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