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                         /*
 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", 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");
 219
 220                 for (l = find_first_zero_bit(bitmap, ca->sb.njournal_buckets);
 221                      l < ca->sb.njournal_buckets;
 222                      l = find_next_zero_bit(bitmap, ca->sb.njournal_buckets,
 223                                             l + 1))
 224                         if (read_bucket(l))
 225                                 goto bsearch;
 226
 227                 /* no journal entries on this device? */
 228                 if (l == ca->sb.njournal_buckets)
 229                         continue;
 230 bsearch:
 231                 BUG_ON(list_empty(list));
 232
 233                 /* Binary search */
 234                 m = l;
 235                 r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 236                 pr_debug("starting binary search, l %u r %u", l, r);
 237
 238                 while (l + 1 < r) {
 239                         seq = list_entry(list->prev, struct journal_replay,
 240                                          list)->j.seq;
 241
 242                         m = (l + r) >> 1;
 243                         read_bucket(m);
 244
 245                         if (seq != list_entry(list->prev, struct journal_replay,
 246                                               list)->j.seq)
 247                                 l = m;
 248                         else
 249                                 r = m;
 250                 }
 251
 252                 /*
 253                  * Read buckets in reverse order until we stop finding more
 254                  * journal entries
 255                  */
 256                 pr_debug("finishing up: m %u njournal_buckets %u",
 257                          m, ca->sb.njournal_buckets);
 258                 l = m;
 259
 260                 while (1) {
 261                         if (!l--)
 262                                 l = ca->sb.njournal_buckets - 1;
 263
 264                         if (l == m)
 265                                 break;
 266
 267                         if (test_bit(l, bitmap))
 268                                 continue;
 269
 270                         if (!read_bucket(l))
 271                                 break;
 272                 }
 273
 274                 seq = 0;
 275
 276                 for (i = 0; i < ca->sb.njournal_buckets; i++)
 277                         if (ja->seq[i] > seq) {
 278                                 seq = ja->seq[i];
 279                                 /*
 280                                  * When journal_reclaim() goes to allocate for
 281                                  * the first time, it'll use the bucket after
 282                                  * ja->cur_idx
 283                                  */
 284                                 ja->cur_idx = i;
 285                                 ja->last_idx = ja->discard_idx = (i + 1) %
 286                                         ca->sb.njournal_buckets;
 287
 288                         }
 289         }
 290
 291         if (!list_empty(list))
 292                 c->journal.seq = list_entry(list->prev,
 293                                             struct journal_replay,
 294                                             list)->j.seq;
 295
 296         return 0;
 297 #undef read_bucket
 298 }
 299
 300 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 301 {
 302         atomic_t p = { 0 };
 303         struct bkey *k;
 304         struct journal_replay *i;
 305         struct journal *j = &c->journal;
 306         uint64_t last = j->seq;
 307
 308         /*
 309          * journal.pin should never fill up - we never write a journal
 310          * entry when it would fill up. But if for some reason it does, we
 311          * iterate over the list in reverse order so that we can just skip that
 312          * refcount instead of bugging.
 313          */
 314
 315         list_for_each_entry_reverse(i, list, list) {
 316                 BUG_ON(last < i->j.seq);
 317                 i->pin = NULL;
 318
 319                 while (last-- != i->j.seq)
 320                         if (fifo_free(&j->pin) > 1) {
 321                                 fifo_push_front(&j->pin, p);
 322                                 atomic_set(&fifo_front(&j->pin), 0);
 323                         }
 324
 325                 if (fifo_free(&j->pin) > 1) {
 326                         fifo_push_front(&j->pin, p);
 327                         i->pin = &fifo_front(&j->pin);
 328                         atomic_set(i->pin, 1);
 329                 }
 330
 331                 for (k = i->j.start;
 332                      k < bset_bkey_last(&i->j);
 333                      k = bkey_next(k))
 334                         if (!__bch_extent_invalid(c, k)) {
 335                                 unsigned int j;
 336
 337                                 for (j = 0; j < KEY_PTRS(k); j++)
 338                                         if (ptr_available(c, k, j))
 339                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 340
 341                                 bch_initial_mark_key(c, 0, k);
 342                         }
 343         }
 344 }
 345
 346 static bool is_discard_enabled(struct cache_set *s)
 347 {
 348         struct cache *ca;
 349         unsigned int i;
 350
 351         for_each_cache(ca, s, i)
 352                 if (ca->discard)
 353                         return true;
 354
 355         return false;
 356 }
 357
 358 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 359 {
 360         int ret = 0, keys = 0, entries = 0;
 361         struct bkey *k;
 362         struct journal_replay *i =
 363                 list_entry(list->prev, struct journal_replay, list);
 364
 365         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 366         struct keylist keylist;
 367
 368         list_for_each_entry(i, list, list) {
 369                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 370
 371                 if (n != i->j.seq) {
 372                         if (n == start && is_discard_enabled(s))
 373                                 pr_info("bcache: journal entries %llu-%llu may be discarded! (replaying %llu-%llu)",
 374                                         n, i->j.seq - 1, start, end);
 375                         else {
 376                                 pr_err("bcache: journal entries %llu-%llu missing! (replaying %llu-%llu)",
 377                                         n, i->j.seq - 1, start, end);
 378                                 ret = -EIO;
 379                                 goto err;
 380                         }
 381                 }
 382
 383                 for (k = i->j.start;
 384                      k < bset_bkey_last(&i->j);
 385                      k = bkey_next(k)) {
 386                         trace_bcache_journal_replay_key(k);
 387
 388                         bch_keylist_init_single(&keylist, k);
 389
 390                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 391                         if (ret)
 392                                 goto err;
 393
 394                         BUG_ON(!bch_keylist_empty(&keylist));
 395                         keys++;
 396
 397                         cond_resched();
 398                 }
 399
 400                 if (i->pin)
 401                         atomic_dec(i->pin);
 402                 n = i->j.seq + 1;
 403                 entries++;
 404         }
 405
 406         pr_info("journal replay done, %i keys in %i entries, seq %llu",
 407                 keys, entries, end);
 408 err:
 409         while (!list_empty(list)) {
 410                 i = list_first_entry(list, struct journal_replay, list);
 411                 list_del(&i->list);
 412                 kfree(i);
 413         }
 414
 415         return ret;
 416 }
 417
 418 /* Journalling */
 419
 420 static void btree_flush_write(struct cache_set *c)
 421 {
 422         struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR];
 423         unsigned int i, nr;
 424         int ref_nr;
 425         atomic_t *fifo_front_p, *now_fifo_front_p;
 426         size_t mask;
 427
 428         if (c->journal.btree_flushing)
 429                 return;
 430
 431         spin_lock(&c->journal.flush_write_lock);
 432         if (c->journal.btree_flushing) {
 433                 spin_unlock(&c->journal.flush_write_lock);
 434                 return;
 435         }
 436         c->journal.btree_flushing = true;
 437         spin_unlock(&c->journal.flush_write_lock);
 438
 439         /* get the oldest journal entry and check its refcount */
 440         spin_lock(&c->journal.lock);
 441         fifo_front_p = &fifo_front(&c->journal.pin);
 442         ref_nr = atomic_read(fifo_front_p);
 443         if (ref_nr <= 0) {
 444                 /*
 445                  * do nothing if no btree node references
 446                  * the oldest journal entry
 447                  */
 448                 spin_unlock(&c->journal.lock);
 449                 goto out;
 450         }
 451         spin_unlock(&c->journal.lock);
 452
 453         mask = c->journal.pin.mask;
 454         nr = 0;
 455         atomic_long_inc(&c->flush_write);
 456         memset(btree_nodes, 0, sizeof(btree_nodes));
 457
 458         mutex_lock(&c->bucket_lock);
 459         list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
 460                 /*
 461                  * It is safe to get now_fifo_front_p without holding
 462                  * c->journal.lock here, because we don't need to know
 463                  * the exactly accurate value, just check whether the
 464                  * front pointer of c->journal.pin is changed.
 465                  */
 466                 now_fifo_front_p = &fifo_front(&c->journal.pin);
 467                 /*
 468                  * If the oldest journal entry is reclaimed and front
 469                  * pointer of c->journal.pin changes, it is unnecessary
 470                  * to scan c->btree_cache anymore, just quit the loop and
 471                  * flush out what we have already.
 472                  */
 473                 if (now_fifo_front_p != fifo_front_p)
 474                         break;
 475                 /*
 476                  * quit this loop if all matching btree nodes are
 477                  * scanned and record in btree_nodes[] already.
 478                  */
 479                 ref_nr = atomic_read(fifo_front_p);
 480                 if (nr >= ref_nr)
 481                         break;
 482
 483                 if (btree_node_journal_flush(b))
 484                         pr_err("BUG: flush_write bit should not be set here!");
 485
 486                 mutex_lock(&b->write_lock);
 487
 488                 if (!btree_node_dirty(b)) {
 489                         mutex_unlock(&b->write_lock);
 490                         continue;
 491                 }
 492
 493                 if (!btree_current_write(b)->journal) {
 494                         mutex_unlock(&b->write_lock);
 495                         continue;
 496                 }
 497
 498                 /*
 499                  * Only select the btree node which exactly references
 500                  * the oldest journal entry.
 501                  *
 502                  * If the journal entry pointed by fifo_front_p is
 503                  * reclaimed in parallel, don't worry:
 504                  * - the list_for_each_xxx loop will quit when checking
 505                  *   next now_fifo_front_p.
 506                  * - If there are matched nodes recorded in btree_nodes[],
 507                  *   they are clean now (this is why and how the oldest
 508                  *   journal entry can be reclaimed). These selected nodes
 509                  *   will be ignored and skipped in the folowing for-loop.
 510                  */
 511                 if (((btree_current_write(b)->journal - fifo_front_p) &
 512                      mask) != 0) {
 513                         mutex_unlock(&b->write_lock);
 514                         continue;
 515                 }
 516
 517                 set_btree_node_journal_flush(b);
 518
 519                 mutex_unlock(&b->write_lock);
 520
 521                 btree_nodes[nr++] = b;
 522                 /*
 523                  * To avoid holding c->bucket_lock too long time,
 524                  * only scan for BTREE_FLUSH_NR matched btree nodes
 525                  * at most. If there are more btree nodes reference
 526                  * the oldest journal entry, try to flush them next
 527                  * time when btree_flush_write() is called.
 528                  */
 529                 if (nr == BTREE_FLUSH_NR)
 530                         break;
 531         }
 532         mutex_unlock(&c->bucket_lock);
 533
 534         for (i = 0; i < nr; i++) {
 535                 b = btree_nodes[i];
 536                 if (!b) {
 537                         pr_err("BUG: btree_nodes[%d] is NULL", i);
 538                         continue;
 539                 }
 540
 541                 /* safe to check without holding b->write_lock */
 542                 if (!btree_node_journal_flush(b)) {
 543                         pr_err("BUG: bnode %p: journal_flush bit cleaned", b);
 544                         continue;
 545                 }
 546
 547                 mutex_lock(&b->write_lock);
 548                 if (!btree_current_write(b)->journal) {
 549                         clear_bit(BTREE_NODE_journal_flush, &b->flags);
 550                         mutex_unlock(&b->write_lock);
 551                         pr_debug("bnode %p: written by others", b);
 552                         continue;
 553                 }
 554
 555                 if (!btree_node_dirty(b)) {
 556                         clear_bit(BTREE_NODE_journal_flush, &b->flags);
 557                         mutex_unlock(&b->write_lock);
 558                         pr_debug("bnode %p: dirty bit cleaned by others", b);
 559                         continue;
 560                 }
 561
 562                 __bch_btree_node_write(b, NULL);
 563                 clear_bit(BTREE_NODE_journal_flush, &b->flags);
 564                 mutex_unlock(&b->write_lock);
 565         }
 566
 567 out:
 568         spin_lock(&c->journal.flush_write_lock);
 569         c->journal.btree_flushing = false;
 570         spin_unlock(&c->journal.flush_write_lock);
 571 }
 572
 573 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 574
 575 static void journal_discard_endio(struct bio *bio)
 576 {
 577         struct journal_device *ja =
 578                 container_of(bio, struct journal_device, discard_bio);
 579         struct cache *ca = container_of(ja, struct cache, journal);
 580
 581         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 582
 583         closure_wake_up(&ca->set->journal.wait);
 584         closure_put(&ca->set->cl);
 585 }
 586
 587 static void journal_discard_work(struct work_struct *work)
 588 {
 589         struct journal_device *ja =
 590                 container_of(work, struct journal_device, discard_work);
 591
 592         submit_bio(&ja->discard_bio);
 593 }
 594
 595 static void do_journal_discard(struct cache *ca)
 596 {
 597         struct journal_device *ja = &ca->journal;
 598         struct bio *bio = &ja->discard_bio;
 599
 600         if (!ca->discard) {
 601                 ja->discard_idx = ja->last_idx;
 602                 return;
 603         }
 604
 605         switch (atomic_read(&ja->discard_in_flight)) {
 606         case DISCARD_IN_FLIGHT:
 607                 return;
 608
 609         case DISCARD_DONE:
 610                 ja->discard_idx = (ja->discard_idx + 1) %
 611                         ca->sb.njournal_buckets;
 612
 613                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 614                 /* fallthrough */
 615
 616         case DISCARD_READY:
 617                 if (ja->discard_idx == ja->last_idx)
 618                         return;
 619
 620                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 621
 622                 bio_init(bio, bio->bi_inline_vecs, 1);
 623                 bio_set_op_attrs(bio, REQ_OP_DISCARD, 0);
 624                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 625                                                 ca->sb.d[ja->discard_idx]);
 626                 bio_set_dev(bio, ca->bdev);
 627                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 628                 bio->bi_end_io          = journal_discard_endio;
 629
 630                 closure_get(&ca->set->cl);
 631                 INIT_WORK(&ja->discard_work, journal_discard_work);
 632                 queue_work(bch_journal_wq, &ja->discard_work);
 633         }
 634 }
 635
 636 static void journal_reclaim(struct cache_set *c)
 637 {
 638         struct bkey *k = &c->journal.key;
 639         struct cache *ca;
 640         uint64_t last_seq;
 641         unsigned int iter, n = 0;
 642         atomic_t p __maybe_unused;
 643
 644         atomic_long_inc(&c->reclaim);
 645
 646         while (!atomic_read(&fifo_front(&c->journal.pin)))
 647                 fifo_pop(&c->journal.pin, p);
 648
 649         last_seq = last_seq(&c->journal);
 650
 651         /* Update last_idx */
 652
 653         for_each_cache(ca, c, iter) {
 654                 struct journal_device *ja = &ca->journal;
 655
 656                 while (ja->last_idx != ja->cur_idx &&
 657                        ja->seq[ja->last_idx] < last_seq)
 658                         ja->last_idx = (ja->last_idx + 1) %
 659                                 ca->sb.njournal_buckets;
 660         }
 661
 662         for_each_cache(ca, c, iter)
 663                 do_journal_discard(ca);
 664
 665         if (c->journal.blocks_free)
 666                 goto out;
 667
 668         /*
 669          * Allocate:
 670          * XXX: Sort by free journal space
 671          */
 672
 673         for_each_cache(ca, c, iter) {
 674                 struct journal_device *ja = &ca->journal;
 675                 unsigned int next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 676
 677                 /* No space available on this device */
 678                 if (next == ja->discard_idx)
 679                         continue;
 680
 681                 ja->cur_idx = next;
 682                 k->ptr[n++] = MAKE_PTR(0,
 683                                   bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 684                                   ca->sb.nr_this_dev);
 685                 atomic_long_inc(&c->reclaimed_journal_buckets);
 686         }
 687
 688         if (n) {
 689                 bkey_init(k);
 690                 SET_KEY_PTRS(k, n);
 691                 c->journal.blocks_free = c->sb.bucket_size >> c->block_bits;
 692         }
 693 out:
 694         if (!journal_full(&c->journal))
 695                 __closure_wake_up(&c->journal.wait);
 696 }
 697
 698 void bch_journal_next(struct journal *j)
 699 {
 700         atomic_t p = { 1 };
 701
 702         j->cur = (j->cur == j->w)
 703                 ? &j->w[1]
 704                 : &j->w[0];
 705
 706         /*
 707          * The fifo_push() needs to happen at the same time as j->seq is
 708          * incremented for last_seq() to be calculated correctly
 709          */
 710         BUG_ON(!fifo_push(&j->pin, p));
 711         atomic_set(&fifo_back(&j->pin), 1);
 712
 713         j->cur->data->seq       = ++j->seq;
 714         j->cur->dirty           = false;
 715         j->cur->need_write      = false;
 716         j->cur->data->keys      = 0;
 717
 718         if (fifo_full(&j->pin))
 719                 pr_debug("journal_pin full (%zu)", fifo_used(&j->pin));
 720 }
 721
 722 static void journal_write_endio(struct bio *bio)
 723 {
 724         struct journal_write *w = bio->bi_private;
 725
 726         cache_set_err_on(bio->bi_status, w->c, "journal io error");
 727         closure_put(&w->c->journal.io);
 728 }
 729
 730 static void journal_write(struct closure *cl);
 731
 732 static void journal_write_done(struct closure *cl)
 733 {
 734         struct journal *j = container_of(cl, struct journal, io);
 735         struct journal_write *w = (j->cur == j->w)
 736                 ? &j->w[1]
 737                 : &j->w[0];
 738
 739         __closure_wake_up(&w->wait);
 740         continue_at_nobarrier(cl, journal_write, bch_journal_wq);
 741 }
 742
 743 static void journal_write_unlock(struct closure *cl)
 744         __releases(&c->journal.lock)
 745 {
 746         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 747
 748         c->journal.io_in_flight = 0;
 749         spin_unlock(&c->journal.lock);
 750 }
 751
 752 static void journal_write_unlocked(struct closure *cl)
 753         __releases(c->journal.lock)
 754 {
 755         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 756         struct cache *ca;
 757         struct journal_write *w = c->journal.cur;
 758         struct bkey *k = &c->journal.key;
 759         unsigned int i, sectors = set_blocks(w->data, block_bytes(c)) *
 760                 c->sb.block_size;
 761
 762         struct bio *bio;
 763         struct bio_list list;
 764
 765         bio_list_init(&list);
 766
 767         if (!w->need_write) {
 768                 closure_return_with_destructor(cl, journal_write_unlock);
 769                 return;
 770         } else if (journal_full(&c->journal)) {
 771                 journal_reclaim(c);
 772                 spin_unlock(&c->journal.lock);
 773
 774                 btree_flush_write(c);
 775                 continue_at(cl, journal_write, bch_journal_wq);
 776                 return;
 777         }
 778
 779         c->journal.blocks_free -= set_blocks(w->data, block_bytes(c));
 780
 781         w->data->btree_level = c->root->level;
 782
 783         bkey_copy(&w->data->btree_root, &c->root->key);
 784         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 785
 786         for_each_cache(ca, c, i)
 787                 w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 788
 789         w->data->magic          = jset_magic(&c->sb);
 790         w->data->version        = BCACHE_JSET_VERSION;
 791         w->data->last_seq       = last_seq(&c->journal);
 792         w->data->csum           = csum_set(w->data);
 793
 794         for (i = 0; i < KEY_PTRS(k); i++) {
 795                 ca = PTR_CACHE(c, k, i);
 796                 bio = &ca->journal.bio;
 797
 798                 atomic_long_add(sectors, &ca->meta_sectors_written);
 799
 800                 bio_reset(bio);
 801                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 802                 bio_set_dev(bio, ca->bdev);
 803                 bio->bi_iter.bi_size = sectors << 9;
 804
 805                 bio->bi_end_io  = journal_write_endio;
 806                 bio->bi_private = w;
 807                 bio_set_op_attrs(bio, REQ_OP_WRITE,
 808                                  REQ_SYNC|REQ_META|REQ_PREFLUSH|REQ_FUA);
 809                 bch_bio_map(bio, w->data);
 810
 811                 trace_bcache_journal_write(bio, w->data->keys);
 812                 bio_list_add(&list, bio);
 813
 814                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 815
 816                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 817         }
 818
 819         /* If KEY_PTRS(k) == 0, this jset gets lost in air */
 820         BUG_ON(i == 0);
 821
 822         atomic_dec_bug(&fifo_back(&c->journal.pin));
 823         bch_journal_next(&c->journal);
 824         journal_reclaim(c);
 825
 826         spin_unlock(&c->journal.lock);
 827
 828         while ((bio = bio_list_pop(&list)))
 829                 closure_bio_submit(c, bio, cl);
 830
 831         continue_at(cl, journal_write_done, NULL);
 832 }
 833
 834 static void journal_write(struct closure *cl)
 835 {
 836         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 837
 838         spin_lock(&c->journal.lock);
 839         journal_write_unlocked(cl);
 840 }
 841
 842 static void journal_try_write(struct cache_set *c)
 843         __releases(c->journal.lock)
 844 {
 845         struct closure *cl = &c->journal.io;
 846         struct journal_write *w = c->journal.cur;
 847
 848         w->need_write = true;
 849
 850         if (!c->journal.io_in_flight) {
 851                 c->journal.io_in_flight = 1;
 852                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 853         } else {
 854                 spin_unlock(&c->journal.lock);
 855         }
 856 }
 857
 858 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 859                                                     unsigned int nkeys)
 860         __acquires(&c->journal.lock)
 861 {
 862         size_t sectors;
 863         struct closure cl;
 864         bool wait = false;
 865
 866         closure_init_stack(&cl);
 867
 868         spin_lock(&c->journal.lock);
 869
 870         while (1) {
 871                 struct journal_write *w = c->journal.cur;
 872
 873                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 874                                        block_bytes(c)) * c->sb.block_size;
 875
 876                 if (sectors <= min_t(size_t,
 877                                      c->journal.blocks_free * c->sb.block_size,
 878                                      PAGE_SECTORS << JSET_BITS))
 879                         return w;
 880
 881                 if (wait)
 882                         closure_wait(&c->journal.wait, &cl);
 883
 884                 if (!journal_full(&c->journal)) {
 885                         if (wait)
 886                                 trace_bcache_journal_entry_full(c);
 887
 888                         /*
 889                          * XXX: If we were inserting so many keys that they
 890                          * won't fit in an _empty_ journal write, we'll
 891                          * deadlock. For now, handle this in
 892                          * bch_keylist_realloc() - but something to think about.
 893                          */
 894                         BUG_ON(!w->data->keys);
 895
 896                         journal_try_write(c); /* unlocks */
 897                 } else {
 898                         if (wait)
 899                                 trace_bcache_journal_full(c);
 900
 901                         journal_reclaim(c);
 902                         spin_unlock(&c->journal.lock);
 903
 904                         btree_flush_write(c);
 905                 }
 906
 907                 closure_sync(&cl);
 908                 spin_lock(&c->journal.lock);
 909                 wait = true;
 910         }
 911 }
 912
 913 static void journal_write_work(struct work_struct *work)
 914 {
 915         struct cache_set *c = container_of(to_delayed_work(work),
 916                                            struct cache_set,
 917                                            journal.work);
 918         spin_lock(&c->journal.lock);
 919         if (c->journal.cur->dirty)
 920                 journal_try_write(c);
 921         else
 922                 spin_unlock(&c->journal.lock);
 923 }
 924
 925 /*
 926  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 927  * pass bch_journal() a list of keys to be journalled, and then
 928  * bch_journal() hands those same keys off to btree_insert_async()
 929  */
 930
 931 atomic_t *bch_journal(struct cache_set *c,
 932                       struct keylist *keys,
 933                       struct closure *parent)
 934 {
 935         struct journal_write *w;
 936         atomic_t *ret;
 937
 938         /* No journaling if CACHE_SET_IO_DISABLE set already */
 939         if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
 940                 return NULL;
 941
 942         if (!CACHE_SYNC(&c->sb))
 943                 return NULL;
 944
 945         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 946
 947         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 948         w->data->keys += bch_keylist_nkeys(keys);
 949
 950         ret = &fifo_back(&c->journal.pin);
 951         atomic_inc(ret);
 952
 953         if (parent) {
 954                 closure_wait(&w->wait, parent);
 955                 journal_try_write(c);
 956         } else if (!w->dirty) {
 957                 w->dirty = true;
 958                 schedule_delayed_work(&c->journal.work,
 959                                       msecs_to_jiffies(c->journal_delay_ms));
 960                 spin_unlock(&c->journal.lock);
 961         } else {
 962                 spin_unlock(&c->journal.lock);
 963         }
 964
 965
 966         return ret;
 967 }
 968
 969 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 970 {
 971         struct keylist keys;
 972         atomic_t *ref;
 973
 974         bch_keylist_init(&keys);
 975
 976         ref = bch_journal(c, &keys, cl);
 977         if (ref)
 978                 atomic_dec_bug(ref);
 979 }
 980
 981 void bch_journal_free(struct cache_set *c)
 982 {
 983         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 984         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 985         free_fifo(&c->journal.pin);
 986 }
 987
 988 int bch_journal_alloc(struct cache_set *c)
 989 {
 990         struct journal *j = &c->journal;
 991
 992         spin_lock_init(&j->lock);
 993         spin_lock_init(&j->flush_write_lock);
 994         INIT_DELAYED_WORK(&j->work, journal_write_work);
 995
 996         c->journal_delay_ms = 100;
 997
 998         j->w[0].c = c;
 999         j->w[1].c = c;
1000
1001         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
1002             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)) ||
1003             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL, JSET_BITS)))
1004                 return -ENOMEM;
1005
1006         return 0;
1007 }