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