drivers/md/bcache/super.c

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