drivers/md/bcache/journal.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * bcache journalling code, for btree insertions
   4  *
   5  * Copyright 2012 Google, Inc.
   6  */
   7
   8 #include "bcache.h"
   9 #include "btree.h"
  10 #include "debug.h"
  11 #include "extents.h"
  12
  13 #include <trace/events/bcache.h>
  14
  15 /*
  16  * Journal replay/recovery:
  17  *
  18  * This code is all driven from run_cache_set(); we first read the journal
  19  * entries, do some other stuff, then we mark all the keys in the journal
  20  * entries (same as garbage collection would), then we replay them - reinserting
  21  * them into the cache in precisely the same order as they appear in the
  22  * journal.
  23  *
  24  * We only journal keys that go in leaf nodes, which simplifies things quite a
  25  * bit.
  26  */
  27
  28 static void journal_read_endio(struct bio *bio)
  29 {
  30         struct closure *cl = bio->bi_private;
  31         closure_put(cl);
  32 }
  33
  34 static int journal_read_bucket(struct cache *ca, struct list_head *list,
  35                                unsigned bucket_index)
  36 {
  37         struct journal_device *ja = &ca->journal;
  38         struct bio *bio = &ja->bio;
  39
  40         struct journal_replay *i;
  41         struct jset *j, *data = ca->set->journal.w[0].data;
  42         struct closure cl;
  43         unsigned len, left, offset = 0;
  44         int ret = 0;
  45         sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
  46
  47         closure_init_stack(&cl);
  48
  49         pr_debug("reading %u", bucket_index);
  50
  51         while (offset < ca->sb.bucket_size) {
  52 reread:         left = ca->sb.bucket_size - offset;
  53                 len = min_t(unsigned, left, PAGE_SECTORS << JSET_BITS);
  54
  55                 bio_reset(bio);
  56                 bio->bi_iter.bi_sector  = bucket + offset;
  57                 bio_set_dev(bio, ca->bdev);
  58                 bio->bi_iter.bi_size    = len << 9;
  59
  60                 bio->bi_end_io  = journal_read_endio;
  61                 bio->bi_private = &cl;
  62                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
  63                 bch_bio_map(bio, data);
  64
  65                 closure_bio_submit(bio, &cl);
  66                 closure_sync(&cl);
  67
  68                 /* This function could be simpler now since we no longer write
  69                  * journal entries that overlap bucket boundaries; this means
  70                  * the start of a bucket will always have a valid journal entry
  71                  * if it has any journal entries at all.
  72                  */
  73
  74                 j = data;
  75                 while (len) {
  76                         struct list_head *where;
  77                         size_t blocks, bytes = set_bytes(j);
  78
  79                         if (j->magic != jset_magic(&ca->sb)) {
  80                                 pr_debug("%u: bad magic", bucket_index);
  81                                 return ret;
  82                         }
  83
  84                         if (bytes > left << 9 ||
  85                             bytes > PAGE_SIZE << JSET_BITS) {
  86                                 pr_info("%u: too big, %zu bytes, offset %u",
  87                                         bucket_index, bytes, offset);
  88                                 return ret;
  89                         }
  90
  91                         if (bytes > len << 9)
  92                                 goto reread;
  93
  94                         if (j->csum != csum_set(j)) {
  95                                 pr_info("%u: bad csum, %zu bytes, offset %u",
  96                                         bucket_index, bytes, offset);
  97                                 return ret;
  98                         }
  99
 100                         blocks = set_blocks(j, block_bytes(ca->set));
 101
 102                         while (!list_empty(list)) {
 103                                 i = list_first_entry(list,
 104                                         struct journal_replay, list);
 105                                 if (i->j.seq >= j->last_seq)
 106                                         break;
 107                                 list_del(&i->list);
 108                                 kfree(i);
 109                         }
 110
 111                         list_for_each_entry_reverse(i, list, list) {
 112                                 if (j->seq == i->j.seq)
 113                                         goto next_set;
 114
 115                                 if (j->seq < i->j.last_seq)
 116                                         goto next_set;
 117
 118                                 if (j->seq > i->j.seq) {
 119                                         where = &i->list;
 120                                         goto add;
 121                                 }
 122                         }
 123
 124                         where = list;
 125 add:
 126                         i = kmalloc(offsetof(struct journal_replay, j) +
 127                                     bytes, GFP_KERNEL);
 128                         if (!i)
 129                                 return -ENOMEM;
 130                         memcpy(&i->j, j, bytes);
 131                         list_add(&i->list, where);
 132                         ret = 1;
 133
 134                         ja->seq[bucket_index] = j->seq;
 135 next_set:
 136                         offset  += blocks * ca->sb.block_size;
 137                         len     -= blocks * ca->sb.block_size;
 138                         j = ((void *) j) + blocks * block_bytes(ca);
 139                 }
 140         }
 141
 142         return ret;
 143 }
 144
 145 int bch_journal_read(struct cache_set *c, struct list_head *list)
 146 {
 147 #define read_bucket(b)                                                  \
 148         ({                                                              \
 149                 int ret = journal_read_bucket(ca, list, b);             \
 150                 __set_bit(b, bitmap);                                   \
 151                 if (ret < 0)                                            \
 152                         return ret;                                     \
 153                 ret;                                                    \
 154         })
 155
 156         struct cache *ca;
 157         unsigned iter;
 158
 159         for_each_cache(ca, c, iter) {
 160                 struct journal_device *ja = &ca->journal;
 161                 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
 162                 unsigned i, l, r, m;
 163                 uint64_t seq;
 164
 165                 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
 166                 pr_debug("%u journal buckets", ca->sb.njournal_buckets);
 167
 168                 /*
 169                  * Read journal buckets ordered by golden ratio hash to quickly
 170                  * find a sequence of buckets with valid journal entries
 171                  */
 172                 for (i = 0; i < ca->sb.njournal_buckets; i++) {
 173                         /*
 174                          * We must try the index l with ZERO first for
 175                          * correctness due to the scenario that the journal
 176                          * bucket is circular buffer which might have wrapped
 177                          */
 178                         l = (i * 2654435769U) % ca->sb.njournal_buckets;
 179
 180                         if (test_bit(l, bitmap))
 181                                 break;
 182
 183                         if (read_bucket(l))
 184                                 goto bsearch;
 185                 }
 186
 187                 /*
 188                  * If that fails, check all the buckets we haven't checked
 189                  * already
 190                  */
 191                 pr_debug("falling back to linear search");
 192
 193                 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
 194                      l < ca->sb.njournal_buckets;
 195                      l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1))
 196                         if (read_bucket(l))
 197                                 goto bsearch;
 198
 199                 /* no journal entries on this device? */
 200                 if (l == ca->sb.njournal_buckets)
 201                         continue;
 202 bsearch:
 203                 BUG_ON(list_empty(list));
 204
 205                 /* Binary search */
 206                 m = l;
 207                 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 208                 pr_debug("starting binary search, l %u r %u", l, r);
 209
 210                 while (l + 1 < r) {
 211                         seq = list_entry(list->prev, struct journal_replay,
 212                                          list)->j.seq;
 213
 214                         m = (l + r) >> 1;
 215                         read_bucket(m);
 216
 217                         if (seq != list_entry(list->prev, struct journal_replay,
 218                                               list)->j.seq)
 219                                 l = m;
 220                         else
 221                                 r = m;
 222                 }
 223
 224                 /*
 225                  * Read buckets in reverse order until we stop finding more
 226                  * journal entries
 227                  */
 228                 pr_debug("finishing up: m %u njournal_buckets %u",
 229                          m, ca->sb.njournal_buckets);
 230                 l = m;
 231
 232                 while (1) {
 233                         if (!l--)
 234                                 l = ca->sb.njournal_buckets - 1;
 235
 236                         if (l == m)
 237                                 break;
 238
 239                         if (test_bit(l, bitmap))
 240                                 continue;
 241
 242                         if (!read_bucket(l))
 243                                 break;
 244                 }
 245
 246                 seq = 0;
 247
 248                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 249                         if (ja->seq[i] > seq) {
 250                                 seq = ja->seq[i];
 251                                 /*
 252                                  * When journal_reclaim() goes to allocate for
 253                                  * the first time, it'll use the bucket after
 254                                  * ja->cur_idx
 255                                  */
 256                                 ja->cur_idx = i;
 257                                 ja->last_idx = ja->discard_idx = (i + 1) %
 258                                         ca->sb.njournal_buckets;
 259
 260                         }
 261         }
 262
 263         if (!list_empty(list))
 264                 c->journal.seq = list_entry(list->prev,
 265                                             struct journal_replay,
 266                                             list)->j.seq;
 267
 268         return 0;
 269 #undef read_bucket
 270 }
 271
 272 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 273 {
 274         atomic_t p = { 0 };
 275         struct bkey *k;
 276         struct journal_replay *i;
 277         struct journal *j = &c->journal;
 278         uint64_t last = j->seq;
 279
 280         /*
 281          * journal.pin should never fill up - we never write a journal
 282          * entry when it would fill up. But if for some reason it does, we
 283          * iterate over the list in reverse order so that we can just skip that
 284          * refcount instead of bugging.
 285          */
 286
 287         list_for_each_entry_reverse(i, list, list) {
 288                 BUG_ON(last < i->j.seq);
 289                 i->pin = NULL;
 290
 291                 while (last-- != i->j.seq)
 292                         if (fifo_free(&j->pin) > 1) {
 293                                 fifo_push_front(&j->pin, p);
 294                                 atomic_set(&fifo_front(&j->pin), 0);
 295                         }
 296
 297                 if (fifo_free(&j->pin) > 1) {
 298                         fifo_push_front(&j->pin, p);
 299                         i->pin = &fifo_front(&j->pin);
 300                         atomic_set(i->pin, 1);
 301                 }
 302
 303                 for (k = i->j.start;
 304                      k < bset_bkey_last(&i->j);
 305                      k = bkey_next(k))
 306                         if (!__bch_extent_invalid(c, k)) {
 307                                 unsigned j;
 308
 309                                 for (j = 0; j < KEY_PTRS(k); j++)
 310                                         if (ptr_available(c, k, j))
 311                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 312
 313                                 bch_initial_mark_key(c, 0, k);
 314                         }
 315         }
 316 }
 317
 318 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 319 {
 320         int ret = 0, keys = 0, entries = 0;
 321         struct bkey *k;
 322         struct journal_replay *i =
 323                 list_entry(list->prev, struct journal_replay, list);
 324
 325         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 326         struct keylist keylist;
 327
 328         list_for_each_entry(i, list, list) {
 329                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 330
 331                 cache_set_err_on(n != i->j.seq, s,
 332 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 333                                  n, i->j.seq - 1, start, end);
 334
 335                 for (k = i->j.start;
 336                      k < bset_bkey_last(&i->j);
 337                      k = bkey_next(k)) {
 338                         trace_bcache_journal_replay_key(k);
 339
 340                         bch_keylist_init_single(&keylist, k);
 341
 342                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 343                         if (ret)
 344                                 goto err;
 345
 346                         BUG_ON(!bch_keylist_empty(&keylist));
 347                         keys++;
 348
 349                         cond_resched();
 350                 }
 351
 352                 if (i->pin)
 353                         atomic_dec(i->pin);
 354                 n = i->j.seq + 1;
 355                 entries++;
 356         }
 357
 358         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 359                 keys, entries, end);
 360 err:
 361         while (!list_empty(list)) {
 362                 i = list_first_entry(list, struct journal_replay, list);
 363                 list_del(&i->list);
 364                 kfree(i);
 365         }
 366
 367         return ret;
 368 }
 369
 370 /* Journalling */
 371
 372 static void btree_flush_write(struct cache_set *c)
 373 {
 374         /*
 375          * Try to find the btree node with that references the oldest journal
 376          * entry, best is our current candidate and is locked if non NULL:
 377          */
 378         struct btree *b, *best;
 379         unsigned i;
 380 retry:
 381         best = NULL;
 382
 383         for_each_cached_btree(b, c, i)
 384                 if (btree_current_write(b)->journal) {
 385                         if (!best)
 386                                 best = b;
 387                         else if (journal_pin_cmp(c,
 388                                         btree_current_write(best)->journal,
 389                                         btree_current_write(b)->journal)) {
 390                                 best = b;
 391                         }
 392                 }
 393
 394         b = best;
 395         if (b) {
 396                 mutex_lock(&b->write_lock);
 397                 if (!btree_current_write(b)->journal) {
 398                         mutex_unlock(&b->write_lock);
 399                         /* We raced */
 400                         goto retry;
 401                 }
 402
 403                 __bch_btree_node_write(b, NULL);
 404                 mutex_unlock(&b->write_lock);
 405         }
 406 }
 407
 408 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 409
 410 static void journal_discard_endio(struct bio *bio)
 411 {
 412         struct journal_device *ja =
 413                 container_of(bio, struct journal_device, discard_bio);
 414         struct cache *ca = container_of(ja, struct cache, journal);
 415
 416         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 417
 418         closure_wake_up(&ca->set->journal.wait);
 419         closure_put(&ca->set->cl);
 420 }
 421
 422 static void journal_discard_work(struct work_struct *work)
 423 {
 424         struct journal_device *ja =
 425                 container_of(work, struct journal_device, discard_work);
 426
 427         submit_bio(&ja->discard_bio);
 428 }
 429
 430 static void do_journal_discard(struct cache *ca)
 431 {
 432         struct journal_device *ja = &ca->journal;
 433         struct bio *bio = &ja->discard_bio;
 434
 435         if (!ca->discard) {
 436                 ja->discard_idx = ja->last_idx;
 437                 return;
 438         }
 439
 440         switch (atomic_read(&ja->discard_in_flight)) {
 441         case DISCARD_IN_FLIGHT:
 442                 return;
 443
 444         case DISCARD_DONE:
 445                 ja->discard_idx = (ja->discard_idx + 1) %
 446                         ca->sb.njournal_buckets;
 447
 448                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 449                 /* fallthrough */
 450
 451         case DISCARD_READY:
 452                 if (ja->discard_idx == ja->last_idx)
 453                         return;
 454
 455                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 456
 457                 bio_init(bio, bio->bi_inline_vecs, 1);
 458                 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
 459                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 460                                                 ca->sb.d[ja->discard_idx]);
 461                 bio_set_dev(bio, ca->bdev);
 462                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 463                 bio->bi_end_io          = journal_discard_endio;
 464
 465                 closure_get(&ca->set->cl);
 466                 INIT_WORK(&ja->discard_work, journal_discard_work);
 467                 schedule_work(&ja->discard_work);
 468         }
 469 }
 470
 471 static void journal_reclaim(struct cache_set *c)
 472 {
 473         struct bkey *k = &c->journal.key;
 474         struct cache *ca;
 475         uint64_t last_seq;
 476         unsigned iter, n = 0;
 477         atomic_t p;
 478
 479         while (!atomic_read(&fifo_front(&c->journal.pin)))
 480                 fifo_pop(&c->journal.pin, p);
 481
 482         last_seq = last_seq(&c->journal);
 483
 484         /* Update last_idx */
 485
 486         for_each_cache(ca, c, iter) {
 487                 struct journal_device *ja = &ca->journal;
 488
 489                 while (ja->last_idx != ja->cur_idx &&
 490                        ja->seq[ja->last_idx] < last_seq)
 491                         ja->last_idx = (ja->last_idx + 1) %
 492                                 ca->sb.njournal_buckets;
 493         }
 494
 495         for_each_cache(ca, c, iter)
 496                 do_journal_discard(ca);
 497
 498         if (c->journal.blocks_free)
 499                 goto out;
 500
 501         /*
 502          * Allocate:
 503          * XXX: Sort by free journal space
 504          */
 505
 506         for_each_cache(ca, c, iter) {
 507                 struct journal_device *ja = &ca->journal;
 508                 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 509
 510                 /* No space available on this device */
 511                 if (next == ja->discard_idx)
 512                         continue;
 513
 514                 ja->cur_idx = next;
 515                 k->ptr[n++] = MAKE_PTR(0,
 516                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 517                                   ca->sb.nr_this_dev);
 518         }
 519
 520         bkey_init(k);
 521         SET_KEY_PTRS(k, n);
 522
 523         if (n)
 524                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 525 out:
 526         if (!journal_full(&c->journal))
 527                 __closure_wake_up(&c->journal.wait);
 528 }
 529
 530 void bch_journal_next(struct journal *j)
 531 {
 532         atomic_t p = { 1 };
 533
 534         j->cur = (j->cur == j->w)
 535                 ? &j->w[1]
 536                 : &j->w[0];
 537
 538         /*
 539          * The fifo_push() needs to happen at the same time as j->seq is
 540          * incremented for last_seq() to be calculated correctly
 541          */
 542         BUG_ON(!fifo_push(&j->pin, p));
 543         atomic_set(&fifo_back(&j->pin), 1);
 544
 545         j->cur->data->seq       = ++j->seq;
 546         j->cur->dirty           = false;
 547         j->cur->need_write      = false;
 548         j->cur->data->keys      = 0;
 549
 550         if (fifo_full(&j->pin))
 551                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 552 }
 553
 554 static void journal_write_endio(struct bio *bio)
 555 {
 556         struct journal_write *w = bio->bi_private;
 557
 558         cache_set_err_on(bio->bi_status, w->c, "journal io error");
 559         closure_put(&w->c->journal.io);
 560 }
 561
 562 static void journal_write(struct closure *);
 563
 564 static void journal_write_done(struct closure *cl)
 565 {
 566         struct journal *j = container_of(cl, struct journal, io);
 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         continue_at_nobarrier(cl, journal_write, system_wq);
 573 }
 574
 575 static void journal_write_unlock(struct closure *cl)
 576 {
 577         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 578
 579         c->journal.io_in_flight = 0;
 580         spin_unlock(&c->journal.lock);
 581 }
 582
 583 static void journal_write_unlocked(struct closure *cl)
 584         __releases(c->journal.lock)
 585 {
 586         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 587         struct cache *ca;
 588         struct journal_write *w = c->journal.cur;
 589         struct bkey *k = &c->journal.key;
 590         unsigned i, sectors = set_blocks(w->data, block_bytes(c)) *
 591                 c->sb.block_size;
 592
 593         struct bio *bio;
 594         struct bio_list list;
 595         bio_list_init(&list);
 596
 597         if (!w->need_write) {
 598                 closure_return_with_destructor(cl, journal_write_unlock);
 599                 return;
 600         } else if (journal_full(&c->journal)) {
 601                 journal_reclaim(c);
 602                 spin_unlock(&c->journal.lock);
 603
 604                 btree_flush_write(c);
 605                 continue_at(cl, journal_write, system_wq);
 606                 return;
 607         }
 608
 609         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 610
 611         w->data->btree_level = c->root->level;
 612
 613         bkey_copy(&w->data->btree_root, &c->root->key);
 614         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 615
 616         for_each_cache(ca, c, i)
 617                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 618
 619         w->data->magic          = jset_magic(&c->sb);
 620         w->data->version        = BCACHE_JSET_VERSION;
 621         w->data->last_seq       = last_seq(&c->journal);
 622         w->data->csum           = csum_set(w->data);
 623
 624         for (i = 0; i < KEY_PTRS(k); i++) {
 625                 ca = PTR_CACHE(c, k, i);
 626                 bio = &ca->journal.bio;
 627
 628                 atomic_long_add(sectors, &ca->meta_sectors_written);
 629
 630                 bio_reset(bio);
 631                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 632                 bio_set_dev(bio, ca->bdev);
 633                 bio->bi_iter.bi_size = sectors << 9;
 634
 635                 bio->bi_end_io  = journal_write_endio;
 636                 bio->bi_private = w;
 637                 bio_set_op_attrs(bio, REQ_OP_WRITE,
 638                                  REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
 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);
 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);
 664
 665         spin_lock(&c->journal.lock);
 666         journal_write_unlocked(cl);
 667 }
 668
 669 static void journal_try_write(struct cache_set *c)
 670         __releases(c->journal.lock)
 671 {
 672         struct closure *cl = &c->journal.io;
 673         struct journal_write *w = c->journal.cur;
 674
 675         w->need_write = true;
 676
 677         if (!c->journal.io_in_flight) {
 678                 c->journal.io_in_flight = 1;
 679                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 680         } else {
 681                 spin_unlock(&c->journal.lock);
 682         }
 683 }
 684
 685 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 686                                                     unsigned nkeys)
 687 {
 688         size_t sectors;
 689         struct closure cl;
 690         bool wait = false;
 691
 692         closure_init_stack(&cl);
 693
 694         spin_lock(&c->journal.lock);
 695
 696         while (1) {
 697                 struct journal_write *w = c->journal.cur;
 698
 699                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 700                                        block_bytes(c)) * c->sb.block_size;
 701
 702                 if (sectors <= min_t(size_t,
 703                                      c->journal.blocks_free * c->sb.block_size,
 704                                      PAGE_SECTORS << JSET_BITS))
 705                         return w;
 706
 707                 if (wait)
 708                         closure_wait(&c->journal.wait, &cl);
 709
 710                 if (!journal_full(&c->journal)) {
 711                         if (wait)
 712                                 trace_bcache_journal_entry_full(c);
 713
 714                         /*
 715                          * XXX: If we were inserting so many keys that they
 716                          * won't fit in an _empty_ journal write, we'll
 717                          * deadlock. For now, handle this in
 718                          * bch_keylist_realloc() - but something to think about.
 719                          */
 720                         BUG_ON(!w->data->keys);
 721
 722                         journal_try_write(c); /* unlocks */
 723                 } else {
 724                         if (wait)
 725                                 trace_bcache_journal_full(c);
 726
 727                         journal_reclaim(c);
 728                         spin_unlock(&c->journal.lock);
 729
 730                         btree_flush_write(c);
 731                 }
 732
 733                 closure_sync(&cl);
 734                 spin_lock(&c->journal.lock);
 735                 wait = true;
 736         }
 737 }
 738
 739 static void journal_write_work(struct work_struct *work)
 740 {
 741         struct cache_set *c = container_of(to_delayed_work(work),
 742                                            struct cache_set,
 743                                            journal.work);
 744         spin_lock(&c->journal.lock);
 745         if (c->journal.cur->dirty)
 746                 journal_try_write(c);
 747         else
 748                 spin_unlock(&c->journal.lock);
 749 }
 750
 751 /*
 752  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 753  * pass bch_journal() a list of keys to be journalled, and then
 754  * bch_journal() hands those same keys off to btree_insert_async()
 755  */
 756
 757 atomic_t *bch_journal(struct cache_set *c,
 758                       struct keylist *keys,
 759                       struct closure *parent)
 760 {
 761         struct journal_write *w;
 762         atomic_t *ret;
 763
 764         if (!CACHE_SYNC(&c->sb))
 765                 return NULL;
 766
 767         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 768
 769         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 770         w->data->keys += bch_keylist_nkeys(keys);
 771
 772         ret = &fifo_back(&c->journal.pin);
 773         atomic_inc(ret);
 774
 775         if (parent) {
 776                 closure_wait(&w->wait, parent);
 777                 journal_try_write(c);
 778         } else if (!w->dirty) {
 779                 w->dirty = true;
 780                 schedule_delayed_work(&c->journal.work,
 781                                       msecs_to_jiffies(c->journal_delay_ms));
 782                 spin_unlock(&c->journal.lock);
 783         } else {
 784                 spin_unlock(&c->journal.lock);
 785         }
 786
 787
 788         return ret;
 789 }
 790
 791 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 792 {
 793         struct keylist keys;
 794         atomic_t *ref;
 795
 796         bch_keylist_init(&keys);
 797
 798         ref = bch_journal(c, &keys, cl);
 799         if (ref)
 800                 atomic_dec_bug(ref);
 801 }
 802
 803 void bch_journal_free(struct cache_set *c)
 804 {
 805         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 806         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 807         free_fifo(&c->journal.pin);
 808 }
 809
 810 int bch_journal_alloc(struct cache_set *c)
 811 {
 812         struct journal *j = &c->journal;
 813
 814         spin_lock_init(&j->lock);
 815         INIT_DELAYED_WORK(&j->work, journal_write_work);
 816
 817         c->journal_delay_ms = 100;
 818
 819         j->w[0].c = c;
 820         j->w[1].c = c;
 821
 822         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 823             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
 824             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
 825                 return -ENOMEM;
 826
 827         return 0;
 828 }