git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_write_buffer.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_locking.h"
   5 #include "btree_update.h"
   6 #include "btree_update_interior.h"
   7 #include "btree_write_buffer.h"
   8 #include "error.h"
   9 #include "journal.h"
  10 #include "journal_io.h"
  11 #include "journal_reclaim.h"
  12
  13 #include <linux/prefetch.h>
  14
  15 static int bch2_btree_write_buffer_journal_flush(struct journal *,
  16                                 struct journal_entry_pin *, u64);
  17
  18 static int bch2_journal_keys_to_write_buffer(struct bch_fs *, struct journal_buf *);
  19
  20 static inline bool __wb_key_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
  21 {
  22         return (cmp_int(l->hi, r->hi) ?:
  23                 cmp_int(l->mi, r->mi) ?:
  24                 cmp_int(l->lo, r->lo)) >= 0;
  25 }
  26
  27 static inline bool wb_key_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
  28 {
  29 #ifdef CONFIG_X86_64
  30         int cmp;
  31
  32         asm("mov   (%[l]), %%rax;"
  33             "sub   (%[r]), %%rax;"
  34             "mov  8(%[l]), %%rax;"
  35             "sbb  8(%[r]), %%rax;"
  36             "mov 16(%[l]), %%rax;"
  37             "sbb 16(%[r]), %%rax;"
  38             : "=@ccae" (cmp)
  39             : [l] "r" (l), [r] "r" (r)
  40             : "rax", "cc");
  41
  42         EBUG_ON(cmp != __wb_key_cmp(l, r));
  43         return cmp;
  44 #else
  45         return __wb_key_cmp(l, r);
  46 #endif
  47 }
  48
  49 /* Compare excluding idx, the low 24 bits: */
  50 static inline bool wb_key_eq(const void *_l, const void *_r)
  51 {
  52         const struct wb_key_ref *l = _l;
  53         const struct wb_key_ref *r = _r;
  54
  55         return !((l->hi ^ r->hi)|
  56                  (l->mi ^ r->mi)|
  57                  ((l->lo >> 24) ^ (r->lo >> 24)));
  58 }
  59
  60 static noinline void wb_sort(struct wb_key_ref *base, size_t num)
  61 {
  62         size_t n = num, a = num / 2;
  63
  64         if (!a)         /* num < 2 || size == 0 */
  65                 return;
  66
  67         for (;;) {
  68                 size_t b, c, d;
  69
  70                 if (a)                  /* Building heap: sift down --a */
  71                         --a;
  72                 else if (--n)           /* Sorting: Extract root to --n */
  73                         swap(base[0], base[n]);
  74                 else                    /* Sort complete */
  75                         break;
  76
  77                 /*
  78                  * Sift element at "a" down into heap.  This is the
  79                  * "bottom-up" variant, which significantly reduces
  80                  * calls to cmp_func(): we find the sift-down path all
  81                  * the way to the leaves (one compare per level), then
  82                  * backtrack to find where to insert the target element.
  83                  *
  84                  * Because elements tend to sift down close to the leaves,
  85                  * this uses fewer compares than doing two per level
  86                  * on the way down.  (A bit more than half as many on
  87                  * average, 3/4 worst-case.)
  88                  */
  89                 for (b = a; c = 2*b + 1, (d = c + 1) < n;)
  90                         b = wb_key_cmp(base + c, base + d) ? c : d;
  91                 if (d == n)             /* Special case last leaf with no sibling */
  92                         b = c;
  93
  94                 /* Now backtrack from "b" to the correct location for "a" */
  95                 while (b != a && wb_key_cmp(base + a, base + b))
  96                         b = (b - 1) / 2;
  97                 c = b;                  /* Where "a" belongs */
  98                 while (b != a) {        /* Shift it into place */
  99                         b = (b - 1) / 2;
 100                         swap(base[b], base[c]);
 101                 }
 102         }
 103 }
 104
 105 static noinline int wb_flush_one_slowpath(struct btree_trans *trans,
 106                                           struct btree_iter *iter,
 107                                           struct btree_write_buffered_key *wb)
 108 {
 109         bch2_btree_node_unlock_write(trans, iter->path, iter->path->l[0].b);
 110
 111         trans->journal_res.seq = wb->journal_seq;
 112
 113         return bch2_trans_update(trans, iter, &wb->k,
 114                                  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 115                 bch2_trans_commit(trans, NULL, NULL,
 116                                   BCH_TRANS_COMMIT_no_enospc|
 117                                   BCH_TRANS_COMMIT_no_check_rw|
 118                                   BCH_TRANS_COMMIT_no_journal_res|
 119                                   BCH_TRANS_COMMIT_journal_reclaim);
 120 }
 121
 122 static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
 123                                struct btree_write_buffered_key *wb,
 124                                bool *write_locked, size_t *fast)
 125 {
 126         struct bch_fs *c = trans->c;
 127         struct btree_path *path;
 128         int ret;
 129
 130         EBUG_ON(!wb->journal_seq);
 131         EBUG_ON(!c->btree_write_buffer.flushing.pin.seq);
 132         EBUG_ON(c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);
 133
 134         ret = bch2_btree_iter_traverse(iter);
 135         if (ret)
 136                 return ret;
 137
 138         /*
 139          * We can't clone a path that has write locks: unshare it now, before
 140          * set_pos and traverse():
 141          */
 142         if (iter->path->ref > 1)
 143                 iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);
 144
 145         path = iter->path;
 146
 147         if (!*write_locked) {
 148                 ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
 149                 if (ret)
 150                         return ret;
 151
 152                 bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
 153                 *write_locked = true;
 154         }
 155
 156         if (unlikely(!bch2_btree_node_insert_fits(c, path->l[0].b, wb->k.k.u64s))) {
 157                 *write_locked = false;
 158                 return wb_flush_one_slowpath(trans, iter, wb);
 159         }
 160
 161         bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
 162         (*fast)++;
 163         return 0;
 164 }
 165
 166 /*
 167  * Update a btree with a write buffered key using the journal seq of the
 168  * original write buffer insert.
 169  *
 170  * It is not safe to rejournal the key once it has been inserted into the write
 171  * buffer because that may break recovery ordering. For example, the key may
 172  * have already been modified in the active write buffer in a seq that comes
 173  * before the current transaction. If we were to journal this key again and
 174  * crash, recovery would process updates in the wrong order.
 175  */
 176 static int
 177 btree_write_buffered_insert(struct btree_trans *trans,
 178                           struct btree_write_buffered_key *wb)
 179 {
 180         struct btree_iter iter;
 181         int ret;
 182
 183         bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
 184                              BTREE_ITER_CACHED|BTREE_ITER_INTENT);
 185
 186         trans->journal_res.seq = wb->journal_seq;
 187
 188         ret   = bch2_btree_iter_traverse(&iter) ?:
 189                 bch2_trans_update(trans, &iter, &wb->k,
 190                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 191         bch2_trans_iter_exit(trans, &iter);
 192         return ret;
 193 }
 194
 195 static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
 196 {
 197         struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
 198         struct journal *j = &c->journal;
 199
 200         if (!wb->inc.keys.nr)
 201                 return;
 202
 203         bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
 204                              bch2_btree_write_buffer_journal_flush);
 205
 206         darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
 207         darray_resize(&wb->sorted, wb->flushing.keys.size);
 208
 209         if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
 210                 swap(wb->flushing.keys, wb->inc.keys);
 211                 goto out;
 212         }
 213
 214         size_t nr = min(darray_room(wb->flushing.keys),
 215                         wb->sorted.size - wb->flushing.keys.nr);
 216         nr = min(nr, wb->inc.keys.nr);
 217
 218         memcpy(&darray_top(wb->flushing.keys),
 219                wb->inc.keys.data,
 220                sizeof(wb->inc.keys.data[0]) * nr);
 221
 222         memmove(wb->inc.keys.data,
 223                 wb->inc.keys.data + nr,
 224                sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));
 225
 226         wb->flushing.keys.nr    += nr;
 227         wb->inc.keys.nr         -= nr;
 228 out:
 229         if (!wb->inc.keys.nr)
 230                 bch2_journal_pin_drop(j, &wb->inc.pin);
 231         else
 232                 bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
 233                                         bch2_btree_write_buffer_journal_flush);
 234
 235         if (j->watermark) {
 236                 spin_lock(&j->lock);
 237                 bch2_journal_set_watermark(j);
 238                 spin_unlock(&j->lock);
 239         }
 240
 241         BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
 242 }
 243
 244 static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
 245 {
 246         struct bch_fs *c = trans->c;
 247         struct journal *j = &c->journal;
 248         struct btree_write_buffer *wb = &c->btree_write_buffer;
 249         struct wb_key_ref *i;
 250         struct btree_iter iter = { NULL };
 251         size_t skipped = 0, fast = 0, slowpath = 0;
 252         bool write_locked = false;
 253         int ret = 0;
 254
 255         bch2_trans_unlock(trans);
 256         bch2_trans_begin(trans);
 257
 258         mutex_lock(&wb->inc.lock);
 259         move_keys_from_inc_to_flushing(wb);
 260         mutex_unlock(&wb->inc.lock);
 261
 262         for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
 263                 wb->sorted.data[i].idx = i;
 264                 wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
 265                 memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
 266         }
 267         wb->sorted.nr = wb->flushing.keys.nr;
 268
 269         /*
 270          * We first sort so that we can detect and skip redundant updates, and
 271          * then we attempt to flush in sorted btree order, as this is most
 272          * efficient.
 273          *
 274          * However, since we're not flushing in the order they appear in the
 275          * journal we won't be able to drop our journal pin until everything is
 276          * flushed - which means this could deadlock the journal if we weren't
 277          * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
 278          * if it would block taking a journal reservation.
 279          *
 280          * If that happens, simply skip the key so we can optimistically insert
 281          * as many keys as possible in the fast path.
 282          */
 283         wb_sort(wb->sorted.data, wb->sorted.nr);
 284
 285         darray_for_each(wb->sorted, i) {
 286                 struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
 287
 288                 for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
 289                         prefetch(&wb->flushing.keys.data[n->idx]);
 290
 291                 BUG_ON(!k->journal_seq);
 292
 293                 if (i + 1 < &darray_top(wb->sorted) &&
 294                     wb_key_eq(i, i + 1)) {
 295                         struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];
 296
 297                         skipped++;
 298                         n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
 299                         k->journal_seq = 0;
 300                         continue;
 301                 }
 302
 303                 if (write_locked &&
 304                     (iter.path->btree_id != k->btree ||
 305                      bpos_gt(k->k.k.p, iter.path->l[0].b->key.k.p))) {
 306                         bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 307                         write_locked = false;
 308                 }
 309
 310                 if (!iter.path || iter.path->btree_id != k->btree) {
 311                         bch2_trans_iter_exit(trans, &iter);
 312                         bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
 313                                              BTREE_ITER_INTENT|BTREE_ITER_ALL_SNAPSHOTS);
 314                 }
 315
 316                 bch2_btree_iter_set_pos(&iter, k->k.k.p);
 317                 iter.path->preserve = false;
 318
 319                 do {
 320                         if (race_fault()) {
 321                                 ret = -BCH_ERR_journal_reclaim_would_deadlock;
 322                                 break;
 323                         }
 324
 325                         ret = wb_flush_one(trans, &iter, k, &write_locked, &fast);
 326                         if (!write_locked)
 327                                 bch2_trans_begin(trans);
 328                 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
 329
 330                 if (!ret) {
 331                         k->journal_seq = 0;
 332                 } else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
 333                         slowpath++;
 334                         ret = 0;
 335                 } else
 336                         break;
 337         }
 338
 339         if (write_locked)
 340                 bch2_btree_node_unlock_write(trans, iter.path, iter.path->l[0].b);
 341         bch2_trans_iter_exit(trans, &iter);
 342
 343         if (ret)
 344                 goto err;
 345
 346         if (slowpath) {
 347                 /*
 348                  * Flush in the order they were present in the journal, so that
 349                  * we can release journal pins:
 350                  * The fastpath zapped the seq of keys that were successfully flushed so
 351                  * we can skip those here.
 352                  */
 353                 trace_write_buffer_flush_slowpath(trans, slowpath, wb->flushing.keys.nr);
 354
 355                 struct btree_write_buffered_key *i;
 356                 darray_for_each(wb->flushing.keys, i) {
 357                         if (!i->journal_seq)
 358                                 continue;
 359
 360                         bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
 361                                                 bch2_btree_write_buffer_journal_flush);
 362
 363                         bch2_trans_begin(trans);
 364
 365                         ret = commit_do(trans, NULL, NULL,
 366                                         BCH_WATERMARK_reclaim|
 367                                         BCH_TRANS_COMMIT_no_check_rw|
 368                                         BCH_TRANS_COMMIT_no_enospc|
 369                                         BCH_TRANS_COMMIT_no_journal_res|
 370                                         BCH_TRANS_COMMIT_journal_reclaim,
 371                                         btree_write_buffered_insert(trans, i));
 372                         if (ret)
 373                                 goto err;
 374                 }
 375         }
 376 err:
 377         bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
 378         trace_write_buffer_flush(trans, wb->flushing.keys.nr, skipped, fast, 0);
 379         bch2_journal_pin_drop(j, &wb->flushing.pin);
 380         wb->flushing.keys.nr = 0;
 381         return ret;
 382 }
 383
 384 static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
 385 {
 386         struct journal *j = &c->journal;
 387         struct journal_buf *buf;
 388         int ret = 0;
 389
 390         mutex_lock(&j->buf_lock);
 391         while ((buf = bch2_next_write_buffer_flush_journal_buf(j, seq)))
 392                 if (bch2_journal_keys_to_write_buffer(c, buf)) {
 393                         ret = -ENOMEM;
 394                         break;
 395                 }
 396         mutex_unlock(&j->buf_lock);
 397
 398         return ret;
 399 }
 400
 401 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 402 {
 403         struct bch_fs *c = trans->c;
 404         struct btree_write_buffer *wb = &c->btree_write_buffer;
 405         int ret = 0, fetch_from_journal_err;
 406
 407         trace_write_buffer_flush_sync(trans, _RET_IP_);
 408 retry:
 409         bch2_trans_unlock(trans);
 410
 411         bch2_journal_block_reservations(&c->journal);
 412         fetch_from_journal_err = fetch_wb_keys_from_journal(c, U64_MAX);
 413         bch2_journal_unblock(&c->journal);
 414
 415         /*
 416          * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
 417          * is not guaranteed to empty wb->inc:
 418          */
 419         mutex_lock(&wb->flushing.lock);
 420         while (!ret &&
 421                (wb->flushing.keys.nr || wb->inc.keys.nr))
 422                 ret = bch2_btree_write_buffer_flush_locked(trans);
 423         mutex_unlock(&wb->flushing.lock);
 424
 425         if (!ret && fetch_from_journal_err)
 426                 goto retry;
 427
 428         return ret;
 429 }
 430
 431 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
 432 {
 433         struct bch_fs *c = trans->c;
 434         struct btree_write_buffer *wb = &c->btree_write_buffer;
 435         int ret = 0;
 436
 437         if (mutex_trylock(&wb->flushing.lock)) {
 438                 ret = bch2_btree_write_buffer_flush_locked(trans);
 439                 mutex_unlock(&wb->flushing.lock);
 440         }
 441
 442         return ret;
 443 }
 444
 445 int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
 446 {
 447         struct bch_fs *c = trans->c;
 448
 449         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer))
 450                 return -BCH_ERR_erofs_no_writes;
 451
 452         int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
 453         bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 454         return ret;
 455 }
 456
 457 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
 458                                 struct journal_entry_pin *_pin, u64 seq)
 459 {
 460         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 461         struct btree_write_buffer *wb = &c->btree_write_buffer;
 462         int ret, fetch_from_journal_err;
 463
 464         do {
 465                 fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
 466
 467                 mutex_lock(&wb->flushing.lock);
 468                 ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
 469                 mutex_unlock(&wb->flushing.lock);
 470         } while (!ret &&
 471                  (fetch_from_journal_err ||
 472                   (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq) ||
 473                   (wb->inc.pin.seq && wb->inc.pin.seq <= seq)));
 474
 475         return ret;
 476 }
 477
 478 static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
 479 {
 480         struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
 481         struct btree_write_buffer *wb = &c->btree_write_buffer;
 482         int ret;
 483
 484         mutex_lock(&wb->flushing.lock);
 485         do {
 486                 ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
 487         } while (!ret && bch2_btree_write_buffer_should_flush(c));
 488         mutex_unlock(&wb->flushing.lock);
 489
 490         bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 491 }
 492
 493 int __bch2_journal_key_to_wb(struct bch_fs *c,
 494                              struct journal_keys_to_wb *dst,
 495                              enum btree_id btree, struct bkey_i *k)
 496 {
 497         struct btree_write_buffer *wb = &c->btree_write_buffer;
 498         int ret;
 499 retry:
 500         ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
 501         if (!ret && dst->wb == &wb->flushing)
 502                 ret = darray_resize(&wb->sorted, wb->flushing.keys.size);
 503
 504         if (unlikely(ret)) {
 505                 if (dst->wb == &c->btree_write_buffer.flushing) {
 506                         mutex_unlock(&dst->wb->lock);
 507                         dst->wb = &c->btree_write_buffer.inc;
 508                         bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
 509                                              bch2_btree_write_buffer_journal_flush);
 510                         goto retry;
 511                 }
 512
 513                 return ret;
 514         }
 515
 516         dst->room = darray_room(dst->wb->keys);
 517         if (dst->wb == &wb->flushing)
 518                 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 519         BUG_ON(!dst->room);
 520         BUG_ON(!dst->seq);
 521
 522         struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
 523         wb_k->journal_seq       = dst->seq;
 524         wb_k->btree             = btree;
 525         bkey_copy(&wb_k->k, k);
 526         dst->wb->keys.nr++;
 527         dst->room--;
 528         return 0;
 529 }
 530
 531 void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
 532 {
 533         struct btree_write_buffer *wb = &c->btree_write_buffer;
 534
 535         if (mutex_trylock(&wb->flushing.lock)) {
 536                 mutex_lock(&wb->inc.lock);
 537                 move_keys_from_inc_to_flushing(wb);
 538
 539                 /*
 540                  * Attempt to skip wb->inc, and add keys directly to
 541                  * wb->flushing, saving us a copy later:
 542                  */
 543
 544                 if (!wb->inc.keys.nr) {
 545                         dst->wb = &wb->flushing;
 546                 } else {
 547                         mutex_unlock(&wb->flushing.lock);
 548                         dst->wb = &wb->inc;
 549                 }
 550         } else {
 551                 mutex_lock(&wb->inc.lock);
 552                 dst->wb = &wb->inc;
 553         }
 554
 555         dst->room = darray_room(dst->wb->keys);
 556         if (dst->wb == &wb->flushing)
 557                 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 558         dst->seq = seq;
 559
 560         bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
 561                              bch2_btree_write_buffer_journal_flush);
 562 }
 563
 564 void bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
 565 {
 566         struct btree_write_buffer *wb = &c->btree_write_buffer;
 567
 568         if (!dst->wb->keys.nr)
 569                 bch2_journal_pin_drop(&c->journal, &dst->wb->pin);
 570
 571         if (bch2_btree_write_buffer_should_flush(c) &&
 572             __bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer) &&
 573             !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
 574                 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 575
 576         if (dst->wb == &wb->flushing)
 577                 mutex_unlock(&wb->flushing.lock);
 578         mutex_unlock(&wb->inc.lock);
 579 }
 580
 581 static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
 582 {
 583         struct journal_keys_to_wb dst;
 584         struct jset_entry *entry;
 585         struct bkey_i *k;
 586         int ret = 0;
 587
 588         bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
 589
 590         for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
 591                 jset_entry_for_each_key(entry, k) {
 592                         ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
 593                         if (ret)
 594                                 goto out;
 595                 }
 596
 597                 entry->type = BCH_JSET_ENTRY_btree_keys;
 598         }
 599
 600         buf->need_flush_to_write_buffer = false;
 601 out:
 602         bch2_journal_keys_to_write_buffer_end(c, &dst);
 603         return ret;
 604 }
 605
 606 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
 607 {
 608         if (wb->keys.size >= new_size)
 609                 return 0;
 610
 611         if (!mutex_trylock(&wb->lock))
 612                 return -EINTR;
 613
 614         int ret = darray_resize(&wb->keys, new_size);
 615         mutex_unlock(&wb->lock);
 616         return ret;
 617 }
 618
 619 int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
 620 {
 621         struct btree_write_buffer *wb = &c->btree_write_buffer;
 622
 623         return wb_keys_resize(&wb->flushing, new_size) ?:
 624                 wb_keys_resize(&wb->inc, new_size);
 625 }
 626
 627 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
 628 {
 629         struct btree_write_buffer *wb = &c->btree_write_buffer;
 630
 631         BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
 632                !bch2_journal_error(&c->journal));
 633
 634         darray_exit(&wb->sorted);
 635         darray_exit(&wb->flushing.keys);
 636         darray_exit(&wb->inc.keys);
 637 }
 638
 639 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
 640 {
 641         struct btree_write_buffer *wb = &c->btree_write_buffer;
 642
 643         mutex_init(&wb->inc.lock);
 644         mutex_init(&wb->flushing.lock);
 645         INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
 646
 647         /* Will be resized by journal as needed: */
 648         unsigned initial_size = 1 << 16;
 649
 650         return  darray_make_room(&wb->inc.keys, initial_size) ?:
 651                 darray_make_room(&wb->flushing.keys, initial_size) ?:
 652                 darray_make_room(&wb->sorted, initial_size);
 653 }