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 "extents.h"
  11
  12 #include <trace/events/bcache.h>
  13
  14 /*
  15  * Journal replay/recovery:
  16  *
  17  * This code is all driven from run_cache_set(); we first read the journal
  18  * entries, do some other stuff, then we mark all the keys in the journal
  19  * entries (same as garbage collection would), then we replay them - reinserting
  20  * them into the cache in precisely the same order as they appear in the
  21  * journal.
  22  *
  23  * We only journal keys that go in leaf nodes, which simplifies things quite a
  24  * bit.
  25  */
  26
  27 static void journal_read_endio(struct bio *bio)
  28 {
  29         struct closure *cl = bio->bi_private;
  30         closure_put(cl);
  31 }
  32
  33 static int journal_read_bucket(struct cache *ca, struct list_head *list,
  34                                unsigned bucket_index)
  35 {
  36         struct journal_device *ja = &ca->journal;
  37         struct bio *bio = &ja->bio;
  38
  39         struct journal_replay *i;
  40         struct jset *j, *data = ca->set->journal.w[0].data;
  41         struct closure cl;
  42         unsigned len, left, offset = 0;
  43         int ret = 0;
  44         sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
  45
  46         closure_init_stack(&cl);
  47
  48         pr_debug("reading %u", bucket_index);
  49
  50         while (offset < ca->sb.bucket_size) {
  51 reread:         left = ca->sb.bucket_size - offset;
  52                 len = min_t(unsigned, left, PAGE_SECTORS << JSET_BITS);
  53
  54                 bio_reset(bio);
  55                 bio->bi_iter.bi_sector  = bucket + offset;
  56                 bio->bi_bdev    = ca->bdev;
  57                 bio->bi_rw      = READ;
  58                 bio->bi_iter.bi_size    = len << 9;
  59
  60                 bio->bi_end_io  = journal_read_endio;
  61                 bio->bi_private = &cl;
  62                 bch_bio_map(bio, data);
  63
  64                 closure_bio_submit(bio, &cl);
  65                 closure_sync(&cl);
  66
  67                 /* This function could be simpler now since we no longer write
  68                  * journal entries that overlap bucket boundaries; this means
  69                  * the start of a bucket will always have a valid journal entry
  70                  * if it has any journal entries at all.
  71                  */
  72
  73                 j = data;
  74                 while (len) {
  75                         struct list_head *where;
  76                         size_t blocks, bytes = set_bytes(j);
  77
  78                         if (j->magic != jset_magic(&ca->sb)) {
  79                                 pr_debug("%u: bad magic", bucket_index);
  80                                 return ret;
  81                         }
  82
  83                         if (bytes > left << 9 ||
  84                             bytes > PAGE_SIZE << JSET_BITS) {
  85                                 pr_info("%u: too big, %zu bytes, offset %u",
  86                                         bucket_index, bytes, offset);
  87                                 return ret;
  88                         }
  89
  90                         if (bytes > len << 9)
  91                                 goto reread;
  92
  93                         if (j->csum != csum_set(j)) {
  94                                 pr_info("%u: bad csum, %zu bytes, offset %u",
  95                                         bucket_index, bytes, offset);
  96                                 return ret;
  97                         }
  98
  99                         blocks = set_blocks(j, block_bytes(ca->set));
 100
 101                         while (!list_empty(list)) {
 102                                 i = list_first_entry(list,
 103                                         struct journal_replay, list);
 104                                 if (i->j.seq >= j->last_seq)
 105                                         break;
 106                                 list_del(&i->list);
 107                                 kfree(i);
 108                         }
 109
 110                         list_for_each_entry_reverse(i, list, list) {
 111                                 if (j->seq == i->j.seq)
 112                                         goto next_set;
 113
 114                                 if (j->seq < i->j.last_seq)
 115                                         goto next_set;
 116
 117                                 if (j->seq > i->j.seq) {
 118                                         where = &i->list;
 119                                         goto add;
 120                                 }
 121                         }
 122
 123                         where = list;
 124 add:
 125                         i = kmalloc(offsetof(struct journal_replay, j) +
 126                                     bytes, GFP_KERNEL);
 127                         if (!i)
 128                                 return -ENOMEM;
 129                         memcpy(&i->j, j, bytes);
 130                         list_add(&i->list, where);
 131                         ret = 1;
 132
 133                         ja->seq[bucket_index] = j->seq;
 134 next_set:
 135                         offset  += blocks * ca->sb.block_size;
 136                         len     -= blocks * ca->sb.block_size;
 137                         j = ((void *) j) + blocks * block_bytes(ca);
 138                 }
 139         }
 140
 141         return ret;
 142 }
 143
 144 int bch_journal_read(struct cache_set *c, struct list_head *list)
 145 {
 146 #define read_bucket(b)                                                  \
 147         ({                                                              \
 148                 int ret = journal_read_bucket(ca, list, b);             \
 149                 __set_bit(b, bitmap);                                   \
 150                 if (ret < 0)                                            \
 151                         return ret;                                     \
 152                 ret;                                                    \
 153         })
 154
 155         struct cache *ca;
 156         unsigned iter;
 157
 158         for_each_cache(ca, c, iter) {
 159                 struct journal_device *ja = &ca->journal;
 160                 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
 161                 unsigned i, l, r, m;
 162                 uint64_t seq;
 163
 164                 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
 165                 pr_debug("%u journal buckets", ca->sb.njournal_buckets);
 166
 167                 /*
 168                  * Read journal buckets ordered by golden ratio hash to quickly
 169                  * find a sequence of buckets with valid journal entries
 170                  */
 171                 for (i = 0; i < ca->sb.njournal_buckets; i++) {
 172                         l = (i * 2654435769U) % ca->sb.njournal_buckets;
 173
 174                         if (test_bit(l, bitmap))
 175                                 break;
 176
 177                         if (read_bucket(l))
 178                                 goto bsearch;
 179                 }
 180
 181                 /*
 182                  * If that fails, check all the buckets we haven't checked
 183                  * already
 184                  */
 185                 pr_debug("falling back to linear search");
 186
 187                 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
 188                      l < ca->sb.njournal_buckets;
 189                      l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1))
 190                         if (read_bucket(l))
 191                                 goto bsearch;
 192
 193                 /* no journal entries on this device? */
 194                 if (l == ca->sb.njournal_buckets)
 195                         continue;
 196 bsearch:
 197                 BUG_ON(list_empty(list));
 198
 199                 /* Binary search */
 200                 m = l;
 201                 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 202                 pr_debug("starting binary search, l %u r %u", l, r);
 203
 204                 while (l + 1 < r) {
 205                         seq = list_entry(list->prev, struct journal_replay,
 206                                          list)->j.seq;
 207
 208                         m = (l + r) >> 1;
 209                         read_bucket(m);
 210
 211                         if (seq != list_entry(list->prev, struct journal_replay,
 212                                               list)->j.seq)
 213                                 l = m;
 214                         else
 215                                 r = m;
 216                 }
 217
 218                 /*
 219                  * Read buckets in reverse order until we stop finding more
 220                  * journal entries
 221                  */
 222                 pr_debug("finishing up: m %u njournal_buckets %u",
 223                          m, ca->sb.njournal_buckets);
 224                 l = m;
 225
 226                 while (1) {
 227                         if (!l--)
 228                                 l = ca->sb.njournal_buckets - 1;
 229
 230                         if (l == m)
 231                                 break;
 232
 233                         if (test_bit(l, bitmap))
 234                                 continue;
 235
 236                         if (!read_bucket(l))
 237                                 break;
 238                 }
 239
 240                 seq = 0;
 241
 242                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 243                         if (ja->seq[i] > seq) {
 244                                 seq = ja->seq[i];
 245                                 /*
 246                                  * When journal_reclaim() goes to allocate for
 247                                  * the first time, it'll use the bucket after
 248                                  * ja->cur_idx
 249                                  */
 250                                 ja->cur_idx = i;
 251                                 ja->last_idx = ja->discard_idx = (i + 1) %
 252                                         ca->sb.njournal_buckets;
 253
 254                         }
 255         }
 256
 257         if (!list_empty(list))
 258                 c->journal.seq = list_entry(list->prev,
 259                                             struct journal_replay,
 260                                             list)->j.seq;
 261
 262         return 0;
 263 #undef read_bucket
 264 }
 265
 266 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 267 {
 268         atomic_t p = { 0 };
 269         struct bkey *k;
 270         struct journal_replay *i;
 271         struct journal *j = &c->journal;
 272         uint64_t last = j->seq;
 273
 274         /*
 275          * journal.pin should never fill up - we never write a journal
 276          * entry when it would fill up. But if for some reason it does, we
 277          * iterate over the list in reverse order so that we can just skip that
 278          * refcount instead of bugging.
 279          */
 280
 281         list_for_each_entry_reverse(i, list, list) {
 282                 BUG_ON(last < i->j.seq);
 283                 i->pin = NULL;
 284
 285                 while (last-- != i->j.seq)
 286                         if (fifo_free(&j->pin) > 1) {
 287                                 fifo_push_front(&j->pin, p);
 288                                 atomic_set(&fifo_front(&j->pin), 0);
 289                         }
 290
 291                 if (fifo_free(&j->pin) > 1) {
 292                         fifo_push_front(&j->pin, p);
 293                         i->pin = &fifo_front(&j->pin);
 294                         atomic_set(i->pin, 1);
 295                 }
 296
 297                 for (k = i->j.start;
 298                      k < bset_bkey_last(&i->j);
 299                      k = bkey_next(k))
 300                         if (!__bch_extent_invalid(c, k)) {
 301                                 unsigned j;
 302
 303                                 for (j = 0; j < KEY_PTRS(k); j++)
 304                                         if (ptr_available(c, k, j))
 305                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 306
 307                                 bch_initial_mark_key(c, 0, k);
 308                         }
 309         }
 310 }
 311
 312 bool is_discard_enabled(struct cache_set *s)
 313 {
 314         struct cache *ca;
 315         unsigned int i;
 316
 317         for_each_cache(ca, s, i)
 318                 if (ca->discard)
 319                         return true;
 320
 321         return false;
 322 }
 323
 324 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 325 {
 326         int ret = 0, keys = 0, entries = 0;
 327         struct bkey *k;
 328         struct journal_replay *i =
 329                 list_entry(list->prev, struct journal_replay, list);
 330
 331         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 332         struct keylist keylist;
 333
 334         list_for_each_entry(i, list, list) {
 335                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 336
 337                 if (n != i->j.seq) {
 338                         if (n == start && is_discard_enabled(s))
 339                                 pr_info("bcache: journal entries %llu-%llu may be discarded! (replaying %llu-%llu)",
 340                                         n, i->j.seq - 1, start, end);
 341                         else {
 342                                 pr_err("bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 343                                         n, i->j.seq - 1, start, end);
 344                                 ret = -EIO;
 345                                 goto err;
 346                         }
 347                 }
 348
 349                 for (k = i->j.start;
 350                      k < bset_bkey_last(&i->j);
 351                      k = bkey_next(k)) {
 352                         trace_bcache_journal_replay_key(k);
 353
 354                         bch_keylist_init_single(&keylist, k);
 355
 356                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 357                         if (ret)
 358                                 goto err;
 359
 360                         BUG_ON(!bch_keylist_empty(&keylist));
 361                         keys++;
 362
 363                         cond_resched();
 364                 }
 365
 366                 if (i->pin)
 367                         atomic_dec(i->pin);
 368                 n = i->j.seq + 1;
 369                 entries++;
 370         }
 371
 372         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 373                 keys, entries, end);
 374 err:
 375         while (!list_empty(list)) {
 376                 i = list_first_entry(list, struct journal_replay, list);
 377                 list_del(&i->list);
 378                 kfree(i);
 379         }
 380
 381         return ret;
 382 }
 383
 384 /* Journalling */
 385
 386 static void btree_flush_write(struct cache_set *c)
 387 {
 388         /*
 389          * Try to find the btree node with that references the oldest journal
 390          * entry, best is our current candidate and is locked if non NULL:
 391          */
 392         struct btree *b, *best;
 393         unsigned i;
 394 retry:
 395         best = NULL;
 396
 397         for_each_cached_btree(b, c, i)
 398                 if (btree_current_write(b)->journal) {
 399                         if (!best)
 400                                 best = b;
 401                         else if (journal_pin_cmp(c,
 402                                         btree_current_write(best)->journal,
 403                                         btree_current_write(b)->journal)) {
 404                                 best = b;
 405                         }
 406                 }
 407
 408         b = best;
 409         if (b) {
 410                 mutex_lock(&b->write_lock);
 411                 if (!btree_current_write(b)->journal) {
 412                         mutex_unlock(&b->write_lock);
 413                         /* We raced */
 414                         goto retry;
 415                 }
 416
 417                 __bch_btree_node_write(b, NULL);
 418                 mutex_unlock(&b->write_lock);
 419         }
 420 }
 421
 422 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 423
 424 static void journal_discard_endio(struct bio *bio)
 425 {
 426         struct journal_device *ja =
 427                 container_of(bio, struct journal_device, discard_bio);
 428         struct cache *ca = container_of(ja, struct cache, journal);
 429
 430         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 431
 432         closure_wake_up(&ca->set->journal.wait);
 433         closure_put(&ca->set->cl);
 434 }
 435
 436 static void journal_discard_work(struct work_struct *work)
 437 {
 438         struct journal_device *ja =
 439                 container_of(work, struct journal_device, discard_work);
 440
 441         submit_bio(0, &ja->discard_bio);
 442 }
 443
 444 static void do_journal_discard(struct cache *ca)
 445 {
 446         struct journal_device *ja = &ca->journal;
 447         struct bio *bio = &ja->discard_bio;
 448
 449         if (!ca->discard) {
 450                 ja->discard_idx = ja->last_idx;
 451                 return;
 452         }
 453
 454         switch (atomic_read(&ja->discard_in_flight)) {
 455         case DISCARD_IN_FLIGHT:
 456                 return;
 457
 458         case DISCARD_DONE:
 459                 ja->discard_idx = (ja->discard_idx + 1) %
 460                         ca->sb.njournal_buckets;
 461
 462                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 463                 /* fallthrough */
 464
 465         case DISCARD_READY:
 466                 if (ja->discard_idx == ja->last_idx)
 467                         return;
 468
 469                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 470
 471                 bio_init(bio);
 472                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 473                                                 ca->sb.d[ja->discard_idx]);
 474                 bio->bi_bdev            = ca->bdev;
 475                 bio->bi_rw              = REQ_WRITE|REQ_DISCARD;
 476                 bio->bi_max_vecs        = 1;
 477                 bio->bi_io_vec          = bio->bi_inline_vecs;
 478                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 479                 bio->bi_end_io          = journal_discard_endio;
 480
 481                 closure_get(&ca->set->cl);
 482                 INIT_WORK(&ja->discard_work, journal_discard_work);
 483                 schedule_work(&ja->discard_work);
 484         }
 485 }
 486
 487 static void journal_reclaim(struct cache_set *c)
 488 {
 489         struct bkey *k = &c->journal.key;
 490         struct cache *ca;
 491         uint64_t last_seq;
 492         unsigned iter, n = 0;
 493         atomic_t p;
 494
 495         while (!atomic_read(&fifo_front(&c->journal.pin)))
 496                 fifo_pop(&c->journal.pin, p);
 497
 498         last_seq = last_seq(&c->journal);
 499
 500         /* Update last_idx */
 501
 502         for_each_cache(ca, c, iter) {
 503                 struct journal_device *ja = &ca->journal;
 504
 505                 while (ja->last_idx != ja->cur_idx &&
 506                        ja->seq[ja->last_idx] < last_seq)
 507                         ja->last_idx = (ja->last_idx + 1) %
 508                                 ca->sb.njournal_buckets;
 509         }
 510
 511         for_each_cache(ca, c, iter)
 512                 do_journal_discard(ca);
 513
 514         if (c->journal.blocks_free)
 515                 goto out;
 516
 517         /*
 518          * Allocate:
 519          * XXX: Sort by free journal space
 520          */
 521
 522         for_each_cache(ca, c, iter) {
 523                 struct journal_device *ja = &ca->journal;
 524                 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 525
 526                 /* No space available on this device */
 527                 if (next == ja->discard_idx)
 528                         continue;
 529
 530                 ja->cur_idx = next;
 531                 k->ptr[n++] = MAKE_PTR(0,
 532                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 533                                   ca->sb.nr_this_dev);
 534         }
 535
 536         if (n) {
 537                 bkey_init(k);
 538                 SET_KEY_PTRS(k, n);
 539                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 540         }
 541 out:
 542         if (!journal_full(&c->journal))
 543                 __closure_wake_up(&c->journal.wait);
 544 }
 545
 546 void bch_journal_next(struct journal *j)
 547 {
 548         atomic_t p = { 1 };
 549
 550         j->cur = (j->cur == j->w)
 551                 ? &j->w[1]
 552                 : &j->w[0];
 553
 554         /*
 555          * The fifo_push() needs to happen at the same time as j->seq is
 556          * incremented for last_seq() to be calculated correctly
 557          */
 558         BUG_ON(!fifo_push(&j->pin, p));
 559         atomic_set(&fifo_back(&j->pin), 1);
 560
 561         j->cur->data->seq       = ++j->seq;
 562         j->cur->dirty           = false;
 563         j->cur->need_write      = false;
 564         j->cur->data->keys      = 0;
 565
 566         if (fifo_full(&j->pin))
 567                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 568 }
 569
 570 static void journal_write_endio(struct bio *bio)
 571 {
 572         struct journal_write *w = bio->bi_private;
 573
 574         cache_set_err_on(bio->bi_error, w->c, "journal io error");
 575         closure_put(&w->c->journal.io);
 576 }
 577
 578 static void journal_write(struct closure *);
 579
 580 static void journal_write_done(struct closure *cl)
 581 {
 582         struct journal *j = container_of(cl, struct journal, io);
 583         struct journal_write *w = (j->cur == j->w)
 584                 ? &j->w[1]
 585                 : &j->w[0];
 586
 587         __closure_wake_up(&w->wait);
 588         continue_at_nobarrier(cl, journal_write, system_wq);
 589 }
 590
 591 static void journal_write_unlock(struct closure *cl)
 592 {
 593         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 594
 595         c->journal.io_in_flight = 0;
 596         spin_unlock(&c->journal.lock);
 597 }
 598
 599 static void journal_write_unlocked(struct closure *cl)
 600         __releases(c->journal.lock)
 601 {
 602         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 603         struct cache *ca;
 604         struct journal_write *w = c->journal.cur;
 605         struct bkey *k = &c->journal.key;
 606         unsigned i, sectors = set_blocks(w->data, block_bytes(c)) *
 607                 c->sb.block_size;
 608
 609         struct bio *bio;
 610         struct bio_list list;
 611         bio_list_init(&list);
 612
 613         if (!w->need_write) {
 614                 closure_return_with_destructor(cl, journal_write_unlock);
 615                 return;
 616         } else if (journal_full(&c->journal)) {
 617                 journal_reclaim(c);
 618                 spin_unlock(&c->journal.lock);
 619
 620                 btree_flush_write(c);
 621                 continue_at(cl, journal_write, system_wq);
 622                 return;
 623         }
 624
 625         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 626
 627         w->data->btree_level = c->root->level;
 628
 629         bkey_copy(&w->data->btree_root, &c->root->key);
 630         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 631
 632         for_each_cache(ca, c, i)
 633                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 634
 635         w->data->magic          = jset_magic(&c->sb);
 636         w->data->version        = BCACHE_JSET_VERSION;
 637         w->data->last_seq       = last_seq(&c->journal);
 638         w->data->csum           = csum_set(w->data);
 639
 640         for (i = 0; i < KEY_PTRS(k); i++) {
 641                 ca = PTR_CACHE(c, k, i);
 642                 bio = &ca->journal.bio;
 643
 644                 atomic_long_add(sectors, &ca->meta_sectors_written);
 645
 646                 bio_reset(bio);
 647                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 648                 bio->bi_bdev    = ca->bdev;
 649                 bio->bi_rw      = REQ_WRITE|REQ_SYNC|REQ_META|REQ_FLUSH|REQ_FUA;
 650                 bio->bi_iter.bi_size = sectors << 9;
 651
 652                 bio->bi_end_io  = journal_write_endio;
 653                 bio->bi_private = w;
 654                 bch_bio_map(bio, w->data);
 655
 656                 trace_bcache_journal_write(bio);
 657                 bio_list_add(&list, bio);
 658
 659                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 660
 661                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 662         }
 663
 664         /* If KEY_PTRS(k) == 0, this jset gets lost in air */
 665         BUG_ON(i == 0);
 666
 667         atomic_dec_bug(&fifo_back(&c->journal.pin));
 668         bch_journal_next(&c->journal);
 669         journal_reclaim(c);
 670
 671         spin_unlock(&c->journal.lock);
 672
 673         while ((bio = bio_list_pop(&list)))
 674                 closure_bio_submit(bio, cl);
 675
 676         continue_at(cl, journal_write_done, NULL);
 677 }
 678
 679 static void journal_write(struct closure *cl)
 680 {
 681         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 682
 683         spin_lock(&c->journal.lock);
 684         journal_write_unlocked(cl);
 685 }
 686
 687 static void journal_try_write(struct cache_set *c)
 688         __releases(c->journal.lock)
 689 {
 690         struct closure *cl = &c->journal.io;
 691         struct journal_write *w = c->journal.cur;
 692
 693         w->need_write = true;
 694
 695         if (!c->journal.io_in_flight) {
 696                 c->journal.io_in_flight = 1;
 697                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 698         } else {
 699                 spin_unlock(&c->journal.lock);
 700         }
 701 }
 702
 703 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 704                                                     unsigned nkeys)
 705 {
 706         size_t sectors;
 707         struct closure cl;
 708         bool wait = false;
 709
 710         closure_init_stack(&cl);
 711
 712         spin_lock(&c->journal.lock);
 713
 714         while (1) {
 715                 struct journal_write *w = c->journal.cur;
 716
 717                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 718                                        block_bytes(c)) * c->sb.block_size;
 719
 720                 if (sectors <= min_t(size_t,
 721                                      c->journal.blocks_free * c->sb.block_size,
 722                                      PAGE_SECTORS << JSET_BITS))
 723                         return w;
 724
 725                 if (wait)
 726                         closure_wait(&c->journal.wait, &cl);
 727
 728                 if (!journal_full(&c->journal)) {
 729                         if (wait)
 730                                 trace_bcache_journal_entry_full(c);
 731
 732                         /*
 733                          * XXX: If we were inserting so many keys that they
 734                          * won't fit in an _empty_ journal write, we'll
 735                          * deadlock. For now, handle this in
 736                          * bch_keylist_realloc() - but something to think about.
 737                          */
 738                         BUG_ON(!w->data->keys);
 739
 740                         journal_try_write(c); /* unlocks */
 741                 } else {
 742                         if (wait)
 743                                 trace_bcache_journal_full(c);
 744
 745                         journal_reclaim(c);
 746                         spin_unlock(&c->journal.lock);
 747
 748                         btree_flush_write(c);
 749                 }
 750
 751                 closure_sync(&cl);
 752                 spin_lock(&c->journal.lock);
 753                 wait = true;
 754         }
 755 }
 756
 757 static void journal_write_work(struct work_struct *work)
 758 {
 759         struct cache_set *c = container_of(to_delayed_work(work),
 760                                            struct cache_set,
 761                                            journal.work);
 762         spin_lock(&c->journal.lock);
 763         if (c->journal.cur->dirty)
 764                 journal_try_write(c);
 765         else
 766                 spin_unlock(&c->journal.lock);
 767 }
 768
 769 /*
 770  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 771  * pass bch_journal() a list of keys to be journalled, and then
 772  * bch_journal() hands those same keys off to btree_insert_async()
 773  */
 774
 775 atomic_t *bch_journal(struct cache_set *c,
 776                       struct keylist *keys,
 777                       struct closure *parent)
 778 {
 779         struct journal_write *w;
 780         atomic_t *ret;
 781
 782         if (!CACHE_SYNC(&c->sb))
 783                 return NULL;
 784
 785         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 786
 787         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 788         w->data->keys += bch_keylist_nkeys(keys);
 789
 790         ret = &fifo_back(&c->journal.pin);
 791         atomic_inc(ret);
 792
 793         if (parent) {
 794                 closure_wait(&w->wait, parent);
 795                 journal_try_write(c);
 796         } else if (!w->dirty) {
 797                 w->dirty = true;
 798                 schedule_delayed_work(&c->journal.work,
 799                                       msecs_to_jiffies(c->journal_delay_ms));
 800                 spin_unlock(&c->journal.lock);
 801         } else {
 802                 spin_unlock(&c->journal.lock);
 803         }
 804
 805
 806         return ret;
 807 }
 808
 809 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 810 {
 811         struct keylist keys;
 812         atomic_t *ref;
 813
 814         bch_keylist_init(&keys);
 815
 816         ref = bch_journal(c, &keys, cl);
 817         if (ref)
 818                 atomic_dec_bug(ref);
 819 }
 820
 821 void bch_journal_free(struct cache_set *c)
 822 {
 823         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 824         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 825         free_fifo(&c->journal.pin);
 826 }
 827
 828 int bch_journal_alloc(struct cache_set *c)
 829 {
 830         struct journal *j = &c->journal;
 831
 832         spin_lock_init(&j->lock);
 833         INIT_DELAYED_WORK(&j->work, journal_write_work);
 834
 835         c->journal_delay_ms = 100;
 836
 837         j->w[0].c = c;
 838         j->w[1].c = c;
 839
 840         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 841             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) ||
 842             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)))
 843                 return -ENOMEM;
 844
 845         return 0;
 846 }