drivers/md/bcache/super.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * bcache setup/teardown code, and some metadata io - read a superblock and
   4  * figure out what to do with it.
   5  *
   6  * Copyright 2010, 2011 Kent Overstreet <kent.overstreet@gmail.com>
   7  * Copyright 2012 Google, Inc.
   8  */
   9
  10 #include "bcache.h"
  11 #include "btree.h"
  12 #include "debug.h"
  13 #include "extents.h"
  14 #include "request.h"
  15 #include "writeback.h"
  16
  17 #include <linux/blkdev.h>
  18 #include <linux/debugfs.h>
  19 #include <linux/genhd.h>
  20 #include <linux/idr.h>
  21 #include <linux/kthread.h>
  22 #include <linux/workqueue.h>
  23 #include <linux/module.h>
  24 #include <linux/random.h>
  25 #include <linux/reboot.h>
  26 #include <linux/sysfs.h>
  27
  28 unsigned int bch_cutoff_writeback;
  29 unsigned int bch_cutoff_writeback_sync;
  30
  31 static const char bcache_magic[] = {
  32         0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
  33         0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
  34 };
  35
  36 static const char invalid_uuid[] = {
  37         0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
  38         0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
  39 };
  40
  41 static struct kobject *bcache_kobj;
  42 struct mutex bch_register_lock;
  43 bool bcache_is_reboot;
  44 LIST_HEAD(bch_cache_sets);
  45 static LIST_HEAD(uncached_devices);
  46
  47 static int bcache_major;
  48 static DEFINE_IDA(bcache_device_idx);
  49 static wait_queue_head_t unregister_wait;
  50 struct workqueue_struct *bcache_wq;
  51 struct workqueue_struct *bch_journal_wq;
  52
  53
  54 #define BTREE_MAX_PAGES         (256 * 1024 / PAGE_SIZE)
  55 /* limitation of partitions number on single bcache device */
  56 #define BCACHE_MINORS           128
  57 /* limitation of bcache devices number on single system */
  58 #define BCACHE_DEVICE_IDX_MAX   ((1U << MINORBITS)/BCACHE_MINORS)
  59
  60 /* Superblock */
  61
  62 static const char *read_super_common(struct cache_sb *sb,  struct block_device *bdev,
  63                                      struct cache_sb_disk *s)
  64 {
  65         const char *err;
  66         unsigned int i;
  67
  68         sb->nbuckets    = le64_to_cpu(s->nbuckets);
  69         sb->bucket_size = le16_to_cpu(s->bucket_size);
  70
  71         sb->nr_in_set   = le16_to_cpu(s->nr_in_set);
  72         sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
  73
  74         err = "Too many buckets";
  75         if (sb->nbuckets > LONG_MAX)
  76                 goto err;
  77
  78         err = "Not enough buckets";
  79         if (sb->nbuckets < 1 << 7)
  80                 goto err;
  81
  82         err = "Bad block size (not power of 2)";
  83         if (!is_power_of_2(sb->block_size))
  84                 goto err;
  85
  86         err = "Bad block size (larger than page size)";
  87         if (sb->block_size > PAGE_SECTORS)
  88                 goto err;
  89
  90         err = "Bad bucket size (not power of 2)";
  91         if (!is_power_of_2(sb->bucket_size))
  92                 goto err;
  93
  94         err = "Bad bucket size (smaller than page size)";
  95         if (sb->bucket_size < PAGE_SECTORS)
  96                 goto err;
  97
  98         err = "Invalid superblock: device too small";
  99         if (get_capacity(bdev->bd_disk) <
 100             sb->bucket_size * sb->nbuckets)
 101                 goto err;
 102
 103         err = "Bad UUID";
 104         if (bch_is_zero(sb->set_uuid, 16))
 105                 goto err;
 106
 107         err = "Bad cache device number in set";
 108         if (!sb->nr_in_set ||
 109             sb->nr_in_set <= sb->nr_this_dev ||
 110             sb->nr_in_set > MAX_CACHES_PER_SET)
 111                 goto err;
 112
 113         err = "Journal buckets not sequential";
 114         for (i = 0; i < sb->keys; i++)
 115                 if (sb->d[i] != sb->first_bucket + i)
 116                         goto err;
 117
 118         err = "Too many journal buckets";
 119         if (sb->first_bucket + sb->keys > sb->nbuckets)
 120                 goto err;
 121
 122         err = "Invalid superblock: first bucket comes before end of super";
 123         if (sb->first_bucket * sb->bucket_size < 16)
 124                 goto err;
 125
 126         err = NULL;
 127 err:
 128         return err;
 129 }
 130
 131
 132 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
 133                               struct cache_sb_disk **res)
 134 {
 135         const char *err;
 136         struct cache_sb_disk *s;
 137         struct page *page;
 138         unsigned int i;
 139
 140         page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
 141                                    SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
 142         if (IS_ERR(page))
 143                 return "IO error";
 144         s = page_address(page) + offset_in_page(SB_OFFSET);
 145
 146         sb->offset              = le64_to_cpu(s->offset);
 147         sb->version             = le64_to_cpu(s->version);
 148
 149         memcpy(sb->magic,       s->magic, 16);
 150         memcpy(sb->uuid,        s->uuid, 16);
 151         memcpy(sb->set_uuid,    s->set_uuid, 16);
 152         memcpy(sb->label,       s->label, SB_LABEL_SIZE);
 153
 154         sb->flags               = le64_to_cpu(s->flags);
 155         sb->seq                 = le64_to_cpu(s->seq);
 156         sb->last_mount          = le32_to_cpu(s->last_mount);
 157         sb->first_bucket        = le16_to_cpu(s->first_bucket);
 158         sb->keys                = le16_to_cpu(s->keys);
 159
 160         for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
 161                 sb->d[i] = le64_to_cpu(s->d[i]);
 162
 163         pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
 164                  sb->version, sb->flags, sb->seq, sb->keys);
 165
 166         err = "Not a bcache superblock (bad offset)";
 167         if (sb->offset != SB_SECTOR)
 168                 goto err;
 169
 170         err = "Not a bcache superblock (bad magic)";
 171         if (memcmp(sb->magic, bcache_magic, 16))
 172                 goto err;
 173
 174         err = "Too many journal buckets";
 175         if (sb->keys > SB_JOURNAL_BUCKETS)
 176                 goto err;
 177
 178         err = "Bad checksum";
 179         if (s->csum != csum_set(s))
 180                 goto err;
 181
 182         err = "Bad UUID";
 183         if (bch_is_zero(sb->uuid, 16))
 184                 goto err;
 185
 186         sb->block_size  = le16_to_cpu(s->block_size);
 187
 188         err = "Superblock block size smaller than device block size";
 189         if (sb->block_size << 9 < bdev_logical_block_size(bdev))
 190                 goto err;
 191
 192         switch (sb->version) {
 193         case BCACHE_SB_VERSION_BDEV:
 194                 sb->data_offset = BDEV_DATA_START_DEFAULT;
 195                 break;
 196         case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
 197                 sb->data_offset = le64_to_cpu(s->data_offset);
 198
 199                 err = "Bad data offset";
 200                 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
 201                         goto err;
 202
 203                 break;
 204         case BCACHE_SB_VERSION_CDEV:
 205         case BCACHE_SB_VERSION_CDEV_WITH_UUID:
 206                 err = read_super_common(sb, bdev, s);
 207                 if (err)
 208                         goto err;
 209                 break;
 210         default:
 211                 err = "Unsupported superblock version";
 212                 goto err;
 213         }
 214
 215         sb->last_mount = (u32)ktime_get_real_seconds();
 216         *res = s;
 217         return NULL;
 218 err:
 219         put_page(page);
 220         return err;
 221 }
 222
 223 static void write_bdev_super_endio(struct bio *bio)
 224 {
 225         struct cached_dev *dc = bio->bi_private;
 226
 227         if (bio->bi_status)
 228                 bch_count_backing_io_errors(dc, bio);
 229
 230         closure_put(&dc->sb_write);
 231 }
 232
 233 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
 234                 struct bio *bio)
 235 {
 236         unsigned int i;
 237
 238         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
 239         bio->bi_iter.bi_sector  = SB_SECTOR;
 240         __bio_add_page(bio, virt_to_page(out), SB_SIZE,
 241                         offset_in_page(out));
 242
 243         out->offset             = cpu_to_le64(sb->offset);
 244         out->version            = cpu_to_le64(sb->version);
 245
 246         memcpy(out->uuid,       sb->uuid, 16);
 247         memcpy(out->set_uuid,   sb->set_uuid, 16);
 248         memcpy(out->label,      sb->label, SB_LABEL_SIZE);
 249
 250         out->flags              = cpu_to_le64(sb->flags);
 251         out->seq                = cpu_to_le64(sb->seq);
 252
 253         out->last_mount         = cpu_to_le32(sb->last_mount);
 254         out->first_bucket       = cpu_to_le16(sb->first_bucket);
 255         out->keys               = cpu_to_le16(sb->keys);
 256
 257         for (i = 0; i < sb->keys; i++)
 258                 out->d[i] = cpu_to_le64(sb->d[i]);
 259
 260         out->csum = csum_set(out);
 261
 262         pr_debug("ver %llu, flags %llu, seq %llu\n",
 263                  sb->version, sb->flags, sb->seq);
 264
 265         submit_bio(bio);
 266 }
 267
 268 static void bch_write_bdev_super_unlock(struct closure *cl)
 269 {
 270         struct cached_dev *dc = container_of(cl, struct cached_dev, sb_write);
 271
 272         up(&dc->sb_write_mutex);
 273 }
 274
 275 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
 276 {
 277         struct closure *cl = &dc->sb_write;
 278         struct bio *bio = &dc->sb_bio;
 279
 280         down(&dc->sb_write_mutex);
 281         closure_init(cl, parent);
 282
 283         bio_init(bio, dc->sb_bv, 1);
 284         bio_set_dev(bio, dc->bdev);
 285         bio->bi_end_io  = write_bdev_super_endio;
 286         bio->bi_private = dc;
 287
 288         closure_get(cl);
 289         /* I/O request sent to backing device */
 290         __write_super(&dc->sb, dc->sb_disk, bio);
 291
 292         closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
 293 }
 294
 295 static void write_super_endio(struct bio *bio)
 296 {
 297         struct cache *ca = bio->bi_private;
 298
 299         /* is_read = 0 */
 300         bch_count_io_errors(ca, bio->bi_status, 0,
 301                             "writing superblock");
 302         closure_put(&ca->set->sb_write);
 303 }
 304
 305 static void bcache_write_super_unlock(struct closure *cl)
 306 {
 307         struct cache_set *c = container_of(cl, struct cache_set, sb_write);
 308
 309         up(&c->sb_write_mutex);
 310 }
 311
 312 void bcache_write_super(struct cache_set *c)
 313 {
 314         struct closure *cl = &c->sb_write;
 315         struct cache *ca;
 316         unsigned int i;
 317
 318         down(&c->sb_write_mutex);
 319         closure_init(cl, &c->cl);
 320
 321         c->sb.seq++;
 322
 323         for_each_cache(ca, c, i) {
 324                 struct bio *bio = &ca->sb_bio;
 325
 326                 ca->sb.version          = BCACHE_SB_VERSION_CDEV_WITH_UUID;
 327                 ca->sb.seq              = c->sb.seq;
 328                 ca->sb.last_mount       = c->sb.last_mount;
 329
 330                 SET_CACHE_SYNC(&ca->sb, CACHE_SYNC(&c->sb));
 331
 332                 bio_init(bio, ca->sb_bv, 1);
 333                 bio_set_dev(bio, ca->bdev);
 334                 bio->bi_end_io  = write_super_endio;
 335                 bio->bi_private = ca;
 336
 337                 closure_get(cl);
 338                 __write_super(&ca->sb, ca->sb_disk, bio);
 339         }
 340
 341         closure_return_with_destructor(cl, bcache_write_super_unlock);
 342 }
 343
 344 /* UUID io */
 345
 346 static void uuid_endio(struct bio *bio)
 347 {
 348         struct closure *cl = bio->bi_private;
 349         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
 350
 351         cache_set_err_on(bio->bi_status, c, "accessing uuids");
 352         bch_bbio_free(bio, c);
 353         closure_put(cl);
 354 }
 355
 356 static void uuid_io_unlock(struct closure *cl)
 357 {
 358         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
 359
 360         up(&c->uuid_write_mutex);
 361 }
 362
 363 static void uuid_io(struct cache_set *c, int op, unsigned long op_flags,
 364                     struct bkey *k, struct closure *parent)
 365 {
 366         struct closure *cl = &c->uuid_write;
 367         struct uuid_entry *u;
 368         unsigned int i;
 369         char buf[80];
 370
 371         BUG_ON(!parent);
 372         down(&c->uuid_write_mutex);
 373         closure_init(cl, parent);
 374
 375         for (i = 0; i < KEY_PTRS(k); i++) {
 376                 struct bio *bio = bch_bbio_alloc(c);
 377
 378                 bio->bi_opf = REQ_SYNC | REQ_META | op_flags;
 379                 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
 380
 381                 bio->bi_end_io  = uuid_endio;
 382                 bio->bi_private = cl;
 383                 bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
 384                 bch_bio_map(bio, c->uuids);
 385
 386                 bch_submit_bbio(bio, c, k, i);
 387
 388                 if (op != REQ_OP_WRITE)
 389                         break;
 390         }
 391
 392         bch_extent_to_text(buf, sizeof(buf), k);
 393         pr_debug("%s UUIDs at %s\n", op == REQ_OP_WRITE ? "wrote" : "read", buf);
 394
 395         for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
 396                 if (!bch_is_zero(u->uuid, 16))
 397                         pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
 398                                  u - c->uuids, u->uuid, u->label,
 399                                  u->first_reg, u->last_reg, u->invalidated);
 400
 401         closure_return_with_destructor(cl, uuid_io_unlock);
 402 }
 403
 404 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
 405 {
 406         struct bkey *k = &j->uuid_bucket;
 407
 408         if (__bch_btree_ptr_invalid(c, k))
 409                 return "bad uuid pointer";
 410
 411         bkey_copy(&c->uuid_bucket, k);
 412         uuid_io(c, REQ_OP_READ, 0, k, cl);
 413
 414         if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
 415                 struct uuid_entry_v0    *u0 = (void *) c->uuids;
 416                 struct uuid_entry       *u1 = (void *) c->uuids;
 417                 int i;
 418
 419                 closure_sync(cl);
 420
 421                 /*
 422                  * Since the new uuid entry is bigger than the old, we have to
 423                  * convert starting at the highest memory address and work down
 424                  * in order to do it in place
 425                  */
 426
 427                 for (i = c->nr_uuids - 1;
 428                      i >= 0;
 429                      --i) {
 430                         memcpy(u1[i].uuid,      u0[i].uuid, 16);
 431                         memcpy(u1[i].label,     u0[i].label, 32);
 432
 433                         u1[i].first_reg         = u0[i].first_reg;
 434                         u1[i].last_reg          = u0[i].last_reg;
 435                         u1[i].invalidated       = u0[i].invalidated;
 436
 437                         u1[i].flags     = 0;
 438                         u1[i].sectors   = 0;
 439                 }
 440         }
 441
 442         return NULL;
 443 }
 444
 445 static int __uuid_write(struct cache_set *c)
 446 {
 447         BKEY_PADDED(key) k;
 448         struct closure cl;
 449         struct cache *ca;
 450
 451         closure_init_stack(&cl);
 452         lockdep_assert_held(&bch_register_lock);
 453
 454         if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, 1, true))
 455                 return 1;
 456
 457         SET_KEY_SIZE(&k.key, c->sb.bucket_size);
 458         uuid_io(c, REQ_OP_WRITE, 0, &k.key, &cl);
 459         closure_sync(&cl);
 460
 461         /* Only one bucket used for uuid write */
 462         ca = PTR_CACHE(c, &k.key, 0);
 463         atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
 464
 465         bkey_copy(&c->uuid_bucket, &k.key);
 466         bkey_put(c, &k.key);
 467         return 0;
 468 }
 469
 470 int bch_uuid_write(struct cache_set *c)
 471 {
 472         int ret = __uuid_write(c);
 473
 474         if (!ret)
 475                 bch_journal_meta(c, NULL);
 476
 477         return ret;
 478 }
 479
 480 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
 481 {
 482         struct uuid_entry *u;
 483
 484         for (u = c->uuids;
 485              u < c->uuids + c->nr_uuids; u++)
 486                 if (!memcmp(u->uuid, uuid, 16))
 487                         return u;
 488
 489         return NULL;
 490 }
 491
 492 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
 493 {
 494         static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
 495
 496         return uuid_find(c, zero_uuid);
 497 }
 498
 499 /*
 500  * Bucket priorities/gens:
 501  *
 502  * For each bucket, we store on disk its
 503  *   8 bit gen
 504  *  16 bit priority
 505  *
 506  * See alloc.c for an explanation of the gen. The priority is used to implement
 507  * lru (and in the future other) cache replacement policies; for most purposes
 508  * it's just an opaque integer.
 509  *
 510  * The gens and the priorities don't have a whole lot to do with each other, and
 511  * it's actually the gens that must be written out at specific times - it's no
 512  * big deal if the priorities don't get written, if we lose them we just reuse
 513  * buckets in suboptimal order.
 514  *
 515  * On disk they're stored in a packed array, and in as many buckets are required
 516  * to fit them all. The buckets we use to store them form a list; the journal
 517  * header points to the first bucket, the first bucket points to the second
 518  * bucket, et cetera.
 519  *
 520  * This code is used by the allocation code; periodically (whenever it runs out
 521  * of buckets to allocate from) the allocation code will invalidate some
 522  * buckets, but it can't use those buckets until their new gens are safely on
 523  * disk.
 524  */
 525
 526 static void prio_endio(struct bio *bio)
 527 {
 528         struct cache *ca = bio->bi_private;
 529
 530         cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
 531         bch_bbio_free(bio, ca->set);
 532         closure_put(&ca->prio);
 533 }
 534
 535 static void prio_io(struct cache *ca, uint64_t bucket, int op,
 536                     unsigned long op_flags)
 537 {
 538         struct closure *cl = &ca->prio;
 539         struct bio *bio = bch_bbio_alloc(ca->set);
 540
 541         closure_init_stack(cl);
 542
 543         bio->bi_iter.bi_sector  = bucket * ca->sb.bucket_size;
 544         bio_set_dev(bio, ca->bdev);
 545         bio->bi_iter.bi_size    = bucket_bytes(ca);
 546
 547         bio->bi_end_io  = prio_endio;
 548         bio->bi_private = ca;
 549         bio_set_op_attrs(bio, op, REQ_SYNC|REQ_META|op_flags);
 550         bch_bio_map(bio, ca->disk_buckets);
 551
 552         closure_bio_submit(ca->set, bio, &ca->prio);
 553         closure_sync(cl);
 554 }
 555
 556 int bch_prio_write(struct cache *ca, bool wait)
 557 {
 558         int i;
 559         struct bucket *b;
 560         struct closure cl;
 561
 562         pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
 563                  fifo_used(&ca->free[RESERVE_PRIO]),
 564                  fifo_used(&ca->free[RESERVE_NONE]),
 565                  fifo_used(&ca->free_inc));
 566
 567         /*
 568          * Pre-check if there are enough free buckets. In the non-blocking
 569          * scenario it's better to fail early rather than starting to allocate
 570          * buckets and do a cleanup later in case of failure.
 571          */
 572         if (!wait) {
 573                 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
 574                                fifo_used(&ca->free[RESERVE_NONE]);
 575                 if (prio_buckets(ca) > avail)
 576                         return -ENOMEM;
 577         }
 578
 579         closure_init_stack(&cl);
 580
 581         lockdep_assert_held(&ca->set->bucket_lock);
 582
 583         ca->disk_buckets->seq++;
 584
 585         atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
 586                         &ca->meta_sectors_written);
 587
 588         for (i = prio_buckets(ca) - 1; i >= 0; --i) {
 589                 long bucket;
 590                 struct prio_set *p = ca->disk_buckets;
 591                 struct bucket_disk *d = p->data;
 592                 struct bucket_disk *end = d + prios_per_bucket(ca);
 593
 594                 for (b = ca->buckets + i * prios_per_bucket(ca);
 595                      b < ca->buckets + ca->sb.nbuckets && d < end;
 596                      b++, d++) {
 597                         d->prio = cpu_to_le16(b->prio);
 598                         d->gen = b->gen;
 599                 }
 600
 601                 p->next_bucket  = ca->prio_buckets[i + 1];
 602                 p->magic        = pset_magic(&ca->sb);
 603                 p->csum         = bch_crc64(&p->magic, bucket_bytes(ca) - 8);
 604
 605                 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
 606                 BUG_ON(bucket == -1);
 607
 608                 mutex_unlock(&ca->set->bucket_lock);
 609                 prio_io(ca, bucket, REQ_OP_WRITE, 0);
 610                 mutex_lock(&ca->set->bucket_lock);
 611
 612                 ca->prio_buckets[i] = bucket;
 613                 atomic_dec_bug(&ca->buckets[bucket].pin);
 614         }
 615
 616         mutex_unlock(&ca->set->bucket_lock);
 617
 618         bch_journal_meta(ca->set, &cl);
 619         closure_sync(&cl);
 620
 621         mutex_lock(&ca->set->bucket_lock);
 622
 623         /*
 624          * Don't want the old priorities to get garbage collected until after we
 625          * finish writing the new ones, and they're journalled
 626          */
 627         for (i = 0; i < prio_buckets(ca); i++) {
 628                 if (ca->prio_last_buckets[i])
 629                         __bch_bucket_free(ca,
 630                                 &ca->buckets[ca->prio_last_buckets[i]]);
 631
 632                 ca->prio_last_buckets[i] = ca->prio_buckets[i];
 633         }
 634         return 0;
 635 }
 636
 637 static int prio_read(struct cache *ca, uint64_t bucket)
 638 {
 639         struct prio_set *p = ca->disk_buckets;
 640         struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
 641         struct bucket *b;
 642         unsigned int bucket_nr = 0;
 643         int ret = -EIO;
 644
 645         for (b = ca->buckets;
 646              b < ca->buckets + ca->sb.nbuckets;
 647              b++, d++) {
 648                 if (d == end) {
 649                         ca->prio_buckets[bucket_nr] = bucket;
 650                         ca->prio_last_buckets[bucket_nr] = bucket;
 651                         bucket_nr++;
 652
 653                         prio_io(ca, bucket, REQ_OP_READ, 0);
 654
 655                         if (p->csum !=
 656                             bch_crc64(&p->magic, bucket_bytes(ca) - 8)) {
 657                                 pr_warn("bad csum reading priorities\n");
 658                                 goto out;
 659                         }
 660
 661                         if (p->magic != pset_magic(&ca->sb)) {
 662                                 pr_warn("bad magic reading priorities\n");
 663                                 goto out;
 664                         }
 665
 666                         bucket = p->next_bucket;
 667                         d = p->data;
 668                 }
 669
 670                 b->prio = le16_to_cpu(d->prio);
 671                 b->gen = b->last_gc = d->gen;
 672         }
 673
 674         ret = 0;
 675 out:
 676         return ret;
 677 }
 678
 679 /* Bcache device */
 680
 681 static int open_dev(struct block_device *b, fmode_t mode)
 682 {
 683         struct bcache_device *d = b->bd_disk->private_data;
 684
 685         if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
 686                 return -ENXIO;
 687
 688         closure_get(&d->cl);
 689         return 0;
 690 }
 691
 692 static void release_dev(struct gendisk *b, fmode_t mode)
 693 {
 694         struct bcache_device *d = b->private_data;
 695
 696         closure_put(&d->cl);
 697 }
 698
 699 static int ioctl_dev(struct block_device *b, fmode_t mode,
 700                      unsigned int cmd, unsigned long arg)
 701 {
 702         struct bcache_device *d = b->bd_disk->private_data;
 703
 704         return d->ioctl(d, mode, cmd, arg);
 705 }
 706
 707 static const struct block_device_operations bcache_cached_ops = {
 708         .submit_bio     = cached_dev_submit_bio,
 709         .open           = open_dev,
 710         .release        = release_dev,
 711         .ioctl          = ioctl_dev,
 712         .owner          = THIS_MODULE,
 713 };
 714
 715 static const struct block_device_operations bcache_flash_ops = {
 716         .submit_bio     = flash_dev_submit_bio,
 717         .open           = open_dev,
 718         .release        = release_dev,
 719         .ioctl          = ioctl_dev,
 720         .owner          = THIS_MODULE,
 721 };
 722
 723 void bcache_device_stop(struct bcache_device *d)
 724 {
 725         if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
 726                 /*
 727                  * closure_fn set to
 728                  * - cached device: cached_dev_flush()
 729                  * - flash dev: flash_dev_flush()
 730                  */
 731                 closure_queue(&d->cl);
 732 }
 733
 734 static void bcache_device_unlink(struct bcache_device *d)
 735 {
 736         lockdep_assert_held(&bch_register_lock);
 737
 738         if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
 739                 unsigned int i;
 740                 struct cache *ca;
 741
 742                 sysfs_remove_link(&d->c->kobj, d->name);
 743                 sysfs_remove_link(&d->kobj, "cache");
 744
 745                 for_each_cache(ca, d->c, i)
 746                         bd_unlink_disk_holder(ca->bdev, d->disk);
 747         }
 748 }
 749
 750 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
 751                                const char *name)
 752 {
 753         unsigned int i;
 754         struct cache *ca;
 755         int ret;
 756
 757         for_each_cache(ca, d->c, i)
 758                 bd_link_disk_holder(ca->bdev, d->disk);
 759
 760         snprintf(d->name, BCACHEDEVNAME_SIZE,
 761                  "%s%u", name, d->id);
 762
 763         ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
 764         if (ret < 0)
 765                 pr_err("Couldn't create device -> cache set symlink\n");
 766
 767         ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
 768         if (ret < 0)
 769                 pr_err("Couldn't create cache set -> device symlink\n");
 770
 771         clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
 772 }
 773
 774 static void bcache_device_detach(struct bcache_device *d)
 775 {
 776         lockdep_assert_held(&bch_register_lock);
 777
 778         atomic_dec(&d->c->attached_dev_nr);
 779
 780         if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
 781                 struct uuid_entry *u = d->c->uuids + d->id;
 782
 783                 SET_UUID_FLASH_ONLY(u, 0);
 784                 memcpy(u->uuid, invalid_uuid, 16);
 785                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
 786                 bch_uuid_write(d->c);
 787         }
 788
 789         bcache_device_unlink(d);
 790
 791         d->c->devices[d->id] = NULL;
 792         closure_put(&d->c->caching);
 793         d->c = NULL;
 794 }
 795
 796 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
 797                                  unsigned int id)
 798 {
 799         d->id = id;
 800         d->c = c;
 801         c->devices[id] = d;
 802
 803         if (id >= c->devices_max_used)
 804                 c->devices_max_used = id + 1;
 805
 806         closure_get(&c->caching);
 807 }
 808
 809 static inline int first_minor_to_idx(int first_minor)
 810 {
 811         return (first_minor/BCACHE_MINORS);
 812 }
 813
 814 static inline int idx_to_first_minor(int idx)
 815 {
 816         return (idx * BCACHE_MINORS);
 817 }
 818
 819 static void bcache_device_free(struct bcache_device *d)
 820 {
 821         struct gendisk *disk = d->disk;
 822
 823         lockdep_assert_held(&bch_register_lock);
 824
 825         if (disk)
 826                 pr_info("%s stopped\n", disk->disk_name);
 827         else
 828                 pr_err("bcache device (NULL gendisk) stopped\n");
 829
 830         if (d->c)
 831                 bcache_device_detach(d);
 832
 833         if (disk) {
 834                 bool disk_added = (disk->flags & GENHD_FL_UP) != 0;
 835
 836                 if (disk_added)
 837                         del_gendisk(disk);
 838
 839                 if (disk->queue)
 840                         blk_cleanup_queue(disk->queue);
 841
 842                 ida_simple_remove(&bcache_device_idx,
 843                                   first_minor_to_idx(disk->first_minor));
 844                 if (disk_added)
 845                         put_disk(disk);
 846         }
 847
 848         bioset_exit(&d->bio_split);
 849         kvfree(d->full_dirty_stripes);
 850         kvfree(d->stripe_sectors_dirty);
 851
 852         closure_debug_destroy(&d->cl);
 853 }
 854
 855 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
 856                 sector_t sectors, struct block_device *cached_bdev,
 857                 const struct block_device_operations *ops)
 858 {
 859         struct request_queue *q;
 860         const size_t max_stripes = min_t(size_t, INT_MAX,
 861                                          SIZE_MAX / sizeof(atomic_t));
 862         uint64_t n;
 863         int idx;
 864
 865         if (!d->stripe_size)
 866                 d->stripe_size = 1 << 31;
 867
 868         n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
 869         if (!n || n > max_stripes) {
 870                 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
 871                         n);
 872                 return -ENOMEM;
 873         }
 874         d->nr_stripes = n;
 875
 876         n = d->nr_stripes * sizeof(atomic_t);
 877         d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
 878         if (!d->stripe_sectors_dirty)
 879                 return -ENOMEM;
 880
 881         n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
 882         d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
 883         if (!d->full_dirty_stripes)
 884                 return -ENOMEM;
 885
 886         idx = ida_simple_get(&bcache_device_idx, 0,
 887                                 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
 888         if (idx < 0)
 889                 return idx;
 890
 891         if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
 892                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
 893                 goto err;
 894
 895         d->disk = alloc_disk(BCACHE_MINORS);
 896         if (!d->disk)
 897                 goto err;
 898
 899         set_capacity(d->disk, sectors);
 900         snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
 901
 902         d->disk->major          = bcache_major;
 903         d->disk->first_minor    = idx_to_first_minor(idx);
 904         d->disk->fops           = ops;
 905         d->disk->private_data   = d;
 906
 907         q = blk_alloc_queue(NUMA_NO_NODE);
 908         if (!q)
 909                 return -ENOMEM;
 910
 911         d->disk->queue                  = q;
 912         q->backing_dev_info->congested_data = d;
 913         q->limits.max_hw_sectors        = UINT_MAX;
 914         q->limits.max_sectors           = UINT_MAX;
 915         q->limits.max_segment_size      = UINT_MAX;
 916         q->limits.max_segments          = BIO_MAX_PAGES;
 917         blk_queue_max_discard_sectors(q, UINT_MAX);
 918         q->limits.discard_granularity   = 512;
 919         q->limits.io_min                = block_size;
 920         q->limits.logical_block_size    = block_size;
 921         q->limits.physical_block_size   = block_size;
 922
 923         if (q->limits.logical_block_size > PAGE_SIZE && cached_bdev) {
 924                 /*
 925                  * This should only happen with BCACHE_SB_VERSION_BDEV.
 926                  * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
 927                  */
 928                 pr_info("%s: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
 929                         d->disk->disk_name, q->limits.logical_block_size,
 930                         PAGE_SIZE, bdev_logical_block_size(cached_bdev));
 931
 932                 /* This also adjusts physical block size/min io size if needed */
 933                 blk_queue_logical_block_size(q, bdev_logical_block_size(cached_bdev));
 934         }
 935
 936         blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
 937         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
 938         blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
 939
 940         blk_queue_write_cache(q, true, true);
 941
 942         return 0;
 943
 944 err:
 945         ida_simple_remove(&bcache_device_idx, idx);
 946         return -ENOMEM;
 947
 948 }
 949
 950 /* Cached device */
 951
 952 static void calc_cached_dev_sectors(struct cache_set *c)
 953 {
 954         uint64_t sectors = 0;
 955         struct cached_dev *dc;
 956
 957         list_for_each_entry(dc, &c->cached_devs, list)
 958                 sectors += bdev_sectors(dc->bdev);
 959
 960         c->cached_dev_sectors = sectors;
 961 }
 962
 963 #define BACKING_DEV_OFFLINE_TIMEOUT 5
 964 static int cached_dev_status_update(void *arg)
 965 {
 966         struct cached_dev *dc = arg;
 967         struct request_queue *q;
 968
 969         /*
 970          * If this delayed worker is stopping outside, directly quit here.
 971          * dc->io_disable might be set via sysfs interface, so check it
 972          * here too.
 973          */
 974         while (!kthread_should_stop() && !dc->io_disable) {
 975                 q = bdev_get_queue(dc->bdev);
 976                 if (blk_queue_dying(q))
 977                         dc->offline_seconds++;
 978                 else
 979                         dc->offline_seconds = 0;
 980
 981                 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
 982                         pr_err("%s: device offline for %d seconds\n",
 983                                dc->backing_dev_name,
 984                                BACKING_DEV_OFFLINE_TIMEOUT);
 985                         pr_err("%s: disable I/O request due to backing device offline\n",
 986                                dc->disk.name);
 987                         dc->io_disable = true;
 988                         /* let others know earlier that io_disable is true */
 989                         smp_mb();
 990                         bcache_device_stop(&dc->disk);
 991                         break;
 992                 }
 993                 schedule_timeout_interruptible(HZ);
 994         }
 995
 996         wait_for_kthread_stop();
 997         return 0;
 998 }
 999
1000
1001 int bch_cached_dev_run(struct cached_dev *dc)
1002 {
1003         struct bcache_device *d = &dc->disk;
1004         char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1005         char *env[] = {
1006                 "DRIVER=bcache",
1007                 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1008                 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1009                 NULL,
1010         };
1011
1012         if (dc->io_disable) {
1013                 pr_err("I/O disabled on cached dev %s\n",
1014                        dc->backing_dev_name);
1015                 kfree(env[1]);
1016                 kfree(env[2]);
1017                 kfree(buf);
1018                 return -EIO;
1019         }
1020
1021         if (atomic_xchg(&dc->running, 1)) {
1022                 kfree(env[1]);
1023                 kfree(env[2]);
1024                 kfree(buf);
1025                 pr_info("cached dev %s is running already\n",
1026                        dc->backing_dev_name);
1027                 return -EBUSY;
1028         }
1029
1030         if (!d->c &&
1031             BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1032                 struct closure cl;
1033
1034                 closure_init_stack(&cl);
1035
1036                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1037                 bch_write_bdev_super(dc, &cl);
1038                 closure_sync(&cl);
1039         }
1040
1041         add_disk(d->disk);
1042         bd_link_disk_holder(dc->bdev, dc->disk.disk);
1043         /*
1044          * won't show up in the uevent file, use udevadm monitor -e instead
1045          * only class / kset properties are persistent
1046          */
1047         kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1048         kfree(env[1]);
1049         kfree(env[2]);
1050         kfree(buf);
1051
1052         if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1053             sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1054                               &d->kobj, "bcache")) {
1055                 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1056                 return -ENOMEM;
1057         }
1058
1059         dc->status_update_thread = kthread_run(cached_dev_status_update,
1060                                                dc, "bcache_status_update");
1061         if (IS_ERR(dc->status_update_thread)) {
1062                 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1063         }
1064
1065         return 0;
1066 }
1067
1068 /*
1069  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1070  * work dc->writeback_rate_update is running. Wait until the routine
1071  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1072  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1073  * seconds, give up waiting here and continue to cancel it too.
1074  */
1075 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1076 {
1077         int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1078
1079         do {
1080                 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1081                               &dc->disk.flags))
1082                         break;
1083                 time_out--;
1084                 schedule_timeout_interruptible(1);
1085         } while (time_out > 0);
1086
1087         if (time_out == 0)
1088                 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1089
1090         cancel_delayed_work_sync(&dc->writeback_rate_update);
1091 }
1092
1093 static void cached_dev_detach_finish(struct work_struct *w)
1094 {
1095         struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1096         struct closure cl;
1097
1098         closure_init_stack(&cl);
1099
1100         BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1101         BUG_ON(refcount_read(&dc->count));
1102
1103
1104         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1105                 cancel_writeback_rate_update_dwork(dc);
1106
1107         if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1108                 kthread_stop(dc->writeback_thread);
1109                 dc->writeback_thread = NULL;
1110         }
1111
1112         memset(&dc->sb.set_uuid, 0, 16);
1113         SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1114
1115         bch_write_bdev_super(dc, &cl);
1116         closure_sync(&cl);
1117
1118         mutex_lock(&bch_register_lock);
1119
1120         calc_cached_dev_sectors(dc->disk.c);
1121         bcache_device_detach(&dc->disk);
1122         list_move(&dc->list, &uncached_devices);
1123
1124         clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1125         clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1126
1127         mutex_unlock(&bch_register_lock);
1128
1129         pr_info("Caching disabled for %s\n", dc->backing_dev_name);
1130
1131         /* Drop ref we took in cached_dev_detach() */
1132         closure_put(&dc->disk.cl);
1133 }
1134
1135 void bch_cached_dev_detach(struct cached_dev *dc)
1136 {
1137         lockdep_assert_held(&bch_register_lock);
1138
1139         if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1140                 return;
1141
1142         if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1143                 return;
1144
1145         /*
1146          * Block the device from being closed and freed until we're finished
1147          * detaching
1148          */
1149         closure_get(&dc->disk.cl);
1150
1151         bch_writeback_queue(dc);
1152
1153         cached_dev_put(dc);
1154 }
1155
1156 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1157                           uint8_t *set_uuid)
1158 {
1159         uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1160         struct uuid_entry *u;
1161         struct cached_dev *exist_dc, *t;
1162         int ret = 0;
1163
1164         if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1165             (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1166                 return -ENOENT;
1167
1168         if (dc->disk.c) {
1169                 pr_err("Can't attach %s: already attached\n",
1170                        dc->backing_dev_name);
1171                 return -EINVAL;
1172         }
1173
1174         if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1175                 pr_err("Can't attach %s: shutting down\n",
1176                        dc->backing_dev_name);
1177                 return -EINVAL;
1178         }
1179
1180         if (dc->sb.block_size < c->sb.block_size) {
1181                 /* Will die */
1182                 pr_err("Couldn't attach %s: block size less than set's block size\n",
1183                        dc->backing_dev_name);
1184                 return -EINVAL;
1185         }
1186
1187         /* Check whether already attached */
1188         list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1189                 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1190                         pr_err("Tried to attach %s but duplicate UUID already attached\n",
1191                                 dc->backing_dev_name);
1192
1193                         return -EINVAL;
1194                 }
1195         }
1196
1197         u = uuid_find(c, dc->sb.uuid);
1198
1199         if (u &&
1200             (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1201              BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1202                 memcpy(u->uuid, invalid_uuid, 16);
1203                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1204                 u = NULL;
1205         }
1206
1207         if (!u) {
1208                 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1209                         pr_err("Couldn't find uuid for %s in set\n",
1210                                dc->backing_dev_name);
1211                         return -ENOENT;
1212                 }
1213
1214                 u = uuid_find_empty(c);
1215                 if (!u) {
1216                         pr_err("Not caching %s, no room for UUID\n",
1217                                dc->backing_dev_name);
1218                         return -EINVAL;
1219                 }
1220         }
1221
1222         /*
1223          * Deadlocks since we're called via sysfs...
1224          * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1225          */
1226
1227         if (bch_is_zero(u->uuid, 16)) {
1228                 struct closure cl;
1229
1230                 closure_init_stack(&cl);
1231
1232                 memcpy(u->uuid, dc->sb.uuid, 16);
1233                 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1234                 u->first_reg = u->last_reg = rtime;
1235                 bch_uuid_write(c);
1236
1237                 memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1238                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1239
1240                 bch_write_bdev_super(dc, &cl);
1241                 closure_sync(&cl);
1242         } else {
1243                 u->last_reg = rtime;
1244                 bch_uuid_write(c);
1245         }
1246
1247         bcache_device_attach(&dc->disk, c, u - c->uuids);
1248         list_move(&dc->list, &c->cached_devs);
1249         calc_cached_dev_sectors(c);
1250
1251         /*
1252          * dc->c must be set before dc->count != 0 - paired with the mb in
1253          * cached_dev_get()
1254          */
1255         smp_wmb();
1256         refcount_set(&dc->count, 1);
1257
1258         /* Block writeback thread, but spawn it */
1259         down_write(&dc->writeback_lock);
1260         if (bch_cached_dev_writeback_start(dc)) {
1261                 up_write(&dc->writeback_lock);
1262                 pr_err("Couldn't start writeback facilities for %s\n",
1263                        dc->disk.disk->disk_name);
1264                 return -ENOMEM;
1265         }
1266
1267         if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1268                 atomic_set(&dc->has_dirty, 1);
1269                 bch_writeback_queue(dc);
1270         }
1271
1272         bch_sectors_dirty_init(&dc->disk);
1273
1274         ret = bch_cached_dev_run(dc);
1275         if (ret && (ret != -EBUSY)) {
1276                 up_write(&dc->writeback_lock);
1277                 /*
1278                  * bch_register_lock is held, bcache_device_stop() is not
1279                  * able to be directly called. The kthread and kworker
1280                  * created previously in bch_cached_dev_writeback_start()
1281                  * have to be stopped manually here.
1282                  */
1283                 kthread_stop(dc->writeback_thread);
1284                 cancel_writeback_rate_update_dwork(dc);
1285                 pr_err("Couldn't run cached device %s\n",
1286                        dc->backing_dev_name);
1287                 return ret;
1288         }
1289
1290         bcache_device_link(&dc->disk, c, "bdev");
1291         atomic_inc(&c->attached_dev_nr);
1292
1293         /* Allow the writeback thread to proceed */
1294         up_write(&dc->writeback_lock);
1295
1296         pr_info("Caching %s as %s on set %pU\n",
1297                 dc->backing_dev_name,
1298                 dc->disk.disk->disk_name,
1299                 dc->disk.c->sb.set_uuid);
1300         return 0;
1301 }
1302
1303 /* when dc->disk.kobj released */
1304 void bch_cached_dev_release(struct kobject *kobj)
1305 {
1306         struct cached_dev *dc = container_of(kobj, struct cached_dev,
1307                                              disk.kobj);
1308         kfree(dc);
1309         module_put(THIS_MODULE);
1310 }
1311
1312 static void cached_dev_free(struct closure *cl)
1313 {
1314         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1315
1316         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1317                 cancel_writeback_rate_update_dwork(dc);
1318
1319         if (!IS_ERR_OR_NULL(dc->writeback_thread))
1320                 kthread_stop(dc->writeback_thread);
1321         if (!IS_ERR_OR_NULL(dc->status_update_thread))
1322                 kthread_stop(dc->status_update_thread);
1323
1324         mutex_lock(&bch_register_lock);
1325
1326         if (atomic_read(&dc->running))
1327                 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1328         bcache_device_free(&dc->disk);
1329         list_del(&dc->list);
1330
1331         mutex_unlock(&bch_register_lock);
1332
1333         if (dc->sb_disk)
1334                 put_page(virt_to_page(dc->sb_disk));
1335
1336         if (!IS_ERR_OR_NULL(dc->bdev))
1337                 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1338
1339         wake_up(&unregister_wait);
1340
1341         kobject_put(&dc->disk.kobj);
1342 }
1343
1344 static void cached_dev_flush(struct closure *cl)
1345 {
1346         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1347         struct bcache_device *d = &dc->disk;
1348
1349         mutex_lock(&bch_register_lock);
1350         bcache_device_unlink(d);
1351         mutex_unlock(&bch_register_lock);
1352
1353         bch_cache_accounting_destroy(&dc->accounting);
1354         kobject_del(&d->kobj);
1355
1356         continue_at(cl, cached_dev_free, system_wq);
1357 }
1358
1359 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1360 {
1361         int ret;
1362         struct io *io;
1363         struct request_queue *q = bdev_get_queue(dc->bdev);
1364
1365         __module_get(THIS_MODULE);
1366         INIT_LIST_HEAD(&dc->list);
1367         closure_init(&dc->disk.cl, NULL);
1368         set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1369         kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1370         INIT_WORK(&dc->detach, cached_dev_detach_finish);
1371         sema_init(&dc->sb_write_mutex, 1);
1372         INIT_LIST_HEAD(&dc->io_lru);
1373         spin_lock_init(&dc->io_lock);
1374         bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1375
1376         dc->sequential_cutoff           = 4 << 20;
1377
1378         for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1379                 list_add(&io->lru, &dc->io_lru);
1380                 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1381         }
1382
1383         dc->disk.stripe_size = q->limits.io_opt >> 9;
1384
1385         if (dc->disk.stripe_size)
1386                 dc->partial_stripes_expensive =
1387                         q->limits.raid_partial_stripes_expensive;
1388
1389         ret = bcache_device_init(&dc->disk, block_size,
1390                          dc->bdev->bd_part->nr_sects - dc->sb.data_offset,
1391                          dc->bdev, &bcache_cached_ops);
1392         if (ret)
1393                 return ret;
1394
1395         dc->disk.disk->queue->backing_dev_info->ra_pages =
1396                 max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1397                     q->backing_dev_info->ra_pages);
1398
1399         atomic_set(&dc->io_errors, 0);
1400         dc->io_disable = false;
1401         dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1402         /* default to auto */
1403         dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1404
1405         bch_cached_dev_request_init(dc);
1406         bch_cached_dev_writeback_init(dc);
1407         return 0;
1408 }
1409
1410 /* Cached device - bcache superblock */
1411
1412 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1413                                  struct block_device *bdev,
1414                                  struct cached_dev *dc)
1415 {
1416         const char *err = "cannot allocate memory";
1417         struct cache_set *c;
1418         int ret = -ENOMEM;
1419
1420         bdevname(bdev, dc->backing_dev_name);
1421         memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1422         dc->bdev = bdev;
1423         dc->bdev->bd_holder = dc;
1424         dc->sb_disk = sb_disk;
1425
1426         if (cached_dev_init(dc, sb->block_size << 9))
1427                 goto err;
1428
1429         err = "error creating kobject";
1430         if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1431                         "bcache"))
1432                 goto err;
1433         if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1434                 goto err;
1435
1436         pr_info("registered backing device %s\n", dc->backing_dev_name);
1437
1438         list_add(&dc->list, &uncached_devices);
1439         /* attach to a matched cache set if it exists */
1440         list_for_each_entry(c, &bch_cache_sets, list)
1441                 bch_cached_dev_attach(dc, c, NULL);
1442
1443         if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1444             BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1445                 err = "failed to run cached device";
1446                 ret = bch_cached_dev_run(dc);
1447                 if (ret)
1448                         goto err;
1449         }
1450
1451         return 0;
1452 err:
1453         pr_notice("error %s: %s\n", dc->backing_dev_name, err);
1454         bcache_device_stop(&dc->disk);
1455         return ret;
1456 }
1457
1458 /* Flash only volumes */
1459
1460 /* When d->kobj released */
1461 void bch_flash_dev_release(struct kobject *kobj)
1462 {
1463         struct bcache_device *d = container_of(kobj, struct bcache_device,
1464                                                kobj);
1465         kfree(d);
1466 }
1467
1468 static void flash_dev_free(struct closure *cl)
1469 {
1470         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1471
1472         mutex_lock(&bch_register_lock);
1473         atomic_long_sub(bcache_dev_sectors_dirty(d),
1474                         &d->c->flash_dev_dirty_sectors);
1475         bcache_device_free(d);
1476         mutex_unlock(&bch_register_lock);
1477         kobject_put(&d->kobj);
1478 }
1479
1480 static void flash_dev_flush(struct closure *cl)
1481 {
1482         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1483
1484         mutex_lock(&bch_register_lock);
1485         bcache_device_unlink(d);
1486         mutex_unlock(&bch_register_lock);
1487         kobject_del(&d->kobj);
1488         continue_at(cl, flash_dev_free, system_wq);
1489 }
1490
1491 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1492 {
1493         struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1494                                           GFP_KERNEL);
1495         if (!d)
1496                 return -ENOMEM;
1497
1498         closure_init(&d->cl, NULL);
1499         set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1500
1501         kobject_init(&d->kobj, &bch_flash_dev_ktype);
1502
1503         if (bcache_device_init(d, block_bytes(c), u->sectors,
1504                         NULL, &bcache_flash_ops))
1505                 goto err;
1506
1507         bcache_device_attach(d, c, u - c->uuids);
1508         bch_sectors_dirty_init(d);
1509         bch_flash_dev_request_init(d);
1510         add_disk(d->disk);
1511
1512         if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1513                 goto err;
1514
1515         bcache_device_link(d, c, "volume");
1516
1517         return 0;
1518 err:
1519         kobject_put(&d->kobj);
1520         return -ENOMEM;
1521 }
1522
1523 static int flash_devs_run(struct cache_set *c)
1524 {
1525         int ret = 0;
1526         struct uuid_entry *u;
1527
1528         for (u = c->uuids;
1529              u < c->uuids + c->nr_uuids && !ret;
1530              u++)
1531                 if (UUID_FLASH_ONLY(u))
1532                         ret = flash_dev_run(c, u);
1533
1534         return ret;
1535 }
1536
1537 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1538 {
1539         struct uuid_entry *u;
1540
1541         if (test_bit(CACHE_SET_STOPPING, &c->flags))
1542                 return -EINTR;
1543
1544         if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1545                 return -EPERM;
1546
1547         u = uuid_find_empty(c);
1548         if (!u) {
1549                 pr_err("Can't create volume, no room for UUID\n");
1550                 return -EINVAL;
1551         }
1552
1553         get_random_bytes(u->uuid, 16);
1554         memset(u->label, 0, 32);
1555         u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1556
1557         SET_UUID_FLASH_ONLY(u, 1);
1558         u->sectors = size >> 9;
1559
1560         bch_uuid_write(c);
1561
1562         return flash_dev_run(c, u);
1563 }
1564
1565 bool bch_cached_dev_error(struct cached_dev *dc)
1566 {
1567         if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1568                 return false;
1569
1570         dc->io_disable = true;
1571         /* make others know io_disable is true earlier */
1572         smp_mb();
1573
1574         pr_err("stop %s: too many IO errors on backing device %s\n",
1575                dc->disk.disk->disk_name, dc->backing_dev_name);
1576
1577         bcache_device_stop(&dc->disk);
1578         return true;
1579 }
1580
1581 /* Cache set */
1582
1583 __printf(2, 3)
1584 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1585 {
1586         struct va_format vaf;
1587         va_list args;
1588
1589         if (c->on_error != ON_ERROR_PANIC &&
1590             test_bit(CACHE_SET_STOPPING, &c->flags))
1591                 return false;
1592
1593         if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1594                 pr_info("CACHE_SET_IO_DISABLE already set\n");
1595
1596         /*
1597          * XXX: we can be called from atomic context
1598          * acquire_console_sem();
1599          */
1600
1601         va_start(args, fmt);
1602
1603         vaf.fmt = fmt;
1604         vaf.va = &args;
1605
1606         pr_err("error on %pU: %pV, disabling caching\n",
1607                c->sb.set_uuid, &vaf);
1608
1609         va_end(args);
1610
1611         if (c->on_error == ON_ERROR_PANIC)
1612                 panic("panic forced after error\n");
1613
1614         bch_cache_set_unregister(c);
1615         return true;
1616 }
1617
1618 /* When c->kobj released */
1619 void bch_cache_set_release(struct kobject *kobj)
1620 {
1621         struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1622
1623         kfree(c);
1624         module_put(THIS_MODULE);
1625 }
1626
1627 static void cache_set_free(struct closure *cl)
1628 {
1629         struct cache_set *c = container_of(cl, struct cache_set, cl);
1630         struct cache *ca;
1631         unsigned int i;
1632
1633         debugfs_remove(c->debug);
1634
1635         bch_open_buckets_free(c);
1636         bch_btree_cache_free(c);
1637         bch_journal_free(c);
1638
1639         mutex_lock(&bch_register_lock);
1640         for_each_cache(ca, c, i)
1641                 if (ca) {
1642                         ca->set = NULL;
1643                         c->cache[ca->sb.nr_this_dev] = NULL;
1644                         kobject_put(&ca->kobj);
1645                 }
1646
1647         bch_bset_sort_state_free(&c->sort);
1648         free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1649
1650         if (c->moving_gc_wq)
1651                 destroy_workqueue(c->moving_gc_wq);
1652         bioset_exit(&c->bio_split);
1653         mempool_exit(&c->fill_iter);
1654         mempool_exit(&c->bio_meta);
1655         mempool_exit(&c->search);
1656         kfree(c->devices);
1657
1658         list_del(&c->list);
1659         mutex_unlock(&bch_register_lock);
1660
1661         pr_info("Cache set %pU unregistered\n", c->sb.set_uuid);
1662         wake_up(&unregister_wait);
1663
1664         closure_debug_destroy(&c->cl);
1665         kobject_put(&c->kobj);
1666 }
1667
1668 static void cache_set_flush(struct closure *cl)
1669 {
1670         struct cache_set *c = container_of(cl, struct cache_set, caching);
1671         struct cache *ca;
1672         struct btree *b;
1673         unsigned int i;
1674
1675         bch_cache_accounting_destroy(&c->accounting);
1676
1677         kobject_put(&c->internal);
1678         kobject_del(&c->kobj);
1679
1680         if (!IS_ERR_OR_NULL(c->gc_thread))
1681                 kthread_stop(c->gc_thread);
1682
1683         if (!IS_ERR_OR_NULL(c->root))
1684                 list_add(&c->root->list, &c->btree_cache);
1685
1686         /*
1687          * Avoid flushing cached nodes if cache set is retiring
1688          * due to too many I/O errors detected.
1689          */
1690         if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1691                 list_for_each_entry(b, &c->btree_cache, list) {
1692                         mutex_lock(&b->write_lock);
1693                         if (btree_node_dirty(b))
1694                                 __bch_btree_node_write(b, NULL);
1695                         mutex_unlock(&b->write_lock);
1696                 }
1697
1698         for_each_cache(ca, c, i)
1699                 if (ca->alloc_thread)
1700                         kthread_stop(ca->alloc_thread);
1701
1702         if (c->journal.cur) {
1703                 cancel_delayed_work_sync(&c->journal.work);
1704                 /* flush last journal entry if needed */
1705                 c->journal.work.work.func(&c->journal.work.work);
1706         }
1707
1708         closure_return(cl);
1709 }
1710
1711 /*
1712  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1713  * cache set is unregistering due to too many I/O errors. In this condition,
1714  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1715  * value and whether the broken cache has dirty data:
1716  *
1717  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1718  *  BCH_CACHED_STOP_AUTO               0               NO
1719  *  BCH_CACHED_STOP_AUTO               1               YES
1720  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1721  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1722  *
1723  * The expected behavior is, if stop_when_cache_set_failed is configured to
1724  * "auto" via sysfs interface, the bcache device will not be stopped if the
1725  * backing device is clean on the broken cache device.
1726  */
1727 static void conditional_stop_bcache_device(struct cache_set *c,
1728                                            struct bcache_device *d,
1729                                            struct cached_dev *dc)
1730 {
1731         if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1732                 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1733                         d->disk->disk_name, c->sb.set_uuid);
1734                 bcache_device_stop(d);
1735         } else if (atomic_read(&dc->has_dirty)) {
1736                 /*
1737                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1738                  * and dc->has_dirty == 1
1739                  */
1740                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1741                         d->disk->disk_name);
1742                 /*
1743                  * There might be a small time gap that cache set is
1744                  * released but bcache device is not. Inside this time
1745                  * gap, regular I/O requests will directly go into
1746                  * backing device as no cache set attached to. This
1747                  * behavior may also introduce potential inconsistence
1748                  * data in writeback mode while cache is dirty.
1749                  * Therefore before calling bcache_device_stop() due
1750                  * to a broken cache device, dc->io_disable should be
1751                  * explicitly set to true.
1752                  */
1753                 dc->io_disable = true;
1754                 /* make others know io_disable is true earlier */
1755                 smp_mb();
1756                 bcache_device_stop(d);
1757         } else {
1758                 /*
1759                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1760                  * and dc->has_dirty == 0
1761                  */
1762                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1763                         d->disk->disk_name);
1764         }
1765 }
1766
1767 static void __cache_set_unregister(struct closure *cl)
1768 {
1769         struct cache_set *c = container_of(cl, struct cache_set, caching);
1770         struct cached_dev *dc;
1771         struct bcache_device *d;
1772         size_t i;
1773
1774         mutex_lock(&bch_register_lock);
1775
1776         for (i = 0; i < c->devices_max_used; i++) {
1777                 d = c->devices[i];
1778                 if (!d)
1779                         continue;
1780
1781                 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1782                     test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1783                         dc = container_of(d, struct cached_dev, disk);
1784                         bch_cached_dev_detach(dc);
1785                         if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1786                                 conditional_stop_bcache_device(c, d, dc);
1787                 } else {
1788                         bcache_device_stop(d);
1789                 }
1790         }
1791
1792         mutex_unlock(&bch_register_lock);
1793
1794         continue_at(cl, cache_set_flush, system_wq);
1795 }
1796
1797 void bch_cache_set_stop(struct cache_set *c)
1798 {
1799         if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1800                 /* closure_fn set to __cache_set_unregister() */
1801                 closure_queue(&c->caching);
1802 }
1803
1804 void bch_cache_set_unregister(struct cache_set *c)
1805 {
1806         set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1807         bch_cache_set_stop(c);
1808 }
1809
1810 #define alloc_bucket_pages(gfp, c)                      \
1811         ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(bucket_pages(c))))
1812
1813 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1814 {
1815         int iter_size;
1816         struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1817
1818         if (!c)
1819                 return NULL;
1820
1821         __module_get(THIS_MODULE);
1822         closure_init(&c->cl, NULL);
1823         set_closure_fn(&c->cl, cache_set_free, system_wq);
1824
1825         closure_init(&c->caching, &c->cl);
1826         set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1827
1828         /* Maybe create continue_at_noreturn() and use it here? */
1829         closure_set_stopped(&c->cl);
1830         closure_put(&c->cl);
1831
1832         kobject_init(&c->kobj, &bch_cache_set_ktype);
1833         kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1834
1835         bch_cache_accounting_init(&c->accounting, &c->cl);
1836
1837         memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1838         c->sb.block_size        = sb->block_size;
1839         c->sb.bucket_size       = sb->bucket_size;
1840         c->sb.nr_in_set         = sb->nr_in_set;
1841         c->sb.last_mount        = sb->last_mount;
1842         c->bucket_bits          = ilog2(sb->bucket_size);
1843         c->block_bits           = ilog2(sb->block_size);
1844         c->nr_uuids             = bucket_bytes(c) / sizeof(struct uuid_entry);
1845         c->devices_max_used     = 0;
1846         atomic_set(&c->attached_dev_nr, 0);
1847         c->btree_pages          = bucket_pages(c);
1848         if (c->btree_pages > BTREE_MAX_PAGES)
1849                 c->btree_pages = max_t(int, c->btree_pages / 4,
1850                                        BTREE_MAX_PAGES);
1851
1852         sema_init(&c->sb_write_mutex, 1);
1853         mutex_init(&c->bucket_lock);
1854         init_waitqueue_head(&c->btree_cache_wait);
1855         spin_lock_init(&c->btree_cannibalize_lock);
1856         init_waitqueue_head(&c->bucket_wait);
1857         init_waitqueue_head(&c->gc_wait);
1858         sema_init(&c->uuid_write_mutex, 1);
1859
1860         spin_lock_init(&c->btree_gc_time.lock);
1861         spin_lock_init(&c->btree_split_time.lock);
1862         spin_lock_init(&c->btree_read_time.lock);
1863
1864         bch_moving_init_cache_set(c);
1865
1866         INIT_LIST_HEAD(&c->list);
1867         INIT_LIST_HEAD(&c->cached_devs);
1868         INIT_LIST_HEAD(&c->btree_cache);
1869         INIT_LIST_HEAD(&c->btree_cache_freeable);
1870         INIT_LIST_HEAD(&c->btree_cache_freed);
1871         INIT_LIST_HEAD(&c->data_buckets);
1872
1873         iter_size = (sb->bucket_size / sb->block_size + 1) *
1874                 sizeof(struct btree_iter_set);
1875
1876         if (!(c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL)) ||
1877             mempool_init_slab_pool(&c->search, 32, bch_search_cache) ||
1878             mempool_init_kmalloc_pool(&c->bio_meta, 2,
1879                                 sizeof(struct bbio) + sizeof(struct bio_vec) *
1880                                 bucket_pages(c)) ||
1881             mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
1882             bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1883                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER) ||
1884             !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1885             !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1886                                                 WQ_MEM_RECLAIM, 0)) ||
1887             bch_journal_alloc(c) ||
1888             bch_btree_cache_alloc(c) ||
1889             bch_open_buckets_alloc(c) ||
1890             bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1891                 goto err;
1892
1893         c->congested_read_threshold_us  = 2000;
1894         c->congested_write_threshold_us = 20000;
1895         c->error_limit  = DEFAULT_IO_ERROR_LIMIT;
1896         c->idle_max_writeback_rate_enabled = 1;
1897         WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1898
1899         return c;
1900 err:
1901         bch_cache_set_unregister(c);
1902         return NULL;
1903 }
1904
1905 static int run_cache_set(struct cache_set *c)
1906 {
1907         const char *err = "cannot allocate memory";
1908         struct cached_dev *dc, *t;
1909         struct cache *ca;
1910         struct closure cl;
1911         unsigned int i;
1912         LIST_HEAD(journal);
1913         struct journal_replay *l;
1914
1915         closure_init_stack(&cl);
1916
1917         for_each_cache(ca, c, i)
1918                 c->nbuckets += ca->sb.nbuckets;
1919         set_gc_sectors(c);
1920
1921         if (CACHE_SYNC(&c->sb)) {
1922                 struct bkey *k;
1923                 struct jset *j;
1924
1925                 err = "cannot allocate memory for journal";
1926                 if (bch_journal_read(c, &journal))
1927                         goto err;
1928
1929                 pr_debug("btree_journal_read() done\n");
1930
1931                 err = "no journal entries found";
1932                 if (list_empty(&journal))
1933                         goto err;
1934
1935                 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1936
1937                 err = "IO error reading priorities";
1938                 for_each_cache(ca, c, i) {
1939                         if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
1940                                 goto err;
1941                 }
1942
1943                 /*
1944                  * If prio_read() fails it'll call cache_set_error and we'll
1945                  * tear everything down right away, but if we perhaps checked
1946                  * sooner we could avoid journal replay.
1947                  */
1948
1949                 k = &j->btree_root;
1950
1951                 err = "bad btree root";
1952                 if (__bch_btree_ptr_invalid(c, k))
1953                         goto err;
1954
1955                 err = "error reading btree root";
1956                 c->root = bch_btree_node_get(c, NULL, k,
1957                                              j->btree_level,
1958                                              true, NULL);
1959                 if (IS_ERR_OR_NULL(c->root))
1960                         goto err;
1961
1962                 list_del_init(&c->root->list);
1963                 rw_unlock(true, c->root);
1964
1965                 err = uuid_read(c, j, &cl);
1966                 if (err)
1967                         goto err;
1968
1969                 err = "error in recovery";
1970                 if (bch_btree_check(c))
1971                         goto err;
1972
1973                 bch_journal_mark(c, &journal);
1974                 bch_initial_gc_finish(c);
1975                 pr_debug("btree_check() done\n");
1976
1977                 /*
1978                  * bcache_journal_next() can't happen sooner, or
1979                  * btree_gc_finish() will give spurious errors about last_gc >
1980                  * gc_gen - this is a hack but oh well.
1981                  */
1982                 bch_journal_next(&c->journal);
1983
1984                 err = "error starting allocator thread";
1985                 for_each_cache(ca, c, i)
1986                         if (bch_cache_allocator_start(ca))
1987                                 goto err;
1988
1989                 /*
1990                  * First place it's safe to allocate: btree_check() and
1991                  * btree_gc_finish() have to run before we have buckets to
1992                  * allocate, and bch_bucket_alloc_set() might cause a journal
1993                  * entry to be written so bcache_journal_next() has to be called
1994                  * first.
1995                  *
1996                  * If the uuids were in the old format we have to rewrite them
1997                  * before the next journal entry is written:
1998                  */
1999                 if (j->version < BCACHE_JSET_VERSION_UUID)
2000                         __uuid_write(c);
2001
2002                 err = "bcache: replay journal failed";
2003                 if (bch_journal_replay(c, &journal))
2004                         goto err;
2005         } else {
2006                 pr_notice("invalidating existing data\n");
2007
2008                 for_each_cache(ca, c, i) {
2009                         unsigned int j;
2010
2011                         ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2012                                               2, SB_JOURNAL_BUCKETS);
2013
2014                         for (j = 0; j < ca->sb.keys; j++)
2015                                 ca->sb.d[j] = ca->sb.first_bucket + j;
2016                 }
2017
2018                 bch_initial_gc_finish(c);
2019
2020                 err = "error starting allocator thread";
2021                 for_each_cache(ca, c, i)
2022                         if (bch_cache_allocator_start(ca))
2023                                 goto err;
2024
2025                 mutex_lock(&c->bucket_lock);
2026                 for_each_cache(ca, c, i)
2027                         bch_prio_write(ca, true);
2028                 mutex_unlock(&c->bucket_lock);
2029
2030                 err = "cannot allocate new UUID bucket";
2031                 if (__uuid_write(c))
2032                         goto err;
2033
2034                 err = "cannot allocate new btree root";
2035                 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2036                 if (IS_ERR_OR_NULL(c->root))
2037                         goto err;
2038
2039                 mutex_lock(&c->root->write_lock);
2040                 bkey_copy_key(&c->root->key, &MAX_KEY);
2041                 bch_btree_node_write(c->root, &cl);
2042                 mutex_unlock(&c->root->write_lock);
2043
2044                 bch_btree_set_root(c->root);
2045                 rw_unlock(true, c->root);
2046
2047                 /*
2048                  * We don't want to write the first journal entry until
2049                  * everything is set up - fortunately journal entries won't be
2050                  * written until the SET_CACHE_SYNC() here:
2051                  */
2052                 SET_CACHE_SYNC(&c->sb, true);
2053
2054                 bch_journal_next(&c->journal);
2055                 bch_journal_meta(c, &cl);
2056         }
2057
2058         err = "error starting gc thread";
2059         if (bch_gc_thread_start(c))
2060                 goto err;
2061
2062         closure_sync(&cl);
2063         c->sb.last_mount = (u32)ktime_get_real_seconds();
2064         bcache_write_super(c);
2065
2066         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2067                 bch_cached_dev_attach(dc, c, NULL);
2068
2069         flash_devs_run(c);
2070
2071         set_bit(CACHE_SET_RUNNING, &c->flags);
2072         return 0;
2073 err:
2074         while (!list_empty(&journal)) {
2075                 l = list_first_entry(&journal, struct journal_replay, list);
2076                 list_del(&l->list);
2077                 kfree(l);
2078         }
2079
2080         closure_sync(&cl);
2081
2082         bch_cache_set_error(c, "%s", err);
2083
2084         return -EIO;
2085 }
2086
2087 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
2088 {
2089         return ca->sb.block_size        == c->sb.block_size &&
2090                 ca->sb.bucket_size      == c->sb.bucket_size &&
2091                 ca->sb.nr_in_set        == c->sb.nr_in_set;
2092 }
2093
2094 static const char *register_cache_set(struct cache *ca)
2095 {
2096         char buf[12];
2097         const char *err = "cannot allocate memory";
2098         struct cache_set *c;
2099
2100         list_for_each_entry(c, &bch_cache_sets, list)
2101                 if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
2102                         if (c->cache[ca->sb.nr_this_dev])
2103                                 return "duplicate cache set member";
2104
2105                         if (!can_attach_cache(ca, c))
2106                                 return "cache sb does not match set";
2107
2108                         if (!CACHE_SYNC(&ca->sb))
2109                                 SET_CACHE_SYNC(&c->sb, false);
2110
2111                         goto found;
2112                 }
2113
2114         c = bch_cache_set_alloc(&ca->sb);
2115         if (!c)
2116                 return err;
2117
2118         err = "error creating kobject";
2119         if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
2120             kobject_add(&c->internal, &c->kobj, "internal"))
2121                 goto err;
2122
2123         if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2124                 goto err;
2125
2126         bch_debug_init_cache_set(c);
2127
2128         list_add(&c->list, &bch_cache_sets);
2129 found:
2130         sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2131         if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2132             sysfs_create_link(&c->kobj, &ca->kobj, buf))
2133                 goto err;
2134
2135         if (ca->sb.seq > c->sb.seq) {
2136                 c->sb.version           = ca->sb.version;
2137                 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2138                 c->sb.flags             = ca->sb.flags;
2139                 c->sb.seq               = ca->sb.seq;
2140                 pr_debug("set version = %llu\n", c->sb.version);
2141         }
2142
2143         kobject_get(&ca->kobj);
2144         ca->set = c;
2145         ca->set->cache[ca->sb.nr_this_dev] = ca;
2146         c->cache_by_alloc[c->caches_loaded++] = ca;
2147
2148         if (c->caches_loaded == c->sb.nr_in_set) {
2149                 err = "failed to run cache set";
2150                 if (run_cache_set(c) < 0)
2151                         goto err;
2152         }
2153
2154         return NULL;
2155 err:
2156         bch_cache_set_unregister(c);
2157         return err;
2158 }
2159
2160 /* Cache device */
2161
2162 /* When ca->kobj released */
2163 void bch_cache_release(struct kobject *kobj)
2164 {
2165         struct cache *ca = container_of(kobj, struct cache, kobj);
2166         unsigned int i;
2167
2168         if (ca->set) {
2169                 BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2170                 ca->set->cache[ca->sb.nr_this_dev] = NULL;
2171         }
2172
2173         free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2174         kfree(ca->prio_buckets);
2175         vfree(ca->buckets);
2176
2177         free_heap(&ca->heap);
2178         free_fifo(&ca->free_inc);
2179
2180         for (i = 0; i < RESERVE_NR; i++)
2181                 free_fifo(&ca->free[i]);
2182
2183         if (ca->sb_disk)
2184                 put_page(virt_to_page(ca->sb_disk));
2185
2186         if (!IS_ERR_OR_NULL(ca->bdev))
2187                 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2188
2189         kfree(ca);
2190         module_put(THIS_MODULE);
2191 }
2192
2193 static int cache_alloc(struct cache *ca)
2194 {
2195         size_t free;
2196         size_t btree_buckets;
2197         struct bucket *b;
2198         int ret = -ENOMEM;
2199         const char *err = NULL;
2200
2201         __module_get(THIS_MODULE);
2202         kobject_init(&ca->kobj, &bch_cache_ktype);
2203
2204         bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2205
2206         /*
2207          * when ca->sb.njournal_buckets is not zero, journal exists,
2208          * and in bch_journal_replay(), tree node may split,
2209          * so bucket of RESERVE_BTREE type is needed,
2210          * the worst situation is all journal buckets are valid journal,
2211          * and all the keys need to replay,
2212          * so the number of  RESERVE_BTREE type buckets should be as much
2213          * as journal buckets
2214          */
2215         btree_buckets = ca->sb.njournal_buckets ?: 8;
2216         free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2217         if (!free) {
2218                 ret = -EPERM;
2219                 err = "ca->sb.nbuckets is too small";
2220                 goto err_free;
2221         }
2222
2223         if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2224                                                 GFP_KERNEL)) {
2225                 err = "ca->free[RESERVE_BTREE] alloc failed";
2226                 goto err_btree_alloc;
2227         }
2228
2229         if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2230                                                         GFP_KERNEL)) {
2231                 err = "ca->free[RESERVE_PRIO] alloc failed";
2232                 goto err_prio_alloc;
2233         }
2234
2235         if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2236                 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2237                 goto err_movinggc_alloc;
2238         }
2239
2240         if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2241                 err = "ca->free[RESERVE_NONE] alloc failed";
2242                 goto err_none_alloc;
2243         }
2244
2245         if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2246                 err = "ca->free_inc alloc failed";
2247                 goto err_free_inc_alloc;
2248         }
2249
2250         if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2251                 err = "ca->heap alloc failed";
2252                 goto err_heap_alloc;
2253         }
2254
2255         ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2256                               ca->sb.nbuckets));
2257         if (!ca->buckets) {
2258                 err = "ca->buckets alloc failed";
2259                 goto err_buckets_alloc;
2260         }
2261
2262         ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2263                                    prio_buckets(ca), 2),
2264                                    GFP_KERNEL);
2265         if (!ca->prio_buckets) {
2266                 err = "ca->prio_buckets alloc failed";
2267                 goto err_prio_buckets_alloc;
2268         }
2269
2270         ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca);
2271         if (!ca->disk_buckets) {
2272                 err = "ca->disk_buckets alloc failed";
2273                 goto err_disk_buckets_alloc;
2274         }
2275
2276         ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2277
2278         for_each_bucket(b, ca)
2279                 atomic_set(&b->pin, 0);
2280         return 0;
2281
2282 err_disk_buckets_alloc:
2283         kfree(ca->prio_buckets);
2284 err_prio_buckets_alloc:
2285         vfree(ca->buckets);
2286 err_buckets_alloc:
2287         free_heap(&ca->heap);
2288 err_heap_alloc:
2289         free_fifo(&ca->free_inc);
2290 err_free_inc_alloc:
2291         free_fifo(&ca->free[RESERVE_NONE]);
2292 err_none_alloc:
2293         free_fifo(&ca->free[RESERVE_MOVINGGC]);
2294 err_movinggc_alloc:
2295         free_fifo(&ca->free[RESERVE_PRIO]);
2296 err_prio_alloc:
2297         free_fifo(&ca->free[RESERVE_BTREE]);
2298 err_btree_alloc:
2299 err_free:
2300         module_put(THIS_MODULE);
2301         if (err)
2302                 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2303         return ret;
2304 }
2305
2306 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2307                                 struct block_device *bdev, struct cache *ca)
2308 {
2309         const char *err = NULL; /* must be set for any error case */
2310         int ret = 0;
2311
2312         bdevname(bdev, ca->cache_dev_name);
2313         memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2314         ca->bdev = bdev;
2315         ca->bdev->bd_holder = ca;
2316         ca->sb_disk = sb_disk;
2317
2318         if (blk_queue_discard(bdev_get_queue(bdev)))
2319                 ca->discard = CACHE_DISCARD(&ca->sb);
2320
2321         ret = cache_alloc(ca);
2322         if (ret != 0) {
2323                 /*
2324                  * If we failed here, it means ca->kobj is not initialized yet,
2325                  * kobject_put() won't be called and there is no chance to
2326                  * call blkdev_put() to bdev in bch_cache_release(). So we
2327                  * explicitly call blkdev_put() here.
2328                  */
2329                 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2330                 if (ret == -ENOMEM)
2331                         err = "cache_alloc(): -ENOMEM";
2332                 else if (ret == -EPERM)
2333                         err = "cache_alloc(): cache device is too small";
2334                 else
2335                         err = "cache_alloc(): unknown error";
2336                 goto err;
2337         }
2338
2339         if (kobject_add(&ca->kobj,
2340                         &part_to_dev(bdev->bd_part)->kobj,
2341                         "bcache")) {
2342                 err = "error calling kobject_add";
2343                 ret = -ENOMEM;
2344                 goto out;
2345         }
2346
2347         mutex_lock(&bch_register_lock);
2348         err = register_cache_set(ca);
2349         mutex_unlock(&bch_register_lock);
2350
2351         if (err) {
2352                 ret = -ENODEV;
2353                 goto out;
2354         }
2355
2356         pr_info("registered cache device %s\n", ca->cache_dev_name);
2357
2358 out:
2359         kobject_put(&ca->kobj);
2360
2361 err:
2362         if (err)
2363                 pr_notice("error %s: %s\n", ca->cache_dev_name, err);
2364
2365         return ret;
2366 }
2367
2368 /* Global interfaces/init */
2369
2370 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2371                                const char *buffer, size_t size);
2372 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2373                                          struct kobj_attribute *attr,
2374                                          const char *buffer, size_t size);
2375
2376 kobj_attribute_write(register,          register_bcache);
2377 kobj_attribute_write(register_quiet,    register_bcache);
2378 kobj_attribute_write(register_async,    register_bcache);
2379 kobj_attribute_write(pendings_cleanup,  bch_pending_bdevs_cleanup);
2380
2381 static bool bch_is_open_backing(struct block_device *bdev)
2382 {
2383         struct cache_set *c, *tc;
2384         struct cached_dev *dc, *t;
2385
2386         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2387                 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2388                         if (dc->bdev == bdev)
2389                                 return true;
2390         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2391                 if (dc->bdev == bdev)
2392                         return true;
2393         return false;
2394 }
2395
2396 static bool bch_is_open_cache(struct block_device *bdev)
2397 {
2398         struct cache_set *c, *tc;
2399         struct cache *ca;
2400         unsigned int i;
2401
2402         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2403                 for_each_cache(ca, c, i)
2404                         if (ca->bdev == bdev)
2405                                 return true;
2406         return false;
2407 }
2408
2409 static bool bch_is_open(struct block_device *bdev)
2410 {
2411         return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2412 }
2413
2414 struct async_reg_args {
2415         struct delayed_work reg_work;
2416         char *path;
2417         struct cache_sb *sb;
2418         struct cache_sb_disk *sb_disk;
2419         struct block_device *bdev;
2420 };
2421
2422 static void register_bdev_worker(struct work_struct *work)
2423 {
2424         int fail = false;
2425         struct async_reg_args *args =
2426                 container_of(work, struct async_reg_args, reg_work.work);
2427         struct cached_dev *dc;
2428
2429         dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2430         if (!dc) {
2431                 fail = true;
2432                 put_page(virt_to_page(args->sb_disk));
2433                 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2434                 goto out;
2435         }
2436
2437         mutex_lock(&bch_register_lock);
2438         if (register_bdev(args->sb, args->sb_disk, args->bdev, dc) < 0)
2439                 fail = true;
2440         mutex_unlock(&bch_register_lock);
2441
2442 out:
2443         if (fail)
2444                 pr_info("error %s: fail to register backing device\n",
2445                         args->path);
2446         kfree(args->sb);
2447         kfree(args->path);
2448         kfree(args);
2449         module_put(THIS_MODULE);
2450 }
2451
2452 static void register_cache_worker(struct work_struct *work)
2453 {
2454         int fail = false;
2455         struct async_reg_args *args =
2456                 container_of(work, struct async_reg_args, reg_work.work);
2457         struct cache *ca;
2458
2459         ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2460         if (!ca) {
2461                 fail = true;
2462                 put_page(virt_to_page(args->sb_disk));
2463                 blkdev_put(args->bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2464                 goto out;
2465         }
2466
2467         /* blkdev_put() will be called in bch_cache_release() */
2468         if (register_cache(args->sb, args->sb_disk, args->bdev, ca) != 0)
2469                 fail = true;
2470
2471 out:
2472         if (fail)
2473                 pr_info("error %s: fail to register cache device\n",
2474                         args->path);
2475         kfree(args->sb);
2476         kfree(args->path);
2477         kfree(args);
2478         module_put(THIS_MODULE);
2479 }
2480
2481 static void register_device_aync(struct async_reg_args *args)
2482 {
2483         if (SB_IS_BDEV(args->sb))
2484                 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2485         else
2486                 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2487
2488         /* 10 jiffies is enough for a delay */
2489         queue_delayed_work(system_wq, &args->reg_work, 10);
2490 }
2491
2492 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2493                                const char *buffer, size_t size)
2494 {
2495         const char *err;
2496         char *path = NULL;
2497         struct cache_sb *sb;
2498         struct cache_sb_disk *sb_disk;
2499         struct block_device *bdev;
2500         ssize_t ret;
2501
2502         ret = -EBUSY;
2503         err = "failed to reference bcache module";
2504         if (!try_module_get(THIS_MODULE))
2505                 goto out;
2506
2507         /* For latest state of bcache_is_reboot */
2508         smp_mb();
2509         err = "bcache is in reboot";
2510         if (bcache_is_reboot)
2511                 goto out_module_put;
2512
2513         ret = -ENOMEM;
2514         err = "cannot allocate memory";
2515         path = kstrndup(buffer, size, GFP_KERNEL);
2516         if (!path)
2517                 goto out_module_put;
2518
2519         sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2520         if (!sb)
2521                 goto out_free_path;
2522
2523         ret = -EINVAL;
2524         err = "failed to open device";
2525         bdev = blkdev_get_by_path(strim(path),
2526                                   FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2527                                   sb);
2528         if (IS_ERR(bdev)) {
2529                 if (bdev == ERR_PTR(-EBUSY)) {
2530                         bdev = lookup_bdev(strim(path));
2531                         mutex_lock(&bch_register_lock);
2532                         if (!IS_ERR(bdev) && bch_is_open(bdev))
2533                                 err = "device already registered";
2534                         else
2535                                 err = "device busy";
2536                         mutex_unlock(&bch_register_lock);
2537                         if (!IS_ERR(bdev))
2538                                 bdput(bdev);
2539                         if (attr == &ksysfs_register_quiet)
2540                                 goto done;
2541                 }
2542                 goto out_free_sb;
2543         }
2544
2545         err = "failed to set blocksize";
2546         if (set_blocksize(bdev, 4096))
2547                 goto out_blkdev_put;
2548
2549         err = read_super(sb, bdev, &sb_disk);
2550         if (err)
2551                 goto out_blkdev_put;
2552
2553         err = "failed to register device";
2554         if (attr == &ksysfs_register_async) {
2555                 /* register in asynchronous way */
2556                 struct async_reg_args *args =
2557                         kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2558
2559                 if (!args) {
2560                         ret = -ENOMEM;
2561                         err = "cannot allocate memory";
2562                         goto out_put_sb_page;
2563                 }
2564
2565                 args->path      = path;
2566                 args->sb        = sb;
2567                 args->sb_disk   = sb_disk;
2568                 args->bdev      = bdev;
2569                 register_device_aync(args);
2570                 /* No wait and returns to user space */
2571                 goto async_done;
2572         }
2573
2574         if (SB_IS_BDEV(sb)) {
2575                 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2576
2577                 if (!dc)
2578                         goto out_put_sb_page;
2579
2580                 mutex_lock(&bch_register_lock);
2581                 ret = register_bdev(sb, sb_disk, bdev, dc);
2582                 mutex_unlock(&bch_register_lock);
2583                 /* blkdev_put() will be called in cached_dev_free() */
2584                 if (ret < 0)
2585                         goto out_free_sb;
2586         } else {
2587                 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2588
2589                 if (!ca)
2590                         goto out_put_sb_page;
2591
2592                 /* blkdev_put() will be called in bch_cache_release() */
2593                 if (register_cache(sb, sb_disk, bdev, ca) != 0)
2594                         goto out_free_sb;
2595         }
2596
2597 done:
2598         kfree(sb);
2599         kfree(path);
2600         module_put(THIS_MODULE);
2601 async_done:
2602         return size;
2603
2604 out_put_sb_page:
2605         put_page(virt_to_page(sb_disk));
2606 out_blkdev_put:
2607         blkdev_put(bdev, FMODE_READ | FMODE_WRITE | FMODE_EXCL);
2608 out_free_sb:
2609         kfree(sb);
2610 out_free_path:
2611         kfree(path);
2612         path = NULL;
2613 out_module_put:
2614         module_put(THIS_MODULE);
2615 out:
2616         pr_info("error %s: %s\n", path?path:"", err);
2617         return ret;
2618 }
2619
2620
2621 struct pdev {
2622         struct list_head list;
2623         struct cached_dev *dc;
2624 };
2625
2626 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2627                                          struct kobj_attribute *attr,
2628                                          const char *buffer,
2629                                          size_t size)
2630 {
2631         LIST_HEAD(pending_devs);
2632         ssize_t ret = size;
2633         struct cached_dev *dc, *tdc;
2634         struct pdev *pdev, *tpdev;
2635         struct cache_set *c, *tc;
2636
2637         mutex_lock(&bch_register_lock);
2638         list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2639                 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2640                 if (!pdev)
2641                         break;
2642                 pdev->dc = dc;
2643                 list_add(&pdev->list, &pending_devs);
2644         }
2645
2646         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2647                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2648                         char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2649                         char *set_uuid = c->sb.uuid;
2650
2651                         if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2652                                 list_del(&pdev->list);
2653                                 kfree(pdev);
2654                                 break;
2655                         }
2656                 }
2657         }
2658         mutex_unlock(&bch_register_lock);
2659
2660         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2661                 pr_info("delete pdev %p\n", pdev);
2662                 list_del(&pdev->list);
2663                 bcache_device_stop(&pdev->dc->disk);
2664                 kfree(pdev);
2665         }
2666
2667         return ret;
2668 }
2669
2670 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2671 {
2672         if (bcache_is_reboot)
2673                 return NOTIFY_DONE;
2674
2675         if (code == SYS_DOWN ||
2676             code == SYS_HALT ||
2677             code == SYS_POWER_OFF) {
2678                 DEFINE_WAIT(wait);
2679                 unsigned long start = jiffies;
2680                 bool stopped = false;
2681
2682                 struct cache_set *c, *tc;
2683                 struct cached_dev *dc, *tdc;
2684
2685                 mutex_lock(&bch_register_lock);
2686
2687                 if (bcache_is_reboot)
2688                         goto out;
2689
2690                 /* New registration is rejected since now */
2691                 bcache_is_reboot = true;
2692                 /*
2693                  * Make registering caller (if there is) on other CPU
2694                  * core know bcache_is_reboot set to true earlier
2695                  */
2696                 smp_mb();
2697
2698                 if (list_empty(&bch_cache_sets) &&
2699                     list_empty(&uncached_devices))
2700                         goto out;
2701
2702                 mutex_unlock(&bch_register_lock);
2703
2704                 pr_info("Stopping all devices:\n");
2705
2706                 /*
2707                  * The reason bch_register_lock is not held to call
2708                  * bch_cache_set_stop() and bcache_device_stop() is to
2709                  * avoid potential deadlock during reboot, because cache
2710                  * set or bcache device stopping process will acqurie
2711                  * bch_register_lock too.
2712                  *
2713                  * We are safe here because bcache_is_reboot sets to
2714                  * true already, register_bcache() will reject new
2715                  * registration now. bcache_is_reboot also makes sure
2716                  * bcache_reboot() won't be re-entered on by other thread,
2717                  * so there is no race in following list iteration by
2718                  * list_for_each_entry_safe().
2719                  */
2720                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2721                         bch_cache_set_stop(c);
2722
2723                 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2724                         bcache_device_stop(&dc->disk);
2725
2726
2727                 /*
2728                  * Give an early chance for other kthreads and
2729                  * kworkers to stop themselves
2730                  */
2731                 schedule();
2732
2733                 /* What's a condition variable? */
2734                 while (1) {
2735                         long timeout = start + 10 * HZ - jiffies;
2736
2737                         mutex_lock(&bch_register_lock);
2738                         stopped = list_empty(&bch_cache_sets) &&
2739                                 list_empty(&uncached_devices);
2740
2741                         if (timeout < 0 || stopped)
2742                                 break;
2743
2744                         prepare_to_wait(&unregister_wait, &wait,
2745                                         TASK_UNINTERRUPTIBLE);
2746
2747                         mutex_unlock(&bch_register_lock);
2748                         schedule_timeout(timeout);
2749                 }
2750
2751                 finish_wait(&unregister_wait, &wait);
2752
2753                 if (stopped)
2754                         pr_info("All devices stopped\n");
2755                 else
2756                         pr_notice("Timeout waiting for devices to be closed\n");
2757 out:
2758                 mutex_unlock(&bch_register_lock);
2759         }
2760
2761         return NOTIFY_DONE;
2762 }
2763
2764 static struct notifier_block reboot = {
2765         .notifier_call  = bcache_reboot,
2766         .priority       = INT_MAX, /* before any real devices */
2767 };
2768
2769 static void bcache_exit(void)
2770 {
2771         bch_debug_exit();
2772         bch_request_exit();
2773         if (bcache_kobj)
2774                 kobject_put(bcache_kobj);
2775         if (bcache_wq)
2776                 destroy_workqueue(bcache_wq);
2777         if (bch_journal_wq)
2778                 destroy_workqueue(bch_journal_wq);
2779
2780         if (bcache_major)
2781                 unregister_blkdev(bcache_major, "bcache");
2782         unregister_reboot_notifier(&reboot);
2783         mutex_destroy(&bch_register_lock);
2784 }
2785
2786 /* Check and fixup module parameters */
2787 static void check_module_parameters(void)
2788 {
2789         if (bch_cutoff_writeback_sync == 0)
2790                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2791         else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2792                 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2793                         bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2794                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2795         }
2796
2797         if (bch_cutoff_writeback == 0)
2798                 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2799         else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2800                 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2801                         bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2802                 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2803         }
2804
2805         if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2806                 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2807                         bch_cutoff_writeback, bch_cutoff_writeback_sync);
2808                 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2809         }
2810 }
2811
2812 static int __init bcache_init(void)
2813 {
2814         static const struct attribute *files[] = {
2815                 &ksysfs_register.attr,
2816                 &ksysfs_register_quiet.attr,
2817 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2818                 &ksysfs_register_async.attr,
2819 #endif
2820                 &ksysfs_pendings_cleanup.attr,
2821                 NULL
2822         };
2823
2824         check_module_parameters();
2825
2826         mutex_init(&bch_register_lock);
2827         init_waitqueue_head(&unregister_wait);
2828         register_reboot_notifier(&reboot);
2829
2830         bcache_major = register_blkdev(0, "bcache");
2831         if (bcache_major < 0) {
2832                 unregister_reboot_notifier(&reboot);
2833                 mutex_destroy(&bch_register_lock);
2834                 return bcache_major;
2835         }
2836
2837         bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2838         if (!bcache_wq)
2839                 goto err;
2840
2841         bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2842         if (!bch_journal_wq)
2843                 goto err;
2844
2845         bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2846         if (!bcache_kobj)
2847                 goto err;
2848
2849         if (bch_request_init() ||
2850             sysfs_create_files(bcache_kobj, files))
2851                 goto err;
2852
2853         bch_debug_init();
2854         closure_debug_init();
2855
2856         bcache_is_reboot = false;
2857
2858         return 0;
2859 err:
2860         bcache_exit();
2861         return -ENOMEM;
2862 }
2863
2864 /*
2865  * Module hooks
2866  */
2867 module_exit(bcache_exit);
2868 module_init(bcache_init);
2869
2870 module_param(bch_cutoff_writeback, uint, 0);
2871 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2872
2873 module_param(bch_cutoff_writeback_sync, uint, 0);
2874 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2875
2876 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2877 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2878 MODULE_LICENSE("GPL");