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