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