drivers/md/bcache/super.c

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