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, ca->bdev, REQ_OP_READ);
  57                 bio->bi_iter.bi_sector  = bucket + offset;
  58                 bio->bi_iter.bi_size    = len << 9;
  59
  60                 bio->bi_end_io  = journal_read_endio;
  61                 bio->bi_private = &cl;
  62                 bch_bio_map(bio, data);
  63
  64                 closure_bio_submit(ca->set, bio, &cl);
  65                 closure_sync(&cl);
  66
  67                 /* This function could be simpler now since we no longer write
  68                  * journal entries that overlap bucket boundaries; this means
  69                  * the start of a bucket will always have a valid journal entry
  70                  * if it has any journal entries at all.
  71                  */
  72
  73                 j = data;
  74                 while (len) {
  75                         struct list_head *where;
  76                         size_t blocks, bytes = set_bytes(j);
  77
  78                         if (j->magic != jset_magic(&ca->sb)) {
  79                                 pr_debug("%u: bad magic\n", bucket_index);
  80                                 return ret;
  81                         }
  82
  83                         if (bytes > left << 9 ||
  84                             bytes > PAGE_SIZE << JSET_BITS) {
  85                                 pr_info("%u: too big, %zu bytes, offset %u\n",
  86                                         bucket_index, bytes, offset);
  87                                 return ret;
  88                         }
  89
  90                         if (bytes > len << 9)
  91                                 goto reread;
  92
  93                         if (j->csum != csum_set(j)) {
  94                                 pr_info("%u: bad csum, %zu bytes, offset %u\n",
  95                                         bucket_index, bytes, offset);
  96                                 return ret;
  97                         }
  98
  99                         blocks = set_blocks(j, block_bytes(ca));
 100
 101                         /*
 102                          * Nodes in 'list' are in linear increasing order of
 103                          * i->j.seq, the node on head has the smallest (oldest)
 104                          * journal seq, the node on tail has the biggest
 105                          * (latest) journal seq.
 106                          */
 107
 108                         /*
 109                          * Check from the oldest jset for last_seq. If
 110                          * i->j.seq < j->last_seq, it means the oldest jset
 111                          * in list is expired and useless, remove it from
 112                          * this list. Otherwise, j is a candidate jset for
 113                          * further following checks.
 114                          */
 115                         while (!list_empty(list)) {
 116                                 i = list_first_entry(list,
 117                                         struct journal_replay, list);
 118                                 if (i->j.seq >= j->last_seq)
 119                                         break;
 120                                 list_del(&i->list);
 121                                 kfree(i);
 122                         }
 123
 124                         /* iterate list in reverse order (from latest jset) */
 125                         list_for_each_entry_reverse(i, list, list) {
 126                                 if (j->seq == i->j.seq)
 127                                         goto next_set;
 128
 129                                 /*
 130                                  * if j->seq is less than any i->j.last_seq
 131                                  * in list, j is an expired and useless jset.
 132                                  */
 133                                 if (j->seq < i->j.last_seq)
 134                                         goto next_set;
 135
 136                                 /*
 137                                  * 'where' points to first jset in list which
 138                                  * is elder then j.
 139                                  */
 140                                 if (j->seq > i->j.seq) {
 141                                         where = &i->list;
 142                                         goto add;
 143                                 }
 144                         }
 145
 146                         where = list;
 147 add:
 148                         i = kmalloc(offsetof(struct journal_replay, j) +
 149                                     bytes, GFP_KERNEL);
 150                         if (!i)
 151                                 return -ENOMEM;
 152                         memcpy(&i->j, j, bytes);
 153                         /* Add to the location after 'where' points to */
 154                         list_add(&i->list, where);
 155                         ret = 1;
 156
 157                         if (j->seq > ja->seq[bucket_index])
 158                                 ja->seq[bucket_index] = j->seq;
 159 next_set:
 160                         offset  += blocks * ca->sb.block_size;
 161                         len     -= blocks * ca->sb.block_size;
 162                         j = ((void *) j) + blocks * block_bytes(ca);
 163                 }
 164         }
 165
 166         return ret;
 167 }
 168
 169 int bch_journal_read(struct cache_set *c, struct list_head *list)
 170 {
 171 #define read_bucket(b)                                                  \
 172         ({                                                              \
 173                 ret = journal_read_bucket(ca, list, b);                 \
 174                 __set_bit(b, bitmap);                                   \
 175                 if (ret < 0)                                            \
 176                         return ret;                                     \
 177                 ret;                                                    \
 178         })
 179
 180         struct cache *ca = c->cache;
 181         int ret = 0;
 182         struct journal_device *ja = &ca->journal;
 183         DECLARE_BITMAP(bitmap, SB_JOURNAL_BUCKETS);
 184         unsigned int i, l, r, m;
 185         uint64_t seq;
 186
 187         bitmap_zero(bitmap, SB_JOURNAL_BUCKETS);
 188         pr_debug("%u journal buckets\n", ca->sb.njournal_buckets);
 189
 190         /*
 191          * Read journal buckets ordered by golden ratio hash to quickly
 192          * find a sequence of buckets with valid journal entries
 193          */
 194         for (i = 0; i < ca->sb.njournal_buckets; i++) {
 195                 /*
 196                  * We must try the index l with ZERO first for
 197                  * correctness due to the scenario that the journal
 198                  * bucket is circular buffer which might have wrapped
 199                  */
 200                 l = (i * 2654435769U) % ca->sb.njournal_buckets;
 201
 202                 if (test_bit(l, bitmap))
 203                         break;
 204
 205                 if (read_bucket(l))
 206                         goto bsearch;
 207         }
 208
 209         /*
 210          * If that fails, check all the buckets we haven't checked
 211          * already
 212          */
 213         pr_debug("falling back to linear search\n");
 214
 215         for_each_clear_bit(l, bitmap, ca->sb.njournal_buckets)
 216                 if (read_bucket(l))
 217                         goto bsearch;
 218
 219         /* no journal entries on this device? */
 220         if (l == ca->sb.njournal_buckets)
 221                 goto out;
 222 bsearch:
 223         BUG_ON(list_empty(list));
 224
 225         /* Binary search */
 226         m = l;
 227         r = find_next_bit(bitmap, ca->sb.njournal_buckets, l + 1);
 228         pr_debug("starting binary search, l %u r %u\n", l, r);
 229
 230         while (l + 1 < r) {
 231                 seq = list_entry(list->prev, struct journal_replay,
 232                                  list)->j.seq;
 233
 234                 m = (l + r) >> 1;
 235                 read_bucket(m);
 236
 237                 if (seq != list_entry(list->prev, struct journal_replay,
 238                                       list)->j.seq)
 239                         l = m;
 240                 else
 241                         r = m;
 242         }
 243
 244         /*
 245          * Read buckets in reverse order until we stop finding more
 246          * journal entries
 247          */
 248         pr_debug("finishing up: m %u njournal_buckets %u\n",
 249                  m, ca->sb.njournal_buckets);
 250         l = m;
 251
 252         while (1) {
 253                 if (!l--)
 254                         l = ca->sb.njournal_buckets - 1;
 255
 256                 if (l == m)
 257                         break;
 258
 259                 if (test_bit(l, bitmap))
 260                         continue;
 261
 262                 if (!read_bucket(l))
 263                         break;
 264         }
 265
 266         seq = 0;
 267
 268         for (i = 0; i < ca->sb.njournal_buckets; i++)
 269                 if (ja->seq[i] > seq) {
 270                         seq = ja->seq[i];
 271                         /*
 272                          * When journal_reclaim() goes to allocate for
 273                          * the first time, it'll use the bucket after
 274                          * ja->cur_idx
 275                          */
 276                         ja->cur_idx = i;
 277                         ja->last_idx = ja->discard_idx = (i + 1) %
 278                                 ca->sb.njournal_buckets;
 279
 280                 }
 281
 282 out:
 283         if (!list_empty(list))
 284                 c->journal.seq = list_entry(list->prev,
 285                                             struct journal_replay,
 286                                             list)->j.seq;
 287
 288         return 0;
 289 #undef read_bucket
 290 }
 291
 292 void bch_journal_mark(struct cache_set *c, struct list_head *list)
 293 {
 294         atomic_t p = { 0 };
 295         struct bkey *k;
 296         struct journal_replay *i;
 297         struct journal *j = &c->journal;
 298         uint64_t last = j->seq;
 299
 300         /*
 301          * journal.pin should never fill up - we never write a journal
 302          * entry when it would fill up. But if for some reason it does, we
 303          * iterate over the list in reverse order so that we can just skip that
 304          * refcount instead of bugging.
 305          */
 306
 307         list_for_each_entry_reverse(i, list, list) {
 308                 BUG_ON(last < i->j.seq);
 309                 i->pin = NULL;
 310
 311                 while (last-- != i->j.seq)
 312                         if (fifo_free(&j->pin) > 1) {
 313                                 fifo_push_front(&j->pin, p);
 314                                 atomic_set(&fifo_front(&j->pin), 0);
 315                         }
 316
 317                 if (fifo_free(&j->pin) > 1) {
 318                         fifo_push_front(&j->pin, p);
 319                         i->pin = &fifo_front(&j->pin);
 320                         atomic_set(i->pin, 1);
 321                 }
 322
 323                 for (k = i->j.start;
 324                      k < bset_bkey_last(&i->j);
 325                      k = bkey_next(k))
 326                         if (!__bch_extent_invalid(c, k)) {
 327                                 unsigned int j;
 328
 329                                 for (j = 0; j < KEY_PTRS(k); j++)
 330                                         if (ptr_available(c, k, j))
 331                                                 atomic_inc(&PTR_BUCKET(c, k, j)->pin);
 332
 333                                 bch_initial_mark_key(c, 0, k);
 334                         }
 335         }
 336 }
 337
 338 static bool is_discard_enabled(struct cache_set *s)
 339 {
 340         struct cache *ca = s->cache;
 341
 342         if (ca->discard)
 343                 return true;
 344
 345         return false;
 346 }
 347
 348 int bch_journal_replay(struct cache_set *s, struct list_head *list)
 349 {
 350         int ret = 0, keys = 0, entries = 0;
 351         struct bkey *k;
 352         struct journal_replay *i =
 353                 list_entry(list->prev, struct journal_replay, list);
 354
 355         uint64_t start = i->j.last_seq, end = i->j.seq, n = start;
 356         struct keylist keylist;
 357
 358         list_for_each_entry(i, list, list) {
 359                 BUG_ON(i->pin && atomic_read(i->pin) != 1);
 360
 361                 if (n != i->j.seq) {
 362                         if (n == start && is_discard_enabled(s))
 363                                 pr_info("journal entries %llu-%llu may be discarded! (replaying %llu-%llu)\n",
 364                                         n, i->j.seq - 1, start, end);
 365                         else {
 366                                 pr_err("journal entries %llu-%llu missing! (replaying %llu-%llu)\n",
 367                                         n, i->j.seq - 1, start, end);
 368                                 ret = -EIO;
 369                                 goto err;
 370                         }
 371                 }
 372
 373                 for (k = i->j.start;
 374                      k < bset_bkey_last(&i->j);
 375                      k = bkey_next(k)) {
 376                         trace_bcache_journal_replay_key(k);
 377
 378                         bch_keylist_init_single(&keylist, k);
 379
 380                         ret = bch_btree_insert(s, &keylist, i->pin, NULL);
 381                         if (ret)
 382                                 goto err;
 383
 384                         BUG_ON(!bch_keylist_empty(&keylist));
 385                         keys++;
 386
 387                         cond_resched();
 388                 }
 389
 390                 if (i->pin)
 391                         atomic_dec(i->pin);
 392                 n = i->j.seq + 1;
 393                 entries++;
 394         }
 395
 396         pr_info("journal replay done, %i keys in %i entries, seq %llu\n",
 397                 keys, entries, end);
 398 err:
 399         while (!list_empty(list)) {
 400                 i = list_first_entry(list, struct journal_replay, list);
 401                 list_del(&i->list);
 402                 kfree(i);
 403         }
 404
 405         return ret;
 406 }
 407
 408 /* Journalling */
 409
 410 static void btree_flush_write(struct cache_set *c)
 411 {
 412         struct btree *b, *t, *btree_nodes[BTREE_FLUSH_NR];
 413         unsigned int i, nr;
 414         int ref_nr;
 415         atomic_t *fifo_front_p, *now_fifo_front_p;
 416         size_t mask;
 417
 418         if (c->journal.btree_flushing)
 419                 return;
 420
 421         spin_lock(&c->journal.flush_write_lock);
 422         if (c->journal.btree_flushing) {
 423                 spin_unlock(&c->journal.flush_write_lock);
 424                 return;
 425         }
 426         c->journal.btree_flushing = true;
 427         spin_unlock(&c->journal.flush_write_lock);
 428
 429         /* get the oldest journal entry and check its refcount */
 430         spin_lock(&c->journal.lock);
 431         fifo_front_p = &fifo_front(&c->journal.pin);
 432         ref_nr = atomic_read(fifo_front_p);
 433         if (ref_nr <= 0) {
 434                 /*
 435                  * do nothing if no btree node references
 436                  * the oldest journal entry
 437                  */
 438                 spin_unlock(&c->journal.lock);
 439                 goto out;
 440         }
 441         spin_unlock(&c->journal.lock);
 442
 443         mask = c->journal.pin.mask;
 444         nr = 0;
 445         atomic_long_inc(&c->flush_write);
 446         memset(btree_nodes, 0, sizeof(btree_nodes));
 447
 448         mutex_lock(&c->bucket_lock);
 449         list_for_each_entry_safe_reverse(b, t, &c->btree_cache, list) {
 450                 /*
 451                  * It is safe to get now_fifo_front_p without holding
 452                  * c->journal.lock here, because we don't need to know
 453                  * the exactly accurate value, just check whether the
 454                  * front pointer of c->journal.pin is changed.
 455                  */
 456                 now_fifo_front_p = &fifo_front(&c->journal.pin);
 457                 /*
 458                  * If the oldest journal entry is reclaimed and front
 459                  * pointer of c->journal.pin changes, it is unnecessary
 460                  * to scan c->btree_cache anymore, just quit the loop and
 461                  * flush out what we have already.
 462                  */
 463                 if (now_fifo_front_p != fifo_front_p)
 464                         break;
 465                 /*
 466                  * quit this loop if all matching btree nodes are
 467                  * scanned and record in btree_nodes[] already.
 468                  */
 469                 ref_nr = atomic_read(fifo_front_p);
 470                 if (nr >= ref_nr)
 471                         break;
 472
 473                 if (btree_node_journal_flush(b))
 474                         pr_err("BUG: flush_write bit should not be set here!\n");
 475
 476                 mutex_lock(&b->write_lock);
 477
 478                 if (!btree_node_dirty(b)) {
 479                         mutex_unlock(&b->write_lock);
 480                         continue;
 481                 }
 482
 483                 if (!btree_current_write(b)->journal) {
 484                         mutex_unlock(&b->write_lock);
 485                         continue;
 486                 }
 487
 488                 /*
 489                  * Only select the btree node which exactly references
 490                  * the oldest journal entry.
 491                  *
 492                  * If the journal entry pointed by fifo_front_p is
 493                  * reclaimed in parallel, don't worry:
 494                  * - the list_for_each_xxx loop will quit when checking
 495                  *   next now_fifo_front_p.
 496                  * - If there are matched nodes recorded in btree_nodes[],
 497                  *   they are clean now (this is why and how the oldest
 498                  *   journal entry can be reclaimed). These selected nodes
 499                  *   will be ignored and skipped in the following for-loop.
 500                  */
 501                 if (((btree_current_write(b)->journal - fifo_front_p) &
 502                      mask) != 0) {
 503                         mutex_unlock(&b->write_lock);
 504                         continue;
 505                 }
 506
 507                 set_btree_node_journal_flush(b);
 508
 509                 mutex_unlock(&b->write_lock);
 510
 511                 btree_nodes[nr++] = b;
 512                 /*
 513                  * To avoid holding c->bucket_lock too long time,
 514                  * only scan for BTREE_FLUSH_NR matched btree nodes
 515                  * at most. If there are more btree nodes reference
 516                  * the oldest journal entry, try to flush them next
 517                  * time when btree_flush_write() is called.
 518                  */
 519                 if (nr == BTREE_FLUSH_NR)
 520                         break;
 521         }
 522         mutex_unlock(&c->bucket_lock);
 523
 524         for (i = 0; i < nr; i++) {
 525                 b = btree_nodes[i];
 526                 if (!b) {
 527                         pr_err("BUG: btree_nodes[%d] is NULL\n", i);
 528                         continue;
 529                 }
 530
 531                 /* safe to check without holding b->write_lock */
 532                 if (!btree_node_journal_flush(b)) {
 533                         pr_err("BUG: bnode %p: journal_flush bit cleaned\n", b);
 534                         continue;
 535                 }
 536
 537                 mutex_lock(&b->write_lock);
 538                 if (!btree_current_write(b)->journal) {
 539                         clear_bit(BTREE_NODE_journal_flush, &b->flags);
 540                         mutex_unlock(&b->write_lock);
 541                         pr_debug("bnode %p: written by others\n", b);
 542                         continue;
 543                 }
 544
 545                 if (!btree_node_dirty(b)) {
 546                         clear_bit(BTREE_NODE_journal_flush, &b->flags);
 547                         mutex_unlock(&b->write_lock);
 548                         pr_debug("bnode %p: dirty bit cleaned by others\n", b);
 549                         continue;
 550                 }
 551
 552                 __bch_btree_node_write(b, NULL);
 553                 clear_bit(BTREE_NODE_journal_flush, &b->flags);
 554                 mutex_unlock(&b->write_lock);
 555         }
 556
 557 out:
 558         spin_lock(&c->journal.flush_write_lock);
 559         c->journal.btree_flushing = false;
 560         spin_unlock(&c->journal.flush_write_lock);
 561 }
 562
 563 #define last_seq(j)     ((j)->seq - fifo_used(&(j)->pin) + 1)
 564
 565 static void journal_discard_endio(struct bio *bio)
 566 {
 567         struct journal_device *ja =
 568                 container_of(bio, struct journal_device, discard_bio);
 569         struct cache *ca = container_of(ja, struct cache, journal);
 570
 571         atomic_set(&ja->discard_in_flight, DISCARD_DONE);
 572
 573         closure_wake_up(&ca->set->journal.wait);
 574         closure_put(&ca->set->cl);
 575 }
 576
 577 static void journal_discard_work(struct work_struct *work)
 578 {
 579         struct journal_device *ja =
 580                 container_of(work, struct journal_device, discard_work);
 581
 582         submit_bio(&ja->discard_bio);
 583 }
 584
 585 static void do_journal_discard(struct cache *ca)
 586 {
 587         struct journal_device *ja = &ca->journal;
 588         struct bio *bio = &ja->discard_bio;
 589
 590         if (!ca->discard) {
 591                 ja->discard_idx = ja->last_idx;
 592                 return;
 593         }
 594
 595         switch (atomic_read(&ja->discard_in_flight)) {
 596         case DISCARD_IN_FLIGHT:
 597                 return;
 598
 599         case DISCARD_DONE:
 600                 ja->discard_idx = (ja->discard_idx + 1) %
 601                         ca->sb.njournal_buckets;
 602
 603                 atomic_set(&ja->discard_in_flight, DISCARD_READY);
 604                 fallthrough;
 605
 606         case DISCARD_READY:
 607                 if (ja->discard_idx == ja->last_idx)
 608                         return;
 609
 610                 atomic_set(&ja->discard_in_flight, DISCARD_IN_FLIGHT);
 611
 612                 bio_init(bio, ca->bdev, bio->bi_inline_vecs, 1, REQ_OP_DISCARD);
 613                 bio->bi_iter.bi_sector  = bucket_to_sector(ca->set,
 614                                                 ca->sb.d[ja->discard_idx]);
 615                 bio->bi_iter.bi_size    = bucket_bytes(ca);
 616                 bio->bi_end_io          = journal_discard_endio;
 617
 618                 closure_get(&ca->set->cl);
 619                 INIT_WORK(&ja->discard_work, journal_discard_work);
 620                 queue_work(bch_journal_wq, &ja->discard_work);
 621         }
 622 }
 623
 624 static void journal_reclaim(struct cache_set *c)
 625 {
 626         struct bkey *k = &c->journal.key;
 627         struct cache *ca = c->cache;
 628         uint64_t last_seq;
 629         unsigned int next;
 630         struct journal_device *ja = &ca->journal;
 631         atomic_t p __maybe_unused;
 632
 633         atomic_long_inc(&c->reclaim);
 634
 635         while (!atomic_read(&fifo_front(&c->journal.pin)))
 636                 fifo_pop(&c->journal.pin, p);
 637
 638         last_seq = last_seq(&c->journal);
 639
 640         /* Update last_idx */
 641
 642         while (ja->last_idx != ja->cur_idx &&
 643                ja->seq[ja->last_idx] < last_seq)
 644                 ja->last_idx = (ja->last_idx + 1) %
 645                         ca->sb.njournal_buckets;
 646
 647         do_journal_discard(ca);
 648
 649         if (c->journal.blocks_free)
 650                 goto out;
 651
 652         next = (ja->cur_idx + 1) % ca->sb.njournal_buckets;
 653         /* No space available on this device */
 654         if (next == ja->discard_idx)
 655                 goto out;
 656
 657         ja->cur_idx = next;
 658         k->ptr[0] = MAKE_PTR(0,
 659                              bucket_to_sector(c, ca->sb.d[ja->cur_idx]),
 660                              ca->sb.nr_this_dev);
 661         atomic_long_inc(&c->reclaimed_journal_buckets);
 662
 663         bkey_init(k);
 664         SET_KEY_PTRS(k, 1);
 665         c->journal.blocks_free = ca->sb.bucket_size >> c->block_bits;
 666
 667 out:
 668         if (!journal_full(&c->journal))
 669                 __closure_wake_up(&c->journal.wait);
 670 }
 671
 672 void bch_journal_next(struct journal *j)
 673 {
 674         atomic_t p = { 1 };
 675
 676         j->cur = (j->cur == j->w)
 677                 ? &j->w[1]
 678                 : &j->w[0];
 679
 680         /*
 681          * The fifo_push() needs to happen at the same time as j->seq is
 682          * incremented for last_seq() to be calculated correctly
 683          */
 684         BUG_ON(!fifo_push(&j->pin, p));
 685         atomic_set(&fifo_back(&j->pin), 1);
 686
 687         j->cur->data->seq       = ++j->seq;
 688         j->cur->dirty           = false;
 689         j->cur->need_write      = false;
 690         j->cur->data->keys      = 0;
 691
 692         if (fifo_full(&j->pin))
 693                 pr_debug("journal_pin full (%zu)\n", fifo_used(&j->pin));
 694 }
 695
 696 static void journal_write_endio(struct bio *bio)
 697 {
 698         struct journal_write *w = bio->bi_private;
 699
 700         cache_set_err_on(bio->bi_status, w->c, "journal io error");
 701         closure_put(&w->c->journal.io);
 702 }
 703
 704 static void journal_write(struct closure *cl);
 705
 706 static void journal_write_done(struct closure *cl)
 707 {
 708         struct journal *j = container_of(cl, struct journal, io);
 709         struct journal_write *w = (j->cur == j->w)
 710                 ? &j->w[1]
 711                 : &j->w[0];
 712
 713         __closure_wake_up(&w->wait);
 714         continue_at_nobarrier(cl, journal_write, bch_journal_wq);
 715 }
 716
 717 static void journal_write_unlock(struct closure *cl)
 718         __releases(&c->journal.lock)
 719 {
 720         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 721
 722         c->journal.io_in_flight = 0;
 723         spin_unlock(&c->journal.lock);
 724 }
 725
 726 static void journal_write_unlocked(struct closure *cl)
 727         __releases(c->journal.lock)
 728 {
 729         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 730         struct cache *ca = c->cache;
 731         struct journal_write *w = c->journal.cur;
 732         struct bkey *k = &c->journal.key;
 733         unsigned int i, sectors = set_blocks(w->data, block_bytes(ca)) *
 734                 ca->sb.block_size;
 735
 736         struct bio *bio;
 737         struct bio_list list;
 738
 739         bio_list_init(&list);
 740
 741         if (!w->need_write) {
 742                 closure_return_with_destructor(cl, journal_write_unlock);
 743                 return;
 744         } else if (journal_full(&c->journal)) {
 745                 journal_reclaim(c);
 746                 spin_unlock(&c->journal.lock);
 747
 748                 btree_flush_write(c);
 749                 continue_at(cl, journal_write, bch_journal_wq);
 750                 return;
 751         }
 752
 753         c->journal.blocks_free -= set_blocks(w->data, block_bytes(ca));
 754
 755         w->data->btree_level = c->root->level;
 756
 757         bkey_copy(&w->data->btree_root, &c->root->key);
 758         bkey_copy(&w->data->uuid_bucket, &c->uuid_bucket);
 759
 760         w->data->prio_bucket[ca->sb.nr_this_dev] = ca->prio_buckets[0];
 761         w->data->magic          = jset_magic(&ca->sb);
 762         w->data->version        = BCACHE_JSET_VERSION;
 763         w->data->last_seq       = last_seq(&c->journal);
 764         w->data->csum           = csum_set(w->data);
 765
 766         for (i = 0; i < KEY_PTRS(k); i++) {
 767                 ca = c->cache;
 768                 bio = &ca->journal.bio;
 769
 770                 atomic_long_add(sectors, &ca->meta_sectors_written);
 771
 772                 bio_reset(bio, ca->bdev, REQ_OP_WRITE |
 773                           REQ_SYNC | REQ_META | REQ_PREFLUSH | REQ_FUA);
 774                 bio->bi_iter.bi_sector  = PTR_OFFSET(k, i);
 775                 bio->bi_iter.bi_size = sectors << 9;
 776
 777                 bio->bi_end_io  = journal_write_endio;
 778                 bio->bi_private = w;
 779                 bch_bio_map(bio, w->data);
 780
 781                 trace_bcache_journal_write(bio, w->data->keys);
 782                 bio_list_add(&list, bio);
 783
 784                 SET_PTR_OFFSET(k, i, PTR_OFFSET(k, i) + sectors);
 785
 786                 ca->journal.seq[ca->journal.cur_idx] = w->data->seq;
 787         }
 788
 789         /* If KEY_PTRS(k) == 0, this jset gets lost in air */
 790         BUG_ON(i == 0);
 791
 792         atomic_dec_bug(&fifo_back(&c->journal.pin));
 793         bch_journal_next(&c->journal);
 794         journal_reclaim(c);
 795
 796         spin_unlock(&c->journal.lock);
 797
 798         while ((bio = bio_list_pop(&list)))
 799                 closure_bio_submit(c, bio, cl);
 800
 801         continue_at(cl, journal_write_done, NULL);
 802 }
 803
 804 static void journal_write(struct closure *cl)
 805 {
 806         struct cache_set *c = container_of(cl, struct cache_set, journal.io);
 807
 808         spin_lock(&c->journal.lock);
 809         journal_write_unlocked(cl);
 810 }
 811
 812 static void journal_try_write(struct cache_set *c)
 813         __releases(c->journal.lock)
 814 {
 815         struct closure *cl = &c->journal.io;
 816         struct journal_write *w = c->journal.cur;
 817
 818         w->need_write = true;
 819
 820         if (!c->journal.io_in_flight) {
 821                 c->journal.io_in_flight = 1;
 822                 closure_call(cl, journal_write_unlocked, NULL, &c->cl);
 823         } else {
 824                 spin_unlock(&c->journal.lock);
 825         }
 826 }
 827
 828 static struct journal_write *journal_wait_for_write(struct cache_set *c,
 829                                                     unsigned int nkeys)
 830         __acquires(&c->journal.lock)
 831 {
 832         size_t sectors;
 833         struct closure cl;
 834         bool wait = false;
 835         struct cache *ca = c->cache;
 836
 837         closure_init_stack(&cl);
 838
 839         spin_lock(&c->journal.lock);
 840
 841         while (1) {
 842                 struct journal_write *w = c->journal.cur;
 843
 844                 sectors = __set_blocks(w->data, w->data->keys + nkeys,
 845                                        block_bytes(ca)) * ca->sb.block_size;
 846
 847                 if (sectors <= min_t(size_t,
 848                                      c->journal.blocks_free * ca->sb.block_size,
 849                                      PAGE_SECTORS << JSET_BITS))
 850                         return w;
 851
 852                 if (wait)
 853                         closure_wait(&c->journal.wait, &cl);
 854
 855                 if (!journal_full(&c->journal)) {
 856                         if (wait)
 857                                 trace_bcache_journal_entry_full(c);
 858
 859                         /*
 860                          * XXX: If we were inserting so many keys that they
 861                          * won't fit in an _empty_ journal write, we'll
 862                          * deadlock. For now, handle this in
 863                          * bch_keylist_realloc() - but something to think about.
 864                          */
 865                         BUG_ON(!w->data->keys);
 866
 867                         journal_try_write(c); /* unlocks */
 868                 } else {
 869                         if (wait)
 870                                 trace_bcache_journal_full(c);
 871
 872                         journal_reclaim(c);
 873                         spin_unlock(&c->journal.lock);
 874
 875                         btree_flush_write(c);
 876                 }
 877
 878                 closure_sync(&cl);
 879                 spin_lock(&c->journal.lock);
 880                 wait = true;
 881         }
 882 }
 883
 884 static void journal_write_work(struct work_struct *work)
 885 {
 886         struct cache_set *c = container_of(to_delayed_work(work),
 887                                            struct cache_set,
 888                                            journal.work);
 889         spin_lock(&c->journal.lock);
 890         if (c->journal.cur->dirty)
 891                 journal_try_write(c);
 892         else
 893                 spin_unlock(&c->journal.lock);
 894 }
 895
 896 /*
 897  * Entry point to the journalling code - bio_insert() and btree_invalidate()
 898  * pass bch_journal() a list of keys to be journalled, and then
 899  * bch_journal() hands those same keys off to btree_insert_async()
 900  */
 901
 902 atomic_t *bch_journal(struct cache_set *c,
 903                       struct keylist *keys,
 904                       struct closure *parent)
 905 {
 906         struct journal_write *w;
 907         atomic_t *ret;
 908
 909         /* No journaling if CACHE_SET_IO_DISABLE set already */
 910         if (unlikely(test_bit(CACHE_SET_IO_DISABLE, &c->flags)))
 911                 return NULL;
 912
 913         if (!CACHE_SYNC(&c->cache->sb))
 914                 return NULL;
 915
 916         w = journal_wait_for_write(c, bch_keylist_nkeys(keys));
 917
 918         memcpy(bset_bkey_last(w->data), keys->keys, bch_keylist_bytes(keys));
 919         w->data->keys += bch_keylist_nkeys(keys);
 920
 921         ret = &fifo_back(&c->journal.pin);
 922         atomic_inc(ret);
 923
 924         if (parent) {
 925                 closure_wait(&w->wait, parent);
 926                 journal_try_write(c);
 927         } else if (!w->dirty) {
 928                 w->dirty = true;
 929                 queue_delayed_work(bch_flush_wq, &c->journal.work,
 930                                    msecs_to_jiffies(c->journal_delay_ms));
 931                 spin_unlock(&c->journal.lock);
 932         } else {
 933                 spin_unlock(&c->journal.lock);
 934         }
 935
 936
 937         return ret;
 938 }
 939
 940 void bch_journal_meta(struct cache_set *c, struct closure *cl)
 941 {
 942         struct keylist keys;
 943         atomic_t *ref;
 944
 945         bch_keylist_init(&keys);
 946
 947         ref = bch_journal(c, &keys, cl);
 948         if (ref)
 949                 atomic_dec_bug(ref);
 950 }
 951
 952 void bch_journal_free(struct cache_set *c)
 953 {
 954         free_pages((unsigned long) c->journal.w[1].data, JSET_BITS);
 955         free_pages((unsigned long) c->journal.w[0].data, JSET_BITS);
 956         free_fifo(&c->journal.pin);
 957 }
 958
 959 int bch_journal_alloc(struct cache_set *c)
 960 {
 961         struct journal *j = &c->journal;
 962
 963         spin_lock_init(&j->lock);
 964         spin_lock_init(&j->flush_write_lock);
 965         INIT_DELAYED_WORK(&j->work, journal_write_work);
 966
 967         c->journal_delay_ms = 100;
 968
 969         j->w[0].c = c;
 970         j->w[1].c = c;
 971
 972         if (!(init_fifo(&j->pin, JOURNAL_PIN, GFP_KERNEL)) ||
 973             !(j->w[0].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)) ||
 974             !(j->w[1].data = (void *) __get_free_pages(GFP_KERNEL|__GFP_COMP, JSET_BITS)))
 975                 return -ENOMEM;
 976
 977         return 0;
 978 }