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_ref_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_ref_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_ref_cmp(l, r));
  43         return cmp;
  44 #else
  45         return __wb_key_ref_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_ref_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_ref_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         struct btree_path *path = btree_iter_path(trans, iter);
 110
 111         bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 112
 113         trans->journal_res.seq = wb->journal_seq;
 114
 115         return bch2_trans_update(trans, iter, &wb->k,
 116                                  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 117                 bch2_trans_commit(trans, NULL, NULL,
 118                                   BCH_TRANS_COMMIT_no_enospc|
 119                                   BCH_TRANS_COMMIT_no_check_rw|
 120                                   BCH_TRANS_COMMIT_no_journal_res|
 121                                   BCH_TRANS_COMMIT_journal_reclaim);
 122 }
 123
 124 static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
 125                                struct btree_write_buffered_key *wb,
 126                                bool *write_locked, size_t *fast)
 127 {
 128         struct bch_fs *c = trans->c;
 129         struct btree_path *path;
 130         int ret;
 131
 132         EBUG_ON(!wb->journal_seq);
 133         EBUG_ON(!c->btree_write_buffer.flushing.pin.seq);
 134         EBUG_ON(c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);
 135
 136         ret = bch2_btree_iter_traverse(iter);
 137         if (ret)
 138                 return ret;
 139
 140         /*
 141          * We can't clone a path that has write locks: unshare it now, before
 142          * set_pos and traverse():
 143          */
 144         if (btree_iter_path(trans, iter)->ref > 1)
 145                 iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);
 146
 147         path = btree_iter_path(trans, iter);
 148
 149         if (!*write_locked) {
 150                 ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
 151                 if (ret)
 152                         return ret;
 153
 154                 bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
 155                 *write_locked = true;
 156         }
 157
 158         if (unlikely(!bch2_btree_node_insert_fits(c, path->l[0].b, wb->k.k.u64s))) {
 159                 *write_locked = false;
 160                 return wb_flush_one_slowpath(trans, iter, wb);
 161         }
 162
 163         bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
 164         (*fast)++;
 165         return 0;
 166 }
 167
 168 /*
 169  * Update a btree with a write buffered key using the journal seq of the
 170  * original write buffer insert.
 171  *
 172  * It is not safe to rejournal the key once it has been inserted into the write
 173  * buffer because that may break recovery ordering. For example, the key may
 174  * have already been modified in the active write buffer in a seq that comes
 175  * before the current transaction. If we were to journal this key again and
 176  * crash, recovery would process updates in the wrong order.
 177  */
 178 static int
 179 btree_write_buffered_insert(struct btree_trans *trans,
 180                           struct btree_write_buffered_key *wb)
 181 {
 182         struct btree_iter iter;
 183         int ret;
 184
 185         bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
 186                              BTREE_ITER_CACHED|BTREE_ITER_INTENT);
 187
 188         trans->journal_res.seq = wb->journal_seq;
 189
 190         ret   = bch2_btree_iter_traverse(&iter) ?:
 191                 bch2_trans_update(trans, &iter, &wb->k,
 192                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 193         bch2_trans_iter_exit(trans, &iter);
 194         return ret;
 195 }
 196
 197 static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
 198 {
 199         struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
 200         struct journal *j = &c->journal;
 201
 202         if (!wb->inc.keys.nr)
 203                 return;
 204
 205         bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
 206                              bch2_btree_write_buffer_journal_flush);
 207
 208         darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
 209         darray_resize(&wb->sorted, wb->flushing.keys.size);
 210
 211         if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
 212                 swap(wb->flushing.keys, wb->inc.keys);
 213                 goto out;
 214         }
 215
 216         size_t nr = min(darray_room(wb->flushing.keys),
 217                         wb->sorted.size - wb->flushing.keys.nr);
 218         nr = min(nr, wb->inc.keys.nr);
 219
 220         memcpy(&darray_top(wb->flushing.keys),
 221                wb->inc.keys.data,
 222                sizeof(wb->inc.keys.data[0]) * nr);
 223
 224         memmove(wb->inc.keys.data,
 225                 wb->inc.keys.data + nr,
 226                sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));
 227
 228         wb->flushing.keys.nr    += nr;
 229         wb->inc.keys.nr         -= nr;
 230 out:
 231         if (!wb->inc.keys.nr)
 232                 bch2_journal_pin_drop(j, &wb->inc.pin);
 233         else
 234                 bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
 235                                         bch2_btree_write_buffer_journal_flush);
 236
 237         if (j->watermark) {
 238                 spin_lock(&j->lock);
 239                 bch2_journal_set_watermark(j);
 240                 spin_unlock(&j->lock);
 241         }
 242
 243         BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
 244 }
 245
 246 static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
 247 {
 248         struct bch_fs *c = trans->c;
 249         struct journal *j = &c->journal;
 250         struct btree_write_buffer *wb = &c->btree_write_buffer;
 251         struct btree_iter iter = { NULL };
 252         size_t skipped = 0, fast = 0, slowpath = 0;
 253         bool write_locked = false;
 254         int ret = 0;
 255
 256         bch2_trans_unlock(trans);
 257         bch2_trans_begin(trans);
 258
 259         mutex_lock(&wb->inc.lock);
 260         move_keys_from_inc_to_flushing(wb);
 261         mutex_unlock(&wb->inc.lock);
 262
 263         for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
 264                 wb->sorted.data[i].idx = i;
 265                 wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
 266                 memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
 267         }
 268         wb->sorted.nr = wb->flushing.keys.nr;
 269
 270         /*
 271          * We first sort so that we can detect and skip redundant updates, and
 272          * then we attempt to flush in sorted btree order, as this is most
 273          * efficient.
 274          *
 275          * However, since we're not flushing in the order they appear in the
 276          * journal we won't be able to drop our journal pin until everything is
 277          * flushed - which means this could deadlock the journal if we weren't
 278          * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
 279          * if it would block taking a journal reservation.
 280          *
 281          * If that happens, simply skip the key so we can optimistically insert
 282          * as many keys as possible in the fast path.
 283          */
 284         wb_sort(wb->sorted.data, wb->sorted.nr);
 285
 286         darray_for_each(wb->sorted, i) {
 287                 struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
 288
 289                 for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
 290                         prefetch(&wb->flushing.keys.data[n->idx]);
 291
 292                 BUG_ON(!k->journal_seq);
 293
 294                 if (i + 1 < &darray_top(wb->sorted) &&
 295                     wb_key_eq(i, i + 1)) {
 296                         struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];
 297
 298                         skipped++;
 299                         n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
 300                         k->journal_seq = 0;
 301                         continue;
 302                 }
 303
 304                 if (write_locked) {
 305                         struct btree_path *path = btree_iter_path(trans, &iter);
 306
 307                         if (path->btree_id != i->btree ||
 308                             bpos_gt(k->k.k.p, path->l[0].b->key.k.p)) {
 309                                 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 310                                 write_locked = false;
 311                         }
 312                 }
 313
 314                 if (!iter.path || iter.btree_id != k->btree) {
 315                         bch2_trans_iter_exit(trans, &iter);
 316                         bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
 317                                              BTREE_ITER_INTENT|BTREE_ITER_ALL_SNAPSHOTS);
 318                 }
 319
 320                 bch2_btree_iter_set_pos(&iter, k->k.k.p);
 321                 btree_iter_path(trans, &iter)->preserve = false;
 322
 323                 do {
 324                         if (race_fault()) {
 325                                 ret = -BCH_ERR_journal_reclaim_would_deadlock;
 326                                 break;
 327                         }
 328
 329                         ret = wb_flush_one(trans, &iter, k, &write_locked, &fast);
 330                         if (!write_locked)
 331                                 bch2_trans_begin(trans);
 332                 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
 333
 334                 if (!ret) {
 335                         k->journal_seq = 0;
 336                 } else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
 337                         slowpath++;
 338                         ret = 0;
 339                 } else
 340                         break;
 341         }
 342
 343         if (write_locked) {
 344                 struct btree_path *path = btree_iter_path(trans, &iter);
 345                 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 346         }
 347         bch2_trans_iter_exit(trans, &iter);
 348
 349         if (ret)
 350                 goto err;
 351
 352         if (slowpath) {
 353                 /*
 354                  * Flush in the order they were present in the journal, so that
 355                  * we can release journal pins:
 356                  * The fastpath zapped the seq of keys that were successfully flushed so
 357                  * we can skip those here.
 358                  */
 359                 trace_and_count(c, write_buffer_flush_slowpath, trans, slowpath, wb->flushing.keys.nr);
 360
 361                 darray_for_each(wb->flushing.keys, i) {
 362                         if (!i->journal_seq)
 363                                 continue;
 364
 365                         bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
 366                                                 bch2_btree_write_buffer_journal_flush);
 367
 368                         bch2_trans_begin(trans);
 369
 370                         ret = commit_do(trans, NULL, NULL,
 371                                         BCH_WATERMARK_reclaim|
 372                                         BCH_TRANS_COMMIT_no_check_rw|
 373                                         BCH_TRANS_COMMIT_no_enospc|
 374                                         BCH_TRANS_COMMIT_no_journal_res|
 375                                         BCH_TRANS_COMMIT_journal_reclaim,
 376                                         btree_write_buffered_insert(trans, i));
 377                         if (ret)
 378                                 goto err;
 379                 }
 380         }
 381 err:
 382         bch2_fs_fatal_err_on(ret, c, "%s: insert error %s", __func__, bch2_err_str(ret));
 383         trace_write_buffer_flush(trans, wb->flushing.keys.nr, skipped, fast, 0);
 384         bch2_journal_pin_drop(j, &wb->flushing.pin);
 385         wb->flushing.keys.nr = 0;
 386         return ret;
 387 }
 388
 389 static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
 390 {
 391         struct journal *j = &c->journal;
 392         struct journal_buf *buf;
 393         int ret = 0;
 394
 395         while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, seq))) {
 396                 ret = bch2_journal_keys_to_write_buffer(c, buf);
 397                 mutex_unlock(&j->buf_lock);
 398         }
 399
 400         return ret;
 401 }
 402
 403 static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq)
 404 {
 405         struct bch_fs *c = trans->c;
 406         struct btree_write_buffer *wb = &c->btree_write_buffer;
 407         int ret = 0, fetch_from_journal_err;
 408
 409         do {
 410                 bch2_trans_unlock(trans);
 411
 412                 fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
 413
 414                 /*
 415                  * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
 416                  * is not guaranteed to empty wb->inc:
 417                  */
 418                 mutex_lock(&wb->flushing.lock);
 419                 ret = bch2_btree_write_buffer_flush_locked(trans);
 420                 mutex_unlock(&wb->flushing.lock);
 421         } while (!ret &&
 422                  (fetch_from_journal_err ||
 423                   (wb->inc.pin.seq && wb->inc.pin.seq <= seq) ||
 424                   (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq)));
 425
 426         return ret;
 427 }
 428
 429 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
 430                                 struct journal_entry_pin *_pin, u64 seq)
 431 {
 432         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 433
 434         return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq));
 435 }
 436
 437 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 438 {
 439         struct bch_fs *c = trans->c;
 440
 441         trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);
 442
 443         return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal));
 444 }
 445
 446 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
 447 {
 448         struct bch_fs *c = trans->c;
 449         struct btree_write_buffer *wb = &c->btree_write_buffer;
 450         int ret = 0;
 451
 452         if (mutex_trylock(&wb->flushing.lock)) {
 453                 ret = bch2_btree_write_buffer_flush_locked(trans);
 454                 mutex_unlock(&wb->flushing.lock);
 455         }
 456
 457         return ret;
 458 }
 459
 460 int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
 461 {
 462         struct bch_fs *c = trans->c;
 463
 464         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer))
 465                 return -BCH_ERR_erofs_no_writes;
 466
 467         int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
 468         bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 469         return ret;
 470 }
 471
 472 static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
 473 {
 474         struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
 475         struct btree_write_buffer *wb = &c->btree_write_buffer;
 476         int ret;
 477
 478         mutex_lock(&wb->flushing.lock);
 479         do {
 480                 ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
 481         } while (!ret && bch2_btree_write_buffer_should_flush(c));
 482         mutex_unlock(&wb->flushing.lock);
 483
 484         bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 485 }
 486
 487 int bch2_journal_key_to_wb_slowpath(struct bch_fs *c,
 488                              struct journal_keys_to_wb *dst,
 489                              enum btree_id btree, struct bkey_i *k)
 490 {
 491         struct btree_write_buffer *wb = &c->btree_write_buffer;
 492         int ret;
 493 retry:
 494         ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
 495         if (!ret && dst->wb == &wb->flushing)
 496                 ret = darray_resize(&wb->sorted, wb->flushing.keys.size);
 497
 498         if (unlikely(ret)) {
 499                 if (dst->wb == &c->btree_write_buffer.flushing) {
 500                         mutex_unlock(&dst->wb->lock);
 501                         dst->wb = &c->btree_write_buffer.inc;
 502                         bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
 503                                              bch2_btree_write_buffer_journal_flush);
 504                         goto retry;
 505                 }
 506
 507                 return ret;
 508         }
 509
 510         dst->room = darray_room(dst->wb->keys);
 511         if (dst->wb == &wb->flushing)
 512                 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 513         BUG_ON(!dst->room);
 514         BUG_ON(!dst->seq);
 515
 516         struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
 517         wb_k->journal_seq       = dst->seq;
 518         wb_k->btree             = btree;
 519         bkey_copy(&wb_k->k, k);
 520         dst->wb->keys.nr++;
 521         dst->room--;
 522         return 0;
 523 }
 524
 525 void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
 526 {
 527         struct btree_write_buffer *wb = &c->btree_write_buffer;
 528
 529         if (mutex_trylock(&wb->flushing.lock)) {
 530                 mutex_lock(&wb->inc.lock);
 531                 move_keys_from_inc_to_flushing(wb);
 532
 533                 /*
 534                  * Attempt to skip wb->inc, and add keys directly to
 535                  * wb->flushing, saving us a copy later:
 536                  */
 537
 538                 if (!wb->inc.keys.nr) {
 539                         dst->wb = &wb->flushing;
 540                 } else {
 541                         mutex_unlock(&wb->flushing.lock);
 542                         dst->wb = &wb->inc;
 543                 }
 544         } else {
 545                 mutex_lock(&wb->inc.lock);
 546                 dst->wb = &wb->inc;
 547         }
 548
 549         dst->room = darray_room(dst->wb->keys);
 550         if (dst->wb == &wb->flushing)
 551                 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 552         dst->seq = seq;
 553
 554         bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
 555                              bch2_btree_write_buffer_journal_flush);
 556 }
 557
 558 void bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
 559 {
 560         struct btree_write_buffer *wb = &c->btree_write_buffer;
 561
 562         if (!dst->wb->keys.nr)
 563                 bch2_journal_pin_drop(&c->journal, &dst->wb->pin);
 564
 565         if (bch2_btree_write_buffer_should_flush(c) &&
 566             __bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer) &&
 567             !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
 568                 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 569
 570         if (dst->wb == &wb->flushing)
 571                 mutex_unlock(&wb->flushing.lock);
 572         mutex_unlock(&wb->inc.lock);
 573 }
 574
 575 static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
 576 {
 577         struct journal_keys_to_wb dst;
 578         struct jset_entry *entry;
 579         struct bkey_i *k;
 580         int ret = 0;
 581
 582         bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
 583
 584         for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
 585                 jset_entry_for_each_key(entry, k) {
 586                         ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
 587                         if (ret)
 588                                 goto out;
 589                 }
 590
 591                 entry->type = BCH_JSET_ENTRY_btree_keys;
 592         }
 593
 594         buf->need_flush_to_write_buffer = false;
 595 out:
 596         bch2_journal_keys_to_write_buffer_end(c, &dst);
 597         return ret;
 598 }
 599
 600 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
 601 {
 602         if (wb->keys.size >= new_size)
 603                 return 0;
 604
 605         if (!mutex_trylock(&wb->lock))
 606                 return -EINTR;
 607
 608         int ret = darray_resize(&wb->keys, new_size);
 609         mutex_unlock(&wb->lock);
 610         return ret;
 611 }
 612
 613 int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
 614 {
 615         struct btree_write_buffer *wb = &c->btree_write_buffer;
 616
 617         return wb_keys_resize(&wb->flushing, new_size) ?:
 618                 wb_keys_resize(&wb->inc, new_size);
 619 }
 620
 621 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
 622 {
 623         struct btree_write_buffer *wb = &c->btree_write_buffer;
 624
 625         BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
 626                !bch2_journal_error(&c->journal));
 627
 628         darray_exit(&wb->sorted);
 629         darray_exit(&wb->flushing.keys);
 630         darray_exit(&wb->inc.keys);
 631 }
 632
 633 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
 634 {
 635         struct btree_write_buffer *wb = &c->btree_write_buffer;
 636
 637         mutex_init(&wb->inc.lock);
 638         mutex_init(&wb->flushing.lock);
 639         INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
 640
 641         /* Will be resized by journal as needed: */
 642         unsigned initial_size = 1 << 16;
 643
 644         return  darray_make_room(&wb->inc.keys, initial_size) ?:
 645                 darray_make_room(&wb->flushing.keys, initial_size) ?:
 646                 darray_make_room(&wb->sorted, initial_size);
 647 }