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                         l = (i * 2654435769U) % ca->sb.njournal_buckets;
 174
 175                         if (test_bit(l, bitmap))
 176                                 break;
 177
 178                         if (read_bucket(l))
 179                                 goto bsearch;
 180                 }
 181
 182                 /*
 183                  * If that fails, check all the buckets we haven't checked
 184                  * already
 185                  */
 186                 pr_debug("falling back to linear search");
 187
 188                 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
 189                      l < ca->sb.njournal_buckets;
 190                      l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets, l + 1))
 191                         if (read_bucket(l))
 192                                 goto bsearch;
 193
 194                 /* no journal entries on this device? */
 195                 if (l == ca->sb.njournal_buckets)
 196                         continue;
 197 bsearch:
 198                 BUG_ON(list_empty(list));
 199
 200                 /* Binary search */
 201                 m = l;
 202                 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 203                 pr_debug("starting binary search, l %u r %u", l, r);
 204
 205                 while (l + 1 < r) {
 206                         seq = list_entry(list->prev, struct journal_replay,
 207                                          list)->j.seq;
 208
 209                         m = (l + r) >> 1;
 210                         read_bucket(m);
 211
 212                         if (seq != list_entry(list->prev, struct journal_replay,
 213                                               list)->j.seq)
 214                                 l = m;
 215                         else
 216                                 r = m;
 217                 }
 218
 219                 /*
 220                  * Read buckets in reverse order until we stop finding more
 221                  * journal entries
 222                  */
 223                 pr_debug("finishing up: m %u njournal_buckets %u",
 224                          m, ca->sb.njournal_buckets);
 225                 l = m;
 226
 227                 while (1) {
 228                         if (!l--)
 229                                 l = ca->sb.njournal_buckets - 1;
 230
 231                         if (l == m)
 232                                 break;
 233
 234                         if (test_bit(l, bitmap))
 235                                 continue;
 236
 237                         if (!read_bucket(l))
 238                                 break;
 239                 }
 240
 241                 seq = 0;
 242
 243                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 244                         if (ja->seq[i] > seq) {
 245                                 seq = ja->seq[i];
 246                                 /*
 247                                  * When journal_reclaim() goes to allocate for
 248                                  * the first time, it'll use the bucket after
 249                                  * ja->cur_idx
 250                                  */
 251                                 ja->cur_idx = i;
 252                                 ja->last_idx = ja->discard_idx = (i + 1) %
 253                                         ca->sb.njournal_buckets;
 254
 255                         }
 256         }
 257
 258         if (!list_empty(list))
 259                 c->journal.seq = list_entry(list->prev,
 260                                             struct journal_replay,
 261                                             list)->j.seq;
 262
 263         return 0;
 264 #undef read_bucket
 265 }
 266
 267 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 268 {
 269         atomic_t p = { 0 };
 270         struct bkey *k;
 271         struct journal_replay *i;
 272         struct journal *j = &c->journal;
 273         uint64_t last = j->seq;
 274
 275         /*
 276          * journal.pin should never fill up - we never write a journal
 277          * entry when it would fill up. But if for some reason it does, we
 278          * iterate over the list in reverse order so that we can just skip that
 279          * refcount instead of bugging.
 280          */
 281
 282         list_for_each_entry_reverse(i, list, list) {
 283                 BUG_ON(last < i->j.seq);
 284                 i->pin = NULL;
 285
 286                 while (last-- != i->j.seq)
 287                         if (fifo_free(&j->pin) > 1) {
 288                                 fifo_push_front(&j->pin, p);
 289                                 atomic_set(&fifo_front(&j->pin), 0);
 290                         }
 291
 292                 if (fifo_free(&j->pin) > 1) {
 293                         fifo_push_front(&j->pin, p);
 294                         i->pin = &fifo_front(&j->pin);
 295                         atomic_set(i->pin, 1);
 296                 }
 297
 298                 for (k = i->j.start;
 299                      k < bset_bkey_last(&i->j);
 300                      k = bkey_next(k))
 301                         if (!__bch_extent_invalid(c, k)) {
 302                                 unsigned j;
 303
 304                                 for (j = 0; j < KEY_PTRS(k); j++)
 305                                         if (ptr_available(c, k, j))
 306                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 307
 308                                 bch_initial_mark_key(c, 0, k);
 309                         }
 310         }
 311 }
 312
 313 bool is_discard_enabled(struct cache_set *s)
 314 {
 315         struct cache *ca;
 316         unsigned int i;
 317
 318         for_each_cache(ca, s, i)
 319                 if (ca->discard)
 320                         return true;
 321
 322         return false;
 323 }
 324
 325 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 326 {
 327         int ret = 0, keys = 0, entries = 0;
 328         struct bkey *k;
 329         struct journal_replay *i =
 330                 list_entry(list->prev, struct journal_replay, list);
 331
 332         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 333         struct keylist keylist;
 334
 335         list_for_each_entry(i, list, list) {
 336                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 337
 338                 if (n != i->j.seq) {
 339                         if (n == start && is_discard_enabled(s))
 340                                 pr_info("bcache: journal entries %llu-%llu may be discarded! (replaying %llu-%llu)",
 341                                         n, i->j.seq - 1, start, end);
 342                         else {
 343                                 pr_err("bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 344                                         n, i->j.seq - 1, start, end);
 345                                 ret = -EIO;
 346                                 goto err;
 347                         }
 348                 }
 349
 350                 for (k = i->j.start;
 351                      k < bset_bkey_last(&i->j);
 352                      k = bkey_next(k)) {
 353                         trace_bcache_journal_replay_key(k);
 354
 355                         bch_keylist_init_single(&keylist, k);
 356
 357                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 358                         if (ret)
 359                                 goto err;
 360
 361                         BUG_ON(!bch_keylist_empty(&keylist));
 362                         keys++;
 363
 364                         cond_resched();
 365                 }
 366
 367                 if (i->pin)
 368                         atomic_dec(i->pin);
 369                 n = i->j.seq + 1;
 370                 entries++;
 371         }
 372
 373         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 374                 keys, entries, end);
 375 err:
 376         while (!list_empty(list)) {
 377                 i = list_first_entry(list, struct journal_replay, list);
 378                 list_del(&i->list);
 379                 kfree(i);
 380         }
 381
 382         return ret;
 383 }
 384
 385 /* Journalling */
 386
 387 static void btree_flush_write(struct cache_set *c)
 388 {
 389         /*
 390          * Try to find the btree node with that references the oldest journal
 391          * entry, best is our current candidate and is locked if non NULL:
 392          */
 393         struct btree *b, *best;
 394         unsigned i;
 395 retry:
 396         best = NULL;
 397
 398         for_each_cached_btree(b, c, i)
 399                 if (btree_current_write(b)->journal) {
 400                         if (!best)
 401                                 best = b;
 402                         else if (journal_pin_cmp(c,
 403                                         btree_current_write(best)->journal,
 404                                         btree_current_write(b)->journal)) {
 405                                 best = b;
 406                         }
 407                 }
 408
 409         b = best;
 410         if (b) {
 411                 mutex_lock(&b->write_lock);
 412                 if (!btree_current_write(b)->journal) {
 413                         mutex_unlock(&b->write_lock);
 414                         /* We raced */
 415                         goto retry;
 416                 }
 417
 418                 __bch_btree_node_write(b, NULL);
 419                 mutex_unlock(&b->write_lock);
 420         }
 421 }
 422
 423 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 424
 425 static void journal_discard_endio(struct bio *bio)
 426 {
 427         struct journal_device *ja =
 428                 container_of(bio, struct journal_device, discard_bio);
 429         struct cache *ca = container_of(ja, struct cache, journal);
 430
 431         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 432
 433         closure_wake_up(&ca->set->journal.wait);
 434         closure_put(&ca->set->cl);
 435 }
 436
 437 static void journal_discard_work(struct work_struct *work)
 438 {
 439         struct journal_device *ja =
 440                 container_of(work, struct journal_device, discard_work);
 441
 442         submit_bio(&ja->discard_bio);
 443 }
 444
 445 static void do_journal_discard(struct cache *ca)
 446 {
 447         struct journal_device *ja = &ca->journal;
 448         struct bio *bio = &ja->discard_bio;
 449
 450         if (!ca->discard) {
 451                 ja->discard_idx = ja->last_idx;
 452                 return;
 453         }
 454
 455         switch (atomic_read(&ja->discard_in_flight)) {
 456         case DISCARD_IN_FLIGHT:
 457                 return;
 458
 459         case DISCARD_DONE:
 460                 ja->discard_idx = (ja->discard_idx + 1) %
 461                         ca->sb.njournal_buckets;
 462
 463                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 464                 /* fallthrough */
 465
 466         case DISCARD_READY:
 467                 if (ja->discard_idx == ja->last_idx)
 468                         return;
 469
 470                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 471
 472                 bio_init(bio, bio->bi_inline_vecs, 1);
 473                 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
 474                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 475                                                 ca->sb.d[ja->discard_idx]);
 476                 bio_set_dev(bio, ca->bdev);
 477                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 478                 bio->bi_end_io          = journal_discard_endio;
 479
 480                 closure_get(&ca->set->cl);
 481                 INIT_WORK(&ja->discard_work, journal_discard_work);
 482                 schedule_work(&ja->discard_work);
 483         }
 484 }
 485
 486 static void journal_reclaim(struct cache_set *c)
 487 {
 488         struct bkey *k = &c->journal.key;
 489         struct cache *ca;
 490         uint64_t last_seq;
 491         unsigned iter, n = 0;
 492         atomic_t p;
 493
 494         while (!atomic_read(&fifo_front(&c->journal.pin)))
 495                 fifo_pop(&c->journal.pin, p);
 496
 497         last_seq = last_seq(&c->journal);
 498
 499         /* Update last_idx */
 500
 501         for_each_cache(ca, c, iter) {
 502                 struct journal_device *ja = &ca->journal;
 503
 504                 while (ja->last_idx != ja->cur_idx &&
 505                        ja->seq[ja->last_idx] < last_seq)
 506                         ja->last_idx = (ja->last_idx + 1) %
 507                                 ca->sb.njournal_buckets;
 508         }
 509
 510         for_each_cache(ca, c, iter)
 511                 do_journal_discard(ca);
 512
 513         if (c->journal.blocks_free)
 514                 goto out;
 515
 516         /*
 517          * Allocate:
 518          * XXX: Sort by free journal space
 519          */
 520
 521         for_each_cache(ca, c, iter) {
 522                 struct journal_device *ja = &ca->journal;
 523                 unsigned next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 524
 525                 /* No space available on this device */
 526                 if (next == ja->discard_idx)
 527                         continue;
 528
 529                 ja->cur_idx = next;
 530                 k->ptr[n++] = MAKE_PTR(0,
 531                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 532                                   ca->sb.nr_this_dev);
 533         }
 534
 535         if (n) {
 536                 bkey_init(k);
 537                 SET_KEY_PTRS(k, n);
 538                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 539         }
 540 out:
 541         if (!journal_full(&c->journal))
 542                 __closure_wake_up(&c->journal.wait);
 543 }
 544
 545 void bch_journal_next(struct journal *j)
 546 {
 547         atomic_t p = { 1 };
 548
 549         j->cur = (j->cur == j->w)
 550                 ? &j->w[1]
 551                 : &j->w[0];
 552
 553         /*
 554          * The fifo_push() needs to happen at the same time as j->seq is
 555          * incremented for last_seq() to be calculated correctly
 556          */
 557         BUG_ON(!fifo_push(&j->pin, p));
 558         atomic_set(&fifo_back(&j->pin), 1);
 559
 560         j->cur->data->seq       = ++j->seq;
 561         j->cur->dirty           = false;
 562         j->cur->need_write      = false;
 563         j->cur->data->keys      = 0;
 564
 565         if (fifo_full(&j->pin))
 566                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 567 }
 568
 569 static void journal_write_endio(struct bio *bio)
 570 {
 571         struct journal_write *w = bio->bi_private;
 572
 573         cache_set_err_on(bio->bi_status, w->c, "journal io error");
 574         closure_put(&w->c->journal.io);
 575 }
 576
 577 static void journal_write(struct closure *);
 578
 579 static void journal_write_done(struct closure *cl)
 580 {
 581         struct journal *j = container_of(cl, struct journal, io);
 582         struct journal_write *w = (j->cur == j->w)
 583                 ? &j->w[1]
 584                 : &j->w[0];
 585
 586         __closure_wake_up(&w->wait);
 587         continue_at_nobarrier(cl, journal_write, system_wq);
 588 }
 589
 590 static void journal_write_unlock(struct closure *cl)
 591 {
 592         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 593
 594         c->journal.io_in_flight = 0;
 595         spin_unlock(&c->journal.lock);
 596 }
 597
 598 static void journal_write_unlocked(struct closure *cl)
 599         __releases(c->journal.lock)
 600 {
 601         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 602         struct cache *ca;
 603         struct journal_write *w = c->journal.cur;
 604         struct bkey *k = &c->journal.key;
 605         unsigned i, sectors = set_blocks(w->data, block_bytes(c)) *
 606                 c->sb.block_size;
 607
 608         struct bio *bio;
 609         struct bio_list list;
 610         bio_list_init(&list);
 611
 612         if (!w->need_write) {
 613                 closure_return_with_destructor(cl, journal_write_unlock);
 614                 return;
 615         } else if (journal_full(&c->journal)) {
 616                 journal_reclaim(c);
 617                 spin_unlock(&c->journal.lock);
 618
 619                 btree_flush_write(c);
 620                 continue_at(cl, journal_write, system_wq);
 621                 return;
 622         }
 623
 624         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 625
 626         w->data->btree_level = c->root->level;
 627
 628         bkey_copy(&w->data->btree_root, &c->root->key);
 629         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 630
 631         for_each_cache(ca, c, i)
 632                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 633
 634         w->data->magic          = jset_magic(&c->sb);
 635         w->data->version        = BCACHE_JSET_VERSION;
 636         w->data->last_seq       = last_seq(&c->journal);
 637         w->data->csum           = csum_set(w->data);
 638
 639         for (i = 0; i < KEY_PTRS(k); i++) {
 640                 ca = PTR_CACHE(c, k, i);
 641                 bio = &ca->journal.bio;
 642
 643                 atomic_long_add(sectors, &ca->meta_sectors_written);
 644
 645                 bio_reset(bio);
 646                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 647                 bio_set_dev(bio, ca->bdev);
 648                 bio->bi_iter.bi_size = sectors << 9;
 649
 650                 bio->bi_end_io  = journal_write_endio;
 651                 bio->bi_private = w;
 652                 bio_set_op_attrs(bio, REQ_OP_WRITE,
 653                                  REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
 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 }