drivers/md/bcache/super.c

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