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