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