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