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                 /*
 666                  * closure_fn set to
 667                  * - cached device: cached_dev_flush()
 668                  * - flash dev: flash_dev_flush()
 669                  */
 670                 closure_queue(&d->cl);
 671 }
 672
 673 static void bcache_device_unlink(struct bcache_device *d)
 674 {
 675         lockdep_assert_held(&bch_register_lock);
 676
 677         if (d->c && !test_and_set_bit(BCACHE_DEV_UNLINK_DONE, &d->flags)) {
 678                 unsigned int i;
 679                 struct cache *ca;
 680
 681                 sysfs_remove_link(&d->c->kobj, d->name);
 682                 sysfs_remove_link(&d->kobj, "cache");
 683
 684                 for_each_cache(ca, d->c, i)
 685                         bd_unlink_disk_holder(ca->bdev, d->disk);
 686         }
 687 }
 688
 689 static void bcache_device_link(struct bcache_device *d, struct cache_set *c,
 690                                const char *name)
 691 {
 692         unsigned int i;
 693         struct cache *ca;
 694
 695         for_each_cache(ca, d->c, i)
 696                 bd_link_disk_holder(ca->bdev, d->disk);
 697
 698         snprintf(d->name, BCACHEDEVNAME_SIZE,
 699                  "%s%u", name, d->id);
 700
 701         WARN(sysfs_create_link(&d->kobj, &c->kobj, "cache") ||
 702              sysfs_create_link(&c->kobj, &d->kobj, d->name),
 703              "Couldn't create device <-> cache set symlinks");
 704
 705         clear_bit(BCACHE_DEV_UNLINK_DONE, &d->flags);
 706 }
 707
 708 static void bcache_device_detach(struct bcache_device *d)
 709 {
 710         lockdep_assert_held(&bch_register_lock);
 711
 712         atomic_dec(&d->c->attached_dev_nr);
 713
 714         if (test_bit(BCACHE_DEV_DETACHING, &d->flags)) {
 715                 struct uuid_entry *u = d->c->uuids + d->id;
 716
 717                 SET_UUID_FLASH_ONLY(u, 0);
 718                 memcpy(u->uuid, invalid_uuid, 16);
 719                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
 720                 bch_uuid_write(d->c);
 721         }
 722
 723         bcache_device_unlink(d);
 724
 725         d->c->devices[d->id] = NULL;
 726         closure_put(&d->c->caching);
 727         d->c = NULL;
 728 }
 729
 730 static void bcache_device_attach(struct bcache_device *d, struct cache_set *c,
 731                                  unsigned int id)
 732 {
 733         d->id = id;
 734         d->c = c;
 735         c->devices[id] = d;
 736
 737         if (id >= c->devices_max_used)
 738                 c->devices_max_used = id + 1;
 739
 740         closure_get(&c->caching);
 741 }
 742
 743 static inline int first_minor_to_idx(int first_minor)
 744 {
 745         return (first_minor/BCACHE_MINORS);
 746 }
 747
 748 static inline int idx_to_first_minor(int idx)
 749 {
 750         return (idx * BCACHE_MINORS);
 751 }
 752
 753 static void bcache_device_free(struct bcache_device *d)
 754 {
 755         lockdep_assert_held(&bch_register_lock);
 756
 757         pr_info("%s stopped", d->disk->disk_name);
 758
 759         if (d->c)
 760                 bcache_device_detach(d);
 761         if (d->disk && d->disk->flags & GENHD_FL_UP)
 762                 del_gendisk(d->disk);
 763         if (d->disk && d->disk->queue)
 764                 blk_cleanup_queue(d->disk->queue);
 765         if (d->disk) {
 766                 ida_simple_remove(&bcache_device_idx,
 767                                   first_minor_to_idx(d->disk->first_minor));
 768                 put_disk(d->disk);
 769         }
 770
 771         bioset_exit(&d->bio_split);
 772         kvfree(d->full_dirty_stripes);
 773         kvfree(d->stripe_sectors_dirty);
 774
 775         closure_debug_destroy(&d->cl);
 776 }
 777
 778 static int bcache_device_init(struct bcache_device *d, unsigned int block_size,
 779                               sector_t sectors)
 780 {
 781         struct request_queue *q;
 782         const size_t max_stripes = min_t(size_t, INT_MAX,
 783                                          SIZE_MAX / sizeof(atomic_t));
 784         size_t n;
 785         int idx;
 786
 787         if (!d->stripe_size)
 788                 d->stripe_size = 1 << 31;
 789
 790         d->nr_stripes = DIV_ROUND_UP_ULL(sectors, d->stripe_size);
 791
 792         if (!d->nr_stripes || d->nr_stripes > max_stripes) {
 793                 pr_err("nr_stripes too large or invalid: %u (start sector beyond end of disk?)",
 794                         (unsigned int)d->nr_stripes);
 795                 return -ENOMEM;
 796         }
 797
 798         n = d->nr_stripes * sizeof(atomic_t);
 799         d->stripe_sectors_dirty = kvzalloc(n, GFP_KERNEL);
 800         if (!d->stripe_sectors_dirty)
 801                 return -ENOMEM;
 802
 803         n = BITS_TO_LONGS(d->nr_stripes) * sizeof(unsigned long);
 804         d->full_dirty_stripes = kvzalloc(n, GFP_KERNEL);
 805         if (!d->full_dirty_stripes)
 806                 return -ENOMEM;
 807
 808         idx = ida_simple_get(&bcache_device_idx, 0,
 809                                 BCACHE_DEVICE_IDX_MAX, GFP_KERNEL);
 810         if (idx < 0)
 811                 return idx;
 812
 813         if (bioset_init(&d->bio_split, 4, offsetof(struct bbio, bio),
 814                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER))
 815                 goto err;
 816
 817         d->disk = alloc_disk(BCACHE_MINORS);
 818         if (!d->disk)
 819                 goto err;
 820
 821         set_capacity(d->disk, sectors);
 822         snprintf(d->disk->disk_name, DISK_NAME_LEN, "bcache%i", idx);
 823
 824         d->disk->major          = bcache_major;
 825         d->disk->first_minor    = idx_to_first_minor(idx);
 826         d->disk->fops           = &bcache_ops;
 827         d->disk->private_data   = d;
 828
 829         q = blk_alloc_queue(GFP_KERNEL);
 830         if (!q)
 831                 return -ENOMEM;
 832
 833         blk_queue_make_request(q, NULL);
 834         d->disk->queue                  = q;
 835         q->queuedata                    = d;
 836         q->backing_dev_info->congested_data = d;
 837         q->limits.max_hw_sectors        = UINT_MAX;
 838         q->limits.max_sectors           = UINT_MAX;
 839         q->limits.max_segment_size      = UINT_MAX;
 840         q->limits.max_segments          = BIO_MAX_PAGES;
 841         blk_queue_max_discard_sectors(q, UINT_MAX);
 842         q->limits.discard_granularity   = 512;
 843         q->limits.io_min                = block_size;
 844         q->limits.logical_block_size    = block_size;
 845         q->limits.physical_block_size   = block_size;
 846         blk_queue_flag_set(QUEUE_FLAG_NONROT, d->disk->queue);
 847         blk_queue_flag_clear(QUEUE_FLAG_ADD_RANDOM, d->disk->queue);
 848         blk_queue_flag_set(QUEUE_FLAG_DISCARD, d->disk->queue);
 849
 850         blk_queue_write_cache(q, true, true);
 851
 852         return 0;
 853
 854 err:
 855         ida_simple_remove(&bcache_device_idx, idx);
 856         return -ENOMEM;
 857
 858 }
 859
 860 /* Cached device */
 861
 862 static void calc_cached_dev_sectors(struct cache_set *c)
 863 {
 864         uint64_t sectors = 0;
 865         struct cached_dev *dc;
 866
 867         list_for_each_entry(dc, &c->cached_devs, list)
 868                 sectors += bdev_sectors(dc->bdev);
 869
 870         c->cached_dev_sectors = sectors;
 871 }
 872
 873 #define BACKING_DEV_OFFLINE_TIMEOUT 5
 874 static int cached_dev_status_update(void *arg)
 875 {
 876         struct cached_dev *dc = arg;
 877         struct request_queue *q;
 878
 879         /*
 880          * If this delayed worker is stopping outside, directly quit here.
 881          * dc->io_disable might be set via sysfs interface, so check it
 882          * here too.
 883          */
 884         while (!kthread_should_stop() && !dc->io_disable) {
 885                 q = bdev_get_queue(dc->bdev);
 886                 if (blk_queue_dying(q))
 887                         dc->offline_seconds++;
 888                 else
 889                         dc->offline_seconds = 0;
 890
 891                 if (dc->offline_seconds >= BACKING_DEV_OFFLINE_TIMEOUT) {
 892                         pr_err("%s: device offline for %d seconds",
 893                                dc->backing_dev_name,
 894                                BACKING_DEV_OFFLINE_TIMEOUT);
 895                         pr_err("%s: disable I/O request due to backing "
 896                                "device offline", dc->disk.name);
 897                         dc->io_disable = true;
 898                         /* let others know earlier that io_disable is true */
 899                         smp_mb();
 900                         bcache_device_stop(&dc->disk);
 901                         break;
 902                 }
 903                 schedule_timeout_interruptible(HZ);
 904         }
 905
 906         wait_for_kthread_stop();
 907         return 0;
 908 }
 909
 910
 911 void bch_cached_dev_run(struct cached_dev *dc)
 912 {
 913         struct bcache_device *d = &dc->disk;
 914         char *buf = kmemdup_nul(dc->sb.label, SB_LABEL_SIZE, GFP_KERNEL);
 915         char *env[] = {
 916                 "DRIVER=bcache",
 917                 kasprintf(GFP_KERNEL, "CACHED_UUID=%pU", dc->sb.uuid),
 918                 kasprintf(GFP_KERNEL, "CACHED_LABEL=%s", buf ? : ""),
 919                 NULL,
 920         };
 921
 922         if (atomic_xchg(&dc->running, 1)) {
 923                 kfree(env[1]);
 924                 kfree(env[2]);
 925                 kfree(buf);
 926                 return;
 927         }
 928
 929         if (!d->c &&
 930             BDEV_STATE(&dc->sb) != BDEV_STATE_NONE) {
 931                 struct closure cl;
 932
 933                 closure_init_stack(&cl);
 934
 935                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_STALE);
 936                 bch_write_bdev_super(dc, &cl);
 937                 closure_sync(&cl);
 938         }
 939
 940         add_disk(d->disk);
 941         bd_link_disk_holder(dc->bdev, dc->disk.disk);
 942         /*
 943          * won't show up in the uevent file, use udevadm monitor -e instead
 944          * only class / kset properties are persistent
 945          */
 946         kobject_uevent_env(&disk_to_dev(d->disk)->kobj, KOBJ_CHANGE, env);
 947         kfree(env[1]);
 948         kfree(env[2]);
 949         kfree(buf);
 950
 951         if (sysfs_create_link(&d->kobj, &disk_to_dev(d->disk)->kobj, "dev") ||
 952             sysfs_create_link(&disk_to_dev(d->disk)->kobj, &d->kobj, "bcache"))
 953                 pr_debug("error creating sysfs link");
 954
 955         dc->status_update_thread = kthread_run(cached_dev_status_update,
 956                                                dc, "bcache_status_update");
 957         if (IS_ERR(dc->status_update_thread)) {
 958                 pr_warn("failed to create bcache_status_update kthread, "
 959                         "continue to run without monitoring backing "
 960                         "device status");
 961         }
 962 }
 963
 964 /*
 965  * If BCACHE_DEV_RATE_DW_RUNNING is set, it means routine of the delayed
 966  * work dc->writeback_rate_update is running. Wait until the routine
 967  * quits (BCACHE_DEV_RATE_DW_RUNNING is clear), then continue to
 968  * cancel it. If BCACHE_DEV_RATE_DW_RUNNING is not clear after time_out
 969  * seconds, give up waiting here and continue to cancel it too.
 970  */
 971 static void cancel_writeback_rate_update_dwork(struct cached_dev *dc)
 972 {
 973         int time_out = WRITEBACK_RATE_UPDATE_SECS_MAX * HZ;
 974
 975         do {
 976                 if (!test_bit(BCACHE_DEV_RATE_DW_RUNNING,
 977                               &dc->disk.flags))
 978                         break;
 979                 time_out--;
 980                 schedule_timeout_interruptible(1);
 981         } while (time_out > 0);
 982
 983         if (time_out == 0)
 984                 pr_warn("give up waiting for dc->writeback_write_update to quit");
 985
 986         cancel_delayed_work_sync(&dc->writeback_rate_update);
 987 }
 988
 989 static void cached_dev_detach_finish(struct work_struct *w)
 990 {
 991         struct cached_dev *dc = container_of(w, struct cached_dev, detach);
 992         struct closure cl;
 993
 994         closure_init_stack(&cl);
 995
 996         BUG_ON(!test_bit(BCACHE_DEV_DETACHING, &dc->disk.flags));
 997         BUG_ON(refcount_read(&dc->count));
 998
 999         mutex_lock(&bch_register_lock);
1000
1001         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1002                 cancel_writeback_rate_update_dwork(dc);
1003
1004         if (!IS_ERR_OR_NULL(dc->writeback_thread)) {
1005                 kthread_stop(dc->writeback_thread);
1006                 dc->writeback_thread = NULL;
1007         }
1008
1009         memset(&dc->sb.set_uuid, 0, 16);
1010         SET_BDEV_STATE(&dc->sb, BDEV_STATE_NONE);
1011
1012         bch_write_bdev_super(dc, &cl);
1013         closure_sync(&cl);
1014
1015         calc_cached_dev_sectors(dc->disk.c);
1016         bcache_device_detach(&dc->disk);
1017         list_move(&dc->list, &uncached_devices);
1018
1019         clear_bit(BCACHE_DEV_DETACHING, &dc->disk.flags);
1020         clear_bit(BCACHE_DEV_UNLINK_DONE, &dc->disk.flags);
1021
1022         mutex_unlock(&bch_register_lock);
1023
1024         pr_info("Caching disabled for %s", dc->backing_dev_name);
1025
1026         /* Drop ref we took in cached_dev_detach() */
1027         closure_put(&dc->disk.cl);
1028 }
1029
1030 void bch_cached_dev_detach(struct cached_dev *dc)
1031 {
1032         lockdep_assert_held(&bch_register_lock);
1033
1034         if (test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1035                 return;
1036
1037         if (test_and_set_bit(BCACHE_DEV_DETACHING, &dc->disk.flags))
1038                 return;
1039
1040         /*
1041          * Block the device from being closed and freed until we're finished
1042          * detaching
1043          */
1044         closure_get(&dc->disk.cl);
1045
1046         bch_writeback_queue(dc);
1047
1048         cached_dev_put(dc);
1049 }
1050
1051 int bch_cached_dev_attach(struct cached_dev *dc, struct cache_set *c,
1052                           uint8_t *set_uuid)
1053 {
1054         uint32_t rtime = cpu_to_le32((u32)ktime_get_real_seconds());
1055         struct uuid_entry *u;
1056         struct cached_dev *exist_dc, *t;
1057
1058         if ((set_uuid && memcmp(set_uuid, c->sb.set_uuid, 16)) ||
1059             (!set_uuid && memcmp(dc->sb.set_uuid, c->sb.set_uuid, 16)))
1060                 return -ENOENT;
1061
1062         if (dc->disk.c) {
1063                 pr_err("Can't attach %s: already attached",
1064                        dc->backing_dev_name);
1065                 return -EINVAL;
1066         }
1067
1068         if (test_bit(CACHE_SET_STOPPING, &c->flags)) {
1069                 pr_err("Can't attach %s: shutting down",
1070                        dc->backing_dev_name);
1071                 return -EINVAL;
1072         }
1073
1074         if (dc->sb.block_size < c->sb.block_size) {
1075                 /* Will die */
1076                 pr_err("Couldn't attach %s: block size less than set's block size",
1077                        dc->backing_dev_name);
1078                 return -EINVAL;
1079         }
1080
1081         /* Check whether already attached */
1082         list_for_each_entry_safe(exist_dc, t, &c->cached_devs, list) {
1083                 if (!memcmp(dc->sb.uuid, exist_dc->sb.uuid, 16)) {
1084                         pr_err("Tried to attach %s but duplicate UUID already attached",
1085                                 dc->backing_dev_name);
1086
1087                         return -EINVAL;
1088                 }
1089         }
1090
1091         u = uuid_find(c, dc->sb.uuid);
1092
1093         if (u &&
1094             (BDEV_STATE(&dc->sb) == BDEV_STATE_STALE ||
1095              BDEV_STATE(&dc->sb) == BDEV_STATE_NONE)) {
1096                 memcpy(u->uuid, invalid_uuid, 16);
1097                 u->invalidated = cpu_to_le32((u32)ktime_get_real_seconds());
1098                 u = NULL;
1099         }
1100
1101         if (!u) {
1102                 if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1103                         pr_err("Couldn't find uuid for %s in set",
1104                                dc->backing_dev_name);
1105                         return -ENOENT;
1106                 }
1107
1108                 u = uuid_find_empty(c);
1109                 if (!u) {
1110                         pr_err("Not caching %s, no room for UUID",
1111                                dc->backing_dev_name);
1112                         return -EINVAL;
1113                 }
1114         }
1115
1116         /*
1117          * Deadlocks since we're called via sysfs...
1118          * sysfs_remove_file(&dc->kobj, &sysfs_attach);
1119          */
1120
1121         if (bch_is_zero(u->uuid, 16)) {
1122                 struct closure cl;
1123
1124                 closure_init_stack(&cl);
1125
1126                 memcpy(u->uuid, dc->sb.uuid, 16);
1127                 memcpy(u->label, dc->sb.label, SB_LABEL_SIZE);
1128                 u->first_reg = u->last_reg = rtime;
1129                 bch_uuid_write(c);
1130
1131                 memcpy(dc->sb.set_uuid, c->sb.set_uuid, 16);
1132                 SET_BDEV_STATE(&dc->sb, BDEV_STATE_CLEAN);
1133
1134                 bch_write_bdev_super(dc, &cl);
1135                 closure_sync(&cl);
1136         } else {
1137                 u->last_reg = rtime;
1138                 bch_uuid_write(c);
1139         }
1140
1141         bcache_device_attach(&dc->disk, c, u - c->uuids);
1142         list_move(&dc->list, &c->cached_devs);
1143         calc_cached_dev_sectors(c);
1144
1145         /*
1146          * dc->c must be set before dc->count != 0 - paired with the mb in
1147          * cached_dev_get()
1148          */
1149         smp_wmb();
1150         refcount_set(&dc->count, 1);
1151
1152         /* Block writeback thread, but spawn it */
1153         down_write(&dc->writeback_lock);
1154         if (bch_cached_dev_writeback_start(dc)) {
1155                 up_write(&dc->writeback_lock);
1156                 return -ENOMEM;
1157         }
1158
1159         if (BDEV_STATE(&dc->sb) == BDEV_STATE_DIRTY) {
1160                 atomic_set(&dc->has_dirty, 1);
1161                 bch_writeback_queue(dc);
1162         }
1163
1164         bch_sectors_dirty_init(&dc->disk);
1165
1166         bch_cached_dev_run(dc);
1167         bcache_device_link(&dc->disk, c, "bdev");
1168         atomic_inc(&c->attached_dev_nr);
1169
1170         /* Allow the writeback thread to proceed */
1171         up_write(&dc->writeback_lock);
1172
1173         pr_info("Caching %s as %s on set %pU",
1174                 dc->backing_dev_name,
1175                 dc->disk.disk->disk_name,
1176                 dc->disk.c->sb.set_uuid);
1177         return 0;
1178 }
1179
1180 /* when dc->disk.kobj released */
1181 void bch_cached_dev_release(struct kobject *kobj)
1182 {
1183         struct cached_dev *dc = container_of(kobj, struct cached_dev,
1184                                              disk.kobj);
1185         kfree(dc);
1186         module_put(THIS_MODULE);
1187 }
1188
1189 static void cached_dev_free(struct closure *cl)
1190 {
1191         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1192
1193         mutex_lock(&bch_register_lock);
1194
1195         if (test_and_clear_bit(BCACHE_DEV_WB_RUNNING, &dc->disk.flags))
1196                 cancel_writeback_rate_update_dwork(dc);
1197
1198         if (!IS_ERR_OR_NULL(dc->writeback_thread))
1199                 kthread_stop(dc->writeback_thread);
1200         if (dc->writeback_write_wq)
1201                 destroy_workqueue(dc->writeback_write_wq);
1202         if (!IS_ERR_OR_NULL(dc->status_update_thread))
1203                 kthread_stop(dc->status_update_thread);
1204
1205         if (atomic_read(&dc->running))
1206                 bd_unlink_disk_holder(dc->bdev, dc->disk.disk);
1207         bcache_device_free(&dc->disk);
1208         list_del(&dc->list);
1209
1210         mutex_unlock(&bch_register_lock);
1211
1212         if (!IS_ERR_OR_NULL(dc->bdev))
1213                 blkdev_put(dc->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
1214
1215         wake_up(&unregister_wait);
1216
1217         kobject_put(&dc->disk.kobj);
1218 }
1219
1220 static void cached_dev_flush(struct closure *cl)
1221 {
1222         struct cached_dev *dc = container_of(cl, struct cached_dev, disk.cl);
1223         struct bcache_device *d = &dc->disk;
1224
1225         mutex_lock(&bch_register_lock);
1226         bcache_device_unlink(d);
1227         mutex_unlock(&bch_register_lock);
1228
1229         bch_cache_accounting_destroy(&dc->accounting);
1230         kobject_del(&d->kobj);
1231
1232         continue_at(cl, cached_dev_free, system_wq);
1233 }
1234
1235 static int cached_dev_init(struct cached_dev *dc, unsigned int block_size)
1236 {
1237         int ret;
1238         struct io *io;
1239         struct request_queue *q = bdev_get_queue(dc->bdev);
1240
1241         __module_get(THIS_MODULE);
1242         INIT_LIST_HEAD(&dc->list);
1243         closure_init(&dc->disk.cl, NULL);
1244         set_closure_fn(&dc->disk.cl, cached_dev_flush, system_wq);
1245         kobject_init(&dc->disk.kobj, &bch_cached_dev_ktype);
1246         INIT_WORK(&dc->detach, cached_dev_detach_finish);
1247         sema_init(&dc->sb_write_mutex, 1);
1248         INIT_LIST_HEAD(&dc->io_lru);
1249         spin_lock_init(&dc->io_lock);
1250         bch_cache_accounting_init(&dc->accounting, &dc->disk.cl);
1251
1252         dc->sequential_cutoff           = 4 << 20;
1253
1254         for (io = dc->io; io < dc->io + RECENT_IO; io++) {
1255                 list_add(&io->lru, &dc->io_lru);
1256                 hlist_add_head(&io->hash, dc->io_hash + RECENT_IO);
1257         }
1258
1259         dc->disk.stripe_size = q->limits.io_opt >> 9;
1260
1261         if (dc->disk.stripe_size)
1262                 dc->partial_stripes_expensive =
1263                         q->limits.raid_partial_stripes_expensive;
1264
1265         ret = bcache_device_init(&dc->disk, block_size,
1266                          dc->bdev->bd_part->nr_sects - dc->sb.data_offset);
1267         if (ret)
1268                 return ret;
1269
1270         dc->disk.disk->queue->backing_dev_info->ra_pages =
1271                 max(dc->disk.disk->queue->backing_dev_info->ra_pages,
1272                     q->backing_dev_info->ra_pages);
1273
1274         atomic_set(&dc->io_errors, 0);
1275         dc->io_disable = false;
1276         dc->error_limit = DEFAULT_CACHED_DEV_ERROR_LIMIT;
1277         /* default to auto */
1278         dc->stop_when_cache_set_failed = BCH_CACHED_DEV_STOP_AUTO;
1279
1280         bch_cached_dev_request_init(dc);
1281         bch_cached_dev_writeback_init(dc);
1282         return 0;
1283 }
1284
1285 /* Cached device - bcache superblock */
1286
1287 static int register_bdev(struct cache_sb *sb, struct page *sb_page,
1288                                  struct block_device *bdev,
1289                                  struct cached_dev *dc)
1290 {
1291         const char *err = "cannot allocate memory";
1292         struct cache_set *c;
1293
1294         bdevname(bdev, dc->backing_dev_name);
1295         memcpy(&dc->sb, sb, sizeof(struct cache_sb));
1296         dc->bdev = bdev;
1297         dc->bdev->bd_holder = dc;
1298
1299         bio_init(&dc->sb_bio, dc->sb_bio.bi_inline_vecs, 1);
1300         bio_first_bvec_all(&dc->sb_bio)->bv_page = sb_page;
1301         get_page(sb_page);
1302
1303
1304         if (cached_dev_init(dc, sb->block_size << 9))
1305                 goto err;
1306
1307         err = "error creating kobject";
1308         if (kobject_add(&dc->disk.kobj, &part_to_dev(bdev->bd_part)->kobj,
1309                         "bcache"))
1310                 goto err;
1311         if (bch_cache_accounting_add_kobjs(&dc->accounting, &dc->disk.kobj))
1312                 goto err;
1313
1314         pr_info("registered backing device %s", dc->backing_dev_name);
1315
1316         list_add(&dc->list, &uncached_devices);
1317         /* attach to a matched cache set if it exists */
1318         list_for_each_entry(c, &bch_cache_sets, list)
1319                 bch_cached_dev_attach(dc, c, NULL);
1320
1321         if (BDEV_STATE(&dc->sb) == BDEV_STATE_NONE ||
1322             BDEV_STATE(&dc->sb) == BDEV_STATE_STALE)
1323                 bch_cached_dev_run(dc);
1324
1325         return 0;
1326 err:
1327         pr_notice("error %s: %s", dc->backing_dev_name, err);
1328         bcache_device_stop(&dc->disk);
1329         return -EIO;
1330 }
1331
1332 /* Flash only volumes */
1333
1334 /* When d->kobj released */
1335 void bch_flash_dev_release(struct kobject *kobj)
1336 {
1337         struct bcache_device *d = container_of(kobj, struct bcache_device,
1338                                                kobj);
1339         kfree(d);
1340 }
1341
1342 static void flash_dev_free(struct closure *cl)
1343 {
1344         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1345
1346         mutex_lock(&bch_register_lock);
1347         atomic_long_sub(bcache_dev_sectors_dirty(d),
1348                         &d->c->flash_dev_dirty_sectors);
1349         bcache_device_free(d);
1350         mutex_unlock(&bch_register_lock);
1351         kobject_put(&d->kobj);
1352 }
1353
1354 static void flash_dev_flush(struct closure *cl)
1355 {
1356         struct bcache_device *d = container_of(cl, struct bcache_device, cl);
1357
1358         mutex_lock(&bch_register_lock);
1359         bcache_device_unlink(d);
1360         mutex_unlock(&bch_register_lock);
1361         kobject_del(&d->kobj);
1362         continue_at(cl, flash_dev_free, system_wq);
1363 }
1364
1365 static int flash_dev_run(struct cache_set *c, struct uuid_entry *u)
1366 {
1367         struct bcache_device *d = kzalloc(sizeof(struct bcache_device),
1368                                           GFP_KERNEL);
1369         if (!d)
1370                 return -ENOMEM;
1371
1372         closure_init(&d->cl, NULL);
1373         set_closure_fn(&d->cl, flash_dev_flush, system_wq);
1374
1375         kobject_init(&d->kobj, &bch_flash_dev_ktype);
1376
1377         if (bcache_device_init(d, block_bytes(c), u->sectors))
1378                 goto err;
1379
1380         bcache_device_attach(d, c, u - c->uuids);
1381         bch_sectors_dirty_init(d);
1382         bch_flash_dev_request_init(d);
1383         add_disk(d->disk);
1384
1385         if (kobject_add(&d->kobj, &disk_to_dev(d->disk)->kobj, "bcache"))
1386                 goto err;
1387
1388         bcache_device_link(d, c, "volume");
1389
1390         return 0;
1391 err:
1392         kobject_put(&d->kobj);
1393         return -ENOMEM;
1394 }
1395
1396 static int flash_devs_run(struct cache_set *c)
1397 {
1398         int ret = 0;
1399         struct uuid_entry *u;
1400
1401         for (u = c->uuids;
1402              u < c->uuids + c->nr_uuids && !ret;
1403              u++)
1404                 if (UUID_FLASH_ONLY(u))
1405                         ret = flash_dev_run(c, u);
1406
1407         return ret;
1408 }
1409
1410 int bch_flash_dev_create(struct cache_set *c, uint64_t size)
1411 {
1412         struct uuid_entry *u;
1413
1414         if (test_bit(CACHE_SET_STOPPING, &c->flags))
1415                 return -EINTR;
1416
1417         if (!test_bit(CACHE_SET_RUNNING, &c->flags))
1418                 return -EPERM;
1419
1420         u = uuid_find_empty(c);
1421         if (!u) {
1422                 pr_err("Can't create volume, no room for UUID");
1423                 return -EINVAL;
1424         }
1425
1426         get_random_bytes(u->uuid, 16);
1427         memset(u->label, 0, 32);
1428         u->first_reg = u->last_reg = cpu_to_le32((u32)ktime_get_real_seconds());
1429
1430         SET_UUID_FLASH_ONLY(u, 1);
1431         u->sectors = size >> 9;
1432
1433         bch_uuid_write(c);
1434
1435         return flash_dev_run(c, u);
1436 }
1437
1438 bool bch_cached_dev_error(struct cached_dev *dc)
1439 {
1440         struct cache_set *c;
1441
1442         if (!dc || test_bit(BCACHE_DEV_CLOSING, &dc->disk.flags))
1443                 return false;
1444
1445         dc->io_disable = true;
1446         /* make others know io_disable is true earlier */
1447         smp_mb();
1448
1449         pr_err("stop %s: too many IO errors on backing device %s\n",
1450                 dc->disk.disk->disk_name, dc->backing_dev_name);
1451
1452         /*
1453          * If the cached device is still attached to a cache set,
1454          * even dc->io_disable is true and no more I/O requests
1455          * accepted, cache device internal I/O (writeback scan or
1456          * garbage collection) may still prevent bcache device from
1457          * being stopped. So here CACHE_SET_IO_DISABLE should be
1458          * set to c->flags too, to make the internal I/O to cache
1459          * device rejected and stopped immediately.
1460          * If c is NULL, that means the bcache device is not attached
1461          * to any cache set, then no CACHE_SET_IO_DISABLE bit to set.
1462          */
1463         c = dc->disk.c;
1464         if (c && test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1465                 pr_info("CACHE_SET_IO_DISABLE already set");
1466
1467         bcache_device_stop(&dc->disk);
1468         return true;
1469 }
1470
1471 /* Cache set */
1472
1473 __printf(2, 3)
1474 bool bch_cache_set_error(struct cache_set *c, const char *fmt, ...)
1475 {
1476         va_list args;
1477
1478         if (c->on_error != ON_ERROR_PANIC &&
1479             test_bit(CACHE_SET_STOPPING, &c->flags))
1480                 return false;
1481
1482         if (test_and_set_bit(CACHE_SET_IO_DISABLE, &c->flags))
1483                 pr_info("CACHE_SET_IO_DISABLE already set");
1484
1485         /*
1486          * XXX: we can be called from atomic context
1487          * acquire_console_sem();
1488          */
1489
1490         pr_err("bcache: error on %pU: ", c->sb.set_uuid);
1491
1492         va_start(args, fmt);
1493         vprintk(fmt, args);
1494         va_end(args);
1495
1496         pr_err(", disabling caching\n");
1497
1498         if (c->on_error == ON_ERROR_PANIC)
1499                 panic("panic forced after error\n");
1500
1501         bch_cache_set_unregister(c);
1502         return true;
1503 }
1504
1505 /* When c->kobj released */
1506 void bch_cache_set_release(struct kobject *kobj)
1507 {
1508         struct cache_set *c = container_of(kobj, struct cache_set, kobj);
1509
1510         kfree(c);
1511         module_put(THIS_MODULE);
1512 }
1513
1514 static void cache_set_free(struct closure *cl)
1515 {
1516         struct cache_set *c = container_of(cl, struct cache_set, cl);
1517         struct cache *ca;
1518         unsigned int i;
1519
1520         debugfs_remove(c->debug);
1521
1522         bch_open_buckets_free(c);
1523         bch_btree_cache_free(c);
1524         bch_journal_free(c);
1525
1526         mutex_lock(&bch_register_lock);
1527         for_each_cache(ca, c, i)
1528                 if (ca) {
1529                         ca->set = NULL;
1530                         c->cache[ca->sb.nr_this_dev] = NULL;
1531                         kobject_put(&ca->kobj);
1532                 }
1533
1534         bch_bset_sort_state_free(&c->sort);
1535         free_pages((unsigned long) c->uuids, ilog2(bucket_pages(c)));
1536
1537         if (c->moving_gc_wq)
1538                 destroy_workqueue(c->moving_gc_wq);
1539         bioset_exit(&c->bio_split);
1540         mempool_exit(&c->fill_iter);
1541         mempool_exit(&c->bio_meta);
1542         mempool_exit(&c->search);
1543         kfree(c->devices);
1544
1545         list_del(&c->list);
1546         mutex_unlock(&bch_register_lock);
1547
1548         pr_info("Cache set %pU unregistered", c->sb.set_uuid);
1549         wake_up(&unregister_wait);
1550
1551         closure_debug_destroy(&c->cl);
1552         kobject_put(&c->kobj);
1553 }
1554
1555 static void cache_set_flush(struct closure *cl)
1556 {
1557         struct cache_set *c = container_of(cl, struct cache_set, caching);
1558         struct cache *ca;
1559         struct btree *b;
1560         unsigned int i;
1561
1562         bch_cache_accounting_destroy(&c->accounting);
1563
1564         kobject_put(&c->internal);
1565         kobject_del(&c->kobj);
1566
1567         if (c->gc_thread)
1568                 kthread_stop(c->gc_thread);
1569
1570         if (!IS_ERR_OR_NULL(c->root))
1571                 list_add(&c->root->list, &c->btree_cache);
1572
1573         /* Should skip this if we're unregistering because of an error */
1574         list_for_each_entry(b, &c->btree_cache, list) {
1575                 mutex_lock(&b->write_lock);
1576                 if (btree_node_dirty(b))
1577                         __bch_btree_node_write(b, NULL);
1578                 mutex_unlock(&b->write_lock);
1579         }
1580
1581         for_each_cache(ca, c, i)
1582                 if (ca->alloc_thread)
1583                         kthread_stop(ca->alloc_thread);
1584
1585         if (c->journal.cur) {
1586                 cancel_delayed_work_sync(&c->journal.work);
1587                 /* flush last journal entry if needed */
1588                 c->journal.work.work.func(&c->journal.work.work);
1589         }
1590
1591         closure_return(cl);
1592 }
1593
1594 /*
1595  * This function is only called when CACHE_SET_IO_DISABLE is set, which means
1596  * cache set is unregistering due to too many I/O errors. In this condition,
1597  * the bcache device might be stopped, it depends on stop_when_cache_set_failed
1598  * value and whether the broken cache has dirty data:
1599  *
1600  * dc->stop_when_cache_set_failed    dc->has_dirty   stop bcache device
1601  *  BCH_CACHED_STOP_AUTO               0               NO
1602  *  BCH_CACHED_STOP_AUTO               1               YES
1603  *  BCH_CACHED_DEV_STOP_ALWAYS         0               YES
1604  *  BCH_CACHED_DEV_STOP_ALWAYS         1               YES
1605  *
1606  * The expected behavior is, if stop_when_cache_set_failed is configured to
1607  * "auto" via sysfs interface, the bcache device will not be stopped if the
1608  * backing device is clean on the broken cache device.
1609  */
1610 static void conditional_stop_bcache_device(struct cache_set *c,
1611                                            struct bcache_device *d,
1612                                            struct cached_dev *dc)
1613 {
1614         if (dc->stop_when_cache_set_failed == BCH_CACHED_DEV_STOP_ALWAYS) {
1615                 pr_warn("stop_when_cache_set_failed of %s is \"always\", stop it for failed cache set %pU.",
1616                         d->disk->disk_name, c->sb.set_uuid);
1617                 bcache_device_stop(d);
1618         } else if (atomic_read(&dc->has_dirty)) {
1619                 /*
1620                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1621                  * and dc->has_dirty == 1
1622                  */
1623                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is dirty, stop it to avoid potential data corruption.",
1624                         d->disk->disk_name);
1625                 /*
1626                  * There might be a small time gap that cache set is
1627                  * released but bcache device is not. Inside this time
1628                  * gap, regular I/O requests will directly go into
1629                  * backing device as no cache set attached to. This
1630                  * behavior may also introduce potential inconsistence
1631                  * data in writeback mode while cache is dirty.
1632                  * Therefore before calling bcache_device_stop() due
1633                  * to a broken cache device, dc->io_disable should be
1634                  * explicitly set to true.
1635                  */
1636                 dc->io_disable = true;
1637                 /* make others know io_disable is true earlier */
1638                 smp_mb();
1639                 bcache_device_stop(d);
1640         } else {
1641                 /*
1642                  * dc->stop_when_cache_set_failed == BCH_CACHED_STOP_AUTO
1643                  * and dc->has_dirty == 0
1644                  */
1645                 pr_warn("stop_when_cache_set_failed of %s is \"auto\" and cache is clean, keep it alive.",
1646                         d->disk->disk_name);
1647         }
1648 }
1649
1650 static void __cache_set_unregister(struct closure *cl)
1651 {
1652         struct cache_set *c = container_of(cl, struct cache_set, caching);
1653         struct cached_dev *dc;
1654         struct bcache_device *d;
1655         size_t i;
1656
1657         mutex_lock(&bch_register_lock);
1658
1659         for (i = 0; i < c->devices_max_used; i++) {
1660                 d = c->devices[i];
1661                 if (!d)
1662                         continue;
1663
1664                 if (!UUID_FLASH_ONLY(&c->uuids[i]) &&
1665                     test_bit(CACHE_SET_UNREGISTERING, &c->flags)) {
1666                         dc = container_of(d, struct cached_dev, disk);
1667                         bch_cached_dev_detach(dc);
1668                         if (test_bit(CACHE_SET_IO_DISABLE, &c->flags))
1669                                 conditional_stop_bcache_device(c, d, dc);
1670                 } else {
1671                         bcache_device_stop(d);
1672                 }
1673         }
1674
1675         mutex_unlock(&bch_register_lock);
1676
1677         continue_at(cl, cache_set_flush, system_wq);
1678 }
1679
1680 void bch_cache_set_stop(struct cache_set *c)
1681 {
1682         if (!test_and_set_bit(CACHE_SET_STOPPING, &c->flags))
1683                 /* closure_fn set to __cache_set_unregister() */
1684                 closure_queue(&c->caching);
1685 }
1686
1687 void bch_cache_set_unregister(struct cache_set *c)
1688 {
1689         set_bit(CACHE_SET_UNREGISTERING, &c->flags);
1690         bch_cache_set_stop(c);
1691 }
1692
1693 #define alloc_bucket_pages(gfp, c)                      \
1694         ((void *) __get_free_pages(__GFP_ZERO|gfp, ilog2(bucket_pages(c))))
1695
1696 struct cache_set *bch_cache_set_alloc(struct cache_sb *sb)
1697 {
1698         int iter_size;
1699         struct cache_set *c = kzalloc(sizeof(struct cache_set), GFP_KERNEL);
1700
1701         if (!c)
1702                 return NULL;
1703
1704         __module_get(THIS_MODULE);
1705         closure_init(&c->cl, NULL);
1706         set_closure_fn(&c->cl, cache_set_free, system_wq);
1707
1708         closure_init(&c->caching, &c->cl);
1709         set_closure_fn(&c->caching, __cache_set_unregister, system_wq);
1710
1711         /* Maybe create continue_at_noreturn() and use it here? */
1712         closure_set_stopped(&c->cl);
1713         closure_put(&c->cl);
1714
1715         kobject_init(&c->kobj, &bch_cache_set_ktype);
1716         kobject_init(&c->internal, &bch_cache_set_internal_ktype);
1717
1718         bch_cache_accounting_init(&c->accounting, &c->cl);
1719
1720         memcpy(c->sb.set_uuid, sb->set_uuid, 16);
1721         c->sb.block_size        = sb->block_size;
1722         c->sb.bucket_size       = sb->bucket_size;
1723         c->sb.nr_in_set         = sb->nr_in_set;
1724         c->sb.last_mount        = sb->last_mount;
1725         c->bucket_bits          = ilog2(sb->bucket_size);
1726         c->block_bits           = ilog2(sb->block_size);
1727         c->nr_uuids             = bucket_bytes(c) / sizeof(struct uuid_entry);
1728         c->devices_max_used     = 0;
1729         atomic_set(&c->attached_dev_nr, 0);
1730         c->btree_pages          = bucket_pages(c);
1731         if (c->btree_pages > BTREE_MAX_PAGES)
1732                 c->btree_pages = max_t(int, c->btree_pages / 4,
1733                                        BTREE_MAX_PAGES);
1734
1735         sema_init(&c->sb_write_mutex, 1);
1736         mutex_init(&c->bucket_lock);
1737         init_waitqueue_head(&c->btree_cache_wait);
1738         init_waitqueue_head(&c->bucket_wait);
1739         init_waitqueue_head(&c->gc_wait);
1740         sema_init(&c->uuid_write_mutex, 1);
1741
1742         spin_lock_init(&c->btree_gc_time.lock);
1743         spin_lock_init(&c->btree_split_time.lock);
1744         spin_lock_init(&c->btree_read_time.lock);
1745
1746         bch_moving_init_cache_set(c);
1747
1748         INIT_LIST_HEAD(&c->list);
1749         INIT_LIST_HEAD(&c->cached_devs);
1750         INIT_LIST_HEAD(&c->btree_cache);
1751         INIT_LIST_HEAD(&c->btree_cache_freeable);
1752         INIT_LIST_HEAD(&c->btree_cache_freed);
1753         INIT_LIST_HEAD(&c->data_buckets);
1754
1755         iter_size = (sb->bucket_size / sb->block_size + 1) *
1756                 sizeof(struct btree_iter_set);
1757
1758         if (!(c->devices = kcalloc(c->nr_uuids, sizeof(void *), GFP_KERNEL)) ||
1759             mempool_init_slab_pool(&c->search, 32, bch_search_cache) ||
1760             mempool_init_kmalloc_pool(&c->bio_meta, 2,
1761                                 sizeof(struct bbio) + sizeof(struct bio_vec) *
1762                                 bucket_pages(c)) ||
1763             mempool_init_kmalloc_pool(&c->fill_iter, 1, iter_size) ||
1764             bioset_init(&c->bio_split, 4, offsetof(struct bbio, bio),
1765                         BIOSET_NEED_BVECS|BIOSET_NEED_RESCUER) ||
1766             !(c->uuids = alloc_bucket_pages(GFP_KERNEL, c)) ||
1767             !(c->moving_gc_wq = alloc_workqueue("bcache_gc",
1768                                                 WQ_MEM_RECLAIM, 0)) ||
1769             bch_journal_alloc(c) ||
1770             bch_btree_cache_alloc(c) ||
1771             bch_open_buckets_alloc(c) ||
1772             bch_bset_sort_state_init(&c->sort, ilog2(c->btree_pages)))
1773                 goto err;
1774
1775         c->congested_read_threshold_us  = 2000;
1776         c->congested_write_threshold_us = 20000;
1777         c->error_limit  = DEFAULT_IO_ERROR_LIMIT;
1778         WARN_ON(test_and_clear_bit(CACHE_SET_IO_DISABLE, &c->flags));
1779
1780         return c;
1781 err:
1782         bch_cache_set_unregister(c);
1783         return NULL;
1784 }
1785
1786 static int run_cache_set(struct cache_set *c)
1787 {
1788         const char *err = "cannot allocate memory";
1789         struct cached_dev *dc, *t;
1790         struct cache *ca;
1791         struct closure cl;
1792         unsigned int i;
1793         LIST_HEAD(journal);
1794         struct journal_replay *l;
1795
1796         closure_init_stack(&cl);
1797
1798         for_each_cache(ca, c, i)
1799                 c->nbuckets += ca->sb.nbuckets;
1800         set_gc_sectors(c);
1801
1802         if (CACHE_SYNC(&c->sb)) {
1803                 struct bkey *k;
1804                 struct jset *j;
1805
1806                 err = "cannot allocate memory for journal";
1807                 if (bch_journal_read(c, &journal))
1808                         goto err;
1809
1810                 pr_debug("btree_journal_read() done");
1811
1812                 err = "no journal entries found";
1813                 if (list_empty(&journal))
1814                         goto err;
1815
1816                 j = &list_entry(journal.prev, struct journal_replay, list)->j;
1817
1818                 err = "IO error reading priorities";
1819                 for_each_cache(ca, c, i)
1820                         prio_read(ca, j->prio_bucket[ca->sb.nr_this_dev]);
1821
1822                 /*
1823                  * If prio_read() fails it'll call cache_set_error and we'll
1824                  * tear everything down right away, but if we perhaps checked
1825                  * sooner we could avoid journal replay.
1826                  */
1827
1828                 k = &j->btree_root;
1829
1830                 err = "bad btree root";
1831                 if (__bch_btree_ptr_invalid(c, k))
1832                         goto err;
1833
1834                 err = "error reading btree root";
1835                 c->root = bch_btree_node_get(c, NULL, k,
1836                                              j->btree_level,
1837                                              true, NULL);
1838                 if (IS_ERR_OR_NULL(c->root))
1839                         goto err;
1840
1841                 list_del_init(&c->root->list);
1842                 rw_unlock(true, c->root);
1843
1844                 err = uuid_read(c, j, &cl);
1845                 if (err)
1846                         goto err;
1847
1848                 err = "error in recovery";
1849                 if (bch_btree_check(c))
1850                         goto err;
1851
1852                 bch_journal_mark(c, &journal);
1853                 bch_initial_gc_finish(c);
1854                 pr_debug("btree_check() done");
1855
1856                 /*
1857                  * bcache_journal_next() can't happen sooner, or
1858                  * btree_gc_finish() will give spurious errors about last_gc >
1859                  * gc_gen - this is a hack but oh well.
1860                  */
1861                 bch_journal_next(&c->journal);
1862
1863                 err = "error starting allocator thread";
1864                 for_each_cache(ca, c, i)
1865                         if (bch_cache_allocator_start(ca))
1866                                 goto err;
1867
1868                 /*
1869                  * First place it's safe to allocate: btree_check() and
1870                  * btree_gc_finish() have to run before we have buckets to
1871                  * allocate, and bch_bucket_alloc_set() might cause a journal
1872                  * entry to be written so bcache_journal_next() has to be called
1873                  * first.
1874                  *
1875                  * If the uuids were in the old format we have to rewrite them
1876                  * before the next journal entry is written:
1877                  */
1878                 if (j->version < BCACHE_JSET_VERSION_UUID)
1879                         __uuid_write(c);
1880
1881                 err = "bcache: replay journal failed";
1882                 if (bch_journal_replay(c, &journal))
1883                         goto err;
1884         } else {
1885                 pr_notice("invalidating existing data");
1886
1887                 for_each_cache(ca, c, i) {
1888                         unsigned int j;
1889
1890                         ca->sb.keys = clamp_t(int, ca->sb.nbuckets >> 7,
1891                                               2, SB_JOURNAL_BUCKETS);
1892
1893                         for (j = 0; j < ca->sb.keys; j++)
1894                                 ca->sb.d[j] = ca->sb.first_bucket + j;
1895                 }
1896
1897                 bch_initial_gc_finish(c);
1898
1899                 err = "error starting allocator thread";
1900                 for_each_cache(ca, c, i)
1901                         if (bch_cache_allocator_start(ca))
1902                                 goto err;
1903
1904                 mutex_lock(&c->bucket_lock);
1905                 for_each_cache(ca, c, i)
1906                         bch_prio_write(ca);
1907                 mutex_unlock(&c->bucket_lock);
1908
1909                 err = "cannot allocate new UUID bucket";
1910                 if (__uuid_write(c))
1911                         goto err;
1912
1913                 err = "cannot allocate new btree root";
1914                 c->root = __bch_btree_node_alloc(c, NULL, 0, true, NULL);
1915                 if (IS_ERR_OR_NULL(c->root))
1916                         goto err;
1917
1918                 mutex_lock(&c->root->write_lock);
1919                 bkey_copy_key(&c->root->key, &MAX_KEY);
1920                 bch_btree_node_write(c->root, &cl);
1921                 mutex_unlock(&c->root->write_lock);
1922
1923                 bch_btree_set_root(c->root);
1924                 rw_unlock(true, c->root);
1925
1926                 /*
1927                  * We don't want to write the first journal entry until
1928                  * everything is set up - fortunately journal entries won't be
1929                  * written until the SET_CACHE_SYNC() here:
1930                  */
1931                 SET_CACHE_SYNC(&c->sb, true);
1932
1933                 bch_journal_next(&c->journal);
1934                 bch_journal_meta(c, &cl);
1935         }
1936
1937         err = "error starting gc thread";
1938         if (bch_gc_thread_start(c))
1939                 goto err;
1940
1941         closure_sync(&cl);
1942         c->sb.last_mount = (u32)ktime_get_real_seconds();
1943         bcache_write_super(c);
1944
1945         list_for_each_entry_safe(dc, t, &uncached_devices, list)
1946                 bch_cached_dev_attach(dc, c, NULL);
1947
1948         flash_devs_run(c);
1949
1950         set_bit(CACHE_SET_RUNNING, &c->flags);
1951         return 0;
1952 err:
1953         while (!list_empty(&journal)) {
1954                 l = list_first_entry(&journal, struct journal_replay, list);
1955                 list_del(&l->list);
1956                 kfree(l);
1957         }
1958
1959         closure_sync(&cl);
1960         /* XXX: test this, it's broken */
1961         bch_cache_set_error(c, "%s", err);
1962
1963         return -EIO;
1964 }
1965
1966 static bool can_attach_cache(struct cache *ca, struct cache_set *c)
1967 {
1968         return ca->sb.block_size        == c->sb.block_size &&
1969                 ca->sb.bucket_size      == c->sb.bucket_size &&
1970                 ca->sb.nr_in_set        == c->sb.nr_in_set;
1971 }
1972
1973 static const char *register_cache_set(struct cache *ca)
1974 {
1975         char buf[12];
1976         const char *err = "cannot allocate memory";
1977         struct cache_set *c;
1978
1979         list_for_each_entry(c, &bch_cache_sets, list)
1980                 if (!memcmp(c->sb.set_uuid, ca->sb.set_uuid, 16)) {
1981                         if (c->cache[ca->sb.nr_this_dev])
1982                                 return "duplicate cache set member";
1983
1984                         if (!can_attach_cache(ca, c))
1985                                 return "cache sb does not match set";
1986
1987                         if (!CACHE_SYNC(&ca->sb))
1988                                 SET_CACHE_SYNC(&c->sb, false);
1989
1990                         goto found;
1991                 }
1992
1993         c = bch_cache_set_alloc(&ca->sb);
1994         if (!c)
1995                 return err;
1996
1997         err = "error creating kobject";
1998         if (kobject_add(&c->kobj, bcache_kobj, "%pU", c->sb.set_uuid) ||
1999             kobject_add(&c->internal, &c->kobj, "internal"))
2000                 goto err;
2001
2002         if (bch_cache_accounting_add_kobjs(&c->accounting, &c->kobj))
2003                 goto err;
2004
2005         bch_debug_init_cache_set(c);
2006
2007         list_add(&c->list, &bch_cache_sets);
2008 found:
2009         sprintf(buf, "cache%i", ca->sb.nr_this_dev);
2010         if (sysfs_create_link(&ca->kobj, &c->kobj, "set") ||
2011             sysfs_create_link(&c->kobj, &ca->kobj, buf))
2012                 goto err;
2013
2014         if (ca->sb.seq > c->sb.seq) {
2015                 c->sb.version           = ca->sb.version;
2016                 memcpy(c->sb.set_uuid, ca->sb.set_uuid, 16);
2017                 c->sb.flags             = ca->sb.flags;
2018                 c->sb.seq               = ca->sb.seq;
2019                 pr_debug("set version = %llu", c->sb.version);
2020         }
2021
2022         kobject_get(&ca->kobj);
2023         ca->set = c;
2024         ca->set->cache[ca->sb.nr_this_dev] = ca;
2025         c->cache_by_alloc[c->caches_loaded++] = ca;
2026
2027         if (c->caches_loaded == c->sb.nr_in_set) {
2028                 err = "failed to run cache set";
2029                 if (run_cache_set(c) < 0)
2030                         goto err;
2031         }
2032
2033         return NULL;
2034 err:
2035         bch_cache_set_unregister(c);
2036         return err;
2037 }
2038
2039 /* Cache device */
2040
2041 /* When ca->kobj released */
2042 void bch_cache_release(struct kobject *kobj)
2043 {
2044         struct cache *ca = container_of(kobj, struct cache, kobj);
2045         unsigned int i;
2046
2047         if (ca->set) {
2048                 BUG_ON(ca->set->cache[ca->sb.nr_this_dev] != ca);
2049                 ca->set->cache[ca->sb.nr_this_dev] = NULL;
2050         }
2051
2052         free_pages((unsigned long) ca->disk_buckets, ilog2(bucket_pages(ca)));
2053         kfree(ca->prio_buckets);
2054         vfree(ca->buckets);
2055
2056         free_heap(&ca->heap);
2057         free_fifo(&ca->free_inc);
2058
2059         for (i = 0; i < RESERVE_NR; i++)
2060                 free_fifo(&ca->free[i]);
2061
2062         if (ca->sb_bio.bi_inline_vecs[0].bv_page)
2063                 put_page(bio_first_page_all(&ca->sb_bio));
2064
2065         if (!IS_ERR_OR_NULL(ca->bdev))
2066                 blkdev_put(ca->bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2067
2068         kfree(ca);
2069         module_put(THIS_MODULE);
2070 }
2071
2072 static int cache_alloc(struct cache *ca)
2073 {
2074         size_t free;
2075         size_t btree_buckets;
2076         struct bucket *b;
2077         int ret = -ENOMEM;
2078         const char *err = NULL;
2079
2080         __module_get(THIS_MODULE);
2081         kobject_init(&ca->kobj, &bch_cache_ktype);
2082
2083         bio_init(&ca->journal.bio, ca->journal.bio.bi_inline_vecs, 8);
2084
2085         /*
2086          * when ca->sb.njournal_buckets is not zero, journal exists,
2087          * and in bch_journal_replay(), tree node may split,
2088          * so bucket of RESERVE_BTREE type is needed,
2089          * the worst situation is all journal buckets are valid journal,
2090          * and all the keys need to replay,
2091          * so the number of  RESERVE_BTREE type buckets should be as much
2092          * as journal buckets
2093          */
2094         btree_buckets = ca->sb.njournal_buckets ?: 8;
2095         free = roundup_pow_of_two(ca->sb.nbuckets) >> 10;
2096         if (!free) {
2097                 ret = -EPERM;
2098                 err = "ca->sb.nbuckets is too small";
2099                 goto err_free;
2100         }
2101
2102         if (!init_fifo(&ca->free[RESERVE_BTREE], btree_buckets,
2103                                                 GFP_KERNEL)) {
2104                 err = "ca->free[RESERVE_BTREE] alloc failed";
2105                 goto err_btree_alloc;
2106         }
2107
2108         if (!init_fifo_exact(&ca->free[RESERVE_PRIO], prio_buckets(ca),
2109                                                         GFP_KERNEL)) {
2110                 err = "ca->free[RESERVE_PRIO] alloc failed";
2111                 goto err_prio_alloc;
2112         }
2113
2114         if (!init_fifo(&ca->free[RESERVE_MOVINGGC], free, GFP_KERNEL)) {
2115                 err = "ca->free[RESERVE_MOVINGGC] alloc failed";
2116                 goto err_movinggc_alloc;
2117         }
2118
2119         if (!init_fifo(&ca->free[RESERVE_NONE], free, GFP_KERNEL)) {
2120                 err = "ca->free[RESERVE_NONE] alloc failed";
2121                 goto err_none_alloc;
2122         }
2123
2124         if (!init_fifo(&ca->free_inc, free << 2, GFP_KERNEL)) {
2125                 err = "ca->free_inc alloc failed";
2126                 goto err_free_inc_alloc;
2127         }
2128
2129         if (!init_heap(&ca->heap, free << 3, GFP_KERNEL)) {
2130                 err = "ca->heap alloc failed";
2131                 goto err_heap_alloc;
2132         }
2133
2134         ca->buckets = vzalloc(array_size(sizeof(struct bucket),
2135                               ca->sb.nbuckets));
2136         if (!ca->buckets) {
2137                 err = "ca->buckets alloc failed";
2138                 goto err_buckets_alloc;
2139         }
2140
2141         ca->prio_buckets = kzalloc(array3_size(sizeof(uint64_t),
2142                                    prio_buckets(ca), 2),
2143                                    GFP_KERNEL);
2144         if (!ca->prio_buckets) {
2145                 err = "ca->prio_buckets alloc failed";
2146                 goto err_prio_buckets_alloc;
2147         }
2148
2149         ca->disk_buckets = alloc_bucket_pages(GFP_KERNEL, ca);
2150         if (!ca->disk_buckets) {
2151                 err = "ca->disk_buckets alloc failed";
2152                 goto err_disk_buckets_alloc;
2153         }
2154
2155         ca->prio_last_buckets = ca->prio_buckets + prio_buckets(ca);
2156
2157         for_each_bucket(b, ca)
2158                 atomic_set(&b->pin, 0);
2159         return 0;
2160
2161 err_disk_buckets_alloc:
2162         kfree(ca->prio_buckets);
2163 err_prio_buckets_alloc:
2164         vfree(ca->buckets);
2165 err_buckets_alloc:
2166         free_heap(&ca->heap);
2167 err_heap_alloc:
2168         free_fifo(&ca->free_inc);
2169 err_free_inc_alloc:
2170         free_fifo(&ca->free[RESERVE_NONE]);
2171 err_none_alloc:
2172         free_fifo(&ca->free[RESERVE_MOVINGGC]);
2173 err_movinggc_alloc:
2174         free_fifo(&ca->free[RESERVE_PRIO]);
2175 err_prio_alloc:
2176         free_fifo(&ca->free[RESERVE_BTREE]);
2177 err_btree_alloc:
2178 err_free:
2179         module_put(THIS_MODULE);
2180         if (err)
2181                 pr_notice("error %s: %s", ca->cache_dev_name, err);
2182         return ret;
2183 }
2184
2185 static int register_cache(struct cache_sb *sb, struct page *sb_page,
2186                                 struct block_device *bdev, struct cache *ca)
2187 {
2188         const char *err = NULL; /* must be set for any error case */
2189         int ret = 0;
2190
2191         bdevname(bdev, ca->cache_dev_name);
2192         memcpy(&ca->sb, sb, sizeof(struct cache_sb));
2193         ca->bdev = bdev;
2194         ca->bdev->bd_holder = ca;
2195
2196         bio_init(&ca->sb_bio, ca->sb_bio.bi_inline_vecs, 1);
2197         bio_first_bvec_all(&ca->sb_bio)->bv_page = sb_page;
2198         get_page(sb_page);
2199
2200         if (blk_queue_discard(bdev_get_queue(bdev)))
2201                 ca->discard = CACHE_DISCARD(&ca->sb);
2202
2203         ret = cache_alloc(ca);
2204         if (ret != 0) {
2205                 /*
2206                  * If we failed here, it means ca->kobj is not initialized yet,
2207                  * kobject_put() won't be called and there is no chance to
2208                  * call blkdev_put() to bdev in bch_cache_release(). So we
2209                  * explicitly call blkdev_put() here.
2210                  */
2211                 blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2212                 if (ret == -ENOMEM)
2213                         err = "cache_alloc(): -ENOMEM";
2214                 else if (ret == -EPERM)
2215                         err = "cache_alloc(): cache device is too small";
2216                 else
2217                         err = "cache_alloc(): unknown error";
2218                 goto err;
2219         }
2220
2221         if (kobject_add(&ca->kobj,
2222                         &part_to_dev(bdev->bd_part)->kobj,
2223                         "bcache")) {
2224                 err = "error calling kobject_add";
2225                 ret = -ENOMEM;
2226                 goto out;
2227         }
2228
2229         mutex_lock(&bch_register_lock);
2230         err = register_cache_set(ca);
2231         mutex_unlock(&bch_register_lock);
2232
2233         if (err) {
2234                 ret = -ENODEV;
2235                 goto out;
2236         }
2237
2238         pr_info("registered cache device %s", ca->cache_dev_name);
2239
2240 out:
2241         kobject_put(&ca->kobj);
2242
2243 err:
2244         if (err)
2245                 pr_notice("error %s: %s", ca->cache_dev_name, err);
2246
2247         return ret;
2248 }
2249
2250 /* Global interfaces/init */
2251
2252 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2253                                const char *buffer, size_t size);
2254
2255 kobj_attribute_write(register,          register_bcache);
2256 kobj_attribute_write(register_quiet,    register_bcache);
2257
2258 static bool bch_is_open_backing(struct block_device *bdev)
2259 {
2260         struct cache_set *c, *tc;
2261         struct cached_dev *dc, *t;
2262
2263         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2264                 list_for_each_entry_safe(dc, t, &c->cached_devs, list)
2265                         if (dc->bdev == bdev)
2266                                 return true;
2267         list_for_each_entry_safe(dc, t, &uncached_devices, list)
2268                 if (dc->bdev == bdev)
2269                         return true;
2270         return false;
2271 }
2272
2273 static bool bch_is_open_cache(struct block_device *bdev)
2274 {
2275         struct cache_set *c, *tc;
2276         struct cache *ca;
2277         unsigned int i;
2278
2279         list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2280                 for_each_cache(ca, c, i)
2281                         if (ca->bdev == bdev)
2282                                 return true;
2283         return false;
2284 }
2285
2286 static bool bch_is_open(struct block_device *bdev)
2287 {
2288         return bch_is_open_cache(bdev) || bch_is_open_backing(bdev);
2289 }
2290
2291 static ssize_t register_bcache(struct kobject *k, struct kobj_attribute *attr,
2292                                const char *buffer, size_t size)
2293 {
2294         ssize_t ret = -EINVAL;
2295         const char *err = "cannot allocate memory";
2296         char *path = NULL;
2297         struct cache_sb *sb = NULL;
2298         struct block_device *bdev = NULL;
2299         struct page *sb_page = NULL;
2300
2301         if (!try_module_get(THIS_MODULE))
2302                 return -EBUSY;
2303
2304         path = kstrndup(buffer, size, GFP_KERNEL);
2305         if (!path)
2306                 goto err;
2307
2308         sb = kmalloc(sizeof(struct cache_sb), GFP_KERNEL);
2309         if (!sb)
2310                 goto err;
2311
2312         err = "failed to open device";
2313         bdev = blkdev_get_by_path(strim(path),
2314                                   FMODE_READ|FMODE_WRITE|FMODE_EXCL,
2315                                   sb);
2316         if (IS_ERR(bdev)) {
2317                 if (bdev == ERR_PTR(-EBUSY)) {
2318                         bdev = lookup_bdev(strim(path));
2319                         mutex_lock(&bch_register_lock);
2320                         if (!IS_ERR(bdev) && bch_is_open(bdev))
2321                                 err = "device already registered";
2322                         else
2323                                 err = "device busy";
2324                         mutex_unlock(&bch_register_lock);
2325                         if (!IS_ERR(bdev))
2326                                 bdput(bdev);
2327                         if (attr == &ksysfs_register_quiet)
2328                                 goto quiet_out;
2329                 }
2330                 goto err;
2331         }
2332
2333         err = "failed to set blocksize";
2334         if (set_blocksize(bdev, 4096))
2335                 goto err_close;
2336
2337         err = read_super(sb, bdev, &sb_page);
2338         if (err)
2339                 goto err_close;
2340
2341         err = "failed to register device";
2342         if (SB_IS_BDEV(sb)) {
2343                 struct cached_dev *dc = kzalloc(sizeof(*dc), GFP_KERNEL);
2344
2345                 if (!dc)
2346                         goto err_close;
2347
2348                 mutex_lock(&bch_register_lock);
2349                 ret = register_bdev(sb, sb_page, bdev, dc);
2350                 mutex_unlock(&bch_register_lock);
2351                 /* blkdev_put() will be called in cached_dev_free() */
2352                 if (ret < 0)
2353                         goto err;
2354         } else {
2355                 struct cache *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
2356
2357                 if (!ca)
2358                         goto err_close;
2359
2360                 /* blkdev_put() will be called in bch_cache_release() */
2361                 if (register_cache(sb, sb_page, bdev, ca) != 0)
2362                         goto err;
2363         }
2364 quiet_out:
2365         ret = size;
2366 out:
2367         if (sb_page)
2368                 put_page(sb_page);
2369         kfree(sb);
2370         kfree(path);
2371         module_put(THIS_MODULE);
2372         return ret;
2373
2374 err_close:
2375         blkdev_put(bdev, FMODE_READ|FMODE_WRITE|FMODE_EXCL);
2376 err:
2377         pr_info("error %s: %s", path, err);
2378         goto out;
2379 }
2380
2381 static int bcache_reboot(struct notifier_block *n, unsigned long code, void *x)
2382 {
2383         if (code == SYS_DOWN ||
2384             code == SYS_HALT ||
2385             code == SYS_POWER_OFF) {
2386                 DEFINE_WAIT(wait);
2387                 unsigned long start = jiffies;
2388                 bool stopped = false;
2389
2390                 struct cache_set *c, *tc;
2391                 struct cached_dev *dc, *tdc;
2392
2393                 mutex_lock(&bch_register_lock);
2394
2395                 if (list_empty(&bch_cache_sets) &&
2396                     list_empty(&uncached_devices))
2397                         goto out;
2398
2399                 pr_info("Stopping all devices:");
2400
2401                 list_for_each_entry_safe(c, tc, &bch_cache_sets, list)
2402                         bch_cache_set_stop(c);
2403
2404                 list_for_each_entry_safe(dc, tdc, &uncached_devices, list)
2405                         bcache_device_stop(&dc->disk);
2406
2407                 mutex_unlock(&bch_register_lock);
2408
2409                 /*
2410                  * Give an early chance for other kthreads and
2411                  * kworkers to stop themselves
2412                  */
2413                 schedule();
2414
2415                 /* What's a condition variable? */
2416                 while (1) {
2417                         long timeout = start + 10 * HZ - jiffies;
2418
2419                         mutex_lock(&bch_register_lock);
2420                         stopped = list_empty(&bch_cache_sets) &&
2421                                 list_empty(&uncached_devices);
2422
2423                         if (timeout < 0 || stopped)
2424                                 break;
2425
2426                         prepare_to_wait(&unregister_wait, &wait,
2427                                         TASK_UNINTERRUPTIBLE);
2428
2429                         mutex_unlock(&bch_register_lock);
2430                         schedule_timeout(timeout);
2431                 }
2432
2433                 finish_wait(&unregister_wait, &wait);
2434
2435                 if (stopped)
2436                         pr_info("All devices stopped");
2437                 else
2438                         pr_notice("Timeout waiting for devices to be closed");
2439 out:
2440                 mutex_unlock(&bch_register_lock);
2441         }
2442
2443         return NOTIFY_DONE;
2444 }
2445
2446 static struct notifier_block reboot = {
2447         .notifier_call  = bcache_reboot,
2448         .priority       = INT_MAX, /* before any real devices */
2449 };
2450
2451 static void bcache_exit(void)
2452 {
2453         bch_debug_exit();
2454         bch_request_exit();
2455         if (bcache_kobj)
2456                 kobject_put(bcache_kobj);
2457         if (bcache_wq)
2458                 destroy_workqueue(bcache_wq);
2459         if (bch_journal_wq)
2460                 destroy_workqueue(bch_journal_wq);
2461
2462         if (bcache_major)
2463                 unregister_blkdev(bcache_major, "bcache");
2464         unregister_reboot_notifier(&reboot);
2465         mutex_destroy(&bch_register_lock);
2466 }
2467
2468 /* Check and fixup module parameters */
2469 static void check_module_parameters(void)
2470 {
2471         if (bch_cutoff_writeback_sync == 0)
2472                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC;
2473         else if (bch_cutoff_writeback_sync > CUTOFF_WRITEBACK_SYNC_MAX) {
2474                 pr_warn("set bch_cutoff_writeback_sync (%u) to max value %u",
2475                         bch_cutoff_writeback_sync, CUTOFF_WRITEBACK_SYNC_MAX);
2476                 bch_cutoff_writeback_sync = CUTOFF_WRITEBACK_SYNC_MAX;
2477         }
2478
2479         if (bch_cutoff_writeback == 0)
2480                 bch_cutoff_writeback = CUTOFF_WRITEBACK;
2481         else if (bch_cutoff_writeback > CUTOFF_WRITEBACK_MAX) {
2482                 pr_warn("set bch_cutoff_writeback (%u) to max value %u",
2483                         bch_cutoff_writeback, CUTOFF_WRITEBACK_MAX);
2484                 bch_cutoff_writeback = CUTOFF_WRITEBACK_MAX;
2485         }
2486
2487         if (bch_cutoff_writeback > bch_cutoff_writeback_sync) {
2488                 pr_warn("set bch_cutoff_writeback (%u) to %u",
2489                         bch_cutoff_writeback, bch_cutoff_writeback_sync);
2490                 bch_cutoff_writeback = bch_cutoff_writeback_sync;
2491         }
2492 }
2493
2494 static int __init bcache_init(void)
2495 {
2496         static const struct attribute *files[] = {
2497                 &ksysfs_register.attr,
2498                 &ksysfs_register_quiet.attr,
2499                 NULL
2500         };
2501
2502         check_module_parameters();
2503
2504         mutex_init(&bch_register_lock);
2505         init_waitqueue_head(&unregister_wait);
2506         register_reboot_notifier(&reboot);
2507
2508         bcache_major = register_blkdev(0, "bcache");
2509         if (bcache_major < 0) {
2510                 unregister_reboot_notifier(&reboot);
2511                 mutex_destroy(&bch_register_lock);
2512                 return bcache_major;
2513         }
2514
2515         bcache_wq = alloc_workqueue("bcache", WQ_MEM_RECLAIM, 0);
2516         if (!bcache_wq)
2517                 goto err;
2518
2519         bch_journal_wq = alloc_workqueue("bch_journal", WQ_MEM_RECLAIM, 0);
2520         if (!bch_journal_wq)
2521                 goto err;
2522
2523         bcache_kobj = kobject_create_and_add("bcache", fs_kobj);
2524         if (!bcache_kobj)
2525                 goto err;
2526
2527         if (bch_request_init() ||
2528             sysfs_create_files(bcache_kobj, files))
2529                 goto err;
2530
2531         bch_debug_init();
2532         closure_debug_init();
2533
2534         return 0;
2535 err:
2536         bcache_exit();
2537         return -ENOMEM;
2538 }
2539
2540 /*
2541  * Module hooks
2542  */
2543 module_exit(bcache_exit);
2544 module_init(bcache_init);
2545
2546 module_param(bch_cutoff_writeback, uint, 0);
2547 MODULE_PARM_DESC(bch_cutoff_writeback, "threshold to cutoff writeback");
2548
2549 module_param(bch_cutoff_writeback_sync, uint, 0);
2550 MODULE_PARM_DESC(bch_cutoff_writeback_sync, "hard threshold to cutoff writeback");
2551
2552 MODULE_DESCRIPTION("Bcache: a Linux block layer cache");
2553 MODULE_AUTHOR("Kent Overstreet <kent.overstreet@gmail.com>");
2554 MODULE_LICENSE("GPL");