drivers/md/bcache/super.c

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