drivers/md/bcache/journal.c

   1 /*
   2  * bcache journalling code, for btree insertions
   3  *
   4  * Copyright 2012 Google, Inc.
   5  */
   6
   7 #include "bcache.h"
   8 #include "btree.h"
   9 #include "debug.h"
  10
  11 #include <trace/events/bcache.h>
  12
  13 /*
  14  * Journal replay/recovery:
  15  *
  16  * This code is all driven from run_cache_set(); we first read the journal
  17  * entries, do some other stuff, then we mark all the keys in the journal
  18  * entries (same as garbage collection would), then we replay them - reinserting
  19  * them into the cache in precisely the same order as they appear in the
  20  * journal.
  21  *
  22  * We only journal keys that go in leaf nodes, which simplifies things quite a
  23  * bit.
  24  */
  25
  26 static void journal_read_endio(struct bio *bio, int error)
  27 {
  28         struct closure *cl = bio->bi_private;
  29         closure_put(cl);
  30 }
  31
  32 static int journal_read_bucket(struct cache *ca, struct list_head *list,
  33                                unsigned bucket_index)
  34 {
  35         struct journal_device *ja = &ca->journal;
  36         struct bio *bio = &ja->bio;
  37
  38         struct journal_replay *i;
  39         struct jset *j, *data = ca->set->journal.w[0].data;
  40         struct closure cl;
  41         unsigned len, left, offset = 0;
  42         int ret = 0;
  43         sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
  44
  45         closure_init_stack(&cl);
  46
  47         pr_debug("reading %u", bucket_index);
  48
  49         while (offset < ca->sb.bucket_size) {
  50 reread:         left = ca->sb.bucket_size - offset;
  51                 len = min_t(unsigned, left, PAGE_SECTORS << JSET_BITS);
  52
  53                 bio_reset(bio);
  54                 bio->bi_iter.bi_sector  = bucket + offset;
  55                 bio->bi_bdev    = ca->bdev;
  56                 bio->bi_rw      = READ;
  57                 bio->bi_iter.bi_size    = len << 9;
  58
  59                 bio->bi_end_io  = journal_read_endio;
  60                 bio->bi_private = &cl;
  61                 bch_bio_map(bio, data);
  62
  63                 closure_bio_submit(bio, &cl, ca);
  64                 closure_sync(&cl);
  65
  66                 /* This function could be simpler now since we no longer write
  67                  * journal entries that overlap bucket boundaries; this means
  68                  * the start of a bucket will always have a valid journal entry
  69                  * if it has any journal entries at all.
  70                  */
  71
  72                 j = data;
  73                 while (len) {
  74                         struct list_head *where;
  75                         size_t blocks, bytes = set_bytes(j);
  76
  77                         if (j->magic != jset_magic(&ca->sb)) {
  78                                 pr_debug("%u: bad magic", bucket_index);
  79                                 return ret;
  80                         }
  81
  82                         if (bytes > left << 9 ||
  83                             bytes > PAGE_SIZE << JSET_BITS) {
  84                                 pr_info("%u: too big, %zu bytes, offset %u",
  85                                         bucket_index, bytes, offset);
  86                                 return ret;
  87                         }
  88
  89                         if (bytes > len << 9)
  90                                 goto reread;
  91
  92                         if (j->csum != csum_set(j)) {
  93                                 pr_info("%u: bad csum, %zu bytes, offset %u",
  94                                         bucket_index, bytes, offset);
  95                                 return ret;
  96                         }
  97
  98                         blocks = set_blocks(j, block_bytes(ca->set));
  99
 100                         while (!list_empty(list)) {
 101                                 i = list_first_entry(list,
 102                                         struct journal_replay, list);
 103                                 if (i->j.seq >= j->last_seq)
 104                                         break;
 105                                 list_del(&i->list);
 106                                 kfree(i);
 107                         }
 108
 109                         list_for_each_entry_reverse(i, list, list) {
 110                                 if (j->seq == i->j.seq)
 111                                         goto next_set;
 112
 113                                 if (j->seq < i->j.last_seq)
 114                                         goto next_set;
 115
 116                                 if (j->seq > i->j.seq) {
 117                                         where = &i->list;
 118                                         goto add;
 119                                 }
 120                         }
 121
 122                         where = list;
 123 add:
 124                         i = kmalloc(offsetof(struct journal_replay, j) +
 125                                     bytes, GFP_KERNEL);
 126                         if (!i)
 127                                 return -ENOMEM;
 128                         memcpy(&i->j, j, bytes);
 129                         list_add(&i->list, where);
 130                         ret = 1;
 131
 132                         ja->seq[bucket_index] = j->seq;
 133 next_set:
 134                         offset  += blocks * ca->sb.block_size;
 135                         len     -= blocks * ca->sb.block_size;
 136                         j = ((void *) j) + blocks * block_bytes(ca);
 137                 }
 138         }
 139
 140         return ret;
 141 }
 142
 143 int bch_journal_read(struct cache_set *c, struct list_head *list)
 144 {
 145 #define read_bucket(b)                                                  \
 146         ({                                                              \
 147                 int ret = journal_read_bucket(ca, list, b);             \
 148                 __set_bit(b, bitmap);                                   \
 149                 if (ret < 0)                                            \
 150                         return ret;                                     \
 151                 ret;                                                    \
 152         })
 153
 154         struct cache *ca;
 155         unsigned iter;
 156
 157         for_each_cache(ca, c, iter) {
 158                 struct journal_device *ja = &ca->journal;
 159                 unsigned long bitmap[SB_JOURNAL_BUCKETS / BITS_PER_LONG];
 160                 unsigned i, l, r, m;
 161                 uint64_t seq;
 162
 163                 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
 164                 pr_debug("%u journal buckets", ca->sb.njournal_buckets);
 165
 166                 /*
 167                  * Read journal buckets ordered by golden ratio hash to quickly
 168                  * find a sequence of buckets with valid journal entries
 169                  */
 170                 for (i = 0; i < ca->sb.njournal_buckets; i++) {
 171                         l = (i * 2654435769U) % ca->sb.njournal_buckets;
 172
 173                         if (test_bit(l, bitmap))
 174                                 break;
 175
 176                         if (read_bucket(l))
 177                                 goto bsearch;
 178                 }
 179
 180                 /*
 181                  * If that fails, check all the buckets we haven't checked
 182                  * already
 183                  */
 184                 pr_debug("falling back to linear search");
 185
 186                 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
 187                      l < ca->sb.njournal_buckets;
 188                      l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1))
 189                         if (read_bucket(l))
 190                                 goto bsearch;
 191
 192                 if (list_empty(list))
 193                         continue;
 194 bsearch:
 195                 /* Binary search */
 196                 m = r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 197                 pr_debug("starting binary search, l %u r %u", l, r);
 198
 199                 while (l + 1 < r) {
 200                         seq = list_entry(list->prev, struct journal_replay,
 201                                          list)->j.seq;
 202
 203                         m = (l + r) >> 1;
 204                         read_bucket(m);
 205
 206                         if (seq != list_entry(list->prev, struct journal_replay,
 207                                               list)->j.seq)
 208                                 l = m;
 209                         else
 210                                 r = m;
 211                 }
 212
 213                 /*
 214                  * Read buckets in reverse order until we stop finding more
 215                  * journal entries
 216                  */
 217                 pr_debug("finishing up: m %u njournal_buckets %u",
 218                          m, ca->sb.njournal_buckets);
 219                 l = m;
 220
 221                 while (1) {
 222                         if (!l--)
 223                                 l = ca->sb.njournal_buckets - 1;
 224
 225                         if (l == m)
 226                                 break;
 227
 228                         if (test_bit(l, bitmap))
 229                                 continue;
 230
 231                         if (!read_bucket(l))
 232                                 break;
 233                 }
 234
 235                 seq = 0;
 236
 237                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 238                         if (ja->seq[i] > seq) {
 239                                 seq = ja->seq[i];
 240                                 ja->cur_idx = ja->discard_idx =
 241                                         ja->last_idx = i;
 242
 243                         }
 244         }
 245
 246         if (!list_empty(list))
 247                 c->journal.seq = list_entry(list->prev,
 248                                             struct journal_replay,
 249                                             list)->j.seq;
 250
 251         return 0;
 252 #undef read_bucket
 253 }
 254
 255 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 256 {
 257         atomic_t p = { 0 };
 258         struct bkey *k;
 259         struct journal_replay *i;
 260         struct journal *j = &c->journal;
 261         uint64_t last = j->seq;
 262
 263         /*
 264          * journal.pin should never fill up - we never write a journal
 265          * entry when it would fill up. But if for some reason it does, we
 266          * iterate over the list in reverse order so that we can just skip that
 267          * refcount instead of bugging.
 268          */
 269
 270         list_for_each_entry_reverse(i, list, list) {
 271                 BUG_ON(last < i->j.seq);
 272                 i->pin = NULL;
 273
 274                 while (last-- != i->j.seq)
 275                         if (fifo_free(&j->pin) > 1) {
 276                                 fifo_push_front(&j->pin, p);
 277                                 atomic_set(&fifo_front(&j->pin), 0);
 278                         }
 279
 280                 if (fifo_free(&j->pin) > 1) {
 281                         fifo_push_front(&j->pin, p);
 282                         i->pin = &fifo_front(&j->pin);
 283                         atomic_set(i->pin, 1);
 284                 }
 285
 286                 for (k = i->j.start;
 287                      k < bset_bkey_last(&i->j);
 288                      k = bkey_next(k)) {
 289                         unsigned j;
 290                         struct bucket *g;
 291
 292                         for (j = 0; j < KEY_PTRS(k); j++) {
 293                                 if (!ptr_available(c, k, j))
 294                                         continue;
 295
 296                                 g = PTR_BUCKET(c, k, j);
 297                                 atomic_inc(&g->pin);
 298
 299                                 if (g->prio == BTREE_PRIO &&
 300                                     !ptr_stale(c, k, j))
 301                                         g->prio = INITIAL_PRIO;
 302                         }
 303
 304                         __bch_btree_mark_key(c, 0, k);
 305                 }
 306         }
 307 }
 308
 309 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 310 {
 311         int ret = 0, keys = 0, entries = 0;
 312         struct bkey *k;
 313         struct journal_replay *i =
 314                 list_entry(list->prev, struct journal_replay, list);
 315
 316         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 317         struct keylist keylist;
 318
 319         bch_keylist_init(&keylist);
 320
 321         list_for_each_entry(i, list, list) {
 322                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 323
 324                 cache_set_err_on(n != i->j.seq, s,
 325 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 326                                  n, i->j.seq - 1, start, end);
 327
 328                 for (k = i->j.start;
 329                      k < bset_bkey_last(&i->j);
 330                      k = bkey_next(k)) {
 331                         trace_bcache_journal_replay_key(k);
 332
 333                         bkey_copy(keylist.top, k);
 334                         bch_keylist_push(&keylist);
 335
 336                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 337                         if (ret)
 338                                 goto err;
 339
 340                         BUG_ON(!bch_keylist_empty(&keylist));
 341                         keys++;
 342
 343                         cond_resched();
 344                 }
 345
 346                 if (i->pin)
 347                         atomic_dec(i->pin);
 348                 n = i->j.seq + 1;
 349                 entries++;
 350         }
 351
 352         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 353                 keys, entries, end);
 354 err:
 355         while (!list_empty(list)) {
 356                 i = list_first_entry(list, struct journal_replay, list);
 357                 list_del(&i->list);
 358                 kfree(i);
 359         }
 360
 361         return ret;
 362 }
 363
 364 /* Journalling */
 365
 366 static void btree_flush_write(struct cache_set *c)
 367 {
 368         /*
 369          * Try to find the btree node with that references the oldest journal
 370          * entry, best is our current candidate and is locked if non NULL:
 371          */
 372         struct btree *b, *best;
 373         unsigned i;
 374 retry:
 375         best = NULL;
 376
 377         for_each_cached_btree(b, c, i)
 378                 if (btree_current_write(b)->journal) {
 379                         if (!best)
 380                                 best = b;
 381                         else if (journal_pin_cmp(c,
 382                                         btree_current_write(best)->journal,
 383                                         btree_current_write(b)->journal)) {
 384                                 best = b;
 385                         }
 386                 }
 387
 388         b = best;
 389         if (b) {
 390                 rw_lock(true, b, b->level);
 391
 392                 if (!btree_current_write(b)->journal) {
 393                         rw_unlock(true, b);
 394                         /* We raced */
 395                         goto retry;
 396                 }
 397
 398                 bch_btree_node_write(b, NULL);
 399                 rw_unlock(true, b);
 400         }
 401 }
 402
 403 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 404
 405 static void journal_discard_endio(struct bio *bio, int error)
 406 {
 407         struct journal_device *ja =
 408                 container_of(bio, struct journal_device, discard_bio);
 409         struct cache *ca = container_of(ja, struct cache, journal);
 410
 411         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 412
 413         closure_wake_up(&ca->set->journal.wait);
 414         closure_put(&ca->set->cl);
 415 }
 416
 417 static void journal_discard_work(struct work_struct *work)
 418 {
 419         struct journal_device *ja =
 420                 container_of(work, struct journal_device, discard_work);
 421
 422         submit_bio(0, &ja->discard_bio);
 423 }
 424
 425 static void do_journal_discard(struct cache *ca)
 426 {
 427         struct journal_device *ja = &ca->journal;
 428         struct bio *bio = &ja->discard_bio;
 429
 430         if (!ca->discard) {
 431                 ja->discard_idx = ja->last_idx;
 432                 return;
 433         }
 434
 435         switch (atomic_read(&ja->discard_in_flight)) {
 436         case DISCARD_IN_FLIGHT:
 437                 return;
 438
 439         case DISCARD_DONE:
 440                 ja->discard_idx = (ja->discard_idx + 1) %
 441                         ca->sb.njournal_buckets;
 442
 443                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 444                 /* fallthrough */
 445
 446         case DISCARD_READY:
 447                 if (ja->discard_idx == ja->last_idx)
 448                         return;
 449
 450                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 451
 452                 bio_init(bio);
 453                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 454                                                 ca->sb.d[ja->discard_idx]);
 455                 bio->bi_bdev            = ca->bdev;
 456                 bio->bi_rw              = REQ_WRITE|REQ_DISCARD;
 457                 bio->bi_max_vecs        = 1;
 458                 bio->bi_io_vec          = bio->bi_inline_vecs;
 459                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 460                 bio->bi_end_io          = journal_discard_endio;
 461
 462                 closure_get(&ca->set->cl);
 463                 INIT_WORK(&ja->discard_work, journal_discard_work);
 464                 schedule_work(&ja->discard_work);
 465         }
 466 }
 467
 468 static void journal_reclaim(struct cache_set *c)
 469 {
 470         struct bkey *k = &c->journal.key;
 471         struct cache *ca;
 472         uint64_t last_seq;
 473         unsigned iter, n = 0;
 474         atomic_t p;
 475
 476         while (!atomic_read(&fifo_front(&c->journal.pin)))
 477                 fifo_pop(&c->journal.pin, p);
 478
 479         last_seq = last_seq(&c->journal);
 480
 481         /* Update last_idx */
 482
 483         for_each_cache(ca, c, iter) {
 484                 struct journal_device *ja = &ca->journal;
 485
 486                 while (ja->last_idx != ja->cur_idx &&
 487                        ja->seq[ja->last_idx] < last_seq)
 488                         ja->last_idx = (ja->last_idx + 1) %
 489                                 ca->sb.njournal_buckets;
 490         }
 491
 492         for_each_cache(ca, c, iter)
 493                 do_journal_discard(ca);
 494
 495         if (c->journal.blocks_free)
 496                 goto out;
 497
 498         /*
 499          * Allocate:
 500          * XXX: Sort by free journal space
 501          */
 502
 503         for_each_cache(ca, c, iter) {
 504                 struct journal_device *ja = &ca->journal;
 505                 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 506
 507                 /* No space available on this device */
 508                 if (next == ja->discard_idx)
 509                         continue;
 510
 511                 ja->cur_idx = next;
 512                 k->ptr[n++] = PTR(0,
 513                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 514                                   ca->sb.nr_this_dev);
 515         }
 516
 517         bkey_init(k);
 518         SET_KEY_PTRS(k, n);
 519
 520         if (n)
 521                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 522 out:
 523         if (!journal_full(&c->journal))
 524                 __closure_wake_up(&c->journal.wait);
 525 }
 526
 527 void bch_journal_next(struct journal *j)
 528 {
 529         atomic_t p = { 1 };
 530
 531         j->cur = (j->cur == j->w)
 532                 ? &j->w[1]
 533                 : &j->w[0];
 534
 535         /*
 536          * The fifo_push() needs to happen at the same time as j->seq is
 537          * incremented for last_seq() to be calculated correctly
 538          */
 539         BUG_ON(!fifo_push(&j->pin, p));
 540         atomic_set(&fifo_back(&j->pin), 1);
 541
 542         j->cur->data->seq       = ++j->seq;
 543         j->cur->dirty           = false;
 544         j->cur->need_write      = false;
 545         j->cur->data->keys      = 0;
 546
 547         if (fifo_full(&j->pin))
 548                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 549 }
 550
 551 static void journal_write_endio(struct bio *bio, int error)
 552 {
 553         struct journal_write *w = bio->bi_private;
 554
 555         cache_set_err_on(error, w->c, "journal io error");
 556         closure_put(&w->c->journal.io);
 557 }
 558
 559 static void journal_write(struct closure *);
 560
 561 static void journal_write_done(struct closure *cl)
 562 {
 563         struct journal *j = container_of(cl, struct journal, io);
 564         struct journal_write *w = (j->cur == j->w)
 565                 ? &j->w[1]
 566                 : &j->w[0];
 567
 568         __closure_wake_up(&w->wait);
 569         continue_at_nobarrier(cl, journal_write, system_wq);
 570 }
 571
 572 static void journal_write_unlock(struct closure *cl)
 573 {
 574         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 575
 576         c->journal.io_in_flight = 0;
 577         spin_unlock(&c->journal.lock);
 578 }
 579
 580 static void journal_write_unlocked(struct closure *cl)
 581         __releases(c->journal.lock)
 582 {
 583         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 584         struct cache *ca;
 585         struct journal_write *w = c->journal.cur;
 586         struct bkey *k = &c->journal.key;
 587         unsigned i, sectors = set_blocks(w->data, block_bytes(c)) *
 588                 c->sb.block_size;
 589
 590         struct bio *bio;
 591         struct bio_list list;
 592         bio_list_init(&list);
 593
 594         if (!w->need_write) {
 595                 closure_return_with_destructor(cl, journal_write_unlock);
 596         } else if (journal_full(&c->journal)) {
 597                 journal_reclaim(c);
 598                 spin_unlock(&c->journal.lock);
 599
 600                 btree_flush_write(c);
 601                 continue_at(cl, journal_write, system_wq);
 602         }
 603
 604         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 605
 606         w->data->btree_level = c->root->level;
 607
 608         bkey_copy(&w->data->btree_root, &c->root->key);
 609         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 610
 611         for_each_cache(ca, c, i)
 612                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 613
 614         w->data->magic          = jset_magic(&c->sb);
 615         w->data->version        = BCACHE_JSET_VERSION;
 616         w->data->last_seq       = last_seq(&c->journal);
 617         w->data->csum           = csum_set(w->data);
 618
 619         for (i = 0; i < KEY_PTRS(k); i++) {
 620                 ca = PTR_CACHE(c, k, i);
 621                 bio = &ca->journal.bio;
 622
 623                 atomic_long_add(sectors, &ca->meta_sectors_written);
 624
 625                 bio_reset(bio);
 626                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 627                 bio->bi_bdev    = ca->bdev;
 628                 bio->bi_rw      = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA;
 629                 bio->bi_iter.bi_size = sectors << 9;
 630
 631                 bio->bi_end_io  = journal_write_endio;
 632                 bio->bi_private = w;
 633                 bch_bio_map(bio, w->data);
 634
 635                 trace_bcache_journal_write(bio);
 636                 bio_list_add(&list, bio);
 637
 638                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 639
 640                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 641         }
 642
 643         atomic_dec_bug(&fifo_back(&c->journal.pin));
 644         bch_journal_next(&c->journal);
 645         journal_reclaim(c);
 646
 647         spin_unlock(&c->journal.lock);
 648
 649         while ((bio = bio_list_pop(&list)))
 650                 closure_bio_submit(bio, cl, c->cache[0]);
 651
 652         continue_at(cl, journal_write_done, NULL);
 653 }
 654
 655 static void journal_write(struct closure *cl)
 656 {
 657         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 658
 659         spin_lock(&c->journal.lock);
 660         journal_write_unlocked(cl);
 661 }
 662
 663 static void journal_try_write(struct cache_set *c)
 664         __releases(c->journal.lock)
 665 {
 666         struct closure *cl = &c->journal.io;
 667         struct journal_write *w = c->journal.cur;
 668
 669         w->need_write = true;
 670
 671         if (!c->journal.io_in_flight) {
 672                 c->journal.io_in_flight = 1;
 673                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 674         } else {
 675                 spin_unlock(&c->journal.lock);
 676         }
 677 }
 678
 679 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 680                                                     unsigned nkeys)
 681 {
 682         size_t sectors;
 683         struct closure cl;
 684         bool wait = false;
 685
 686         closure_init_stack(&cl);
 687
 688         spin_lock(&c->journal.lock);
 689
 690         while (1) {
 691                 struct journal_write *w = c->journal.cur;
 692
 693                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 694                                        block_bytes(c)) * c->sb.block_size;
 695
 696                 if (sectors <= min_t(size_t,
 697                                      c->journal.blocks_free * c->sb.block_size,
 698                                      PAGE_SECTORS << JSET_BITS))
 699                         return w;
 700
 701                 if (wait)
 702                         closure_wait(&c->journal.wait, &cl);
 703
 704                 if (!journal_full(&c->journal)) {
 705                         if (wait)
 706                                 trace_bcache_journal_entry_full(c);
 707
 708                         /*
 709                          * XXX: If we were inserting so many keys that they
 710                          * won't fit in an _empty_ journal write, we'll
 711                          * deadlock. For now, handle this in
 712                          * bch_keylist_realloc() - but something to think about.
 713                          */
 714                         BUG_ON(!w->data->keys);
 715
 716                         journal_try_write(c); /* unlocks */
 717                 } else {
 718                         if (wait)
 719                                 trace_bcache_journal_full(c);
 720
 721                         journal_reclaim(c);
 722                         spin_unlock(&c->journal.lock);
 723
 724                         btree_flush_write(c);
 725                 }
 726
 727                 closure_sync(&cl);
 728                 spin_lock(&c->journal.lock);
 729                 wait = true;
 730         }
 731 }
 732
 733 static void journal_write_work(struct work_struct *work)
 734 {
 735         struct cache_set *c = container_of(to_delayed_work(work),
 736                                            struct cache_set,
 737                                            journal.work);
 738         spin_lock(&c->journal.lock);
 739         if (c->journal.cur->dirty)
 740                 journal_try_write(c);
 741         else
 742                 spin_unlock(&c->journal.lock);
 743 }
 744
 745 /*
 746  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 747  * pass bch_journal() a list of keys to be journalled, and then
 748  * bch_journal() hands those same keys off to btree_insert_async()
 749  */
 750
 751 atomic_t *bch_journal(struct cache_set *c,
 752                       struct keylist *keys,
 753                       struct closure *parent)
 754 {
 755         struct journal_write *w;
 756         atomic_t *ret;
 757
 758         if (!CACHE_SYNC(&c->sb))
 759                 return NULL;
 760
 761         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 762
 763         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 764         w->data->keys += bch_keylist_nkeys(keys);
 765
 766         ret = &fifo_back(&c->journal.pin);
 767         atomic_inc(ret);
 768
 769         if (parent) {
 770                 closure_wait(&w->wait, parent);
 771                 journal_try_write(c);
 772         } else if (!w->dirty) {
 773                 w->dirty = true;
 774                 schedule_delayed_work(&c->journal.work,
 775                                       msecs_to_jiffies(c->journal_delay_ms));
 776                 spin_unlock(&c->journal.lock);
 777         } else {
 778                 spin_unlock(&c->journal.lock);
 779         }
 780
 781
 782         return ret;
 783 }
 784
 785 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 786 {
 787         struct keylist keys;
 788         atomic_t *ref;
 789
 790         bch_keylist_init(&keys);
 791
 792         ref = bch_journal(c, &keys, cl);
 793         if (ref)
 794                 atomic_dec_bug(ref);
 795 }
 796
 797 void bch_journal_free(struct cache_set *c)
 798 {
 799         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 800         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 801         free_fifo(&c->journal.pin);
 802 }
 803
 804 int bch_journal_alloc(struct cache_set *c)
 805 {
 806         struct journal *j = &c->journal;
 807
 808         spin_lock_init(&j->lock);
 809         INIT_DELAYED_WORK(&j->work, journal_write_work);
 810
 811         c->journal_delay_ms = 100;
 812
 813         j->w[0].c = c;
 814         j->w[1].c = c;
 815
 816         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 817             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
 818             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
 819                 return -ENOMEM;
 820
 821         return 0;
 822 }