interconnect: qcom: icc-rpm: Fix peak rate calculation
[linux-2.6-microblaze.git] / drivers / md / bcache / super.c
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
29 unsigned int bch_cutoff_writeback;
30 unsigned int bch_cutoff_writeback_sync;
31
32 static const char bcache_magic[] = {
33         0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
34         0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
35 };
36
37 static const char invalid_uuid[] = {
38         0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
39         0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
40 };
41
42 static struct kobject *bcache_kobj;
43 struct mutex bch_register_lock;
44 bool bcache_is_reboot;
45 LIST_HEAD(bch_cache_sets);
46 static LIST_HEAD(uncached_devices);
47
48 static int bcache_major;
49 static DEFINE_IDA(bcache_device_idx);
50 static wait_queue_head_t unregister_wait;
51 struct workqueue_struct *bcache_wq;
52 struct workqueue_struct *bch_flush_wq;
53 struct workqueue_struct *bch_journal_wq;
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
64 static 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
91 static 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;
161 err:
162         return err;
163 }
164
165
166 static 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;
274 err:
275         put_page(page);
276         return err;
277 }
278
279 static 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
289 static 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
330 static void bch_write_bdev_super_unlock(struct closure *cl)
331 {
332         struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
333
334         up(&dc->sb_write_mutex);
335 }
336
337 void 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
356 static 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
366 static void bcache_write_super_unlock(struct closure *cl)
367 {
368         struct cache_set *c = container_of(cl, struct cache_set, sb_write);
369
370         up(&c->sb_write_mutex);
371 }
372
373 void 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
400 static 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
410 static void uuid_io_unlock(struct closure *cl)
411 {
412         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
413
414         up(&c->uuid_write_mutex);
415 }
416
417 static 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
458 static 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
499 static 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
525 int 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
535 static 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
547 static 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
581 static 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
590 static 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
610 int 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
691 static 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;
729 out:
730         return ret;
731 }
732
733 /* Bcache device */
734
735 static 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
746 static void release_dev(struct gendisk *b)
747 {
748         struct bcache_device *d = b->private_data;
749
750         closure_put(&d->cl);
751 }
752
753 static 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
761 static 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
769 static 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
777 void 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
788 static 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
802 static 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
824 static 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
846 static 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
859 static inline int first_minor_to_idx(int first_minor)
860 {
861         return (first_minor/BCACHE_MINORS);
862 }
863
864 static inline int idx_to_first_minor(int idx)
865 {
866         return (idx * BCACHE_MINORS);
867 }
868
869 static 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
896 static 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
909         n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
910         if (!n || n > max_stripes) {
911                 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
912                         n);
913                 return -ENOMEM;
914         }
915         d->nr_stripes = n;
916
917         n = d->nr_stripes * sizeof(atomic_t);
918         d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
919         if (!d->stripe_sectors_dirty)
920                 return -ENOMEM;
921
922         n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
923         d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
924         if (!d->full_dirty_stripes)
925                 goto out_free_stripe_sectors_dirty;
926
927         idx = ida_simple_get(&bcache_device_idx, 0,
928                                 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
929         if (idx < 0)
930                 goto out_free_full_dirty_stripes;
931
932         if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
933                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
934                 goto out_ida_remove;
935
936         d->disk = blk_alloc_disk(NUMA_NO_NODE);
937         if (!d->disk)
938                 goto out_bioset_exit;
939
940         set_capacity(d->disk, sectors);
941         snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
942
943         d->disk->major          = bcache_major;
944         d->disk->first_minor    = idx_to_first_minor(idx);
945         d->disk->minors         = BCACHE_MINORS;
946         d->disk->fops           = ops;
947         d->disk->private_data   = d;
948
949         q = d->disk->queue;
950         q->limits.max_hw_sectors        = UINT_MAX;
951         q->limits.max_sectors           = UINT_MAX;
952         q->limits.max_segment_size      = UINT_MAX;
953         q->limits.max_segments          = BIO_MAX_VECS;
954         blk_queue_max_discard_sectors(q, UINT_MAX);
955         q->limits.discard_granularity   = 512;
956         q->limits.io_min                = block_size;
957         q->limits.logical_block_size    = block_size;
958         q->limits.physical_block_size   = block_size;
959
960         if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
961                 /*
962                  * This should only happen with BCACHE_SB_VERSION_BDEV.
963                  * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
964                  */
965                 pr_info("%s: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
966                         d->disk->disk_name, q->limits.logical_block_size,
967                         PAGE_SIZE, bdev_logical_block_size(cached_bdev));
968
969                 /* This also adjusts physical block size/min io size if needed */
970                 blk_queue_logical_block_size(q, bdev_logical_block_size(cached_bdev));
971         }
972
973         blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
974
975         blk_queue_write_cache(q, true, true);
976
977         return 0;
978
979 out_bioset_exit:
980         bioset_exit(&d->bio_split);
981 out_ida_remove:
982         ida_simple_remove(&bcache_device_idx, idx);
983 out_free_full_dirty_stripes:
984         kvfree(d->full_dirty_stripes);
985 out_free_stripe_sectors_dirty:
986         kvfree(d->stripe_sectors_dirty);
987         return -ENOMEM;
988
989 }
990
991 /* Cached device */
992
993 static void calc_cached_dev_sectors(struct cache_set *c)
994 {
995         uint64_t sectors = 0;
996         struct cached_dev *dc;
997
998         list_for_each_entry(dc, &c->cached_devs, list)
999                 sectors += bdev_nr_sectors(dc->bdev);
1000
1001         c->cached_dev_sectors = sectors;
1002 }
1003
1004 #define BACKING_DEV_OFFLINE_TIMEOUT 5
1005 static int cached_dev_status_update(void *arg)
1006 {
1007         struct cached_dev *dc = arg;
1008         struct request_queue *q;
1009
1010         /*
1011          * If this delayed worker is stopping outside, directly quit here.
1012          * dc->io_disable might be set via sysfs interface, so check it
1013          * here too.
1014          */
1015         while (!kthread_should_stop() && !dc->io_disable) {
1016                 q = bdev_get_queue(dc->bdev);
1017                 if (blk_queue_dying(q))
1018                         dc->offline_seconds++;
1019                 else
1020                         dc->offline_seconds = 0;
1021
1022                 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1023                         pr_err("%pg: device offline for %d seconds\n",
1024                                dc->bdev,
1025                                BACKING_DEV_OFFLINE_TIMEOUT);
1026                         pr_err("%s: disable I/O request due to backing device offline\n",
1027                                dc->disk.name);
1028                         dc->io_disable = true;
1029                         /* let others know earlier that io_disable is true */
1030                         smp_mb();
1031                         bcache_device_stop(&dc->disk);
1032                         break;
1033                 }
1034                 schedule_timeout_interruptible(HZ);
1035         }
1036
1037         wait_for_kthread_stop();
1038         return 0;
1039 }
1040
1041
1042 int bch_cached_dev_run(struct cached_dev *dc)
1043 {
1044         int ret = 0;
1045         struct bcache_device *d = &dc->disk;
1046         char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1047         char *env[] = {
1048                 "DRIVER=bcache",
1049                 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1050                 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1051                 NULL,
1052         };
1053
1054         if (dc->io_disable) {
1055                 pr_err("I/O disabled on cached dev %pg\n", dc->bdev);
1056                 ret = -EIO;
1057                 goto out;
1058         }
1059
1060         if (atomic_xchg(&dc->running, 1)) {
1061                 pr_info("cached dev %pg is running already\n", dc->bdev);
1062                 ret = -EBUSY;
1063                 goto out;
1064         }
1065
1066         if (!d->c &&
1067             BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1068                 struct closure cl;
1069
1070                 closure_init_stack(&cl);
1071
1072                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1073                 bch_write_bdev_super(dc, &cl);
1074                 closure_sync(&cl);
1075         }
1076
1077         ret = add_disk(d->disk);
1078         if (ret)
1079                 goto out;
1080         bd_link_disk_holder(dc->bdev, dc->disk.disk);
1081         /*
1082          * won't show up in the uevent file, use udevadm monitor -e instead
1083          * only class / kset properties are persistent
1084          */
1085         kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1086
1087         if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1088             sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1089                               &d->kobj, "bcache")) {
1090                 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1091                 ret = -ENOMEM;
1092                 goto out;
1093         }
1094
1095         dc->status_update_thread = kthread_run(cached_dev_status_update,
1096                                                dc, "bcache_status_update");
1097         if (IS_ERR(dc->status_update_thread)) {
1098                 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1099         }
1100
1101 out:
1102         kfree(env[1]);
1103         kfree(env[2]);
1104         kfree(buf);
1105         return ret;
1106 }
1107
1108 /*
1109  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1110  * work dc->writeback_rate_update is running. Wait until the routine
1111  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1112  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1113  * seconds, give up waiting here and continue to cancel it too.
1114  */
1115 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1116 {
1117         int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1118
1119         do {
1120                 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1121                               &dc->disk.flags))
1122                         break;
1123                 time_out--;
1124                 schedule_timeout_interruptible(1);
1125         } while (time_out > 0);
1126
1127         if (time_out == 0)
1128                 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1129
1130         cancel_delayed_work_sync(&dc->writeback_rate_update);
1131 }
1132
1133 static void cached_dev_detach_finish(struct work_struct *w)
1134 {
1135         struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1136         struct cache_set *c = dc->disk.c;
1137
1138         BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1139         BUG_ON(refcount_read(&dc->count));
1140
1141
1142         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1143                 cancel_writeback_rate_update_dwork(dc);
1144
1145         if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1146                 kthread_stop(dc->writeback_thread);
1147                 dc->writeback_thread = NULL;
1148         }
1149
1150         mutex_lock(&bch_register_lock);
1151
1152         bcache_device_detach(&dc->disk);
1153         list_move(&dc->list, &uncached_devices);
1154         calc_cached_dev_sectors(c);
1155
1156         clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1157         clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1158
1159         mutex_unlock(&bch_register_lock);
1160
1161         pr_info("Caching disabled for %pg\n", dc->bdev);
1162
1163         /* Drop ref we took in cached_dev_detach() */
1164         closure_put(&dc->disk.cl);
1165 }
1166
1167 void bch_cached_dev_detach(struct cached_dev *dc)
1168 {
1169         lockdep_assert_held(&bch_register_lock);
1170
1171         if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1172                 return;
1173
1174         if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1175                 return;
1176
1177         /*
1178          * Block the device from being closed and freed until we're finished
1179          * detaching
1180          */
1181         closure_get(&dc->disk.cl);
1182
1183         bch_writeback_queue(dc);
1184
1185         cached_dev_put(dc);
1186 }
1187
1188 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1189                           uint8_t *set_uuid)
1190 {
1191         uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1192         struct uuid_entry *u;
1193         struct cached_dev *exist_dc, *t;
1194         int ret = 0;
1195
1196         if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1197             (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1198                 return -ENOENT;
1199
1200         if (dc->disk.c) {
1201                 pr_err("Can't attach %pg: already attached\n", dc->bdev);
1202                 return -EINVAL;
1203         }
1204
1205         if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1206                 pr_err("Can't attach %pg: shutting down\n", dc->bdev);
1207                 return -EINVAL;
1208         }
1209
1210         if (dc->sb.block_size < c->cache->sb.block_size) {
1211                 /* Will die */
1212                 pr_err("Couldn't attach %pg: block size less than set's block size\n",
1213                        dc->bdev);
1214                 return -EINVAL;
1215         }
1216
1217         /* Check whether already attached */
1218         list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1219                 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1220                         pr_err("Tried to attach %pg but duplicate UUID already attached\n",
1221                                 dc->bdev);
1222
1223                         return -EINVAL;
1224                 }
1225         }
1226
1227         u = uuid_find(c, dc->sb.uuid);
1228
1229         if (u &&
1230             (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1231              BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1232                 memcpy(u->uuid, invalid_uuid, 16);
1233                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1234                 u = NULL;
1235         }
1236
1237         if (!u) {
1238                 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1239                         pr_err("Couldn't find uuid for %pg in set\n", dc->bdev);
1240                         return -ENOENT;
1241                 }
1242
1243                 u = uuid_find_empty(c);
1244                 if (!u) {
1245                         pr_err("Not caching %pg, no room for UUID\n", dc->bdev);
1246                         return -EINVAL;
1247                 }
1248         }
1249
1250         /*
1251          * Deadlocks since we're called via sysfs...
1252          * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1253          */
1254
1255         if (bch_is_zero(u->uuid, 16)) {
1256                 struct closure cl;
1257
1258                 closure_init_stack(&cl);
1259
1260                 memcpy(u->uuid, dc->sb.uuid, 16);
1261                 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1262                 u->first_reg = u->last_reg = rtime;
1263                 bch_uuid_write(c);
1264
1265                 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1266                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1267
1268                 bch_write_bdev_super(dc, &cl);
1269                 closure_sync(&cl);
1270         } else {
1271                 u->last_reg = rtime;
1272                 bch_uuid_write(c);
1273         }
1274
1275         bcache_device_attach(&dc->disk, c, u - c->uuids);
1276         list_move(&dc->list, &c->cached_devs);
1277         calc_cached_dev_sectors(c);
1278
1279         /*
1280          * dc->c must be set before dc->count != 0 - paired with the mb in
1281          * cached_dev_get()
1282          */
1283         smp_wmb();
1284         refcount_set(&dc->count, 1);
1285
1286         /* Block writeback thread, but spawn it */
1287         down_write(&dc->writeback_lock);
1288         if (bch_cached_dev_writeback_start(dc)) {
1289                 up_write(&dc->writeback_lock);
1290                 pr_err("Couldn't start writeback facilities for %s\n",
1291                        dc->disk.disk->disk_name);
1292                 return -ENOMEM;
1293         }
1294
1295         if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1296                 atomic_set(&dc->has_dirty, 1);
1297                 bch_writeback_queue(dc);
1298         }
1299
1300         bch_sectors_dirty_init(&dc->disk);
1301
1302         ret = bch_cached_dev_run(dc);
1303         if (ret && (ret != -EBUSY)) {
1304                 up_write(&dc->writeback_lock);
1305                 /*
1306                  * bch_register_lock is held, bcache_device_stop() is not
1307                  * able to be directly called. The kthread and kworker
1308                  * created previously in bch_cached_dev_writeback_start()
1309                  * have to be stopped manually here.
1310                  */
1311                 kthread_stop(dc->writeback_thread);
1312                 cancel_writeback_rate_update_dwork(dc);
1313                 pr_err("Couldn't run cached device %pg\n", dc->bdev);
1314                 return ret;
1315         }
1316
1317         bcache_device_link(&dc->disk, c, "bdev");
1318         atomic_inc(&c->attached_dev_nr);
1319
1320         if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1321                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1322                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1323                 set_disk_ro(dc->disk.disk, 1);
1324         }
1325
1326         /* Allow the writeback thread to proceed */
1327         up_write(&dc->writeback_lock);
1328
1329         pr_info("Caching %pg as %s on set %pU\n",
1330                 dc->bdev,
1331                 dc->disk.disk->disk_name,
1332                 dc->disk.c->set_uuid);
1333         return 0;
1334 }
1335
1336 /* when dc->disk.kobj released */
1337 void bch_cached_dev_release(struct kobject *kobj)
1338 {
1339         struct cached_dev *dc = container_of(kobj, struct cached_dev,
1340                                              disk.kobj);
1341         kfree(dc);
1342         module_put(THIS_MODULE);
1343 }
1344
1345 static void cached_dev_free(struct closure *cl)
1346 {
1347         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1348
1349         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1350                 cancel_writeback_rate_update_dwork(dc);
1351
1352         if (!IS_ERR_OR_NULL(dc->writeback_thread))
1353                 kthread_stop(dc->writeback_thread);
1354         if (!IS_ERR_OR_NULL(dc->status_update_thread))
1355                 kthread_stop(dc->status_update_thread);
1356
1357         mutex_lock(&bch_register_lock);
1358
1359         if (atomic_read(&dc->running)) {
1360                 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1361                 del_gendisk(dc->disk.disk);
1362         }
1363         bcache_device_free(&dc->disk);
1364         list_del(&dc->list);
1365
1366         mutex_unlock(&bch_register_lock);
1367
1368         if (dc->sb_disk)
1369                 put_page(virt_to_page(dc->sb_disk));
1370
1371         if (dc->bdev_handle)
1372                 bdev_release(dc->bdev_handle);
1373
1374         wake_up(&unregister_wait);
1375
1376         kobject_put(&dc->disk.kobj);
1377 }
1378
1379 static void cached_dev_flush(struct closure *cl)
1380 {
1381         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1382         struct bcache_device *d = &dc->disk;
1383
1384         mutex_lock(&bch_register_lock);
1385         bcache_device_unlink(d);
1386         mutex_unlock(&bch_register_lock);
1387
1388         bch_cache_accounting_destroy(&dc->accounting);
1389         kobject_del(&d->kobj);
1390
1391         continue_at(cl, cached_dev_free, system_wq);
1392 }
1393
1394 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1395 {
1396         int ret;
1397         struct io *io;
1398         struct request_queue *q = bdev_get_queue(dc->bdev);
1399
1400         __module_get(THIS_MODULE);
1401         INIT_LIST_HEAD(&dc->list);
1402         closure_init(&dc->disk.cl, NULL);
1403         set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1404         kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1405         INIT_WORK(&dc->detach, cached_dev_detach_finish);
1406         sema_init(&dc->sb_write_mutex, 1);
1407         INIT_LIST_HEAD(&dc->io_lru);
1408         spin_lock_init(&dc->io_lock);
1409         bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1410
1411         dc->sequential_cutoff           = 4 << 20;
1412
1413         for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1414                 list_add(&io->lru, &dc->io_lru);
1415                 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1416         }
1417
1418         dc->disk.stripe_size = q->limits.io_opt >> 9;
1419
1420         if (dc->disk.stripe_size)
1421                 dc->partial_stripes_expensive =
1422                         q->limits.raid_partial_stripes_expensive;
1423
1424         ret = bcache_device_init(&dc->disk, block_size,
1425                          bdev_nr_sectors(dc->bdev) - dc->sb.data_offset,
1426                          dc->bdev, &bcache_cached_ops);
1427         if (ret)
1428                 return ret;
1429
1430         blk_queue_io_opt(dc->disk.disk->queue,
1431                 max(queue_io_opt(dc->disk.disk->queue), queue_io_opt(q)));
1432
1433         atomic_set(&dc->io_errors, 0);
1434         dc->io_disable = false;
1435         dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1436         /* default to auto */
1437         dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1438
1439         bch_cached_dev_request_init(dc);
1440         bch_cached_dev_writeback_init(dc);
1441         return 0;
1442 }
1443
1444 /* Cached device - bcache superblock */
1445
1446 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1447                                  struct bdev_handle *bdev_handle,
1448                                  struct cached_dev *dc)
1449 {
1450         const char *err = "cannot allocate memory";
1451         struct cache_set *c;
1452         int ret = -ENOMEM;
1453
1454         memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1455         dc->bdev_handle = bdev_handle;
1456         dc->bdev = bdev_handle->bdev;
1457         dc->sb_disk = sb_disk;
1458
1459         if (cached_dev_init(dc, sb->block_size << 9))
1460                 goto err;
1461
1462         err = "error creating kobject";
1463         if (kobject_add(&dc->disk.kobj, bdev_kobj(dc->bdev), "bcache"))
1464                 goto err;
1465         if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1466                 goto err;
1467
1468         pr_info("registered backing device %pg\n", dc->bdev);
1469
1470         list_add(&dc->list, &uncached_devices);
1471         /* attach to a matched cache set if it exists */
1472         list_for_each_entry(c, &bch_cache_sets, list)
1473                 bch_cached_dev_attach(dc, c, NULL);
1474
1475         if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1476             BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1477                 err = "failed to run cached device";
1478                 ret = bch_cached_dev_run(dc);
1479                 if (ret)
1480                         goto err;
1481         }
1482
1483         return 0;
1484 err:
1485         pr_notice("error %pg: %s\n", dc->bdev, err);
1486         bcache_device_stop(&dc->disk);
1487         return ret;
1488 }
1489
1490 /* Flash only volumes */
1491
1492 /* When d->kobj released */
1493 void bch_flash_dev_release(struct kobject *kobj)
1494 {
1495         struct bcache_device *d = container_of(kobj, struct bcache_device,
1496                                                kobj);
1497         kfree(d);
1498 }
1499
1500 static void flash_dev_free(struct closure *cl)
1501 {
1502         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1503
1504         mutex_lock(&bch_register_lock);
1505         atomic_long_sub(bcache_dev_sectors_dirty(d),
1506                         &d->c->flash_dev_dirty_sectors);
1507         del_gendisk(d->disk);
1508         bcache_device_free(d);
1509         mutex_unlock(&bch_register_lock);
1510         kobject_put(&d->kobj);
1511 }
1512
1513 static void flash_dev_flush(struct closure *cl)
1514 {
1515         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1516
1517         mutex_lock(&bch_register_lock);
1518         bcache_device_unlink(d);
1519         mutex_unlock(&bch_register_lock);
1520         kobject_del(&d->kobj);
1521         continue_at(cl, flash_dev_free, system_wq);
1522 }
1523
1524 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1525 {
1526         int err = -ENOMEM;
1527         struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1528                                           GFP_KERNEL);
1529         if (!d)
1530                 goto err_ret;
1531
1532         closure_init(&d->cl, NULL);
1533         set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1534
1535         kobject_init(&d->kobj, &bch_flash_dev_ktype);
1536
1537         if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1538                         NULL, &bcache_flash_ops))
1539                 goto err;
1540
1541         bcache_device_attach(d, c, u - c->uuids);
1542         bch_sectors_dirty_init(d);
1543         bch_flash_dev_request_init(d);
1544         err = add_disk(d->disk);
1545         if (err)
1546                 goto err;
1547
1548         err = kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache");
1549         if (err)
1550                 goto err;
1551
1552         bcache_device_link(d, c, "volume");
1553
1554         if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1555                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1556                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1557                 set_disk_ro(d->disk, 1);
1558         }
1559
1560         return 0;
1561 err:
1562         kobject_put(&d->kobj);
1563 err_ret:
1564         return err;
1565 }
1566
1567 static int flash_devs_run(struct cache_set *c)
1568 {
1569         int ret = 0;
1570         struct uuid_entry *u;
1571
1572         for (u = c->uuids;
1573              u < c->uuids + c->nr_uuids && !ret;
1574              u++)
1575                 if (UUID_FLASH_ONLY(u))
1576                         ret = flash_dev_run(c, u);
1577
1578         return ret;
1579 }
1580
1581 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1582 {
1583         struct uuid_entry *u;
1584
1585         if (test_bit(CACHE_SET_STOPPING, &c->flags))
1586                 return -EINTR;
1587
1588         if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1589                 return -EPERM;
1590
1591         u = uuid_find_empty(c);
1592         if (!u) {
1593                 pr_err("Can't create volume, no room for UUID\n");
1594                 return -EINVAL;
1595         }
1596
1597         get_random_bytes(u->uuid, 16);
1598         memset(u->label, 0, 32);
1599         u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1600
1601         SET_UUID_FLASH_ONLY(u, 1);
1602         u->sectors = size >> 9;
1603
1604         bch_uuid_write(c);
1605
1606         return flash_dev_run(c, u);
1607 }
1608
1609 bool bch_cached_dev_error(struct cached_dev *dc)
1610 {
1611         if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1612                 return false;
1613
1614         dc->io_disable = true;
1615         /* make others know io_disable is true earlier */
1616         smp_mb();
1617
1618         pr_err("stop %s: too many IO errors on backing device %pg\n",
1619                dc->disk.disk->disk_name, dc->bdev);
1620
1621         bcache_device_stop(&dc->disk);
1622         return true;
1623 }
1624
1625 /* Cache set */
1626
1627 __printf(2, 3)
1628 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1629 {
1630         struct va_format vaf;
1631         va_list args;
1632
1633         if (c->on_error != ON_ERROR_PANIC &&
1634             test_bit(CACHE_SET_STOPPING, &c->flags))
1635                 return false;
1636
1637         if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1638                 pr_info("CACHE_SET_IO_DISABLE already set\n");
1639
1640         /*
1641          * XXX: we can be called from atomic context
1642          * acquire_console_sem();
1643          */
1644
1645         va_start(args, fmt);
1646
1647         vaf.fmt = fmt;
1648         vaf.va = &args;
1649
1650         pr_err("error on %pU: %pV, disabling caching\n",
1651                c->set_uuid, &vaf);
1652
1653         va_end(args);
1654
1655         if (c->on_error == ON_ERROR_PANIC)
1656                 panic("panic forced after error\n");
1657
1658         bch_cache_set_unregister(c);
1659         return true;
1660 }
1661
1662 /* When c->kobj released */
1663 void bch_cache_set_release(struct kobject *kobj)
1664 {
1665         struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1666
1667         kfree(c);
1668         module_put(THIS_MODULE);
1669 }
1670
1671 static void cache_set_free(struct closure *cl)
1672 {
1673         struct cache_set *c = container_of(cl, struct cache_set, cl);
1674         struct cache *ca;
1675
1676         debugfs_remove(c->debug);
1677
1678         bch_open_buckets_free(c);
1679         bch_btree_cache_free(c);
1680         bch_journal_free(c);
1681
1682         mutex_lock(&bch_register_lock);
1683         bch_bset_sort_state_free(&c->sort);
1684         free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1685
1686         ca = c->cache;
1687         if (ca) {
1688                 ca->set = NULL;
1689                 c->cache = NULL;
1690                 kobject_put(&ca->kobj);
1691         }
1692
1693
1694         if (c->moving_gc_wq)
1695                 destroy_workqueue(c->moving_gc_wq);
1696         bioset_exit(&c->bio_split);
1697         mempool_exit(&c->fill_iter);
1698         mempool_exit(&c->bio_meta);
1699         mempool_exit(&c->search);
1700         kfree(c->devices);
1701
1702         list_del(&c->list);
1703         mutex_unlock(&bch_register_lock);
1704
1705         pr_info("Cache set %pU unregistered\n", c->set_uuid);
1706         wake_up(&unregister_wait);
1707
1708         closure_debug_destroy(&c->cl);
1709         kobject_put(&c->kobj);
1710 }
1711
1712 static void cache_set_flush(struct closure *cl)
1713 {
1714         struct cache_set *c = container_of(cl, struct cache_set, caching);
1715         struct cache *ca = c->cache;
1716         struct btree *b;
1717
1718         bch_cache_accounting_destroy(&c->accounting);
1719
1720         kobject_put(&c->internal);
1721         kobject_del(&c->kobj);
1722
1723         if (!IS_ERR_OR_NULL(c->gc_thread))
1724                 kthread_stop(c->gc_thread);
1725
1726         if (!IS_ERR(c->root))
1727                 list_add(&c->root->list, &c->btree_cache);
1728
1729         /*
1730          * Avoid flushing cached nodes if cache set is retiring
1731          * due to too many I/O errors detected.
1732          */
1733         if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1734                 list_for_each_entry(b, &c->btree_cache, list) {
1735                         mutex_lock(&b->write_lock);
1736                         if (btree_node_dirty(b))
1737                                 __bch_btree_node_write(b, NULL);
1738                         mutex_unlock(&b->write_lock);
1739                 }
1740
1741         if (ca->alloc_thread)
1742                 kthread_stop(ca->alloc_thread);
1743
1744         if (c->journal.cur) {
1745                 cancel_delayed_work_sync(&c->journal.work);
1746                 /* flush last journal entry if needed */
1747                 c->journal.work.work.func(&c->journal.work.work);
1748         }
1749
1750         closure_return(cl);
1751 }
1752
1753 /*
1754  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1755  * cache set is unregistering due to too many I/O errors. In this condition,
1756  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1757  * value and whether the broken cache has dirty data:
1758  *
1759  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1760  *  BCH_CACHED_STOP_AUTO               0               NO
1761  *  BCH_CACHED_STOP_AUTO               1               YES
1762  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1763  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1764  *
1765  * The expected behavior is, if stop_when_cache_set_failed is configured to
1766  * "auto" via sysfs interface, the bcache device will not be stopped if the
1767  * backing device is clean on the broken cache device.
1768  */
1769 static void conditional_stop_bcache_device(struct cache_set *c,
1770                                            struct bcache_device *d,
1771                                            struct cached_dev *dc)
1772 {
1773         if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1774                 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1775                         d->disk->disk_name, c->set_uuid);
1776                 bcache_device_stop(d);
1777         } else if (atomic_read(&dc->has_dirty)) {
1778                 /*
1779                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1780                  * and dc->has_dirty == 1
1781                  */
1782                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1783                         d->disk->disk_name);
1784                 /*
1785                  * There might be a small time gap that cache set is
1786                  * released but bcache device is not. Inside this time
1787                  * gap, regular I/O requests will directly go into
1788                  * backing device as no cache set attached to. This
1789                  * behavior may also introduce potential inconsistence
1790                  * data in writeback mode while cache is dirty.
1791                  * Therefore before calling bcache_device_stop() due
1792                  * to a broken cache device, dc->io_disable should be
1793                  * explicitly set to true.
1794                  */
1795                 dc->io_disable = true;
1796                 /* make others know io_disable is true earlier */
1797                 smp_mb();
1798                 bcache_device_stop(d);
1799         } else {
1800                 /*
1801                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1802                  * and dc->has_dirty == 0
1803                  */
1804                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1805                         d->disk->disk_name);
1806         }
1807 }
1808
1809 static void __cache_set_unregister(struct closure *cl)
1810 {
1811         struct cache_set *c = container_of(cl, struct cache_set, caching);
1812         struct cached_dev *dc;
1813         struct bcache_device *d;
1814         size_t i;
1815
1816         mutex_lock(&bch_register_lock);
1817
1818         for (i = 0; i < c->devices_max_used; i++) {
1819                 d = c->devices[i];
1820                 if (!d)
1821                         continue;
1822
1823                 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1824                     test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1825                         dc = container_of(d, struct cached_dev, disk);
1826                         bch_cached_dev_detach(dc);
1827                         if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1828                                 conditional_stop_bcache_device(c, d, dc);
1829                 } else {
1830                         bcache_device_stop(d);
1831                 }
1832         }
1833
1834         mutex_unlock(&bch_register_lock);
1835
1836         continue_at(cl, cache_set_flush, system_wq);
1837 }
1838
1839 void bch_cache_set_stop(struct cache_set *c)
1840 {
1841         if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1842                 /* closure_fn set to __cache_set_unregister() */
1843                 closure_queue(&c->caching);
1844 }
1845
1846 void bch_cache_set_unregister(struct cache_set *c)
1847 {
1848         set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1849         bch_cache_set_stop(c);
1850 }
1851
1852 #define alloc_meta_bucket_pages(gfp, sb)                \
1853         ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1854
1855 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1856 {
1857         int iter_size;
1858         struct cache *ca = container_of(sb, struct cache, sb);
1859         struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1860
1861         if (!c)
1862                 return NULL;
1863
1864         __module_get(THIS_MODULE);
1865         closure_init(&c->cl, NULL);
1866         set_closure_fn(&c->cl, cache_set_free, system_wq);
1867
1868         closure_init(&c->caching, &c->cl);
1869         set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1870
1871         /* Maybe create continue_at_noreturn() and use it here? */
1872         closure_set_stopped(&c->cl);
1873         closure_put(&c->cl);
1874
1875         kobject_init(&c->kobj, &bch_cache_set_ktype);
1876         kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1877
1878         bch_cache_accounting_init(&c->accounting, &c->cl);
1879
1880         memcpy(c->set_uuid, sb->set_uuid, 16);
1881
1882         c->cache                = ca;
1883         c->cache->set           = c;
1884         c->bucket_bits          = ilog2(sb->bucket_size);
1885         c->block_bits           = ilog2(sb->block_size);
1886         c->nr_uuids             = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1887         c->devices_max_used     = 0;
1888         atomic_set(&c->attached_dev_nr, 0);
1889         c->btree_pages          = meta_bucket_pages(sb);
1890         if (c->btree_pages > BTREE_MAX_PAGES)
1891                 c->btree_pages = max_t(int, c->btree_pages / 4,
1892                                        BTREE_MAX_PAGES);
1893
1894         sema_init(&c->sb_write_mutex, 1);
1895         mutex_init(&c->bucket_lock);
1896         init_waitqueue_head(&c->btree_cache_wait);
1897         spin_lock_init(&c->btree_cannibalize_lock);
1898         init_waitqueue_head(&c->bucket_wait);
1899         init_waitqueue_head(&c->gc_wait);
1900         sema_init(&c->uuid_write_mutex, 1);
1901
1902         spin_lock_init(&c->btree_gc_time.lock);
1903         spin_lock_init(&c->btree_split_time.lock);
1904         spin_lock_init(&c->btree_read_time.lock);
1905
1906         bch_moving_init_cache_set(c);
1907
1908         INIT_LIST_HEAD(&c->list);
1909         INIT_LIST_HEAD(&c->cached_devs);
1910         INIT_LIST_HEAD(&c->btree_cache);
1911         INIT_LIST_HEAD(&c->btree_cache_freeable);
1912         INIT_LIST_HEAD(&c->btree_cache_freed);
1913         INIT_LIST_HEAD(&c->data_buckets);
1914
1915         iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1916                 sizeof(struct btree_iter_set);
1917
1918         c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1919         if (!c->devices)
1920                 goto err;
1921
1922         if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1923                 goto err;
1924
1925         if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1926                         sizeof(struct bbio) +
1927                         sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1928                 goto err;
1929
1930         if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1931                 goto err;
1932
1933         if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1934                         BIOSET_NEED_RESCUER))
1935                 goto err;
1936
1937         c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1938         if (!c->uuids)
1939                 goto err;
1940
1941         c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1942         if (!c->moving_gc_wq)
1943                 goto err;
1944
1945         if (bch_journal_alloc(c))
1946                 goto err;
1947
1948         if (bch_btree_cache_alloc(c))
1949                 goto err;
1950
1951         if (bch_open_buckets_alloc(c))
1952                 goto err;
1953
1954         if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1955                 goto err;
1956
1957         c->congested_read_threshold_us  = 2000;
1958         c->congested_write_threshold_us = 20000;
1959         c->error_limit  = DEFAULT_IO_ERROR_LIMIT;
1960         c->idle_max_writeback_rate_enabled = 1;
1961         WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1962
1963         return c;
1964 err:
1965         bch_cache_set_unregister(c);
1966         return NULL;
1967 }
1968
1969 static int run_cache_set(struct cache_set *c)
1970 {
1971         const char *err = "cannot allocate memory";
1972         struct cached_dev *dc, *t;
1973         struct cache *ca = c->cache;
1974         struct closure cl;
1975         LIST_HEAD(journal);
1976         struct journal_replay *l;
1977
1978         closure_init_stack(&cl);
1979
1980         c->nbuckets = ca->sb.nbuckets;
1981         set_gc_sectors(c);
1982
1983         if (CACHE_SYNC(&c->cache->sb)) {
1984                 struct bkey *k;
1985                 struct jset *j;
1986
1987                 err = "cannot allocate memory for journal";
1988                 if (bch_journal_read(c, &journal))
1989                         goto err;
1990
1991                 pr_debug("btree_journal_read() done\n");
1992
1993                 err = "no journal entries found";
1994                 if (list_empty(&journal))
1995                         goto err;
1996
1997                 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1998
1999                 err = "IO error reading priorities";
2000                 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
2001                         goto err;
2002
2003                 /*
2004                  * If prio_read() fails it'll call cache_set_error and we'll
2005                  * tear everything down right away, but if we perhaps checked
2006                  * sooner we could avoid journal replay.
2007                  */
2008
2009                 k = &j->btree_root;
2010
2011                 err = "bad btree root";
2012                 if (__bch_btree_ptr_invalid(c, k))
2013                         goto err;
2014
2015                 err = "error reading btree root";
2016                 c->root = bch_btree_node_get(c, NULL, k,
2017                                              j->btree_level,
2018                                              true, NULL);
2019                 if (IS_ERR_OR_NULL(c->root))
2020                         goto err;
2021
2022                 list_del_init(&c->root->list);
2023                 rw_unlock(true, c->root);
2024
2025                 err = uuid_read(c, j, &cl);
2026                 if (err)
2027                         goto err;
2028
2029                 err = "error in recovery";
2030                 if (bch_btree_check(c))
2031                         goto err;
2032
2033                 bch_journal_mark(c, &journal);
2034                 bch_initial_gc_finish(c);
2035                 pr_debug("btree_check() done\n");
2036
2037                 /*
2038                  * bcache_journal_next() can't happen sooner, or
2039                  * btree_gc_finish() will give spurious errors about last_gc >
2040                  * gc_gen - this is a hack but oh well.
2041                  */
2042                 bch_journal_next(&c->journal);
2043
2044                 err = "error starting allocator thread";
2045                 if (bch_cache_allocator_start(ca))
2046                         goto err;
2047
2048                 /*
2049                  * First place it's safe to allocate: btree_check() and
2050                  * btree_gc_finish() have to run before we have buckets to
2051                  * allocate, and bch_bucket_alloc_set() might cause a journal
2052                  * entry to be written so bcache_journal_next() has to be called
2053                  * first.
2054                  *
2055                  * If the uuids were in the old format we have to rewrite them
2056                  * before the next journal entry is written:
2057                  */
2058                 if (j->version < BCACHE_JSET_VERSION_UUID)
2059                         __uuid_write(c);
2060
2061                 err = "bcache: replay journal failed";
2062                 if (bch_journal_replay(c, &journal))
2063                         goto err;
2064         } else {
2065                 unsigned int j;
2066
2067                 pr_notice("invalidating existing data\n");
2068                 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2069                                         2, SB_JOURNAL_BUCKETS);
2070
2071                 for (j = 0; j < ca->sb.keys; j++)
2072                         ca->sb.d[j] = ca->sb.first_bucket + j;
2073
2074                 bch_initial_gc_finish(c);
2075
2076                 err = "error starting allocator thread";
2077                 if (bch_cache_allocator_start(ca))
2078                         goto err;
2079
2080                 mutex_lock(&c->bucket_lock);
2081                 bch_prio_write(ca, true);
2082                 mutex_unlock(&c->bucket_lock);
2083
2084                 err = "cannot allocate new UUID bucket";
2085                 if (__uuid_write(c))
2086                         goto err;
2087
2088                 err = "cannot allocate new btree root";
2089                 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2090                 if (IS_ERR(c->root))
2091                         goto err;
2092
2093                 mutex_lock(&c->root->write_lock);
2094                 bkey_copy_key(&c->root->key, &MAX_KEY);
2095                 bch_btree_node_write(c->root, &cl);
2096                 mutex_unlock(&c->root->write_lock);
2097
2098                 bch_btree_set_root(c->root);
2099                 rw_unlock(true, c->root);
2100
2101                 /*
2102                  * We don't want to write the first journal entry until
2103                  * everything is set up - fortunately journal entries won't be
2104                  * written until the SET_CACHE_SYNC() here:
2105                  */
2106                 SET_CACHE_SYNC(&c->cache->sb, true);
2107
2108                 bch_journal_next(&c->journal);
2109                 bch_journal_meta(c, &cl);
2110         }
2111
2112         err = "error starting gc thread";
2113         if (bch_gc_thread_start(c))
2114                 goto err;
2115
2116         closure_sync(&cl);
2117         c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2118         bcache_write_super(c);
2119
2120         if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2121                 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2122
2123         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2124                 bch_cached_dev_attach(dc, c, NULL);
2125
2126         flash_devs_run(c);
2127
2128         bch_journal_space_reserve(&c->journal);
2129         set_bit(CACHE_SET_RUNNING, &c->flags);
2130         return 0;
2131 err:
2132         while (!list_empty(&journal)) {
2133                 l = list_first_entry(&journal, struct journal_replay, list);
2134                 list_del(&l->list);
2135                 kfree(l);
2136         }
2137
2138         closure_sync(&cl);
2139
2140         bch_cache_set_error(c, "%s", err);
2141
2142         return -EIO;
2143 }
2144
2145 static const char *register_cache_set(struct cache *ca)
2146 {
2147         char buf[12];
2148         const char *err = "cannot allocate memory";
2149         struct cache_set *c;
2150
2151         list_for_each_entry(c, &bch_cache_sets, list)
2152                 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2153                         if (c->cache)
2154                                 return "duplicate cache set member";
2155
2156                         goto found;
2157                 }
2158
2159         c = bch_cache_set_alloc(&ca->sb);
2160         if (!c)
2161                 return err;
2162
2163         err = "error creating kobject";
2164         if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2165             kobject_add(&c->internal, &c->kobj, "internal"))
2166                 goto err;
2167
2168         if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2169                 goto err;
2170
2171         bch_debug_init_cache_set(c);
2172
2173         list_add(&c->list, &bch_cache_sets);
2174 found:
2175         sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2176         if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2177             sysfs_create_link(&c->kobj, &ca->kobj, buf))
2178                 goto err;
2179
2180         kobject_get(&ca->kobj);
2181         ca->set = c;
2182         ca->set->cache = ca;
2183
2184         err = "failed to run cache set";
2185         if (run_cache_set(c) < 0)
2186                 goto err;
2187
2188         return NULL;
2189 err:
2190         bch_cache_set_unregister(c);
2191         return err;
2192 }
2193
2194 /* Cache device */
2195
2196 /* When ca->kobj released */
2197 void bch_cache_release(struct kobject *kobj)
2198 {
2199         struct cache *ca = container_of(kobj, struct cache, kobj);
2200         unsigned int i;
2201
2202         if (ca->set) {
2203                 BUG_ON(ca->set->cache != ca);
2204                 ca->set->cache = NULL;
2205         }
2206
2207         free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2208         kfree(ca->prio_buckets);
2209         vfree(ca->buckets);
2210
2211         free_heap(&ca->heap);
2212         free_fifo(&ca->free_inc);
2213
2214         for (i = 0; i < RESERVE_NR; i++)
2215                 free_fifo(&ca->free[i]);
2216
2217         if (ca->sb_disk)
2218                 put_page(virt_to_page(ca->sb_disk));
2219
2220         if (ca->bdev_handle)
2221                 bdev_release(ca->bdev_handle);
2222
2223         kfree(ca);
2224         module_put(THIS_MODULE);
2225 }
2226
2227 static int cache_alloc(struct cache *ca)
2228 {
2229         size_t free;
2230         size_t btree_buckets;
2231         struct bucket *b;
2232         int ret = -ENOMEM;
2233         const char *err = NULL;
2234
2235         __module_get(THIS_MODULE);
2236         kobject_init(&ca->kobj, &bch_cache_ktype);
2237
2238         bio_init(&ca->journal.bio, NULL, ca->journal.bio.bi_inline_vecs, 8, 0);
2239
2240         /*
2241          * when ca->sb.njournal_buckets is not zero, journal exists,
2242          * and in bch_journal_replay(), tree node may split,
2243          * so bucket of RESERVE_BTREE type is needed,
2244          * the worst situation is all journal buckets are valid journal,
2245          * and all the keys need to replay,
2246          * so the number of  RESERVE_BTREE type buckets should be as much
2247          * as journal buckets
2248          */
2249         btree_buckets = ca->sb.njournal_buckets ?: 8;
2250         free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2251         if (!free) {
2252                 ret = -EPERM;
2253                 err = "ca->sb.nbuckets is too small";
2254                 goto err_free;
2255         }
2256
2257         if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2258                                                 GFP_KERNEL)) {
2259                 err = "ca->free[RESERVE_BTREE] alloc failed";
2260                 goto err_btree_alloc;
2261         }
2262
2263         if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2264                                                         GFP_KERNEL)) {
2265                 err = "ca->free[RESERVE_PRIO] alloc failed";
2266                 goto err_prio_alloc;
2267         }
2268
2269         if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2270                 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2271                 goto err_movinggc_alloc;
2272         }
2273
2274         if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2275                 err = "ca->free[RESERVE_NONE] alloc failed";
2276                 goto err_none_alloc;
2277         }
2278
2279         if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2280                 err = "ca->free_inc alloc failed";
2281                 goto err_free_inc_alloc;
2282         }
2283
2284         if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2285                 err = "ca->heap alloc failed";
2286                 goto err_heap_alloc;
2287         }
2288
2289         ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2290                               ca->sb.nbuckets));
2291         if (!ca->buckets) {
2292                 err = "ca->buckets alloc failed";
2293                 goto err_buckets_alloc;
2294         }
2295
2296         ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2297                                    prio_buckets(ca), 2),
2298                                    GFP_KERNEL);
2299         if (!ca->prio_buckets) {
2300                 err = "ca->prio_buckets alloc failed";
2301                 goto err_prio_buckets_alloc;
2302         }
2303
2304         ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2305         if (!ca->disk_buckets) {
2306                 err = "ca->disk_buckets alloc failed";
2307                 goto err_disk_buckets_alloc;
2308         }
2309
2310         ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2311
2312         for_each_bucket(b, ca)
2313                 atomic_set(&b->pin, 0);
2314         return 0;
2315
2316 err_disk_buckets_alloc:
2317         kfree(ca->prio_buckets);
2318 err_prio_buckets_alloc:
2319         vfree(ca->buckets);
2320 err_buckets_alloc:
2321         free_heap(&ca->heap);
2322 err_heap_alloc:
2323         free_fifo(&ca->free_inc);
2324 err_free_inc_alloc:
2325         free_fifo(&ca->free[RESERVE_NONE]);
2326 err_none_alloc:
2327         free_fifo(&ca->free[RESERVE_MOVINGGC]);
2328 err_movinggc_alloc:
2329         free_fifo(&ca->free[RESERVE_PRIO]);
2330 err_prio_alloc:
2331         free_fifo(&ca->free[RESERVE_BTREE]);
2332 err_btree_alloc:
2333 err_free:
2334         module_put(THIS_MODULE);
2335         if (err)
2336                 pr_notice("error %pg: %s\n", ca->bdev, err);
2337         return ret;
2338 }
2339
2340 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2341                                 struct bdev_handle *bdev_handle,
2342                                 struct cache *ca)
2343 {
2344         const char *err = NULL; /* must be set for any error case */
2345         int ret = 0;
2346
2347         memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2348         ca->bdev_handle = bdev_handle;
2349         ca->bdev = bdev_handle->bdev;
2350         ca->sb_disk = sb_disk;
2351
2352         if (bdev_max_discard_sectors((bdev_handle->bdev)))
2353                 ca->discard = CACHE_DISCARD(&ca->sb);
2354
2355         ret = cache_alloc(ca);
2356         if (ret != 0) {
2357                 if (ret == -ENOMEM)
2358                         err = "cache_alloc(): -ENOMEM";
2359                 else if (ret == -EPERM)
2360                         err = "cache_alloc(): cache device is too small";
2361                 else
2362                         err = "cache_alloc(): unknown error";
2363                 pr_notice("error %pg: %s\n", bdev_handle->bdev, err);
2364                 /*
2365                  * If we failed here, it means ca->kobj is not initialized yet,
2366                  * kobject_put() won't be called and there is no chance to
2367                  * call bdev_release() to bdev in bch_cache_release(). So
2368                  * we explicitly call bdev_release() here.
2369                  */
2370                 bdev_release(bdev_handle);
2371                 return ret;
2372         }
2373
2374         if (kobject_add(&ca->kobj, bdev_kobj(bdev_handle->bdev), "bcache")) {
2375                 pr_notice("error %pg: error calling kobject_add\n",
2376                           bdev_handle->bdev);
2377                 ret = -ENOMEM;
2378                 goto out;
2379         }
2380
2381         mutex_lock(&bch_register_lock);
2382         err = register_cache_set(ca);
2383         mutex_unlock(&bch_register_lock);
2384
2385         if (err) {
2386                 ret = -ENODEV;
2387                 goto out;
2388         }
2389
2390         pr_info("registered cache device %pg\n", ca->bdev_handle->bdev);
2391
2392 out:
2393         kobject_put(&ca->kobj);
2394         return ret;
2395 }
2396
2397 /* Global interfaces/init */
2398
2399 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2400                                const char *buffer, size_t size);
2401 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2402                                          struct kobj_attribute *attr,
2403                                          const char *buffer, size_t size);
2404
2405 kobj_attribute_write(register,          register_bcache);
2406 kobj_attribute_write(register_quiet,    register_bcache);
2407 kobj_attribute_write(pendings_cleanup,  bch_pending_bdevs_cleanup);
2408
2409 static bool bch_is_open_backing(dev_t dev)
2410 {
2411         struct cache_set *c, *tc;
2412         struct cached_dev *dc, *t;
2413
2414         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2415                 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2416                         if (dc->bdev->bd_dev == dev)
2417                                 return true;
2418         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2419                 if (dc->bdev->bd_dev == dev)
2420                         return true;
2421         return false;
2422 }
2423
2424 static bool bch_is_open_cache(dev_t dev)
2425 {
2426         struct cache_set *c, *tc;
2427
2428         list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2429                 struct cache *ca = c->cache;
2430
2431                 if (ca->bdev->bd_dev == dev)
2432                         return true;
2433         }
2434
2435         return false;
2436 }
2437
2438 static bool bch_is_open(dev_t dev)
2439 {
2440         return bch_is_open_cache(dev) || bch_is_open_backing(dev);
2441 }
2442
2443 struct async_reg_args {
2444         struct delayed_work reg_work;
2445         char *path;
2446         struct cache_sb *sb;
2447         struct cache_sb_disk *sb_disk;
2448         struct bdev_handle *bdev_handle;
2449         void *holder;
2450 };
2451
2452 static void register_bdev_worker(struct work_struct *work)
2453 {
2454         int fail = false;
2455         struct async_reg_args *args =
2456                 container_of(work, struct async_reg_args, reg_work.work);
2457
2458         mutex_lock(&bch_register_lock);
2459         if (register_bdev(args->sb, args->sb_disk, args->bdev_handle,
2460                           args->holder) < 0)
2461                 fail = true;
2462         mutex_unlock(&bch_register_lock);
2463
2464         if (fail)
2465                 pr_info("error %s: fail to register backing device\n",
2466                         args->path);
2467         kfree(args->sb);
2468         kfree(args->path);
2469         kfree(args);
2470         module_put(THIS_MODULE);
2471 }
2472
2473 static void register_cache_worker(struct work_struct *work)
2474 {
2475         int fail = false;
2476         struct async_reg_args *args =
2477                 container_of(work, struct async_reg_args, reg_work.work);
2478
2479         /* blkdev_put() will be called in bch_cache_release() */
2480         if (register_cache(args->sb, args->sb_disk, args->bdev_handle,
2481                            args->holder))
2482                 fail = true;
2483
2484         if (fail)
2485                 pr_info("error %s: fail to register cache device\n",
2486                         args->path);
2487         kfree(args->sb);
2488         kfree(args->path);
2489         kfree(args);
2490         module_put(THIS_MODULE);
2491 }
2492
2493 static void register_device_async(struct async_reg_args *args)
2494 {
2495         if (SB_IS_BDEV(args->sb))
2496                 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2497         else
2498                 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2499
2500         /* 10 jiffies is enough for a delay */
2501         queue_delayed_work(system_wq, &args->reg_work, 10);
2502 }
2503
2504 static void *alloc_holder_object(struct cache_sb *sb)
2505 {
2506         if (SB_IS_BDEV(sb))
2507                 return kzalloc(sizeof(struct cached_dev), GFP_KERNEL);
2508         return kzalloc(sizeof(struct cache), GFP_KERNEL);
2509 }
2510
2511 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2512                                const char *buffer, size_t size)
2513 {
2514         const char *err;
2515         char *path = NULL;
2516         struct cache_sb *sb;
2517         struct cache_sb_disk *sb_disk;
2518         struct bdev_handle *bdev_handle, *bdev_handle2;
2519         void *holder = NULL;
2520         ssize_t ret;
2521         bool async_registration = false;
2522         bool quiet = false;
2523
2524 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2525         async_registration = true;
2526 #endif
2527
2528         ret = -EBUSY;
2529         err = "failed to reference bcache module";
2530         if (!try_module_get(THIS_MODULE))
2531                 goto out;
2532
2533         /* For latest state of bcache_is_reboot */
2534         smp_mb();
2535         err = "bcache is in reboot";
2536         if (bcache_is_reboot)
2537                 goto out_module_put;
2538
2539         ret = -ENOMEM;
2540         err = "cannot allocate memory";
2541         path = kstrndup(buffer, size, GFP_KERNEL);
2542         if (!path)
2543                 goto out_module_put;
2544
2545         sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2546         if (!sb)
2547                 goto out_free_path;
2548
2549         ret = -EINVAL;
2550         err = "failed to open device";
2551         bdev_handle = bdev_open_by_path(strim(path), BLK_OPEN_READ, NULL, NULL);
2552         if (IS_ERR(bdev_handle))
2553                 goto out_free_sb;
2554
2555         err = "failed to set blocksize";
2556         if (set_blocksize(bdev_handle->bdev, 4096))
2557                 goto out_blkdev_put;
2558
2559         err = read_super(sb, bdev_handle->bdev, &sb_disk);
2560         if (err)
2561                 goto out_blkdev_put;
2562
2563         holder = alloc_holder_object(sb);
2564         if (!holder) {
2565                 ret = -ENOMEM;
2566                 err = "cannot allocate memory";
2567                 goto out_put_sb_page;
2568         }
2569
2570         /* Now reopen in exclusive mode with proper holder */
2571         bdev_handle2 = bdev_open_by_dev(bdev_handle->bdev->bd_dev,
2572                         BLK_OPEN_READ | BLK_OPEN_WRITE, holder, NULL);
2573         bdev_release(bdev_handle);
2574         bdev_handle = bdev_handle2;
2575         if (IS_ERR(bdev_handle)) {
2576                 ret = PTR_ERR(bdev_handle);
2577                 bdev_handle = NULL;
2578                 if (ret == -EBUSY) {
2579                         dev_t dev;
2580
2581                         mutex_lock(&bch_register_lock);
2582                         if (lookup_bdev(strim(path), &dev) == 0 &&
2583                             bch_is_open(dev))
2584                                 err = "device already registered";
2585                         else
2586                                 err = "device busy";
2587                         mutex_unlock(&bch_register_lock);
2588                         if (attr == &ksysfs_register_quiet) {
2589                                 quiet = true;
2590                                 ret = size;
2591                         }
2592                 }
2593                 goto out_free_holder;
2594         }
2595
2596         err = "failed to register device";
2597
2598         if (async_registration) {
2599                 /* register in asynchronous way */
2600                 struct async_reg_args *args =
2601                         kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2602
2603                 if (!args) {
2604                         ret = -ENOMEM;
2605                         err = "cannot allocate memory";
2606                         goto out_free_holder;
2607                 }
2608
2609                 args->path      = path;
2610                 args->sb        = sb;
2611                 args->sb_disk   = sb_disk;
2612                 args->bdev_handle       = bdev_handle;
2613                 args->holder    = holder;
2614                 register_device_async(args);
2615                 /* No wait and returns to user space */
2616                 goto async_done;
2617         }
2618
2619         if (SB_IS_BDEV(sb)) {
2620                 mutex_lock(&bch_register_lock);
2621                 ret = register_bdev(sb, sb_disk, bdev_handle, holder);
2622                 mutex_unlock(&bch_register_lock);
2623                 /* blkdev_put() will be called in cached_dev_free() */
2624                 if (ret < 0)
2625                         goto out_free_sb;
2626         } else {
2627                 /* blkdev_put() will be called in bch_cache_release() */
2628                 ret = register_cache(sb, sb_disk, bdev_handle, holder);
2629                 if (ret)
2630                         goto out_free_sb;
2631         }
2632
2633         kfree(sb);
2634         kfree(path);
2635         module_put(THIS_MODULE);
2636 async_done:
2637         return size;
2638
2639 out_free_holder:
2640         kfree(holder);
2641 out_put_sb_page:
2642         put_page(virt_to_page(sb_disk));
2643 out_blkdev_put:
2644         if (bdev_handle)
2645                 bdev_release(bdev_handle);
2646 out_free_sb:
2647         kfree(sb);
2648 out_free_path:
2649         kfree(path);
2650         path = NULL;
2651 out_module_put:
2652         module_put(THIS_MODULE);
2653 out:
2654         if (!quiet)
2655                 pr_info("error %s: %s\n", path?path:"", err);
2656         return ret;
2657 }
2658
2659
2660 struct pdev {
2661         struct list_head list;
2662         struct cached_dev *dc;
2663 };
2664
2665 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2666                                          struct kobj_attribute *attr,
2667                                          const char *buffer,
2668                                          size_t size)
2669 {
2670         LIST_HEAD(pending_devs);
2671         ssize_t ret = size;
2672         struct cached_dev *dc, *tdc;
2673         struct pdev *pdev, *tpdev;
2674         struct cache_set *c, *tc;
2675
2676         mutex_lock(&bch_register_lock);
2677         list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2678                 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2679                 if (!pdev)
2680                         break;
2681                 pdev->dc = dc;
2682                 list_add(&pdev->list, &pending_devs);
2683         }
2684
2685         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2686                 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2687                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2688                         char *set_uuid = c->set_uuid;
2689
2690                         if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2691                                 list_del(&pdev->list);
2692                                 kfree(pdev);
2693                                 break;
2694                         }
2695                 }
2696         }
2697         mutex_unlock(&bch_register_lock);
2698
2699         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2700                 pr_info("delete pdev %p\n", pdev);
2701                 list_del(&pdev->list);
2702                 bcache_device_stop(&pdev->dc->disk);
2703                 kfree(pdev);
2704         }
2705
2706         return ret;
2707 }
2708
2709 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2710 {
2711         if (bcache_is_reboot)
2712                 return NOTIFY_DONE;
2713
2714         if (code == SYS_DOWN ||
2715             code == SYS_HALT ||
2716             code == SYS_POWER_OFF) {
2717                 DEFINE_WAIT(wait);
2718                 unsigned long start = jiffies;
2719                 bool stopped = false;
2720
2721                 struct cache_set *c, *tc;
2722                 struct cached_dev *dc, *tdc;
2723
2724                 mutex_lock(&bch_register_lock);
2725
2726                 if (bcache_is_reboot)
2727                         goto out;
2728
2729                 /* New registration is rejected since now */
2730                 bcache_is_reboot = true;
2731                 /*
2732                  * Make registering caller (if there is) on other CPU
2733                  * core know bcache_is_reboot set to true earlier
2734                  */
2735                 smp_mb();
2736
2737                 if (list_empty(&bch_cache_sets) &&
2738                     list_empty(&uncached_devices))
2739                         goto out;
2740
2741                 mutex_unlock(&bch_register_lock);
2742
2743                 pr_info("Stopping all devices:\n");
2744
2745                 /*
2746                  * The reason bch_register_lock is not held to call
2747                  * bch_cache_set_stop() and bcache_device_stop() is to
2748                  * avoid potential deadlock during reboot, because cache
2749                  * set or bcache device stopping process will acquire
2750                  * bch_register_lock too.
2751                  *
2752                  * We are safe here because bcache_is_reboot sets to
2753                  * true already, register_bcache() will reject new
2754                  * registration now. bcache_is_reboot also makes sure
2755                  * bcache_reboot() won't be re-entered on by other thread,
2756                  * so there is no race in following list iteration by
2757                  * list_for_each_entry_safe().
2758                  */
2759                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2760                         bch_cache_set_stop(c);
2761
2762                 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2763                         bcache_device_stop(&dc->disk);
2764
2765
2766                 /*
2767                  * Give an early chance for other kthreads and
2768                  * kworkers to stop themselves
2769                  */
2770                 schedule();
2771
2772                 /* What's a condition variable? */
2773                 while (1) {
2774                         long timeout = start + 10 * HZ - jiffies;
2775
2776                         mutex_lock(&bch_register_lock);
2777                         stopped = list_empty(&bch_cache_sets) &&
2778                                 list_empty(&uncached_devices);
2779
2780                         if (timeout < 0 || stopped)
2781                                 break;
2782
2783                         prepare_to_wait(&unregister_wait, &wait,
2784                                         TASK_UNINTERRUPTIBLE);
2785
2786                         mutex_unlock(&bch_register_lock);
2787                         schedule_timeout(timeout);
2788                 }
2789
2790                 finish_wait(&unregister_wait, &wait);
2791
2792                 if (stopped)
2793                         pr_info("All devices stopped\n");
2794                 else
2795                         pr_notice("Timeout waiting for devices to be closed\n");
2796 out:
2797                 mutex_unlock(&bch_register_lock);
2798         }
2799
2800         return NOTIFY_DONE;
2801 }
2802
2803 static struct notifier_block reboot = {
2804         .notifier_call  = bcache_reboot,
2805         .priority       = INT_MAX, /* before any real devices */
2806 };
2807
2808 static void bcache_exit(void)
2809 {
2810         bch_debug_exit();
2811         bch_request_exit();
2812         if (bcache_kobj)
2813                 kobject_put(bcache_kobj);
2814         if (bcache_wq)
2815                 destroy_workqueue(bcache_wq);
2816         if (bch_journal_wq)
2817                 destroy_workqueue(bch_journal_wq);
2818         if (bch_flush_wq)
2819                 destroy_workqueue(bch_flush_wq);
2820         bch_btree_exit();
2821
2822         if (bcache_major)
2823                 unregister_blkdev(bcache_major, "bcache");
2824         unregister_reboot_notifier(&reboot);
2825         mutex_destroy(&bch_register_lock);
2826 }
2827
2828 /* Check and fixup module parameters */
2829 static void check_module_parameters(void)
2830 {
2831         if (bch_cutoff_writeback_sync == 0)
2832                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2833         else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2834                 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2835                         bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2836                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2837         }
2838
2839         if (bch_cutoff_writeback == 0)
2840                 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2841         else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2842                 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2843                         bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2844                 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2845         }
2846
2847         if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2848                 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2849                         bch_cutoff_writeback, bch_cutoff_writeback_sync);
2850                 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2851         }
2852 }
2853
2854 static int __init bcache_init(void)
2855 {
2856         static const struct attribute *files[] = {
2857                 &ksysfs_register.attr,
2858                 &ksysfs_register_quiet.attr,
2859                 &ksysfs_pendings_cleanup.attr,
2860                 NULL
2861         };
2862
2863         check_module_parameters();
2864
2865         mutex_init(&bch_register_lock);
2866         init_waitqueue_head(&unregister_wait);
2867         register_reboot_notifier(&reboot);
2868
2869         bcache_major = register_blkdev(0, "bcache");
2870         if (bcache_major < 0) {
2871                 unregister_reboot_notifier(&reboot);
2872                 mutex_destroy(&bch_register_lock);
2873                 return bcache_major;
2874         }
2875
2876         if (bch_btree_init())
2877                 goto err;
2878
2879         bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2880         if (!bcache_wq)
2881                 goto err;
2882
2883         /*
2884          * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2885          *
2886          * 1. It used `system_wq` before which also does no memory reclaim.
2887          * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2888          *    reduced throughput can be observed.
2889          *
2890          * We still want to user our own queue to not congest the `system_wq`.
2891          */
2892         bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2893         if (!bch_flush_wq)
2894                 goto err;
2895
2896         bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2897         if (!bch_journal_wq)
2898                 goto err;
2899
2900         bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2901         if (!bcache_kobj)
2902                 goto err;
2903
2904         if (bch_request_init() ||
2905             sysfs_create_files(bcache_kobj, files))
2906                 goto err;
2907
2908         bch_debug_init();
2909
2910         bcache_is_reboot = false;
2911
2912         return 0;
2913 err:
2914         bcache_exit();
2915         return -ENOMEM;
2916 }
2917
2918 /*
2919  * Module hooks
2920  */
2921 module_exit(bcache_exit);
2922 module_init(bcache_init);
2923
2924 module_param(bch_cutoff_writeback, uint, 0);
2925 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2926
2927 module_param(bch_cutoff_writeback_sync, uint, 0);
2928 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2929
2930 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2931 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2932 MODULE_LICENSE("GPL");