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