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