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                                 /*
 241                                  * When journal_reclaim() goes to allocate for
 242                                  * the first time, it'll use the bucket after
 243                                  * ja->cur_idx
 244                                  */
 245                                 ja->cur_idx = i;
 246                                 ja->last_idx = ja->discard_idx = (i + 1) %
 247                                         ca->sb.njournal_buckets;
 248
 249                         }
 250         }
 251
 252         if (!list_empty(list))
 253                 c->journal.seq = list_entry(list->prev,
 254                                             struct journal_replay,
 255                                             list)->j.seq;
 256
 257         return 0;
 258 #undef read_bucket
 259 }
 260
 261 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 262 {
 263         atomic_t p = { 0 };
 264         struct bkey *k;
 265         struct journal_replay *i;
 266         struct journal *j = &c->journal;
 267         uint64_t last = j->seq;
 268
 269         /*
 270          * journal.pin should never fill up - we never write a journal
 271          * entry when it would fill up. But if for some reason it does, we
 272          * iterate over the list in reverse order so that we can just skip that
 273          * refcount instead of bugging.
 274          */
 275
 276         list_for_each_entry_reverse(i, list, list) {
 277                 BUG_ON(last < i->j.seq);
 278                 i->pin = NULL;
 279
 280                 while (last-- != i->j.seq)
 281                         if (fifo_free(&j->pin) > 1) {
 282                                 fifo_push_front(&j->pin, p);
 283                                 atomic_set(&fifo_front(&j->pin), 0);
 284                         }
 285
 286                 if (fifo_free(&j->pin) > 1) {
 287                         fifo_push_front(&j->pin, p);
 288                         i->pin = &fifo_front(&j->pin);
 289                         atomic_set(i->pin, 1);
 290                 }
 291
 292                 for (k = i->j.start;
 293                      k < bset_bkey_last(&i->j);
 294                      k = bkey_next(k)) {
 295                         unsigned j;
 296
 297                         for (j = 0; j < KEY_PTRS(k); j++)
 298                                 if (ptr_available(c, k, j))
 299                                         atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 300
 301                         bch_initial_mark_key(c, 0, k);
 302                 }
 303         }
 304 }
 305
 306 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 307 {
 308         int ret = 0, keys = 0, entries = 0;
 309         struct bkey *k;
 310         struct journal_replay *i =
 311                 list_entry(list->prev, struct journal_replay, list);
 312
 313         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 314         struct keylist keylist;
 315
 316         list_for_each_entry(i, list, list) {
 317                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 318
 319                 cache_set_err_on(n != i->j.seq, s,
 320 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 321                                  n, i->j.seq - 1, start, end);
 322
 323                 for (k = i->j.start;
 324                      k < bset_bkey_last(&i->j);
 325                      k = bkey_next(k)) {
 326                         trace_bcache_journal_replay_key(k);
 327
 328                         bch_keylist_init_single(&keylist, k);
 329
 330                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 331                         if (ret)
 332                                 goto err;
 333
 334                         BUG_ON(!bch_keylist_empty(&keylist));
 335                         keys++;
 336
 337                         cond_resched();
 338                 }
 339
 340                 if (i->pin)
 341                         atomic_dec(i->pin);
 342                 n = i->j.seq + 1;
 343                 entries++;
 344         }
 345
 346         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 347                 keys, entries, end);
 348 err:
 349         while (!list_empty(list)) {
 350                 i = list_first_entry(list, struct journal_replay, list);
 351                 list_del(&i->list);
 352                 kfree(i);
 353         }
 354
 355         return ret;
 356 }
 357
 358 /* Journalling */
 359
 360 static void btree_flush_write(struct cache_set *c)
 361 {
 362         /*
 363          * Try to find the btree node with that references the oldest journal
 364          * entry, best is our current candidate and is locked if non NULL:
 365          */
 366         struct btree *b, *best;
 367         unsigned i;
 368 retry:
 369         best = NULL;
 370
 371         for_each_cached_btree(b, c, i)
 372                 if (btree_current_write(b)->journal) {
 373                         if (!best)
 374                                 best = b;
 375                         else if (journal_pin_cmp(c,
 376                                         btree_current_write(best)->journal,
 377                                         btree_current_write(b)->journal)) {
 378                                 best = b;
 379                         }
 380                 }
 381
 382         b = best;
 383         if (b) {
 384                 mutex_lock(&b->write_lock);
 385                 if (!btree_current_write(b)->journal) {
 386                         mutex_unlock(&b->write_lock);
 387                         /* We raced */
 388                         goto retry;
 389                 }
 390
 391                 __bch_btree_node_write(b, NULL);
 392                 mutex_unlock(&b->write_lock);
 393         }
 394 }
 395
 396 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 397
 398 static void journal_discard_endio(struct bio *bio, int error)
 399 {
 400         struct journal_device *ja =
 401                 container_of(bio, struct journal_device, discard_bio);
 402         struct cache *ca = container_of(ja, struct cache, journal);
 403
 404         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 405
 406         closure_wake_up(&ca->set->journal.wait);
 407         closure_put(&ca->set->cl);
 408 }
 409
 410 static void journal_discard_work(struct work_struct *work)
 411 {
 412         struct journal_device *ja =
 413                 container_of(work, struct journal_device, discard_work);
 414
 415         submit_bio(0, &ja->discard_bio);
 416 }
 417
 418 static void do_journal_discard(struct cache *ca)
 419 {
 420         struct journal_device *ja = &ca->journal;
 421         struct bio *bio = &ja->discard_bio;
 422
 423         if (!ca->discard) {
 424                 ja->discard_idx = ja->last_idx;
 425                 return;
 426         }
 427
 428         switch (atomic_read(&ja->discard_in_flight)) {
 429         case DISCARD_IN_FLIGHT:
 430                 return;
 431
 432         case DISCARD_DONE:
 433                 ja->discard_idx = (ja->discard_idx + 1) %
 434                         ca->sb.njournal_buckets;
 435
 436                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 437                 /* fallthrough */
 438
 439         case DISCARD_READY:
 440                 if (ja->discard_idx == ja->last_idx)
 441                         return;
 442
 443                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 444
 445                 bio_init(bio);
 446                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 447                                                 ca->sb.d[ja->discard_idx]);
 448                 bio->bi_bdev            = ca->bdev;
 449                 bio->bi_rw              = REQ_WRITE|REQ_DISCARD;
 450                 bio->bi_max_vecs        = 1;
 451                 bio->bi_io_vec          = bio->bi_inline_vecs;
 452                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 453                 bio->bi_end_io          = journal_discard_endio;
 454
 455                 closure_get(&ca->set->cl);
 456                 INIT_WORK(&ja->discard_work, journal_discard_work);
 457                 schedule_work(&ja->discard_work);
 458         }
 459 }
 460
 461 static void journal_reclaim(struct cache_set *c)
 462 {
 463         struct bkey *k = &c->journal.key;
 464         struct cache *ca;
 465         uint64_t last_seq;
 466         unsigned iter, n = 0;
 467         atomic_t p;
 468
 469         while (!atomic_read(&fifo_front(&c->journal.pin)))
 470                 fifo_pop(&c->journal.pin, p);
 471
 472         last_seq = last_seq(&c->journal);
 473
 474         /* Update last_idx */
 475
 476         for_each_cache(ca, c, iter) {
 477                 struct journal_device *ja = &ca->journal;
 478
 479                 while (ja->last_idx != ja->cur_idx &&
 480                        ja->seq[ja->last_idx] < last_seq)
 481                         ja->last_idx = (ja->last_idx + 1) %
 482                                 ca->sb.njournal_buckets;
 483         }
 484
 485         for_each_cache(ca, c, iter)
 486                 do_journal_discard(ca);
 487
 488         if (c->journal.blocks_free)
 489                 goto out;
 490
 491         /*
 492          * Allocate:
 493          * XXX: Sort by free journal space
 494          */
 495
 496         for_each_cache(ca, c, iter) {
 497                 struct journal_device *ja = &ca->journal;
 498                 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 499
 500                 /* No space available on this device */
 501                 if (next == ja->discard_idx)
 502                         continue;
 503
 504                 ja->cur_idx = next;
 505                 k->ptr[n++] = PTR(0,
 506                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 507                                   ca->sb.nr_this_dev);
 508         }
 509
 510         bkey_init(k);
 511         SET_KEY_PTRS(k, n);
 512
 513         if (n)
 514                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 515 out:
 516         if (!journal_full(&c->journal))
 517                 __closure_wake_up(&c->journal.wait);
 518 }
 519
 520 void bch_journal_next(struct journal *j)
 521 {
 522         atomic_t p = { 1 };
 523
 524         j->cur = (j->cur == j->w)
 525                 ? &j->w[1]
 526                 : &j->w[0];
 527
 528         /*
 529          * The fifo_push() needs to happen at the same time as j->seq is
 530          * incremented for last_seq() to be calculated correctly
 531          */
 532         BUG_ON(!fifo_push(&j->pin, p));
 533         atomic_set(&fifo_back(&j->pin), 1);
 534
 535         j->cur->data->seq       = ++j->seq;
 536         j->cur->dirty           = false;
 537         j->cur->need_write      = false;
 538         j->cur->data->keys      = 0;
 539
 540         if (fifo_full(&j->pin))
 541                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 542 }
 543
 544 static void journal_write_endio(struct bio *bio, int error)
 545 {
 546         struct journal_write *w = bio->bi_private;
 547
 548         cache_set_err_on(error, w->c, "journal io error");
 549         closure_put(&w->c->journal.io);
 550 }
 551
 552 static void journal_write(struct closure *);
 553
 554 static void journal_write_done(struct closure *cl)
 555 {
 556         struct journal *j = container_of(cl, struct journal, io);
 557         struct journal_write *w = (j->cur == j->w)
 558                 ? &j->w[1]
 559                 : &j->w[0];
 560
 561         __closure_wake_up(&w->wait);
 562         continue_at_nobarrier(cl, journal_write, system_wq);
 563 }
 564
 565 static void journal_write_unlock(struct closure *cl)
 566 {
 567         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 568
 569         c->journal.io_in_flight = 0;
 570         spin_unlock(&c->journal.lock);
 571 }
 572
 573 static void journal_write_unlocked(struct closure *cl)
 574         __releases(c->journal.lock)
 575 {
 576         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 577         struct cache *ca;
 578         struct journal_write *w = c->journal.cur;
 579         struct bkey *k = &c->journal.key;
 580         unsigned i, sectors = set_blocks(w->data, block_bytes(c)) *
 581                 c->sb.block_size;
 582
 583         struct bio *bio;
 584         struct bio_list list;
 585         bio_list_init(&list);
 586
 587         if (!w->need_write) {
 588                 closure_return_with_destructor(cl, journal_write_unlock);
 589         } else if (journal_full(&c->journal)) {
 590                 journal_reclaim(c);
 591                 spin_unlock(&c->journal.lock);
 592
 593                 btree_flush_write(c);
 594                 continue_at(cl, journal_write, system_wq);
 595         }
 596
 597         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 598
 599         w->data->btree_level = c->root->level;
 600
 601         bkey_copy(&w->data->btree_root, &c->root->key);
 602         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 603
 604         for_each_cache(ca, c, i)
 605                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 606
 607         w->data->magic          = jset_magic(&c->sb);
 608         w->data->version        = BCACHE_JSET_VERSION;
 609         w->data->last_seq       = last_seq(&c->journal);
 610         w->data->csum           = csum_set(w->data);
 611
 612         for (i = 0; i < KEY_PTRS(k); i++) {
 613                 ca = PTR_CACHE(c, k, i);
 614                 bio = &ca->journal.bio;
 615
 616                 atomic_long_add(sectors, &ca->meta_sectors_written);
 617
 618                 bio_reset(bio);
 619                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 620                 bio->bi_bdev    = ca->bdev;
 621                 bio->bi_rw      = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA;
 622                 bio->bi_iter.bi_size = sectors << 9;
 623
 624                 bio->bi_end_io  = journal_write_endio;
 625                 bio->bi_private = w;
 626                 bch_bio_map(bio, w->data);
 627
 628                 trace_bcache_journal_write(bio);
 629                 bio_list_add(&list, bio);
 630
 631                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 632
 633                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 634         }
 635
 636         atomic_dec_bug(&fifo_back(&c->journal.pin));
 637         bch_journal_next(&c->journal);
 638         journal_reclaim(c);
 639
 640         spin_unlock(&c->journal.lock);
 641
 642         while ((bio = bio_list_pop(&list)))
 643                 closure_bio_submit(bio, cl, c->cache[0]);
 644
 645         continue_at(cl, journal_write_done, NULL);
 646 }
 647
 648 static void journal_write(struct closure *cl)
 649 {
 650         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 651
 652         spin_lock(&c->journal.lock);
 653         journal_write_unlocked(cl);
 654 }
 655
 656 static void journal_try_write(struct cache_set *c)
 657         __releases(c->journal.lock)
 658 {
 659         struct closure *cl = &c->journal.io;
 660         struct journal_write *w = c->journal.cur;
 661
 662         w->need_write = true;
 663
 664         if (!c->journal.io_in_flight) {
 665                 c->journal.io_in_flight = 1;
 666                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 667         } else {
 668                 spin_unlock(&c->journal.lock);
 669         }
 670 }
 671
 672 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 673                                                     unsigned nkeys)
 674 {
 675         size_t sectors;
 676         struct closure cl;
 677         bool wait = false;
 678
 679         closure_init_stack(&cl);
 680
 681         spin_lock(&c->journal.lock);
 682
 683         while (1) {
 684                 struct journal_write *w = c->journal.cur;
 685
 686                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 687                                        block_bytes(c)) * c->sb.block_size;
 688
 689                 if (sectors <= min_t(size_t,
 690                                      c->journal.blocks_free * c->sb.block_size,
 691                                      PAGE_SECTORS << JSET_BITS))
 692                         return w;
 693
 694                 if (wait)
 695                         closure_wait(&c->journal.wait, &cl);
 696
 697                 if (!journal_full(&c->journal)) {
 698                         if (wait)
 699                                 trace_bcache_journal_entry_full(c);
 700
 701                         /*
 702                          * XXX: If we were inserting so many keys that they
 703                          * won't fit in an _empty_ journal write, we'll
 704                          * deadlock. For now, handle this in
 705                          * bch_keylist_realloc() - but something to think about.
 706                          */
 707                         BUG_ON(!w->data->keys);
 708
 709                         journal_try_write(c); /* unlocks */
 710                 } else {
 711                         if (wait)
 712                                 trace_bcache_journal_full(c);
 713
 714                         journal_reclaim(c);
 715                         spin_unlock(&c->journal.lock);
 716
 717                         btree_flush_write(c);
 718                 }
 719
 720                 closure_sync(&cl);
 721                 spin_lock(&c->journal.lock);
 722                 wait = true;
 723         }
 724 }
 725
 726 static void journal_write_work(struct work_struct *work)
 727 {
 728         struct cache_set *c = container_of(to_delayed_work(work),
 729                                            struct cache_set,
 730                                            journal.work);
 731         spin_lock(&c->journal.lock);
 732         if (c->journal.cur->dirty)
 733                 journal_try_write(c);
 734         else
 735                 spin_unlock(&c->journal.lock);
 736 }
 737
 738 /*
 739  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 740  * pass bch_journal() a list of keys to be journalled, and then
 741  * bch_journal() hands those same keys off to btree_insert_async()
 742  */
 743
 744 atomic_t *bch_journal(struct cache_set *c,
 745                       struct keylist *keys,
 746                       struct closure *parent)
 747 {
 748         struct journal_write *w;
 749         atomic_t *ret;
 750
 751         if (!CACHE_SYNC(&c->sb))
 752                 return NULL;
 753
 754         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 755
 756         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 757         w->data->keys += bch_keylist_nkeys(keys);
 758
 759         ret = &fifo_back(&c->journal.pin);
 760         atomic_inc(ret);
 761
 762         if (parent) {
 763                 closure_wait(&w->wait, parent);
 764                 journal_try_write(c);
 765         } else if (!w->dirty) {
 766                 w->dirty = true;
 767                 schedule_delayed_work(&c->journal.work,
 768                                       msecs_to_jiffies(c->journal_delay_ms));
 769                 spin_unlock(&c->journal.lock);
 770         } else {
 771                 spin_unlock(&c->journal.lock);
 772         }
 773
 774
 775         return ret;
 776 }
 777
 778 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 779 {
 780         struct keylist keys;
 781         atomic_t *ref;
 782
 783         bch_keylist_init(&keys);
 784
 785         ref = bch_journal(c, &keys, cl);
 786         if (ref)
 787                 atomic_dec_bug(ref);
 788 }
 789
 790 void bch_journal_free(struct cache_set *c)
 791 {
 792         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 793         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 794         free_fifo(&c->journal.pin);
 795 }
 796
 797 int bch_journal_alloc(struct cache_set *c)
 798 {
 799         struct journal *j = &c->journal;
 800
 801         spin_lock_init(&j->lock);
 802         INIT_DELAYED_WORK(&j->work, journal_write_work);
 803
 804         c->journal_delay_ms = 100;
 805
 806         j->w[0].c = c;
 807         j->w[1].c = c;
 808
 809         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 810             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
 811             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
 812                 return -ENOMEM;
 813
 814         return 0;
 815 }