Loading...
Note: File does not exist in v3.5.6.
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * bcache setup/teardown code, and some metadata io - read a superblock and
4 * figure out what to do with it.
5 *
6 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
7 * Copyright 2012 Google, Inc.
8 */
9
10#include "bcache.h"
11#include "btree.h"
12#include "debug.h"
13#include "extents.h"
14#include "request.h"
15#include "writeback.h"
16
17#include <linux/blkdev.h>
18#include <linux/buffer_head.h>
19#include <linux/debugfs.h>
20#include <linux/genhd.h>
21#include <linux/idr.h>
22#include <linux/kthread.h>
23#include <linux/module.h>
24#include <linux/random.h>
25#include <linux/reboot.h>
26#include <linux/sysfs.h>
27
28unsigned int bch_cutoff_writeback;
29unsigned int bch_cutoff_writeback_sync;
30
31static const char bcache_magic[] = {
32 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
33 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
34};
35
36static const char invalid_uuid[] = {
37 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
38 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
39};
40
41static struct kobject *bcache_kobj;
42struct mutex bch_register_lock;
43bool bcache_is_reboot;
44LIST_HEAD(bch_cache_sets);
45static LIST_HEAD(uncached_devices);
46
47static int bcache_major;
48static DEFINE_IDA(bcache_device_idx);
49static wait_queue_head_t unregister_wait;
50struct workqueue_struct *bcache_wq;
51struct workqueue_struct *bch_journal_wq;
52
53
54#define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
55/* limitation of partitions number on single bcache device */
56#define BCACHE_MINORS 128
57/* limitation of bcache devices number on single system */
58#define BCACHE_DEVICE_IDX_MAX ((1U << MINORBITS)/BCACHE_MINORS)
59
60/* Superblock */
61
62static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
63 struct page **res)
64{
65 const char *err;
66 struct cache_sb *s;
67 struct buffer_head *bh = __bread(bdev, 1, SB_SIZE);
68 unsigned int i;
69
70 if (!bh)
71 return "IO error";
72
73 s = (struct cache_sb *) bh->b_data;
74
75 sb->offset = le64_to_cpu(s->offset);
76 sb->version = le64_to_cpu(s->version);
77
78 memcpy(sb->magic, s->magic, 16);
79 memcpy(sb->uuid, s->uuid, 16);
80 memcpy(sb->set_uuid, s->set_uuid, 16);
81 memcpy(sb->label, s->label, SB_LABEL_SIZE);
82
83 sb->flags = le64_to_cpu(s->flags);
84 sb->seq = le64_to_cpu(s->seq);
85 sb->last_mount = le32_to_cpu(s->last_mount);
86 sb->first_bucket = le16_to_cpu(s->first_bucket);
87 sb->keys = le16_to_cpu(s->keys);
88
89 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
90 sb->d[i] = le64_to_cpu(s->d[i]);
91
92 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u",
93 sb->version, sb->flags, sb->seq, sb->keys);
94
95 err = "Not a bcache superblock";
96 if (sb->offset != SB_SECTOR)
97 goto err;
98
99 if (memcmp(sb->magic, bcache_magic, 16))
100 goto err;
101
102 err = "Too many journal buckets";
103 if (sb->keys > SB_JOURNAL_BUCKETS)
104 goto err;
105
106 err = "Bad checksum";
107 if (s->csum != csum_set(s))
108 goto err;
109
110 err = "Bad UUID";
111 if (bch_is_zero(sb->uuid, 16))
112 goto err;
113
114 sb->block_size = le16_to_cpu(s->block_size);
115
116 err = "Superblock block size smaller than device block size";
117 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
118 goto err;
119
120 switch (sb->version) {
121 case BCACHE_SB_VERSION_BDEV:
122 sb->data_offset = BDEV_DATA_START_DEFAULT;
123 break;
124 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
125 sb->data_offset = le64_to_cpu(s->data_offset);
126
127 err = "Bad data offset";
128 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
129 goto err;
130
131 break;
132 case BCACHE_SB_VERSION_CDEV:
133 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
134 sb->nbuckets = le64_to_cpu(s->nbuckets);
135 sb->bucket_size = le16_to_cpu(s->bucket_size);
136
137 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
138 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
139
140 err = "Too many buckets";
141 if (sb->nbuckets > LONG_MAX)
142 goto err;
143
144 err = "Not enough buckets";
145 if (sb->nbuckets < 1 << 7)
146 goto err;
147
148 err = "Bad block/bucket size";
149 if (!is_power_of_2(sb->block_size) ||
150 sb->block_size > PAGE_SECTORS ||
151 !is_power_of_2(sb->bucket_size) ||
152 sb->bucket_size < PAGE_SECTORS)
153 goto err;
154
155 err = "Invalid superblock: device too small";
156 if (get_capacity(bdev->bd_disk) <
157 sb->bucket_size * sb->nbuckets)
158 goto err;
159
160 err = "Bad UUID";
161 if (bch_is_zero(sb->set_uuid, 16))
162 goto err;
163
164 err = "Bad cache device number in set";
165 if (!sb->nr_in_set ||
166 sb->nr_in_set <= sb->nr_this_dev ||
167 sb->nr_in_set > MAX_CACHES_PER_SET)
168 goto err;
169
170 err = "Journal buckets not sequential";
171 for (i = 0; i < sb->keys; i++)
172 if (sb->d[i] != sb->first_bucket + i)
173 goto err;
174
175 err = "Too many journal buckets";
176 if (sb->first_bucket + sb->keys > sb->nbuckets)
177 goto err;
178
179 err = "Invalid superblock: first bucket comes before end of super";
180 if (sb->first_bucket * sb->bucket_size < 16)
181 goto err;
182
183 break;
184 default:
185 err = "Unsupported superblock version";
186 goto err;
187 }
188
189 sb->last_mount = (u32)ktime_get_real_seconds();
190 err = NULL;
191
192 get_page(bh->b_page);
193 *res = bh->b_page;
194err:
195 put_bh(bh);
196 return err;
197}
198
199static void write_bdev_super_endio(struct bio *bio)
200{
201 struct cached_dev *dc = bio->bi_private;
202
203 if (bio->bi_status)
204 bch_count_backing_io_errors(dc, bio);
205
206 closure_put(&dc->sb_write);
207}
208
209static void __write_super(struct cache_sb *sb, struct bio *bio)
210{
211 struct cache_sb *out = page_address(bio_first_page_all(bio));
212 unsigned int i;
213
214 bio->bi_iter.bi_sector = SB_SECTOR;
215 bio->bi_iter.bi_size = SB_SIZE;
216 bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC|REQ_META);
217 bch_bio_map(bio, NULL);
218
219 out->offset = cpu_to_le64(sb->offset);
220 out->version = cpu_to_le64(sb->version);
221
222 memcpy(out->uuid, sb->uuid, 16);
223 memcpy(out->set_uuid, sb->set_uuid, 16);
224 memcpy(out->label, sb->label, SB_LABEL_SIZE);
225
226 out->flags = cpu_to_le64(sb->flags);
227 out->seq = cpu_to_le64(sb->seq);
228
229 out->last_mount = cpu_to_le32(sb->last_mount);
230 out->first_bucket = cpu_to_le16(sb->first_bucket);
231 out->keys = cpu_to_le16(sb->keys);
232
233 for (i = 0; i < sb->keys; i++)
234 out->d[i] = cpu_to_le64(sb->d[i]);
235
236 out->csum = csum_set(out);
237
238 pr_debug("ver %llu, flags %llu, seq %llu",
239 sb->version, sb->flags, sb->seq);
240
241 submit_bio(bio);
242}
243
244static void bch_write_bdev_super_unlock(struct closure *cl)
245{
246 struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
247
248 up(&dc->sb_write_mutex);
249}
250
251void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
252{
253 struct closure *cl = &dc->sb_write;
254 struct bio *bio = &dc->sb_bio;
255
256 down(&dc->sb_write_mutex);
257 closure_init(cl, parent);
258
259 bio_reset(bio);
260 bio_set_dev(bio, dc->bdev);
261 bio->bi_end_io = write_bdev_super_endio;
262 bio->bi_private = dc;
263
264 closure_get(cl);
265 /* I/O request sent to backing device */
266 __write_super(&dc->sb, bio);
267
268 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
269}
270
271static void write_super_endio(struct bio *bio)
272{
273 struct cache *ca = bio->bi_private;
274
275 /* is_read = 0 */
276 bch_count_io_errors(ca, bio->bi_status, 0,
277 "writing superblock");
278 closure_put(&ca->set->sb_write);
279}
280
281static void bcache_write_super_unlock(struct closure *cl)
282{
283 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
284
285 up(&c->sb_write_mutex);
286}
287
288void bcache_write_super(struct cache_set *c)
289{
290 struct closure *cl = &c->sb_write;
291 struct cache *ca;
292 unsigned int i;
293
294 down(&c->sb_write_mutex);
295 closure_init(cl, &c->cl);
296
297 c->sb.seq++;
298
299 for_each_cache(ca, c, i) {
300 struct bio *bio = &ca->sb_bio;
301
302 ca->sb.version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
303 ca->sb.seq = c->sb.seq;
304 ca->sb.last_mount = c->sb.last_mount;
305
306 SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
307
308 bio_reset(bio);
309 bio_set_dev(bio, ca->bdev);
310 bio->bi_end_io = write_super_endio;
311 bio->bi_private = ca;
312
313 closure_get(cl);
314 __write_super(&ca->sb, bio);
315 }
316
317 closure_return_with_destructor(cl, bcache_write_super_unlock);
318}
319
320/* UUID io */
321
322static void uuid_endio(struct bio *bio)
323{
324 struct closure *cl = bio->bi_private;
325 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
326
327 cache_set_err_on(bio->bi_status, c, "accessing uuids");
328 bch_bbio_free(bio, c);
329 closure_put(cl);
330}
331
332static void uuid_io_unlock(struct closure *cl)
333{
334 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
335
336 up(&c->uuid_write_mutex);
337}
338
339static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
340 struct bkey *k, struct closure *parent)
341{
342 struct closure *cl = &c->uuid_write;
343 struct uuid_entry *u;
344 unsigned int i;
345 char buf[80];
346
347 BUG_ON(!parent);
348 down(&c->uuid_write_mutex);
349 closure_init(cl, parent);
350
351 for (i = 0; i < KEY_PTRS(k); i++) {
352 struct bio *bio = bch_bbio_alloc(c);
353
354 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
355 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
356
357 bio->bi_end_io = uuid_endio;
358 bio->bi_private = cl;
359 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
360 bch_bio_map(bio, c->uuids);
361
362 bch_submit_bbio(bio, c, k, i);
363
364 if (op != REQ_OP_WRITE)
365 break;
366 }
367
368 bch_extent_to_text(buf, sizeof(buf), k);
369 pr_debug("%s UUIDs at %s", op == REQ_OP_WRITE ? "wrote" : "read", buf);
370
371 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
372 if (!bch_is_zero(u->uuid, 16))
373 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
374 u - c->uuids, u->uuid, u->label,
375 u->first_reg, u->last_reg, u->invalidated);
376
377 closure_return_with_destructor(cl, uuid_io_unlock);
378}
379
380static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
381{
382 struct bkey *k = &j->uuid_bucket;
383
384 if (__bch_btree_ptr_invalid(c, k))
385 return "bad uuid pointer";
386
387 bkey_copy(&c->uuid_bucket, k);
388 uuid_io(c, REQ_OP_READ, 0, k, cl);
389
390 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
391 struct uuid_entry_v0 *u0 = (void *) c->uuids;
392 struct uuid_entry *u1 = (void *) c->uuids;
393 int i;
394
395 closure_sync(cl);
396
397 /*
398 * Since the new uuid entry is bigger than the old, we have to
399 * convert starting at the highest memory address and work down
400 * in order to do it in place
401 */
402
403 for (i = c->nr_uuids - 1;
404 i >= 0;
405 --i) {
406 memcpy(u1[i].uuid, u0[i].uuid, 16);
407 memcpy(u1[i].label, u0[i].label, 32);
408
409 u1[i].first_reg = u0[i].first_reg;
410 u1[i].last_reg = u0[i].last_reg;
411 u1[i].invalidated = u0[i].invalidated;
412
413 u1[i].flags = 0;
414 u1[i].sectors = 0;
415 }
416 }
417
418 return NULL;
419}
420
421static int __uuid_write(struct cache_set *c)
422{
423 BKEY_PADDED(key) k;
424 struct closure cl;
425 struct cache *ca;
426
427 closure_init_stack(&cl);
428 lockdep_assert_held(&bch_register_lock);
429
430 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
431 return 1;
432
433 SET_KEY_SIZE(&k.key, c->sb.bucket_size);
434 uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
435 closure_sync(&cl);
436
437 /* Only one bucket used for uuid write */
438 ca = PTR_CACHE(c, &k.key, 0);
439 atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
440
441 bkey_copy(&c->uuid_bucket, &k.key);
442 bkey_put(c, &k.key);
443 return 0;
444}
445
446int bch_uuid_write(struct cache_set *c)
447{
448 int ret = __uuid_write(c);
449
450 if (!ret)
451 bch_journal_meta(c, NULL);
452
453 return ret;
454}
455
456static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
457{
458 struct uuid_entry *u;
459
460 for (u = c->uuids;
461 u < c->uuids + c->nr_uuids; u++)
462 if (!memcmp(u->uuid, uuid, 16))
463 return u;
464
465 return NULL;
466}
467
468static struct uuid_entry *uuid_find_empty(struct cache_set *c)
469{
470 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
471
472 return uuid_find(c, zero_uuid);
473}
474
475/*
476 * Bucket priorities/gens:
477 *
478 * For each bucket, we store on disk its
479 * 8 bit gen
480 * 16 bit priority
481 *
482 * See alloc.c for an explanation of the gen. The priority is used to implement
483 * lru (and in the future other) cache replacement policies; for most purposes
484 * it's just an opaque integer.
485 *
486 * The gens and the priorities don't have a whole lot to do with each other, and
487 * it's actually the gens that must be written out at specific times - it's no
488 * big deal if the priorities don't get written, if we lose them we just reuse
489 * buckets in suboptimal order.
490 *
491 * On disk they're stored in a packed array, and in as many buckets are required
492 * to fit them all. The buckets we use to store them form a list; the journal
493 * header points to the first bucket, the first bucket points to the second
494 * bucket, et cetera.
495 *
496 * This code is used by the allocation code; periodically (whenever it runs out
497 * of buckets to allocate from) the allocation code will invalidate some
498 * buckets, but it can't use those buckets until their new gens are safely on
499 * disk.
500 */
501
502static void prio_endio(struct bio *bio)
503{
504 struct cache *ca = bio->bi_private;
505
506 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
507 bch_bbio_free(bio, ca->set);
508 closure_put(&ca->prio);
509}
510
511static void prio_io(struct cache *ca, uint64_t bucket, int op,
512 unsigned long op_flags)
513{
514 struct closure *cl = &ca->prio;
515 struct bio *bio = bch_bbio_alloc(ca->set);
516
517 closure_init_stack(cl);
518
519 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
520 bio_set_dev(bio, ca->bdev);
521 bio->bi_iter.bi_size = bucket_bytes(ca);
522
523 bio->bi_end_io = prio_endio;
524 bio->bi_private = ca;
525 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
526 bch_bio_map(bio, ca->disk_buckets);
527
528 closure_bio_submit(ca->set, bio, &ca->prio);
529 closure_sync(cl);
530}
531
532void bch_prio_write(struct cache *ca)
533{
534 int i;
535 struct bucket *b;
536 struct closure cl;
537
538 closure_init_stack(&cl);
539
540 lockdep_assert_held(&ca->set->bucket_lock);
541
542 ca->disk_buckets->seq++;
543
544 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
545 &ca->meta_sectors_written);
546
547 //pr_debug("free %zu, free_inc %zu, unused %zu", fifo_used(&ca->free),
548 // fifo_used(&ca->free_inc), fifo_used(&ca->unused));
549
550 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
551 long bucket;
552 struct prio_set *p = ca->disk_buckets;
553 struct bucket_disk *d = p->data;
554 struct bucket_disk *end = d + prios_per_bucket(ca);
555
556 for (b = ca->buckets + i * prios_per_bucket(ca);
557 b < ca->buckets + ca->sb.nbuckets && d < end;
558 b++, d++) {
559 d->prio = cpu_to_le16(b->prio);
560 d->gen = b->gen;
561 }
562
563 p->next_bucket = ca->prio_buckets[i + 1];
564 p->magic = pset_magic(&ca->sb);
565 p->csum = bch_crc64(&p->magic, bucket_bytes(ca) - 8);
566
567 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, true);
568 BUG_ON(bucket == -1);
569
570 mutex_unlock(&ca->set->bucket_lock);
571 prio_io(ca, bucket, REQ_OP_WRITE, 0);
572 mutex_lock(&ca->set->bucket_lock);
573
574 ca->prio_buckets[i] = bucket;
575 atomic_dec_bug(&ca->buckets[bucket].pin);
576 }
577
578 mutex_unlock(&ca->set->bucket_lock);
579
580 bch_journal_meta(ca->set, &cl);
581 closure_sync(&cl);
582
583 mutex_lock(&ca->set->bucket_lock);
584
585 /*
586 * Don't want the old priorities to get garbage collected until after we
587 * finish writing the new ones, and they're journalled
588 */
589 for (i = 0; i < prio_buckets(ca); i++) {
590 if (ca->prio_last_buckets[i])
591 __bch_bucket_free(ca,
592 &ca->buckets[ca->prio_last_buckets[i]]);
593
594 ca->prio_last_buckets[i] = ca->prio_buckets[i];
595 }
596}
597
598static void prio_read(struct cache *ca, uint64_t bucket)
599{
600 struct prio_set *p = ca->disk_buckets;
601 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
602 struct bucket *b;
603 unsigned int bucket_nr = 0;
604
605 for (b = ca->buckets;
606 b < ca->buckets + ca->sb.nbuckets;
607 b++, d++) {
608 if (d == end) {
609 ca->prio_buckets[bucket_nr] = bucket;
610 ca->prio_last_buckets[bucket_nr] = bucket;
611 bucket_nr++;
612
613 prio_io(ca, bucket, REQ_OP_READ, 0);
614
615 if (p->csum !=
616 bch_crc64(&p->magic, bucket_bytes(ca) - 8))
617 pr_warn("bad csum reading priorities");
618
619 if (p->magic != pset_magic(&ca->sb))
620 pr_warn("bad magic reading priorities");
621
622 bucket = p->next_bucket;
623 d = p->data;
624 }
625
626 b->prio = le16_to_cpu(d->prio);
627 b->gen = b->last_gc = d->gen;
628 }
629}
630
631/* Bcache device */
632
633static int open_dev(struct block_device *b, fmode_t mode)
634{
635 struct bcache_device *d = b->bd_disk->private_data;
636
637 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
638 return -ENXIO;
639
640 closure_get(&d->cl);
641 return 0;
642}
643
644static void release_dev(struct gendisk *b, fmode_t mode)
645{
646 struct bcache_device *d = b->private_data;
647
648 closure_put(&d->cl);
649}
650
651static int ioctl_dev(struct block_device *b, fmode_t mode,
652 unsigned int cmd, unsigned long arg)
653{
654 struct bcache_device *d = b->bd_disk->private_data;
655
656 return d->ioctl(d, mode, cmd, arg);
657}
658
659static const struct block_device_operations bcache_ops = {
660 .open = open_dev,
661 .release = release_dev,
662 .ioctl = ioctl_dev,
663 .owner = THIS_MODULE,
664};
665
666void bcache_device_stop(struct bcache_device *d)
667{
668 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
669 /*
670 * closure_fn set to
671 * - cached device: cached_dev_flush()
672 * - flash dev: flash_dev_flush()
673 */
674 closure_queue(&d->cl);
675}
676
677static void bcache_device_unlink(struct bcache_device *d)
678{
679 lockdep_assert_held(&bch_register_lock);
680
681 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
682 unsigned int i;
683 struct cache *ca;
684
685 sysfs_remove_link(&d->c->kobj, d->name);
686 sysfs_remove_link(&d->kobj, "cache");
687
688 for_each_cache(ca, d->c, i)
689 bd_unlink_disk_holder(ca->bdev, d->disk);
690 }
691}
692
693static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
694 const char *name)
695{
696 unsigned int i;
697 struct cache *ca;
698 int ret;
699
700 for_each_cache(ca, d->c, i)
701 bd_link_disk_holder(ca->bdev, d->disk);
702
703 snprintf(d->name, BCACHEDEVNAME_SIZE,
704 "%s%u", name, d->id);
705
706 ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
707 if (ret < 0)
708 pr_err("Couldn't create device -> cache set symlink");
709
710 ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
711 if (ret < 0)
712 pr_err("Couldn't create cache set -> device symlink");
713
714 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
715}
716
717static void bcache_device_detach(struct bcache_device *d)
718{
719 lockdep_assert_held(&bch_register_lock);
720
721 atomic_dec(&d->c->attached_dev_nr);
722
723 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
724 struct uuid_entry *u = d->c->uuids + d->id;
725
726 SET_UUID_FLASH_ONLY(u, 0);
727 memcpy(u->uuid, invalid_uuid, 16);
728 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
729 bch_uuid_write(d->c);
730 }
731
732 bcache_device_unlink(d);
733
734 d->c->devices[d->id] = NULL;
735 closure_put(&d->c->caching);
736 d->c = NULL;
737}
738
739static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
740 unsigned int id)
741{
742 d->id = id;
743 d->c = c;
744 c->devices[id] = d;
745
746 if (id >= c->devices_max_used)
747 c->devices_max_used = id + 1;
748
749 closure_get(&c->caching);
750}
751
752static inline int first_minor_to_idx(int first_minor)
753{
754 return (first_minor/BCACHE_MINORS);
755}
756
757static inline int idx_to_first_minor(int idx)
758{
759 return (idx * BCACHE_MINORS);
760}
761
762static void bcache_device_free(struct bcache_device *d)
763{
764 lockdep_assert_held(&bch_register_lock);
765
766 pr_info("%s stopped", d->disk->disk_name);
767
768 if (d->c)
769 bcache_device_detach(d);
770 if (d->disk && d->disk->flags & GENHD_FL_UP)
771 del_gendisk(d->disk);
772 if (d->disk && d->disk->queue)
773 blk_cleanup_queue(d->disk->queue);
774 if (d->disk) {
775 ida_simple_remove(&bcache_device_idx,
776 first_minor_to_idx(d->disk->first_minor));
777 put_disk(d->disk);
778 }
779
780 bioset_exit(&d->bio_split);
781 kvfree(d->full_dirty_stripes);
782 kvfree(d->stripe_sectors_dirty);
783
784 closure_debug_destroy(&d->cl);
785}
786
787static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
788 sector_t sectors)
789{
790 struct request_queue *q;
791 const size_t max_stripes = min_t(size_t, INT_MAX,
792 SIZE_MAX / sizeof(atomic_t));
793 size_t n;
794 int idx;
795
796 if (!d->stripe_size)
797 d->stripe_size = 1 << 31;
798
799 d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
800
801 if (!d->nr_stripes || d->nr_stripes > max_stripes) {
802 pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
803 (unsigned int)d->nr_stripes);
804 return -ENOMEM;
805 }
806
807 n = d->nr_stripes * sizeof(atomic_t);
808 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
809 if (!d->stripe_sectors_dirty)
810 return -ENOMEM;
811
812 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
813 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
814 if (!d->full_dirty_stripes)
815 return -ENOMEM;
816
817 idx = ida_simple_get(&bcache_device_idx, 0,
818 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
819 if (idx < 0)
820 return idx;
821
822 if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
823 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
824 goto err;
825
826 d->disk = alloc_disk(BCACHE_MINORS);
827 if (!d->disk)
828 goto err;
829
830 set_capacity(d->disk, sectors);
831 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
832
833 d->disk->major = bcache_major;
834 d->disk->first_minor = idx_to_first_minor(idx);
835 d->disk->fops = &bcache_ops;
836 d->disk->private_data = d;
837
838 q = blk_alloc_queue(GFP_KERNEL);
839 if (!q)
840 return -ENOMEM;
841
842 blk_queue_make_request(q, NULL);
843 d->disk->queue = q;
844 q->queuedata = d;
845 q->backing_dev_info->congested_data = d;
846 q->limits.max_hw_sectors = UINT_MAX;
847 q->limits.max_sectors = UINT_MAX;
848 q->limits.max_segment_size = UINT_MAX;
849 q->limits.max_segments = BIO_MAX_PAGES;
850 blk_queue_max_discard_sectors(q, UINT_MAX);
851 q->limits.discard_granularity = 512;
852 q->limits.io_min = block_size;
853 q->limits.logical_block_size = block_size;
854 q->limits.physical_block_size = block_size;
855 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
856 blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
857 blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
858
859 blk_queue_write_cache(q, true, true);
860
861 return 0;
862
863err:
864 ida_simple_remove(&bcache_device_idx, idx);
865 return -ENOMEM;
866
867}
868
869/* Cached device */
870
871static void calc_cached_dev_sectors(struct cache_set *c)
872{
873 uint64_t sectors = 0;
874 struct cached_dev *dc;
875
876 list_for_each_entry(dc, &c->cached_devs, list)
877 sectors += bdev_sectors(dc->bdev);
878
879 c->cached_dev_sectors = sectors;
880}
881
882#define BACKING_DEV_OFFLINE_TIMEOUT 5
883static int cached_dev_status_update(void *arg)
884{
885 struct cached_dev *dc = arg;
886 struct request_queue *q;
887
888 /*
889 * If this delayed worker is stopping outside, directly quit here.
890 * dc->io_disable might be set via sysfs interface, so check it
891 * here too.
892 */
893 while (!kthread_should_stop() && !dc->io_disable) {
894 q = bdev_get_queue(dc->bdev);
895 if (blk_queue_dying(q))
896 dc->offline_seconds++;
897 else
898 dc->offline_seconds = 0;
899
900 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
901 pr_err("%s: device offline for %d seconds",
902 dc->backing_dev_name,
903 BACKING_DEV_OFFLINE_TIMEOUT);
904 pr_err("%s: disable I/O request due to backing "
905 "device offline", dc->disk.name);
906 dc->io_disable = true;
907 /* let others know earlier that io_disable is true */
908 smp_mb();
909 bcache_device_stop(&dc->disk);
910 break;
911 }
912 schedule_timeout_interruptible(HZ);
913 }
914
915 wait_for_kthread_stop();
916 return 0;
917}
918
919
920int bch_cached_dev_run(struct cached_dev *dc)
921{
922 struct bcache_device *d = &dc->disk;
923 char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
924 char *env[] = {
925 "DRIVER=bcache",
926 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
927 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
928 NULL,
929 };
930
931 if (dc->io_disable) {
932 pr_err("I/O disabled on cached dev %s",
933 dc->backing_dev_name);
934 kfree(env[1]);
935 kfree(env[2]);
936 kfree(buf);
937 return -EIO;
938 }
939
940 if (atomic_xchg(&dc->running, 1)) {
941 kfree(env[1]);
942 kfree(env[2]);
943 kfree(buf);
944 pr_info("cached dev %s is running already",
945 dc->backing_dev_name);
946 return -EBUSY;
947 }
948
949 if (!d->c &&
950 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
951 struct closure cl;
952
953 closure_init_stack(&cl);
954
955 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
956 bch_write_bdev_super(dc, &cl);
957 closure_sync(&cl);
958 }
959
960 add_disk(d->disk);
961 bd_link_disk_holder(dc->bdev, dc->disk.disk);
962 /*
963 * won't show up in the uevent file, use udevadm monitor -e instead
964 * only class / kset properties are persistent
965 */
966 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
967 kfree(env[1]);
968 kfree(env[2]);
969 kfree(buf);
970
971 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
972 sysfs_create_link(&disk_to_dev(d->disk)->kobj,
973 &d->kobj, "bcache")) {
974 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks");
975 return -ENOMEM;
976 }
977
978 dc->status_update_thread = kthread_run(cached_dev_status_update,
979 dc, "bcache_status_update");
980 if (IS_ERR(dc->status_update_thread)) {
981 pr_warn("failed to create bcache_status_update kthread, "
982 "continue to run without monitoring backing "
983 "device status");
984 }
985
986 return 0;
987}
988
989/*
990 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
991 * work dc->writeback_rate_update is running. Wait until the routine
992 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
993 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
994 * seconds, give up waiting here and continue to cancel it too.
995 */
996static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
997{
998 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
999
1000 do {
1001 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1002 &dc->disk.flags))
1003 break;
1004 time_out--;
1005 schedule_timeout_interruptible(1);
1006 } while (time_out > 0);
1007
1008 if (time_out == 0)
1009 pr_warn("give up waiting for dc->writeback_write_update to quit");
1010
1011 cancel_delayed_work_sync(&dc->writeback_rate_update);
1012}
1013
1014static void cached_dev_detach_finish(struct work_struct *w)
1015{
1016 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1017 struct closure cl;
1018
1019 closure_init_stack(&cl);
1020
1021 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1022 BUG_ON(refcount_read(&dc->count));
1023
1024
1025 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1026 cancel_writeback_rate_update_dwork(dc);
1027
1028 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1029 kthread_stop(dc->writeback_thread);
1030 dc->writeback_thread = NULL;
1031 }
1032
1033 memset(&dc->sb.set_uuid, 0, 16);
1034 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1035
1036 bch_write_bdev_super(dc, &cl);
1037 closure_sync(&cl);
1038
1039 mutex_lock(&bch_register_lock);
1040
1041 calc_cached_dev_sectors(dc->disk.c);
1042 bcache_device_detach(&dc->disk);
1043 list_move(&dc->list, &uncached_devices);
1044
1045 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1046 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1047
1048 mutex_unlock(&bch_register_lock);
1049
1050 pr_info("Caching disabled for %s", dc->backing_dev_name);
1051
1052 /* Drop ref we took in cached_dev_detach() */
1053 closure_put(&dc->disk.cl);
1054}
1055
1056void bch_cached_dev_detach(struct cached_dev *dc)
1057{
1058 lockdep_assert_held(&bch_register_lock);
1059
1060 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1061 return;
1062
1063 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1064 return;
1065
1066 /*
1067 * Block the device from being closed and freed until we're finished
1068 * detaching
1069 */
1070 closure_get(&dc->disk.cl);
1071
1072 bch_writeback_queue(dc);
1073
1074 cached_dev_put(dc);
1075}
1076
1077int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1078 uint8_t *set_uuid)
1079{
1080 uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1081 struct uuid_entry *u;
1082 struct cached_dev *exist_dc, *t;
1083 int ret = 0;
1084
1085 if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1086 (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1087 return -ENOENT;
1088
1089 if (dc->disk.c) {
1090 pr_err("Can't attach %s: already attached",
1091 dc->backing_dev_name);
1092 return -EINVAL;
1093 }
1094
1095 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1096 pr_err("Can't attach %s: shutting down",
1097 dc->backing_dev_name);
1098 return -EINVAL;
1099 }
1100
1101 if (dc->sb.block_size < c->sb.block_size) {
1102 /* Will die */
1103 pr_err("Couldn't attach %s: block size less than set's block size",
1104 dc->backing_dev_name);
1105 return -EINVAL;
1106 }
1107
1108 /* Check whether already attached */
1109 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1110 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1111 pr_err("Tried to attach %s but duplicate UUID already attached",
1112 dc->backing_dev_name);
1113
1114 return -EINVAL;
1115 }
1116 }
1117
1118 u = uuid_find(c, dc->sb.uuid);
1119
1120 if (u &&
1121 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1122 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1123 memcpy(u->uuid, invalid_uuid, 16);
1124 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1125 u = NULL;
1126 }
1127
1128 if (!u) {
1129 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1130 pr_err("Couldn't find uuid for %s in set",
1131 dc->backing_dev_name);
1132 return -ENOENT;
1133 }
1134
1135 u = uuid_find_empty(c);
1136 if (!u) {
1137 pr_err("Not caching %s, no room for UUID",
1138 dc->backing_dev_name);
1139 return -EINVAL;
1140 }
1141 }
1142
1143 /*
1144 * Deadlocks since we're called via sysfs...
1145 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1146 */
1147
1148 if (bch_is_zero(u->uuid, 16)) {
1149 struct closure cl;
1150
1151 closure_init_stack(&cl);
1152
1153 memcpy(u->uuid, dc->sb.uuid, 16);
1154 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1155 u->first_reg = u->last_reg = rtime;
1156 bch_uuid_write(c);
1157
1158 memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1159 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1160
1161 bch_write_bdev_super(dc, &cl);
1162 closure_sync(&cl);
1163 } else {
1164 u->last_reg = rtime;
1165 bch_uuid_write(c);
1166 }
1167
1168 bcache_device_attach(&dc->disk, c, u - c->uuids);
1169 list_move(&dc->list, &c->cached_devs);
1170 calc_cached_dev_sectors(c);
1171
1172 /*
1173 * dc->c must be set before dc->count != 0 - paired with the mb in
1174 * cached_dev_get()
1175 */
1176 smp_wmb();
1177 refcount_set(&dc->count, 1);
1178
1179 /* Block writeback thread, but spawn it */
1180 down_write(&dc->writeback_lock);
1181 if (bch_cached_dev_writeback_start(dc)) {
1182 up_write(&dc->writeback_lock);
1183 pr_err("Couldn't start writeback facilities for %s",
1184 dc->disk.disk->disk_name);
1185 return -ENOMEM;
1186 }
1187
1188 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1189 atomic_set(&dc->has_dirty, 1);
1190 bch_writeback_queue(dc);
1191 }
1192
1193 bch_sectors_dirty_init(&dc->disk);
1194
1195 ret = bch_cached_dev_run(dc);
1196 if (ret && (ret != -EBUSY)) {
1197 up_write(&dc->writeback_lock);
1198 /*
1199 * bch_register_lock is held, bcache_device_stop() is not
1200 * able to be directly called. The kthread and kworker
1201 * created previously in bch_cached_dev_writeback_start()
1202 * have to be stopped manually here.
1203 */
1204 kthread_stop(dc->writeback_thread);
1205 cancel_writeback_rate_update_dwork(dc);
1206 pr_err("Couldn't run cached device %s",
1207 dc->backing_dev_name);
1208 return ret;
1209 }
1210
1211 bcache_device_link(&dc->disk, c, "bdev");
1212 atomic_inc(&c->attached_dev_nr);
1213
1214 /* Allow the writeback thread to proceed */
1215 up_write(&dc->writeback_lock);
1216
1217 pr_info("Caching %s as %s on set %pU",
1218 dc->backing_dev_name,
1219 dc->disk.disk->disk_name,
1220 dc->disk.c->sb.set_uuid);
1221 return 0;
1222}
1223
1224/* when dc->disk.kobj released */
1225void bch_cached_dev_release(struct kobject *kobj)
1226{
1227 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1228 disk.kobj);
1229 kfree(dc);
1230 module_put(THIS_MODULE);
1231}
1232
1233static void cached_dev_free(struct closure *cl)
1234{
1235 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1236
1237 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1238 cancel_writeback_rate_update_dwork(dc);
1239
1240 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1241 kthread_stop(dc->writeback_thread);
1242 if (!IS_ERR_OR_NULL(dc->status_update_thread))
1243 kthread_stop(dc->status_update_thread);
1244
1245 mutex_lock(&bch_register_lock);
1246
1247 if (atomic_read(&dc->running))
1248 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1249 bcache_device_free(&dc->disk);
1250 list_del(&dc->list);
1251
1252 mutex_unlock(&bch_register_lock);
1253
1254 if (!IS_ERR_OR_NULL(dc->bdev))
1255 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1256
1257 wake_up(&unregister_wait);
1258
1259 kobject_put(&dc->disk.kobj);
1260}
1261
1262static void cached_dev_flush(struct closure *cl)
1263{
1264 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1265 struct bcache_device *d = &dc->disk;
1266
1267 mutex_lock(&bch_register_lock);
1268 bcache_device_unlink(d);
1269 mutex_unlock(&bch_register_lock);
1270
1271 bch_cache_accounting_destroy(&dc->accounting);
1272 kobject_del(&d->kobj);
1273
1274 continue_at(cl, cached_dev_free, system_wq);
1275}
1276
1277static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1278{
1279 int ret;
1280 struct io *io;
1281 struct request_queue *q = bdev_get_queue(dc->bdev);
1282
1283 __module_get(THIS_MODULE);
1284 INIT_LIST_HEAD(&dc->list);
1285 closure_init(&dc->disk.cl, NULL);
1286 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1287 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1288 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1289 sema_init(&dc->sb_write_mutex, 1);
1290 INIT_LIST_HEAD(&dc->io_lru);
1291 spin_lock_init(&dc->io_lock);
1292 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1293
1294 dc->sequential_cutoff = 4 << 20;
1295
1296 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1297 list_add(&io->lru, &dc->io_lru);
1298 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1299 }
1300
1301 dc->disk.stripe_size = q->limits.io_opt >> 9;
1302
1303 if (dc->disk.stripe_size)
1304 dc->partial_stripes_expensive =
1305 q->limits.raid_partial_stripes_expensive;
1306
1307 ret = bcache_device_init(&dc->disk, block_size,
1308 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1309 if (ret)
1310 return ret;
1311
1312 dc->disk.disk->queue->backing_dev_info->ra_pages =
1313 max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1314 q->backing_dev_info->ra_pages);
1315
1316 atomic_set(&dc->io_errors, 0);
1317 dc->io_disable = false;
1318 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1319 /* default to auto */
1320 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1321
1322 bch_cached_dev_request_init(dc);
1323 bch_cached_dev_writeback_init(dc);
1324 return 0;
1325}
1326
1327/* Cached device - bcache superblock */
1328
1329static int register_bdev(struct cache_sb *sb, struct page *sb_page,
1330 struct block_device *bdev,
1331 struct cached_dev *dc)
1332{
1333 const char *err = "cannot allocate memory";
1334 struct cache_set *c;
1335 int ret = -ENOMEM;
1336
1337 bdevname(bdev, dc->backing_dev_name);
1338 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1339 dc->bdev = bdev;
1340 dc->bdev->bd_holder = dc;
1341
1342 bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1343 bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1344 get_page(sb_page);
1345
1346
1347 if (cached_dev_init(dc, sb->block_size << 9))
1348 goto err;
1349
1350 err = "error creating kobject";
1351 if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1352 "bcache"))
1353 goto err;
1354 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1355 goto err;
1356
1357 pr_info("registered backing device %s", dc->backing_dev_name);
1358
1359 list_add(&dc->list, &uncached_devices);
1360 /* attach to a matched cache set if it exists */
1361 list_for_each_entry(c, &bch_cache_sets, list)
1362 bch_cached_dev_attach(dc, c, NULL);
1363
1364 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1365 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1366 err = "failed to run cached device";
1367 ret = bch_cached_dev_run(dc);
1368 if (ret)
1369 goto err;
1370 }
1371
1372 return 0;
1373err:
1374 pr_notice("error %s: %s", dc->backing_dev_name, err);
1375 bcache_device_stop(&dc->disk);
1376 return ret;
1377}
1378
1379/* Flash only volumes */
1380
1381/* When d->kobj released */
1382void bch_flash_dev_release(struct kobject *kobj)
1383{
1384 struct bcache_device *d = container_of(kobj, struct bcache_device,
1385 kobj);
1386 kfree(d);
1387}
1388
1389static void flash_dev_free(struct closure *cl)
1390{
1391 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1392
1393 mutex_lock(&bch_register_lock);
1394 atomic_long_sub(bcache_dev_sectors_dirty(d),
1395 &d->c->flash_dev_dirty_sectors);
1396 bcache_device_free(d);
1397 mutex_unlock(&bch_register_lock);
1398 kobject_put(&d->kobj);
1399}
1400
1401static void flash_dev_flush(struct closure *cl)
1402{
1403 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1404
1405 mutex_lock(&bch_register_lock);
1406 bcache_device_unlink(d);
1407 mutex_unlock(&bch_register_lock);
1408 kobject_del(&d->kobj);
1409 continue_at(cl, flash_dev_free, system_wq);
1410}
1411
1412static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1413{
1414 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1415 GFP_KERNEL);
1416 if (!d)
1417 return -ENOMEM;
1418
1419 closure_init(&d->cl, NULL);
1420 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1421
1422 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1423
1424 if (bcache_device_init(d, block_bytes(c), u->sectors))
1425 goto err;
1426
1427 bcache_device_attach(d, c, u - c->uuids);
1428 bch_sectors_dirty_init(d);
1429 bch_flash_dev_request_init(d);
1430 add_disk(d->disk);
1431
1432 if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1433 goto err;
1434
1435 bcache_device_link(d, c, "volume");
1436
1437 return 0;
1438err:
1439 kobject_put(&d->kobj);
1440 return -ENOMEM;
1441}
1442
1443static int flash_devs_run(struct cache_set *c)
1444{
1445 int ret = 0;
1446 struct uuid_entry *u;
1447
1448 for (u = c->uuids;
1449 u < c->uuids + c->nr_uuids && !ret;
1450 u++)
1451 if (UUID_FLASH_ONLY(u))
1452 ret = flash_dev_run(c, u);
1453
1454 return ret;
1455}
1456
1457int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1458{
1459 struct uuid_entry *u;
1460
1461 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1462 return -EINTR;
1463
1464 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1465 return -EPERM;
1466
1467 u = uuid_find_empty(c);
1468 if (!u) {
1469 pr_err("Can't create volume, no room for UUID");
1470 return -EINVAL;
1471 }
1472
1473 get_random_bytes(u->uuid, 16);
1474 memset(u->label, 0, 32);
1475 u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1476
1477 SET_UUID_FLASH_ONLY(u, 1);
1478 u->sectors = size >> 9;
1479
1480 bch_uuid_write(c);
1481
1482 return flash_dev_run(c, u);
1483}
1484
1485bool bch_cached_dev_error(struct cached_dev *dc)
1486{
1487 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1488 return false;
1489
1490 dc->io_disable = true;
1491 /* make others know io_disable is true earlier */
1492 smp_mb();
1493
1494 pr_err("stop %s: too many IO errors on backing device %s\n",
1495 dc->disk.disk->disk_name, dc->backing_dev_name);
1496
1497 bcache_device_stop(&dc->disk);
1498 return true;
1499}
1500
1501/* Cache set */
1502
1503__printf(2, 3)
1504bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1505{
1506 va_list args;
1507
1508 if (c->on_error != ON_ERROR_PANIC &&
1509 test_bit(CACHE_SET_STOPPING, &c->flags))
1510 return false;
1511
1512 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1513 pr_info("CACHE_SET_IO_DISABLE already set");
1514
1515 /*
1516 * XXX: we can be called from atomic context
1517 * acquire_console_sem();
1518 */
1519
1520 pr_err("bcache: error on %pU: ", c->sb.set_uuid);
1521
1522 va_start(args, fmt);
1523 vprintk(fmt, args);
1524 va_end(args);
1525
1526 pr_err(", disabling caching\n");
1527
1528 if (c->on_error == ON_ERROR_PANIC)
1529 panic("panic forced after error\n");
1530
1531 bch_cache_set_unregister(c);
1532 return true;
1533}
1534
1535/* When c->kobj released */
1536void bch_cache_set_release(struct kobject *kobj)
1537{
1538 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1539
1540 kfree(c);
1541 module_put(THIS_MODULE);
1542}
1543
1544static void cache_set_free(struct closure *cl)
1545{
1546 struct cache_set *c = container_of(cl, struct cache_set, cl);
1547 struct cache *ca;
1548 unsigned int i;
1549
1550 debugfs_remove(c->debug);
1551
1552 bch_open_buckets_free(c);
1553 bch_btree_cache_free(c);
1554 bch_journal_free(c);
1555
1556 mutex_lock(&bch_register_lock);
1557 for_each_cache(ca, c, i)
1558 if (ca) {
1559 ca->set = NULL;
1560 c->cache[ca->sb.nr_this_dev] = NULL;
1561 kobject_put(&ca->kobj);
1562 }
1563
1564 bch_bset_sort_state_free(&c->sort);
1565 free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1566
1567 if (c->moving_gc_wq)
1568 destroy_workqueue(c->moving_gc_wq);
1569 bioset_exit(&c->bio_split);
1570 mempool_exit(&c->fill_iter);
1571 mempool_exit(&c->bio_meta);
1572 mempool_exit(&c->search);
1573 kfree(c->devices);
1574
1575 list_del(&c->list);
1576 mutex_unlock(&bch_register_lock);
1577
1578 pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1579 wake_up(&unregister_wait);
1580
1581 closure_debug_destroy(&c->cl);
1582 kobject_put(&c->kobj);
1583}
1584
1585static void cache_set_flush(struct closure *cl)
1586{
1587 struct cache_set *c = container_of(cl, struct cache_set, caching);
1588 struct cache *ca;
1589 struct btree *b;
1590 unsigned int i;
1591
1592 bch_cache_accounting_destroy(&c->accounting);
1593
1594 kobject_put(&c->internal);
1595 kobject_del(&c->kobj);
1596
1597 if (!IS_ERR_OR_NULL(c->gc_thread))
1598 kthread_stop(c->gc_thread);
1599
1600 if (!IS_ERR_OR_NULL(c->root))
1601 list_add(&c->root->list, &c->btree_cache);
1602
1603 /*
1604 * Avoid flushing cached nodes if cache set is retiring
1605 * due to too many I/O errors detected.
1606 */
1607 if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1608 list_for_each_entry(b, &c->btree_cache, list) {
1609 mutex_lock(&b->write_lock);
1610 if (btree_node_dirty(b))
1611 __bch_btree_node_write(b, NULL);
1612 mutex_unlock(&b->write_lock);
1613 }
1614
1615 for_each_cache(ca, c, i)
1616 if (ca->alloc_thread)
1617 kthread_stop(ca->alloc_thread);
1618
1619 if (c->journal.cur) {
1620 cancel_delayed_work_sync(&c->journal.work);
1621 /* flush last journal entry if needed */
1622 c->journal.work.work.func(&c->journal.work.work);
1623 }
1624
1625 closure_return(cl);
1626}
1627
1628/*
1629 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1630 * cache set is unregistering due to too many I/O errors. In this condition,
1631 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1632 * value and whether the broken cache has dirty data:
1633 *
1634 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1635 * BCH_CACHED_STOP_AUTO 0 NO
1636 * BCH_CACHED_STOP_AUTO 1 YES
1637 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1638 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1639 *
1640 * The expected behavior is, if stop_when_cache_set_failed is configured to
1641 * "auto" via sysfs interface, the bcache device will not be stopped if the
1642 * backing device is clean on the broken cache device.
1643 */
1644static void conditional_stop_bcache_device(struct cache_set *c,
1645 struct bcache_device *d,
1646 struct cached_dev *dc)
1647{
1648 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1649 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1650 d->disk->disk_name, c->sb.set_uuid);
1651 bcache_device_stop(d);
1652 } else if (atomic_read(&dc->has_dirty)) {
1653 /*
1654 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1655 * and dc->has_dirty == 1
1656 */
1657 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1658 d->disk->disk_name);
1659 /*
1660 * There might be a small time gap that cache set is
1661 * released but bcache device is not. Inside this time
1662 * gap, regular I/O requests will directly go into
1663 * backing device as no cache set attached to. This
1664 * behavior may also introduce potential inconsistence
1665 * data in writeback mode while cache is dirty.
1666 * Therefore before calling bcache_device_stop() due
1667 * to a broken cache device, dc->io_disable should be
1668 * explicitly set to true.
1669 */
1670 dc->io_disable = true;
1671 /* make others know io_disable is true earlier */
1672 smp_mb();
1673 bcache_device_stop(d);
1674 } else {
1675 /*
1676 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1677 * and dc->has_dirty == 0
1678 */
1679 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1680 d->disk->disk_name);
1681 }
1682}
1683
1684static void __cache_set_unregister(struct closure *cl)
1685{
1686 struct cache_set *c = container_of(cl, struct cache_set, caching);
1687 struct cached_dev *dc;
1688 struct bcache_device *d;
1689 size_t i;
1690
1691 mutex_lock(&bch_register_lock);
1692
1693 for (i = 0; i < c->devices_max_used; i++) {
1694 d = c->devices[i];
1695 if (!d)
1696 continue;
1697
1698 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1699 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1700 dc = container_of(d, struct cached_dev, disk);
1701 bch_cached_dev_detach(dc);
1702 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1703 conditional_stop_bcache_device(c, d, dc);
1704 } else {
1705 bcache_device_stop(d);
1706 }
1707 }
1708
1709 mutex_unlock(&bch_register_lock);
1710
1711 continue_at(cl, cache_set_flush, system_wq);
1712}
1713
1714void bch_cache_set_stop(struct cache_set *c)
1715{
1716 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1717 /* closure_fn set to __cache_set_unregister() */
1718 closure_queue(&c->caching);
1719}
1720
1721void bch_cache_set_unregister(struct cache_set *c)
1722{
1723 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1724 bch_cache_set_stop(c);
1725}
1726
1727#define alloc_bucket_pages(gfp, c) \
1728 ((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1729
1730struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1731{
1732 int iter_size;
1733 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1734
1735 if (!c)
1736 return NULL;
1737
1738 __module_get(THIS_MODULE);
1739 closure_init(&c->cl, NULL);
1740 set_closure_fn(&c->cl, cache_set_free, system_wq);
1741
1742 closure_init(&c->caching, &c->cl);
1743 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1744
1745 /* Maybe create continue_at_noreturn() and use it here? */
1746 closure_set_stopped(&c->cl);
1747 closure_put(&c->cl);
1748
1749 kobject_init(&c->kobj, &bch_cache_set_ktype);
1750 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1751
1752 bch_cache_accounting_init(&c->accounting, &c->cl);
1753
1754 memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1755 c->sb.block_size = sb->block_size;
1756 c->sb.bucket_size = sb->bucket_size;
1757 c->sb.nr_in_set = sb->nr_in_set;
1758 c->sb.last_mount = sb->last_mount;
1759 c->bucket_bits = ilog2(sb->bucket_size);
1760 c->block_bits = ilog2(sb->block_size);
1761 c->nr_uuids = bucket_bytes(c) / sizeof(struct uuid_entry);
1762 c->devices_max_used = 0;
1763 atomic_set(&c->attached_dev_nr, 0);
1764 c->btree_pages = bucket_pages(c);
1765 if (c->btree_pages > BTREE_MAX_PAGES)
1766 c->btree_pages = max_t(int, c->btree_pages / 4,
1767 BTREE_MAX_PAGES);
1768
1769 sema_init(&c->sb_write_mutex, 1);
1770 mutex_init(&c->bucket_lock);
1771 init_waitqueue_head(&c->btree_cache_wait);
1772 init_waitqueue_head(&c->bucket_wait);
1773 init_waitqueue_head(&c->gc_wait);
1774 sema_init(&c->uuid_write_mutex, 1);
1775
1776 spin_lock_init(&c->btree_gc_time.lock);
1777 spin_lock_init(&c->btree_split_time.lock);
1778 spin_lock_init(&c->btree_read_time.lock);
1779
1780 bch_moving_init_cache_set(c);
1781
1782 INIT_LIST_HEAD(&c->list);
1783 INIT_LIST_HEAD(&c->cached_devs);
1784 INIT_LIST_HEAD(&c->btree_cache);
1785 INIT_LIST_HEAD(&c->btree_cache_freeable);
1786 INIT_LIST_HEAD(&c->btree_cache_freed);
1787 INIT_LIST_HEAD(&c->data_buckets);
1788
1789 iter_size = (sb->bucket_size / sb->block_size + 1) *
1790 sizeof(struct btree_iter_set);
1791
1792 if (!(c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL)) ||
1793 mempool_init_slab_pool(&c->search, 32, bch_search_cache) ||
1794 mempool_init_kmalloc_pool(&c->bio_meta, 2,
1795 sizeof(struct bbio) + sizeof(struct bio_vec) *
1796 bucket_pages(c)) ||
1797 mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
1798 bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1799 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER) ||
1800 !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1801 !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1802 WQ_MEM_RECLAIM, 0)) ||
1803 bch_journal_alloc(c) ||
1804 bch_btree_cache_alloc(c) ||
1805 bch_open_buckets_alloc(c) ||
1806 bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1807 goto err;
1808
1809 c->congested_read_threshold_us = 2000;
1810 c->congested_write_threshold_us = 20000;
1811 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1812 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1813
1814 return c;
1815err:
1816 bch_cache_set_unregister(c);
1817 return NULL;
1818}
1819
1820static int run_cache_set(struct cache_set *c)
1821{
1822 const char *err = "cannot allocate memory";
1823 struct cached_dev *dc, *t;
1824 struct cache *ca;
1825 struct closure cl;
1826 unsigned int i;
1827 LIST_HEAD(journal);
1828 struct journal_replay *l;
1829
1830 closure_init_stack(&cl);
1831
1832 for_each_cache(ca, c, i)
1833 c->nbuckets += ca->sb.nbuckets;
1834 set_gc_sectors(c);
1835
1836 if (CACHE_SYNC(&c->sb)) {
1837 struct bkey *k;
1838 struct jset *j;
1839
1840 err = "cannot allocate memory for journal";
1841 if (bch_journal_read(c, &journal))
1842 goto err;
1843
1844 pr_debug("btree_journal_read() done");
1845
1846 err = "no journal entries found";
1847 if (list_empty(&journal))
1848 goto err;
1849
1850 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1851
1852 err = "IO error reading priorities";
1853 for_each_cache(ca, c, i)
1854 prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1855
1856 /*
1857 * If prio_read() fails it'll call cache_set_error and we'll
1858 * tear everything down right away, but if we perhaps checked
1859 * sooner we could avoid journal replay.
1860 */
1861
1862 k = &j->btree_root;
1863
1864 err = "bad btree root";
1865 if (__bch_btree_ptr_invalid(c, k))
1866 goto err;
1867
1868 err = "error reading btree root";
1869 c->root = bch_btree_node_get(c, NULL, k,
1870 j->btree_level,
1871 true, NULL);
1872 if (IS_ERR_OR_NULL(c->root))
1873 goto err;
1874
1875 list_del_init(&c->root->list);
1876 rw_unlock(true, c->root);
1877
1878 err = uuid_read(c, j, &cl);
1879 if (err)
1880 goto err;
1881
1882 err = "error in recovery";
1883 if (bch_btree_check(c))
1884 goto err;
1885
1886 /*
1887 * bch_btree_check() may occupy too much system memory which
1888 * has negative effects to user space application (e.g. data
1889 * base) performance. Shrink the mca cache memory proactively
1890 * here to avoid competing memory with user space workloads..
1891 */
1892 if (!c->shrinker_disabled) {
1893 struct shrink_control sc;
1894
1895 sc.gfp_mask = GFP_KERNEL;
1896 sc.nr_to_scan = c->btree_cache_used * c->btree_pages;
1897 /* first run to clear b->accessed tag */
1898 c->shrink.scan_objects(&c->shrink, &sc);
1899 /* second run to reap non-accessed nodes */
1900 c->shrink.scan_objects(&c->shrink, &sc);
1901 }
1902
1903 bch_journal_mark(c, &journal);
1904 bch_initial_gc_finish(c);
1905 pr_debug("btree_check() done");
1906
1907 /*
1908 * bcache_journal_next() can't happen sooner, or
1909 * btree_gc_finish() will give spurious errors about last_gc >
1910 * gc_gen - this is a hack but oh well.
1911 */
1912 bch_journal_next(&c->journal);
1913
1914 err = "error starting allocator thread";
1915 for_each_cache(ca, c, i)
1916 if (bch_cache_allocator_start(ca))
1917 goto err;
1918
1919 /*
1920 * First place it's safe to allocate: btree_check() and
1921 * btree_gc_finish() have to run before we have buckets to
1922 * allocate, and bch_bucket_alloc_set() might cause a journal
1923 * entry to be written so bcache_journal_next() has to be called
1924 * first.
1925 *
1926 * If the uuids were in the old format we have to rewrite them
1927 * before the next journal entry is written:
1928 */
1929 if (j->version < BCACHE_JSET_VERSION_UUID)
1930 __uuid_write(c);
1931
1932 err = "bcache: replay journal failed";
1933 if (bch_journal_replay(c, &journal))
1934 goto err;
1935 } else {
1936 pr_notice("invalidating existing data");
1937
1938 for_each_cache(ca, c, i) {
1939 unsigned int j;
1940
1941 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1942 2, SB_JOURNAL_BUCKETS);
1943
1944 for (j = 0; j < ca->sb.keys; j++)
1945 ca->sb.d[j] = ca->sb.first_bucket + j;
1946 }
1947
1948 bch_initial_gc_finish(c);
1949
1950 err = "error starting allocator thread";
1951 for_each_cache(ca, c, i)
1952 if (bch_cache_allocator_start(ca))
1953 goto err;
1954
1955 mutex_lock(&c->bucket_lock);
1956 for_each_cache(ca, c, i)
1957 bch_prio_write(ca);
1958 mutex_unlock(&c->bucket_lock);
1959
1960 err = "cannot allocate new UUID bucket";
1961 if (__uuid_write(c))
1962 goto err;
1963
1964 err = "cannot allocate new btree root";
1965 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1966 if (IS_ERR_OR_NULL(c->root))
1967 goto err;
1968
1969 mutex_lock(&c->root->write_lock);
1970 bkey_copy_key(&c->root->key, &MAX_KEY);
1971 bch_btree_node_write(c->root, &cl);
1972 mutex_unlock(&c->root->write_lock);
1973
1974 bch_btree_set_root(c->root);
1975 rw_unlock(true, c->root);
1976
1977 /*
1978 * We don't want to write the first journal entry until
1979 * everything is set up - fortunately journal entries won't be
1980 * written until the SET_CACHE_SYNC() here:
1981 */
1982 SET_CACHE_SYNC(&c->sb, true);
1983
1984 bch_journal_next(&c->journal);
1985 bch_journal_meta(c, &cl);
1986 }
1987
1988 err = "error starting gc thread";
1989 if (bch_gc_thread_start(c))
1990 goto err;
1991
1992 closure_sync(&cl);
1993 c->sb.last_mount = (u32)ktime_get_real_seconds();
1994 bcache_write_super(c);
1995
1996 list_for_each_entry_safe(dc, t, &uncached_devices, list)
1997 bch_cached_dev_attach(dc, c, NULL);
1998
1999 flash_devs_run(c);
2000
2001 set_bit(CACHE_SET_RUNNING, &c->flags);
2002 return 0;
2003err:
2004 while (!list_empty(&journal)) {
2005 l = list_first_entry(&journal, struct journal_replay, list);
2006 list_del(&l->list);
2007 kfree(l);
2008 }
2009
2010 closure_sync(&cl);
2011
2012 bch_cache_set_error(c, "%s", err);
2013
2014 return -EIO;
2015}
2016
2017static bool can_attach_cache(struct cache *ca, struct cache_set *c)
2018{
2019 return ca->sb.block_size == c->sb.block_size &&
2020 ca->sb.bucket_size == c->sb.bucket_size &&
2021 ca->sb.nr_in_set == c->sb.nr_in_set;
2022}
2023
2024static const char *register_cache_set(struct cache *ca)
2025{
2026 char buf[12];
2027 const char *err = "cannot allocate memory";
2028 struct cache_set *c;
2029
2030 list_for_each_entry(c, &bch_cache_sets, list)
2031 if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
2032 if (c->cache[ca->sb.nr_this_dev])
2033 return "duplicate cache set member";
2034
2035 if (!can_attach_cache(ca, c))
2036 return "cache sb does not match set";
2037
2038 if (!CACHE_SYNC(&ca->sb))
2039 SET_CACHE_SYNC(&c->sb, false);
2040
2041 goto found;
2042 }
2043
2044 c = bch_cache_set_alloc(&ca->sb);
2045 if (!c)
2046 return err;
2047
2048 err = "error creating kobject";
2049 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
2050 kobject_add(&c->internal, &c->kobj, "internal"))
2051 goto err;
2052
2053 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2054 goto err;
2055
2056 bch_debug_init_cache_set(c);
2057
2058 list_add(&c->list, &bch_cache_sets);
2059found:
2060 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2061 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2062 sysfs_create_link(&c->kobj, &ca->kobj, buf))
2063 goto err;
2064
2065 if (ca->sb.seq > c->sb.seq) {
2066 c->sb.version = ca->sb.version;
2067 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2068 c->sb.flags = ca->sb.flags;
2069 c->sb.seq = ca->sb.seq;
2070 pr_debug("set version = %llu", c->sb.version);
2071 }
2072
2073 kobject_get(&ca->kobj);
2074 ca->set = c;
2075 ca->set->cache[ca->sb.nr_this_dev] = ca;
2076 c->cache_by_alloc[c->caches_loaded++] = ca;
2077
2078 if (c->caches_loaded == c->sb.nr_in_set) {
2079 err = "failed to run cache set";
2080 if (run_cache_set(c) < 0)
2081 goto err;
2082 }
2083
2084 return NULL;
2085err:
2086 bch_cache_set_unregister(c);
2087 return err;
2088}
2089
2090/* Cache device */
2091
2092/* When ca->kobj released */
2093void bch_cache_release(struct kobject *kobj)
2094{
2095 struct cache *ca = container_of(kobj, struct cache, kobj);
2096 unsigned int i;
2097
2098 if (ca->set) {
2099 BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2100 ca->set->cache[ca->sb.nr_this_dev] = NULL;
2101 }
2102
2103 free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2104 kfree(ca->prio_buckets);
2105 vfree(ca->buckets);
2106
2107 free_heap(&ca->heap);
2108 free_fifo(&ca->free_inc);
2109
2110 for (i = 0; i < RESERVE_NR; i++)
2111 free_fifo(&ca->free[i]);
2112
2113 if (ca->sb_bio.bi_inline_vecs[0].bv_page)
2114 put_page(bio_first_page_all(&ca->sb_bio));
2115
2116 if (!IS_ERR_OR_NULL(ca->bdev))
2117 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2118
2119 kfree(ca);
2120 module_put(THIS_MODULE);
2121}
2122
2123static int cache_alloc(struct cache *ca)
2124{
2125 size_t free;
2126 size_t btree_buckets;
2127 struct bucket *b;
2128 int ret = -ENOMEM;
2129 const char *err = NULL;
2130
2131 __module_get(THIS_MODULE);
2132 kobject_init(&ca->kobj, &bch_cache_ktype);
2133
2134 bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2135
2136 /*
2137 * when ca->sb.njournal_buckets is not zero, journal exists,
2138 * and in bch_journal_replay(), tree node may split,
2139 * so bucket of RESERVE_BTREE type is needed,
2140 * the worst situation is all journal buckets are valid journal,
2141 * and all the keys need to replay,
2142 * so the number of RESERVE_BTREE type buckets should be as much
2143 * as journal buckets
2144 */
2145 btree_buckets = ca->sb.njournal_buckets ?: 8;
2146 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2147 if (!free) {
2148 ret = -EPERM;
2149 err = "ca->sb.nbuckets is too small";
2150 goto err_free;
2151 }
2152
2153 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2154 GFP_KERNEL)) {
2155 err = "ca->free[RESERVE_BTREE] alloc failed";
2156 goto err_btree_alloc;
2157 }
2158
2159 if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2160 GFP_KERNEL)) {
2161 err = "ca->free[RESERVE_PRIO] alloc failed";
2162 goto err_prio_alloc;
2163 }
2164
2165 if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2166 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2167 goto err_movinggc_alloc;
2168 }
2169
2170 if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2171 err = "ca->free[RESERVE_NONE] alloc failed";
2172 goto err_none_alloc;
2173 }
2174
2175 if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2176 err = "ca->free_inc alloc failed";
2177 goto err_free_inc_alloc;
2178 }
2179
2180 if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2181 err = "ca->heap alloc failed";
2182 goto err_heap_alloc;
2183 }
2184
2185 ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2186 ca->sb.nbuckets));
2187 if (!ca->buckets) {
2188 err = "ca->buckets alloc failed";
2189 goto err_buckets_alloc;
2190 }
2191
2192 ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2193 prio_buckets(ca), 2),
2194 GFP_KERNEL);
2195 if (!ca->prio_buckets) {
2196 err = "ca->prio_buckets alloc failed";
2197 goto err_prio_buckets_alloc;
2198 }
2199
2200 ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca);
2201 if (!ca->disk_buckets) {
2202 err = "ca->disk_buckets alloc failed";
2203 goto err_disk_buckets_alloc;
2204 }
2205
2206 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2207
2208 for_each_bucket(b, ca)
2209 atomic_set(&b->pin, 0);
2210 return 0;
2211
2212err_disk_buckets_alloc:
2213 kfree(ca->prio_buckets);
2214err_prio_buckets_alloc:
2215 vfree(ca->buckets);
2216err_buckets_alloc:
2217 free_heap(&ca->heap);
2218err_heap_alloc:
2219 free_fifo(&ca->free_inc);
2220err_free_inc_alloc:
2221 free_fifo(&ca->free[RESERVE_NONE]);
2222err_none_alloc:
2223 free_fifo(&ca->free[RESERVE_MOVINGGC]);
2224err_movinggc_alloc:
2225 free_fifo(&ca->free[RESERVE_PRIO]);
2226err_prio_alloc:
2227 free_fifo(&ca->free[RESERVE_BTREE]);
2228err_btree_alloc:
2229err_free:
2230 module_put(THIS_MODULE);
2231 if (err)
2232 pr_notice("error %s: %s", ca->cache_dev_name, err);
2233 return ret;
2234}
2235
2236static int register_cache(struct cache_sb *sb, struct page *sb_page,
2237 struct block_device *bdev, struct cache *ca)
2238{
2239 const char *err = NULL; /* must be set for any error case */
2240 int ret = 0;
2241
2242 bdevname(bdev, ca->cache_dev_name);
2243 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2244 ca->bdev = bdev;
2245 ca->bdev->bd_holder = ca;
2246
2247 bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2248 bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2249 get_page(sb_page);
2250
2251 if (blk_queue_discard(bdev_get_queue(bdev)))
2252 ca->discard = CACHE_DISCARD(&ca->sb);
2253
2254 ret = cache_alloc(ca);
2255 if (ret != 0) {
2256 /*
2257 * If we failed here, it means ca->kobj is not initialized yet,
2258 * kobject_put() won't be called and there is no chance to
2259 * call blkdev_put() to bdev in bch_cache_release(). So we
2260 * explicitly call blkdev_put() here.
2261 */
2262 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2263 if (ret == -ENOMEM)
2264 err = "cache_alloc(): -ENOMEM";
2265 else if (ret == -EPERM)
2266 err = "cache_alloc(): cache device is too small";
2267 else
2268 err = "cache_alloc(): unknown error";
2269 goto err;
2270 }
2271
2272 if (kobject_add(&ca->kobj,
2273 &part_to_dev(bdev->bd_part)->kobj,
2274 "bcache")) {
2275 err = "error calling kobject_add";
2276 ret = -ENOMEM;
2277 goto out;
2278 }
2279
2280 mutex_lock(&bch_register_lock);
2281 err = register_cache_set(ca);
2282 mutex_unlock(&bch_register_lock);
2283
2284 if (err) {
2285 ret = -ENODEV;
2286 goto out;
2287 }
2288
2289 pr_info("registered cache device %s", ca->cache_dev_name);
2290
2291out:
2292 kobject_put(&ca->kobj);
2293
2294err:
2295 if (err)
2296 pr_notice("error %s: %s", ca->cache_dev_name, err);
2297
2298 return ret;
2299}
2300
2301/* Global interfaces/init */
2302
2303static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2304 const char *buffer, size_t size);
2305static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2306 struct kobj_attribute *attr,
2307 const char *buffer, size_t size);
2308
2309kobj_attribute_write(register, register_bcache);
2310kobj_attribute_write(register_quiet, register_bcache);
2311kobj_attribute_write(pendings_cleanup, bch_pending_bdevs_cleanup);
2312
2313static bool bch_is_open_backing(struct block_device *bdev)
2314{
2315 struct cache_set *c, *tc;
2316 struct cached_dev *dc, *t;
2317
2318 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2319 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2320 if (dc->bdev == bdev)
2321 return true;
2322 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2323 if (dc->bdev == bdev)
2324 return true;
2325 return false;
2326}
2327
2328static bool bch_is_open_cache(struct block_device *bdev)
2329{
2330 struct cache_set *c, *tc;
2331 struct cache *ca;
2332 unsigned int i;
2333
2334 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2335 for_each_cache(ca, c, i)
2336 if (ca->bdev == bdev)
2337 return true;
2338 return false;
2339}
2340
2341static bool bch_is_open(struct block_device *bdev)
2342{
2343 return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2344}
2345
2346static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2347 const char *buffer, size_t size)
2348{
2349 ssize_t ret = -EINVAL;
2350 const char *err = "cannot allocate memory";
2351 char *path = NULL;
2352 struct cache_sb *sb = NULL;
2353 struct block_device *bdev = NULL;
2354 struct page *sb_page = NULL;
2355
2356 if (!try_module_get(THIS_MODULE))
2357 return -EBUSY;
2358
2359 /* For latest state of bcache_is_reboot */
2360 smp_mb();
2361 if (bcache_is_reboot)
2362 return -EBUSY;
2363
2364 path = kstrndup(buffer, size, GFP_KERNEL);
2365 if (!path)
2366 goto err;
2367
2368 sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2369 if (!sb)
2370 goto err;
2371
2372 err = "failed to open device";
2373 bdev = blkdev_get_by_path(strim(path),
2374 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2375 sb);
2376 if (IS_ERR(bdev)) {
2377 if (bdev == ERR_PTR(-EBUSY)) {
2378 bdev = lookup_bdev(strim(path));
2379 mutex_lock(&bch_register_lock);
2380 if (!IS_ERR(bdev) && bch_is_open(bdev))
2381 err = "device already registered";
2382 else
2383 err = "device busy";
2384 mutex_unlock(&bch_register_lock);
2385 if (!IS_ERR(bdev))
2386 bdput(bdev);
2387 if (attr == &ksysfs_register_quiet)
2388 goto quiet_out;
2389 }
2390 goto err;
2391 }
2392
2393 err = "failed to set blocksize";
2394 if (set_blocksize(bdev, 4096))
2395 goto err_close;
2396
2397 err = read_super(sb, bdev, &sb_page);
2398 if (err)
2399 goto err_close;
2400
2401 err = "failed to register device";
2402 if (SB_IS_BDEV(sb)) {
2403 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2404
2405 if (!dc)
2406 goto err_close;
2407
2408 mutex_lock(&bch_register_lock);
2409 ret = register_bdev(sb, sb_page, bdev, dc);
2410 mutex_unlock(&bch_register_lock);
2411 /* blkdev_put() will be called in cached_dev_free() */
2412 if (ret < 0)
2413 goto err;
2414 } else {
2415 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2416
2417 if (!ca)
2418 goto err_close;
2419
2420 /* blkdev_put() will be called in bch_cache_release() */
2421 if (register_cache(sb, sb_page, bdev, ca) != 0)
2422 goto err;
2423 }
2424quiet_out:
2425 ret = size;
2426out:
2427 if (sb_page)
2428 put_page(sb_page);
2429 kfree(sb);
2430 kfree(path);
2431 module_put(THIS_MODULE);
2432 return ret;
2433
2434err_close:
2435 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2436err:
2437 pr_info("error %s: %s", path, err);
2438 goto out;
2439}
2440
2441
2442struct pdev {
2443 struct list_head list;
2444 struct cached_dev *dc;
2445};
2446
2447static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2448 struct kobj_attribute *attr,
2449 const char *buffer,
2450 size_t size)
2451{
2452 LIST_HEAD(pending_devs);
2453 ssize_t ret = size;
2454 struct cached_dev *dc, *tdc;
2455 struct pdev *pdev, *tpdev;
2456 struct cache_set *c, *tc;
2457
2458 mutex_lock(&bch_register_lock);
2459 list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2460 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2461 if (!pdev)
2462 break;
2463 pdev->dc = dc;
2464 list_add(&pdev->list, &pending_devs);
2465 }
2466
2467 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2468 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2469 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2470 char *set_uuid = c->sb.uuid;
2471
2472 if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2473 list_del(&pdev->list);
2474 kfree(pdev);
2475 break;
2476 }
2477 }
2478 }
2479 mutex_unlock(&bch_register_lock);
2480
2481 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2482 pr_info("delete pdev %p", pdev);
2483 list_del(&pdev->list);
2484 bcache_device_stop(&pdev->dc->disk);
2485 kfree(pdev);
2486 }
2487
2488 return ret;
2489}
2490
2491static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2492{
2493 if (bcache_is_reboot)
2494 return NOTIFY_DONE;
2495
2496 if (code == SYS_DOWN ||
2497 code == SYS_HALT ||
2498 code == SYS_POWER_OFF) {
2499 DEFINE_WAIT(wait);
2500 unsigned long start = jiffies;
2501 bool stopped = false;
2502
2503 struct cache_set *c, *tc;
2504 struct cached_dev *dc, *tdc;
2505
2506 mutex_lock(&bch_register_lock);
2507
2508 if (bcache_is_reboot)
2509 goto out;
2510
2511 /* New registration is rejected since now */
2512 bcache_is_reboot = true;
2513 /*
2514 * Make registering caller (if there is) on other CPU
2515 * core know bcache_is_reboot set to true earlier
2516 */
2517 smp_mb();
2518
2519 if (list_empty(&bch_cache_sets) &&
2520 list_empty(&uncached_devices))
2521 goto out;
2522
2523 mutex_unlock(&bch_register_lock);
2524
2525 pr_info("Stopping all devices:");
2526
2527 /*
2528 * The reason bch_register_lock is not held to call
2529 * bch_cache_set_stop() and bcache_device_stop() is to
2530 * avoid potential deadlock during reboot, because cache
2531 * set or bcache device stopping process will acqurie
2532 * bch_register_lock too.
2533 *
2534 * We are safe here because bcache_is_reboot sets to
2535 * true already, register_bcache() will reject new
2536 * registration now. bcache_is_reboot also makes sure
2537 * bcache_reboot() won't be re-entered on by other thread,
2538 * so there is no race in following list iteration by
2539 * list_for_each_entry_safe().
2540 */
2541 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2542 bch_cache_set_stop(c);
2543
2544 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2545 bcache_device_stop(&dc->disk);
2546
2547
2548 /*
2549 * Give an early chance for other kthreads and
2550 * kworkers to stop themselves
2551 */
2552 schedule();
2553
2554 /* What's a condition variable? */
2555 while (1) {
2556 long timeout = start + 10 * HZ - jiffies;
2557
2558 mutex_lock(&bch_register_lock);
2559 stopped = list_empty(&bch_cache_sets) &&
2560 list_empty(&uncached_devices);
2561
2562 if (timeout < 0 || stopped)
2563 break;
2564
2565 prepare_to_wait(&unregister_wait, &wait,
2566 TASK_UNINTERRUPTIBLE);
2567
2568 mutex_unlock(&bch_register_lock);
2569 schedule_timeout(timeout);
2570 }
2571
2572 finish_wait(&unregister_wait, &wait);
2573
2574 if (stopped)
2575 pr_info("All devices stopped");
2576 else
2577 pr_notice("Timeout waiting for devices to be closed");
2578out:
2579 mutex_unlock(&bch_register_lock);
2580 }
2581
2582 return NOTIFY_DONE;
2583}
2584
2585static struct notifier_block reboot = {
2586 .notifier_call = bcache_reboot,
2587 .priority = INT_MAX, /* before any real devices */
2588};
2589
2590static void bcache_exit(void)
2591{
2592 bch_debug_exit();
2593 bch_request_exit();
2594 if (bcache_kobj)
2595 kobject_put(bcache_kobj);
2596 if (bcache_wq)
2597 destroy_workqueue(bcache_wq);
2598 if (bch_journal_wq)
2599 destroy_workqueue(bch_journal_wq);
2600
2601 if (bcache_major)
2602 unregister_blkdev(bcache_major, "bcache");
2603 unregister_reboot_notifier(&reboot);
2604 mutex_destroy(&bch_register_lock);
2605}
2606
2607/* Check and fixup module parameters */
2608static void check_module_parameters(void)
2609{
2610 if (bch_cutoff_writeback_sync == 0)
2611 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2612 else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2613 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u",
2614 bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2615 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2616 }
2617
2618 if (bch_cutoff_writeback == 0)
2619 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2620 else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2621 pr_warn("set bch_cutoff_writeback (%u) to max value %u",
2622 bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2623 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2624 }
2625
2626 if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2627 pr_warn("set bch_cutoff_writeback (%u) to %u",
2628 bch_cutoff_writeback, bch_cutoff_writeback_sync);
2629 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2630 }
2631}
2632
2633static int __init bcache_init(void)
2634{
2635 static const struct attribute *files[] = {
2636 &ksysfs_register.attr,
2637 &ksysfs_register_quiet.attr,
2638 &ksysfs_pendings_cleanup.attr,
2639 NULL
2640 };
2641
2642 check_module_parameters();
2643
2644 mutex_init(&bch_register_lock);
2645 init_waitqueue_head(&unregister_wait);
2646 register_reboot_notifier(&reboot);
2647
2648 bcache_major = register_blkdev(0, "bcache");
2649 if (bcache_major < 0) {
2650 unregister_reboot_notifier(&reboot);
2651 mutex_destroy(&bch_register_lock);
2652 return bcache_major;
2653 }
2654
2655 bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2656 if (!bcache_wq)
2657 goto err;
2658
2659 bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2660 if (!bch_journal_wq)
2661 goto err;
2662
2663 bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2664 if (!bcache_kobj)
2665 goto err;
2666
2667 if (bch_request_init() ||
2668 sysfs_create_files(bcache_kobj, files))
2669 goto err;
2670
2671 bch_debug_init();
2672 closure_debug_init();
2673
2674 bcache_is_reboot = false;
2675
2676 return 0;
2677err:
2678 bcache_exit();
2679 return -ENOMEM;
2680}
2681
2682/*
2683 * Module hooks
2684 */
2685module_exit(bcache_exit);
2686module_init(bcache_init);
2687
2688module_param(bch_cutoff_writeback, uint, 0);
2689MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2690
2691module_param(bch_cutoff_writeback_sync, uint, 0);
2692MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2693
2694MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2695MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2696MODULE_LICENSE("GPL");