drivers/md/bcache/super.c

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