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
  32         closure_put(cl);
  33 }
  34
  35 static int journal_read_bucket(struct cache *ca, struct list_head *list,
  36                                unsigned int bucket_index)
  37 {
  38         struct journal_device *ja = &ca->journal;
  39         struct bio *bio = &ja->bio;
  40
  41         struct journal_replay *i;
  42         struct jset *j, *data = ca->set->journal.w[0].data;
  43         struct closure cl;
  44         unsigned int len, left, offset = 0;
  45         int ret = 0;
  46         sector_t bucket = bucket_to_sector(ca->set, ca->sb.d[bucket_index]);
  47
  48         closure_init_stack(&cl);
  49
  50         pr_debug("reading %u", bucket_index);
  51
  52         while (offset < ca->sb.bucket_size) {
  53 reread:         left = ca->sb.bucket_size - offset;
  54                 len = min_t(unsigned int, left, PAGE_SECTORS << JSET_BITS);
  55
  56                 bio_reset(bio);
  57                 bio->bi_iter.bi_sector  = bucket + offset;
  58                 bio_set_dev(bio, ca->bdev);
  59                 bio->bi_iter.bi_size    = len << 9;
  60
  61                 bio->bi_end_io  = journal_read_endio;
  62                 bio->bi_private = &cl;
  63                 bio_set_op_attrs(bio, REQ_OP_READ, 0);
  64                 bch_bio_map(bio, data);
  65
  66                 closure_bio_submit(ca->set, bio, &cl);
  67                 closure_sync(&cl);
  68
  69                 /* This function could be simpler now since we no longer write
  70                  * journal entries that overlap bucket boundaries; this means
  71                  * the start of a bucket will always have a valid journal entry
  72                  * if it has any journal entries at all.
  73                  */
  74
  75                 j = data;
  76                 while (len) {
  77                         struct list_head *where;
  78                         size_t blocks, bytes = set_bytes(j);
  79
  80                         if (j->magic != jset_magic(&ca->sb)) {
  81                                 pr_debug("%u: bad magic", bucket_index);
  82                                 return ret;
  83                         }
  84
  85                         if (bytes > left << 9 ||
  86                             bytes > PAGE_SIZE << JSET_BITS) {
  87                                 pr_info("%u: too big, %zu bytes, offset %u",
  88                                         bucket_index, bytes, offset);
  89                                 return ret;
  90                         }
  91
  92                         if (bytes > len << 9)
  93                                 goto reread;
  94
  95                         if (j->csum != csum_set(j)) {
  96                                 pr_info("%u: bad csum, %zu bytes, offset %u",
  97                                         bucket_index, bytes, offset);
  98                                 return ret;
  99                         }
 100
 101                         blocks = set_blocks(j, block_bytes(ca->set));
 102
 103                         while (!list_empty(list)) {
 104                                 i = list_first_entry(list,
 105                                         struct journal_replay, list);
 106                                 if (i->j.seq >= j->last_seq)
 107                                         break;
 108                                 list_del(&i->list);
 109                                 kfree(i);
 110                         }
 111
 112                         list_for_each_entry_reverse(i, list, list) {
 113                                 if (j->seq == i->j.seq)
 114                                         goto next_set;
 115
 116                                 if (j->seq < i->j.last_seq)
 117                                         goto next_set;
 118
 119                                 if (j->seq > i->j.seq) {
 120                                         where = &i->list;
 121                                         goto add;
 122                                 }
 123                         }
 124
 125                         where = list;
 126 add:
 127                         i = kmalloc(offsetof(struct journal_replay, j) +
 128                                     bytes, GFP_KERNEL);
 129                         if (!i)
 130                                 return -ENOMEM;
 131                         memcpy(&i->j, j, bytes);
 132                         list_add(&i->list, where);
 133                         ret = 1;
 134
 135                         ja->seq[bucket_index] = j->seq;
 136 next_set:
 137                         offset  += blocks * ca->sb.block_size;
 138                         len     -= blocks * ca->sb.block_size;
 139                         j = ((void *) j) + blocks * block_bytes(ca);
 140                 }
 141         }
 142
 143         return ret;
 144 }
 145
 146 int bch_journal_read(struct cache_set *c, struct list_head *list)
 147 {
 148 #define read_bucket(b)                                                  \
 149         ({                                                              \
 150                 int ret = journal_read_bucket(ca, list, b);             \
 151                 __set_bit(b, bitmap);                                   \
 152                 if (ret < 0)                                            \
 153                         return ret;                                     \
 154                 ret;                                                    \
 155         })
 156
 157         struct cache *ca;
 158         unsigned int iter;
 159
 160         for_each_cache(ca, c, iter) {
 161                 struct journal_device *ja = &ca->journal;
 162                 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
 163                 unsigned int i, l, r, m;
 164                 uint64_t seq;
 165
 166                 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
 167                 pr_debug("%u journal buckets", ca->sb.njournal_buckets);
 168
 169                 /*
 170                  * Read journal buckets ordered by golden ratio hash to quickly
 171                  * find a sequence of buckets with valid journal entries
 172                  */
 173                 for (i = 0; i < ca->sb.njournal_buckets; i++) {
 174                         /*
 175                          * We must try the index l with ZERO first for
 176                          * correctness due to the scenario that the journal
 177                          * bucket is circular buffer which might have wrapped
 178                          */
 179                         l = (i * 2654435769U) % ca->sb.njournal_buckets;
 180
 181                         if (test_bit(l, bitmap))
 182                                 break;
 183
 184                         if (read_bucket(l))
 185                                 goto bsearch;
 186                 }
 187
 188                 /*
 189                  * If that fails, check all the buckets we haven't checked
 190                  * already
 191                  */
 192                 pr_debug("falling back to linear search");
 193
 194                 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
 195                      l < ca->sb.njournal_buckets;
 196                      l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets,
 197                                             l + 1))
 198                         if (read_bucket(l))
 199                                 goto bsearch;
 200
 201                 /* no journal entries on this device? */
 202                 if (l == ca->sb.njournal_buckets)
 203                         continue;
 204 bsearch:
 205                 BUG_ON(list_empty(list));
 206
 207                 /* Binary search */
 208                 m = l;
 209                 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 210                 pr_debug("starting binary search, l %u r %u", l, r);
 211
 212                 while (l + 1 < r) {
 213                         seq = list_entry(list->prev, struct journal_replay,
 214                                          list)->j.seq;
 215
 216                         m = (l + r) >> 1;
 217                         read_bucket(m);
 218
 219                         if (seq != list_entry(list->prev, struct journal_replay,
 220                                               list)->j.seq)
 221                                 l = m;
 222                         else
 223                                 r = m;
 224                 }
 225
 226                 /*
 227                  * Read buckets in reverse order until we stop finding more
 228                  * journal entries
 229                  */
 230                 pr_debug("finishing up: m %u njournal_buckets %u",
 231                          m, ca->sb.njournal_buckets);
 232                 l = m;
 233
 234                 while (1) {
 235                         if (!l--)
 236                                 l = ca->sb.njournal_buckets - 1;
 237
 238                         if (l == m)
 239                                 break;
 240
 241                         if (test_bit(l, bitmap))
 242                                 continue;
 243
 244                         if (!read_bucket(l))
 245                                 break;
 246                 }
 247
 248                 seq = 0;
 249
 250                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 251                         if (ja->seq[i] > seq) {
 252                                 seq = ja->seq[i];
 253                                 /*
 254                                  * When journal_reclaim() goes to allocate for
 255                                  * the first time, it'll use the bucket after
 256                                  * ja->cur_idx
 257                                  */
 258                                 ja->cur_idx = i;
 259                                 ja->last_idx = ja->discard_idx = (i + 1) %
 260                                         ca->sb.njournal_buckets;
 261
 262                         }
 263         }
 264
 265         if (!list_empty(list))
 266                 c->journal.seq = list_entry(list->prev,
 267                                             struct journal_replay,
 268                                             list)->j.seq;
 269
 270         return 0;
 271 #undef read_bucket
 272 }
 273
 274 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 275 {
 276         atomic_t p = { 0 };
 277         struct bkey *k;
 278         struct journal_replay *i;
 279         struct journal *j = &c->journal;
 280         uint64_t last = j->seq;
 281
 282         /*
 283          * journal.pin should never fill up - we never write a journal
 284          * entry when it would fill up. But if for some reason it does, we
 285          * iterate over the list in reverse order so that we can just skip that
 286          * refcount instead of bugging.
 287          */
 288
 289         list_for_each_entry_reverse(i, list, list) {
 290                 BUG_ON(last < i->j.seq);
 291                 i->pin = NULL;
 292
 293                 while (last-- != i->j.seq)
 294                         if (fifo_free(&j->pin) > 1) {
 295                                 fifo_push_front(&j->pin, p);
 296                                 atomic_set(&fifo_front(&j->pin), 0);
 297                         }
 298
 299                 if (fifo_free(&j->pin) > 1) {
 300                         fifo_push_front(&j->pin, p);
 301                         i->pin = &fifo_front(&j->pin);
 302                         atomic_set(i->pin, 1);
 303                 }
 304
 305                 for (k = i->j.start;
 306                      k < bset_bkey_last(&i->j);
 307                      k = bkey_next(k))
 308                         if (!__bch_extent_invalid(c, k)) {
 309                                 unsigned int j;
 310
 311                                 for (j = 0; j < KEY_PTRS(k); j++)
 312                                         if (ptr_available(c, k, j))
 313                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 314
 315                                 bch_initial_mark_key(c, 0, k);
 316                         }
 317         }
 318 }
 319
 320 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 321 {
 322         int ret = 0, keys = 0, entries = 0;
 323         struct bkey *k;
 324         struct journal_replay *i =
 325                 list_entry(list->prev, struct journal_replay, list);
 326
 327         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 328         struct keylist keylist;
 329
 330         list_for_each_entry(i, list, list) {
 331                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 332
 333                 cache_set_err_on(n != i->j.seq, s,
 334 "bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 335                                  n, i->j.seq - 1, start, end);
 336
 337                 for (k = i->j.start;
 338                      k < bset_bkey_last(&i->j);
 339                      k = bkey_next(k)) {
 340                         trace_bcache_journal_replay_key(k);
 341
 342                         bch_keylist_init_single(&keylist, k);
 343
 344                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 345                         if (ret)
 346                                 goto err;
 347
 348                         BUG_ON(!bch_keylist_empty(&keylist));
 349                         keys++;
 350
 351                         cond_resched();
 352                 }
 353
 354                 if (i->pin)
 355                         atomic_dec(i->pin);
 356                 n = i->j.seq + 1;
 357                 entries++;
 358         }
 359
 360         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 361                 keys, entries, end);
 362 err:
 363         while (!list_empty(list)) {
 364                 i = list_first_entry(list, struct journal_replay, list);
 365                 list_del(&i->list);
 366                 kfree(i);
 367         }
 368
 369         return ret;
 370 }
 371
 372 /* Journalling */
 373 #define journal_max_cmp(l, r) \
 374         (fifo_idx(&c->journal.pin, btree_current_write(l)->journal) < \
 375          fifo_idx(&(c)->journal.pin, btree_current_write(r)->journal))
 376 #define journal_min_cmp(l, r) \
 377         (fifo_idx(&c->journal.pin, btree_current_write(l)->journal) > \
 378          fifo_idx(&(c)->journal.pin, btree_current_write(r)->journal))
 379
 380 static void btree_flush_write(struct cache_set *c)
 381 {
 382         /*
 383          * Try to find the btree node with that references the oldest journal
 384          * entry, best is our current candidate and is locked if non NULL:
 385          */
 386         struct btree *b;
 387         int i;
 388
 389         atomic_long_inc(&c->flush_write);
 390
 391 retry:
 392         spin_lock(&c->journal.lock);
 393         if (heap_empty(&c->flush_btree)) {
 394                 for_each_cached_btree(b, c, i)
 395                         if (btree_current_write(b)->journal) {
 396                                 if (!heap_full(&c->flush_btree))
 397                                         heap_add(&c->flush_btree, b,
 398                                                  journal_max_cmp);
 399                                 else if (journal_max_cmp(b,
 400                                          heap_peek(&c->flush_btree))) {
 401                                         c->flush_btree.data[0] = b;
 402                                         heap_sift(&c->flush_btree, 0,
 403                                                   journal_max_cmp);
 404                                 }
 405                         }
 406
 407                 for (i = c->flush_btree.used / 2 - 1; i >= 0; --i)
 408                         heap_sift(&c->flush_btree, i, journal_min_cmp);
 409         }
 410
 411         b = NULL;
 412         heap_pop(&c->flush_btree, b, journal_min_cmp);
 413         spin_unlock(&c->journal.lock);
 414
 415         if (b) {
 416                 mutex_lock(&b->write_lock);
 417                 if (!btree_current_write(b)->journal) {
 418                         mutex_unlock(&b->write_lock);
 419                         /* We raced */
 420                         atomic_long_inc(&c->retry_flush_write);
 421                         goto retry;
 422                 }
 423
 424                 __bch_btree_node_write(b, NULL);
 425                 mutex_unlock(&b->write_lock);
 426         }
 427 }
 428
 429 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 430
 431 static void journal_discard_endio(struct bio *bio)
 432 {
 433         struct journal_device *ja =
 434                 container_of(bio, struct journal_device, discard_bio);
 435         struct cache *ca = container_of(ja, struct cache, journal);
 436
 437         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 438
 439         closure_wake_up(&ca->set->journal.wait);
 440         closure_put(&ca->set->cl);
 441 }
 442
 443 static void journal_discard_work(struct work_struct *work)
 444 {
 445         struct journal_device *ja =
 446                 container_of(work, struct journal_device, discard_work);
 447
 448         submit_bio(&ja->discard_bio);
 449 }
 450
 451 static void do_journal_discard(struct cache *ca)
 452 {
 453         struct journal_device *ja = &ca->journal;
 454         struct bio *bio = &ja->discard_bio;
 455
 456         if (!ca->discard) {
 457                 ja->discard_idx = ja->last_idx;
 458                 return;
 459         }
 460
 461         switch (atomic_read(&ja->discard_in_flight)) {
 462         case DISCARD_IN_FLIGHT:
 463                 return;
 464
 465         case DISCARD_DONE:
 466                 ja->discard_idx = (ja->discard_idx + 1) %
 467                         ca->sb.njournal_buckets;
 468
 469                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 470                 /* fallthrough */
 471
 472         case DISCARD_READY:
 473                 if (ja->discard_idx == ja->last_idx)
 474                         return;
 475
 476                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 477
 478                 bio_init(bio, bio->bi_inline_vecs, 1);
 479                 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
 480                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 481                                                 ca->sb.d[ja->discard_idx]);
 482                 bio_set_dev(bio, ca->bdev);
 483                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 484                 bio->bi_end_io          = journal_discard_endio;
 485
 486                 closure_get(&ca->set->cl);
 487                 INIT_WORK(&ja->discard_work, journal_discard_work);
 488                 queue_work(bch_journal_wq, &ja->discard_work);
 489         }
 490 }
 491
 492 static void journal_reclaim(struct cache_set *c)
 493 {
 494         struct bkey *k = &c->journal.key;
 495         struct cache *ca;
 496         uint64_t last_seq;
 497         unsigned int iter, n = 0;
 498         atomic_t p __maybe_unused;
 499
 500         atomic_long_inc(&c->reclaim);
 501
 502         while (!atomic_read(&fifo_front(&c->journal.pin)))
 503                 fifo_pop(&c->journal.pin, p);
 504
 505         last_seq = last_seq(&c->journal);
 506
 507         /* Update last_idx */
 508
 509         for_each_cache(ca, c, iter) {
 510                 struct journal_device *ja = &ca->journal;
 511
 512                 while (ja->last_idx != ja->cur_idx &&
 513                        ja->seq[ja->last_idx] < last_seq)
 514                         ja->last_idx = (ja->last_idx + 1) %
 515                                 ca->sb.njournal_buckets;
 516         }
 517
 518         for_each_cache(ca, c, iter)
 519                 do_journal_discard(ca);
 520
 521         if (c->journal.blocks_free)
 522                 goto out;
 523
 524         /*
 525          * Allocate:
 526          * XXX: Sort by free journal space
 527          */
 528
 529         for_each_cache(ca, c, iter) {
 530                 struct journal_device *ja = &ca->journal;
 531                 unsigned int next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 532
 533                 /* No space available on this device */
 534                 if (next == ja->discard_idx)
 535                         continue;
 536
 537                 ja->cur_idx = next;
 538                 k->ptr[n++] = MAKE_PTR(0,
 539                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 540                                   ca->sb.nr_this_dev);
 541         }
 542
 543         bkey_init(k);
 544         SET_KEY_PTRS(k, n);
 545
 546         if (n)
 547                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 548 out:
 549         if (!journal_full(&c->journal))
 550                 __closure_wake_up(&c->journal.wait);
 551 }
 552
 553 void bch_journal_next(struct journal *j)
 554 {
 555         atomic_t p = { 1 };
 556
 557         j->cur = (j->cur == j->w)
 558                 ? &j->w[1]
 559                 : &j->w[0];
 560
 561         /*
 562          * The fifo_push() needs to happen at the same time as j->seq is
 563          * incremented for last_seq() to be calculated correctly
 564          */
 565         BUG_ON(!fifo_push(&j->pin, p));
 566         atomic_set(&fifo_back(&j->pin), 1);
 567
 568         j->cur->data->seq       = ++j->seq;
 569         j->cur->dirty           = false;
 570         j->cur->need_write      = false;
 571         j->cur->data->keys      = 0;
 572
 573         if (fifo_full(&j->pin))
 574                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 575 }
 576
 577 static void journal_write_endio(struct bio *bio)
 578 {
 579         struct journal_write *w = bio->bi_private;
 580
 581         cache_set_err_on(bio->bi_status, w->c, "journal io error");
 582         closure_put(&w->c->journal.io);
 583 }
 584
 585 static void journal_write(struct closure *cl);
 586
 587 static void journal_write_done(struct closure *cl)
 588 {
 589         struct journal *j = container_of(cl, struct journal, io);
 590         struct journal_write *w = (j->cur == j->w)
 591                 ? &j->w[1]
 592                 : &j->w[0];
 593
 594         __closure_wake_up(&w->wait);
 595         continue_at_nobarrier(cl, journal_write, bch_journal_wq);
 596 }
 597
 598 static void journal_write_unlock(struct closure *cl)
 599         __releases(&c->journal.lock)
 600 {
 601         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 602
 603         c->journal.io_in_flight = 0;
 604         spin_unlock(&c->journal.lock);
 605 }
 606
 607 static void journal_write_unlocked(struct closure *cl)
 608         __releases(c->journal.lock)
 609 {
 610         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 611         struct cache *ca;
 612         struct journal_write *w = c->journal.cur;
 613         struct bkey *k = &c->journal.key;
 614         unsigned int i, sectors = set_blocks(w->data, block_bytes(c)) *
 615                 c->sb.block_size;
 616
 617         struct bio *bio;
 618         struct bio_list list;
 619
 620         bio_list_init(&list);
 621
 622         if (!w->need_write) {
 623                 closure_return_with_destructor(cl, journal_write_unlock);
 624                 return;
 625         } else if (journal_full(&c->journal)) {
 626                 journal_reclaim(c);
 627                 spin_unlock(&c->journal.lock);
 628
 629                 btree_flush_write(c);
 630                 continue_at(cl, journal_write, bch_journal_wq);
 631                 return;
 632         }
 633
 634         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 635
 636         w->data->btree_level = c->root->level;
 637
 638         bkey_copy(&w->data->btree_root, &c->root->key);
 639         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 640
 641         for_each_cache(ca, c, i)
 642                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 643
 644         w->data->magic          = jset_magic(&c->sb);
 645         w->data->version        = BCACHE_JSET_VERSION;
 646         w->data->last_seq       = last_seq(&c->journal);
 647         w->data->csum           = csum_set(w->data);
 648
 649         for (i = 0; i < KEY_PTRS(k); i++) {
 650                 ca = PTR_CACHE(c, k, i);
 651                 bio = &ca->journal.bio;
 652
 653                 atomic_long_add(sectors, &ca->meta_sectors_written);
 654
 655                 bio_reset(bio);
 656                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 657                 bio_set_dev(bio, ca->bdev);
 658                 bio->bi_iter.bi_size = sectors << 9;
 659
 660                 bio->bi_end_io  = journal_write_endio;
 661                 bio->bi_private = w;
 662                 bio_set_op_attrs(bio, REQ_OP_WRITE,
 663                                  REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
 664                 bch_bio_map(bio, w->data);
 665
 666                 trace_bcache_journal_write(bio);
 667                 bio_list_add(&list, bio);
 668
 669                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 670
 671                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 672         }
 673
 674         atomic_dec_bug(&fifo_back(&c->journal.pin));
 675         bch_journal_next(&c->journal);
 676         journal_reclaim(c);
 677
 678         spin_unlock(&c->journal.lock);
 679
 680         while ((bio = bio_list_pop(&list)))
 681                 closure_bio_submit(c, bio, cl);
 682
 683         continue_at(cl, journal_write_done, NULL);
 684 }
 685
 686 static void journal_write(struct closure *cl)
 687 {
 688         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 689
 690         spin_lock(&c->journal.lock);
 691         journal_write_unlocked(cl);
 692 }
 693
 694 static void journal_try_write(struct cache_set *c)
 695         __releases(c->journal.lock)
 696 {
 697         struct closure *cl = &c->journal.io;
 698         struct journal_write *w = c->journal.cur;
 699
 700         w->need_write = true;
 701
 702         if (!c->journal.io_in_flight) {
 703                 c->journal.io_in_flight = 1;
 704                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 705         } else {
 706                 spin_unlock(&c->journal.lock);
 707         }
 708 }
 709
 710 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 711                                                     unsigned int nkeys)
 712         __acquires(&c->journal.lock)
 713 {
 714         size_t sectors;
 715         struct closure cl;
 716         bool wait = false;
 717
 718         closure_init_stack(&cl);
 719
 720         spin_lock(&c->journal.lock);
 721
 722         while (1) {
 723                 struct journal_write *w = c->journal.cur;
 724
 725                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 726                                        block_bytes(c)) * c->sb.block_size;
 727
 728                 if (sectors <= min_t(size_t,
 729                                      c->journal.blocks_free * c->sb.block_size,
 730                                      PAGE_SECTORS << JSET_BITS))
 731                         return w;
 732
 733                 if (wait)
 734                         closure_wait(&c->journal.wait, &cl);
 735
 736                 if (!journal_full(&c->journal)) {
 737                         if (wait)
 738                                 trace_bcache_journal_entry_full(c);
 739
 740                         /*
 741                          * XXX: If we were inserting so many keys that they
 742                          * won't fit in an _empty_ journal write, we'll
 743                          * deadlock. For now, handle this in
 744                          * bch_keylist_realloc() - but something to think about.
 745                          */
 746                         BUG_ON(!w->data->keys);
 747
 748                         journal_try_write(c); /* unlocks */
 749                 } else {
 750                         if (wait)
 751                                 trace_bcache_journal_full(c);
 752
 753                         journal_reclaim(c);
 754                         spin_unlock(&c->journal.lock);
 755
 756                         btree_flush_write(c);
 757                 }
 758
 759                 closure_sync(&cl);
 760                 spin_lock(&c->journal.lock);
 761                 wait = true;
 762         }
 763 }
 764
 765 static void journal_write_work(struct work_struct *work)
 766 {
 767         struct cache_set *c = container_of(to_delayed_work(work),
 768                                            struct cache_set,
 769                                            journal.work);
 770         spin_lock(&c->journal.lock);
 771         if (c->journal.cur->dirty)
 772                 journal_try_write(c);
 773         else
 774                 spin_unlock(&c->journal.lock);
 775 }
 776
 777 /*
 778  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 779  * pass bch_journal() a list of keys to be journalled, and then
 780  * bch_journal() hands those same keys off to btree_insert_async()
 781  */
 782
 783 atomic_t *bch_journal(struct cache_set *c,
 784                       struct keylist *keys,
 785                       struct closure *parent)
 786 {
 787         struct journal_write *w;
 788         atomic_t *ret;
 789
 790         if (!CACHE_SYNC(&c->sb))
 791                 return NULL;
 792
 793         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 794
 795         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 796         w->data->keys += bch_keylist_nkeys(keys);
 797
 798         ret = &fifo_back(&c->journal.pin);
 799         atomic_inc(ret);
 800
 801         if (parent) {
 802                 closure_wait(&w->wait, parent);
 803                 journal_try_write(c);
 804         } else if (!w->dirty) {
 805                 w->dirty = true;
 806                 schedule_delayed_work(&c->journal.work,
 807                                       msecs_to_jiffies(c->journal_delay_ms));
 808                 spin_unlock(&c->journal.lock);
 809         } else {
 810                 spin_unlock(&c->journal.lock);
 811         }
 812
 813
 814         return ret;
 815 }
 816
 817 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 818 {
 819         struct keylist keys;
 820         atomic_t *ref;
 821
 822         bch_keylist_init(&keys);
 823
 824         ref = bch_journal(c, &keys, cl);
 825         if (ref)
 826                 atomic_dec_bug(ref);
 827 }
 828
 829 void bch_journal_free(struct cache_set *c)
 830 {
 831         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 832         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 833         free_fifo(&c->journal.pin);
 834         free_heap(&c->flush_btree);
 835 }
 836
 837 int bch_journal_alloc(struct cache_set *c)
 838 {
 839         struct journal *j = &c->journal;
 840
 841         spin_lock_init(&j->lock);
 842         INIT_DELAYED_WORK(&j->work, journal_write_work);
 843
 844         c->journal_delay_ms = 100;
 845
 846         j->w[0].c = c;
 847         j->w[1].c = c;
 848
 849         if (!(init_heap(&c->flush_btree, 128, GFP_KERNEL)) ||
 850             !(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 851             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
 852             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
 853                 return -ENOMEM;
 854
 855         return 0;
 856 }