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\n", 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\n", 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\n",
  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\n",
  97                                         bucket_index, bytes, offset);
  98                                 return ret;
  99                         }
 100
 101                         blocks = set_blocks(j, block_bytes(ca->set));
 102
 103                         /*
 104                          * Nodes in 'list' are in linear increasing order of
 105                          * i->j.seq, the node on head has the smallest (oldest)
 106                          * journal seq, the node on tail has the biggest
 107                          * (latest) journal seq.
 108                          */
 109
 110                         /*
 111                          * Check from the oldest jset for last_seq. If
 112                          * i->j.seq < j->last_seq, it means the oldest jset
 113                          * in list is expired and useless, remove it from
 114                          * this list. Otherwise, j is a condidate jset for
 115                          * further following checks.
 116                          */
 117                         while (!list_empty(list)) {
 118                                 i = list_first_entry(list,
 119                                         struct journal_replay, list);
 120                                 if (i->j.seq >= j->last_seq)
 121                                         break;
 122                                 list_del(&i->list);
 123                                 kfree(i);
 124                         }
 125
 126                         /* iterate list in reverse order (from latest jset) */
 127                         list_for_each_entry_reverse(i, list, list) {
 128                                 if (j->seq == i->j.seq)
 129                                         goto next_set;
 130
 131                                 /*
 132                                  * if j->seq is less than any i->j.last_seq
 133                                  * in list, j is an expired and useless jset.
 134                                  */
 135                                 if (j->seq < i->j.last_seq)
 136                                         goto next_set;
 137
 138                                 /*
 139                                  * 'where' points to first jset in list which
 140                                  * is elder then j.
 141                                  */
 142                                 if (j->seq > i->j.seq) {
 143                                         where = &i->list;
 144                                         goto add;
 145                                 }
 146                         }
 147
 148                         where = list;
 149 add:
 150                         i = kmalloc(offsetof(struct journal_replay, j) +
 151                                     bytes, GFP_KERNEL);
 152                         if (!i)
 153                                 return -ENOMEM;
 154                         memcpy(&i->j, j, bytes);
 155                         /* Add to the location after 'where' points to */
 156                         list_add(&i->list, where);
 157                         ret = 1;
 158
 159                         if (j->seq > ja->seq[bucket_index])
 160                                 ja->seq[bucket_index] = j->seq;
 161 next_set:
 162                         offset  += blocks * ca->sb.block_size;
 163                         len     -= blocks * ca->sb.block_size;
 164                         j = ((void *) j) + blocks * block_bytes(ca);
 165                 }
 166         }
 167
 168         return ret;
 169 }
 170
 171 int bch_journal_read(struct cache_set *c, struct list_head *list)
 172 {
 173 #define read_bucket(b)                                                  \
 174         ({                                                              \
 175                 ret = journal_read_bucket(ca, list, b);                 \
 176                 __set_bit(b, bitmap);                                   \
 177                 if (ret < 0)                                            \
 178                         return ret;                                     \
 179                 ret;                                                    \
 180         })
 181
 182         struct cache *ca;
 183         unsigned int iter;
 184         int ret = 0;
 185
 186         for_each_cache(ca, c, iter) {
 187                 struct journal_device *ja = &ca->journal;
 188                 DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
 189                 unsigned int i, l, r, m;
 190                 uint64_t seq;
 191
 192                 bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
 193                 pr_debug("%u journal buckets\n", ca->sb.njournal_buckets);
 194
 195                 /*
 196                  * Read journal buckets ordered by golden ratio hash to quickly
 197                  * find a sequence of buckets with valid journal entries
 198                  */
 199                 for (i = 0; i < ca->sb.njournal_buckets; i++) {
 200                         /*
 201                          * We must try the index l with ZERO first for
 202                          * correctness due to the scenario that the journal
 203                          * bucket is circular buffer which might have wrapped
 204                          */
 205                         l = (i * 2654435769U) % ca->sb.njournal_buckets;
 206
 207                         if (test_bit(l, bitmap))
 208                                 break;
 209
 210                         if (read_bucket(l))
 211                                 goto bsearch;
 212                 }
 213
 214                 /*
 215                  * If that fails, check all the buckets we haven't checked
 216                  * already
 217                  */
 218                 pr_debug("falling back to linear search\n");
 219
 220                 for_each_clear_bit(l, bitmap, ca->sb.njournal_buckets)
 221                         if (read_bucket(l))
 222                                 goto bsearch;
 223
 224                 /* no journal entries on this device? */
 225                 if (l == ca->sb.njournal_buckets)
 226                         continue;
 227 bsearch:
 228                 BUG_ON(list_empty(list));
 229
 230                 /* Binary search */
 231                 m = l;
 232                 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 233                 pr_debug("starting binary search, l %u r %u\n", l, r);
 234
 235                 while (l + 1 < r) {
 236                         seq = list_entry(list->prev, struct journal_replay,
 237                                          list)->j.seq;
 238
 239                         m = (l + r) >> 1;
 240                         read_bucket(m);
 241
 242                         if (seq != list_entry(list->prev, struct journal_replay,
 243                                               list)->j.seq)
 244                                 l = m;
 245                         else
 246                                 r = m;
 247                 }
 248
 249                 /*
 250                  * Read buckets in reverse order until we stop finding more
 251                  * journal entries
 252                  */
 253                 pr_debug("finishing up: m %u njournal_buckets %u\n",
 254                          m, ca->sb.njournal_buckets);
 255                 l = m;
 256
 257                 while (1) {
 258                         if (!l--)
 259                                 l = ca->sb.njournal_buckets - 1;
 260
 261                         if (l == m)
 262                                 break;
 263
 264                         if (test_bit(l, bitmap))
 265                                 continue;
 266
 267                         if (!read_bucket(l))
 268                                 break;
 269                 }
 270
 271                 seq = 0;
 272
 273                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 274                         if (ja->seq[i] > seq) {
 275                                 seq = ja->seq[i];
 276                                 /*
 277                                  * When journal_reclaim() goes to allocate for
 278                                  * the first time, it'll use the bucket after
 279                                  * ja->cur_idx
 280                                  */
 281                                 ja->cur_idx = i;
 282                                 ja->last_idx = ja->discard_idx = (i + 1) %
 283                                         ca->sb.njournal_buckets;
 284
 285                         }
 286         }
 287
 288         if (!list_empty(list))
 289                 c->journal.seq = list_entry(list->prev,
 290                                             struct journal_replay,
 291                                             list)->j.seq;
 292
 293         return 0;
 294 #undef read_bucket
 295 }
 296
 297 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 298 {
 299         atomic_t p = { 0 };
 300         struct bkey *k;
 301         struct journal_replay *i;
 302         struct journal *j = &c->journal;
 303         uint64_t last = j->seq;
 304
 305         /*
 306          * journal.pin should never fill up - we never write a journal
 307          * entry when it would fill up. But if for some reason it does, we
 308          * iterate over the list in reverse order so that we can just skip that
 309          * refcount instead of bugging.
 310          */
 311
 312         list_for_each_entry_reverse(i, list, list) {
 313                 BUG_ON(last < i->j.seq);
 314                 i->pin = NULL;
 315
 316                 while (last-- != i->j.seq)
 317                         if (fifo_free(&j->pin) > 1) {
 318                                 fifo_push_front(&j->pin, p);
 319                                 atomic_set(&fifo_front(&j->pin), 0);
 320                         }
 321
 322                 if (fifo_free(&j->pin) > 1) {
 323                         fifo_push_front(&j->pin, p);
 324                         i->pin = &fifo_front(&j->pin);
 325                         atomic_set(i->pin, 1);
 326                 }
 327
 328                 for (k = i->j.start;
 329                      k < bset_bkey_last(&i->j);
 330                      k = bkey_next(k))
 331                         if (!__bch_extent_invalid(c, k)) {
 332                                 unsigned int j;
 333
 334                                 for (j = 0; j < KEY_PTRS(k); j++)
 335                                         if (ptr_available(c, k, j))
 336                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 337
 338                                 bch_initial_mark_key(c, 0, k);
 339                         }
 340         }
 341 }
 342
 343 static bool is_discard_enabled(struct cache_set *s)
 344 {
 345         struct cache *ca;
 346         unsigned int i;
 347
 348         for_each_cache(ca, s, i)
 349                 if (ca->discard)
 350                         return true;
 351
 352         return false;
 353 }
 354
 355 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 356 {
 357         int ret = 0, keys = 0, entries = 0;
 358         struct bkey *k;
 359         struct journal_replay *i =
 360                 list_entry(list->prev, struct journal_replay, list);
 361
 362         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 363         struct keylist keylist;
 364
 365         list_for_each_entry(i, list, list) {
 366                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 367
 368                 if (n != i->j.seq) {
 369                         if (n == start && is_discard_enabled(s))
 370                                 pr_info("journal entries %llu-%llu may be discarded! (replaying %llu-%llu)\n",
 371                                         n, i->j.seq - 1, start, end);
 372                         else {
 373                                 pr_err("journal entries %llu-%llu missing! (replaying %llu-%llu)\n",
 374                                         n, i->j.seq - 1, start, end);
 375                                 ret = -EIO;
 376                                 goto err;
 377                         }
 378                 }
 379
 380                 for (k = i->j.start;
 381                      k < bset_bkey_last(&i->j);
 382                      k = bkey_next(k)) {
 383                         trace_bcache_journal_replay_key(k);
 384
 385                         bch_keylist_init_single(&keylist, k);
 386
 387                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 388                         if (ret)
 389                                 goto err;
 390
 391                         BUG_ON(!bch_keylist_empty(&keylist));
 392                         keys++;
 393
 394                         cond_resched();
 395                 }
 396
 397                 if (i->pin)
 398                         atomic_dec(i->pin);
 399                 n = i->j.seq + 1;
 400                 entries++;
 401         }
 402
 403         pr_info("journal replay done, %i keys in %i entries, seq %llu\n",
 404                 keys, entries, end);
 405 err:
 406         while (!list_empty(list)) {
 407                 i = list_first_entry(list, struct journal_replay, list);
 408                 list_del(&i->list);
 409                 kfree(i);
 410         }
 411
 412         return ret;
 413 }
 414
 415 /* Journalling */
 416
 417 static void btree_flush_write(struct cache_set *c)
 418 {
 419         struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR];
 420         unsigned int i, nr;
 421         int ref_nr;
 422         atomic_t *fifo_front_p, *now_fifo_front_p;
 423         size_t mask;
 424
 425         if (c->journal.btree_flushing)
 426                 return;
 427
 428         spin_lock(&c->journal.flush_write_lock);
 429         if (c->journal.btree_flushing) {
 430                 spin_unlock(&c->journal.flush_write_lock);
 431                 return;
 432         }
 433         c->journal.btree_flushing = true;
 434         spin_unlock(&c->journal.flush_write_lock);
 435
 436         /* get the oldest journal entry and check its refcount */
 437         spin_lock(&c->journal.lock);
 438         fifo_front_p = &fifo_front(&c->journal.pin);
 439         ref_nr = atomic_read(fifo_front_p);
 440         if (ref_nr <= 0) {
 441                 /*
 442                  * do nothing if no btree node references
 443                  * the oldest journal entry
 444                  */
 445                 spin_unlock(&c->journal.lock);
 446                 goto out;
 447         }
 448         spin_unlock(&c->journal.lock);
 449
 450         mask = c->journal.pin.mask;
 451         nr = 0;
 452         atomic_long_inc(&c->flush_write);
 453         memset(btree_nodes, 0, sizeof(btree_nodes));
 454
 455         mutex_lock(&c->bucket_lock);
 456         list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
 457                 /*
 458                  * It is safe to get now_fifo_front_p without holding
 459                  * c->journal.lock here, because we don't need to know
 460                  * the exactly accurate value, just check whether the
 461                  * front pointer of c->journal.pin is changed.
 462                  */
 463                 now_fifo_front_p = &fifo_front(&c->journal.pin);
 464                 /*
 465                  * If the oldest journal entry is reclaimed and front
 466                  * pointer of c->journal.pin changes, it is unnecessary
 467                  * to scan c->btree_cache anymore, just quit the loop and
 468                  * flush out what we have already.
 469                  */
 470                 if (now_fifo_front_p != fifo_front_p)
 471                         break;
 472                 /*
 473                  * quit this loop if all matching btree nodes are
 474                  * scanned and record in btree_nodes[] already.
 475                  */
 476                 ref_nr = atomic_read(fifo_front_p);
 477                 if (nr >= ref_nr)
 478                         break;
 479
 480                 if (btree_node_journal_flush(b))
 481                         pr_err("BUG: flush_write bit should not be set here!\n");
 482
 483                 mutex_lock(&b->write_lock);
 484
 485                 if (!btree_node_dirty(b)) {
 486                         mutex_unlock(&b->write_lock);
 487                         continue;
 488                 }
 489
 490                 if (!btree_current_write(b)->journal) {
 491                         mutex_unlock(&b->write_lock);
 492                         continue;
 493                 }
 494
 495                 /*
 496                  * Only select the btree node which exactly references
 497                  * the oldest journal entry.
 498                  *
 499                  * If the journal entry pointed by fifo_front_p is
 500                  * reclaimed in parallel, don't worry:
 501                  * - the list_for_each_xxx loop will quit when checking
 502                  *   next now_fifo_front_p.
 503                  * - If there are matched nodes recorded in btree_nodes[],
 504                  *   they are clean now (this is why and how the oldest
 505                  *   journal entry can be reclaimed). These selected nodes
 506                  *   will be ignored and skipped in the folowing for-loop.
 507                  */
 508                 if (((btree_current_write(b)->journal - fifo_front_p) &
 509                      mask) != 0) {
 510                         mutex_unlock(&b->write_lock);
 511                         continue;
 512                 }
 513
 514                 set_btree_node_journal_flush(b);
 515
 516                 mutex_unlock(&b->write_lock);
 517
 518                 btree_nodes[nr++] = b;
 519                 /*
 520                  * To avoid holding c->bucket_lock too long time,
 521                  * only scan for BTREE_FLUSH_NR matched btree nodes
 522                  * at most. If there are more btree nodes reference
 523                  * the oldest journal entry, try to flush them next
 524                  * time when btree_flush_write() is called.
 525                  */
 526                 if (nr == BTREE_FLUSH_NR)
 527                         break;
 528         }
 529         mutex_unlock(&c->bucket_lock);
 530
 531         for (i = 0; i < nr; i++) {
 532                 b = btree_nodes[i];
 533                 if (!b) {
 534                         pr_err("BUG: btree_nodes[%d] is NULL\n", i);
 535                         continue;
 536                 }
 537
 538                 /* safe to check without holding b->write_lock */
 539                 if (!btree_node_journal_flush(b)) {
 540                         pr_err("BUG: bnode %p: journal_flush bit cleaned\n", b);
 541                         continue;
 542                 }
 543
 544                 mutex_lock(&b->write_lock);
 545                 if (!btree_current_write(b)->journal) {
 546                         clear_bit(BTREE_NODE_journal_flush, &b->flags);
 547                         mutex_unlock(&b->write_lock);
 548                         pr_debug("bnode %p: written by others\n", b);
 549                         continue;
 550                 }
 551
 552                 if (!btree_node_dirty(b)) {
 553                         clear_bit(BTREE_NODE_journal_flush, &b->flags);
 554                         mutex_unlock(&b->write_lock);
 555                         pr_debug("bnode %p: dirty bit cleaned by others\n", b);
 556                         continue;
 557                 }
 558
 559                 __bch_btree_node_write(b, NULL);
 560                 clear_bit(BTREE_NODE_journal_flush, &b->flags);
 561                 mutex_unlock(&b->write_lock);
 562         }
 563
 564 out:
 565         spin_lock(&c->journal.flush_write_lock);
 566         c->journal.btree_flushing = false;
 567         spin_unlock(&c->journal.flush_write_lock);
 568 }
 569
 570 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 571
 572 static void journal_discard_endio(struct bio *bio)
 573 {
 574         struct journal_device *ja =
 575                 container_of(bio, struct journal_device, discard_bio);
 576         struct cache *ca = container_of(ja, struct cache, journal);
 577
 578         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 579
 580         closure_wake_up(&ca->set->journal.wait);
 581         closure_put(&ca->set->cl);
 582 }
 583
 584 static void journal_discard_work(struct work_struct *work)
 585 {
 586         struct journal_device *ja =
 587                 container_of(work, struct journal_device, discard_work);
 588
 589         submit_bio(&ja->discard_bio);
 590 }
 591
 592 static void do_journal_discard(struct cache *ca)
 593 {
 594         struct journal_device *ja = &ca->journal;
 595         struct bio *bio = &ja->discard_bio;
 596
 597         if (!ca->discard) {
 598                 ja->discard_idx = ja->last_idx;
 599                 return;
 600         }
 601
 602         switch (atomic_read(&ja->discard_in_flight)) {
 603         case DISCARD_IN_FLIGHT:
 604                 return;
 605
 606         case DISCARD_DONE:
 607                 ja->discard_idx = (ja->discard_idx + 1) %
 608                         ca->sb.njournal_buckets;
 609
 610                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 611                 fallthrough;
 612
 613         case DISCARD_READY:
 614                 if (ja->discard_idx == ja->last_idx)
 615                         return;
 616
 617                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 618
 619                 bio_init(bio, bio->bi_inline_vecs, 1);
 620                 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
 621                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 622                                                 ca->sb.d[ja->discard_idx]);
 623                 bio_set_dev(bio, ca->bdev);
 624                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 625                 bio->bi_end_io          = journal_discard_endio;
 626
 627                 closure_get(&ca->set->cl);
 628                 INIT_WORK(&ja->discard_work, journal_discard_work);
 629                 queue_work(bch_journal_wq, &ja->discard_work);
 630         }
 631 }
 632
 633 static void journal_reclaim(struct cache_set *c)
 634 {
 635         struct bkey *k = &c->journal.key;
 636         struct cache *ca;
 637         uint64_t last_seq;
 638         unsigned int iter, n = 0;
 639         atomic_t p __maybe_unused;
 640
 641         atomic_long_inc(&c->reclaim);
 642
 643         while (!atomic_read(&fifo_front(&c->journal.pin)))
 644                 fifo_pop(&c->journal.pin, p);
 645
 646         last_seq = last_seq(&c->journal);
 647
 648         /* Update last_idx */
 649
 650         for_each_cache(ca, c, iter) {
 651                 struct journal_device *ja = &ca->journal;
 652
 653                 while (ja->last_idx != ja->cur_idx &&
 654                        ja->seq[ja->last_idx] < last_seq)
 655                         ja->last_idx = (ja->last_idx + 1) %
 656                                 ca->sb.njournal_buckets;
 657         }
 658
 659         for_each_cache(ca, c, iter)
 660                 do_journal_discard(ca);
 661
 662         if (c->journal.blocks_free)
 663                 goto out;
 664
 665         /*
 666          * Allocate:
 667          * XXX: Sort by free journal space
 668          */
 669
 670         for_each_cache(ca, c, iter) {
 671                 struct journal_device *ja = &ca->journal;
 672                 unsigned int next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 673
 674                 /* No space available on this device */
 675                 if (next == ja->discard_idx)
 676                         continue;
 677
 678                 ja->cur_idx = next;
 679                 k->ptr[n++] = MAKE_PTR(0,
 680                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 681                                   ca->sb.nr_this_dev);
 682                 atomic_long_inc(&c->reclaimed_journal_buckets);
 683         }
 684
 685         if (n) {
 686                 bkey_init(k);
 687                 SET_KEY_PTRS(k, n);
 688                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 689         }
 690 out:
 691         if (!journal_full(&c->journal))
 692                 __closure_wake_up(&c->journal.wait);
 693 }
 694
 695 void bch_journal_next(struct journal *j)
 696 {
 697         atomic_t p = { 1 };
 698
 699         j->cur = (j->cur == j->w)
 700                 ? &j->w[1]
 701                 : &j->w[0];
 702
 703         /*
 704          * The fifo_push() needs to happen at the same time as j->seq is
 705          * incremented for last_seq() to be calculated correctly
 706          */
 707         BUG_ON(!fifo_push(&j->pin, p));
 708         atomic_set(&fifo_back(&j->pin), 1);
 709
 710         j->cur->data->seq       = ++j->seq;
 711         j->cur->dirty           = false;
 712         j->cur->need_write      = false;
 713         j->cur->data->keys      = 0;
 714
 715         if (fifo_full(&j->pin))
 716                 pr_debug("journal_pin full (%zu)\n", fifo_used(&j->pin));
 717 }
 718
 719 static void journal_write_endio(struct bio *bio)
 720 {
 721         struct journal_write *w = bio->bi_private;
 722
 723         cache_set_err_on(bio->bi_status, w->c, "journal io error");
 724         closure_put(&w->c->journal.io);
 725 }
 726
 727 static void journal_write(struct closure *cl);
 728
 729 static void journal_write_done(struct closure *cl)
 730 {
 731         struct journal *j = container_of(cl, struct journal, io);
 732         struct journal_write *w = (j->cur == j->w)
 733                 ? &j->w[1]
 734                 : &j->w[0];
 735
 736         __closure_wake_up(&w->wait);
 737         continue_at_nobarrier(cl, journal_write, bch_journal_wq);
 738 }
 739
 740 static void journal_write_unlock(struct closure *cl)
 741         __releases(&c->journal.lock)
 742 {
 743         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 744
 745         c->journal.io_in_flight = 0;
 746         spin_unlock(&c->journal.lock);
 747 }
 748
 749 static void journal_write_unlocked(struct closure *cl)
 750         __releases(c->journal.lock)
 751 {
 752         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 753         struct cache *ca;
 754         struct journal_write *w = c->journal.cur;
 755         struct bkey *k = &c->journal.key;
 756         unsigned int i, sectors = set_blocks(w->data, block_bytes(c)) *
 757                 c->sb.block_size;
 758
 759         struct bio *bio;
 760         struct bio_list list;
 761
 762         bio_list_init(&list);
 763
 764         if (!w->need_write) {
 765                 closure_return_with_destructor(cl, journal_write_unlock);
 766                 return;
 767         } else if (journal_full(&c->journal)) {
 768                 journal_reclaim(c);
 769                 spin_unlock(&c->journal.lock);
 770
 771                 btree_flush_write(c);
 772                 continue_at(cl, journal_write, bch_journal_wq);
 773                 return;
 774         }
 775
 776         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 777
 778         w->data->btree_level = c->root->level;
 779
 780         bkey_copy(&w->data->btree_root, &c->root->key);
 781         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 782
 783         for_each_cache(ca, c, i)
 784                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 785
 786         w->data->magic          = jset_magic(&c->sb);
 787         w->data->version        = BCACHE_JSET_VERSION;
 788         w->data->last_seq       = last_seq(&c->journal);
 789         w->data->csum           = csum_set(w->data);
 790
 791         for (i = 0; i < KEY_PTRS(k); i++) {
 792                 ca = PTR_CACHE(c, k, i);
 793                 bio = &ca->journal.bio;
 794
 795                 atomic_long_add(sectors, &ca->meta_sectors_written);
 796
 797                 bio_reset(bio);
 798                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 799                 bio_set_dev(bio, ca->bdev);
 800                 bio->bi_iter.bi_size = sectors << 9;
 801
 802                 bio->bi_end_io  = journal_write_endio;
 803                 bio->bi_private = w;
 804                 bio_set_op_attrs(bio, REQ_OP_WRITE,
 805                                  REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
 806                 bch_bio_map(bio, w->data);
 807
 808                 trace_bcache_journal_write(bio, w->data->keys);
 809                 bio_list_add(&list, bio);
 810
 811                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 812
 813                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 814         }
 815
 816         /* If KEY_PTRS(k) == 0, this jset gets lost in air */
 817         BUG_ON(i == 0);
 818
 819         atomic_dec_bug(&fifo_back(&c->journal.pin));
 820         bch_journal_next(&c->journal);
 821         journal_reclaim(c);
 822
 823         spin_unlock(&c->journal.lock);
 824
 825         while ((bio = bio_list_pop(&list)))
 826                 closure_bio_submit(c, bio, cl);
 827
 828         continue_at(cl, journal_write_done, NULL);
 829 }
 830
 831 static void journal_write(struct closure *cl)
 832 {
 833         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 834
 835         spin_lock(&c->journal.lock);
 836         journal_write_unlocked(cl);
 837 }
 838
 839 static void journal_try_write(struct cache_set *c)
 840         __releases(c->journal.lock)
 841 {
 842         struct closure *cl = &c->journal.io;
 843         struct journal_write *w = c->journal.cur;
 844
 845         w->need_write = true;
 846
 847         if (!c->journal.io_in_flight) {
 848                 c->journal.io_in_flight = 1;
 849                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 850         } else {
 851                 spin_unlock(&c->journal.lock);
 852         }
 853 }
 854
 855 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 856                                                     unsigned int nkeys)
 857         __acquires(&c->journal.lock)
 858 {
 859         size_t sectors;
 860         struct closure cl;
 861         bool wait = false;
 862
 863         closure_init_stack(&cl);
 864
 865         spin_lock(&c->journal.lock);
 866
 867         while (1) {
 868                 struct journal_write *w = c->journal.cur;
 869
 870                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 871                                        block_bytes(c)) * c->sb.block_size;
 872
 873                 if (sectors <= min_t(size_t,
 874                                      c->journal.blocks_free * c->sb.block_size,
 875                                      PAGE_SECTORS << JSET_BITS))
 876                         return w;
 877
 878                 if (wait)
 879                         closure_wait(&c->journal.wait, &cl);
 880
 881                 if (!journal_full(&c->journal)) {
 882                         if (wait)
 883                                 trace_bcache_journal_entry_full(c);
 884
 885                         /*
 886                          * XXX: If we were inserting so many keys that they
 887                          * won't fit in an _empty_ journal write, we'll
 888                          * deadlock. For now, handle this in
 889                          * bch_keylist_realloc() - but something to think about.
 890                          */
 891                         BUG_ON(!w->data->keys);
 892
 893                         journal_try_write(c); /* unlocks */
 894                 } else {
 895                         if (wait)
 896                                 trace_bcache_journal_full(c);
 897
 898                         journal_reclaim(c);
 899                         spin_unlock(&c->journal.lock);
 900
 901                         btree_flush_write(c);
 902                 }
 903
 904                 closure_sync(&cl);
 905                 spin_lock(&c->journal.lock);
 906                 wait = true;
 907         }
 908 }
 909
 910 static void journal_write_work(struct work_struct *work)
 911 {
 912         struct cache_set *c = container_of(to_delayed_work(work),
 913                                            struct cache_set,
 914                                            journal.work);
 915         spin_lock(&c->journal.lock);
 916         if (c->journal.cur->dirty)
 917                 journal_try_write(c);
 918         else
 919                 spin_unlock(&c->journal.lock);
 920 }
 921
 922 /*
 923  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 924  * pass bch_journal() a list of keys to be journalled, and then
 925  * bch_journal() hands those same keys off to btree_insert_async()
 926  */
 927
 928 atomic_t *bch_journal(struct cache_set *c,
 929                       struct keylist *keys,
 930                       struct closure *parent)
 931 {
 932         struct journal_write *w;
 933         atomic_t *ret;
 934
 935         /* No journaling if CACHE_SET_IO_DISABLE set already */
 936         if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
 937                 return NULL;
 938
 939         if (!CACHE_SYNC(&c->sb))
 940                 return NULL;
 941
 942         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 943
 944         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 945         w->data->keys += bch_keylist_nkeys(keys);
 946
 947         ret = &fifo_back(&c->journal.pin);
 948         atomic_inc(ret);
 949
 950         if (parent) {
 951                 closure_wait(&w->wait, parent);
 952                 journal_try_write(c);
 953         } else if (!w->dirty) {
 954                 w->dirty = true;
 955                 schedule_delayed_work(&c->journal.work,
 956                                       msecs_to_jiffies(c->journal_delay_ms));
 957                 spin_unlock(&c->journal.lock);
 958         } else {
 959                 spin_unlock(&c->journal.lock);
 960         }
 961
 962
 963         return ret;
 964 }
 965
 966 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 967 {
 968         struct keylist keys;
 969         atomic_t *ref;
 970
 971         bch_keylist_init(&keys);
 972
 973         ref = bch_journal(c, &keys, cl);
 974         if (ref)
 975                 atomic_dec_bug(ref);
 976 }
 977
 978 void bch_journal_free(struct cache_set *c)
 979 {
 980         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 981         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 982         free_fifo(&c->journal.pin);
 983 }
 984
 985 int bch_journal_alloc(struct cache_set *c)
 986 {
 987         struct journal *j = &c->journal;
 988
 989         spin_lock_init(&j->lock);
 990         spin_lock_init(&j->flush_write_lock);
 991         INIT_DELAYED_WORK(&j->work, journal_write_work);
 992
 993         c->journal_delay_ms = 100;
 994
 995         j->w[0].c = c;
 996         j->w[1].c = c;
 997
 998         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 999             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) ||
1000             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)))
1001                 return -ENOMEM;
1002
1003         return 0;
1004 }