys; i++)
313 out->d[i] = cpu_to_le64(sb->d[i]);
315 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
316 out->feature_compat = cpu_to_le64(sb->feature_compat);
317 out->feature_incompat = cpu_to_le64(sb->feature_incompat);
318 out->feature_ro_compat = cpu_to_le64(sb->feature_ro_compat);
321 out->version = cpu_to_le64(sb->version);
322 out->csum = csum_set(out);
324 pr_debug("ver %llu, flags %llu, seq %llu\n",
325 sb->version, sb->flags, sb->seq);
330 static void bch_write_bdev_super_unlock(struct closure *cl)
332 struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
334 up(&dc->sb_write_mutex);
337 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
339 struct closure *cl = &dc->sb_write;
340 struct bio *bio = &dc->sb_bio;
342 down(&dc->sb_write_mutex);
343 closure_init(cl, parent);
345 bio_init(bio, dc->bdev, dc->sb_bv, 1, 0);
346 bio->bi_end_io = write_bdev_super_endio;
347 bio->bi_private = dc;
350 /* I/O request sent to backing device */
351 __write_super(&dc->sb, dc->sb_disk, bio);
353 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
356 static void write_super_endio(struct bio *bio)
358 struct cache *ca = bio->bi_private;
361 bch_count_io_errors(ca, bio->bi_status, 0,
362 "writing superblock");
363 closure_put(&ca->set->sb_write);
366 static void bcache_write_super_unlock(struct closure *cl)
368 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
370 up(&c->sb_write_mutex);
373 void bcache_write_super(struct cache_set *c)
375 struct closure *cl = &c->sb_write;
376 struct cache *ca = c->cache;
377 struct bio *bio = &ca->sb_bio;
378 unsigned int version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
380 down(&c->sb_write_mutex);
381 closure_init(cl, &c->cl);
385 if (ca->sb.version < version)
386 ca->sb.version = version;
388 bio_init(bio, ca->bdev, ca->sb_bv, 1, 0);
389 bio->bi_end_io = write_super_endio;
390 bio->bi_private = ca;
393 __write_super(&ca->sb, ca->sb_disk, bio);
395 closure_return_with_destructor(cl, bcache_write_super_unlock);
400 static void uuid_endio(struct bio *bio)
402 struct closure *cl = bio->bi_private;
403 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
405 cache_set_err_on(bio->bi_status, c, "accessing uuids");
406 bch_bbio_free(bio, c);
410 static void uuid_io_unlock(struct closure *cl)
412 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
414 up(&c->uuid_write_mutex);
417 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
418 struct bkey *k, struct closure *parent)
420 struct closure *cl = &c->uuid_write;
421 struct uuid_entry *u;
426 down(&c->uuid_write_mutex);
427 closure_init(cl, parent);
429 for (i = 0; i < KEY_PTRS(k); i++) {
430 struct bio *bio = bch_bbio_alloc(c);
432 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
433 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
435 bio->bi_end_io = uuid_endio;
436 bio->bi_private = cl;
437 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
438 bch_bio_map(bio, c->uuids);
440 bch_submit_bbio(bio, c, k, i);
442 if (op != REQ_OP_WRITE)
446 bch_extent_to_text(buf, sizeof(buf), k);
447 pr_debug("%s UUIDs at %s\n", op == REQ_OP_WRITE ? "wrote" : "read", buf);
449 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
450 if (!bch_is_zero(u->uuid, 16))
451 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
452 u - c->uuids, u->uuid, u->label,
453 u->first_reg, u->last_reg, u->invalidated);
455 closure_return_with_destructor(cl, uuid_io_unlock);
458 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
460 struct bkey *k = &j->uuid_bucket;
462 if (__bch_btree_ptr_invalid(c, k))
463 return "bad uuid pointer";
465 bkey_copy(&c->uuid_bucket, k);
466 uuid_io(c, REQ_OP_READ, 0, k, cl);
468 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
469 struct uuid_entry_v0 *u0 = (void *) c->uuids;
470 struct uuid_entry *u1 = (void *) c->uuids;
476 * Since the new uuid entry is bigger than the old, we have to
477 * convert starting at the highest memory address and work down
478 * in order to do it in place
481 for (i = c->nr_uuids - 1;
484 memcpy(u1[i].uuid, u0[i].uuid, 16);
485 memcpy(u1[i].label, u0[i].label, 32);
487 u1[i].first_reg = u0[i].first_reg;
488 u1[i].last_reg = u0[i].last_reg;
489 u1[i].invalidated = u0[i].invalidated;
499 static int __uuid_write(struct cache_set *c)
503 struct cache *ca = c->cache;
506 closure_init_stack(&cl);
507 lockdep_assert_held(&bch_register_lock);
509 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
512 size = meta_bucket_pages(&ca->sb) * PAGE_SECTORS;
513 SET_KEY_SIZE(&k.key, size);
514 uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
517 /* Only one bucket used for uuid write */
518 atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
520 bkey_copy(&c->uuid_bucket, &k.key);
525 int bch_uuid_write(struct cache_set *c)
527 int ret = __uuid_write(c);
530 bch_journal_meta(c, NULL);
535 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
537 struct uuid_entry *u;
540 u < c->uuids + c->nr_uuids; u++)
541 if (!memcmp(u->uuid, uuid, 16))
547 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
549 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
551 return uuid_find(c, zero_uuid);
555 * Bucket priorities/gens:
557 * For each bucket, we store on disk its
561 * See alloc.c for an explanation of the gen. The priority is used to implement
562 * lru (and in the future other) cache replacement policies; for most purposes
563 * it's just an opaque integer.
565 * The gens and the priorities don't have a whole lot to do with each other, and
566 * it's actually the gens that must be written out at specific times - it's no
567 * big deal if the priorities don't get written, if we lose them we just reuse
568 * buckets in suboptimal order.
570 * On disk they're stored in a packed array, and in as many buckets are required
571 * to fit them all. The buckets we use to store them form a list; the journal
572 * header points to the first bucket, the first bucket points to the second
575 * This code is used by the allocation code; periodically (whenever it runs out
576 * of buckets to allocate from) the allocation code will invalidate some
577 * buckets, but it can't use those buckets until their new gens are safely on
581 static void prio_endio(struct bio *bio)
583 struct cache *ca = bio->bi_private;
585 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
586 bch_bbio_free(bio, ca->set);
587 closure_put(&ca->prio);
590 static void prio_io(struct cache *ca, uint64_t bucket, int op,
591 unsigned long op_flags)
593 struct closure *cl = &ca->prio;
594 struct bio *bio = bch_bbio_alloc(ca->set);
596 closure_init_stack(cl);
598 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
599 bio_set_dev(bio, ca->bdev);
600 bio->bi_iter.bi_size = meta_bucket_bytes(&ca->sb);
602 bio->bi_end_io = prio_endio;
603 bio->bi_private = ca;
604 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
605 bch_bio_map(bio, ca->disk_buckets);
607 closure_bio_submit(ca->set, bio, &ca->prio);
611 int bch_prio_write(struct cache *ca, bool wait)
617 pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
618 fifo_used(&ca->free[RESERVE_PRIO]),
619 fifo_used(&ca->free[RESERVE_NONE]),
620 fifo_used(&ca->free_inc));
623 * Pre-check if there are enough free buckets. In the non-blocking
624 * scenario it's better to fail early rather than starting to allocate
625 * buckets and do a cleanup later in case of failure.
628 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
629 fifo_used(&ca->free[RESERVE_NONE]);
630 if (prio_buckets(ca) > avail)
634 closure_init_stack(&cl);
636 lockdep_assert_held(&ca->set->bucket_lock);
638 ca->disk_buckets->seq++;
640 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
641 &ca->meta_sectors_written);
643 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
645 struct prio_set *p = ca->disk_buckets;
646 struct bucket_disk *d = p->data;
647 struct bucket_disk *end = d + prios_per_bucket(ca);
649 for (b = ca->buckets + i * prios_per_bucket(ca);
650 b < ca->buckets + ca->sb.nbuckets && d < end;
652 d->prio = cpu_to_le16(b->prio);
656 p->next_bucket = ca->prio_buckets[i + 1];
657 p->magic = pset_magic(&ca->sb);
658 p->csum = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
660 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
661 BUG_ON(bucket == -1);
663 mutex_unlock(&ca->set->bucket_lock);
664 prio_io(ca, bucket, REQ_OP_WRITE, 0);
665 mutex_lock(&ca->set->bucket_lock);
667 ca->prio_buckets[i] = bucket;
668 atomic_dec_bug(&ca->buckets[bucket].pin);
671 mutex_unlock(&ca->set->bucket_lock);
673 bch_journal_meta(ca->set, &cl);
676 mutex_lock(&ca->set->bucket_lock);
679 * Don't want the old priorities to get garbage collected until after we
680 * finish writing the new ones, and they're journalled
682 for (i = 0; i < prio_buckets(ca); i++) {
683 if (ca->prio_last_buckets[i])
684 __bch_bucket_free(ca,
685 &ca->buckets[ca->prio_last_buckets[i]]);
687 ca->prio_last_buckets[i] = ca->prio_buckets[i];
692 static int prio_read(struct cache *ca, uint64_t bucket)
694 struct prio_set *p = ca->disk_buckets;
695 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
697 unsigned int bucket_nr = 0;
700 for (b = ca->buckets;
701 b < ca->buckets + ca->sb.nbuckets;
704 ca->prio_buckets[bucket_nr] = bucket;
705 ca->prio_last_buckets[bucket_nr] = bucket;
708 prio_io(ca, bucket, REQ_OP_READ, 0);
711 bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
712 pr_warn("bad csum reading priorities\n");
716 if (p->magic != pset_magic(&ca->sb)) {
717 pr_warn("bad magic reading priorities\n");
721 bucket = p->next_bucket;
725 b->prio = le16_to_cpu(d->prio);
726 b->gen = b->last_gc = d->gen;
736 static int open_dev(struct block_device *b, fmode_t mode)
738 struct bcache_device *d = b->bd_disk->private_data;
740 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
747 static void release_dev(struct gendisk *b, fmode_t mode)
749 struct bcache_device *d = b->private_data;
754 static int ioctl_dev(struct block_device *b, fmode_t mode,
755 unsigned int cmd, unsigned long arg)
757 struct bcache_device *d = b->bd_disk->private_data;
759 return d->ioctl(d, mode, cmd, arg);
762 static const struct block_device_operations bcache_cached_ops = {
763 .submit_bio = cached_dev_submit_bio,
765 .release = release_dev,
767 .owner = THIS_MODULE,
770 static const struct block_device_operations bcache_flash_ops = {
771 .submit_bio = flash_dev_submit_bio,
773 .release = release_dev,
775 .owner = THIS_MODULE,
778 void bcache_device_stop(struct bcache_device *d)
780 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
783 * - cached device: cached_dev_flush()
784 * - flash dev: flash_dev_flush()
786 closure_queue(&d->cl);
789 static void bcache_device_unlink(struct bcache_device *d)
791 lockdep_assert_held(&bch_register_lock);
793 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
794 struct cache *ca = d->c->cache;
796 sysfs_remove_link(&d->c->kobj, d->name);
797 sysfs_remove_link(&d->kobj, "cache");
799 bd_unlink_disk_holder(ca->bdev, d->disk);
803 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
806 struct cache *ca = c->cache;
809 bd_link_disk_holder(ca->bdev, d->disk);
811 snprintf(d->name, BCACHEDEVNAME_SIZE,
812 "%s%u", name, d->id);
814 ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
816 pr_err("Couldn't create device -> cache set symlink\n");
818 ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
820 pr_err("Couldn't create cache set -> device symlink\n");
822 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
825 static void bcache_device_detach(struct bcache_device *d)
827 lockdep_assert_held(&bch_register_lock);
829 atomic_dec(&d->c->attached_dev_nr);
831 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
832 struct uuid_entry *u = d->c->uuids + d->id;
834 SET_UUID_FLASH_ONLY(u, 0);
835 memcpy(u->uuid, invalid_uuid, 16);
836 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
837 bch_uuid_write(d->c);
840 bcache_device_unlink(d);
842 d->c->devices[d->id] = NULL;
843 closure_put(&d->c->caching);
847 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
854 if (id >= c->devices_max_used)
855 c->devices_max_used = id + 1;
857 closure_get(&c->caching);
860 static inline int first_minor_to_idx(int first_minor)
862 return (first_minor/BCACHE_MINORS);
865 static inline int idx_to_first_minor(int idx)
867 return (idx * BCACHE_MINORS);
870 static void bcache_device_free(struct bcache_device *d)
872 struct gendisk *disk = d->disk;
874 lockdep_assert_held(&bch_register_lock);
877 pr_info("%s stopped\n", disk->disk_name);
879 pr_err("bcache device (NULL gendisk) stopped\n");
882 bcache_device_detach(d);
885 ida_simple_remove(&bcache_device_idx,
886 first_minor_to_idx(disk->first_minor));
887 blk_cleanup_disk(disk);
890 bioset_exit(&d->bio_split);
891 kvfree(d->full_dirty_stripes);
892 kvfree(d->stripe_sectors_dirty);
894 closure_debug_destroy(&d->cl);
897 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
898 sector_t sectors, struct block_device *cached_bdev,
899 const struct block_device_operations *ops)
901 struct request_queue *q;
902 const size_t max_stripes = min_t(size_t, INT_MAX,
903 SIZE_MAX / sizeof(atomic_t));
908 d->stripe_size = 1 << 31;
910 n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
911 if (!n || n > max_stripes) {
912 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
918 n = d->nr_stripes * sizeof(atomic_t);
919 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
920 if (!d->stripe_sectors_dirty)
923 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
924 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
925 if (!d->full_dirty_stripes)
926 goto out_free_stripe_sectors_dirty;
928 idx = ida_simple_get(&bcache_device_idx, 0,
929 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
931 goto out_free_full_dirty_stripes;
933 if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
934 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
937 d->disk = blk_alloc_disk(NUMA_NO_NODE);
939 goto out_bioset_exit;
941 set_capacity(d->disk, sectors);
942 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
944 d->disk->major = bcache_major;
945 d->disk->first_minor = idx_to_first_minor(idx);
946 d->disk->minors = BCACHE_MINORS;
948 d->disk->private_data = d;
951 q->limits.max_hw_sectors = UINT_MAX;
952 q->limits.max_sectors = UINT_MAX;
953 q->limits.max_segment_size = UINT_MAX;
954 q->limits.max_segments = BIO_MAX_VECS;
955 blk_queue_max_discard_sectors(q, UINT_MAX);
956 q->limits.discard_granularity = 512;
957 q->limits.io_min = block_size;
958 q->limits.logical_block_size = block_size;
959 q->limits.physical_block_size = block_size;
961 if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
963 * This should only happen with BCACHE_SB_VERSION_BDEV.
964 * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
966 pr_info("%s: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
967 d->disk->disk_name, q->limits.logical_block_size,
968 PAGE_SIZE, bdev_logical_block_size(cached_bdev));
970 /* This also adjusts physical block size/min io size if needed */
971 blk_queue_logical_block_size(q, bdev_logical_block_size(cached_bdev));
974 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
975 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
977 blk_queue_write_cache(q, true, true);
982 bioset_exit(&d->bio_split);
984 ida_simple_remove(&bcache_device_idx, idx);
985 out_free_full_dirty_stripes:
986 kvfree(d->full_dirty_stripes);
987 out_free_stripe_sectors_dirty:
988 kvfree(d->stripe_sectors_dirty);
995 static void calc_cached_dev_sectors(struct cache_set *c)
997 uint64_t sectors = 0;
998 struct cached_dev *dc;
1000 list_for_each_entry(dc, &c->cached_devs, list)
1001 sectors += bdev_nr_sectors(dc->bdev);
1003 c->cached_dev_sectors = sectors;
1006 #define BACKING_DEV_OFFLINE_TIMEOUT 5
1007 static int cached_dev_status_update(void *arg)
1009 struct cached_dev *dc = arg;
1010 struct request_queue *q;
1013 * If this delayed worker is stopping outside, directly quit here.
1014 * dc->io_disable might be set via sysfs interface, so check it
1017 while (!kthread_should_stop() && !dc->io_disable) {
1018 q = bdev_get_queue(dc->bdev);
1019 if (blk_queue_dying(q))
1020 dc->offline_seconds++;
1022 dc->offline_seconds = 0;
1024 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1025 pr_err("%pg: device offline for %d seconds\n",
1027 BACKING_DEV_OFFLINE_TIMEOUT);
1028 pr_err("%s: disable I/O request due to backing device offline\n",
1030 dc->io_disable = true;
1031 /* let others know earlier that io_disable is true */
1033 bcache_device_stop(&dc->disk);
1036 schedule_timeout_interruptible(HZ);
1039 wait_for_kthread_stop();
1044 int bch_cached_dev_run(struct cached_dev *dc)
1047 struct bcache_device *d = &dc->disk;
1048 char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1051 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1052 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1056 if (dc->io_disable) {
1057 pr_err("I/O disabled on cached dev %pg\n", dc->bdev);
1062 if (atomic_xchg(&dc->running, 1)) {
1063 pr_info("cached dev %pg is running already\n", dc->bdev);
1069 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1072 closure_init_stack(&cl);
1074 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1075 bch_write_bdev_super(dc, &cl);
1079 ret = add_disk(d->disk);
1082 bd_link_disk_holder(dc->bdev, dc->disk.disk);
1084 * won't show up in the uevent file, use udevadm monitor -e instead
1085 * only class / kset properties are persistent
1087 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1089 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1090 sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1091 &d->kobj, "bcache")) {
1092 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1097 dc->status_update_thread = kthread_run(cached_dev_status_update,
1098 dc, "bcache_status_update");
1099 if (IS_ERR(dc->status_update_thread)) {
1100 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1111 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1112 * work dc->writeback_rate_update is running. Wait until the routine
1113 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1114 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1115 * seconds, give up waiting here and continue to cancel it too.
1117 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1119 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1122 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1126 schedule_timeout_interruptible(1);
1127 } while (time_out > 0);
1130 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1132 cancel_delayed_work_sync(&dc->writeback_rate_update);
1135 static void cached_dev_detach_finish(struct work_struct *w)
1137 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1138 struct cache_set *c = dc->disk.c;
1140 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1141 BUG_ON(refcount_read(&dc->count));
1144 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1145 cancel_writeback_rate_update_dwork(dc);
1147 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1148 kthread_stop(dc->writeback_thread);
1149 dc->writeback_thread = NULL;
1152 mutex_lock(&bch_register_lock);
1154 bcache_device_detach(&dc->disk);
1155 list_move(&dc->list, &uncached_devices);
1156 calc_cached_dev_sectors(c);
1158 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1159 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1161 mutex_unlock(&bch_register_lock);
1163 pr_info("Caching disabled for %pg\n", dc->bdev);
1165 /* Drop ref we took in cached_dev_detach() */
1166 closure_put(&dc->disk.cl);
1169 void bch_cached_dev_detach(struct cached_dev *dc)
1171 lockdep_assert_held(&bch_register_lock);
1173 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1176 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1180 * Block the device from being closed and freed until we're finished
1183 closure_get(&dc->disk.cl);
1185 bch_writeback_queue(dc);
1190 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1193 uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1194 struct uuid_entry *u;
1195 struct cached_dev *exist_dc, *t;
1198 if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1199 (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1203 pr_err("Can't attach %pg: already attached\n", dc->bdev);
1207 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1208 pr_err("Can't attach %pg: shutting down\n", dc->bdev);
1212 if (dc->sb.block_size < c->cache->sb.block_size) {
1214 pr_err("Couldn't attach %pg: block size less than set's block size\n",
1219 /* Check whether already attached */
1220 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1221 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1222 pr_err("Tried to attach %pg but duplicate UUID already attached\n",
1229 u = uuid_find(c, dc->sb.uuid);
1232 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1233 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1234 memcpy(u->uuid, invalid_uuid, 16);
1235 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1240 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1241 pr_err("Couldn't find uuid for %pg in set\n", dc->bdev);
1245 u = uuid_find_empty(c);
1247 pr_err("Not caching %pg, no room for UUID\n", dc->bdev);
1253 * Deadlocks since we're called via sysfs...
1254 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1257 if (bch_is_zero(u->uuid, 16)) {
1260 closure_init_stack(&cl);
1262 memcpy(u->uuid, dc->sb.uuid, 16);
1263 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1264 u->first_reg = u->last_reg = rtime;
1267 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1268 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1270 bch_write_bdev_super(dc, &cl);
1273 u->last_reg = rtime;
1277 bcache_device_attach(&dc->disk, c, u - c->uuids);
1278 list_move(&dc->list, &c->cached_devs);
1279 calc_cached_dev_sectors(c);
1282 * dc->c must be set before dc->count != 0 - paired with the mb in
1286 refcount_set(&dc->count, 1);
1288 /* Block writeback thread, but spawn it */
1289 down_write(&dc->writeback_lock);
1290 if (bch_cached_dev_writeback_start(dc)) {
1291 up_write(&dc->writeback_lock);
1292 pr_err("Couldn't start writeback facilities for %s\n",
1293 dc->disk.disk->disk_name);
1297 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1298 atomic_set(&dc->has_dirty, 1);
1299 bch_writeback_queue(dc);
1302 bch_sectors_dirty_init(&dc->disk);
1304 ret = bch_cached_dev_run(dc);
1305 if (ret && (ret != -EBUSY)) {
1306 up_write(&dc->writeback_lock);
1308 * bch_register_lock is held, bcache_device_stop() is not
1309 * able to be directly called. The kthread and kworker
1310 * created previously in bch_cached_dev_writeback_start()
1311 * have to be stopped manually here.
1313 kthread_stop(dc->writeback_thread);
1314 cancel_writeback_rate_update_dwork(dc);
1315 pr_err("Couldn't run cached device %pg\n", dc->bdev);
1319 bcache_device_link(&dc->disk, c, "bdev");
1320 atomic_inc(&c->attached_dev_nr);
1322 if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1323 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1324 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1325 set_disk_ro(dc->disk.disk, 1);
1328 /* Allow the writeback thread to proceed */
1329 up_write(&dc->writeback_lock);
1331 pr_info("Caching %pg as %s on set %pU\n",
1333 dc->disk.disk->disk_name,
1334 dc->disk.c->set_uuid);
1338 /* when dc->disk.kobj released */
1339 void bch_cached_dev_release(struct kobject *kobj)
1341 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1344 module_put(THIS_MODULE);
1347 static void cached_dev_free(struct closure *cl)
1349 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1351 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1352 cancel_writeback_rate_update_dwork(dc);
1354 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1355 kthread_stop(dc->writeback_thread);
1356 if (!IS_ERR_OR_NULL(dc->status_update_thread))
1357 kthread_stop(dc->status_update_thread);
1359 mutex_lock(&bch_register_lock);
1361 if (atomic_read(&dc->running)) {
1362 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1363 del_gendisk(dc->disk.disk);
1365 bcache_device_free(&dc->disk);
1366 list_del(&dc->list);
1368 mutex_unlock(&bch_register_lock);
1371 put_page(virt_to_page(dc->sb_disk));
1373 if (!IS_ERR_OR_NULL(dc->bdev))
1374 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1376 wake_up(&unregister_wait);
1378 kobject_put(&dc->disk.kobj);
1381 static void cached_dev_flush(struct closure *cl)
1383 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1384 struct bcache_device *d = &dc->disk;
1386 mutex_lock(&bch_register_lock);
1387 bcache_device_unlink(d);
1388 mutex_unlock(&bch_register_lock);
1390 bch_cache_accounting_destroy(&dc->accounting);
1391 kobject_del(&d->kobj);
1393 continue_at(cl, cached_dev_free, system_wq);
1396 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1400 struct request_queue *q = bdev_get_queue(dc->bdev);
1402 __module_get(THIS_MODULE);
1403 INIT_LIST_HEAD(&dc->list);
1404 closure_init(&dc->disk.cl, NULL);
1405 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1406 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1407 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1408 sema_init(&dc->sb_write_mutex, 1);
1409 INIT_LIST_HEAD(&dc->io_lru);
1410 spin_lock_init(&dc->io_lock);
1411 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1413 dc->sequential_cutoff = 4 << 20;
1415 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1416 list_add(&io->lru, &dc->io_lru);
1417 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1420 dc->disk.stripe_size = q->limits.io_opt >> 9;
1422 if (dc->disk.stripe_size)
1423 dc->partial_stripes_expensive =
1424 q->limits.raid_partial_stripes_expensive;
1426 ret = bcache_device_init(&dc->disk, block_size,
1427 bdev_nr_sectors(dc->bdev) - dc->sb.data_offset,
1428 dc->bdev, &bcache_cached_ops);
1432 blk_queue_io_opt(dc->disk.disk->queue,
1433 max(queue_io_opt(dc->disk.disk->queue), queue_io_opt(q)));
1435 atomic_set(&dc->io_errors, 0);
1436 dc->io_disable = false;
1437 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1438 /* default to auto */
1439 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1441 bch_cached_dev_request_init(dc);
1442 bch_cached_dev_writeback_init(dc);
1446 /* Cached device - bcache superblock */
1448 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1449 struct block_device *bdev,
1450 struct cached_dev *dc)
1452 const char *err = "cannot allocate memory";
1453 struct cache_set *c;
1456 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1458 dc->bdev->bd_holder = dc;
1459 dc->sb_disk = sb_disk;
1461 if (cached_dev_init(dc, sb->block_size << 9))
1464 err = "error creating kobject";
1465 if (kobject_add(&dc->disk.kobj, bdev_kobj(bdev), "bcache"))
1467 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1470 pr_info("registered backing device %pg\n", dc->bdev);
1472 list_add(&dc->list, &uncached_devices);
1473 /* attach to a matched cache set if it exists */
1474 list_for_each_entry(c, &bch_cache_sets, list)
1475 bch_cached_dev_attach(dc, c, NULL);
1477 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1478 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1479 err = "failed to run cached device";
1480 ret = bch_cached_dev_run(dc);
1487 pr_notice("error %pg: %s\n", dc->bdev, err);
1488 bcache_device_stop(&dc->disk);
1492 /* Flash only volumes */
1494 /* When d->kobj released */
1495 void bch_flash_dev_release(struct kobject *kobj)
1497 struct bcache_device *d = container_of(kobj, struct bcache_device,
1502 static void flash_dev_free(struct closure *cl)
1504 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1506 mutex_lock(&bch_register_lock);
1507 atomic_long_sub(bcache_dev_sectors_dirty(d),
1508 &d->c->flash_dev_dirty_sectors);
1509 del_gendisk(d->disk);
1510 bcache_device_free(d);
1511 mutex_unlock(&bch_register_lock);
1512 kobject_put(&d->kobj);
1515 static void flash_dev_flush(struct closure *cl)
1517 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1519 mutex_lock(&bch_register_lock);
1520 bcache_device_unlink(d);
1521 mutex_unlock(&bch_register_lock);
1522 kobject_del(&d->kobj);
1523 continue_at(cl, flash_dev_free, system_wq);
1526 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1529 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1534 closure_init(&d->cl, NULL);
1535 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1537 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1539 if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1540 NULL, &bcache_flash_ops))
1543 bcache_device_attach(d, c, u - c->uuids);
1544 bch_sectors_dirty_init(d);
1545 bch_flash_dev_request_init(d);
1546 err = add_disk(d->disk);
1550 err = kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache");
1554 bcache_device_link(d, c, "volume");
1556 if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1557 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1558 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1559 set_disk_ro(d->disk, 1);
1564 kobject_put(&d->kobj);
1569 static int flash_devs_run(struct cache_set *c)
1572 struct uuid_entry *u;
1575 u < c->uuids + c->nr_uuids && !ret;
1577 if (UUID_FLASH_ONLY(u))
1578 ret = flash_dev_run(c, u);
1583 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1585 struct uuid_entry *u;
1587 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1590 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1593 u = uuid_find_empty(c);
1595 pr_err("Can't create volume, no room for UUID\n");
1599 get_random_bytes(u->uuid, 16);
1600 memset(u->label, 0, 32);
1601 u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1603 SET_UUID_FLASH_ONLY(u, 1);
1604 u->sectors = size >> 9;
1608 return flash_dev_run(c, u);
1611 bool bch_cached_dev_error(struct cached_dev *dc)
1613 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1616 dc->io_disable = true;
1617 /* make others know io_disable is true earlier */
1620 pr_err("stop %s: too many IO errors on backing device %pg\n",
1621 dc->disk.disk->disk_name, dc->bdev);
1623 bcache_device_stop(&dc->disk);
1630 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1632 struct va_format vaf;
1635 if (c->on_error != ON_ERROR_PANIC &&
1636 test_bit(CACHE_SET_STOPPING, &c->flags))
1639 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1640 pr_info("CACHE_SET_IO_DISABLE already set\n");
1643 * XXX: we can be called from atomic context
1644 * acquire_console_sem();
1647 va_start(args, fmt);
1652 pr_err("error on %pU: %pV, disabling caching\n",
1657 if (c->on_error == ON_ERROR_PANIC)
1658 panic("panic forced after error\n");
1660 bch_cache_set_unregister(c);
1664 /* When c->kobj released */
1665 void bch_cache_set_release(struct kobject *kobj)
1667 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1670 module_put(THIS_MODULE);
1673 static void cache_set_free(struct closure *cl)
1675 struct cache_set *c = container_of(cl, struct cache_set, cl);
1678 debugfs_remove(c->debug);
1680 bch_open_buckets_free(c);
1681 bch_btree_cache_free(c);
1682 bch_journal_free(c);
1684 mutex_lock(&bch_register_lock);
1685 bch_bset_sort_state_free(&c->sort);
1686 free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1692 kobject_put(&ca->kobj);
1696 if (c->moving_gc_wq)
1697 destroy_workqueue(c->moving_gc_wq);
1698 bioset_exit(&c->bio_split);
1699 mempool_exit(&c->fill_iter);
1700 mempool_exit(&c->bio_meta);
1701 mempool_exit(&c->search);
1705 mutex_unlock(&bch_register_lock);
1707 pr_info("Cache set %pU unregistered\n", c->set_uuid);
1708 wake_up(&unregister_wait);
1710 closure_debug_destroy(&c->cl);
1711 kobject_put(&c->kobj);
1714 static void cache_set_flush(struct closure *cl)
1716 struct cache_set *c = container_of(cl, struct cache_set, caching);
1717 struct cache *ca = c->cache;
1720 bch_cache_accounting_destroy(&c->accounting);
1722 kobject_put(&c->internal);
1723 kobject_del(&c->kobj);
1725 if (!IS_ERR_OR_NULL(c->gc_thread))
1726 kthread_stop(c->gc_thread);
1728 if (!IS_ERR_OR_NULL(c->root))
1729 list_add(&c->root->list, &c->btree_cache);
1732 * Avoid flushing cached nodes if cache set is retiring
1733 * due to too many I/O errors detected.
1735 if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1736 list_for_each_entry(b, &c->btree_cache, list) {
1737 mutex_lock(&b->write_lock);
1738 if (btree_node_dirty(b))
1739 __bch_btree_node_write(b, NULL);
1740 mutex_unlock(&b->write_lock);
1743 if (ca->alloc_thread)
1744 kthread_stop(ca->alloc_thread);
1746 if (c->journal.cur) {
1747 cancel_delayed_work_sync(&c->journal.work);
1748 /* flush last journal entry if needed */
1749 c->journal.work.work.func(&c->journal.work.work);
1756 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1757 * cache set is unregistering due to too many I/O errors. In this condition,
1758 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1759 * value and whether the broken cache has dirty data:
1761 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1762 * BCH_CACHED_STOP_AUTO 0 NO
1763 * BCH_CACHED_STOP_AUTO 1 YES
1764 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1765 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1767 * The expected behavior is, if stop_when_cache_set_failed is configured to
1768 * "auto" via sysfs interface, the bcache device will not be stopped if the
1769 * backing device is clean on the broken cache device.
1771 static void conditional_stop_bcache_device(struct cache_set *c,
1772 struct bcache_device *d,
1773 struct cached_dev *dc)
1775 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1776 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1777 d->disk->disk_name, c->set_uuid);
1778 bcache_device_stop(d);
1779 } else if (atomic_read(&dc->has_dirty)) {
1781 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1782 * and dc->has_dirty == 1
1784 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1785 d->disk->disk_name);
1787 * There might be a small time gap that cache set is
1788 * released but bcache device is not. Inside this time
1789 * gap, regular I/O requests will directly go into
1790 * backing device as no cache set attached to. This
1791 * behavior may also introduce potential inconsistence
1792 * data in writeback mode while cache is dirty.
1793 * Therefore before calling bcache_device_stop() due
1794 * to a broken cache device, dc->io_disable should be
1795 * explicitly set to true.
1797 dc->io_disable = true;
1798 /* make others know io_disable is true earlier */
1800 bcache_device_stop(d);
1803 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1804 * and dc->has_dirty == 0
1806 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1807 d->disk->disk_name);
1811 static void __cache_set_unregister(struct closure *cl)
1813 struct cache_set *c = container_of(cl, struct cache_set, caching);
1814 struct cached_dev *dc;
1815 struct bcache_device *d;
1818 mutex_lock(&bch_register_lock);
1820 for (i = 0; i < c->devices_max_used; i++) {
1825 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1826 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1827 dc = container_of(d, struct cached_dev, disk);
1828 bch_cached_dev_detach(dc);
1829 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1830 conditional_stop_bcache_device(c, d, dc);
1832 bcache_device_stop(d);
1836 mutex_unlock(&bch_register_lock);
1838 continue_at(cl, cache_set_flush, system_wq);
1841 void bch_cache_set_stop(struct cache_set *c)
1843 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1844 /* closure_fn set to __cache_set_unregister() */
1845 closure_queue(&c->caching);
1848 void bch_cache_set_unregister(struct cache_set *c)
1850 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1851 bch_cache_set_stop(c);
1854 #define alloc_meta_bucket_pages(gfp, sb) \
1855 ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1857 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1860 struct cache *ca = container_of(sb, struct cache, sb);
1861 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1866 __module_get(THIS_MODULE);
1867 closure_init(&c->cl, NULL);
1868 set_closure_fn(&c->cl, cache_set_free, system_wq);
1870 closure_init(&c->caching, &c->cl);
1871 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1873 /* Maybe create continue_at_noreturn() and use it here? */
1874 closure_set_stopped(&c->cl);
1875 closure_put(&c->cl);
1877 kobject_init(&c->kobj, &bch_cache_set_ktype);
1878 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1880 bch_cache_accounting_init(&c->accounting, &c->cl);
1882 memcpy(c->set_uuid, sb->set_uuid, 16);
1886 c->bucket_bits = ilog2(sb->bucket_size);
1887 c->block_bits = ilog2(sb->block_size);
1888 c->nr_uuids = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1889 c->devices_max_used = 0;
1890 atomic_set(&c->attached_dev_nr, 0);
1891 c->btree_pages = meta_bucket_pages(sb);
1892 if (c->btree_pages > BTREE_MAX_PAGES)
1893 c->btree_pages = max_t(int, c->btree_pages / 4,
1896 sema_init(&c->sb_write_mutex, 1);
1897 mutex_init(&c->bucket_lock);
1898 init_waitqueue_head(&c->btree_cache_wait);
1899 spin_lock_init(&c->btree_cannibalize_lock);
1900 init_waitqueue_head(&c->bucket_wait);
1901 init_waitqueue_head(&c->gc_wait);
1902 sema_init(&c->uuid_write_mutex, 1);
1904 spin_lock_init(&c->btree_gc_time.lock);
1905 spin_lock_init(&c->btree_split_time.lock);
1906 spin_lock_init(&c->btree_read_time.lock);
1908 bch_moving_init_cache_set(c);
1910 INIT_LIST_HEAD(&c->list);
1911 INIT_LIST_HEAD(&c->cached_devs);
1912 INIT_LIST_HEAD(&c->btree_cache);
1913 INIT_LIST_HEAD(&c->btree_cache_freeable);
1914 INIT_LIST_HEAD(&c->btree_cache_freed);
1915 INIT_LIST_HEAD(&c->data_buckets);
1917 iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1918 sizeof(struct btree_iter_set);
1920 c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1924 if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1927 if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1928 sizeof(struct bbio) +
1929 sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1932 if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1935 if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1936 BIOSET_NEED_RESCUER))
1939 c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1943 c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1944 if (!c->moving_gc_wq)
1947 if (bch_journal_alloc(c))
1950 if (bch_btree_cache_alloc(c))
1953 if (bch_open_buckets_alloc(c))
1956 if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1959 c->congested_read_threshold_us = 2000;
1960 c->congested_write_threshold_us = 20000;
1961 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1962 c->idle_max_writeback_rate_enabled = 1;
1963 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1967 bch_cache_set_unregister(c);
1971 static int run_cache_set(struct cache_set *c)
1973 const char *err = "cannot allocate memory";
1974 struct cached_dev *dc, *t;
1975 struct cache *ca = c->cache;
1978 struct journal_replay *l;
1980 closure_init_stack(&cl);
1982 c->nbuckets = ca->sb.nbuckets;
1985 if (CACHE_SYNC(&c->cache->sb)) {
1989 err = "cannot allocate memory for journal";
1990 if (bch_journal_read(c, &journal))
1993 pr_debug("btree_journal_read() done\n");
1995 err = "no journal entries found";
1996 if (list_empty(&journal))
1999 j = &list_entry(journal.prev, struct journal_replay, list)->j;
2001 err = "IO error reading priorities";
2002 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
2006 * If prio_read() fails it'll call cache_set_error and we'll
2007 * tear everything down right away, but if we perhaps checked
2008 * sooner we could avoid journal replay.
2013 err = "bad btree root";
2014 if (__bch_btree_ptr_invalid(c, k))
2017 err = "error reading btree root";
2018 c->root = bch_btree_node_get(c, NULL, k,
2021 if (IS_ERR_OR_NULL(c->root))
2024 list_del_init(&c->root->list);
2025 rw_unlock(true, c->root);
2027 err = uuid_read(c, j, &cl);
2031 err = "error in recovery";
2032 if (bch_btree_check(c))
2035 bch_journal_mark(c, &journal);
2036 bch_initial_gc_finish(c);
2037 pr_debug("btree_check() done\n");
2040 * bcache_journal_next() can't happen sooner, or
2041 * btree_gc_finish() will give spurious errors about last_gc >
2042 * gc_gen - this is a hack but oh well.
2044 bch_journal_next(&c->journal);
2046 err = "error starting allocator thread";
2047 if (bch_cache_allocator_start(ca))
2051 * First place it's safe to allocate: btree_check() and
2052 * btree_gc_finish() have to run before we have buckets to
2053 * allocate, and bch_bucket_alloc_set() might cause a journal
2054 * entry to be written so bcache_journal_next() has to be called
2057 * If the uuids were in the old format we have to rewrite them
2058 * before the next journal entry is written:
2060 if (j->version < BCACHE_JSET_VERSION_UUID)
2063 err = "bcache: replay journal failed";
2064 if (bch_journal_replay(c, &journal))
2069 pr_notice("invalidating existing data\n");
2070 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2071 2, SB_JOURNAL_BUCKETS);
2073 for (j = 0; j < ca->sb.keys; j++)
2074 ca->sb.d[j] = ca->sb.first_bucket + j;
2076 bch_initial_gc_finish(c);
2078 err = "error starting allocator thread";
2079 if (bch_cache_allocator_start(ca))
2082 mutex_lock(&c->bucket_lock);
2083 bch_prio_write(ca, true);
2084 mutex_unlock(&c->bucket_lock);
2086 err = "cannot allocate new UUID bucket";
2087 if (__uuid_write(c))
2090 err = "cannot allocate new btree root";
2091 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2092 if (IS_ERR_OR_NULL(c->root))
2095 mutex_lock(&c->root->write_lock);
2096 bkey_copy_key(&c->root->key, &MAX_KEY);
2097 bch_btree_node_write(c->root, &cl);
2098 mutex_unlock(&c->root->write_lock);
2100 bch_btree_set_root(c->root);
2101 rw_unlock(true, c->root);
2104 * We don't want to write the first journal entry until
2105 * everything is set up - fortunately journal entries won't be
2106 * written until the SET_CACHE_SYNC() here:
2108 SET_CACHE_SYNC(&c->cache->sb, true);
2110 bch_journal_next(&c->journal);
2111 bch_journal_meta(c, &cl);
2114 err = "error starting gc thread";
2115 if (bch_gc_thread_start(c))
2119 c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2120 bcache_write_super(c);
2122 if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2123 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2125 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2126 bch_cached_dev_attach(dc, c, NULL);
2130 set_bit(CACHE_SET_RUNNING, &c->flags);
2133 while (!list_empty(&journal)) {
2134 l = list_first_entry(&journal, struct journal_replay, list);
2141 bch_cache_set_error(c, "%s", err);
2146 static const char *register_cache_set(struct cache *ca)
2149 const char *err = "cannot allocate memory";
2150 struct cache_set *c;
2152 list_for_each_entry(c, &bch_cache_sets, list)
2153 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2155 return "duplicate cache set member";
2160 c = bch_cache_set_alloc(&ca->sb);
2164 err = "error creating kobject";
2165 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2166 kobject_add(&c->internal, &c->kobj, "internal"))
2169 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2172 bch_debug_init_cache_set(c);
2174 list_add(&c->list, &bch_cache_sets);
2176 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2177 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2178 sysfs_create_link(&c->kobj, &ca->kobj, buf))
2181 kobject_get(&ca->kobj);
2183 ca->set->cache = ca;
2185 err = "failed to run cache set";
2186 if (run_cache_set(c) < 0)
2191 bch_cache_set_unregister(c);
2197 /* When ca->kobj released */
2198 void bch_cache_release(struct kobject *kobj)
2200 struct cache *ca = container_of(kobj, struct cache, kobj);
2204 BUG_ON(ca->set->cache != ca);
2205 ca->set->cache = NULL;
2208 free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2209 kfree(ca->prio_buckets);
2212 free_heap(&ca->heap);
2213 free_fifo(&ca->free_inc);
2215 for (i = 0; i < RESERVE_NR; i++)
2216 free_fifo(&ca->free[i]);
2219 put_page(virt_to_page(ca->sb_disk));
2221 if (!IS_ERR_OR_NULL(ca->bdev))
2222 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2225 module_put(THIS_MODULE);
2228 static int cache_alloc(struct cache *ca)
2231 size_t btree_buckets;
2234 const char *err = NULL;
2236 __module_get(THIS_MODULE);
2237 kobject_init(&ca->kobj, &bch_cache_ktype);
2239 bio_init(&ca->journal.bio, NULL, ca->journal.bio.bi_inline_vecs, 8, 0);
2242 * when ca->sb.njournal_buckets is not zero, journal exists,
2243 * and in bch_journal_replay(), tree node may split,
2244 * so bucket of RESERVE_BTREE type is needed,
2245 * the worst situation is all journal buckets are valid journal,
2246 * and all the keys need to replay,
2247 * so the number of RESERVE_BTREE type buckets should be as much
2248 * as journal buckets
2250 btree_buckets = ca->sb.njournal_buckets ?: 8;
2251 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2254 err = "ca->sb.nbuckets is too small";
2258 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2260 err = "ca->free[RESERVE_BTREE] alloc failed";
2261 goto err_btree_alloc;
2264 if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2266 err = "ca->free[RESERVE_PRIO] alloc failed";
2267 goto err_prio_alloc;
2270 if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2271 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2272 goto err_movinggc_alloc;
2275 if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2276 err = "ca->free[RESERVE_NONE] alloc failed";
2277 goto err_none_alloc;
2280 if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2281 err = "ca->free_inc alloc failed";
2282 goto err_free_inc_alloc;
2285 if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2286 err = "ca->heap alloc failed";
2287 goto err_heap_alloc;
2290 ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2293 err = "ca->buckets alloc failed";
2294 goto err_buckets_alloc;
2297 ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2298 prio_buckets(ca), 2),
2300 if (!ca->prio_buckets) {
2301 err = "ca->prio_buckets alloc failed";
2302 goto err_prio_buckets_alloc;
2305 ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2306 if (!ca->disk_buckets) {
2307 err = "ca->disk_buckets alloc failed";
2308 goto err_disk_buckets_alloc;
2311 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2313 for_each_bucket(b, ca)
2314 atomic_set(&b->pin, 0);
2317 err_disk_buckets_alloc:
2318 kfree(ca->prio_buckets);
2319 err_prio_buckets_alloc:
2322 free_heap(&ca->heap);
2324 free_fifo(&ca->free_inc);
2326 free_fifo(&ca->free[RESERVE_NONE]);
2328 free_fifo(&ca->free[RESERVE_MOVINGGC]);
2330 free_fifo(&ca->free[RESERVE_PRIO]);
2332 free_fifo(&ca->free[RESERVE_BTREE]);
2335 module_put(THIS_MODULE);
2337 pr_notice("error %pg: %s\n", ca->bdev, err);
2341 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2342 struct block_device *bdev, struct cache *ca)
2344 const char *err = NULL; /* must be set for any error case */
2347 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2349 ca->bdev->bd_holder = ca;
2350 ca->sb_disk = sb_disk;
2352 if (bdev_max_discard_sectors((bdev)))
2353 ca->discard = CACHE_DISCARD(&ca->sb);
2355 ret = cache_alloc(ca);
2358 * If we failed here, it means ca->kobj is not initialized yet,
2359 * kobject_put() won't be called and there is no chance to
2360 * call blkdev_put() to bdev in bch_cache_release(). So we
2361 * explicitly call blkdev_put() here.
2363 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2365 err = "cache_alloc(): -ENOMEM";
2366 else if (ret == -EPERM)
2367 err = "cache_alloc(): cache device is too small";
2369 err = "cache_alloc(): unknown error";
2373 if (kobject_add(&ca->kobj, bdev_kobj(bdev), "bcache")) {
2374 err = "error calling kobject_add";
2379 mutex_lock(&bch_register_lock);
2380 err = register_cache_set(ca);
2381 mutex_unlock(&bch_register_lock);
2388 pr_info("registered cache device %pg\n", ca->bdev);
2391 kobject_put(&ca->kobj);
2395 pr_notice("error %pg: %s\n", ca->bdev, err);
2400 /* Global interfaces/init */
2402 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2403 const char *buffer, size_t size);
2404 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2405 struct kobj_attribute *attr,
2406 const char *buffer, size_t size);
2408 kobj_attribute_write(register, register_bcache);
2409 kobj_attribute_write(register_quiet, register_bcache);
2410 kobj_attribute_write(pendings_cleanup, bch_pending_bdevs_cleanup);
2412 static bool bch_is_open_backing(dev_t dev)
2414 struct cache_set *c, *tc;
2415 struct cached_dev *dc, *t;
2417 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2418 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2419 if (dc->bdev->bd_dev == dev)
2421 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2422 if (dc->bdev->bd_dev == dev)
2427 static bool bch_is_open_cache(dev_t dev)
2429 struct cache_set *c, *tc;
2431 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2432 struct cache *ca = c->cache;
2434 if (ca->bdev->bd_dev == dev)
2441 static bool bch_is_open(dev_t dev)
2443 return bch_is_open_cache(dev) || bch_is_open_backing(dev);
2446 struct async_reg_args {
2447 struct delayed_work reg_work;
2449 struct cache_sb *sb;
2450 struct cache_sb_disk *sb_disk;
2451 struct block_device *bdev;
2454 static void register_bdev_worker(struct work_struct *work)
2457 struct async_reg_args *args =
2458 container_of(work, struct async_reg_args, reg_work.work);
2459 struct cached_dev *dc;
2461 dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2464 put_page(virt_to_page(args->sb_disk));
2465 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2469 mutex_lock(&bch_register_lock);
2470 if (register_bdev(args->sb, args->sb_disk, args->bdev, dc) < 0)
2472 mutex_unlock(&bch_register_lock);
2476 pr_info("error %s: fail to register backing device\n",
2481 module_put(THIS_MODULE);
2484 static void register_cache_worker(struct work_struct *work)
2487 struct async_reg_args *args =
2488 container_of(work, struct async_reg_args, reg_work.work);
2491 ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2494 put_page(virt_to_page(args->sb_disk));
2495 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2499 /* blkdev_put() will be called in bch_cache_release() */
2500 if (register_cache(args->sb, args->sb_disk, args->bdev, ca) != 0)
2505 pr_info("error %s: fail to register cache device\n",
2510 module_put(THIS_MODULE);
2513 static void register_device_async(struct async_reg_args *args)
2515 if (SB_IS_BDEV(args->sb))
2516 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2518 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2520 /* 10 jiffies is enough for a delay */
2521 queue_delayed_work(system_wq, &args->reg_work, 10);
2524 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2525 const char *buffer, size_t size)
2529 struct cache_sb *sb;
2530 struct cache_sb_disk *sb_disk;
2531 struct block_device *bdev;
2533 bool async_registration = false;
2535 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2536 async_registration = true;
2540 err = "failed to reference bcache module";
2541 if (!try_module_get(THIS_MODULE))
2544 /* For latest state of bcache_is_reboot */
2546 err = "bcache is in reboot";
2547 if (bcache_is_reboot)
2548 goto out_module_put;
2551 err = "cannot allocate memory";
2552 path = kstrndup(buffer, size, GFP_KERNEL);
2554 goto out_module_put;
2556 sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2561 err = "failed to open device";
2562 bdev = blkdev_get_by_path(strim(path),
2563 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2566 if (bdev == ERR_PTR(-EBUSY)) {
2569 mutex_lock(&bch_register_lock);
2570 if (lookup_bdev(strim(path), &dev) == 0 &&
2572 err = "device already registered";
2574 err = "device busy";
2575 mutex_unlock(&bch_register_lock);
2576 if (attr == &ksysfs_register_quiet)
2582 err = "failed to set blocksize";
2583 if (set_blocksize(bdev, 4096))
2584 goto out_blkdev_put;
2586 err = read_super(sb, bdev, &sb_disk);
2588 goto out_blkdev_put;
2590 err = "failed to register device";
2592 if (async_registration) {
2593 /* register in asynchronous way */
2594 struct async_reg_args *args =
2595 kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2599 err = "cannot allocate memory";
2600 goto out_put_sb_page;
2605 args->sb_disk = sb_disk;
2607 register_device_async(args);
2608 /* No wait and returns to user space */
2612 if (SB_IS_BDEV(sb)) {
2613 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2617 err = "cannot allocate memory";
2618 goto out_put_sb_page;
2621 mutex_lock(&bch_register_lock);
2622 ret = register_bdev(sb, sb_disk, bdev, dc);
2623 mutex_unlock(&bch_register_lock);
2624 /* blkdev_put() will be called in cached_dev_free() */
2628 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2632 err = "cannot allocate memory";
2633 goto out_put_sb_page;
2636 /* blkdev_put() will be called in bch_cache_release() */
2637 ret = register_cache(sb, sb_disk, bdev, ca);
2645 module_put(THIS_MODULE);
2650 put_page(virt_to_page(sb_disk));
2652 blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2659 module_put(THIS_MODULE);
2661 pr_info("error %s: %s\n", path?path:"", err);
2667 struct list_head list;
2668 struct cached_dev *dc;
2671 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2672 struct kobj_attribute *attr,
2676 LIST_HEAD(pending_devs);
2678 struct cached_dev *dc, *tdc;
2679 struct pdev *pdev, *tpdev;
2680 struct cache_set *c, *tc;
2682 mutex_lock(&bch_register_lock);
2683 list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2684 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2688 list_add(&pdev->list, &pending_devs);
2691 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2692 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2693 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2694 char *set_uuid = c->set_uuid;
2696 if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2697 list_del(&pdev->list);
2703 mutex_unlock(&bch_register_lock);
2705 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2706 pr_info("delete pdev %p\n", pdev);
2707 list_del(&pdev->list);
2708 bcache_device_stop(&pdev->dc->disk);
2715 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2717 if (bcache_is_reboot)
2720 if (code == SYS_DOWN ||
2722 code == SYS_POWER_OFF) {
2724 unsigned long start = jiffies;
2725 bool stopped = false;
2727 struct cache_set *c, *tc;
2728 struct cached_dev *dc, *tdc;
2730 mutex_lock(&bch_register_lock);
2732 if (bcache_is_reboot)
2735 /* New registration is rejected since now */
2736 bcache_is_reboot = true;
2738 * Make registering caller (if there is) on other CPU
2739 * core know bcache_is_reboot set to true earlier
2743 if (list_empty(&bch_cache_sets) &&
2744 list_empty(&uncached_devices))
2747 mutex_unlock(&bch_register_lock);
2749 pr_info("Stopping all devices:\n");
2752 * The reason bch_register_lock is not held to call
2753 * bch_cache_set_stop() and bcache_device_stop() is to
2754 * avoid potential deadlock during reboot, because cache
2755 * set or bcache device stopping process will acquire
2756 * bch_register_lock too.
2758 * We are safe here because bcache_is_reboot sets to
2759 * true already, register_bcache() will reject new
2760 * registration now. bcache_is_reboot also makes sure
2761 * bcache_reboot() won't be re-entered on by other thread,
2762 * so there is no race in following list iteration by
2763 * list_for_each_entry_safe().
2765 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2766 bch_cache_set_stop(c);
2768 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2769 bcache_device_stop(&dc->disk);
2773 * Give an early chance for other kthreads and
2774 * kworkers to stop themselves
2778 /* What's a condition variable? */
2780 long timeout = start + 10 * HZ - jiffies;
2782 mutex_lock(&bch_register_lock);
2783 stopped = list_empty(&bch_cache_sets) &&
2784 list_empty(&uncached_devices);
2786 if (timeout < 0 || stopped)
2789 prepare_to_wait(&unregister_wait, &wait,
2790 TASK_UNINTERRUPTIBLE);
2792 mutex_unlock(&bch_register_lock);
2793 schedule_timeout(timeout);
2796 finish_wait(&unregister_wait, &wait);
2799 pr_info("All devices stopped\n");
2801 pr_notice("Timeout waiting for devices to be closed\n");
2803 mutex_unlock(&bch_register_lock);
2809 static struct notifier_block reboot = {
2810 .notifier_call = bcache_reboot,
2811 .priority = INT_MAX, /* before any real devices */
2814 static void bcache_exit(void)
2819 kobject_put(bcache_kobj);
2821 destroy_workqueue(bcache_wq);
2823 destroy_workqueue(bch_journal_wq);
2825 destroy_workqueue(bch_flush_wq);
2829 unregister_blkdev(bcache_major, "bcache");
2830 unregister_reboot_notifier(&reboot);
2831 mutex_destroy(&bch_register_lock);
2834 /* Check and fixup module parameters */
2835 static void check_module_parameters(void)
2837 if (bch_cutoff_writeback_sync == 0)
2838 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2839 else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2840 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2841 bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2842 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2845 if (bch_cutoff_writeback == 0)
2846 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2847 else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2848 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2849 bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2850 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2853 if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2854 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2855 bch_cutoff_writeback, bch_cutoff_writeback_sync);
2856 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2860 static int __init bcache_init(void)
2862 static const struct attribute *files[] = {
2863 &ksysfs_register.attr,
2864 &ksysfs_register_quiet.attr,
2865 &ksysfs_pendings_cleanup.attr,
2869 check_module_parameters();
2871 mutex_init(&bch_register_lock);
2872 init_waitqueue_head(&unregister_wait);
2873 register_reboot_notifier(&reboot);
2875 bcache_major = register_blkdev(0, "bcache");
2876 if (bcache_major < 0) {
2877 unregister_reboot_notifier(&reboot);
2878 mutex_destroy(&bch_register_lock);
2879 return bcache_major;
2882 if (bch_btree_init())
2885 bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2890 * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2892 * 1. It used `system_wq` before which also does no memory reclaim.
2893 * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2894 * reduced throughput can be observed.
2896 * We still want to user our own queue to not congest the `system_wq`.
2898 bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2902 bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2903 if (!bch_journal_wq)
2906 bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2910 if (bch_request_init() ||
2911 sysfs_create_files(bcache_kobj, files))
2915 closure_debug_init();
2917 bcache_is_reboot = false;
2928 module_exit(bcache_exit);
2929 module_init(bcache_init);
2931 module_param(bch_cutoff_writeback, uint, 0);
2932 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2934 module_param(bch_cutoff_writeback_sync, uint, 0);
2935 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2937 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2938 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2939 MODULE_LICENSE("GPL");