Loading...
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#include "features.h"
17
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>
28
29unsigned int bch_cutoff_writeback;
30unsigned int bch_cutoff_writeback_sync;
31
32static const char bcache_magic[] = {
33 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
34 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
35};
36
37static const char invalid_uuid[] = {
38 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
39 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
40};
41
42static struct kobject *bcache_kobj;
43struct mutex bch_register_lock;
44bool bcache_is_reboot;
45LIST_HEAD(bch_cache_sets);
46static LIST_HEAD(uncached_devices);
47
48static int bcache_major;
49static DEFINE_IDA(bcache_device_idx);
50static wait_queue_head_t unregister_wait;
51struct workqueue_struct *bcache_wq;
52struct workqueue_struct *bch_flush_wq;
53struct workqueue_struct *bch_journal_wq;
54
55
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)
61
62/* Superblock */
63
64static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
65{
66 unsigned int bucket_size = le16_to_cpu(s->bucket_size);
67
68 if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
69 if (bch_has_feature_large_bucket(sb)) {
70 unsigned int max, order;
71
72 max = sizeof(unsigned int) * BITS_PER_BYTE - 1;
73 order = le16_to_cpu(s->bucket_size);
74 /*
75 * bcache tool will make sure the overflow won't
76 * happen, an error message here is enough.
77 */
78 if (order > max)
79 pr_err("Bucket size (1 << %u) overflows\n",
80 order);
81 bucket_size = 1 << order;
82 } else if (bch_has_feature_obso_large_bucket(sb)) {
83 bucket_size +=
84 le16_to_cpu(s->obso_bucket_size_hi) << 16;
85 }
86 }
87
88 return bucket_size;
89}
90
91static const char *read_super_common(struct cache_sb *sb, struct block_device *bdev,
92 struct cache_sb_disk *s)
93{
94 const char *err;
95 unsigned int i;
96
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);
100
101 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
102 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
103
104 err = "Too many journal buckets";
105 if (sb->keys > SB_JOURNAL_BUCKETS)
106 goto err;
107
108 err = "Too many buckets";
109 if (sb->nbuckets > LONG_MAX)
110 goto err;
111
112 err = "Not enough buckets";
113 if (sb->nbuckets < 1 << 7)
114 goto err;
115
116 err = "Bad block size (not power of 2)";
117 if (!is_power_of_2(sb->block_size))
118 goto err;
119
120 err = "Bad block size (larger than page size)";
121 if (sb->block_size > PAGE_SECTORS)
122 goto err;
123
124 err = "Bad bucket size (not power of 2)";
125 if (!is_power_of_2(sb->bucket_size))
126 goto err;
127
128 err = "Bad bucket size (smaller than page size)";
129 if (sb->bucket_size < PAGE_SECTORS)
130 goto err;
131
132 err = "Invalid superblock: device too small";
133 if (get_capacity(bdev->bd_disk) <
134 sb->bucket_size * sb->nbuckets)
135 goto err;
136
137 err = "Bad UUID";
138 if (bch_is_zero(sb->set_uuid, 16))
139 goto err;
140
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)
145 goto err;
146
147 err = "Journal buckets not sequential";
148 for (i = 0; i < sb->keys; i++)
149 if (sb->d[i] != sb->first_bucket + i)
150 goto err;
151
152 err = "Too many journal buckets";
153 if (sb->first_bucket + sb->keys > sb->nbuckets)
154 goto err;
155
156 err = "Invalid superblock: first bucket comes before end of super";
157 if (sb->first_bucket * sb->bucket_size < 16)
158 goto err;
159
160 err = NULL;
161err:
162 return err;
163}
164
165
166static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
167 struct cache_sb_disk **res)
168{
169 const char *err;
170 struct cache_sb_disk *s;
171 struct page *page;
172 unsigned int i;
173
174 page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
175 SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
176 if (IS_ERR(page))
177 return "IO error";
178 s = page_address(page) + offset_in_page(SB_OFFSET);
179
180 sb->offset = le64_to_cpu(s->offset);
181 sb->version = le64_to_cpu(s->version);
182
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);
187
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);
192
193 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
194 sb->d[i] = le64_to_cpu(s->d[i]);
195
196 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
197 sb->version, sb->flags, sb->seq, sb->keys);
198
199 err = "Not a bcache superblock (bad offset)";
200 if (sb->offset != SB_SECTOR)
201 goto err;
202
203 err = "Not a bcache superblock (bad magic)";
204 if (memcmp(sb->magic, bcache_magic, 16))
205 goto err;
206
207 err = "Bad checksum";
208 if (s->csum != csum_set(s))
209 goto err;
210
211 err = "Bad UUID";
212 if (bch_is_zero(sb->uuid, 16))
213 goto err;
214
215 sb->block_size = le16_to_cpu(s->block_size);
216
217 err = "Superblock block size smaller than device block size";
218 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
219 goto err;
220
221 switch (sb->version) {
222 case BCACHE_SB_VERSION_BDEV:
223 sb->data_offset = BDEV_DATA_START_DEFAULT;
224 break;
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);
228
229 err = "Bad data offset";
230 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
231 goto err;
232
233 break;
234 case BCACHE_SB_VERSION_CDEV:
235 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
236 err = read_super_common(sb, bdev, s);
237 if (err)
238 goto err;
239 break;
240 case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
241 /*
242 * Feature bits are needed in read_super_common(),
243 * convert them firstly.
244 */
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);
248
249 /* Check incompatible features */
250 err = "Unsupported compatible feature found";
251 if (bch_has_unknown_compat_features(sb))
252 goto err;
253
254 err = "Unsupported read-only compatible feature found";
255 if (bch_has_unknown_ro_compat_features(sb))
256 goto err;
257
258 err = "Unsupported incompatible feature found";
259 if (bch_has_unknown_incompat_features(sb))
260 goto err;
261
262 err = read_super_common(sb, bdev, s);
263 if (err)
264 goto err;
265 break;
266 default:
267 err = "Unsupported superblock version";
268 goto err;
269 }
270
271 sb->last_mount = (u32)ktime_get_real_seconds();
272 *res = s;
273 return NULL;
274err:
275 put_page(page);
276 return err;
277}
278
279static void write_bdev_super_endio(struct bio *bio)
280{
281 struct cached_dev *dc = bio->bi_private;
282
283 if (bio->bi_status)
284 bch_count_backing_io_errors(dc, bio);
285
286 closure_put(&dc->sb_write);
287}
288
289static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
290 struct bio *bio)
291{
292 unsigned int i;
293
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));
298
299 out->offset = cpu_to_le64(sb->offset);
300
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);
304
305 out->flags = cpu_to_le64(sb->flags);
306 out->seq = cpu_to_le64(sb->seq);
307
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);
311
312 for (i = 0; i < sb->keys; i++)
313 out->d[i] = cpu_to_le64(sb->d[i]);
314
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);
319 }
320
321 out->version = cpu_to_le64(sb->version);
322 out->csum = csum_set(out);
323
324 pr_debug("ver %llu, flags %llu, seq %llu\n",
325 sb->version, sb->flags, sb->seq);
326
327 submit_bio(bio);
328}
329
330static CLOSURE_CALLBACK(bch_write_bdev_super_unlock)
331{
332 closure_type(dc, struct cached_dev, sb_write);
333
334 up(&dc->sb_write_mutex);
335}
336
337void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
338{
339 struct closure *cl = &dc->sb_write;
340 struct bio *bio = &dc->sb_bio;
341
342 down(&dc->sb_write_mutex);
343 closure_init(cl, parent);
344
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;
348
349 closure_get(cl);
350 /* I/O request sent to backing device */
351 __write_super(&dc->sb, dc->sb_disk, bio);
352
353 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
354}
355
356static void write_super_endio(struct bio *bio)
357{
358 struct cache *ca = bio->bi_private;
359
360 /* is_read = 0 */
361 bch_count_io_errors(ca, bio->bi_status, 0,
362 "writing superblock");
363 closure_put(&ca->set->sb_write);
364}
365
366static CLOSURE_CALLBACK(bcache_write_super_unlock)
367{
368 closure_type(c, struct cache_set, sb_write);
369
370 up(&c->sb_write_mutex);
371}
372
373void bcache_write_super(struct cache_set *c)
374{
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;
379
380 down(&c->sb_write_mutex);
381 closure_init(cl, &c->cl);
382
383 ca->sb.seq++;
384
385 if (ca->sb.version < version)
386 ca->sb.version = version;
387
388 bio_init(bio, ca->bdev, ca->sb_bv, 1, 0);
389 bio->bi_end_io = write_super_endio;
390 bio->bi_private = ca;
391
392 closure_get(cl);
393 __write_super(&ca->sb, ca->sb_disk, bio);
394
395 closure_return_with_destructor(cl, bcache_write_super_unlock);
396}
397
398/* UUID io */
399
400static void uuid_endio(struct bio *bio)
401{
402 struct closure *cl = bio->bi_private;
403 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
404
405 cache_set_err_on(bio->bi_status, c, "accessing uuids");
406 bch_bbio_free(bio, c);
407 closure_put(cl);
408}
409
410static CLOSURE_CALLBACK(uuid_io_unlock)
411{
412 closure_type(c, struct cache_set, uuid_write);
413
414 up(&c->uuid_write_mutex);
415}
416
417static void uuid_io(struct cache_set *c, blk_opf_t opf, struct bkey *k,
418 struct closure *parent)
419{
420 struct closure *cl = &c->uuid_write;
421 struct uuid_entry *u;
422 unsigned int i;
423 char buf[80];
424
425 BUG_ON(!parent);
426 down(&c->uuid_write_mutex);
427 closure_init(cl, parent);
428
429 for (i = 0; i < KEY_PTRS(k); i++) {
430 struct bio *bio = bch_bbio_alloc(c);
431
432 bio->bi_opf = opf | REQ_SYNC | REQ_META;
433 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
434
435 bio->bi_end_io = uuid_endio;
436 bio->bi_private = cl;
437 bch_bio_map(bio, c->uuids);
438
439 bch_submit_bbio(bio, c, k, i);
440
441 if ((opf & REQ_OP_MASK) != REQ_OP_WRITE)
442 break;
443 }
444
445 bch_extent_to_text(buf, sizeof(buf), k);
446 pr_debug("%s UUIDs at %s\n", (opf & REQ_OP_MASK) == REQ_OP_WRITE ?
447 "wrote" : "read", buf);
448
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);
454
455 closure_return_with_destructor(cl, uuid_io_unlock);
456}
457
458static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
459{
460 struct bkey *k = &j->uuid_bucket;
461
462 if (__bch_btree_ptr_invalid(c, k))
463 return "bad uuid pointer";
464
465 bkey_copy(&c->uuid_bucket, k);
466 uuid_io(c, REQ_OP_READ, k, cl);
467
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;
471 int i;
472
473 closure_sync(cl);
474
475 /*
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
479 */
480
481 for (i = c->nr_uuids - 1;
482 i >= 0;
483 --i) {
484 memcpy(u1[i].uuid, u0[i].uuid, 16);
485 memcpy(u1[i].label, u0[i].label, 32);
486
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;
490
491 u1[i].flags = 0;
492 u1[i].sectors = 0;
493 }
494 }
495
496 return NULL;
497}
498
499static int __uuid_write(struct cache_set *c)
500{
501 BKEY_PADDED(key) k;
502 struct closure cl;
503 struct cache *ca = c->cache;
504 unsigned int size;
505
506 closure_init_stack(&cl);
507 lockdep_assert_held(&bch_register_lock);
508
509 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
510 return 1;
511
512 size = meta_bucket_pages(&ca->sb) * PAGE_SECTORS;
513 SET_KEY_SIZE(&k.key, size);
514 uuid_io(c, REQ_OP_WRITE, &k.key, &cl);
515 closure_sync(&cl);
516
517 /* Only one bucket used for uuid write */
518 atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
519
520 bkey_copy(&c->uuid_bucket, &k.key);
521 bkey_put(c, &k.key);
522 return 0;
523}
524
525int bch_uuid_write(struct cache_set *c)
526{
527 int ret = __uuid_write(c);
528
529 if (!ret)
530 bch_journal_meta(c, NULL);
531
532 return ret;
533}
534
535static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
536{
537 struct uuid_entry *u;
538
539 for (u = c->uuids;
540 u < c->uuids + c->nr_uuids; u++)
541 if (!memcmp(u->uuid, uuid, 16))
542 return u;
543
544 return NULL;
545}
546
547static struct uuid_entry *uuid_find_empty(struct cache_set *c)
548{
549 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
550
551 return uuid_find(c, zero_uuid);
552}
553
554/*
555 * Bucket priorities/gens:
556 *
557 * For each bucket, we store on disk its
558 * 8 bit gen
559 * 16 bit priority
560 *
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.
564 *
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.
569 *
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
573 * bucket, et cetera.
574 *
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
578 * disk.
579 */
580
581static void prio_endio(struct bio *bio)
582{
583 struct cache *ca = bio->bi_private;
584
585 cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
586 bch_bbio_free(bio, ca->set);
587 closure_put(&ca->prio);
588}
589
590static void prio_io(struct cache *ca, uint64_t bucket, blk_opf_t opf)
591{
592 struct closure *cl = &ca->prio;
593 struct bio *bio = bch_bbio_alloc(ca->set);
594
595 closure_init_stack(cl);
596
597 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
598 bio_set_dev(bio, ca->bdev);
599 bio->bi_iter.bi_size = meta_bucket_bytes(&ca->sb);
600
601 bio->bi_end_io = prio_endio;
602 bio->bi_private = ca;
603 bio->bi_opf = opf | REQ_SYNC | REQ_META;
604 bch_bio_map(bio, ca->disk_buckets);
605
606 closure_bio_submit(ca->set, bio, &ca->prio);
607 closure_sync(cl);
608}
609
610int bch_prio_write(struct cache *ca, bool wait)
611{
612 int i;
613 struct bucket *b;
614 struct closure cl;
615
616 pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
617 fifo_used(&ca->free[RESERVE_PRIO]),
618 fifo_used(&ca->free[RESERVE_NONE]),
619 fifo_used(&ca->free_inc));
620
621 /*
622 * Pre-check if there are enough free buckets. In the non-blocking
623 * scenario it's better to fail early rather than starting to allocate
624 * buckets and do a cleanup later in case of failure.
625 */
626 if (!wait) {
627 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
628 fifo_used(&ca->free[RESERVE_NONE]);
629 if (prio_buckets(ca) > avail)
630 return -ENOMEM;
631 }
632
633 closure_init_stack(&cl);
634
635 lockdep_assert_held(&ca->set->bucket_lock);
636
637 ca->disk_buckets->seq++;
638
639 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
640 &ca->meta_sectors_written);
641
642 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
643 long bucket;
644 struct prio_set *p = ca->disk_buckets;
645 struct bucket_disk *d = p->data;
646 struct bucket_disk *end = d + prios_per_bucket(ca);
647
648 for (b = ca->buckets + i * prios_per_bucket(ca);
649 b < ca->buckets + ca->sb.nbuckets && d < end;
650 b++, d++) {
651 d->prio = cpu_to_le16(b->prio);
652 d->gen = b->gen;
653 }
654
655 p->next_bucket = ca->prio_buckets[i + 1];
656 p->magic = pset_magic(&ca->sb);
657 p->csum = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
658
659 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
660 BUG_ON(bucket == -1);
661
662 mutex_unlock(&ca->set->bucket_lock);
663 prio_io(ca, bucket, REQ_OP_WRITE);
664 mutex_lock(&ca->set->bucket_lock);
665
666 ca->prio_buckets[i] = bucket;
667 atomic_dec_bug(&ca->buckets[bucket].pin);
668 }
669
670 mutex_unlock(&ca->set->bucket_lock);
671
672 bch_journal_meta(ca->set, &cl);
673 closure_sync(&cl);
674
675 mutex_lock(&ca->set->bucket_lock);
676
677 /*
678 * Don't want the old priorities to get garbage collected until after we
679 * finish writing the new ones, and they're journalled
680 */
681 for (i = 0; i < prio_buckets(ca); i++) {
682 if (ca->prio_last_buckets[i])
683 __bch_bucket_free(ca,
684 &ca->buckets[ca->prio_last_buckets[i]]);
685
686 ca->prio_last_buckets[i] = ca->prio_buckets[i];
687 }
688 return 0;
689}
690
691static int prio_read(struct cache *ca, uint64_t bucket)
692{
693 struct prio_set *p = ca->disk_buckets;
694 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
695 struct bucket *b;
696 unsigned int bucket_nr = 0;
697 int ret = -EIO;
698
699 for (b = ca->buckets;
700 b < ca->buckets + ca->sb.nbuckets;
701 b++, d++) {
702 if (d == end) {
703 ca->prio_buckets[bucket_nr] = bucket;
704 ca->prio_last_buckets[bucket_nr] = bucket;
705 bucket_nr++;
706
707 prio_io(ca, bucket, REQ_OP_READ);
708
709 if (p->csum !=
710 bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
711 pr_warn("bad csum reading priorities\n");
712 goto out;
713 }
714
715 if (p->magic != pset_magic(&ca->sb)) {
716 pr_warn("bad magic reading priorities\n");
717 goto out;
718 }
719
720 bucket = p->next_bucket;
721 d = p->data;
722 }
723
724 b->prio = le16_to_cpu(d->prio);
725 b->gen = b->last_gc = d->gen;
726 }
727
728 ret = 0;
729out:
730 return ret;
731}
732
733/* Bcache device */
734
735static int open_dev(struct gendisk *disk, blk_mode_t mode)
736{
737 struct bcache_device *d = disk->private_data;
738
739 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
740 return -ENXIO;
741
742 closure_get(&d->cl);
743 return 0;
744}
745
746static void release_dev(struct gendisk *b)
747{
748 struct bcache_device *d = b->private_data;
749
750 closure_put(&d->cl);
751}
752
753static int ioctl_dev(struct block_device *b, blk_mode_t mode,
754 unsigned int cmd, unsigned long arg)
755{
756 struct bcache_device *d = b->bd_disk->private_data;
757
758 return d->ioctl(d, mode, cmd, arg);
759}
760
761static const struct block_device_operations bcache_cached_ops = {
762 .submit_bio = cached_dev_submit_bio,
763 .open = open_dev,
764 .release = release_dev,
765 .ioctl = ioctl_dev,
766 .owner = THIS_MODULE,
767};
768
769static const struct block_device_operations bcache_flash_ops = {
770 .submit_bio = flash_dev_submit_bio,
771 .open = open_dev,
772 .release = release_dev,
773 .ioctl = ioctl_dev,
774 .owner = THIS_MODULE,
775};
776
777void bcache_device_stop(struct bcache_device *d)
778{
779 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
780 /*
781 * closure_fn set to
782 * - cached device: cached_dev_flush()
783 * - flash dev: flash_dev_flush()
784 */
785 closure_queue(&d->cl);
786}
787
788static void bcache_device_unlink(struct bcache_device *d)
789{
790 lockdep_assert_held(&bch_register_lock);
791
792 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
793 struct cache *ca = d->c->cache;
794
795 sysfs_remove_link(&d->c->kobj, d->name);
796 sysfs_remove_link(&d->kobj, "cache");
797
798 bd_unlink_disk_holder(ca->bdev, d->disk);
799 }
800}
801
802static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
803 const char *name)
804{
805 struct cache *ca = c->cache;
806 int ret;
807
808 bd_link_disk_holder(ca->bdev, d->disk);
809
810 snprintf(d->name, BCACHEDEVNAME_SIZE,
811 "%s%u", name, d->id);
812
813 ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
814 if (ret < 0)
815 pr_err("Couldn't create device -> cache set symlink\n");
816
817 ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
818 if (ret < 0)
819 pr_err("Couldn't create cache set -> device symlink\n");
820
821 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
822}
823
824static void bcache_device_detach(struct bcache_device *d)
825{
826 lockdep_assert_held(&bch_register_lock);
827
828 atomic_dec(&d->c->attached_dev_nr);
829
830 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
831 struct uuid_entry *u = d->c->uuids + d->id;
832
833 SET_UUID_FLASH_ONLY(u, 0);
834 memcpy(u->uuid, invalid_uuid, 16);
835 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
836 bch_uuid_write(d->c);
837 }
838
839 bcache_device_unlink(d);
840
841 d->c->devices[d->id] = NULL;
842 closure_put(&d->c->caching);
843 d->c = NULL;
844}
845
846static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
847 unsigned int id)
848{
849 d->id = id;
850 d->c = c;
851 c->devices[id] = d;
852
853 if (id >= c->devices_max_used)
854 c->devices_max_used = id + 1;
855
856 closure_get(&c->caching);
857}
858
859static inline int first_minor_to_idx(int first_minor)
860{
861 return (first_minor/BCACHE_MINORS);
862}
863
864static inline int idx_to_first_minor(int idx)
865{
866 return (idx * BCACHE_MINORS);
867}
868
869static void bcache_device_free(struct bcache_device *d)
870{
871 struct gendisk *disk = d->disk;
872
873 lockdep_assert_held(&bch_register_lock);
874
875 if (disk)
876 pr_info("%s stopped\n", disk->disk_name);
877 else
878 pr_err("bcache device (NULL gendisk) stopped\n");
879
880 if (d->c)
881 bcache_device_detach(d);
882
883 if (disk) {
884 ida_simple_remove(&bcache_device_idx,
885 first_minor_to_idx(disk->first_minor));
886 put_disk(disk);
887 }
888
889 bioset_exit(&d->bio_split);
890 kvfree(d->full_dirty_stripes);
891 kvfree(d->stripe_sectors_dirty);
892
893 closure_debug_destroy(&d->cl);
894}
895
896static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
897 sector_t sectors, struct block_device *cached_bdev,
898 const struct block_device_operations *ops)
899{
900 struct request_queue *q;
901 const size_t max_stripes = min_t(size_t, INT_MAX,
902 SIZE_MAX / sizeof(atomic_t));
903 uint64_t n;
904 int idx;
905
906 if (!d->stripe_size)
907 d->stripe_size = 1 << 31;
908 else if (d->stripe_size < BCH_MIN_STRIPE_SZ)
909 d->stripe_size = roundup(BCH_MIN_STRIPE_SZ, d->stripe_size);
910
911 n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
912 if (!n || n > max_stripes) {
913 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
914 n);
915 return -ENOMEM;
916 }
917 d->nr_stripes = n;
918
919 n = d->nr_stripes * sizeof(atomic_t);
920 d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
921 if (!d->stripe_sectors_dirty)
922 return -ENOMEM;
923
924 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
925 d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
926 if (!d->full_dirty_stripes)
927 goto out_free_stripe_sectors_dirty;
928
929 idx = ida_simple_get(&bcache_device_idx, 0,
930 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
931 if (idx < 0)
932 goto out_free_full_dirty_stripes;
933
934 if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
935 BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
936 goto out_ida_remove;
937
938 d->disk = blk_alloc_disk(NUMA_NO_NODE);
939 if (!d->disk)
940 goto out_bioset_exit;
941
942 set_capacity(d->disk, sectors);
943 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
944
945 d->disk->major = bcache_major;
946 d->disk->first_minor = idx_to_first_minor(idx);
947 d->disk->minors = BCACHE_MINORS;
948 d->disk->fops = ops;
949 d->disk->private_data = d;
950
951 q = d->disk->queue;
952 q->limits.max_hw_sectors = UINT_MAX;
953 q->limits.max_sectors = UINT_MAX;
954 q->limits.max_segment_size = UINT_MAX;
955 q->limits.max_segments = BIO_MAX_VECS;
956 blk_queue_max_discard_sectors(q, UINT_MAX);
957 q->limits.io_min = block_size;
958 q->limits.logical_block_size = block_size;
959 q->limits.physical_block_size = block_size;
960
961 if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
962 /*
963 * This should only happen with BCACHE_SB_VERSION_BDEV.
964 * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
965 */
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));
969
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));
972 }
973
974 blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
975
976 blk_queue_write_cache(q, true, true);
977
978 return 0;
979
980out_bioset_exit:
981 bioset_exit(&d->bio_split);
982out_ida_remove:
983 ida_simple_remove(&bcache_device_idx, idx);
984out_free_full_dirty_stripes:
985 kvfree(d->full_dirty_stripes);
986out_free_stripe_sectors_dirty:
987 kvfree(d->stripe_sectors_dirty);
988 return -ENOMEM;
989
990}
991
992/* Cached device */
993
994static void calc_cached_dev_sectors(struct cache_set *c)
995{
996 uint64_t sectors = 0;
997 struct cached_dev *dc;
998
999 list_for_each_entry(dc, &c->cached_devs, list)
1000 sectors += bdev_nr_sectors(dc->bdev);
1001
1002 c->cached_dev_sectors = sectors;
1003}
1004
1005#define BACKING_DEV_OFFLINE_TIMEOUT 5
1006static int cached_dev_status_update(void *arg)
1007{
1008 struct cached_dev *dc = arg;
1009 struct request_queue *q;
1010
1011 /*
1012 * If this delayed worker is stopping outside, directly quit here.
1013 * dc->io_disable might be set via sysfs interface, so check it
1014 * here too.
1015 */
1016 while (!kthread_should_stop() && !dc->io_disable) {
1017 q = bdev_get_queue(dc->bdev);
1018 if (blk_queue_dying(q))
1019 dc->offline_seconds++;
1020 else
1021 dc->offline_seconds = 0;
1022
1023 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1024 pr_err("%pg: device offline for %d seconds\n",
1025 dc->bdev,
1026 BACKING_DEV_OFFLINE_TIMEOUT);
1027 pr_err("%s: disable I/O request due to backing device offline\n",
1028 dc->disk.name);
1029 dc->io_disable = true;
1030 /* let others know earlier that io_disable is true */
1031 smp_mb();
1032 bcache_device_stop(&dc->disk);
1033 break;
1034 }
1035 schedule_timeout_interruptible(HZ);
1036 }
1037
1038 wait_for_kthread_stop();
1039 return 0;
1040}
1041
1042
1043int bch_cached_dev_run(struct cached_dev *dc)
1044{
1045 int ret = 0;
1046 struct bcache_device *d = &dc->disk;
1047 char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1048 char *env[] = {
1049 "DRIVER=bcache",
1050 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1051 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1052 NULL,
1053 };
1054
1055 if (dc->io_disable) {
1056 pr_err("I/O disabled on cached dev %pg\n", dc->bdev);
1057 ret = -EIO;
1058 goto out;
1059 }
1060
1061 if (atomic_xchg(&dc->running, 1)) {
1062 pr_info("cached dev %pg is running already\n", dc->bdev);
1063 ret = -EBUSY;
1064 goto out;
1065 }
1066
1067 if (!d->c &&
1068 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1069 struct closure cl;
1070
1071 closure_init_stack(&cl);
1072
1073 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1074 bch_write_bdev_super(dc, &cl);
1075 closure_sync(&cl);
1076 }
1077
1078 ret = add_disk(d->disk);
1079 if (ret)
1080 goto out;
1081 bd_link_disk_holder(dc->bdev, dc->disk.disk);
1082 /*
1083 * won't show up in the uevent file, use udevadm monitor -e instead
1084 * only class / kset properties are persistent
1085 */
1086 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1087
1088 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1089 sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1090 &d->kobj, "bcache")) {
1091 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1092 ret = -ENOMEM;
1093 goto out;
1094 }
1095
1096 dc->status_update_thread = kthread_run(cached_dev_status_update,
1097 dc, "bcache_status_update");
1098 if (IS_ERR(dc->status_update_thread)) {
1099 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1100 }
1101
1102out:
1103 kfree(env[1]);
1104 kfree(env[2]);
1105 kfree(buf);
1106 return ret;
1107}
1108
1109/*
1110 * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1111 * work dc->writeback_rate_update is running. Wait until the routine
1112 * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1113 * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1114 * seconds, give up waiting here and continue to cancel it too.
1115 */
1116static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1117{
1118 int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1119
1120 do {
1121 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1122 &dc->disk.flags))
1123 break;
1124 time_out--;
1125 schedule_timeout_interruptible(1);
1126 } while (time_out > 0);
1127
1128 if (time_out == 0)
1129 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1130
1131 cancel_delayed_work_sync(&dc->writeback_rate_update);
1132}
1133
1134static void cached_dev_detach_finish(struct work_struct *w)
1135{
1136 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1137 struct cache_set *c = dc->disk.c;
1138
1139 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1140 BUG_ON(refcount_read(&dc->count));
1141
1142
1143 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1144 cancel_writeback_rate_update_dwork(dc);
1145
1146 if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1147 kthread_stop(dc->writeback_thread);
1148 dc->writeback_thread = NULL;
1149 }
1150
1151 mutex_lock(&bch_register_lock);
1152
1153 bcache_device_detach(&dc->disk);
1154 list_move(&dc->list, &uncached_devices);
1155 calc_cached_dev_sectors(c);
1156
1157 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1158 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1159
1160 mutex_unlock(&bch_register_lock);
1161
1162 pr_info("Caching disabled for %pg\n", dc->bdev);
1163
1164 /* Drop ref we took in cached_dev_detach() */
1165 closure_put(&dc->disk.cl);
1166}
1167
1168void bch_cached_dev_detach(struct cached_dev *dc)
1169{
1170 lockdep_assert_held(&bch_register_lock);
1171
1172 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1173 return;
1174
1175 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1176 return;
1177
1178 /*
1179 * Block the device from being closed and freed until we're finished
1180 * detaching
1181 */
1182 closure_get(&dc->disk.cl);
1183
1184 bch_writeback_queue(dc);
1185
1186 cached_dev_put(dc);
1187}
1188
1189int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1190 uint8_t *set_uuid)
1191{
1192 uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1193 struct uuid_entry *u;
1194 struct cached_dev *exist_dc, *t;
1195 int ret = 0;
1196
1197 if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1198 (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1199 return -ENOENT;
1200
1201 if (dc->disk.c) {
1202 pr_err("Can't attach %pg: already attached\n", dc->bdev);
1203 return -EINVAL;
1204 }
1205
1206 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1207 pr_err("Can't attach %pg: shutting down\n", dc->bdev);
1208 return -EINVAL;
1209 }
1210
1211 if (dc->sb.block_size < c->cache->sb.block_size) {
1212 /* Will die */
1213 pr_err("Couldn't attach %pg: block size less than set's block size\n",
1214 dc->bdev);
1215 return -EINVAL;
1216 }
1217
1218 /* Check whether already attached */
1219 list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1220 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1221 pr_err("Tried to attach %pg but duplicate UUID already attached\n",
1222 dc->bdev);
1223
1224 return -EINVAL;
1225 }
1226 }
1227
1228 u = uuid_find(c, dc->sb.uuid);
1229
1230 if (u &&
1231 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1232 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1233 memcpy(u->uuid, invalid_uuid, 16);
1234 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1235 u = NULL;
1236 }
1237
1238 if (!u) {
1239 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1240 pr_err("Couldn't find uuid for %pg in set\n", dc->bdev);
1241 return -ENOENT;
1242 }
1243
1244 u = uuid_find_empty(c);
1245 if (!u) {
1246 pr_err("Not caching %pg, no room for UUID\n", dc->bdev);
1247 return -EINVAL;
1248 }
1249 }
1250
1251 /*
1252 * Deadlocks since we're called via sysfs...
1253 * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1254 */
1255
1256 if (bch_is_zero(u->uuid, 16)) {
1257 struct closure cl;
1258
1259 closure_init_stack(&cl);
1260
1261 memcpy(u->uuid, dc->sb.uuid, 16);
1262 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1263 u->first_reg = u->last_reg = rtime;
1264 bch_uuid_write(c);
1265
1266 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1267 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1268
1269 bch_write_bdev_super(dc, &cl);
1270 closure_sync(&cl);
1271 } else {
1272 u->last_reg = rtime;
1273 bch_uuid_write(c);
1274 }
1275
1276 bcache_device_attach(&dc->disk, c, u - c->uuids);
1277 list_move(&dc->list, &c->cached_devs);
1278 calc_cached_dev_sectors(c);
1279
1280 /*
1281 * dc->c must be set before dc->count != 0 - paired with the mb in
1282 * cached_dev_get()
1283 */
1284 smp_wmb();
1285 refcount_set(&dc->count, 1);
1286
1287 /* Block writeback thread, but spawn it */
1288 down_write(&dc->writeback_lock);
1289 if (bch_cached_dev_writeback_start(dc)) {
1290 up_write(&dc->writeback_lock);
1291 pr_err("Couldn't start writeback facilities for %s\n",
1292 dc->disk.disk->disk_name);
1293 return -ENOMEM;
1294 }
1295
1296 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1297 atomic_set(&dc->has_dirty, 1);
1298 bch_writeback_queue(dc);
1299 }
1300
1301 bch_sectors_dirty_init(&dc->disk);
1302
1303 ret = bch_cached_dev_run(dc);
1304 if (ret && (ret != -EBUSY)) {
1305 up_write(&dc->writeback_lock);
1306 /*
1307 * bch_register_lock is held, bcache_device_stop() is not
1308 * able to be directly called. The kthread and kworker
1309 * created previously in bch_cached_dev_writeback_start()
1310 * have to be stopped manually here.
1311 */
1312 kthread_stop(dc->writeback_thread);
1313 cancel_writeback_rate_update_dwork(dc);
1314 pr_err("Couldn't run cached device %pg\n", dc->bdev);
1315 return ret;
1316 }
1317
1318 bcache_device_link(&dc->disk, c, "bdev");
1319 atomic_inc(&c->attached_dev_nr);
1320
1321 if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1322 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1323 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1324 set_disk_ro(dc->disk.disk, 1);
1325 }
1326
1327 /* Allow the writeback thread to proceed */
1328 up_write(&dc->writeback_lock);
1329
1330 pr_info("Caching %pg as %s on set %pU\n",
1331 dc->bdev,
1332 dc->disk.disk->disk_name,
1333 dc->disk.c->set_uuid);
1334 return 0;
1335}
1336
1337/* when dc->disk.kobj released */
1338void bch_cached_dev_release(struct kobject *kobj)
1339{
1340 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1341 disk.kobj);
1342 kfree(dc);
1343 module_put(THIS_MODULE);
1344}
1345
1346static CLOSURE_CALLBACK(cached_dev_free)
1347{
1348 closure_type(dc, struct cached_dev, disk.cl);
1349
1350 if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1351 cancel_writeback_rate_update_dwork(dc);
1352
1353 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1354 kthread_stop(dc->writeback_thread);
1355 if (!IS_ERR_OR_NULL(dc->status_update_thread))
1356 kthread_stop(dc->status_update_thread);
1357
1358 mutex_lock(&bch_register_lock);
1359
1360 if (atomic_read(&dc->running)) {
1361 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1362 del_gendisk(dc->disk.disk);
1363 }
1364 bcache_device_free(&dc->disk);
1365 list_del(&dc->list);
1366
1367 mutex_unlock(&bch_register_lock);
1368
1369 if (dc->sb_disk)
1370 put_page(virt_to_page(dc->sb_disk));
1371
1372 if (dc->bdev_handle)
1373 bdev_release(dc->bdev_handle);
1374
1375 wake_up(&unregister_wait);
1376
1377 kobject_put(&dc->disk.kobj);
1378}
1379
1380static CLOSURE_CALLBACK(cached_dev_flush)
1381{
1382 closure_type(dc, struct cached_dev, disk.cl);
1383 struct bcache_device *d = &dc->disk;
1384
1385 mutex_lock(&bch_register_lock);
1386 bcache_device_unlink(d);
1387 mutex_unlock(&bch_register_lock);
1388
1389 bch_cache_accounting_destroy(&dc->accounting);
1390 kobject_del(&d->kobj);
1391
1392 continue_at(cl, cached_dev_free, system_wq);
1393}
1394
1395static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1396{
1397 int ret;
1398 struct io *io;
1399 struct request_queue *q = bdev_get_queue(dc->bdev);
1400
1401 __module_get(THIS_MODULE);
1402 INIT_LIST_HEAD(&dc->list);
1403 closure_init(&dc->disk.cl, NULL);
1404 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1405 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1406 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1407 sema_init(&dc->sb_write_mutex, 1);
1408 INIT_LIST_HEAD(&dc->io_lru);
1409 spin_lock_init(&dc->io_lock);
1410 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1411
1412 dc->sequential_cutoff = 4 << 20;
1413
1414 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1415 list_add(&io->lru, &dc->io_lru);
1416 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1417 }
1418
1419 dc->disk.stripe_size = q->limits.io_opt >> 9;
1420
1421 if (dc->disk.stripe_size)
1422 dc->partial_stripes_expensive =
1423 q->limits.raid_partial_stripes_expensive;
1424
1425 ret = bcache_device_init(&dc->disk, block_size,
1426 bdev_nr_sectors(dc->bdev) - dc->sb.data_offset,
1427 dc->bdev, &bcache_cached_ops);
1428 if (ret)
1429 return ret;
1430
1431 blk_queue_io_opt(dc->disk.disk->queue,
1432 max(queue_io_opt(dc->disk.disk->queue), queue_io_opt(q)));
1433
1434 atomic_set(&dc->io_errors, 0);
1435 dc->io_disable = false;
1436 dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1437 /* default to auto */
1438 dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1439
1440 bch_cached_dev_request_init(dc);
1441 bch_cached_dev_writeback_init(dc);
1442 return 0;
1443}
1444
1445/* Cached device - bcache superblock */
1446
1447static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1448 struct bdev_handle *bdev_handle,
1449 struct cached_dev *dc)
1450{
1451 const char *err = "cannot allocate memory";
1452 struct cache_set *c;
1453 int ret = -ENOMEM;
1454
1455 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1456 dc->bdev_handle = bdev_handle;
1457 dc->bdev = bdev_handle->bdev;
1458 dc->sb_disk = sb_disk;
1459
1460 if (cached_dev_init(dc, sb->block_size << 9))
1461 goto err;
1462
1463 err = "error creating kobject";
1464 if (kobject_add(&dc->disk.kobj, bdev_kobj(dc->bdev), "bcache"))
1465 goto err;
1466 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1467 goto err;
1468
1469 pr_info("registered backing device %pg\n", dc->bdev);
1470
1471 list_add(&dc->list, &uncached_devices);
1472 /* attach to a matched cache set if it exists */
1473 list_for_each_entry(c, &bch_cache_sets, list)
1474 bch_cached_dev_attach(dc, c, NULL);
1475
1476 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1477 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1478 err = "failed to run cached device";
1479 ret = bch_cached_dev_run(dc);
1480 if (ret)
1481 goto err;
1482 }
1483
1484 return 0;
1485err:
1486 pr_notice("error %pg: %s\n", dc->bdev, err);
1487 bcache_device_stop(&dc->disk);
1488 return ret;
1489}
1490
1491/* Flash only volumes */
1492
1493/* When d->kobj released */
1494void bch_flash_dev_release(struct kobject *kobj)
1495{
1496 struct bcache_device *d = container_of(kobj, struct bcache_device,
1497 kobj);
1498 kfree(d);
1499}
1500
1501static CLOSURE_CALLBACK(flash_dev_free)
1502{
1503 closure_type(d, struct bcache_device, cl);
1504
1505 mutex_lock(&bch_register_lock);
1506 atomic_long_sub(bcache_dev_sectors_dirty(d),
1507 &d->c->flash_dev_dirty_sectors);
1508 del_gendisk(d->disk);
1509 bcache_device_free(d);
1510 mutex_unlock(&bch_register_lock);
1511 kobject_put(&d->kobj);
1512}
1513
1514static CLOSURE_CALLBACK(flash_dev_flush)
1515{
1516 closure_type(d, struct bcache_device, cl);
1517
1518 mutex_lock(&bch_register_lock);
1519 bcache_device_unlink(d);
1520 mutex_unlock(&bch_register_lock);
1521 kobject_del(&d->kobj);
1522 continue_at(cl, flash_dev_free, system_wq);
1523}
1524
1525static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1526{
1527 int err = -ENOMEM;
1528 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1529 GFP_KERNEL);
1530 if (!d)
1531 goto err_ret;
1532
1533 closure_init(&d->cl, NULL);
1534 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1535
1536 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1537
1538 if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1539 NULL, &bcache_flash_ops))
1540 goto err;
1541
1542 bcache_device_attach(d, c, u - c->uuids);
1543 bch_sectors_dirty_init(d);
1544 bch_flash_dev_request_init(d);
1545 err = add_disk(d->disk);
1546 if (err)
1547 goto err;
1548
1549 err = kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache");
1550 if (err)
1551 goto err;
1552
1553 bcache_device_link(d, c, "volume");
1554
1555 if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1556 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1557 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1558 set_disk_ro(d->disk, 1);
1559 }
1560
1561 return 0;
1562err:
1563 kobject_put(&d->kobj);
1564err_ret:
1565 return err;
1566}
1567
1568static int flash_devs_run(struct cache_set *c)
1569{
1570 int ret = 0;
1571 struct uuid_entry *u;
1572
1573 for (u = c->uuids;
1574 u < c->uuids + c->nr_uuids && !ret;
1575 u++)
1576 if (UUID_FLASH_ONLY(u))
1577 ret = flash_dev_run(c, u);
1578
1579 return ret;
1580}
1581
1582int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1583{
1584 struct uuid_entry *u;
1585
1586 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1587 return -EINTR;
1588
1589 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1590 return -EPERM;
1591
1592 u = uuid_find_empty(c);
1593 if (!u) {
1594 pr_err("Can't create volume, no room for UUID\n");
1595 return -EINVAL;
1596 }
1597
1598 get_random_bytes(u->uuid, 16);
1599 memset(u->label, 0, 32);
1600 u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1601
1602 SET_UUID_FLASH_ONLY(u, 1);
1603 u->sectors = size >> 9;
1604
1605 bch_uuid_write(c);
1606
1607 return flash_dev_run(c, u);
1608}
1609
1610bool bch_cached_dev_error(struct cached_dev *dc)
1611{
1612 if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1613 return false;
1614
1615 dc->io_disable = true;
1616 /* make others know io_disable is true earlier */
1617 smp_mb();
1618
1619 pr_err("stop %s: too many IO errors on backing device %pg\n",
1620 dc->disk.disk->disk_name, dc->bdev);
1621
1622 bcache_device_stop(&dc->disk);
1623 return true;
1624}
1625
1626/* Cache set */
1627
1628__printf(2, 3)
1629bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1630{
1631 struct va_format vaf;
1632 va_list args;
1633
1634 if (c->on_error != ON_ERROR_PANIC &&
1635 test_bit(CACHE_SET_STOPPING, &c->flags))
1636 return false;
1637
1638 if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1639 pr_info("CACHE_SET_IO_DISABLE already set\n");
1640
1641 /*
1642 * XXX: we can be called from atomic context
1643 * acquire_console_sem();
1644 */
1645
1646 va_start(args, fmt);
1647
1648 vaf.fmt = fmt;
1649 vaf.va = &args;
1650
1651 pr_err("error on %pU: %pV, disabling caching\n",
1652 c->set_uuid, &vaf);
1653
1654 va_end(args);
1655
1656 if (c->on_error == ON_ERROR_PANIC)
1657 panic("panic forced after error\n");
1658
1659 bch_cache_set_unregister(c);
1660 return true;
1661}
1662
1663/* When c->kobj released */
1664void bch_cache_set_release(struct kobject *kobj)
1665{
1666 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1667
1668 kfree(c);
1669 module_put(THIS_MODULE);
1670}
1671
1672static CLOSURE_CALLBACK(cache_set_free)
1673{
1674 closure_type(c, struct cache_set, cl);
1675 struct cache *ca;
1676
1677 debugfs_remove(c->debug);
1678
1679 bch_open_buckets_free(c);
1680 bch_btree_cache_free(c);
1681 bch_journal_free(c);
1682
1683 mutex_lock(&bch_register_lock);
1684 bch_bset_sort_state_free(&c->sort);
1685 free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1686
1687 ca = c->cache;
1688 if (ca) {
1689 ca->set = NULL;
1690 c->cache = NULL;
1691 kobject_put(&ca->kobj);
1692 }
1693
1694
1695 if (c->moving_gc_wq)
1696 destroy_workqueue(c->moving_gc_wq);
1697 bioset_exit(&c->bio_split);
1698 mempool_exit(&c->fill_iter);
1699 mempool_exit(&c->bio_meta);
1700 mempool_exit(&c->search);
1701 kfree(c->devices);
1702
1703 list_del(&c->list);
1704 mutex_unlock(&bch_register_lock);
1705
1706 pr_info("Cache set %pU unregistered\n", c->set_uuid);
1707 wake_up(&unregister_wait);
1708
1709 closure_debug_destroy(&c->cl);
1710 kobject_put(&c->kobj);
1711}
1712
1713static CLOSURE_CALLBACK(cache_set_flush)
1714{
1715 closure_type(c, struct cache_set, caching);
1716 struct cache *ca = c->cache;
1717 struct btree *b;
1718
1719 bch_cache_accounting_destroy(&c->accounting);
1720
1721 kobject_put(&c->internal);
1722 kobject_del(&c->kobj);
1723
1724 if (!IS_ERR_OR_NULL(c->gc_thread))
1725 kthread_stop(c->gc_thread);
1726
1727 if (!IS_ERR(c->root))
1728 list_add(&c->root->list, &c->btree_cache);
1729
1730 /*
1731 * Avoid flushing cached nodes if cache set is retiring
1732 * due to too many I/O errors detected.
1733 */
1734 if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1735 list_for_each_entry(b, &c->btree_cache, list) {
1736 mutex_lock(&b->write_lock);
1737 if (btree_node_dirty(b))
1738 __bch_btree_node_write(b, NULL);
1739 mutex_unlock(&b->write_lock);
1740 }
1741
1742 if (ca->alloc_thread)
1743 kthread_stop(ca->alloc_thread);
1744
1745 if (c->journal.cur) {
1746 cancel_delayed_work_sync(&c->journal.work);
1747 /* flush last journal entry if needed */
1748 c->journal.work.work.func(&c->journal.work.work);
1749 }
1750
1751 closure_return(cl);
1752}
1753
1754/*
1755 * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1756 * cache set is unregistering due to too many I/O errors. In this condition,
1757 * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1758 * value and whether the broken cache has dirty data:
1759 *
1760 * dc->stop_when_cache_set_failed dc->has_dirty stop bcache device
1761 * BCH_CACHED_STOP_AUTO 0 NO
1762 * BCH_CACHED_STOP_AUTO 1 YES
1763 * BCH_CACHED_DEV_STOP_ALWAYS 0 YES
1764 * BCH_CACHED_DEV_STOP_ALWAYS 1 YES
1765 *
1766 * The expected behavior is, if stop_when_cache_set_failed is configured to
1767 * "auto" via sysfs interface, the bcache device will not be stopped if the
1768 * backing device is clean on the broken cache device.
1769 */
1770static void conditional_stop_bcache_device(struct cache_set *c,
1771 struct bcache_device *d,
1772 struct cached_dev *dc)
1773{
1774 if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1775 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1776 d->disk->disk_name, c->set_uuid);
1777 bcache_device_stop(d);
1778 } else if (atomic_read(&dc->has_dirty)) {
1779 /*
1780 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1781 * and dc->has_dirty == 1
1782 */
1783 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1784 d->disk->disk_name);
1785 /*
1786 * There might be a small time gap that cache set is
1787 * released but bcache device is not. Inside this time
1788 * gap, regular I/O requests will directly go into
1789 * backing device as no cache set attached to. This
1790 * behavior may also introduce potential inconsistence
1791 * data in writeback mode while cache is dirty.
1792 * Therefore before calling bcache_device_stop() due
1793 * to a broken cache device, dc->io_disable should be
1794 * explicitly set to true.
1795 */
1796 dc->io_disable = true;
1797 /* make others know io_disable is true earlier */
1798 smp_mb();
1799 bcache_device_stop(d);
1800 } else {
1801 /*
1802 * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1803 * and dc->has_dirty == 0
1804 */
1805 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1806 d->disk->disk_name);
1807 }
1808}
1809
1810static CLOSURE_CALLBACK(__cache_set_unregister)
1811{
1812 closure_type(c, struct cache_set, caching);
1813 struct cached_dev *dc;
1814 struct bcache_device *d;
1815 size_t i;
1816
1817 mutex_lock(&bch_register_lock);
1818
1819 for (i = 0; i < c->devices_max_used; i++) {
1820 d = c->devices[i];
1821 if (!d)
1822 continue;
1823
1824 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1825 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1826 dc = container_of(d, struct cached_dev, disk);
1827 bch_cached_dev_detach(dc);
1828 if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1829 conditional_stop_bcache_device(c, d, dc);
1830 } else {
1831 bcache_device_stop(d);
1832 }
1833 }
1834
1835 mutex_unlock(&bch_register_lock);
1836
1837 continue_at(cl, cache_set_flush, system_wq);
1838}
1839
1840void bch_cache_set_stop(struct cache_set *c)
1841{
1842 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1843 /* closure_fn set to __cache_set_unregister() */
1844 closure_queue(&c->caching);
1845}
1846
1847void bch_cache_set_unregister(struct cache_set *c)
1848{
1849 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1850 bch_cache_set_stop(c);
1851}
1852
1853#define alloc_meta_bucket_pages(gfp, sb) \
1854 ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1855
1856struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1857{
1858 int iter_size;
1859 struct cache *ca = container_of(sb, struct cache, sb);
1860 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1861
1862 if (!c)
1863 return NULL;
1864
1865 __module_get(THIS_MODULE);
1866 closure_init(&c->cl, NULL);
1867 set_closure_fn(&c->cl, cache_set_free, system_wq);
1868
1869 closure_init(&c->caching, &c->cl);
1870 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1871
1872 /* Maybe create continue_at_noreturn() and use it here? */
1873 closure_set_stopped(&c->cl);
1874 closure_put(&c->cl);
1875
1876 kobject_init(&c->kobj, &bch_cache_set_ktype);
1877 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1878
1879 bch_cache_accounting_init(&c->accounting, &c->cl);
1880
1881 memcpy(c->set_uuid, sb->set_uuid, 16);
1882
1883 c->cache = ca;
1884 c->cache->set = c;
1885 c->bucket_bits = ilog2(sb->bucket_size);
1886 c->block_bits = ilog2(sb->block_size);
1887 c->nr_uuids = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1888 c->devices_max_used = 0;
1889 atomic_set(&c->attached_dev_nr, 0);
1890 c->btree_pages = meta_bucket_pages(sb);
1891 if (c->btree_pages > BTREE_MAX_PAGES)
1892 c->btree_pages = max_t(int, c->btree_pages / 4,
1893 BTREE_MAX_PAGES);
1894
1895 sema_init(&c->sb_write_mutex, 1);
1896 mutex_init(&c->bucket_lock);
1897 init_waitqueue_head(&c->btree_cache_wait);
1898 spin_lock_init(&c->btree_cannibalize_lock);
1899 init_waitqueue_head(&c->bucket_wait);
1900 init_waitqueue_head(&c->gc_wait);
1901 sema_init(&c->uuid_write_mutex, 1);
1902
1903 spin_lock_init(&c->btree_gc_time.lock);
1904 spin_lock_init(&c->btree_split_time.lock);
1905 spin_lock_init(&c->btree_read_time.lock);
1906
1907 bch_moving_init_cache_set(c);
1908
1909 INIT_LIST_HEAD(&c->list);
1910 INIT_LIST_HEAD(&c->cached_devs);
1911 INIT_LIST_HEAD(&c->btree_cache);
1912 INIT_LIST_HEAD(&c->btree_cache_freeable);
1913 INIT_LIST_HEAD(&c->btree_cache_freed);
1914 INIT_LIST_HEAD(&c->data_buckets);
1915
1916 iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1917 sizeof(struct btree_iter_set);
1918
1919 c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1920 if (!c->devices)
1921 goto err;
1922
1923 if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1924 goto err;
1925
1926 if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1927 sizeof(struct bbio) +
1928 sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1929 goto err;
1930
1931 if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1932 goto err;
1933
1934 if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1935 BIOSET_NEED_RESCUER))
1936 goto err;
1937
1938 c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1939 if (!c->uuids)
1940 goto err;
1941
1942 c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1943 if (!c->moving_gc_wq)
1944 goto err;
1945
1946 if (bch_journal_alloc(c))
1947 goto err;
1948
1949 if (bch_btree_cache_alloc(c))
1950 goto err;
1951
1952 if (bch_open_buckets_alloc(c))
1953 goto err;
1954
1955 if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1956 goto err;
1957
1958 c->congested_read_threshold_us = 2000;
1959 c->congested_write_threshold_us = 20000;
1960 c->error_limit = DEFAULT_IO_ERROR_LIMIT;
1961 c->idle_max_writeback_rate_enabled = 1;
1962 WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1963
1964 return c;
1965err:
1966 bch_cache_set_unregister(c);
1967 return NULL;
1968}
1969
1970static int run_cache_set(struct cache_set *c)
1971{
1972 const char *err = "cannot allocate memory";
1973 struct cached_dev *dc, *t;
1974 struct cache *ca = c->cache;
1975 struct closure cl;
1976 LIST_HEAD(journal);
1977 struct journal_replay *l;
1978
1979 closure_init_stack(&cl);
1980
1981 c->nbuckets = ca->sb.nbuckets;
1982 set_gc_sectors(c);
1983
1984 if (CACHE_SYNC(&c->cache->sb)) {
1985 struct bkey *k;
1986 struct jset *j;
1987
1988 err = "cannot allocate memory for journal";
1989 if (bch_journal_read(c, &journal))
1990 goto err;
1991
1992 pr_debug("btree_journal_read() done\n");
1993
1994 err = "no journal entries found";
1995 if (list_empty(&journal))
1996 goto err;
1997
1998 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1999
2000 err = "IO error reading priorities";
2001 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
2002 goto err;
2003
2004 /*
2005 * If prio_read() fails it'll call cache_set_error and we'll
2006 * tear everything down right away, but if we perhaps checked
2007 * sooner we could avoid journal replay.
2008 */
2009
2010 k = &j->btree_root;
2011
2012 err = "bad btree root";
2013 if (__bch_btree_ptr_invalid(c, k))
2014 goto err;
2015
2016 err = "error reading btree root";
2017 c->root = bch_btree_node_get(c, NULL, k,
2018 j->btree_level,
2019 true, NULL);
2020 if (IS_ERR(c->root))
2021 goto err;
2022
2023 list_del_init(&c->root->list);
2024 rw_unlock(true, c->root);
2025
2026 err = uuid_read(c, j, &cl);
2027 if (err)
2028 goto err;
2029
2030 err = "error in recovery";
2031 if (bch_btree_check(c))
2032 goto err;
2033
2034 bch_journal_mark(c, &journal);
2035 bch_initial_gc_finish(c);
2036 pr_debug("btree_check() done\n");
2037
2038 /*
2039 * bcache_journal_next() can't happen sooner, or
2040 * btree_gc_finish() will give spurious errors about last_gc >
2041 * gc_gen - this is a hack but oh well.
2042 */
2043 bch_journal_next(&c->journal);
2044
2045 err = "error starting allocator thread";
2046 if (bch_cache_allocator_start(ca))
2047 goto err;
2048
2049 /*
2050 * First place it's safe to allocate: btree_check() and
2051 * btree_gc_finish() have to run before we have buckets to
2052 * allocate, and bch_bucket_alloc_set() might cause a journal
2053 * entry to be written so bcache_journal_next() has to be called
2054 * first.
2055 *
2056 * If the uuids were in the old format we have to rewrite them
2057 * before the next journal entry is written:
2058 */
2059 if (j->version < BCACHE_JSET_VERSION_UUID)
2060 __uuid_write(c);
2061
2062 err = "bcache: replay journal failed";
2063 if (bch_journal_replay(c, &journal))
2064 goto err;
2065 } else {
2066 unsigned int j;
2067
2068 pr_notice("invalidating existing data\n");
2069 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2070 2, SB_JOURNAL_BUCKETS);
2071
2072 for (j = 0; j < ca->sb.keys; j++)
2073 ca->sb.d[j] = ca->sb.first_bucket + j;
2074
2075 bch_initial_gc_finish(c);
2076
2077 err = "error starting allocator thread";
2078 if (bch_cache_allocator_start(ca))
2079 goto err;
2080
2081 mutex_lock(&c->bucket_lock);
2082 bch_prio_write(ca, true);
2083 mutex_unlock(&c->bucket_lock);
2084
2085 err = "cannot allocate new UUID bucket";
2086 if (__uuid_write(c))
2087 goto err;
2088
2089 err = "cannot allocate new btree root";
2090 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2091 if (IS_ERR(c->root))
2092 goto err;
2093
2094 mutex_lock(&c->root->write_lock);
2095 bkey_copy_key(&c->root->key, &MAX_KEY);
2096 bch_btree_node_write(c->root, &cl);
2097 mutex_unlock(&c->root->write_lock);
2098
2099 bch_btree_set_root(c->root);
2100 rw_unlock(true, c->root);
2101
2102 /*
2103 * We don't want to write the first journal entry until
2104 * everything is set up - fortunately journal entries won't be
2105 * written until the SET_CACHE_SYNC() here:
2106 */
2107 SET_CACHE_SYNC(&c->cache->sb, true);
2108
2109 bch_journal_next(&c->journal);
2110 bch_journal_meta(c, &cl);
2111 }
2112
2113 err = "error starting gc thread";
2114 if (bch_gc_thread_start(c))
2115 goto err;
2116
2117 closure_sync(&cl);
2118 c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2119 bcache_write_super(c);
2120
2121 if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2122 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2123
2124 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2125 bch_cached_dev_attach(dc, c, NULL);
2126
2127 flash_devs_run(c);
2128
2129 bch_journal_space_reserve(&c->journal);
2130 set_bit(CACHE_SET_RUNNING, &c->flags);
2131 return 0;
2132err:
2133 while (!list_empty(&journal)) {
2134 l = list_first_entry(&journal, struct journal_replay, list);
2135 list_del(&l->list);
2136 kfree(l);
2137 }
2138
2139 closure_sync(&cl);
2140
2141 bch_cache_set_error(c, "%s", err);
2142
2143 return -EIO;
2144}
2145
2146static const char *register_cache_set(struct cache *ca)
2147{
2148 char buf[12];
2149 const char *err = "cannot allocate memory";
2150 struct cache_set *c;
2151
2152 list_for_each_entry(c, &bch_cache_sets, list)
2153 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2154 if (c->cache)
2155 return "duplicate cache set member";
2156
2157 goto found;
2158 }
2159
2160 c = bch_cache_set_alloc(&ca->sb);
2161 if (!c)
2162 return err;
2163
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"))
2167 goto err;
2168
2169 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2170 goto err;
2171
2172 bch_debug_init_cache_set(c);
2173
2174 list_add(&c->list, &bch_cache_sets);
2175found:
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))
2179 goto err;
2180
2181 kobject_get(&ca->kobj);
2182 ca->set = c;
2183 ca->set->cache = ca;
2184
2185 err = "failed to run cache set";
2186 if (run_cache_set(c) < 0)
2187 goto err;
2188
2189 return NULL;
2190err:
2191 bch_cache_set_unregister(c);
2192 return err;
2193}
2194
2195/* Cache device */
2196
2197/* When ca->kobj released */
2198void bch_cache_release(struct kobject *kobj)
2199{
2200 struct cache *ca = container_of(kobj, struct cache, kobj);
2201 unsigned int i;
2202
2203 if (ca->set) {
2204 BUG_ON(ca->set->cache != ca);
2205 ca->set->cache = NULL;
2206 }
2207
2208 free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2209 kfree(ca->prio_buckets);
2210 vfree(ca->buckets);
2211
2212 free_heap(&ca->heap);
2213 free_fifo(&ca->free_inc);
2214
2215 for (i = 0; i < RESERVE_NR; i++)
2216 free_fifo(&ca->free[i]);
2217
2218 if (ca->sb_disk)
2219 put_page(virt_to_page(ca->sb_disk));
2220
2221 if (ca->bdev_handle)
2222 bdev_release(ca->bdev_handle);
2223
2224 kfree(ca);
2225 module_put(THIS_MODULE);
2226}
2227
2228static int cache_alloc(struct cache *ca)
2229{
2230 size_t free;
2231 size_t btree_buckets;
2232 struct bucket *b;
2233 int ret = -ENOMEM;
2234 const char *err = NULL;
2235
2236 __module_get(THIS_MODULE);
2237 kobject_init(&ca->kobj, &bch_cache_ktype);
2238
2239 bio_init(&ca->journal.bio, NULL, ca->journal.bio.bi_inline_vecs, 8, 0);
2240
2241 /*
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
2249 */
2250 btree_buckets = ca->sb.njournal_buckets ?: 8;
2251 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2252 if (!free) {
2253 ret = -EPERM;
2254 err = "ca->sb.nbuckets is too small";
2255 goto err_free;
2256 }
2257
2258 if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2259 GFP_KERNEL)) {
2260 err = "ca->free[RESERVE_BTREE] alloc failed";
2261 goto err_btree_alloc;
2262 }
2263
2264 if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2265 GFP_KERNEL)) {
2266 err = "ca->free[RESERVE_PRIO] alloc failed";
2267 goto err_prio_alloc;
2268 }
2269
2270 if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2271 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2272 goto err_movinggc_alloc;
2273 }
2274
2275 if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2276 err = "ca->free[RESERVE_NONE] alloc failed";
2277 goto err_none_alloc;
2278 }
2279
2280 if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2281 err = "ca->free_inc alloc failed";
2282 goto err_free_inc_alloc;
2283 }
2284
2285 if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2286 err = "ca->heap alloc failed";
2287 goto err_heap_alloc;
2288 }
2289
2290 ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2291 ca->sb.nbuckets));
2292 if (!ca->buckets) {
2293 err = "ca->buckets alloc failed";
2294 goto err_buckets_alloc;
2295 }
2296
2297 ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2298 prio_buckets(ca), 2),
2299 GFP_KERNEL);
2300 if (!ca->prio_buckets) {
2301 err = "ca->prio_buckets alloc failed";
2302 goto err_prio_buckets_alloc;
2303 }
2304
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;
2309 }
2310
2311 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2312
2313 for_each_bucket(b, ca)
2314 atomic_set(&b->pin, 0);
2315 return 0;
2316
2317err_disk_buckets_alloc:
2318 kfree(ca->prio_buckets);
2319err_prio_buckets_alloc:
2320 vfree(ca->buckets);
2321err_buckets_alloc:
2322 free_heap(&ca->heap);
2323err_heap_alloc:
2324 free_fifo(&ca->free_inc);
2325err_free_inc_alloc:
2326 free_fifo(&ca->free[RESERVE_NONE]);
2327err_none_alloc:
2328 free_fifo(&ca->free[RESERVE_MOVINGGC]);
2329err_movinggc_alloc:
2330 free_fifo(&ca->free[RESERVE_PRIO]);
2331err_prio_alloc:
2332 free_fifo(&ca->free[RESERVE_BTREE]);
2333err_btree_alloc:
2334err_free:
2335 module_put(THIS_MODULE);
2336 if (err)
2337 pr_notice("error %pg: %s\n", ca->bdev, err);
2338 return ret;
2339}
2340
2341static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2342 struct bdev_handle *bdev_handle,
2343 struct cache *ca)
2344{
2345 const char *err = NULL; /* must be set for any error case */
2346 int ret = 0;
2347
2348 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2349 ca->bdev_handle = bdev_handle;
2350 ca->bdev = bdev_handle->bdev;
2351 ca->sb_disk = sb_disk;
2352
2353 if (bdev_max_discard_sectors((bdev_handle->bdev)))
2354 ca->discard = CACHE_DISCARD(&ca->sb);
2355
2356 ret = cache_alloc(ca);
2357 if (ret != 0) {
2358 if (ret == -ENOMEM)
2359 err = "cache_alloc(): -ENOMEM";
2360 else if (ret == -EPERM)
2361 err = "cache_alloc(): cache device is too small";
2362 else
2363 err = "cache_alloc(): unknown error";
2364 pr_notice("error %pg: %s\n", bdev_handle->bdev, err);
2365 /*
2366 * If we failed here, it means ca->kobj is not initialized yet,
2367 * kobject_put() won't be called and there is no chance to
2368 * call bdev_release() to bdev in bch_cache_release(). So
2369 * we explicitly call bdev_release() here.
2370 */
2371 bdev_release(bdev_handle);
2372 return ret;
2373 }
2374
2375 if (kobject_add(&ca->kobj, bdev_kobj(bdev_handle->bdev), "bcache")) {
2376 pr_notice("error %pg: error calling kobject_add\n",
2377 bdev_handle->bdev);
2378 ret = -ENOMEM;
2379 goto out;
2380 }
2381
2382 mutex_lock(&bch_register_lock);
2383 err = register_cache_set(ca);
2384 mutex_unlock(&bch_register_lock);
2385
2386 if (err) {
2387 ret = -ENODEV;
2388 goto out;
2389 }
2390
2391 pr_info("registered cache device %pg\n", ca->bdev_handle->bdev);
2392
2393out:
2394 kobject_put(&ca->kobj);
2395 return ret;
2396}
2397
2398/* Global interfaces/init */
2399
2400static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2401 const char *buffer, size_t size);
2402static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2403 struct kobj_attribute *attr,
2404 const char *buffer, size_t size);
2405
2406kobj_attribute_write(register, register_bcache);
2407kobj_attribute_write(register_quiet, register_bcache);
2408kobj_attribute_write(pendings_cleanup, bch_pending_bdevs_cleanup);
2409
2410static bool bch_is_open_backing(dev_t dev)
2411{
2412 struct cache_set *c, *tc;
2413 struct cached_dev *dc, *t;
2414
2415 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2416 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2417 if (dc->bdev->bd_dev == dev)
2418 return true;
2419 list_for_each_entry_safe(dc, t, &uncached_devices, list)
2420 if (dc->bdev->bd_dev == dev)
2421 return true;
2422 return false;
2423}
2424
2425static bool bch_is_open_cache(dev_t dev)
2426{
2427 struct cache_set *c, *tc;
2428
2429 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2430 struct cache *ca = c->cache;
2431
2432 if (ca->bdev->bd_dev == dev)
2433 return true;
2434 }
2435
2436 return false;
2437}
2438
2439static bool bch_is_open(dev_t dev)
2440{
2441 return bch_is_open_cache(dev) || bch_is_open_backing(dev);
2442}
2443
2444struct async_reg_args {
2445 struct delayed_work reg_work;
2446 char *path;
2447 struct cache_sb *sb;
2448 struct cache_sb_disk *sb_disk;
2449 struct bdev_handle *bdev_handle;
2450 void *holder;
2451};
2452
2453static void register_bdev_worker(struct work_struct *work)
2454{
2455 int fail = false;
2456 struct async_reg_args *args =
2457 container_of(work, struct async_reg_args, reg_work.work);
2458
2459 mutex_lock(&bch_register_lock);
2460 if (register_bdev(args->sb, args->sb_disk, args->bdev_handle,
2461 args->holder) < 0)
2462 fail = true;
2463 mutex_unlock(&bch_register_lock);
2464
2465 if (fail)
2466 pr_info("error %s: fail to register backing device\n",
2467 args->path);
2468 kfree(args->sb);
2469 kfree(args->path);
2470 kfree(args);
2471 module_put(THIS_MODULE);
2472}
2473
2474static void register_cache_worker(struct work_struct *work)
2475{
2476 int fail = false;
2477 struct async_reg_args *args =
2478 container_of(work, struct async_reg_args, reg_work.work);
2479
2480 /* blkdev_put() will be called in bch_cache_release() */
2481 if (register_cache(args->sb, args->sb_disk, args->bdev_handle,
2482 args->holder))
2483 fail = true;
2484
2485 if (fail)
2486 pr_info("error %s: fail to register cache device\n",
2487 args->path);
2488 kfree(args->sb);
2489 kfree(args->path);
2490 kfree(args);
2491 module_put(THIS_MODULE);
2492}
2493
2494static void register_device_async(struct async_reg_args *args)
2495{
2496 if (SB_IS_BDEV(args->sb))
2497 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2498 else
2499 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2500
2501 /* 10 jiffies is enough for a delay */
2502 queue_delayed_work(system_wq, &args->reg_work, 10);
2503}
2504
2505static void *alloc_holder_object(struct cache_sb *sb)
2506{
2507 if (SB_IS_BDEV(sb))
2508 return kzalloc(sizeof(struct cached_dev), GFP_KERNEL);
2509 return kzalloc(sizeof(struct cache), GFP_KERNEL);
2510}
2511
2512static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2513 const char *buffer, size_t size)
2514{
2515 const char *err;
2516 char *path = NULL;
2517 struct cache_sb *sb;
2518 struct cache_sb_disk *sb_disk;
2519 struct bdev_handle *bdev_handle, *bdev_handle2;
2520 void *holder = NULL;
2521 ssize_t ret;
2522 bool async_registration = false;
2523 bool quiet = false;
2524
2525#ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2526 async_registration = true;
2527#endif
2528
2529 ret = -EBUSY;
2530 err = "failed to reference bcache module";
2531 if (!try_module_get(THIS_MODULE))
2532 goto out;
2533
2534 /* For latest state of bcache_is_reboot */
2535 smp_mb();
2536 err = "bcache is in reboot";
2537 if (bcache_is_reboot)
2538 goto out_module_put;
2539
2540 ret = -ENOMEM;
2541 err = "cannot allocate memory";
2542 path = kstrndup(buffer, size, GFP_KERNEL);
2543 if (!path)
2544 goto out_module_put;
2545
2546 sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2547 if (!sb)
2548 goto out_free_path;
2549
2550 ret = -EINVAL;
2551 err = "failed to open device";
2552 bdev_handle = bdev_open_by_path(strim(path), BLK_OPEN_READ, NULL, NULL);
2553 if (IS_ERR(bdev_handle))
2554 goto out_free_sb;
2555
2556 err = "failed to set blocksize";
2557 if (set_blocksize(bdev_handle->bdev, 4096))
2558 goto out_blkdev_put;
2559
2560 err = read_super(sb, bdev_handle->bdev, &sb_disk);
2561 if (err)
2562 goto out_blkdev_put;
2563
2564 holder = alloc_holder_object(sb);
2565 if (!holder) {
2566 ret = -ENOMEM;
2567 err = "cannot allocate memory";
2568 goto out_put_sb_page;
2569 }
2570
2571 /* Now reopen in exclusive mode with proper holder */
2572 bdev_handle2 = bdev_open_by_dev(bdev_handle->bdev->bd_dev,
2573 BLK_OPEN_READ | BLK_OPEN_WRITE, holder, NULL);
2574 bdev_release(bdev_handle);
2575 bdev_handle = bdev_handle2;
2576 if (IS_ERR(bdev_handle)) {
2577 ret = PTR_ERR(bdev_handle);
2578 bdev_handle = NULL;
2579 if (ret == -EBUSY) {
2580 dev_t dev;
2581
2582 mutex_lock(&bch_register_lock);
2583 if (lookup_bdev(strim(path), &dev) == 0 &&
2584 bch_is_open(dev))
2585 err = "device already registered";
2586 else
2587 err = "device busy";
2588 mutex_unlock(&bch_register_lock);
2589 if (attr == &ksysfs_register_quiet) {
2590 quiet = true;
2591 ret = size;
2592 }
2593 }
2594 goto out_free_holder;
2595 }
2596
2597 err = "failed to register device";
2598
2599 if (async_registration) {
2600 /* register in asynchronous way */
2601 struct async_reg_args *args =
2602 kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2603
2604 if (!args) {
2605 ret = -ENOMEM;
2606 err = "cannot allocate memory";
2607 goto out_free_holder;
2608 }
2609
2610 args->path = path;
2611 args->sb = sb;
2612 args->sb_disk = sb_disk;
2613 args->bdev_handle = bdev_handle;
2614 args->holder = holder;
2615 register_device_async(args);
2616 /* No wait and returns to user space */
2617 goto async_done;
2618 }
2619
2620 if (SB_IS_BDEV(sb)) {
2621 mutex_lock(&bch_register_lock);
2622 ret = register_bdev(sb, sb_disk, bdev_handle, holder);
2623 mutex_unlock(&bch_register_lock);
2624 /* blkdev_put() will be called in cached_dev_free() */
2625 if (ret < 0)
2626 goto out_free_sb;
2627 } else {
2628 /* blkdev_put() will be called in bch_cache_release() */
2629 ret = register_cache(sb, sb_disk, bdev_handle, holder);
2630 if (ret)
2631 goto out_free_sb;
2632 }
2633
2634 kfree(sb);
2635 kfree(path);
2636 module_put(THIS_MODULE);
2637async_done:
2638 return size;
2639
2640out_free_holder:
2641 kfree(holder);
2642out_put_sb_page:
2643 put_page(virt_to_page(sb_disk));
2644out_blkdev_put:
2645 if (bdev_handle)
2646 bdev_release(bdev_handle);
2647out_free_sb:
2648 kfree(sb);
2649out_free_path:
2650 kfree(path);
2651 path = NULL;
2652out_module_put:
2653 module_put(THIS_MODULE);
2654out:
2655 if (!quiet)
2656 pr_info("error %s: %s\n", path?path:"", err);
2657 return ret;
2658}
2659
2660
2661struct pdev {
2662 struct list_head list;
2663 struct cached_dev *dc;
2664};
2665
2666static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2667 struct kobj_attribute *attr,
2668 const char *buffer,
2669 size_t size)
2670{
2671 LIST_HEAD(pending_devs);
2672 ssize_t ret = size;
2673 struct cached_dev *dc, *tdc;
2674 struct pdev *pdev, *tpdev;
2675 struct cache_set *c, *tc;
2676
2677 mutex_lock(&bch_register_lock);
2678 list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2679 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2680 if (!pdev)
2681 break;
2682 pdev->dc = dc;
2683 list_add(&pdev->list, &pending_devs);
2684 }
2685
2686 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2687 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2688 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2689 char *set_uuid = c->set_uuid;
2690
2691 if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2692 list_del(&pdev->list);
2693 kfree(pdev);
2694 break;
2695 }
2696 }
2697 }
2698 mutex_unlock(&bch_register_lock);
2699
2700 list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2701 pr_info("delete pdev %p\n", pdev);
2702 list_del(&pdev->list);
2703 bcache_device_stop(&pdev->dc->disk);
2704 kfree(pdev);
2705 }
2706
2707 return ret;
2708}
2709
2710static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2711{
2712 if (bcache_is_reboot)
2713 return NOTIFY_DONE;
2714
2715 if (code == SYS_DOWN ||
2716 code == SYS_HALT ||
2717 code == SYS_POWER_OFF) {
2718 DEFINE_WAIT(wait);
2719 unsigned long start = jiffies;
2720 bool stopped = false;
2721
2722 struct cache_set *c, *tc;
2723 struct cached_dev *dc, *tdc;
2724
2725 mutex_lock(&bch_register_lock);
2726
2727 if (bcache_is_reboot)
2728 goto out;
2729
2730 /* New registration is rejected since now */
2731 bcache_is_reboot = true;
2732 /*
2733 * Make registering caller (if there is) on other CPU
2734 * core know bcache_is_reboot set to true earlier
2735 */
2736 smp_mb();
2737
2738 if (list_empty(&bch_cache_sets) &&
2739 list_empty(&uncached_devices))
2740 goto out;
2741
2742 mutex_unlock(&bch_register_lock);
2743
2744 pr_info("Stopping all devices:\n");
2745
2746 /*
2747 * The reason bch_register_lock is not held to call
2748 * bch_cache_set_stop() and bcache_device_stop() is to
2749 * avoid potential deadlock during reboot, because cache
2750 * set or bcache device stopping process will acquire
2751 * bch_register_lock too.
2752 *
2753 * We are safe here because bcache_is_reboot sets to
2754 * true already, register_bcache() will reject new
2755 * registration now. bcache_is_reboot also makes sure
2756 * bcache_reboot() won't be re-entered on by other thread,
2757 * so there is no race in following list iteration by
2758 * list_for_each_entry_safe().
2759 */
2760 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2761 bch_cache_set_stop(c);
2762
2763 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2764 bcache_device_stop(&dc->disk);
2765
2766
2767 /*
2768 * Give an early chance for other kthreads and
2769 * kworkers to stop themselves
2770 */
2771 schedule();
2772
2773 /* What's a condition variable? */
2774 while (1) {
2775 long timeout = start + 10 * HZ - jiffies;
2776
2777 mutex_lock(&bch_register_lock);
2778 stopped = list_empty(&bch_cache_sets) &&
2779 list_empty(&uncached_devices);
2780
2781 if (timeout < 0 || stopped)
2782 break;
2783
2784 prepare_to_wait(&unregister_wait, &wait,
2785 TASK_UNINTERRUPTIBLE);
2786
2787 mutex_unlock(&bch_register_lock);
2788 schedule_timeout(timeout);
2789 }
2790
2791 finish_wait(&unregister_wait, &wait);
2792
2793 if (stopped)
2794 pr_info("All devices stopped\n");
2795 else
2796 pr_notice("Timeout waiting for devices to be closed\n");
2797out:
2798 mutex_unlock(&bch_register_lock);
2799 }
2800
2801 return NOTIFY_DONE;
2802}
2803
2804static struct notifier_block reboot = {
2805 .notifier_call = bcache_reboot,
2806 .priority = INT_MAX, /* before any real devices */
2807};
2808
2809static void bcache_exit(void)
2810{
2811 bch_debug_exit();
2812 bch_request_exit();
2813 if (bcache_kobj)
2814 kobject_put(bcache_kobj);
2815 if (bcache_wq)
2816 destroy_workqueue(bcache_wq);
2817 if (bch_journal_wq)
2818 destroy_workqueue(bch_journal_wq);
2819 if (bch_flush_wq)
2820 destroy_workqueue(bch_flush_wq);
2821 bch_btree_exit();
2822
2823 if (bcache_major)
2824 unregister_blkdev(bcache_major, "bcache");
2825 unregister_reboot_notifier(&reboot);
2826 mutex_destroy(&bch_register_lock);
2827}
2828
2829/* Check and fixup module parameters */
2830static void check_module_parameters(void)
2831{
2832 if (bch_cutoff_writeback_sync == 0)
2833 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2834 else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2835 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2836 bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2837 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2838 }
2839
2840 if (bch_cutoff_writeback == 0)
2841 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2842 else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2843 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2844 bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2845 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2846 }
2847
2848 if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2849 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2850 bch_cutoff_writeback, bch_cutoff_writeback_sync);
2851 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2852 }
2853}
2854
2855static int __init bcache_init(void)
2856{
2857 static const struct attribute *files[] = {
2858 &ksysfs_register.attr,
2859 &ksysfs_register_quiet.attr,
2860 &ksysfs_pendings_cleanup.attr,
2861 NULL
2862 };
2863
2864 check_module_parameters();
2865
2866 mutex_init(&bch_register_lock);
2867 init_waitqueue_head(&unregister_wait);
2868 register_reboot_notifier(&reboot);
2869
2870 bcache_major = register_blkdev(0, "bcache");
2871 if (bcache_major < 0) {
2872 unregister_reboot_notifier(&reboot);
2873 mutex_destroy(&bch_register_lock);
2874 return bcache_major;
2875 }
2876
2877 if (bch_btree_init())
2878 goto err;
2879
2880 bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2881 if (!bcache_wq)
2882 goto err;
2883
2884 /*
2885 * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2886 *
2887 * 1. It used `system_wq` before which also does no memory reclaim.
2888 * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2889 * reduced throughput can be observed.
2890 *
2891 * We still want to user our own queue to not congest the `system_wq`.
2892 */
2893 bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2894 if (!bch_flush_wq)
2895 goto err;
2896
2897 bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2898 if (!bch_journal_wq)
2899 goto err;
2900
2901 bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2902 if (!bcache_kobj)
2903 goto err;
2904
2905 if (bch_request_init() ||
2906 sysfs_create_files(bcache_kobj, files))
2907 goto err;
2908
2909 bch_debug_init();
2910
2911 bcache_is_reboot = false;
2912
2913 return 0;
2914err:
2915 bcache_exit();
2916 return -ENOMEM;
2917}
2918
2919/*
2920 * Module hooks
2921 */
2922module_exit(bcache_exit);
2923module_init(bcache_init);
2924
2925module_param(bch_cutoff_writeback, uint, 0);
2926MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2927
2928module_param(bch_cutoff_writeback_sync, uint, 0);
2929MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2930
2931MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2932MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2933MODULE_LICENSE("GPL");
1/*
2 * bcache setup/teardown code, and some metadata io - read a superblock and
3 * figure out what to do with it.
4 *
5 * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
6 * Copyright 2012 Google, Inc.
7 */
8
9#include "bcache.h"
10#include "btree.h"
11#include "debug.h"
12#include "extents.h"
13#include "request.h"
14#include "writeback.h"
15
16#include <linux/blkdev.h>
17#include <linux/buffer_head.h>
18#include <linux/debugfs.h>
19#include <linux/genhd.h>
20#include <linux/idr.h>
21#include <linux/kthread.h>
22#include <linux/module.h>
23#include <linux/random.h>
24#include <linux/reboot.h>
25#include <linux/sysfs.h>
26
27MODULE_LICENSE("GPL");
28MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
29
30static const char bcache_magic[] = {
31 0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
32 0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
33};
34
35static const char invalid_uuid[] = {
36 0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
37 0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
38};
39
40/* Default is -1; we skip past it for struct cached_dev's cache mode */
41const char * const bch_cache_modes[] = {
42 "default",
43 "writethrough",
44 "writeback",
45 "writearound",
46 "none",
47 NULL
48};
49
50static struct kobject *bcache_kobj;
51struct mutex bch_register_lock;
52LIST_HEAD(bch_cache_sets);
53static LIST_HEAD(uncached_devices);
54
55static int bcache_major;
56static DEFINE_IDA(bcache_minor);
57static wait_queue_head_t unregister_wait;
58struct workqueue_struct *bcache_wq;
59
60#define BTREE_MAX_PAGES (256 * 1024 / PAGE_SIZE)
61
62/* Superblock */
63
64static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
65 struct page **res)
66{
67 const char *err;
68 struct cache_sb *s;
69 struct buffer_head *bh = __bread(bdev, 1, SB_SIZE);
70 unsigned i;
71
72 if (!bh)
73 return "IO error";
74
75 s = (struct cache_sb *) bh->b_data;
76
77 sb->offset = le64_to_cpu(s->offset);
78 sb->version = le64_to_cpu(s->version);
79
80 memcpy(sb->magic, s->magic, 16);
81 memcpy(sb->uuid, s->uuid, 16);
82 memcpy(sb->set_uuid, s->set_uuid, 16);
83 memcpy(sb->label, s->label, SB_LABEL_SIZE);
84
85 sb->flags = le64_to_cpu(s->flags);
86 sb->seq = le64_to_cpu(s->seq);
87 sb->last_mount = le32_to_cpu(s->last_mount);
88 sb->first_bucket = le16_to_cpu(s->first_bucket);
89 sb->keys = le16_to_cpu(s->keys);
90
91 for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
92 sb->d[i] = le64_to_cpu(s->d[i]);
93
94 pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u",
95 sb->version, sb->flags, sb->seq, sb->keys);
96
97 err = "Not a bcache superblock";
98 if (sb->offset != SB_SECTOR)
99 goto err;
100
101 if (memcmp(sb->magic, bcache_magic, 16))
102 goto err;
103
104 err = "Too many journal buckets";
105 if (sb->keys > SB_JOURNAL_BUCKETS)
106 goto err;
107
108 err = "Bad checksum";
109 if (s->csum != csum_set(s))
110 goto err;
111
112 err = "Bad UUID";
113 if (bch_is_zero(sb->uuid, 16))
114 goto err;
115
116 sb->block_size = le16_to_cpu(s->block_size);
117
118 err = "Superblock block size smaller than device block size";
119 if (sb->block_size << 9 < bdev_logical_block_size(bdev))
120 goto err;
121
122 switch (sb->version) {
123 case BCACHE_SB_VERSION_BDEV:
124 sb->data_offset = BDEV_DATA_START_DEFAULT;
125 break;
126 case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
127 sb->data_offset = le64_to_cpu(s->data_offset);
128
129 err = "Bad data offset";
130 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
131 goto err;
132
133 break;
134 case BCACHE_SB_VERSION_CDEV:
135 case BCACHE_SB_VERSION_CDEV_WITH_UUID:
136 sb->nbuckets = le64_to_cpu(s->nbuckets);
137 sb->block_size = le16_to_cpu(s->block_size);
138 sb->bucket_size = le16_to_cpu(s->bucket_size);
139
140 sb->nr_in_set = le16_to_cpu(s->nr_in_set);
141 sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
142
143 err = "Too many buckets";
144 if (sb->nbuckets > LONG_MAX)
145 goto err;
146
147 err = "Not enough buckets";
148 if (sb->nbuckets < 1 << 7)
149 goto err;
150
151 err = "Bad block/bucket size";
152 if (!is_power_of_2(sb->block_size) ||
153 sb->block_size > PAGE_SECTORS ||
154 !is_power_of_2(sb->bucket_size) ||
155 sb->bucket_size < PAGE_SECTORS)
156 goto err;
157
158 err = "Invalid superblock: device too small";
159 if (get_capacity(bdev->bd_disk) < sb->bucket_size * sb->nbuckets)
160 goto err;
161
162 err = "Bad UUID";
163 if (bch_is_zero(sb->set_uuid, 16))
164 goto err;
165
166 err = "Bad cache device number in set";
167 if (!sb->nr_in_set ||
168 sb->nr_in_set <= sb->nr_this_dev ||
169 sb->nr_in_set > MAX_CACHES_PER_SET)
170 goto err;
171
172 err = "Journal buckets not sequential";
173 for (i = 0; i < sb->keys; i++)
174 if (sb->d[i] != sb->first_bucket + i)
175 goto err;
176
177 err = "Too many journal buckets";
178 if (sb->first_bucket + sb->keys > sb->nbuckets)
179 goto err;
180
181 err = "Invalid superblock: first bucket comes before end of super";
182 if (sb->first_bucket * sb->bucket_size < 16)
183 goto err;
184
185 break;
186 default:
187 err = "Unsupported superblock version";
188 goto err;
189 }
190
191 sb->last_mount = get_seconds();
192 err = NULL;
193
194 get_page(bh->b_page);
195 *res = bh->b_page;
196err:
197 put_bh(bh);
198 return err;
199}
200
201static void write_bdev_super_endio(struct bio *bio)
202{
203 struct cached_dev *dc = bio->bi_private;
204 /* XXX: error checking */
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->bi_io_vec[0].bv_page);
212 unsigned i;
213
214 bio->bi_iter.bi_sector = SB_SECTOR;
215 bio->bi_rw = REQ_SYNC|REQ_META;
216 bio->bi_iter.bi_size = SB_SIZE;
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(REQ_WRITE, 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->bi_bdev = dc->bdev;
261 bio->bi_end_io = write_bdev_super_endio;
262 bio->bi_private = dc;
263
264 closure_get(cl);
265 __write_super(&dc->sb, bio);
266
267 closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
268}
269
270static void write_super_endio(struct bio *bio)
271{
272 struct cache *ca = bio->bi_private;
273
274 bch_count_io_errors(ca, bio->bi_error, "writing superblock");
275 closure_put(&ca->set->sb_write);
276}
277
278static void bcache_write_super_unlock(struct closure *cl)
279{
280 struct cache_set *c = container_of(cl, struct cache_set, sb_write);
281
282 up(&c->sb_write_mutex);
283}
284
285void bcache_write_super(struct cache_set *c)
286{
287 struct closure *cl = &c->sb_write;
288 struct cache *ca;
289 unsigned i;
290
291 down(&c->sb_write_mutex);
292 closure_init(cl, &c->cl);
293
294 c->sb.seq++;
295
296 for_each_cache(ca, c, i) {
297 struct bio *bio = &ca->sb_bio;
298
299 ca->sb.version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
300 ca->sb.seq = c->sb.seq;
301 ca->sb.last_mount = c->sb.last_mount;
302
303 SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
304
305 bio_reset(bio);
306 bio->bi_bdev = ca->bdev;
307 bio->bi_end_io = write_super_endio;
308 bio->bi_private = ca;
309
310 closure_get(cl);
311 __write_super(&ca->sb, bio);
312 }
313
314 closure_return_with_destructor(cl, bcache_write_super_unlock);
315}
316
317/* UUID io */
318
319static void uuid_endio(struct bio *bio)
320{
321 struct closure *cl = bio->bi_private;
322 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
323
324 cache_set_err_on(bio->bi_error, c, "accessing uuids");
325 bch_bbio_free(bio, c);
326 closure_put(cl);
327}
328
329static void uuid_io_unlock(struct closure *cl)
330{
331 struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
332
333 up(&c->uuid_write_mutex);
334}
335
336static void uuid_io(struct cache_set *c, unsigned long rw,
337 struct bkey *k, struct closure *parent)
338{
339 struct closure *cl = &c->uuid_write;
340 struct uuid_entry *u;
341 unsigned i;
342 char buf[80];
343
344 BUG_ON(!parent);
345 down(&c->uuid_write_mutex);
346 closure_init(cl, parent);
347
348 for (i = 0; i < KEY_PTRS(k); i++) {
349 struct bio *bio = bch_bbio_alloc(c);
350
351 bio->bi_rw = REQ_SYNC|REQ_META|rw;
352 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
353
354 bio->bi_end_io = uuid_endio;
355 bio->bi_private = cl;
356 bch_bio_map(bio, c->uuids);
357
358 bch_submit_bbio(bio, c, k, i);
359
360 if (!(rw & WRITE))
361 break;
362 }
363
364 bch_extent_to_text(buf, sizeof(buf), k);
365 pr_debug("%s UUIDs at %s", rw & REQ_WRITE ? "wrote" : "read", buf);
366
367 for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
368 if (!bch_is_zero(u->uuid, 16))
369 pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u",
370 u - c->uuids, u->uuid, u->label,
371 u->first_reg, u->last_reg, u->invalidated);
372
373 closure_return_with_destructor(cl, uuid_io_unlock);
374}
375
376static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
377{
378 struct bkey *k = &j->uuid_bucket;
379
380 if (__bch_btree_ptr_invalid(c, k))
381 return "bad uuid pointer";
382
383 bkey_copy(&c->uuid_bucket, k);
384 uuid_io(c, READ_SYNC, k, cl);
385
386 if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
387 struct uuid_entry_v0 *u0 = (void *) c->uuids;
388 struct uuid_entry *u1 = (void *) c->uuids;
389 int i;
390
391 closure_sync(cl);
392
393 /*
394 * Since the new uuid entry is bigger than the old, we have to
395 * convert starting at the highest memory address and work down
396 * in order to do it in place
397 */
398
399 for (i = c->nr_uuids - 1;
400 i >= 0;
401 --i) {
402 memcpy(u1[i].uuid, u0[i].uuid, 16);
403 memcpy(u1[i].label, u0[i].label, 32);
404
405 u1[i].first_reg = u0[i].first_reg;
406 u1[i].last_reg = u0[i].last_reg;
407 u1[i].invalidated = u0[i].invalidated;
408
409 u1[i].flags = 0;
410 u1[i].sectors = 0;
411 }
412 }
413
414 return NULL;
415}
416
417static int __uuid_write(struct cache_set *c)
418{
419 BKEY_PADDED(key) k;
420 struct closure cl;
421 closure_init_stack(&cl);
422
423 lockdep_assert_held(&bch_register_lock);
424
425 if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
426 return 1;
427
428 SET_KEY_SIZE(&k.key, c->sb.bucket_size);
429 uuid_io(c, REQ_WRITE, &k.key, &cl);
430 closure_sync(&cl);
431
432 bkey_copy(&c->uuid_bucket, &k.key);
433 bkey_put(c, &k.key);
434 return 0;
435}
436
437int bch_uuid_write(struct cache_set *c)
438{
439 int ret = __uuid_write(c);
440
441 if (!ret)
442 bch_journal_meta(c, NULL);
443
444 return ret;
445}
446
447static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
448{
449 struct uuid_entry *u;
450
451 for (u = c->uuids;
452 u < c->uuids + c->nr_uuids; u++)
453 if (!memcmp(u->uuid, uuid, 16))
454 return u;
455
456 return NULL;
457}
458
459static struct uuid_entry *uuid_find_empty(struct cache_set *c)
460{
461 static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
462 return uuid_find(c, zero_uuid);
463}
464
465/*
466 * Bucket priorities/gens:
467 *
468 * For each bucket, we store on disk its
469 * 8 bit gen
470 * 16 bit priority
471 *
472 * See alloc.c for an explanation of the gen. The priority is used to implement
473 * lru (and in the future other) cache replacement policies; for most purposes
474 * it's just an opaque integer.
475 *
476 * The gens and the priorities don't have a whole lot to do with each other, and
477 * it's actually the gens that must be written out at specific times - it's no
478 * big deal if the priorities don't get written, if we lose them we just reuse
479 * buckets in suboptimal order.
480 *
481 * On disk they're stored in a packed array, and in as many buckets are required
482 * to fit them all. The buckets we use to store them form a list; the journal
483 * header points to the first bucket, the first bucket points to the second
484 * bucket, et cetera.
485 *
486 * This code is used by the allocation code; periodically (whenever it runs out
487 * of buckets to allocate from) the allocation code will invalidate some
488 * buckets, but it can't use those buckets until their new gens are safely on
489 * disk.
490 */
491
492static void prio_endio(struct bio *bio)
493{
494 struct cache *ca = bio->bi_private;
495
496 cache_set_err_on(bio->bi_error, ca->set, "accessing priorities");
497 bch_bbio_free(bio, ca->set);
498 closure_put(&ca->prio);
499}
500
501static void prio_io(struct cache *ca, uint64_t bucket, unsigned long rw)
502{
503 struct closure *cl = &ca->prio;
504 struct bio *bio = bch_bbio_alloc(ca->set);
505
506 closure_init_stack(cl);
507
508 bio->bi_iter.bi_sector = bucket * ca->sb.bucket_size;
509 bio->bi_bdev = ca->bdev;
510 bio->bi_rw = REQ_SYNC|REQ_META|rw;
511 bio->bi_iter.bi_size = bucket_bytes(ca);
512
513 bio->bi_end_io = prio_endio;
514 bio->bi_private = ca;
515 bch_bio_map(bio, ca->disk_buckets);
516
517 closure_bio_submit(bio, &ca->prio);
518 closure_sync(cl);
519}
520
521void bch_prio_write(struct cache *ca)
522{
523 int i;
524 struct bucket *b;
525 struct closure cl;
526
527 closure_init_stack(&cl);
528
529 lockdep_assert_held(&ca->set->bucket_lock);
530
531 ca->disk_buckets->seq++;
532
533 atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
534 &ca->meta_sectors_written);
535
536 //pr_debug("free %zu, free_inc %zu, unused %zu", fifo_used(&ca->free),
537 // fifo_used(&ca->free_inc), fifo_used(&ca->unused));
538
539 for (i = prio_buckets(ca) - 1; i >= 0; --i) {
540 long bucket;
541 struct prio_set *p = ca->disk_buckets;
542 struct bucket_disk *d = p->data;
543 struct bucket_disk *end = d + prios_per_bucket(ca);
544
545 for (b = ca->buckets + i * prios_per_bucket(ca);
546 b < ca->buckets + ca->sb.nbuckets && d < end;
547 b++, d++) {
548 d->prio = cpu_to_le16(b->prio);
549 d->gen = b->gen;
550 }
551
552 p->next_bucket = ca->prio_buckets[i + 1];
553 p->magic = pset_magic(&ca->sb);
554 p->csum = bch_crc64(&p->magic, bucket_bytes(ca) - 8);
555
556 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, true);
557 BUG_ON(bucket == -1);
558
559 mutex_unlock(&ca->set->bucket_lock);
560 prio_io(ca, bucket, REQ_WRITE);
561 mutex_lock(&ca->set->bucket_lock);
562
563 ca->prio_buckets[i] = bucket;
564 atomic_dec_bug(&ca->buckets[bucket].pin);
565 }
566
567 mutex_unlock(&ca->set->bucket_lock);
568
569 bch_journal_meta(ca->set, &cl);
570 closure_sync(&cl);
571
572 mutex_lock(&ca->set->bucket_lock);
573
574 /*
575 * Don't want the old priorities to get garbage collected until after we
576 * finish writing the new ones, and they're journalled
577 */
578 for (i = 0; i < prio_buckets(ca); i++) {
579 if (ca->prio_last_buckets[i])
580 __bch_bucket_free(ca,
581 &ca->buckets[ca->prio_last_buckets[i]]);
582
583 ca->prio_last_buckets[i] = ca->prio_buckets[i];
584 }
585}
586
587static void prio_read(struct cache *ca, uint64_t bucket)
588{
589 struct prio_set *p = ca->disk_buckets;
590 struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
591 struct bucket *b;
592 unsigned bucket_nr = 0;
593
594 for (b = ca->buckets;
595 b < ca->buckets + ca->sb.nbuckets;
596 b++, d++) {
597 if (d == end) {
598 ca->prio_buckets[bucket_nr] = bucket;
599 ca->prio_last_buckets[bucket_nr] = bucket;
600 bucket_nr++;
601
602 prio_io(ca, bucket, READ_SYNC);
603
604 if (p->csum != bch_crc64(&p->magic, bucket_bytes(ca) - 8))
605 pr_warn("bad csum reading priorities");
606
607 if (p->magic != pset_magic(&ca->sb))
608 pr_warn("bad magic reading priorities");
609
610 bucket = p->next_bucket;
611 d = p->data;
612 }
613
614 b->prio = le16_to_cpu(d->prio);
615 b->gen = b->last_gc = d->gen;
616 }
617}
618
619/* Bcache device */
620
621static int open_dev(struct block_device *b, fmode_t mode)
622{
623 struct bcache_device *d = b->bd_disk->private_data;
624 if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
625 return -ENXIO;
626
627 closure_get(&d->cl);
628 return 0;
629}
630
631static void release_dev(struct gendisk *b, fmode_t mode)
632{
633 struct bcache_device *d = b->private_data;
634 closure_put(&d->cl);
635}
636
637static int ioctl_dev(struct block_device *b, fmode_t mode,
638 unsigned int cmd, unsigned long arg)
639{
640 struct bcache_device *d = b->bd_disk->private_data;
641 return d->ioctl(d, mode, cmd, arg);
642}
643
644static const struct block_device_operations bcache_ops = {
645 .open = open_dev,
646 .release = release_dev,
647 .ioctl = ioctl_dev,
648 .owner = THIS_MODULE,
649};
650
651void bcache_device_stop(struct bcache_device *d)
652{
653 if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
654 closure_queue(&d->cl);
655}
656
657static void bcache_device_unlink(struct bcache_device *d)
658{
659 lockdep_assert_held(&bch_register_lock);
660
661 if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
662 unsigned i;
663 struct cache *ca;
664
665 sysfs_remove_link(&d->c->kobj, d->name);
666 sysfs_remove_link(&d->kobj, "cache");
667
668 for_each_cache(ca, d->c, i)
669 bd_unlink_disk_holder(ca->bdev, d->disk);
670 }
671}
672
673static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
674 const char *name)
675{
676 unsigned i;
677 struct cache *ca;
678
679 for_each_cache(ca, d->c, i)
680 bd_link_disk_holder(ca->bdev, d->disk);
681
682 snprintf(d->name, BCACHEDEVNAME_SIZE,
683 "%s%u", name, d->id);
684
685 WARN(sysfs_create_link(&d->kobj, &c->kobj, "cache") ||
686 sysfs_create_link(&c->kobj, &d->kobj, d->name),
687 "Couldn't create device <-> cache set symlinks");
688
689 clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
690}
691
692static void bcache_device_detach(struct bcache_device *d)
693{
694 lockdep_assert_held(&bch_register_lock);
695
696 if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
697 struct uuid_entry *u = d->c->uuids + d->id;
698
699 SET_UUID_FLASH_ONLY(u, 0);
700 memcpy(u->uuid, invalid_uuid, 16);
701 u->invalidated = cpu_to_le32(get_seconds());
702 bch_uuid_write(d->c);
703 }
704
705 bcache_device_unlink(d);
706
707 d->c->devices[d->id] = NULL;
708 closure_put(&d->c->caching);
709 d->c = NULL;
710}
711
712static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
713 unsigned id)
714{
715 d->id = id;
716 d->c = c;
717 c->devices[id] = d;
718
719 closure_get(&c->caching);
720}
721
722static void bcache_device_free(struct bcache_device *d)
723{
724 lockdep_assert_held(&bch_register_lock);
725
726 pr_info("%s stopped", d->disk->disk_name);
727
728 if (d->c)
729 bcache_device_detach(d);
730 if (d->disk && d->disk->flags & GENHD_FL_UP)
731 del_gendisk(d->disk);
732 if (d->disk && d->disk->queue)
733 blk_cleanup_queue(d->disk->queue);
734 if (d->disk) {
735 ida_simple_remove(&bcache_minor, d->disk->first_minor);
736 put_disk(d->disk);
737 }
738
739 if (d->bio_split)
740 bioset_free(d->bio_split);
741 kvfree(d->full_dirty_stripes);
742 kvfree(d->stripe_sectors_dirty);
743
744 closure_debug_destroy(&d->cl);
745}
746
747static int bcache_device_init(struct bcache_device *d, unsigned block_size,
748 sector_t sectors)
749{
750 struct request_queue *q;
751 size_t n;
752 int minor;
753
754 if (!d->stripe_size)
755 d->stripe_size = 1 << 31;
756
757 d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
758
759 if (!d->nr_stripes ||
760 d->nr_stripes > INT_MAX ||
761 d->nr_stripes > SIZE_MAX / sizeof(atomic_t)) {
762 pr_err("nr_stripes too large");
763 return -ENOMEM;
764 }
765
766 n = d->nr_stripes * sizeof(atomic_t);
767 d->stripe_sectors_dirty = n < PAGE_SIZE << 6
768 ? kzalloc(n, GFP_KERNEL)
769 : vzalloc(n);
770 if (!d->stripe_sectors_dirty)
771 return -ENOMEM;
772
773 n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
774 d->full_dirty_stripes = n < PAGE_SIZE << 6
775 ? kzalloc(n, GFP_KERNEL)
776 : vzalloc(n);
777 if (!d->full_dirty_stripes)
778 return -ENOMEM;
779
780 minor = ida_simple_get(&bcache_minor, 0, MINORMASK + 1, GFP_KERNEL);
781 if (minor < 0)
782 return minor;
783
784 if (!(d->bio_split = bioset_create(4, offsetof(struct bbio, bio))) ||
785 !(d->disk = alloc_disk(1))) {
786 ida_simple_remove(&bcache_minor, minor);
787 return -ENOMEM;
788 }
789
790 set_capacity(d->disk, sectors);
791 snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", minor);
792
793 d->disk->major = bcache_major;
794 d->disk->first_minor = minor;
795 d->disk->fops = &bcache_ops;
796 d->disk->private_data = d;
797
798 q = blk_alloc_queue(GFP_KERNEL);
799 if (!q)
800 return -ENOMEM;
801
802 blk_queue_make_request(q, NULL);
803 d->disk->queue = q;
804 q->queuedata = d;
805 q->backing_dev_info.congested_data = d;
806 q->limits.max_hw_sectors = UINT_MAX;
807 q->limits.max_sectors = UINT_MAX;
808 q->limits.max_segment_size = UINT_MAX;
809 q->limits.max_segments = BIO_MAX_PAGES;
810 blk_queue_max_discard_sectors(q, UINT_MAX);
811 q->limits.discard_granularity = 512;
812 q->limits.io_min = block_size;
813 q->limits.logical_block_size = block_size;
814 q->limits.physical_block_size = block_size;
815 set_bit(QUEUE_FLAG_NONROT, &d->disk->queue->queue_flags);
816 clear_bit(QUEUE_FLAG_ADD_RANDOM, &d->disk->queue->queue_flags);
817 set_bit(QUEUE_FLAG_DISCARD, &d->disk->queue->queue_flags);
818
819 blk_queue_flush(q, REQ_FLUSH|REQ_FUA);
820
821 return 0;
822}
823
824/* Cached device */
825
826static void calc_cached_dev_sectors(struct cache_set *c)
827{
828 uint64_t sectors = 0;
829 struct cached_dev *dc;
830
831 list_for_each_entry(dc, &c->cached_devs, list)
832 sectors += bdev_sectors(dc->bdev);
833
834 c->cached_dev_sectors = sectors;
835}
836
837void bch_cached_dev_run(struct cached_dev *dc)
838{
839 struct bcache_device *d = &dc->disk;
840 char buf[SB_LABEL_SIZE + 1];
841 char *env[] = {
842 "DRIVER=bcache",
843 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
844 NULL,
845 NULL,
846 };
847
848 memcpy(buf, dc->sb.label, SB_LABEL_SIZE);
849 buf[SB_LABEL_SIZE] = '\0';
850 env[2] = kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf);
851
852 if (atomic_xchg(&dc->running, 1)) {
853 kfree(env[1]);
854 kfree(env[2]);
855 return;
856 }
857
858 if (!d->c &&
859 BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
860 struct closure cl;
861 closure_init_stack(&cl);
862
863 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
864 bch_write_bdev_super(dc, &cl);
865 closure_sync(&cl);
866 }
867
868 add_disk(d->disk);
869 bd_link_disk_holder(dc->bdev, dc->disk.disk);
870 /* won't show up in the uevent file, use udevadm monitor -e instead
871 * only class / kset properties are persistent */
872 kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
873 kfree(env[1]);
874 kfree(env[2]);
875
876 if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
877 sysfs_create_link(&disk_to_dev(d->disk)->kobj, &d->kobj, "bcache"))
878 pr_debug("error creating sysfs link");
879}
880
881static void cached_dev_detach_finish(struct work_struct *w)
882{
883 struct cached_dev *dc = container_of(w, struct cached_dev, detach);
884 char buf[BDEVNAME_SIZE];
885 struct closure cl;
886 closure_init_stack(&cl);
887
888 BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
889 BUG_ON(atomic_read(&dc->count));
890
891 mutex_lock(&bch_register_lock);
892
893 memset(&dc->sb.set_uuid, 0, 16);
894 SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
895
896 bch_write_bdev_super(dc, &cl);
897 closure_sync(&cl);
898
899 bcache_device_detach(&dc->disk);
900 list_move(&dc->list, &uncached_devices);
901
902 clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
903 clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
904
905 mutex_unlock(&bch_register_lock);
906
907 pr_info("Caching disabled for %s", bdevname(dc->bdev, buf));
908
909 /* Drop ref we took in cached_dev_detach() */
910 closure_put(&dc->disk.cl);
911}
912
913void bch_cached_dev_detach(struct cached_dev *dc)
914{
915 lockdep_assert_held(&bch_register_lock);
916
917 if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
918 return;
919
920 if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
921 return;
922
923 /*
924 * Block the device from being closed and freed until we're finished
925 * detaching
926 */
927 closure_get(&dc->disk.cl);
928
929 bch_writeback_queue(dc);
930 cached_dev_put(dc);
931}
932
933int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c)
934{
935 uint32_t rtime = cpu_to_le32(get_seconds());
936 struct uuid_entry *u;
937 char buf[BDEVNAME_SIZE];
938
939 bdevname(dc->bdev, buf);
940
941 if (memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16))
942 return -ENOENT;
943
944 if (dc->disk.c) {
945 pr_err("Can't attach %s: already attached", buf);
946 return -EINVAL;
947 }
948
949 if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
950 pr_err("Can't attach %s: shutting down", buf);
951 return -EINVAL;
952 }
953
954 if (dc->sb.block_size < c->sb.block_size) {
955 /* Will die */
956 pr_err("Couldn't attach %s: block size less than set's block size",
957 buf);
958 return -EINVAL;
959 }
960
961 u = uuid_find(c, dc->sb.uuid);
962
963 if (u &&
964 (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
965 BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
966 memcpy(u->uuid, invalid_uuid, 16);
967 u->invalidated = cpu_to_le32(get_seconds());
968 u = NULL;
969 }
970
971 if (!u) {
972 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
973 pr_err("Couldn't find uuid for %s in set", buf);
974 return -ENOENT;
975 }
976
977 u = uuid_find_empty(c);
978 if (!u) {
979 pr_err("Not caching %s, no room for UUID", buf);
980 return -EINVAL;
981 }
982 }
983
984 /* Deadlocks since we're called via sysfs...
985 sysfs_remove_file(&dc->kobj, &sysfs_attach);
986 */
987
988 if (bch_is_zero(u->uuid, 16)) {
989 struct closure cl;
990 closure_init_stack(&cl);
991
992 memcpy(u->uuid, dc->sb.uuid, 16);
993 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
994 u->first_reg = u->last_reg = rtime;
995 bch_uuid_write(c);
996
997 memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
998 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
999
1000 bch_write_bdev_super(dc, &cl);
1001 closure_sync(&cl);
1002 } else {
1003 u->last_reg = rtime;
1004 bch_uuid_write(c);
1005 }
1006
1007 bcache_device_attach(&dc->disk, c, u - c->uuids);
1008 list_move(&dc->list, &c->cached_devs);
1009 calc_cached_dev_sectors(c);
1010
1011 smp_wmb();
1012 /*
1013 * dc->c must be set before dc->count != 0 - paired with the mb in
1014 * cached_dev_get()
1015 */
1016 atomic_set(&dc->count, 1);
1017
1018 /* Block writeback thread, but spawn it */
1019 down_write(&dc->writeback_lock);
1020 if (bch_cached_dev_writeback_start(dc)) {
1021 up_write(&dc->writeback_lock);
1022 return -ENOMEM;
1023 }
1024
1025 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1026 bch_sectors_dirty_init(dc);
1027 atomic_set(&dc->has_dirty, 1);
1028 atomic_inc(&dc->count);
1029 bch_writeback_queue(dc);
1030 }
1031
1032 bch_cached_dev_run(dc);
1033 bcache_device_link(&dc->disk, c, "bdev");
1034
1035 /* Allow the writeback thread to proceed */
1036 up_write(&dc->writeback_lock);
1037
1038 pr_info("Caching %s as %s on set %pU",
1039 bdevname(dc->bdev, buf), dc->disk.disk->disk_name,
1040 dc->disk.c->sb.set_uuid);
1041 return 0;
1042}
1043
1044void bch_cached_dev_release(struct kobject *kobj)
1045{
1046 struct cached_dev *dc = container_of(kobj, struct cached_dev,
1047 disk.kobj);
1048 kfree(dc);
1049 module_put(THIS_MODULE);
1050}
1051
1052static void cached_dev_free(struct closure *cl)
1053{
1054 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1055
1056 cancel_delayed_work_sync(&dc->writeback_rate_update);
1057 if (!IS_ERR_OR_NULL(dc->writeback_thread))
1058 kthread_stop(dc->writeback_thread);
1059
1060 mutex_lock(&bch_register_lock);
1061
1062 if (atomic_read(&dc->running))
1063 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1064 bcache_device_free(&dc->disk);
1065 list_del(&dc->list);
1066
1067 mutex_unlock(&bch_register_lock);
1068
1069 if (!IS_ERR_OR_NULL(dc->bdev))
1070 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1071
1072 wake_up(&unregister_wait);
1073
1074 kobject_put(&dc->disk.kobj);
1075}
1076
1077static void cached_dev_flush(struct closure *cl)
1078{
1079 struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1080 struct bcache_device *d = &dc->disk;
1081
1082 mutex_lock(&bch_register_lock);
1083 bcache_device_unlink(d);
1084 mutex_unlock(&bch_register_lock);
1085
1086 bch_cache_accounting_destroy(&dc->accounting);
1087 kobject_del(&d->kobj);
1088
1089 continue_at(cl, cached_dev_free, system_wq);
1090}
1091
1092static int cached_dev_init(struct cached_dev *dc, unsigned block_size)
1093{
1094 int ret;
1095 struct io *io;
1096 struct request_queue *q = bdev_get_queue(dc->bdev);
1097
1098 __module_get(THIS_MODULE);
1099 INIT_LIST_HEAD(&dc->list);
1100 closure_init(&dc->disk.cl, NULL);
1101 set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1102 kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1103 INIT_WORK(&dc->detach, cached_dev_detach_finish);
1104 sema_init(&dc->sb_write_mutex, 1);
1105 INIT_LIST_HEAD(&dc->io_lru);
1106 spin_lock_init(&dc->io_lock);
1107 bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1108
1109 dc->sequential_cutoff = 4 << 20;
1110
1111 for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1112 list_add(&io->lru, &dc->io_lru);
1113 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1114 }
1115
1116 dc->disk.stripe_size = q->limits.io_opt >> 9;
1117
1118 if (dc->disk.stripe_size)
1119 dc->partial_stripes_expensive =
1120 q->limits.raid_partial_stripes_expensive;
1121
1122 ret = bcache_device_init(&dc->disk, block_size,
1123 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1124 if (ret)
1125 return ret;
1126
1127 set_capacity(dc->disk.disk,
1128 dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1129
1130 dc->disk.disk->queue->backing_dev_info.ra_pages =
1131 max(dc->disk.disk->queue->backing_dev_info.ra_pages,
1132 q->backing_dev_info.ra_pages);
1133
1134 bch_cached_dev_request_init(dc);
1135 bch_cached_dev_writeback_init(dc);
1136 return 0;
1137}
1138
1139/* Cached device - bcache superblock */
1140
1141static void register_bdev(struct cache_sb *sb, struct page *sb_page,
1142 struct block_device *bdev,
1143 struct cached_dev *dc)
1144{
1145 char name[BDEVNAME_SIZE];
1146 const char *err = "cannot allocate memory";
1147 struct cache_set *c;
1148
1149 memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1150 dc->bdev = bdev;
1151 dc->bdev->bd_holder = dc;
1152
1153 bio_init(&dc->sb_bio);
1154 dc->sb_bio.bi_max_vecs = 1;
1155 dc->sb_bio.bi_io_vec = dc->sb_bio.bi_inline_vecs;
1156 dc->sb_bio.bi_io_vec[0].bv_page = sb_page;
1157 get_page(sb_page);
1158
1159 if (cached_dev_init(dc, sb->block_size << 9))
1160 goto err;
1161
1162 err = "error creating kobject";
1163 if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1164 "bcache"))
1165 goto err;
1166 if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1167 goto err;
1168
1169 pr_info("registered backing device %s", bdevname(bdev, name));
1170
1171 list_add(&dc->list, &uncached_devices);
1172 list_for_each_entry(c, &bch_cache_sets, list)
1173 bch_cached_dev_attach(dc, c);
1174
1175 if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1176 BDEV_STATE(&dc->sb) == BDEV_STATE_STALE)
1177 bch_cached_dev_run(dc);
1178
1179 return;
1180err:
1181 pr_notice("error opening %s: %s", bdevname(bdev, name), err);
1182 bcache_device_stop(&dc->disk);
1183}
1184
1185/* Flash only volumes */
1186
1187void bch_flash_dev_release(struct kobject *kobj)
1188{
1189 struct bcache_device *d = container_of(kobj, struct bcache_device,
1190 kobj);
1191 kfree(d);
1192}
1193
1194static void flash_dev_free(struct closure *cl)
1195{
1196 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1197 mutex_lock(&bch_register_lock);
1198 bcache_device_free(d);
1199 mutex_unlock(&bch_register_lock);
1200 kobject_put(&d->kobj);
1201}
1202
1203static void flash_dev_flush(struct closure *cl)
1204{
1205 struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1206
1207 mutex_lock(&bch_register_lock);
1208 bcache_device_unlink(d);
1209 mutex_unlock(&bch_register_lock);
1210 kobject_del(&d->kobj);
1211 continue_at(cl, flash_dev_free, system_wq);
1212}
1213
1214static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1215{
1216 struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1217 GFP_KERNEL);
1218 if (!d)
1219 return -ENOMEM;
1220
1221 closure_init(&d->cl, NULL);
1222 set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1223
1224 kobject_init(&d->kobj, &bch_flash_dev_ktype);
1225
1226 if (bcache_device_init(d, block_bytes(c), u->sectors))
1227 goto err;
1228
1229 bcache_device_attach(d, c, u - c->uuids);
1230 bch_flash_dev_request_init(d);
1231 add_disk(d->disk);
1232
1233 if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1234 goto err;
1235
1236 bcache_device_link(d, c, "volume");
1237
1238 return 0;
1239err:
1240 kobject_put(&d->kobj);
1241 return -ENOMEM;
1242}
1243
1244static int flash_devs_run(struct cache_set *c)
1245{
1246 int ret = 0;
1247 struct uuid_entry *u;
1248
1249 for (u = c->uuids;
1250 u < c->uuids + c->nr_uuids && !ret;
1251 u++)
1252 if (UUID_FLASH_ONLY(u))
1253 ret = flash_dev_run(c, u);
1254
1255 return ret;
1256}
1257
1258int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1259{
1260 struct uuid_entry *u;
1261
1262 if (test_bit(CACHE_SET_STOPPING, &c->flags))
1263 return -EINTR;
1264
1265 if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1266 return -EPERM;
1267
1268 u = uuid_find_empty(c);
1269 if (!u) {
1270 pr_err("Can't create volume, no room for UUID");
1271 return -EINVAL;
1272 }
1273
1274 get_random_bytes(u->uuid, 16);
1275 memset(u->label, 0, 32);
1276 u->first_reg = u->last_reg = cpu_to_le32(get_seconds());
1277
1278 SET_UUID_FLASH_ONLY(u, 1);
1279 u->sectors = size >> 9;
1280
1281 bch_uuid_write(c);
1282
1283 return flash_dev_run(c, u);
1284}
1285
1286/* Cache set */
1287
1288__printf(2, 3)
1289bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1290{
1291 va_list args;
1292
1293 if (c->on_error != ON_ERROR_PANIC &&
1294 test_bit(CACHE_SET_STOPPING, &c->flags))
1295 return false;
1296
1297 /* XXX: we can be called from atomic context
1298 acquire_console_sem();
1299 */
1300
1301 printk(KERN_ERR "bcache: error on %pU: ", c->sb.set_uuid);
1302
1303 va_start(args, fmt);
1304 vprintk(fmt, args);
1305 va_end(args);
1306
1307 printk(", disabling caching\n");
1308
1309 if (c->on_error == ON_ERROR_PANIC)
1310 panic("panic forced after error\n");
1311
1312 bch_cache_set_unregister(c);
1313 return true;
1314}
1315
1316void bch_cache_set_release(struct kobject *kobj)
1317{
1318 struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1319 kfree(c);
1320 module_put(THIS_MODULE);
1321}
1322
1323static void cache_set_free(struct closure *cl)
1324{
1325 struct cache_set *c = container_of(cl, struct cache_set, cl);
1326 struct cache *ca;
1327 unsigned i;
1328
1329 if (!IS_ERR_OR_NULL(c->debug))
1330 debugfs_remove(c->debug);
1331
1332 bch_open_buckets_free(c);
1333 bch_btree_cache_free(c);
1334 bch_journal_free(c);
1335
1336 for_each_cache(ca, c, i)
1337 if (ca) {
1338 ca->set = NULL;
1339 c->cache[ca->sb.nr_this_dev] = NULL;
1340 kobject_put(&ca->kobj);
1341 }
1342
1343 bch_bset_sort_state_free(&c->sort);
1344 free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1345
1346 if (c->moving_gc_wq)
1347 destroy_workqueue(c->moving_gc_wq);
1348 if (c->bio_split)
1349 bioset_free(c->bio_split);
1350 if (c->fill_iter)
1351 mempool_destroy(c->fill_iter);
1352 if (c->bio_meta)
1353 mempool_destroy(c->bio_meta);
1354 if (c->search)
1355 mempool_destroy(c->search);
1356 kfree(c->devices);
1357
1358 mutex_lock(&bch_register_lock);
1359 list_del(&c->list);
1360 mutex_unlock(&bch_register_lock);
1361
1362 pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1363 wake_up(&unregister_wait);
1364
1365 closure_debug_destroy(&c->cl);
1366 kobject_put(&c->kobj);
1367}
1368
1369static void cache_set_flush(struct closure *cl)
1370{
1371 struct cache_set *c = container_of(cl, struct cache_set, caching);
1372 struct cache *ca;
1373 struct btree *b;
1374 unsigned i;
1375
1376 if (!c)
1377 closure_return(cl);
1378
1379 bch_cache_accounting_destroy(&c->accounting);
1380
1381 kobject_put(&c->internal);
1382 kobject_del(&c->kobj);
1383
1384 if (c->gc_thread)
1385 kthread_stop(c->gc_thread);
1386
1387 if (!IS_ERR_OR_NULL(c->root))
1388 list_add(&c->root->list, &c->btree_cache);
1389
1390 /* Should skip this if we're unregistering because of an error */
1391 list_for_each_entry(b, &c->btree_cache, list) {
1392 mutex_lock(&b->write_lock);
1393 if (btree_node_dirty(b))
1394 __bch_btree_node_write(b, NULL);
1395 mutex_unlock(&b->write_lock);
1396 }
1397
1398 for_each_cache(ca, c, i)
1399 if (ca->alloc_thread)
1400 kthread_stop(ca->alloc_thread);
1401
1402 if (c->journal.cur) {
1403 cancel_delayed_work_sync(&c->journal.work);
1404 /* flush last journal entry if needed */
1405 c->journal.work.work.func(&c->journal.work.work);
1406 }
1407
1408 closure_return(cl);
1409}
1410
1411static void __cache_set_unregister(struct closure *cl)
1412{
1413 struct cache_set *c = container_of(cl, struct cache_set, caching);
1414 struct cached_dev *dc;
1415 size_t i;
1416
1417 mutex_lock(&bch_register_lock);
1418
1419 for (i = 0; i < c->nr_uuids; i++)
1420 if (c->devices[i]) {
1421 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1422 test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1423 dc = container_of(c->devices[i],
1424 struct cached_dev, disk);
1425 bch_cached_dev_detach(dc);
1426 } else {
1427 bcache_device_stop(c->devices[i]);
1428 }
1429 }
1430
1431 mutex_unlock(&bch_register_lock);
1432
1433 continue_at(cl, cache_set_flush, system_wq);
1434}
1435
1436void bch_cache_set_stop(struct cache_set *c)
1437{
1438 if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1439 closure_queue(&c->caching);
1440}
1441
1442void bch_cache_set_unregister(struct cache_set *c)
1443{
1444 set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1445 bch_cache_set_stop(c);
1446}
1447
1448#define alloc_bucket_pages(gfp, c) \
1449 ((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1450
1451struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1452{
1453 int iter_size;
1454 struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1455 if (!c)
1456 return NULL;
1457
1458 __module_get(THIS_MODULE);
1459 closure_init(&c->cl, NULL);
1460 set_closure_fn(&c->cl, cache_set_free, system_wq);
1461
1462 closure_init(&c->caching, &c->cl);
1463 set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1464
1465 /* Maybe create continue_at_noreturn() and use it here? */
1466 closure_set_stopped(&c->cl);
1467 closure_put(&c->cl);
1468
1469 kobject_init(&c->kobj, &bch_cache_set_ktype);
1470 kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1471
1472 bch_cache_accounting_init(&c->accounting, &c->cl);
1473
1474 memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1475 c->sb.block_size = sb->block_size;
1476 c->sb.bucket_size = sb->bucket_size;
1477 c->sb.nr_in_set = sb->nr_in_set;
1478 c->sb.last_mount = sb->last_mount;
1479 c->bucket_bits = ilog2(sb->bucket_size);
1480 c->block_bits = ilog2(sb->block_size);
1481 c->nr_uuids = bucket_bytes(c) / sizeof(struct uuid_entry);
1482
1483 c->btree_pages = bucket_pages(c);
1484 if (c->btree_pages > BTREE_MAX_PAGES)
1485 c->btree_pages = max_t(int, c->btree_pages / 4,
1486 BTREE_MAX_PAGES);
1487
1488 sema_init(&c->sb_write_mutex, 1);
1489 mutex_init(&c->bucket_lock);
1490 init_waitqueue_head(&c->btree_cache_wait);
1491 init_waitqueue_head(&c->bucket_wait);
1492 sema_init(&c->uuid_write_mutex, 1);
1493
1494 spin_lock_init(&c->btree_gc_time.lock);
1495 spin_lock_init(&c->btree_split_time.lock);
1496 spin_lock_init(&c->btree_read_time.lock);
1497
1498 bch_moving_init_cache_set(c);
1499
1500 INIT_LIST_HEAD(&c->list);
1501 INIT_LIST_HEAD(&c->cached_devs);
1502 INIT_LIST_HEAD(&c->btree_cache);
1503 INIT_LIST_HEAD(&c->btree_cache_freeable);
1504 INIT_LIST_HEAD(&c->btree_cache_freed);
1505 INIT_LIST_HEAD(&c->data_buckets);
1506
1507 c->search = mempool_create_slab_pool(32, bch_search_cache);
1508 if (!c->search)
1509 goto err;
1510
1511 iter_size = (sb->bucket_size / sb->block_size + 1) *
1512 sizeof(struct btree_iter_set);
1513
1514 if (!(c->devices = kzalloc(c->nr_uuids * sizeof(void *), GFP_KERNEL)) ||
1515 !(c->bio_meta = mempool_create_kmalloc_pool(2,
1516 sizeof(struct bbio) + sizeof(struct bio_vec) *
1517 bucket_pages(c))) ||
1518 !(c->fill_iter = mempool_create_kmalloc_pool(1, iter_size)) ||
1519 !(c->bio_split = bioset_create(4, offsetof(struct bbio, bio))) ||
1520 !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1521 !(c->moving_gc_wq = create_workqueue("bcache_gc")) ||
1522 bch_journal_alloc(c) ||
1523 bch_btree_cache_alloc(c) ||
1524 bch_open_buckets_alloc(c) ||
1525 bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1526 goto err;
1527
1528 c->congested_read_threshold_us = 2000;
1529 c->congested_write_threshold_us = 20000;
1530 c->error_limit = 8 << IO_ERROR_SHIFT;
1531
1532 return c;
1533err:
1534 bch_cache_set_unregister(c);
1535 return NULL;
1536}
1537
1538static void run_cache_set(struct cache_set *c)
1539{
1540 const char *err = "cannot allocate memory";
1541 struct cached_dev *dc, *t;
1542 struct cache *ca;
1543 struct closure cl;
1544 unsigned i;
1545
1546 closure_init_stack(&cl);
1547
1548 for_each_cache(ca, c, i)
1549 c->nbuckets += ca->sb.nbuckets;
1550
1551 if (CACHE_SYNC(&c->sb)) {
1552 LIST_HEAD(journal);
1553 struct bkey *k;
1554 struct jset *j;
1555
1556 err = "cannot allocate memory for journal";
1557 if (bch_journal_read(c, &journal))
1558 goto err;
1559
1560 pr_debug("btree_journal_read() done");
1561
1562 err = "no journal entries found";
1563 if (list_empty(&journal))
1564 goto err;
1565
1566 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1567
1568 err = "IO error reading priorities";
1569 for_each_cache(ca, c, i)
1570 prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1571
1572 /*
1573 * If prio_read() fails it'll call cache_set_error and we'll
1574 * tear everything down right away, but if we perhaps checked
1575 * sooner we could avoid journal replay.
1576 */
1577
1578 k = &j->btree_root;
1579
1580 err = "bad btree root";
1581 if (__bch_btree_ptr_invalid(c, k))
1582 goto err;
1583
1584 err = "error reading btree root";
1585 c->root = bch_btree_node_get(c, NULL, k, j->btree_level, true, NULL);
1586 if (IS_ERR_OR_NULL(c->root))
1587 goto err;
1588
1589 list_del_init(&c->root->list);
1590 rw_unlock(true, c->root);
1591
1592 err = uuid_read(c, j, &cl);
1593 if (err)
1594 goto err;
1595
1596 err = "error in recovery";
1597 if (bch_btree_check(c))
1598 goto err;
1599
1600 bch_journal_mark(c, &journal);
1601 bch_initial_gc_finish(c);
1602 pr_debug("btree_check() done");
1603
1604 /*
1605 * bcache_journal_next() can't happen sooner, or
1606 * btree_gc_finish() will give spurious errors about last_gc >
1607 * gc_gen - this is a hack but oh well.
1608 */
1609 bch_journal_next(&c->journal);
1610
1611 err = "error starting allocator thread";
1612 for_each_cache(ca, c, i)
1613 if (bch_cache_allocator_start(ca))
1614 goto err;
1615
1616 /*
1617 * First place it's safe to allocate: btree_check() and
1618 * btree_gc_finish() have to run before we have buckets to
1619 * allocate, and bch_bucket_alloc_set() might cause a journal
1620 * entry to be written so bcache_journal_next() has to be called
1621 * first.
1622 *
1623 * If the uuids were in the old format we have to rewrite them
1624 * before the next journal entry is written:
1625 */
1626 if (j->version < BCACHE_JSET_VERSION_UUID)
1627 __uuid_write(c);
1628
1629 bch_journal_replay(c, &journal);
1630 } else {
1631 pr_notice("invalidating existing data");
1632
1633 for_each_cache(ca, c, i) {
1634 unsigned j;
1635
1636 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1637 2, SB_JOURNAL_BUCKETS);
1638
1639 for (j = 0; j < ca->sb.keys; j++)
1640 ca->sb.d[j] = ca->sb.first_bucket + j;
1641 }
1642
1643 bch_initial_gc_finish(c);
1644
1645 err = "error starting allocator thread";
1646 for_each_cache(ca, c, i)
1647 if (bch_cache_allocator_start(ca))
1648 goto err;
1649
1650 mutex_lock(&c->bucket_lock);
1651 for_each_cache(ca, c, i)
1652 bch_prio_write(ca);
1653 mutex_unlock(&c->bucket_lock);
1654
1655 err = "cannot allocate new UUID bucket";
1656 if (__uuid_write(c))
1657 goto err;
1658
1659 err = "cannot allocate new btree root";
1660 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1661 if (IS_ERR_OR_NULL(c->root))
1662 goto err;
1663
1664 mutex_lock(&c->root->write_lock);
1665 bkey_copy_key(&c->root->key, &MAX_KEY);
1666 bch_btree_node_write(c->root, &cl);
1667 mutex_unlock(&c->root->write_lock);
1668
1669 bch_btree_set_root(c->root);
1670 rw_unlock(true, c->root);
1671
1672 /*
1673 * We don't want to write the first journal entry until
1674 * everything is set up - fortunately journal entries won't be
1675 * written until the SET_CACHE_SYNC() here:
1676 */
1677 SET_CACHE_SYNC(&c->sb, true);
1678
1679 bch_journal_next(&c->journal);
1680 bch_journal_meta(c, &cl);
1681 }
1682
1683 err = "error starting gc thread";
1684 if (bch_gc_thread_start(c))
1685 goto err;
1686
1687 closure_sync(&cl);
1688 c->sb.last_mount = get_seconds();
1689 bcache_write_super(c);
1690
1691 list_for_each_entry_safe(dc, t, &uncached_devices, list)
1692 bch_cached_dev_attach(dc, c);
1693
1694 flash_devs_run(c);
1695
1696 set_bit(CACHE_SET_RUNNING, &c->flags);
1697 return;
1698err:
1699 closure_sync(&cl);
1700 /* XXX: test this, it's broken */
1701 bch_cache_set_error(c, "%s", err);
1702}
1703
1704static bool can_attach_cache(struct cache *ca, struct cache_set *c)
1705{
1706 return ca->sb.block_size == c->sb.block_size &&
1707 ca->sb.bucket_size == c->sb.bucket_size &&
1708 ca->sb.nr_in_set == c->sb.nr_in_set;
1709}
1710
1711static const char *register_cache_set(struct cache *ca)
1712{
1713 char buf[12];
1714 const char *err = "cannot allocate memory";
1715 struct cache_set *c;
1716
1717 list_for_each_entry(c, &bch_cache_sets, list)
1718 if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
1719 if (c->cache[ca->sb.nr_this_dev])
1720 return "duplicate cache set member";
1721
1722 if (!can_attach_cache(ca, c))
1723 return "cache sb does not match set";
1724
1725 if (!CACHE_SYNC(&ca->sb))
1726 SET_CACHE_SYNC(&c->sb, false);
1727
1728 goto found;
1729 }
1730
1731 c = bch_cache_set_alloc(&ca->sb);
1732 if (!c)
1733 return err;
1734
1735 err = "error creating kobject";
1736 if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
1737 kobject_add(&c->internal, &c->kobj, "internal"))
1738 goto err;
1739
1740 if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
1741 goto err;
1742
1743 bch_debug_init_cache_set(c);
1744
1745 list_add(&c->list, &bch_cache_sets);
1746found:
1747 sprintf(buf, "cache%i", ca->sb.nr_this_dev);
1748 if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
1749 sysfs_create_link(&c->kobj, &ca->kobj, buf))
1750 goto err;
1751
1752 if (ca->sb.seq > c->sb.seq) {
1753 c->sb.version = ca->sb.version;
1754 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
1755 c->sb.flags = ca->sb.flags;
1756 c->sb.seq = ca->sb.seq;
1757 pr_debug("set version = %llu", c->sb.version);
1758 }
1759
1760 kobject_get(&ca->kobj);
1761 ca->set = c;
1762 ca->set->cache[ca->sb.nr_this_dev] = ca;
1763 c->cache_by_alloc[c->caches_loaded++] = ca;
1764
1765 if (c->caches_loaded == c->sb.nr_in_set)
1766 run_cache_set(c);
1767
1768 return NULL;
1769err:
1770 bch_cache_set_unregister(c);
1771 return err;
1772}
1773
1774/* Cache device */
1775
1776void bch_cache_release(struct kobject *kobj)
1777{
1778 struct cache *ca = container_of(kobj, struct cache, kobj);
1779 unsigned i;
1780
1781 if (ca->set) {
1782 BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
1783 ca->set->cache[ca->sb.nr_this_dev] = NULL;
1784 }
1785
1786 free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
1787 kfree(ca->prio_buckets);
1788 vfree(ca->buckets);
1789
1790 free_heap(&ca->heap);
1791 free_fifo(&ca->free_inc);
1792
1793 for (i = 0; i < RESERVE_NR; i++)
1794 free_fifo(&ca->free[i]);
1795
1796 if (ca->sb_bio.bi_inline_vecs[0].bv_page)
1797 put_page(ca->sb_bio.bi_io_vec[0].bv_page);
1798
1799 if (!IS_ERR_OR_NULL(ca->bdev))
1800 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1801
1802 kfree(ca);
1803 module_put(THIS_MODULE);
1804}
1805
1806static int cache_alloc(struct cache_sb *sb, struct cache *ca)
1807{
1808 size_t free;
1809 struct bucket *b;
1810
1811 __module_get(THIS_MODULE);
1812 kobject_init(&ca->kobj, &bch_cache_ktype);
1813
1814 bio_init(&ca->journal.bio);
1815 ca->journal.bio.bi_max_vecs = 8;
1816 ca->journal.bio.bi_io_vec = ca->journal.bio.bi_inline_vecs;
1817
1818 free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
1819
1820 if (!init_fifo(&ca->free[RESERVE_BTREE], 8, GFP_KERNEL) ||
1821 !init_fifo(&ca->free[RESERVE_PRIO], prio_buckets(ca), GFP_KERNEL) ||
1822 !init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL) ||
1823 !init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL) ||
1824 !init_fifo(&ca->free_inc, free << 2, GFP_KERNEL) ||
1825 !init_heap(&ca->heap, free << 3, GFP_KERNEL) ||
1826 !(ca->buckets = vzalloc(sizeof(struct bucket) *
1827 ca->sb.nbuckets)) ||
1828 !(ca->prio_buckets = kzalloc(sizeof(uint64_t) * prio_buckets(ca) *
1829 2, GFP_KERNEL)) ||
1830 !(ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca)))
1831 return -ENOMEM;
1832
1833 ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
1834
1835 for_each_bucket(b, ca)
1836 atomic_set(&b->pin, 0);
1837
1838 return 0;
1839}
1840
1841static int register_cache(struct cache_sb *sb, struct page *sb_page,
1842 struct block_device *bdev, struct cache *ca)
1843{
1844 char name[BDEVNAME_SIZE];
1845 const char *err = NULL;
1846 int ret = 0;
1847
1848 memcpy(&ca->sb, sb, sizeof(struct cache_sb));
1849 ca->bdev = bdev;
1850 ca->bdev->bd_holder = ca;
1851
1852 bio_init(&ca->sb_bio);
1853 ca->sb_bio.bi_max_vecs = 1;
1854 ca->sb_bio.bi_io_vec = ca->sb_bio.bi_inline_vecs;
1855 ca->sb_bio.bi_io_vec[0].bv_page = sb_page;
1856 get_page(sb_page);
1857
1858 if (blk_queue_discard(bdev_get_queue(ca->bdev)))
1859 ca->discard = CACHE_DISCARD(&ca->sb);
1860
1861 ret = cache_alloc(sb, ca);
1862 if (ret != 0)
1863 goto err;
1864
1865 if (kobject_add(&ca->kobj, &part_to_dev(bdev->bd_part)->kobj, "bcache")) {
1866 err = "error calling kobject_add";
1867 ret = -ENOMEM;
1868 goto out;
1869 }
1870
1871 mutex_lock(&bch_register_lock);
1872 err = register_cache_set(ca);
1873 mutex_unlock(&bch_register_lock);
1874
1875 if (err) {
1876 ret = -ENODEV;
1877 goto out;
1878 }
1879
1880 pr_info("registered cache device %s", bdevname(bdev, name));
1881
1882out:
1883 kobject_put(&ca->kobj);
1884
1885err:
1886 if (err)
1887 pr_notice("error opening %s: %s", bdevname(bdev, name), err);
1888
1889 return ret;
1890}
1891
1892/* Global interfaces/init */
1893
1894static ssize_t register_bcache(struct kobject *, struct kobj_attribute *,
1895 const char *, size_t);
1896
1897kobj_attribute_write(register, register_bcache);
1898kobj_attribute_write(register_quiet, register_bcache);
1899
1900static bool bch_is_open_backing(struct block_device *bdev) {
1901 struct cache_set *c, *tc;
1902 struct cached_dev *dc, *t;
1903
1904 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
1905 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
1906 if (dc->bdev == bdev)
1907 return true;
1908 list_for_each_entry_safe(dc, t, &uncached_devices, list)
1909 if (dc->bdev == bdev)
1910 return true;
1911 return false;
1912}
1913
1914static bool bch_is_open_cache(struct block_device *bdev) {
1915 struct cache_set *c, *tc;
1916 struct cache *ca;
1917 unsigned i;
1918
1919 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
1920 for_each_cache(ca, c, i)
1921 if (ca->bdev == bdev)
1922 return true;
1923 return false;
1924}
1925
1926static bool bch_is_open(struct block_device *bdev) {
1927 return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
1928}
1929
1930static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
1931 const char *buffer, size_t size)
1932{
1933 ssize_t ret = size;
1934 const char *err = "cannot allocate memory";
1935 char *path = NULL;
1936 struct cache_sb *sb = NULL;
1937 struct block_device *bdev = NULL;
1938 struct page *sb_page = NULL;
1939
1940 if (!try_module_get(THIS_MODULE))
1941 return -EBUSY;
1942
1943 if (!(path = kstrndup(buffer, size, GFP_KERNEL)) ||
1944 !(sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL)))
1945 goto err;
1946
1947 err = "failed to open device";
1948 bdev = blkdev_get_by_path(strim(path),
1949 FMODE_READ|FMODE_WRITE|FMODE_EXCL,
1950 sb);
1951 if (IS_ERR(bdev)) {
1952 if (bdev == ERR_PTR(-EBUSY)) {
1953 bdev = lookup_bdev(strim(path));
1954 mutex_lock(&bch_register_lock);
1955 if (!IS_ERR(bdev) && bch_is_open(bdev))
1956 err = "device already registered";
1957 else
1958 err = "device busy";
1959 mutex_unlock(&bch_register_lock);
1960 if (attr == &ksysfs_register_quiet)
1961 goto out;
1962 }
1963 goto err;
1964 }
1965
1966 err = "failed to set blocksize";
1967 if (set_blocksize(bdev, 4096))
1968 goto err_close;
1969
1970 err = read_super(sb, bdev, &sb_page);
1971 if (err)
1972 goto err_close;
1973
1974 if (SB_IS_BDEV(sb)) {
1975 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
1976 if (!dc)
1977 goto err_close;
1978
1979 mutex_lock(&bch_register_lock);
1980 register_bdev(sb, sb_page, bdev, dc);
1981 mutex_unlock(&bch_register_lock);
1982 } else {
1983 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
1984 if (!ca)
1985 goto err_close;
1986
1987 if (register_cache(sb, sb_page, bdev, ca) != 0)
1988 goto err_close;
1989 }
1990out:
1991 if (sb_page)
1992 put_page(sb_page);
1993 kfree(sb);
1994 kfree(path);
1995 module_put(THIS_MODULE);
1996 return ret;
1997
1998err_close:
1999 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2000err:
2001 pr_info("error opening %s: %s", path, err);
2002 ret = -EINVAL;
2003 goto out;
2004}
2005
2006static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2007{
2008 if (code == SYS_DOWN ||
2009 code == SYS_HALT ||
2010 code == SYS_POWER_OFF) {
2011 DEFINE_WAIT(wait);
2012 unsigned long start = jiffies;
2013 bool stopped = false;
2014
2015 struct cache_set *c, *tc;
2016 struct cached_dev *dc, *tdc;
2017
2018 mutex_lock(&bch_register_lock);
2019
2020 if (list_empty(&bch_cache_sets) &&
2021 list_empty(&uncached_devices))
2022 goto out;
2023
2024 pr_info("Stopping all devices:");
2025
2026 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2027 bch_cache_set_stop(c);
2028
2029 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2030 bcache_device_stop(&dc->disk);
2031
2032 /* What's a condition variable? */
2033 while (1) {
2034 long timeout = start + 2 * HZ - jiffies;
2035
2036 stopped = list_empty(&bch_cache_sets) &&
2037 list_empty(&uncached_devices);
2038
2039 if (timeout < 0 || stopped)
2040 break;
2041
2042 prepare_to_wait(&unregister_wait, &wait,
2043 TASK_UNINTERRUPTIBLE);
2044
2045 mutex_unlock(&bch_register_lock);
2046 schedule_timeout(timeout);
2047 mutex_lock(&bch_register_lock);
2048 }
2049
2050 finish_wait(&unregister_wait, &wait);
2051
2052 if (stopped)
2053 pr_info("All devices stopped");
2054 else
2055 pr_notice("Timeout waiting for devices to be closed");
2056out:
2057 mutex_unlock(&bch_register_lock);
2058 }
2059
2060 return NOTIFY_DONE;
2061}
2062
2063static struct notifier_block reboot = {
2064 .notifier_call = bcache_reboot,
2065 .priority = INT_MAX, /* before any real devices */
2066};
2067
2068static void bcache_exit(void)
2069{
2070 bch_debug_exit();
2071 bch_request_exit();
2072 if (bcache_kobj)
2073 kobject_put(bcache_kobj);
2074 if (bcache_wq)
2075 destroy_workqueue(bcache_wq);
2076 if (bcache_major)
2077 unregister_blkdev(bcache_major, "bcache");
2078 unregister_reboot_notifier(&reboot);
2079}
2080
2081static int __init bcache_init(void)
2082{
2083 static const struct attribute *files[] = {
2084 &ksysfs_register.attr,
2085 &ksysfs_register_quiet.attr,
2086 NULL
2087 };
2088
2089 mutex_init(&bch_register_lock);
2090 init_waitqueue_head(&unregister_wait);
2091 register_reboot_notifier(&reboot);
2092 closure_debug_init();
2093
2094 bcache_major = register_blkdev(0, "bcache");
2095 if (bcache_major < 0) {
2096 unregister_reboot_notifier(&reboot);
2097 return bcache_major;
2098 }
2099
2100 if (!(bcache_wq = create_workqueue("bcache")) ||
2101 !(bcache_kobj = kobject_create_and_add("bcache", fs_kobj)) ||
2102 sysfs_create_files(bcache_kobj, files) ||
2103 bch_request_init() ||
2104 bch_debug_init(bcache_kobj))
2105 goto err;
2106
2107 return 0;
2108err:
2109 bcache_exit();
2110 return -ENOMEM;
2111}
2112
2113module_exit(bcache_exit);
2114module_init(bcache_init);