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