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 #include "features.h"
  17
  18 #include <linux/blkdev.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/debugfs.h>
  21 #include <linux/idr.h>
  22 #include <linux/kthread.h>
  23 #include <linux/workqueue.h>
  24 #include <linux/module.h>
  25 #include <linux/random.h>
  26 #include <linux/reboot.h>
  27 #include <linux/sysfs.h>
  28
  29 unsigned int bch_cutoff_writeback;
  30 unsigned int bch_cutoff_writeback_sync;
  31
  32 static const char bcache_magic[] = {
  33         0xc6, 0x85, 0x73, 0xf6, 0x4e, 0x1a, 0x45, 0xca,
  34         0x82, 0x65, 0xf5, 0x7f, 0x48, 0xba, 0x6d, 0x81
  35 };
  36
  37 static const char invalid_uuid[] = {
  38         0xa0, 0x3e, 0xf8, 0xed, 0x3e, 0xe1, 0xb8, 0x78,
  39         0xc8, 0x50, 0xfc, 0x5e, 0xcb, 0x16, 0xcd, 0x99
  40 };
  41
  42 static struct kobject *bcache_kobj;
  43 struct mutex bch_register_lock;
  44 bool bcache_is_reboot;
  45 LIST_HEAD(bch_cache_sets);
  46 static LIST_HEAD(uncached_devices);
  47
  48 static int bcache_major;
  49 static DEFINE_IDA(bcache_device_idx);
  50 static wait_queue_head_t unregister_wait;
  51 struct workqueue_struct *bcache_wq;
  52 struct workqueue_struct *bch_flush_wq;
  53 struct workqueue_struct *bch_journal_wq;
  54
  55
  56 #define BTREE_MAX_PAGES         (256 * 1024 / PAGE_SIZE)
  57 /* limitation of partitions number on single bcache device */
  58 #define BCACHE_MINORS           128
  59 /* limitation of bcache devices number on single system */
  60 #define BCACHE_DEVICE_IDX_MAX   ((1U << MINORBITS)/BCACHE_MINORS)
  61
  62 /* Superblock */
  63
  64 static unsigned int get_bucket_size(struct cache_sb *sb, struct cache_sb_disk *s)
  65 {
  66         unsigned int bucket_size = le16_to_cpu(s->bucket_size);
  67
  68         if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
  69                 if (bch_has_feature_large_bucket(sb)) {
  70                         unsigned int max, order;
  71
  72                         max = sizeof(unsigned int) * BITS_PER_BYTE - 1;
  73                         order = le16_to_cpu(s->bucket_size);
  74                         /*
  75                          * bcache tool will make sure the overflow won't
  76                          * happen, an error message here is enough.
  77                          */
  78                         if (order > max)
  79                                 pr_err("Bucket size (1 << %u) overflows\n",
  80                                         order);
  81                         bucket_size = 1 << order;
  82                 } else if (bch_has_feature_obso_large_bucket(sb)) {
  83                         bucket_size +=
  84                                 le16_to_cpu(s->obso_bucket_size_hi) << 16;
  85                 }
  86         }
  87
  88         return bucket_size;
  89 }
  90
  91 static const char *read_super_common(struct cache_sb *sb,  struct block_device *bdev,
  92                                      struct cache_sb_disk *s)
  93 {
  94         const char *err;
  95         unsigned int i;
  96
  97         sb->first_bucket= le16_to_cpu(s->first_bucket);
  98         sb->nbuckets    = le64_to_cpu(s->nbuckets);
  99         sb->bucket_size = get_bucket_size(sb, s);
 100
 101         sb->nr_in_set   = le16_to_cpu(s->nr_in_set);
 102         sb->nr_this_dev = le16_to_cpu(s->nr_this_dev);
 103
 104         err = "Too many journal buckets";
 105         if (sb->keys > SB_JOURNAL_BUCKETS)
 106                 goto err;
 107
 108         err = "Too many buckets";
 109         if (sb->nbuckets > LONG_MAX)
 110                 goto err;
 111
 112         err = "Not enough buckets";
 113         if (sb->nbuckets < 1 << 7)
 114                 goto err;
 115
 116         err = "Bad block size (not power of 2)";
 117         if (!is_power_of_2(sb->block_size))
 118                 goto err;
 119
 120         err = "Bad block size (larger than page size)";
 121         if (sb->block_size > PAGE_SECTORS)
 122                 goto err;
 123
 124         err = "Bad bucket size (not power of 2)";
 125         if (!is_power_of_2(sb->bucket_size))
 126                 goto err;
 127
 128         err = "Bad bucket size (smaller than page size)";
 129         if (sb->bucket_size < PAGE_SECTORS)
 130                 goto err;
 131
 132         err = "Invalid superblock: device too small";
 133         if (get_capacity(bdev->bd_disk) <
 134             sb->bucket_size * sb->nbuckets)
 135                 goto err;
 136
 137         err = "Bad UUID";
 138         if (bch_is_zero(sb->set_uuid, 16))
 139                 goto err;
 140
 141         err = "Bad cache device number in set";
 142         if (!sb->nr_in_set ||
 143             sb->nr_in_set <= sb->nr_this_dev ||
 144             sb->nr_in_set > MAX_CACHES_PER_SET)
 145                 goto err;
 146
 147         err = "Journal buckets not sequential";
 148         for (i = 0; i < sb->keys; i++)
 149                 if (sb->d[i] != sb->first_bucket + i)
 150                         goto err;
 151
 152         err = "Too many journal buckets";
 153         if (sb->first_bucket + sb->keys > sb->nbuckets)
 154                 goto err;
 155
 156         err = "Invalid superblock: first bucket comes before end of super";
 157         if (sb->first_bucket * sb->bucket_size < 16)
 158                 goto err;
 159
 160         err = NULL;
 161 err:
 162         return err;
 163 }
 164
 165
 166 static const char *read_super(struct cache_sb *sb, struct block_device *bdev,
 167                               struct cache_sb_disk **res)
 168 {
 169         const char *err;
 170         struct cache_sb_disk *s;
 171         struct page *page;
 172         unsigned int i;
 173
 174         page = read_cache_page_gfp(bdev->bd_inode->i_mapping,
 175                                    SB_OFFSET >> PAGE_SHIFT, GFP_KERNEL);
 176         if (IS_ERR(page))
 177                 return "IO error";
 178         s = page_address(page) + offset_in_page(SB_OFFSET);
 179
 180         sb->offset              = le64_to_cpu(s->offset);
 181         sb->version             = le64_to_cpu(s->version);
 182
 183         memcpy(sb->magic,       s->magic, 16);
 184         memcpy(sb->uuid,        s->uuid, 16);
 185         memcpy(sb->set_uuid,    s->set_uuid, 16);
 186         memcpy(sb->label,       s->label, SB_LABEL_SIZE);
 187
 188         sb->flags               = le64_to_cpu(s->flags);
 189         sb->seq                 = le64_to_cpu(s->seq);
 190         sb->last_mount          = le32_to_cpu(s->last_mount);
 191         sb->keys                = le16_to_cpu(s->keys);
 192
 193         for (i = 0; i < SB_JOURNAL_BUCKETS; i++)
 194                 sb->d[i] = le64_to_cpu(s->d[i]);
 195
 196         pr_debug("read sb version %llu, flags %llu, seq %llu, journal size %u\n",
 197                  sb->version, sb->flags, sb->seq, sb->keys);
 198
 199         err = "Not a bcache superblock (bad offset)";
 200         if (sb->offset != SB_SECTOR)
 201                 goto err;
 202
 203         err = "Not a bcache superblock (bad magic)";
 204         if (memcmp(sb->magic, bcache_magic, 16))
 205                 goto err;
 206
 207         err = "Bad checksum";
 208         if (s->csum != csum_set(s))
 209                 goto err;
 210
 211         err = "Bad UUID";
 212         if (bch_is_zero(sb->uuid, 16))
 213                 goto err;
 214
 215         sb->block_size  = le16_to_cpu(s->block_size);
 216
 217         err = "Superblock block size smaller than device block size";
 218         if (sb->block_size << 9 < bdev_logical_block_size(bdev))
 219                 goto err;
 220
 221         switch (sb->version) {
 222         case BCACHE_SB_VERSION_BDEV:
 223                 sb->data_offset = BDEV_DATA_START_DEFAULT;
 224                 break;
 225         case BCACHE_SB_VERSION_BDEV_WITH_OFFSET:
 226         case BCACHE_SB_VERSION_BDEV_WITH_FEATURES:
 227                 sb->data_offset = le64_to_cpu(s->data_offset);
 228
 229                 err = "Bad data offset";
 230                 if (sb->data_offset < BDEV_DATA_START_DEFAULT)
 231                         goto err;
 232
 233                 break;
 234         case BCACHE_SB_VERSION_CDEV:
 235         case BCACHE_SB_VERSION_CDEV_WITH_UUID:
 236                 err = read_super_common(sb, bdev, s);
 237                 if (err)
 238                         goto err;
 239                 break;
 240         case BCACHE_SB_VERSION_CDEV_WITH_FEATURES:
 241                 /*
 242                  * Feature bits are needed in read_super_common(),
 243                  * convert them firstly.
 244                  */
 245                 sb->feature_compat = le64_to_cpu(s->feature_compat);
 246                 sb->feature_incompat = le64_to_cpu(s->feature_incompat);
 247                 sb->feature_ro_compat = le64_to_cpu(s->feature_ro_compat);
 248
 249                 /* Check incompatible features */
 250                 err = "Unsupported compatible feature found";
 251                 if (bch_has_unknown_compat_features(sb))
 252                         goto err;
 253
 254                 err = "Unsupported read-only compatible feature found";
 255                 if (bch_has_unknown_ro_compat_features(sb))
 256                         goto err;
 257
 258                 err = "Unsupported incompatible feature found";
 259                 if (bch_has_unknown_incompat_features(sb))
 260                         goto err;
 261
 262                 err = read_super_common(sb, bdev, s);
 263                 if (err)
 264                         goto err;
 265                 break;
 266         default:
 267                 err = "Unsupported superblock version";
 268                 goto err;
 269         }
 270
 271         sb->last_mount = (u32)ktime_get_real_seconds();
 272         *res = s;
 273         return NULL;
 274 err:
 275         put_page(page);
 276         return err;
 277 }
 278
 279 static void write_bdev_super_endio(struct bio *bio)
 280 {
 281         struct cached_dev *dc = bio->bi_private;
 282
 283         if (bio->bi_status)
 284                 bch_count_backing_io_errors(dc, bio);
 285
 286         closure_put(&dc->sb_write);
 287 }
 288
 289 static void __write_super(struct cache_sb *sb, struct cache_sb_disk *out,
 290                 struct bio *bio)
 291 {
 292         unsigned int i;
 293
 294         bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_META;
 295         bio->bi_iter.bi_sector  = SB_SECTOR;
 296         __bio_add_page(bio, virt_to_page(out), SB_SIZE,
 297                         offset_in_page(out));
 298
 299         out->offset             = cpu_to_le64(sb->offset);
 300
 301         memcpy(out->uuid,       sb->uuid, 16);
 302         memcpy(out->set_uuid,   sb->set_uuid, 16);
 303         memcpy(out->label,      sb->label, SB_LABEL_SIZE);
 304
 305         out->flags              = cpu_to_le64(sb->flags);
 306         out->seq                = cpu_to_le64(sb->seq);
 307
 308         out->last_mount         = cpu_to_le32(sb->last_mount);
 309         out->first_bucket       = cpu_to_le16(sb->first_bucket);
 310         out->keys               = cpu_to_le16(sb->keys);
 311
 312         for (i = 0; i < sb->keys; i++)
 313                 out->d[i] = cpu_to_le64(sb->d[i]);
 314
 315         if (sb->version >= BCACHE_SB_VERSION_CDEV_WITH_FEATURES) {
 316                 out->feature_compat    = cpu_to_le64(sb->feature_compat);
 317                 out->feature_incompat  = cpu_to_le64(sb->feature_incompat);
 318                 out->feature_ro_compat = cpu_to_le64(sb->feature_ro_compat);
 319         }
 320
 321         out->version            = cpu_to_le64(sb->version);
 322         out->csum = csum_set(out);
 323
 324         pr_debug("ver %llu, flags %llu, seq %llu\n",
 325                  sb->version, sb->flags, sb->seq);
 326
 327         submit_bio(bio);
 328 }
 329
 330 static CLOSURE_CALLBACK(bch_write_bdev_super_unlock)
 331 {
 332         closure_type(dc, struct cached_dev, sb_write);
 333
 334         up(&dc->sb_write_mutex);
 335 }
 336
 337 void bch_write_bdev_super(struct cached_dev *dc, struct closure *parent)
 338 {
 339         struct closure *cl = &dc->sb_write;
 340         struct bio *bio = &dc->sb_bio;
 341
 342         down(&dc->sb_write_mutex);
 343         closure_init(cl, parent);
 344
 345         bio_init(bio, dc->bdev, dc->sb_bv, 1, 0);
 346         bio->bi_end_io  = write_bdev_super_endio;
 347         bio->bi_private = dc;
 348
 349         closure_get(cl);
 350         /* I/O request sent to backing device */
 351         __write_super(&dc->sb, dc->sb_disk, bio);
 352
 353         closure_return_with_destructor(cl, bch_write_bdev_super_unlock);
 354 }
 355
 356 static void write_super_endio(struct bio *bio)
 357 {
 358         struct cache *ca = bio->bi_private;
 359
 360         /* is_read = 0 */
 361         bch_count_io_errors(ca, bio->bi_status, 0,
 362                             "writing superblock");
 363         closure_put(&ca->set->sb_write);
 364 }
 365
 366 static CLOSURE_CALLBACK(bcache_write_super_unlock)
 367 {
 368         closure_type(c, struct cache_set, sb_write);
 369
 370         up(&c->sb_write_mutex);
 371 }
 372
 373 void bcache_write_super(struct cache_set *c)
 374 {
 375         struct closure *cl = &c->sb_write;
 376         struct cache *ca = c->cache;
 377         struct bio *bio = &ca->sb_bio;
 378         unsigned int version = BCACHE_SB_VERSION_CDEV_WITH_UUID;
 379
 380         down(&c->sb_write_mutex);
 381         closure_init(cl, &c->cl);
 382
 383         ca->sb.seq++;
 384
 385         if (ca->sb.version < version)
 386                 ca->sb.version = version;
 387
 388         bio_init(bio, ca->bdev, ca->sb_bv, 1, 0);
 389         bio->bi_end_io  = write_super_endio;
 390         bio->bi_private = ca;
 391
 392         closure_get(cl);
 393         __write_super(&ca->sb, ca->sb_disk, bio);
 394
 395         closure_return_with_destructor(cl, bcache_write_super_unlock);
 396 }
 397
 398 /* UUID io */
 399
 400 static void uuid_endio(struct bio *bio)
 401 {
 402         struct closure *cl = bio->bi_private;
 403         struct cache_set *c = container_of(cl, struct cache_set, uuid_write);
 404
 405         cache_set_err_on(bio->bi_status, c, "accessing uuids");
 406         bch_bbio_free(bio, c);
 407         closure_put(cl);
 408 }
 409
 410 static CLOSURE_CALLBACK(uuid_io_unlock)
 411 {
 412         closure_type(c, struct cache_set, uuid_write);
 413
 414         up(&c->uuid_write_mutex);
 415 }
 416
 417 static void uuid_io(struct cache_set *c, blk_opf_t opf, struct bkey *k,
 418                     struct closure *parent)
 419 {
 420         struct closure *cl = &c->uuid_write;
 421         struct uuid_entry *u;
 422         unsigned int i;
 423         char buf[80];
 424
 425         BUG_ON(!parent);
 426         down(&c->uuid_write_mutex);
 427         closure_init(cl, parent);
 428
 429         for (i = 0; i < KEY_PTRS(k); i++) {
 430                 struct bio *bio = bch_bbio_alloc(c);
 431
 432                 bio->bi_opf = opf | REQ_SYNC | REQ_META;
 433                 bio->bi_iter.bi_size = KEY_SIZE(k) << 9;
 434
 435                 bio->bi_end_io  = uuid_endio;
 436                 bio->bi_private = cl;
 437                 bch_bio_map(bio, c->uuids);
 438
 439                 bch_submit_bbio(bio, c, k, i);
 440
 441                 if ((opf & REQ_OP_MASK) != REQ_OP_WRITE)
 442                         break;
 443         }
 444
 445         bch_extent_to_text(buf, sizeof(buf), k);
 446         pr_debug("%s UUIDs at %s\n", (opf & REQ_OP_MASK) == REQ_OP_WRITE ?
 447                  "wrote" : "read", buf);
 448
 449         for (u = c->uuids; u < c->uuids + c->nr_uuids; u++)
 450                 if (!bch_is_zero(u->uuid, 16))
 451                         pr_debug("Slot %zi: %pU: %s: 1st: %u last: %u inv: %u\n",
 452                                  u - c->uuids, u->uuid, u->label,
 453                                  u->first_reg, u->last_reg, u->invalidated);
 454
 455         closure_return_with_destructor(cl, uuid_io_unlock);
 456 }
 457
 458 static char *uuid_read(struct cache_set *c, struct jset *j, struct closure *cl)
 459 {
 460         struct bkey *k = &j->uuid_bucket;
 461
 462         if (__bch_btree_ptr_invalid(c, k))
 463                 return "bad uuid pointer";
 464
 465         bkey_copy(&c->uuid_bucket, k);
 466         uuid_io(c, REQ_OP_READ, k, cl);
 467
 468         if (j->version < BCACHE_JSET_VERSION_UUIDv1) {
 469                 struct uuid_entry_v0    *u0 = (void *) c->uuids;
 470                 struct uuid_entry       *u1 = (void *) c->uuids;
 471                 int i;
 472
 473                 closure_sync(cl);
 474
 475                 /*
 476                  * Since the new uuid entry is bigger than the old, we have to
 477                  * convert starting at the highest memory address and work down
 478                  * in order to do it in place
 479                  */
 480
 481                 for (i = c->nr_uuids - 1;
 482                      i >= 0;
 483                      --i) {
 484                         memcpy(u1[i].uuid,      u0[i].uuid, 16);
 485                         memcpy(u1[i].label,     u0[i].label, 32);
 486
 487                         u1[i].first_reg         = u0[i].first_reg;
 488                         u1[i].last_reg          = u0[i].last_reg;
 489                         u1[i].invalidated       = u0[i].invalidated;
 490
 491                         u1[i].flags     = 0;
 492                         u1[i].sectors   = 0;
 493                 }
 494         }
 495
 496         return NULL;
 497 }
 498
 499 static int __uuid_write(struct cache_set *c)
 500 {
 501         BKEY_PADDED(key) k;
 502         struct closure cl;
 503         struct cache *ca = c->cache;
 504         unsigned int size;
 505
 506         closure_init_stack(&cl);
 507         lockdep_assert_held(&bch_register_lock);
 508
 509         if (bch_bucket_alloc_set(c, RESERVE_BTREE, &k.key, true))
 510                 return 1;
 511
 512         size =  meta_bucket_pages(&ca->sb) * PAGE_SECTORS;
 513         SET_KEY_SIZE(&k.key, size);
 514         uuid_io(c, REQ_OP_WRITE, &k.key, &cl);
 515         closure_sync(&cl);
 516
 517         /* Only one bucket used for uuid write */
 518         atomic_long_add(ca->sb.bucket_size, &ca->meta_sectors_written);
 519
 520         bkey_copy(&c->uuid_bucket, &k.key);
 521         bkey_put(c, &k.key);
 522         return 0;
 523 }
 524
 525 int bch_uuid_write(struct cache_set *c)
 526 {
 527         int ret = __uuid_write(c);
 528
 529         if (!ret)
 530                 bch_journal_meta(c, NULL);
 531
 532         return ret;
 533 }
 534
 535 static struct uuid_entry *uuid_find(struct cache_set *c, const char *uuid)
 536 {
 537         struct uuid_entry *u;
 538
 539         for (u = c->uuids;
 540              u < c->uuids + c->nr_uuids; u++)
 541                 if (!memcmp(u->uuid, uuid, 16))
 542                         return u;
 543
 544         return NULL;
 545 }
 546
 547 static struct uuid_entry *uuid_find_empty(struct cache_set *c)
 548 {
 549         static const char zero_uuid[16] = "\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0\0";
 550
 551         return uuid_find(c, zero_uuid);
 552 }
 553
 554 /*
 555  * Bucket priorities/gens:
 556  *
 557  * For each bucket, we store on disk its
 558  *   8 bit gen
 559  *  16 bit priority
 560  *
 561  * See alloc.c for an explanation of the gen. The priority is used to implement
 562  * lru (and in the future other) cache replacement policies; for most purposes
 563  * it's just an opaque integer.
 564  *
 565  * The gens and the priorities don't have a whole lot to do with each other, and
 566  * it's actually the gens that must be written out at specific times - it's no
 567  * big deal if the priorities don't get written, if we lose them we just reuse
 568  * buckets in suboptimal order.
 569  *
 570  * On disk they're stored in a packed array, and in as many buckets are required
 571  * to fit them all. The buckets we use to store them form a list; the journal
 572  * header points to the first bucket, the first bucket points to the second
 573  * bucket, et cetera.
 574  *
 575  * This code is used by the allocation code; periodically (whenever it runs out
 576  * of buckets to allocate from) the allocation code will invalidate some
 577  * buckets, but it can't use those buckets until their new gens are safely on
 578  * disk.
 579  */
 580
 581 static void prio_endio(struct bio *bio)
 582 {
 583         struct cache *ca = bio->bi_private;
 584
 585         cache_set_err_on(bio->bi_status, ca->set, "accessing priorities");
 586         bch_bbio_free(bio, ca->set);
 587         closure_put(&ca->prio);
 588 }
 589
 590 static void prio_io(struct cache *ca, uint64_t bucket, blk_opf_t opf)
 591 {
 592         struct closure *cl = &ca->prio;
 593         struct bio *bio = bch_bbio_alloc(ca->set);
 594
 595         closure_init_stack(cl);
 596
 597         bio->bi_iter.bi_sector  = bucket * ca->sb.bucket_size;
 598         bio_set_dev(bio, ca->bdev);
 599         bio->bi_iter.bi_size    = meta_bucket_bytes(&ca->sb);
 600
 601         bio->bi_end_io  = prio_endio;
 602         bio->bi_private = ca;
 603         bio->bi_opf = opf | REQ_SYNC | REQ_META;
 604         bch_bio_map(bio, ca->disk_buckets);
 605
 606         closure_bio_submit(ca->set, bio, &ca->prio);
 607         closure_sync(cl);
 608 }
 609
 610 int bch_prio_write(struct cache *ca, bool wait)
 611 {
 612         int i;
 613         struct bucket *b;
 614         struct closure cl;
 615
 616         pr_debug("free_prio=%zu, free_none=%zu, free_inc=%zu\n",
 617                  fifo_used(&ca->free[RESERVE_PRIO]),
 618                  fifo_used(&ca->free[RESERVE_NONE]),
 619                  fifo_used(&ca->free_inc));
 620
 621         /*
 622          * Pre-check if there are enough free buckets. In the non-blocking
 623          * scenario it's better to fail early rather than starting to allocate
 624          * buckets and do a cleanup later in case of failure.
 625          */
 626         if (!wait) {
 627                 size_t avail = fifo_used(&ca->free[RESERVE_PRIO]) +
 628                                fifo_used(&ca->free[RESERVE_NONE]);
 629                 if (prio_buckets(ca) > avail)
 630                         return -ENOMEM;
 631         }
 632
 633         closure_init_stack(&cl);
 634
 635         lockdep_assert_held(&ca->set->bucket_lock);
 636
 637         ca->disk_buckets->seq++;
 638
 639         atomic_long_add(ca->sb.bucket_size * prio_buckets(ca),
 640                         &ca->meta_sectors_written);
 641
 642         for (i = prio_buckets(ca) - 1; i >= 0; --i) {
 643                 long bucket;
 644                 struct prio_set *p = ca->disk_buckets;
 645                 struct bucket_disk *d = p->data;
 646                 struct bucket_disk *end = d + prios_per_bucket(ca);
 647
 648                 for (b = ca->buckets + i * prios_per_bucket(ca);
 649                      b < ca->buckets + ca->sb.nbuckets && d < end;
 650                      b++, d++) {
 651                         d->prio = cpu_to_le16(b->prio);
 652                         d->gen = b->gen;
 653                 }
 654
 655                 p->next_bucket  = ca->prio_buckets[i + 1];
 656                 p->magic        = pset_magic(&ca->sb);
 657                 p->csum         = bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8);
 658
 659                 bucket = bch_bucket_alloc(ca, RESERVE_PRIO, wait);
 660                 BUG_ON(bucket == -1);
 661
 662                 mutex_unlock(&ca->set->bucket_lock);
 663                 prio_io(ca, bucket, REQ_OP_WRITE);
 664                 mutex_lock(&ca->set->bucket_lock);
 665
 666                 ca->prio_buckets[i] = bucket;
 667                 atomic_dec_bug(&ca->buckets[bucket].pin);
 668         }
 669
 670         mutex_unlock(&ca->set->bucket_lock);
 671
 672         bch_journal_meta(ca->set, &cl);
 673         closure_sync(&cl);
 674
 675         mutex_lock(&ca->set->bucket_lock);
 676
 677         /*
 678          * Don't want the old priorities to get garbage collected until after we
 679          * finish writing the new ones, and they're journalled
 680          */
 681         for (i = 0; i < prio_buckets(ca); i++) {
 682                 if (ca->prio_last_buckets[i])
 683                         __bch_bucket_free(ca,
 684                                 &ca->buckets[ca->prio_last_buckets[i]]);
 685
 686                 ca->prio_last_buckets[i] = ca->prio_buckets[i];
 687         }
 688         return 0;
 689 }
 690
 691 static int prio_read(struct cache *ca, uint64_t bucket)
 692 {
 693         struct prio_set *p = ca->disk_buckets;
 694         struct bucket_disk *d = p->data + prios_per_bucket(ca), *end = d;
 695         struct bucket *b;
 696         unsigned int bucket_nr = 0;
 697         int ret = -EIO;
 698
 699         for (b = ca->buckets;
 700              b < ca->buckets + ca->sb.nbuckets;
 701              b++, d++) {
 702                 if (d == end) {
 703                         ca->prio_buckets[bucket_nr] = bucket;
 704                         ca->prio_last_buckets[bucket_nr] = bucket;
 705                         bucket_nr++;
 706
 707                         prio_io(ca, bucket, REQ_OP_READ);
 708
 709                         if (p->csum !=
 710                             bch_crc64(&p->magic, meta_bucket_bytes(&ca->sb) - 8)) {
 711                                 pr_warn("bad csum reading priorities\n");
 712                                 goto out;
 713                         }
 714
 715                         if (p->magic != pset_magic(&ca->sb)) {
 716                                 pr_warn("bad magic reading priorities\n");
 717                                 goto out;
 718                         }
 719
 720                         bucket = p->next_bucket;
 721                         d = p->data;
 722                 }
 723
 724                 b->prio = le16_to_cpu(d->prio);
 725                 b->gen = b->last_gc = d->gen;
 726         }
 727
 728         ret = 0;
 729 out:
 730         return ret;
 731 }
 732
 733 /* Bcache device */
 734
 735 static int open_dev(struct gendisk *disk, blk_mode_t mode)
 736 {
 737         struct bcache_device *d = disk->private_data;
 738
 739         if (test_bit(BCACHE_DEV_CLOSING, &d->flags))
 740                 return -ENXIO;
 741
 742         closure_get(&d->cl);
 743         return 0;
 744 }
 745
 746 static void release_dev(struct gendisk *b)
 747 {
 748         struct bcache_device *d = b->private_data;
 749
 750         closure_put(&d->cl);
 751 }
 752
 753 static int ioctl_dev(struct block_device *b, blk_mode_t mode,
 754                      unsigned int cmd, unsigned long arg)
 755 {
 756         struct bcache_device *d = b->bd_disk->private_data;
 757
 758         return d->ioctl(d, mode, cmd, arg);
 759 }
 760
 761 static const struct block_device_operations bcache_cached_ops = {
 762         .submit_bio     = cached_dev_submit_bio,
 763         .open           = open_dev,
 764         .release        = release_dev,
 765         .ioctl          = ioctl_dev,
 766         .owner          = THIS_MODULE,
 767 };
 768
 769 static const struct block_device_operations bcache_flash_ops = {
 770         .submit_bio     = flash_dev_submit_bio,
 771         .open           = open_dev,
 772         .release        = release_dev,
 773         .ioctl          = ioctl_dev,
 774         .owner          = THIS_MODULE,
 775 };
 776
 777 void bcache_device_stop(struct bcache_device *d)
 778 {
 779         if (!test_and_set_bit(BCACHE_DEV_CLOSING, &d->flags))
 780                 /*
 781                  * closure_fn set to
 782                  * - cached device: cached_dev_flush()
 783                  * - flash dev: flash_dev_flush()
 784                  */
 785                 closure_queue(&d->cl);
 786 }
 787
 788 static void bcache_device_unlink(struct bcache_device *d)
 789 {
 790         lockdep_assert_held(&bch_register_lock);
 791
 792         if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
 793                 struct cache *ca = d->c->cache;
 794
 795                 sysfs_remove_link(&d->c->kobj, d->name);
 796                 sysfs_remove_link(&d->kobj, "cache");
 797
 798                 bd_unlink_disk_holder(ca->bdev, d->disk);
 799         }
 800 }
 801
 802 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
 803                                const char *name)
 804 {
 805         struct cache *ca = c->cache;
 806         int ret;
 807
 808         bd_link_disk_holder(ca->bdev, d->disk);
 809
 810         snprintf(d->name, BCACHEDEVNAME_SIZE,
 811                  "%s%u", name, d->id);
 812
 813         ret = sysfs_create_link(&d->kobj, &c->kobj, "cache");
 814         if (ret < 0)
 815                 pr_err("Couldn't create device -> cache set symlink\n");
 816
 817         ret = sysfs_create_link(&c->kobj, &d->kobj, d->name);
 818         if (ret < 0)
 819                 pr_err("Couldn't create cache set -> device symlink\n");
 820
 821         clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
 822 }
 823
 824 static void bcache_device_detach(struct bcache_device *d)
 825 {
 826         lockdep_assert_held(&bch_register_lock);
 827
 828         atomic_dec(&d->c->attached_dev_nr);
 829
 830         if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
 831                 struct uuid_entry *u = d->c->uuids + d->id;
 832
 833                 SET_UUID_FLASH_ONLY(u, 0);
 834                 memcpy(u->uuid, invalid_uuid, 16);
 835                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
 836                 bch_uuid_write(d->c);
 837         }
 838
 839         bcache_device_unlink(d);
 840
 841         d->c->devices[d->id] = NULL;
 842         closure_put(&d->c->caching);
 843         d->c = NULL;
 844 }
 845
 846 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
 847                                  unsigned int id)
 848 {
 849         d->id = id;
 850         d->c = c;
 851         c->devices[id] = d;
 852
 853         if (id >= c->devices_max_used)
 854                 c->devices_max_used = id + 1;
 855
 856         closure_get(&c->caching);
 857 }
 858
 859 static inline int first_minor_to_idx(int first_minor)
 860 {
 861         return (first_minor/BCACHE_MINORS);
 862 }
 863
 864 static inline int idx_to_first_minor(int idx)
 865 {
 866         return (idx * BCACHE_MINORS);
 867 }
 868
 869 static void bcache_device_free(struct bcache_device *d)
 870 {
 871         struct gendisk *disk = d->disk;
 872
 873         lockdep_assert_held(&bch_register_lock);
 874
 875         if (disk)
 876                 pr_info("%s stopped\n", disk->disk_name);
 877         else
 878                 pr_err("bcache device (NULL gendisk) stopped\n");
 879
 880         if (d->c)
 881                 bcache_device_detach(d);
 882
 883         if (disk) {
 884                 ida_simple_remove(&bcache_device_idx,
 885                                   first_minor_to_idx(disk->first_minor));
 886                 put_disk(disk);
 887         }
 888
 889         bioset_exit(&d->bio_split);
 890         kvfree(d->full_dirty_stripes);
 891         kvfree(d->stripe_sectors_dirty);
 892
 893         closure_debug_destroy(&d->cl);
 894 }
 895
 896 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
 897                 sector_t sectors, struct block_device *cached_bdev,
 898                 const struct block_device_operations *ops)
 899 {
 900         struct request_queue *q;
 901         const size_t max_stripes = min_t(size_t, INT_MAX,
 902                                          SIZE_MAX / sizeof(atomic_t));
 903         struct queue_limits lim = {
 904                 .max_hw_sectors         = UINT_MAX,
 905                 .max_sectors            = UINT_MAX,
 906                 .max_segment_size       = UINT_MAX,
 907                 .max_segments           = BIO_MAX_VECS,
 908                 .max_hw_discard_sectors = UINT_MAX,
 909                 .io_min                 = block_size,
 910                 .logical_block_size     = block_size,
 911                 .physical_block_size    = block_size,
 912         };
 913         uint64_t n;
 914         int idx;
 915
 916         if (cached_bdev) {
 917                 d->stripe_size = bdev_io_opt(cached_bdev) >> SECTOR_SHIFT;
 918                 lim.io_opt = umax(block_size, bdev_io_opt(cached_bdev));
 919         }
 920         if (!d->stripe_size)
 921                 d->stripe_size = 1 << 31;
 922         else if (d->stripe_size < BCH_MIN_STRIPE_SZ)
 923                 d->stripe_size = roundup(BCH_MIN_STRIPE_SZ, d->stripe_size);
 924
 925         n = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
 926         if (!n || n > max_stripes) {
 927                 pr_err("nr_stripes too large or invalid: %llu (start sector beyond end of disk?)\n",
 928                         n);
 929                 return -ENOMEM;
 930         }
 931         d->nr_stripes = n;
 932
 933         n = d->nr_stripes * sizeof(atomic_t);
 934         d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
 935         if (!d->stripe_sectors_dirty)
 936                 return -ENOMEM;
 937
 938         n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
 939         d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
 940         if (!d->full_dirty_stripes)
 941                 goto out_free_stripe_sectors_dirty;
 942
 943         idx = ida_simple_get(&bcache_device_idx, 0,
 944                                 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
 945         if (idx < 0)
 946                 goto out_free_full_dirty_stripes;
 947
 948         if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
 949                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
 950                 goto out_ida_remove;
 951
 952         if (lim.logical_block_size > PAGE_SIZE && cached_bdev) {
 953                 /*
 954                  * This should only happen with BCACHE_SB_VERSION_BDEV.
 955                  * Block/page size is checked for BCACHE_SB_VERSION_CDEV.
 956                  */
 957                 pr_info("bcache%i: sb/logical block size (%u) greater than page size (%lu) falling back to device logical block size (%u)\n",
 958                         idx, lim.logical_block_size,
 959                         PAGE_SIZE, bdev_logical_block_size(cached_bdev));
 960
 961                 /* This also adjusts physical block size/min io size if needed */
 962                 lim.logical_block_size = bdev_logical_block_size(cached_bdev);
 963         }
 964
 965         d->disk = blk_alloc_disk(&lim, NUMA_NO_NODE);
 966         if (IS_ERR(d->disk))
 967                 goto out_bioset_exit;
 968
 969         set_capacity(d->disk, sectors);
 970         snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
 971
 972         d->disk->major          = bcache_major;
 973         d->disk->first_minor    = idx_to_first_minor(idx);
 974         d->disk->minors         = BCACHE_MINORS;
 975         d->disk->fops           = ops;
 976         d->disk->private_data   = d;
 977
 978         q = d->disk->queue;
 979
 980         blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
 981
 982         blk_queue_write_cache(q, true, true);
 983
 984         return 0;
 985
 986 out_bioset_exit:
 987         bioset_exit(&d->bio_split);
 988 out_ida_remove:
 989         ida_simple_remove(&bcache_device_idx, idx);
 990 out_free_full_dirty_stripes:
 991         kvfree(d->full_dirty_stripes);
 992 out_free_stripe_sectors_dirty:
 993         kvfree(d->stripe_sectors_dirty);
 994         return -ENOMEM;
 995
 996 }
 997
 998 /* Cached device */
 999
1000 static void calc_cached_dev_sectors(struct cache_set *c)
1001 {
1002         uint64_t sectors = 0;
1003         struct cached_dev *dc;
1004
1005         list_for_each_entry(dc, &c->cached_devs, list)
1006                 sectors += bdev_nr_sectors(dc->bdev);
1007
1008         c->cached_dev_sectors = sectors;
1009 }
1010
1011 #define BACKING_DEV_OFFLINE_TIMEOUT 5
1012 static int cached_dev_status_update(void *arg)
1013 {
1014         struct cached_dev *dc = arg;
1015         struct request_queue *q;
1016
1017         /*
1018          * If this delayed worker is stopping outside, directly quit here.
1019          * dc->io_disable might be set via sysfs interface, so check it
1020          * here too.
1021          */
1022         while (!kthread_should_stop() && !dc->io_disable) {
1023                 q = bdev_get_queue(dc->bdev);
1024                 if (blk_queue_dying(q))
1025                         dc->offline_seconds++;
1026                 else
1027                         dc->offline_seconds = 0;
1028
1029                 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
1030                         pr_err("%pg: device offline for %d seconds\n",
1031                                dc->bdev,
1032                                BACKING_DEV_OFFLINE_TIMEOUT);
1033                         pr_err("%s: disable I/O request due to backing device offline\n",
1034                                dc->disk.name);
1035                         dc->io_disable = true;
1036                         /* let others know earlier that io_disable is true */
1037                         smp_mb();
1038                         bcache_device_stop(&dc->disk);
1039                         break;
1040                 }
1041                 schedule_timeout_interruptible(HZ);
1042         }
1043
1044         wait_for_kthread_stop();
1045         return 0;
1046 }
1047
1048
1049 int bch_cached_dev_run(struct cached_dev *dc)
1050 {
1051         int ret = 0;
1052         struct bcache_device *d = &dc->disk;
1053         char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
1054         char *env[] = {
1055                 "DRIVER=bcache",
1056                 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
1057                 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
1058                 NULL,
1059         };
1060
1061         if (dc->io_disable) {
1062                 pr_err("I/O disabled on cached dev %pg\n", dc->bdev);
1063                 ret = -EIO;
1064                 goto out;
1065         }
1066
1067         if (atomic_xchg(&dc->running, 1)) {
1068                 pr_info("cached dev %pg is running already\n", dc->bdev);
1069                 ret = -EBUSY;
1070                 goto out;
1071         }
1072
1073         if (!d->c &&
1074             BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
1075                 struct closure cl;
1076
1077                 closure_init_stack(&cl);
1078
1079                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
1080                 bch_write_bdev_super(dc, &cl);
1081                 closure_sync(&cl);
1082         }
1083
1084         ret = add_disk(d->disk);
1085         if (ret)
1086                 goto out;
1087         bd_link_disk_holder(dc->bdev, dc->disk.disk);
1088         /*
1089          * won't show up in the uevent file, use udevadm monitor -e instead
1090          * only class / kset properties are persistent
1091          */
1092         kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
1093
1094         if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
1095             sysfs_create_link(&disk_to_dev(d->disk)->kobj,
1096                               &d->kobj, "bcache")) {
1097                 pr_err("Couldn't create bcache dev <-> disk sysfs symlinks\n");
1098                 ret = -ENOMEM;
1099                 goto out;
1100         }
1101
1102         dc->status_update_thread = kthread_run(cached_dev_status_update,
1103                                                dc, "bcache_status_update");
1104         if (IS_ERR(dc->status_update_thread)) {
1105                 pr_warn("failed to create bcache_status_update kthread, continue to run without monitoring backing device status\n");
1106         }
1107
1108 out:
1109         kfree(env[1]);
1110         kfree(env[2]);
1111         kfree(buf);
1112         return ret;
1113 }
1114
1115 /*
1116  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
1117  * work dc->writeback_rate_update is running. Wait until the routine
1118  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
1119  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
1120  * seconds, give up waiting here and continue to cancel it too.
1121  */
1122 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
1123 {
1124         int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
1125
1126         do {
1127                 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
1128                               &dc->disk.flags))
1129                         break;
1130                 time_out--;
1131                 schedule_timeout_interruptible(1);
1132         } while (time_out > 0);
1133
1134         if (time_out == 0)
1135                 pr_warn("give up waiting for dc->writeback_write_update to quit\n");
1136
1137         cancel_delayed_work_sync(&dc->writeback_rate_update);
1138 }
1139
1140 static void cached_dev_detach_finish(struct work_struct *w)
1141 {
1142         struct cached_dev *dc = container_of(w, struct cached_dev, detach);
1143         struct cache_set *c = dc->disk.c;
1144
1145         BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
1146         BUG_ON(refcount_read(&dc->count));
1147
1148
1149         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1150                 cancel_writeback_rate_update_dwork(dc);
1151
1152         if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1153                 kthread_stop(dc->writeback_thread);
1154                 dc->writeback_thread = NULL;
1155         }
1156
1157         mutex_lock(&bch_register_lock);
1158
1159         bcache_device_detach(&dc->disk);
1160         list_move(&dc->list, &uncached_devices);
1161         calc_cached_dev_sectors(c);
1162
1163         clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1164         clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1165
1166         mutex_unlock(&bch_register_lock);
1167
1168         pr_info("Caching disabled for %pg\n", dc->bdev);
1169
1170         /* Drop ref we took in cached_dev_detach() */
1171         closure_put(&dc->disk.cl);
1172 }
1173
1174 void bch_cached_dev_detach(struct cached_dev *dc)
1175 {
1176         lockdep_assert_held(&bch_register_lock);
1177
1178         if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1179                 return;
1180
1181         if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1182                 return;
1183
1184         /*
1185          * Block the device from being closed and freed until we're finished
1186          * detaching
1187          */
1188         closure_get(&dc->disk.cl);
1189
1190         bch_writeback_queue(dc);
1191
1192         cached_dev_put(dc);
1193 }
1194
1195 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1196                           uint8_t *set_uuid)
1197 {
1198         uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1199         struct uuid_entry *u;
1200         struct cached_dev *exist_dc, *t;
1201         int ret = 0;
1202
1203         if ((set_uuid && memcmp(set_uuid, c->set_uuid, 16)) ||
1204             (!set_uuid && memcmp(dc->sb.set_uuid, c->set_uuid, 16)))
1205                 return -ENOENT;
1206
1207         if (dc->disk.c) {
1208                 pr_err("Can't attach %pg: already attached\n", dc->bdev);
1209                 return -EINVAL;
1210         }
1211
1212         if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1213                 pr_err("Can't attach %pg: shutting down\n", dc->bdev);
1214                 return -EINVAL;
1215         }
1216
1217         if (dc->sb.block_size < c->cache->sb.block_size) {
1218                 /* Will die */
1219                 pr_err("Couldn't attach %pg: block size less than set's block size\n",
1220                        dc->bdev);
1221                 return -EINVAL;
1222         }
1223
1224         /* Check whether already attached */
1225         list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1226                 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1227                         pr_err("Tried to attach %pg but duplicate UUID already attached\n",
1228                                 dc->bdev);
1229
1230                         return -EINVAL;
1231                 }
1232         }
1233
1234         u = uuid_find(c, dc->sb.uuid);
1235
1236         if (u &&
1237             (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1238              BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1239                 memcpy(u->uuid, invalid_uuid, 16);
1240                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1241                 u = NULL;
1242         }
1243
1244         if (!u) {
1245                 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1246                         pr_err("Couldn't find uuid for %pg in set\n", dc->bdev);
1247                         return -ENOENT;
1248                 }
1249
1250                 u = uuid_find_empty(c);
1251                 if (!u) {
1252                         pr_err("Not caching %pg, no room for UUID\n", dc->bdev);
1253                         return -EINVAL;
1254                 }
1255         }
1256
1257         /*
1258          * Deadlocks since we're called via sysfs...
1259          * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1260          */
1261
1262         if (bch_is_zero(u->uuid, 16)) {
1263                 struct closure cl;
1264
1265                 closure_init_stack(&cl);
1266
1267                 memcpy(u->uuid, dc->sb.uuid, 16);
1268                 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1269                 u->first_reg = u->last_reg = rtime;
1270                 bch_uuid_write(c);
1271
1272                 memcpy(dc->sb.set_uuid, c->set_uuid, 16);
1273                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1274
1275                 bch_write_bdev_super(dc, &cl);
1276                 closure_sync(&cl);
1277         } else {
1278                 u->last_reg = rtime;
1279                 bch_uuid_write(c);
1280         }
1281
1282         bcache_device_attach(&dc->disk, c, u - c->uuids);
1283         list_move(&dc->list, &c->cached_devs);
1284         calc_cached_dev_sectors(c);
1285
1286         /*
1287          * dc->c must be set before dc->count != 0 - paired with the mb in
1288          * cached_dev_get()
1289          */
1290         smp_wmb();
1291         refcount_set(&dc->count, 1);
1292
1293         /* Block writeback thread, but spawn it */
1294         down_write(&dc->writeback_lock);
1295         if (bch_cached_dev_writeback_start(dc)) {
1296                 up_write(&dc->writeback_lock);
1297                 pr_err("Couldn't start writeback facilities for %s\n",
1298                        dc->disk.disk->disk_name);
1299                 return -ENOMEM;
1300         }
1301
1302         if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1303                 atomic_set(&dc->has_dirty, 1);
1304                 bch_writeback_queue(dc);
1305         }
1306
1307         bch_sectors_dirty_init(&dc->disk);
1308
1309         ret = bch_cached_dev_run(dc);
1310         if (ret && (ret != -EBUSY)) {
1311                 up_write(&dc->writeback_lock);
1312                 /*
1313                  * bch_register_lock is held, bcache_device_stop() is not
1314                  * able to be directly called. The kthread and kworker
1315                  * created previously in bch_cached_dev_writeback_start()
1316                  * have to be stopped manually here.
1317                  */
1318                 kthread_stop(dc->writeback_thread);
1319                 cancel_writeback_rate_update_dwork(dc);
1320                 pr_err("Couldn't run cached device %pg\n", dc->bdev);
1321                 return ret;
1322         }
1323
1324         bcache_device_link(&dc->disk, c, "bdev");
1325         atomic_inc(&c->attached_dev_nr);
1326
1327         if (bch_has_feature_obso_large_bucket(&(c->cache->sb))) {
1328                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1329                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1330                 set_disk_ro(dc->disk.disk, 1);
1331         }
1332
1333         /* Allow the writeback thread to proceed */
1334         up_write(&dc->writeback_lock);
1335
1336         pr_info("Caching %pg as %s on set %pU\n",
1337                 dc->bdev,
1338                 dc->disk.disk->disk_name,
1339                 dc->disk.c->set_uuid);
1340         return 0;
1341 }
1342
1343 /* when dc->disk.kobj released */
1344 void bch_cached_dev_release(struct kobject *kobj)
1345 {
1346         struct cached_dev *dc = container_of(kobj, struct cached_dev,
1347                                              disk.kobj);
1348         kfree(dc);
1349         module_put(THIS_MODULE);
1350 }
1351
1352 static CLOSURE_CALLBACK(cached_dev_free)
1353 {
1354         closure_type(dc, struct cached_dev, disk.cl);
1355
1356         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1357                 cancel_writeback_rate_update_dwork(dc);
1358
1359         if (!IS_ERR_OR_NULL(dc->writeback_thread))
1360                 kthread_stop(dc->writeback_thread);
1361         if (!IS_ERR_OR_NULL(dc->status_update_thread))
1362                 kthread_stop(dc->status_update_thread);
1363
1364         mutex_lock(&bch_register_lock);
1365
1366         if (atomic_read(&dc->running)) {
1367                 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1368                 del_gendisk(dc->disk.disk);
1369         }
1370         bcache_device_free(&dc->disk);
1371         list_del(&dc->list);
1372
1373         mutex_unlock(&bch_register_lock);
1374
1375         if (dc->sb_disk)
1376                 put_page(virt_to_page(dc->sb_disk));
1377
1378         if (dc->bdev_file)
1379                 fput(dc->bdev_file);
1380
1381         wake_up(&unregister_wait);
1382
1383         kobject_put(&dc->disk.kobj);
1384 }
1385
1386 static CLOSURE_CALLBACK(cached_dev_flush)
1387 {
1388         closure_type(dc, struct cached_dev, disk.cl);
1389         struct bcache_device *d = &dc->disk;
1390
1391         mutex_lock(&bch_register_lock);
1392         bcache_device_unlink(d);
1393         mutex_unlock(&bch_register_lock);
1394
1395         bch_cache_accounting_destroy(&dc->accounting);
1396         kobject_del(&d->kobj);
1397
1398         continue_at(cl, cached_dev_free, system_wq);
1399 }
1400
1401 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1402 {
1403         int ret;
1404         struct io *io;
1405         struct request_queue *q = bdev_get_queue(dc->bdev);
1406
1407         __module_get(THIS_MODULE);
1408         INIT_LIST_HEAD(&dc->list);
1409         closure_init(&dc->disk.cl, NULL);
1410         set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1411         kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1412         INIT_WORK(&dc->detach, cached_dev_detach_finish);
1413         sema_init(&dc->sb_write_mutex, 1);
1414         INIT_LIST_HEAD(&dc->io_lru);
1415         spin_lock_init(&dc->io_lock);
1416         bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1417
1418         dc->sequential_cutoff           = 4 << 20;
1419
1420         for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1421                 list_add(&io->lru, &dc->io_lru);
1422                 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1423         }
1424
1425         if (bdev_io_opt(dc->bdev))
1426                 dc->partial_stripes_expensive =
1427                         q->limits.raid_partial_stripes_expensive;
1428
1429         ret = bcache_device_init(&dc->disk, block_size,
1430                          bdev_nr_sectors(dc->bdev) - dc->sb.data_offset,
1431                          dc->bdev, &bcache_cached_ops);
1432         if (ret)
1433                 return ret;
1434
1435         atomic_set(&dc->io_errors, 0);
1436         dc->io_disable = false;
1437         dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1438         /* default to auto */
1439         dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1440
1441         bch_cached_dev_request_init(dc);
1442         bch_cached_dev_writeback_init(dc);
1443         return 0;
1444 }
1445
1446 /* Cached device - bcache superblock */
1447
1448 static int register_bdev(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
1449                                  struct file *bdev_file,
1450                                  struct cached_dev *dc)
1451 {
1452         const char *err = "cannot allocate memory";
1453         struct cache_set *c;
1454         int ret = -ENOMEM;
1455
1456         memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1457         dc->bdev_file = bdev_file;
1458         dc->bdev = file_bdev(bdev_file);
1459         dc->sb_disk = sb_disk;
1460
1461         if (cached_dev_init(dc, sb->block_size << 9))
1462                 goto err;
1463
1464         err = "error creating kobject";
1465         if (kobject_add(&dc->disk.kobj, bdev_kobj(dc->bdev), "bcache"))
1466                 goto err;
1467         if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1468                 goto err;
1469
1470         pr_info("registered backing device %pg\n", dc->bdev);
1471
1472         list_add(&dc->list, &uncached_devices);
1473         /* attach to a matched cache set if it exists */
1474         list_for_each_entry(c, &bch_cache_sets, list)
1475                 bch_cached_dev_attach(dc, c, NULL);
1476
1477         if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1478             BDEV_STATE(&dc->sb) == BDEV_STATE_STALE) {
1479                 err = "failed to run cached device";
1480                 ret = bch_cached_dev_run(dc);
1481                 if (ret)
1482                         goto err;
1483         }
1484
1485         return 0;
1486 err:
1487         pr_notice("error %pg: %s\n", dc->bdev, err);
1488         bcache_device_stop(&dc->disk);
1489         return ret;
1490 }
1491
1492 /* Flash only volumes */
1493
1494 /* When d->kobj released */
1495 void bch_flash_dev_release(struct kobject *kobj)
1496 {
1497         struct bcache_device *d = container_of(kobj, struct bcache_device,
1498                                                kobj);
1499         kfree(d);
1500 }
1501
1502 static CLOSURE_CALLBACK(flash_dev_free)
1503 {
1504         closure_type(d, struct bcache_device, cl);
1505
1506         mutex_lock(&bch_register_lock);
1507         atomic_long_sub(bcache_dev_sectors_dirty(d),
1508                         &d->c->flash_dev_dirty_sectors);
1509         del_gendisk(d->disk);
1510         bcache_device_free(d);
1511         mutex_unlock(&bch_register_lock);
1512         kobject_put(&d->kobj);
1513 }
1514
1515 static CLOSURE_CALLBACK(flash_dev_flush)
1516 {
1517         closure_type(d, struct bcache_device, cl);
1518
1519         mutex_lock(&bch_register_lock);
1520         bcache_device_unlink(d);
1521         mutex_unlock(&bch_register_lock);
1522         kobject_del(&d->kobj);
1523         continue_at(cl, flash_dev_free, system_wq);
1524 }
1525
1526 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1527 {
1528         int err = -ENOMEM;
1529         struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1530                                           GFP_KERNEL);
1531         if (!d)
1532                 goto err_ret;
1533
1534         closure_init(&d->cl, NULL);
1535         set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1536
1537         kobject_init(&d->kobj, &bch_flash_dev_ktype);
1538
1539         if (bcache_device_init(d, block_bytes(c->cache), u->sectors,
1540                         NULL, &bcache_flash_ops))
1541                 goto err;
1542
1543         bcache_device_attach(d, c, u - c->uuids);
1544         bch_sectors_dirty_init(d);
1545         bch_flash_dev_request_init(d);
1546         err = add_disk(d->disk);
1547         if (err)
1548                 goto err;
1549
1550         err = kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache");
1551         if (err)
1552                 goto err;
1553
1554         bcache_device_link(d, c, "volume");
1555
1556         if (bch_has_feature_obso_large_bucket(&c->cache->sb)) {
1557                 pr_err("The obsoleted large bucket layout is unsupported, set the bcache device into read-only\n");
1558                 pr_err("Please update to the latest bcache-tools to create the cache device\n");
1559                 set_disk_ro(d->disk, 1);
1560         }
1561
1562         return 0;
1563 err:
1564         kobject_put(&d->kobj);
1565 err_ret:
1566         return err;
1567 }
1568
1569 static int flash_devs_run(struct cache_set *c)
1570 {
1571         int ret = 0;
1572         struct uuid_entry *u;
1573
1574         for (u = c->uuids;
1575              u < c->uuids + c->nr_uuids && !ret;
1576              u++)
1577                 if (UUID_FLASH_ONLY(u))
1578                         ret = flash_dev_run(c, u);
1579
1580         return ret;
1581 }
1582
1583 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1584 {
1585         struct uuid_entry *u;
1586
1587         if (test_bit(CACHE_SET_STOPPING, &c->flags))
1588                 return -EINTR;
1589
1590         if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1591                 return -EPERM;
1592
1593         u = uuid_find_empty(c);
1594         if (!u) {
1595                 pr_err("Can't create volume, no room for UUID\n");
1596                 return -EINVAL;
1597         }
1598
1599         get_random_bytes(u->uuid, 16);
1600         memset(u->label, 0, 32);
1601         u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1602
1603         SET_UUID_FLASH_ONLY(u, 1);
1604         u->sectors = size >> 9;
1605
1606         bch_uuid_write(c);
1607
1608         return flash_dev_run(c, u);
1609 }
1610
1611 bool bch_cached_dev_error(struct cached_dev *dc)
1612 {
1613         if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1614                 return false;
1615
1616         dc->io_disable = true;
1617         /* make others know io_disable is true earlier */
1618         smp_mb();
1619
1620         pr_err("stop %s: too many IO errors on backing device %pg\n",
1621                dc->disk.disk->disk_name, dc->bdev);
1622
1623         bcache_device_stop(&dc->disk);
1624         return true;
1625 }
1626
1627 /* Cache set */
1628
1629 __printf(2, 3)
1630 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1631 {
1632         struct va_format vaf;
1633         va_list args;
1634
1635         if (c->on_error != ON_ERROR_PANIC &&
1636             test_bit(CACHE_SET_STOPPING, &c->flags))
1637                 return false;
1638
1639         if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1640                 pr_info("CACHE_SET_IO_DISABLE already set\n");
1641
1642         /*
1643          * XXX: we can be called from atomic context
1644          * acquire_console_sem();
1645          */
1646
1647         va_start(args, fmt);
1648
1649         vaf.fmt = fmt;
1650         vaf.va = &args;
1651
1652         pr_err("error on %pU: %pV, disabling caching\n",
1653                c->set_uuid, &vaf);
1654
1655         va_end(args);
1656
1657         if (c->on_error == ON_ERROR_PANIC)
1658                 panic("panic forced after error\n");
1659
1660         bch_cache_set_unregister(c);
1661         return true;
1662 }
1663
1664 /* When c->kobj released */
1665 void bch_cache_set_release(struct kobject *kobj)
1666 {
1667         struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1668
1669         kfree(c);
1670         module_put(THIS_MODULE);
1671 }
1672
1673 static CLOSURE_CALLBACK(cache_set_free)
1674 {
1675         closure_type(c, struct cache_set, cl);
1676         struct cache *ca;
1677
1678         debugfs_remove(c->debug);
1679
1680         bch_open_buckets_free(c);
1681         bch_btree_cache_free(c);
1682         bch_journal_free(c);
1683
1684         mutex_lock(&bch_register_lock);
1685         bch_bset_sort_state_free(&c->sort);
1686         free_pages((unsigned long) c->uuids, ilog2(meta_bucket_pages(&c->cache->sb)));
1687
1688         ca = c->cache;
1689         if (ca) {
1690                 ca->set = NULL;
1691                 c->cache = NULL;
1692                 kobject_put(&ca->kobj);
1693         }
1694
1695
1696         if (c->moving_gc_wq)
1697                 destroy_workqueue(c->moving_gc_wq);
1698         bioset_exit(&c->bio_split);
1699         mempool_exit(&c->fill_iter);
1700         mempool_exit(&c->bio_meta);
1701         mempool_exit(&c->search);
1702         kfree(c->devices);
1703
1704         list_del(&c->list);
1705         mutex_unlock(&bch_register_lock);
1706
1707         pr_info("Cache set %pU unregistered\n", c->set_uuid);
1708         wake_up(&unregister_wait);
1709
1710         closure_debug_destroy(&c->cl);
1711         kobject_put(&c->kobj);
1712 }
1713
1714 static CLOSURE_CALLBACK(cache_set_flush)
1715 {
1716         closure_type(c, struct cache_set, caching);
1717         struct cache *ca = c->cache;
1718         struct btree *b;
1719
1720         bch_cache_accounting_destroy(&c->accounting);
1721
1722         kobject_put(&c->internal);
1723         kobject_del(&c->kobj);
1724
1725         if (!IS_ERR_OR_NULL(c->gc_thread))
1726                 kthread_stop(c->gc_thread);
1727
1728         if (!IS_ERR(c->root))
1729                 list_add(&c->root->list, &c->btree_cache);
1730
1731         /*
1732          * Avoid flushing cached nodes if cache set is retiring
1733          * due to too many I/O errors detected.
1734          */
1735         if (!test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1736                 list_for_each_entry(b, &c->btree_cache, list) {
1737                         mutex_lock(&b->write_lock);
1738                         if (btree_node_dirty(b))
1739                                 __bch_btree_node_write(b, NULL);
1740                         mutex_unlock(&b->write_lock);
1741                 }
1742
1743         if (ca->alloc_thread)
1744                 kthread_stop(ca->alloc_thread);
1745
1746         if (c->journal.cur) {
1747                 cancel_delayed_work_sync(&c->journal.work);
1748                 /* flush last journal entry if needed */
1749                 c->journal.work.work.func(&c->journal.work.work);
1750         }
1751
1752         closure_return(cl);
1753 }
1754
1755 /*
1756  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1757  * cache set is unregistering due to too many I/O errors. In this condition,
1758  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1759  * value and whether the broken cache has dirty data:
1760  *
1761  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1762  *  BCH_CACHED_STOP_AUTO               0               NO
1763  *  BCH_CACHED_STOP_AUTO               1               YES
1764  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1765  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1766  *
1767  * The expected behavior is, if stop_when_cache_set_failed is configured to
1768  * "auto" via sysfs interface, the bcache device will not be stopped if the
1769  * backing device is clean on the broken cache device.
1770  */
1771 static void conditional_stop_bcache_device(struct cache_set *c,
1772                                            struct bcache_device *d,
1773                                            struct cached_dev *dc)
1774 {
1775         if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1776                 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.\n",
1777                         d->disk->disk_name, c->set_uuid);
1778                 bcache_device_stop(d);
1779         } else if (atomic_read(&dc->has_dirty)) {
1780                 /*
1781                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1782                  * and dc->has_dirty == 1
1783                  */
1784                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.\n",
1785                         d->disk->disk_name);
1786                 /*
1787                  * There might be a small time gap that cache set is
1788                  * released but bcache device is not. Inside this time
1789                  * gap, regular I/O requests will directly go into
1790                  * backing device as no cache set attached to. This
1791                  * behavior may also introduce potential inconsistence
1792                  * data in writeback mode while cache is dirty.
1793                  * Therefore before calling bcache_device_stop() due
1794                  * to a broken cache device, dc->io_disable should be
1795                  * explicitly set to true.
1796                  */
1797                 dc->io_disable = true;
1798                 /* make others know io_disable is true earlier */
1799                 smp_mb();
1800                 bcache_device_stop(d);
1801         } else {
1802                 /*
1803                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1804                  * and dc->has_dirty == 0
1805                  */
1806                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.\n",
1807                         d->disk->disk_name);
1808         }
1809 }
1810
1811 static CLOSURE_CALLBACK(__cache_set_unregister)
1812 {
1813         closure_type(c, struct cache_set, caching);
1814         struct cached_dev *dc;
1815         struct bcache_device *d;
1816         size_t i;
1817
1818         mutex_lock(&bch_register_lock);
1819
1820         for (i = 0; i < c->devices_max_used; i++) {
1821                 d = c->devices[i];
1822                 if (!d)
1823                         continue;
1824
1825                 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1826                     test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1827                         dc = container_of(d, struct cached_dev, disk);
1828                         bch_cached_dev_detach(dc);
1829                         if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1830                                 conditional_stop_bcache_device(c, d, dc);
1831                 } else {
1832                         bcache_device_stop(d);
1833                 }
1834         }
1835
1836         mutex_unlock(&bch_register_lock);
1837
1838         continue_at(cl, cache_set_flush, system_wq);
1839 }
1840
1841 void bch_cache_set_stop(struct cache_set *c)
1842 {
1843         if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1844                 /* closure_fn set to __cache_set_unregister() */
1845                 closure_queue(&c->caching);
1846 }
1847
1848 void bch_cache_set_unregister(struct cache_set *c)
1849 {
1850         set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1851         bch_cache_set_stop(c);
1852 }
1853
1854 #define alloc_meta_bucket_pages(gfp, sb)                \
1855         ((void *) __get_free_pages(__GFP_ZERO|__GFP_COMP|gfp, ilog2(meta_bucket_pages(sb))))
1856
1857 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1858 {
1859         int iter_size;
1860         struct cache *ca = container_of(sb, struct cache, sb);
1861         struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1862
1863         if (!c)
1864                 return NULL;
1865
1866         __module_get(THIS_MODULE);
1867         closure_init(&c->cl, NULL);
1868         set_closure_fn(&c->cl, cache_set_free, system_wq);
1869
1870         closure_init(&c->caching, &c->cl);
1871         set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1872
1873         /* Maybe create continue_at_noreturn() and use it here? */
1874         closure_set_stopped(&c->cl);
1875         closure_put(&c->cl);
1876
1877         kobject_init(&c->kobj, &bch_cache_set_ktype);
1878         kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1879
1880         bch_cache_accounting_init(&c->accounting, &c->cl);
1881
1882         memcpy(c->set_uuid, sb->set_uuid, 16);
1883
1884         c->cache                = ca;
1885         c->cache->set           = c;
1886         c->bucket_bits          = ilog2(sb->bucket_size);
1887         c->block_bits           = ilog2(sb->block_size);
1888         c->nr_uuids             = meta_bucket_bytes(sb) / sizeof(struct uuid_entry);
1889         c->devices_max_used     = 0;
1890         atomic_set(&c->attached_dev_nr, 0);
1891         c->btree_pages          = meta_bucket_pages(sb);
1892         if (c->btree_pages > BTREE_MAX_PAGES)
1893                 c->btree_pages = max_t(int, c->btree_pages / 4,
1894                                        BTREE_MAX_PAGES);
1895
1896         sema_init(&c->sb_write_mutex, 1);
1897         mutex_init(&c->bucket_lock);
1898         init_waitqueue_head(&c->btree_cache_wait);
1899         spin_lock_init(&c->btree_cannibalize_lock);
1900         init_waitqueue_head(&c->bucket_wait);
1901         init_waitqueue_head(&c->gc_wait);
1902         sema_init(&c->uuid_write_mutex, 1);
1903
1904         spin_lock_init(&c->btree_gc_time.lock);
1905         spin_lock_init(&c->btree_split_time.lock);
1906         spin_lock_init(&c->btree_read_time.lock);
1907
1908         bch_moving_init_cache_set(c);
1909
1910         INIT_LIST_HEAD(&c->list);
1911         INIT_LIST_HEAD(&c->cached_devs);
1912         INIT_LIST_HEAD(&c->btree_cache);
1913         INIT_LIST_HEAD(&c->btree_cache_freeable);
1914         INIT_LIST_HEAD(&c->btree_cache_freed);
1915         INIT_LIST_HEAD(&c->data_buckets);
1916
1917         iter_size = ((meta_bucket_pages(sb) * PAGE_SECTORS) / sb->block_size + 1) *
1918                 sizeof(struct btree_iter_set);
1919
1920         c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL);
1921         if (!c->devices)
1922                 goto err;
1923
1924         if (mempool_init_slab_pool(&c->search, 32, bch_search_cache))
1925                 goto err;
1926
1927         if (mempool_init_kmalloc_pool(&c->bio_meta, 2,
1928                         sizeof(struct bbio) +
1929                         sizeof(struct bio_vec) * meta_bucket_pages(sb)))
1930                 goto err;
1931
1932         if (mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size))
1933                 goto err;
1934
1935         if (bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1936                         BIOSET_NEED_RESCUER))
1937                 goto err;
1938
1939         c->uuids = alloc_meta_bucket_pages(GFP_KERNEL, sb);
1940         if (!c->uuids)
1941                 goto err;
1942
1943         c->moving_gc_wq = alloc_workqueue("bcache_gc", WQ_MEM_RECLAIM, 0);
1944         if (!c->moving_gc_wq)
1945                 goto err;
1946
1947         if (bch_journal_alloc(c))
1948                 goto err;
1949
1950         if (bch_btree_cache_alloc(c))
1951                 goto err;
1952
1953         if (bch_open_buckets_alloc(c))
1954                 goto err;
1955
1956         if (bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1957                 goto err;
1958
1959         c->congested_read_threshold_us  = 2000;
1960         c->congested_write_threshold_us = 20000;
1961         c->error_limit  = DEFAULT_IO_ERROR_LIMIT;
1962         c->idle_max_writeback_rate_enabled = 1;
1963         WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1964
1965         return c;
1966 err:
1967         bch_cache_set_unregister(c);
1968         return NULL;
1969 }
1970
1971 static int run_cache_set(struct cache_set *c)
1972 {
1973         const char *err = "cannot allocate memory";
1974         struct cached_dev *dc, *t;
1975         struct cache *ca = c->cache;
1976         struct closure cl;
1977         LIST_HEAD(journal);
1978         struct journal_replay *l;
1979
1980         closure_init_stack(&cl);
1981
1982         c->nbuckets = ca->sb.nbuckets;
1983         set_gc_sectors(c);
1984
1985         if (CACHE_SYNC(&c->cache->sb)) {
1986                 struct bkey *k;
1987                 struct jset *j;
1988
1989                 err = "cannot allocate memory for journal";
1990                 if (bch_journal_read(c, &journal))
1991                         goto err;
1992
1993                 pr_debug("btree_journal_read() done\n");
1994
1995                 err = "no journal entries found";
1996                 if (list_empty(&journal))
1997                         goto err;
1998
1999                 j = &list_entry(journal.prev, struct journal_replay, list)->j;
2000
2001                 err = "IO error reading priorities";
2002                 if (prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]))
2003                         goto err;
2004
2005                 /*
2006                  * If prio_read() fails it'll call cache_set_error and we'll
2007                  * tear everything down right away, but if we perhaps checked
2008                  * sooner we could avoid journal replay.
2009                  */
2010
2011                 k = &j->btree_root;
2012
2013                 err = "bad btree root";
2014                 if (__bch_btree_ptr_invalid(c, k))
2015                         goto err;
2016
2017                 err = "error reading btree root";
2018                 c->root = bch_btree_node_get(c, NULL, k,
2019                                              j->btree_level,
2020                                              true, NULL);
2021                 if (IS_ERR(c->root))
2022                         goto err;
2023
2024                 list_del_init(&c->root->list);
2025                 rw_unlock(true, c->root);
2026
2027                 err = uuid_read(c, j, &cl);
2028                 if (err)
2029                         goto err;
2030
2031                 err = "error in recovery";
2032                 if (bch_btree_check(c))
2033                         goto err;
2034
2035                 bch_journal_mark(c, &journal);
2036                 bch_initial_gc_finish(c);
2037                 pr_debug("btree_check() done\n");
2038
2039                 /*
2040                  * bcache_journal_next() can't happen sooner, or
2041                  * btree_gc_finish() will give spurious errors about last_gc >
2042                  * gc_gen - this is a hack but oh well.
2043                  */
2044                 bch_journal_next(&c->journal);
2045
2046                 err = "error starting allocator thread";
2047                 if (bch_cache_allocator_start(ca))
2048                         goto err;
2049
2050                 /*
2051                  * First place it's safe to allocate: btree_check() and
2052                  * btree_gc_finish() have to run before we have buckets to
2053                  * allocate, and bch_bucket_alloc_set() might cause a journal
2054                  * entry to be written so bcache_journal_next() has to be called
2055                  * first.
2056                  *
2057                  * If the uuids were in the old format we have to rewrite them
2058                  * before the next journal entry is written:
2059                  */
2060                 if (j->version < BCACHE_JSET_VERSION_UUID)
2061                         __uuid_write(c);
2062
2063                 err = "bcache: replay journal failed";
2064                 if (bch_journal_replay(c, &journal))
2065                         goto err;
2066         } else {
2067                 unsigned int j;
2068
2069                 pr_notice("invalidating existing data\n");
2070                 ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
2071                                         2, SB_JOURNAL_BUCKETS);
2072
2073                 for (j = 0; j < ca->sb.keys; j++)
2074                         ca->sb.d[j] = ca->sb.first_bucket + j;
2075
2076                 bch_initial_gc_finish(c);
2077
2078                 err = "error starting allocator thread";
2079                 if (bch_cache_allocator_start(ca))
2080                         goto err;
2081
2082                 mutex_lock(&c->bucket_lock);
2083                 bch_prio_write(ca, true);
2084                 mutex_unlock(&c->bucket_lock);
2085
2086                 err = "cannot allocate new UUID bucket";
2087                 if (__uuid_write(c))
2088                         goto err;
2089
2090                 err = "cannot allocate new btree root";
2091                 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
2092                 if (IS_ERR(c->root))
2093                         goto err;
2094
2095                 mutex_lock(&c->root->write_lock);
2096                 bkey_copy_key(&c->root->key, &MAX_KEY);
2097                 bch_btree_node_write(c->root, &cl);
2098                 mutex_unlock(&c->root->write_lock);
2099
2100                 bch_btree_set_root(c->root);
2101                 rw_unlock(true, c->root);
2102
2103                 /*
2104                  * We don't want to write the first journal entry until
2105                  * everything is set up - fortunately journal entries won't be
2106                  * written until the SET_CACHE_SYNC() here:
2107                  */
2108                 SET_CACHE_SYNC(&c->cache->sb, true);
2109
2110                 bch_journal_next(&c->journal);
2111                 bch_journal_meta(c, &cl);
2112         }
2113
2114         err = "error starting gc thread";
2115         if (bch_gc_thread_start(c))
2116                 goto err;
2117
2118         closure_sync(&cl);
2119         c->cache->sb.last_mount = (u32)ktime_get_real_seconds();
2120         bcache_write_super(c);
2121
2122         if (bch_has_feature_obso_large_bucket(&c->cache->sb))
2123                 pr_err("Detect obsoleted large bucket layout, all attached bcache device will be read-only\n");
2124
2125         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2126                 bch_cached_dev_attach(dc, c, NULL);
2127
2128         flash_devs_run(c);
2129
2130         bch_journal_space_reserve(&c->journal);
2131         set_bit(CACHE_SET_RUNNING, &c->flags);
2132         return 0;
2133 err:
2134         while (!list_empty(&journal)) {
2135                 l = list_first_entry(&journal, struct journal_replay, list);
2136                 list_del(&l->list);
2137                 kfree(l);
2138         }
2139
2140         closure_sync(&cl);
2141
2142         bch_cache_set_error(c, "%s", err);
2143
2144         return -EIO;
2145 }
2146
2147 static const char *register_cache_set(struct cache *ca)
2148 {
2149         char buf[12];
2150         const char *err = "cannot allocate memory";
2151         struct cache_set *c;
2152
2153         list_for_each_entry(c, &bch_cache_sets, list)
2154                 if (!memcmp(c->set_uuid, ca->sb.set_uuid, 16)) {
2155                         if (c->cache)
2156                                 return "duplicate cache set member";
2157
2158                         goto found;
2159                 }
2160
2161         c = bch_cache_set_alloc(&ca->sb);
2162         if (!c)
2163                 return err;
2164
2165         err = "error creating kobject";
2166         if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->set_uuid) ||
2167             kobject_add(&c->internal, &c->kobj, "internal"))
2168                 goto err;
2169
2170         if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2171                 goto err;
2172
2173         bch_debug_init_cache_set(c);
2174
2175         list_add(&c->list, &bch_cache_sets);
2176 found:
2177         sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2178         if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2179             sysfs_create_link(&c->kobj, &ca->kobj, buf))
2180                 goto err;
2181
2182         kobject_get(&ca->kobj);
2183         ca->set = c;
2184         ca->set->cache = ca;
2185
2186         err = "failed to run cache set";
2187         if (run_cache_set(c) < 0)
2188                 goto err;
2189
2190         return NULL;
2191 err:
2192         bch_cache_set_unregister(c);
2193         return err;
2194 }
2195
2196 /* Cache device */
2197
2198 /* When ca->kobj released */
2199 void bch_cache_release(struct kobject *kobj)
2200 {
2201         struct cache *ca = container_of(kobj, struct cache, kobj);
2202         unsigned int i;
2203
2204         if (ca->set) {
2205                 BUG_ON(ca->set->cache != ca);
2206                 ca->set->cache = NULL;
2207         }
2208
2209         free_pages((unsigned long) ca->disk_buckets, ilog2(meta_bucket_pages(&ca->sb)));
2210         kfree(ca->prio_buckets);
2211         vfree(ca->buckets);
2212
2213         free_heap(&ca->heap);
2214         free_fifo(&ca->free_inc);
2215
2216         for (i = 0; i < RESERVE_NR; i++)
2217                 free_fifo(&ca->free[i]);
2218
2219         if (ca->sb_disk)
2220                 put_page(virt_to_page(ca->sb_disk));
2221
2222         if (ca->bdev_file)
2223                 fput(ca->bdev_file);
2224
2225         kfree(ca);
2226         module_put(THIS_MODULE);
2227 }
2228
2229 static int cache_alloc(struct cache *ca)
2230 {
2231         size_t free;
2232         size_t btree_buckets;
2233         struct bucket *b;
2234         int ret = -ENOMEM;
2235         const char *err = NULL;
2236
2237         __module_get(THIS_MODULE);
2238         kobject_init(&ca->kobj, &bch_cache_ktype);
2239
2240         bio_init(&ca->journal.bio, NULL, ca->journal.bio.bi_inline_vecs, 8, 0);
2241
2242         /*
2243          * when ca->sb.njournal_buckets is not zero, journal exists,
2244          * and in bch_journal_replay(), tree node may split,
2245          * so bucket of RESERVE_BTREE type is needed,
2246          * the worst situation is all journal buckets are valid journal,
2247          * and all the keys need to replay,
2248          * so the number of  RESERVE_BTREE type buckets should be as much
2249          * as journal buckets
2250          */
2251         btree_buckets = ca->sb.njournal_buckets ?: 8;
2252         free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2253         if (!free) {
2254                 ret = -EPERM;
2255                 err = "ca->sb.nbuckets is too small";
2256                 goto err_free;
2257         }
2258
2259         if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2260                                                 GFP_KERNEL)) {
2261                 err = "ca->free[RESERVE_BTREE] alloc failed";
2262                 goto err_btree_alloc;
2263         }
2264
2265         if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2266                                                         GFP_KERNEL)) {
2267                 err = "ca->free[RESERVE_PRIO] alloc failed";
2268                 goto err_prio_alloc;
2269         }
2270
2271         if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2272                 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2273                 goto err_movinggc_alloc;
2274         }
2275
2276         if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2277                 err = "ca->free[RESERVE_NONE] alloc failed";
2278                 goto err_none_alloc;
2279         }
2280
2281         if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2282                 err = "ca->free_inc alloc failed";
2283                 goto err_free_inc_alloc;
2284         }
2285
2286         if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2287                 err = "ca->heap alloc failed";
2288                 goto err_heap_alloc;
2289         }
2290
2291         ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2292                               ca->sb.nbuckets));
2293         if (!ca->buckets) {
2294                 err = "ca->buckets alloc failed";
2295                 goto err_buckets_alloc;
2296         }
2297
2298         ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2299                                    prio_buckets(ca), 2),
2300                                    GFP_KERNEL);
2301         if (!ca->prio_buckets) {
2302                 err = "ca->prio_buckets alloc failed";
2303                 goto err_prio_buckets_alloc;
2304         }
2305
2306         ca->disk_buckets = alloc_meta_bucket_pages(GFP_KERNEL, &ca->sb);
2307         if (!ca->disk_buckets) {
2308                 err = "ca->disk_buckets alloc failed";
2309                 goto err_disk_buckets_alloc;
2310         }
2311
2312         ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2313
2314         for_each_bucket(b, ca)
2315                 atomic_set(&b->pin, 0);
2316         return 0;
2317
2318 err_disk_buckets_alloc:
2319         kfree(ca->prio_buckets);
2320 err_prio_buckets_alloc:
2321         vfree(ca->buckets);
2322 err_buckets_alloc:
2323         free_heap(&ca->heap);
2324 err_heap_alloc:
2325         free_fifo(&ca->free_inc);
2326 err_free_inc_alloc:
2327         free_fifo(&ca->free[RESERVE_NONE]);
2328 err_none_alloc:
2329         free_fifo(&ca->free[RESERVE_MOVINGGC]);
2330 err_movinggc_alloc:
2331         free_fifo(&ca->free[RESERVE_PRIO]);
2332 err_prio_alloc:
2333         free_fifo(&ca->free[RESERVE_BTREE]);
2334 err_btree_alloc:
2335 err_free:
2336         module_put(THIS_MODULE);
2337         if (err)
2338                 pr_notice("error %pg: %s\n", ca->bdev, err);
2339         return ret;
2340 }
2341
2342 static int register_cache(struct cache_sb *sb, struct cache_sb_disk *sb_disk,
2343                                 struct file *bdev_file,
2344                                 struct cache *ca)
2345 {
2346         const char *err = NULL; /* must be set for any error case */
2347         int ret = 0;
2348
2349         memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2350         ca->bdev_file = bdev_file;
2351         ca->bdev = file_bdev(bdev_file);
2352         ca->sb_disk = sb_disk;
2353
2354         if (bdev_max_discard_sectors(file_bdev(bdev_file)))
2355                 ca->discard = CACHE_DISCARD(&ca->sb);
2356
2357         ret = cache_alloc(ca);
2358         if (ret != 0) {
2359                 if (ret == -ENOMEM)
2360                         err = "cache_alloc(): -ENOMEM";
2361                 else if (ret == -EPERM)
2362                         err = "cache_alloc(): cache device is too small";
2363                 else
2364                         err = "cache_alloc(): unknown error";
2365                 pr_notice("error %pg: %s\n", file_bdev(bdev_file), err);
2366                 /*
2367                  * If we failed here, it means ca->kobj is not initialized yet,
2368                  * kobject_put() won't be called and there is no chance to
2369                  * call fput() to bdev in bch_cache_release(). So
2370                  * we explicitly call fput() on the block device here.
2371                  */
2372                 fput(bdev_file);
2373                 return ret;
2374         }
2375
2376         if (kobject_add(&ca->kobj, bdev_kobj(file_bdev(bdev_file)), "bcache")) {
2377                 pr_notice("error %pg: error calling kobject_add\n",
2378                           file_bdev(bdev_file));
2379                 ret = -ENOMEM;
2380                 goto out;
2381         }
2382
2383         mutex_lock(&bch_register_lock);
2384         err = register_cache_set(ca);
2385         mutex_unlock(&bch_register_lock);
2386
2387         if (err) {
2388                 ret = -ENODEV;
2389                 goto out;
2390         }
2391
2392         pr_info("registered cache device %pg\n", file_bdev(ca->bdev_file));
2393
2394 out:
2395         kobject_put(&ca->kobj);
2396         return ret;
2397 }
2398
2399 /* Global interfaces/init */
2400
2401 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2402                                const char *buffer, size_t size);
2403 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2404                                          struct kobj_attribute *attr,
2405                                          const char *buffer, size_t size);
2406
2407 kobj_attribute_write(register,          register_bcache);
2408 kobj_attribute_write(register_quiet,    register_bcache);
2409 kobj_attribute_write(pendings_cleanup,  bch_pending_bdevs_cleanup);
2410
2411 static bool bch_is_open_backing(dev_t dev)
2412 {
2413         struct cache_set *c, *tc;
2414         struct cached_dev *dc, *t;
2415
2416         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2417                 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2418                         if (dc->bdev->bd_dev == dev)
2419                                 return true;
2420         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2421                 if (dc->bdev->bd_dev == dev)
2422                         return true;
2423         return false;
2424 }
2425
2426 static bool bch_is_open_cache(dev_t dev)
2427 {
2428         struct cache_set *c, *tc;
2429
2430         list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2431                 struct cache *ca = c->cache;
2432
2433                 if (ca->bdev->bd_dev == dev)
2434                         return true;
2435         }
2436
2437         return false;
2438 }
2439
2440 static bool bch_is_open(dev_t dev)
2441 {
2442         return bch_is_open_cache(dev) || bch_is_open_backing(dev);
2443 }
2444
2445 struct async_reg_args {
2446         struct delayed_work reg_work;
2447         char *path;
2448         struct cache_sb *sb;
2449         struct cache_sb_disk *sb_disk;
2450         struct file *bdev_file;
2451         void *holder;
2452 };
2453
2454 static void register_bdev_worker(struct work_struct *work)
2455 {
2456         int fail = false;
2457         struct async_reg_args *args =
2458                 container_of(work, struct async_reg_args, reg_work.work);
2459
2460         mutex_lock(&bch_register_lock);
2461         if (register_bdev(args->sb, args->sb_disk, args->bdev_file,
2462                           args->holder) < 0)
2463                 fail = true;
2464         mutex_unlock(&bch_register_lock);
2465
2466         if (fail)
2467                 pr_info("error %s: fail to register backing device\n",
2468                         args->path);
2469         kfree(args->sb);
2470         kfree(args->path);
2471         kfree(args);
2472         module_put(THIS_MODULE);
2473 }
2474
2475 static void register_cache_worker(struct work_struct *work)
2476 {
2477         int fail = false;
2478         struct async_reg_args *args =
2479                 container_of(work, struct async_reg_args, reg_work.work);
2480
2481         /* blkdev_put() will be called in bch_cache_release() */
2482         if (register_cache(args->sb, args->sb_disk, args->bdev_file,
2483                            args->holder))
2484                 fail = true;
2485
2486         if (fail)
2487                 pr_info("error %s: fail to register cache device\n",
2488                         args->path);
2489         kfree(args->sb);
2490         kfree(args->path);
2491         kfree(args);
2492         module_put(THIS_MODULE);
2493 }
2494
2495 static void register_device_async(struct async_reg_args *args)
2496 {
2497         if (SB_IS_BDEV(args->sb))
2498                 INIT_DELAYED_WORK(&args->reg_work, register_bdev_worker);
2499         else
2500                 INIT_DELAYED_WORK(&args->reg_work, register_cache_worker);
2501
2502         /* 10 jiffies is enough for a delay */
2503         queue_delayed_work(system_wq, &args->reg_work, 10);
2504 }
2505
2506 static void *alloc_holder_object(struct cache_sb *sb)
2507 {
2508         if (SB_IS_BDEV(sb))
2509                 return kzalloc(sizeof(struct cached_dev), GFP_KERNEL);
2510         return kzalloc(sizeof(struct cache), GFP_KERNEL);
2511 }
2512
2513 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2514                                const char *buffer, size_t size)
2515 {
2516         const char *err;
2517         char *path = NULL;
2518         struct cache_sb *sb;
2519         struct cache_sb_disk *sb_disk;
2520         struct file *bdev_file, *bdev_file2;
2521         void *holder = NULL;
2522         ssize_t ret;
2523         bool async_registration = false;
2524         bool quiet = false;
2525
2526 #ifdef CONFIG_BCACHE_ASYNC_REGISTRATION
2527         async_registration = true;
2528 #endif
2529
2530         ret = -EBUSY;
2531         err = "failed to reference bcache module";
2532         if (!try_module_get(THIS_MODULE))
2533                 goto out;
2534
2535         /* For latest state of bcache_is_reboot */
2536         smp_mb();
2537         err = "bcache is in reboot";
2538         if (bcache_is_reboot)
2539                 goto out_module_put;
2540
2541         ret = -ENOMEM;
2542         err = "cannot allocate memory";
2543         path = kstrndup(buffer, size, GFP_KERNEL);
2544         if (!path)
2545                 goto out_module_put;
2546
2547         sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2548         if (!sb)
2549                 goto out_free_path;
2550
2551         ret = -EINVAL;
2552         err = "failed to open device";
2553         bdev_file = bdev_file_open_by_path(strim(path), BLK_OPEN_READ, NULL, NULL);
2554         if (IS_ERR(bdev_file))
2555                 goto out_free_sb;
2556
2557         err = "failed to set blocksize";
2558         if (set_blocksize(file_bdev(bdev_file), 4096))
2559                 goto out_blkdev_put;
2560
2561         err = read_super(sb, file_bdev(bdev_file), &sb_disk);
2562         if (err)
2563                 goto out_blkdev_put;
2564
2565         holder = alloc_holder_object(sb);
2566         if (!holder) {
2567                 ret = -ENOMEM;
2568                 err = "cannot allocate memory";
2569                 goto out_put_sb_page;
2570         }
2571
2572         /* Now reopen in exclusive mode with proper holder */
2573         bdev_file2 = bdev_file_open_by_dev(file_bdev(bdev_file)->bd_dev,
2574                         BLK_OPEN_READ | BLK_OPEN_WRITE, holder, NULL);
2575         fput(bdev_file);
2576         bdev_file = bdev_file2;
2577         if (IS_ERR(bdev_file)) {
2578                 ret = PTR_ERR(bdev_file);
2579                 bdev_file = NULL;
2580                 if (ret == -EBUSY) {
2581                         dev_t dev;
2582
2583                         mutex_lock(&bch_register_lock);
2584                         if (lookup_bdev(strim(path), &dev) == 0 &&
2585                             bch_is_open(dev))
2586                                 err = "device already registered";
2587                         else
2588                                 err = "device busy";
2589                         mutex_unlock(&bch_register_lock);
2590                         if (attr == &ksysfs_register_quiet) {
2591                                 quiet = true;
2592                                 ret = size;
2593                         }
2594                 }
2595                 goto out_free_holder;
2596         }
2597
2598         err = "failed to register device";
2599
2600         if (async_registration) {
2601                 /* register in asynchronous way */
2602                 struct async_reg_args *args =
2603                         kzalloc(sizeof(struct async_reg_args), GFP_KERNEL);
2604
2605                 if (!args) {
2606                         ret = -ENOMEM;
2607                         err = "cannot allocate memory";
2608                         goto out_free_holder;
2609                 }
2610
2611                 args->path      = path;
2612                 args->sb        = sb;
2613                 args->sb_disk   = sb_disk;
2614                 args->bdev_file = bdev_file;
2615                 args->holder    = holder;
2616                 register_device_async(args);
2617                 /* No wait and returns to user space */
2618                 goto async_done;
2619         }
2620
2621         if (SB_IS_BDEV(sb)) {
2622                 mutex_lock(&bch_register_lock);
2623                 ret = register_bdev(sb, sb_disk, bdev_file, holder);
2624                 mutex_unlock(&bch_register_lock);
2625                 /* blkdev_put() will be called in cached_dev_free() */
2626                 if (ret < 0)
2627                         goto out_free_sb;
2628         } else {
2629                 /* blkdev_put() will be called in bch_cache_release() */
2630                 ret = register_cache(sb, sb_disk, bdev_file, holder);
2631                 if (ret)
2632                         goto out_free_sb;
2633         }
2634
2635         kfree(sb);
2636         kfree(path);
2637         module_put(THIS_MODULE);
2638 async_done:
2639         return size;
2640
2641 out_free_holder:
2642         kfree(holder);
2643 out_put_sb_page:
2644         put_page(virt_to_page(sb_disk));
2645 out_blkdev_put:
2646         if (bdev_file)
2647                 fput(bdev_file);
2648 out_free_sb:
2649         kfree(sb);
2650 out_free_path:
2651         kfree(path);
2652         path = NULL;
2653 out_module_put:
2654         module_put(THIS_MODULE);
2655 out:
2656         if (!quiet)
2657                 pr_info("error %s: %s\n", path?path:"", err);
2658         return ret;
2659 }
2660
2661
2662 struct pdev {
2663         struct list_head list;
2664         struct cached_dev *dc;
2665 };
2666
2667 static ssize_t bch_pending_bdevs_cleanup(struct kobject *k,
2668                                          struct kobj_attribute *attr,
2669                                          const char *buffer,
2670                                          size_t size)
2671 {
2672         LIST_HEAD(pending_devs);
2673         ssize_t ret = size;
2674         struct cached_dev *dc, *tdc;
2675         struct pdev *pdev, *tpdev;
2676         struct cache_set *c, *tc;
2677
2678         mutex_lock(&bch_register_lock);
2679         list_for_each_entry_safe(dc, tdc, &uncached_devices, list) {
2680                 pdev = kmalloc(sizeof(struct pdev), GFP_KERNEL);
2681                 if (!pdev)
2682                         break;
2683                 pdev->dc = dc;
2684                 list_add(&pdev->list, &pending_devs);
2685         }
2686
2687         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2688                 char *pdev_set_uuid = pdev->dc->sb.set_uuid;
2689                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list) {
2690                         char *set_uuid = c->set_uuid;
2691
2692                         if (!memcmp(pdev_set_uuid, set_uuid, 16)) {
2693                                 list_del(&pdev->list);
2694                                 kfree(pdev);
2695                                 break;
2696                         }
2697                 }
2698         }
2699         mutex_unlock(&bch_register_lock);
2700
2701         list_for_each_entry_safe(pdev, tpdev, &pending_devs, list) {
2702                 pr_info("delete pdev %p\n", pdev);
2703                 list_del(&pdev->list);
2704                 bcache_device_stop(&pdev->dc->disk);
2705                 kfree(pdev);
2706         }
2707
2708         return ret;
2709 }
2710
2711 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2712 {
2713         if (bcache_is_reboot)
2714                 return NOTIFY_DONE;
2715
2716         if (code == SYS_DOWN ||
2717             code == SYS_HALT ||
2718             code == SYS_POWER_OFF) {
2719                 DEFINE_WAIT(wait);
2720                 unsigned long start = jiffies;
2721                 bool stopped = false;
2722
2723                 struct cache_set *c, *tc;
2724                 struct cached_dev *dc, *tdc;
2725
2726                 mutex_lock(&bch_register_lock);
2727
2728                 if (bcache_is_reboot)
2729                         goto out;
2730
2731                 /* New registration is rejected since now */
2732                 bcache_is_reboot = true;
2733                 /*
2734                  * Make registering caller (if there is) on other CPU
2735                  * core know bcache_is_reboot set to true earlier
2736                  */
2737                 smp_mb();
2738
2739                 if (list_empty(&bch_cache_sets) &&
2740                     list_empty(&uncached_devices))
2741                         goto out;
2742
2743                 mutex_unlock(&bch_register_lock);
2744
2745                 pr_info("Stopping all devices:\n");
2746
2747                 /*
2748                  * The reason bch_register_lock is not held to call
2749                  * bch_cache_set_stop() and bcache_device_stop() is to
2750                  * avoid potential deadlock during reboot, because cache
2751                  * set or bcache device stopping process will acquire
2752                  * bch_register_lock too.
2753                  *
2754                  * We are safe here because bcache_is_reboot sets to
2755                  * true already, register_bcache() will reject new
2756                  * registration now. bcache_is_reboot also makes sure
2757                  * bcache_reboot() won't be re-entered on by other thread,
2758                  * so there is no race in following list iteration by
2759                  * list_for_each_entry_safe().
2760                  */
2761                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2762                         bch_cache_set_stop(c);
2763
2764                 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2765                         bcache_device_stop(&dc->disk);
2766
2767
2768                 /*
2769                  * Give an early chance for other kthreads and
2770                  * kworkers to stop themselves
2771                  */
2772                 schedule();
2773
2774                 /* What's a condition variable? */
2775                 while (1) {
2776                         long timeout = start + 10 * HZ - jiffies;
2777
2778                         mutex_lock(&bch_register_lock);
2779                         stopped = list_empty(&bch_cache_sets) &&
2780                                 list_empty(&uncached_devices);
2781
2782                         if (timeout < 0 || stopped)
2783                                 break;
2784
2785                         prepare_to_wait(&unregister_wait, &wait,
2786                                         TASK_UNINTERRUPTIBLE);
2787
2788                         mutex_unlock(&bch_register_lock);
2789                         schedule_timeout(timeout);
2790                 }
2791
2792                 finish_wait(&unregister_wait, &wait);
2793
2794                 if (stopped)
2795                         pr_info("All devices stopped\n");
2796                 else
2797                         pr_notice("Timeout waiting for devices to be closed\n");
2798 out:
2799                 mutex_unlock(&bch_register_lock);
2800         }
2801
2802         return NOTIFY_DONE;
2803 }
2804
2805 static struct notifier_block reboot = {
2806         .notifier_call  = bcache_reboot,
2807         .priority       = INT_MAX, /* before any real devices */
2808 };
2809
2810 static void bcache_exit(void)
2811 {
2812         bch_debug_exit();
2813         bch_request_exit();
2814         if (bcache_kobj)
2815                 kobject_put(bcache_kobj);
2816         if (bcache_wq)
2817                 destroy_workqueue(bcache_wq);
2818         if (bch_journal_wq)
2819                 destroy_workqueue(bch_journal_wq);
2820         if (bch_flush_wq)
2821                 destroy_workqueue(bch_flush_wq);
2822         bch_btree_exit();
2823
2824         if (bcache_major)
2825                 unregister_blkdev(bcache_major, "bcache");
2826         unregister_reboot_notifier(&reboot);
2827         mutex_destroy(&bch_register_lock);
2828 }
2829
2830 /* Check and fixup module parameters */
2831 static void check_module_parameters(void)
2832 {
2833         if (bch_cutoff_writeback_sync == 0)
2834                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2835         else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2836                 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u\n",
2837                         bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2838                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2839         }
2840
2841         if (bch_cutoff_writeback == 0)
2842                 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2843         else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2844                 pr_warn("set bch_cutoff_writeback (%u) to max value %u\n",
2845                         bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2846                 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2847         }
2848
2849         if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2850                 pr_warn("set bch_cutoff_writeback (%u) to %u\n",
2851                         bch_cutoff_writeback, bch_cutoff_writeback_sync);
2852                 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2853         }
2854 }
2855
2856 static int __init bcache_init(void)
2857 {
2858         static const struct attribute *files[] = {
2859                 &ksysfs_register.attr,
2860                 &ksysfs_register_quiet.attr,
2861                 &ksysfs_pendings_cleanup.attr,
2862                 NULL
2863         };
2864
2865         check_module_parameters();
2866
2867         mutex_init(&bch_register_lock);
2868         init_waitqueue_head(&unregister_wait);
2869         register_reboot_notifier(&reboot);
2870
2871         bcache_major = register_blkdev(0, "bcache");
2872         if (bcache_major < 0) {
2873                 unregister_reboot_notifier(&reboot);
2874                 mutex_destroy(&bch_register_lock);
2875                 return bcache_major;
2876         }
2877
2878         if (bch_btree_init())
2879                 goto err;
2880
2881         bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2882         if (!bcache_wq)
2883                 goto err;
2884
2885         /*
2886          * Let's not make this `WQ_MEM_RECLAIM` for the following reasons:
2887          *
2888          * 1. It used `system_wq` before which also does no memory reclaim.
2889          * 2. With `WQ_MEM_RECLAIM` desktop stalls, increased boot times, and
2890          *    reduced throughput can be observed.
2891          *
2892          * We still want to user our own queue to not congest the `system_wq`.
2893          */
2894         bch_flush_wq = alloc_workqueue("bch_flush", 0, 0);
2895         if (!bch_flush_wq)
2896                 goto err;
2897
2898         bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2899         if (!bch_journal_wq)
2900                 goto err;
2901
2902         bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2903         if (!bcache_kobj)
2904                 goto err;
2905
2906         if (bch_request_init() ||
2907             sysfs_create_files(bcache_kobj, files))
2908                 goto err;
2909
2910         bch_debug_init();
2911
2912         bcache_is_reboot = false;
2913
2914         return 0;
2915 err:
2916         bcache_exit();
2917         return -ENOMEM;
2918 }
2919
2920 /*
2921  * Module hooks
2922  */
2923 module_exit(bcache_exit);
2924 module_init(bcache_init);
2925
2926 module_param(bch_cutoff_writeback, uint, 0);
2927 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2928
2929 module_param(bch_cutoff_writeback_sync, uint, 0);
2930 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2931
2932 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2933 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2934 MODULE_LICENSE("GPL");