fs/bcachefs/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 #include <linux/sort.h>
  15
  16 static int bch2_btree_write_buffer_journal_flush(struct journal *,
  17                                 struct journal_entry_pin *, u64);
  18
  19 static int bch2_journal_keys_to_write_buffer(struct bch_fs *, struct journal_buf *);
  20
  21 static inline bool __wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
  22 {
  23         return (cmp_int(l->hi, r->hi) ?:
  24                 cmp_int(l->mi, r->mi) ?:
  25                 cmp_int(l->lo, r->lo)) >= 0;
  26 }
  27
  28 static inline bool wb_key_ref_cmp(const struct wb_key_ref *l, const struct wb_key_ref *r)
  29 {
  30 #ifdef CONFIG_X86_64
  31         int cmp;
  32
  33         asm("mov   (%[l]), %%rax;"
  34             "sub   (%[r]), %%rax;"
  35             "mov  8(%[l]), %%rax;"
  36             "sbb  8(%[r]), %%rax;"
  37             "mov 16(%[l]), %%rax;"
  38             "sbb 16(%[r]), %%rax;"
  39             : "=@ccae" (cmp)
  40             : [l] "r" (l), [r] "r" (r)
  41             : "rax", "cc");
  42
  43         EBUG_ON(cmp != __wb_key_ref_cmp(l, r));
  44         return cmp;
  45 #else
  46         return __wb_key_ref_cmp(l, r);
  47 #endif
  48 }
  49
  50 static int wb_key_seq_cmp(const void *_l, const void *_r)
  51 {
  52         const struct btree_write_buffered_key *l = _l;
  53         const struct btree_write_buffered_key *r = _r;
  54
  55         return cmp_int(l->journal_seq, r->journal_seq);
  56 }
  57
  58 /* Compare excluding idx, the low 24 bits: */
  59 static inline bool wb_key_eq(const void *_l, const void *_r)
  60 {
  61         const struct wb_key_ref *l = _l;
  62         const struct wb_key_ref *r = _r;
  63
  64         return !((l->hi ^ r->hi)|
  65                  (l->mi ^ r->mi)|
  66                  ((l->lo >> 24) ^ (r->lo >> 24)));
  67 }
  68
  69 static noinline void wb_sort(struct wb_key_ref *base, size_t num)
  70 {
  71         size_t n = num, a = num / 2;
  72
  73         if (!a)         /* num < 2 || size == 0 */
  74                 return;
  75
  76         for (;;) {
  77                 size_t b, c, d;
  78
  79                 if (a)                  /* Building heap: sift down --a */
  80                         --a;
  81                 else if (--n)           /* Sorting: Extract root to --n */
  82                         swap(base[0], base[n]);
  83                 else                    /* Sort complete */
  84                         break;
  85
  86                 /*
  87                  * Sift element at "a" down into heap.  This is the
  88                  * "bottom-up" variant, which significantly reduces
  89                  * calls to cmp_func(): we find the sift-down path all
  90                  * the way to the leaves (one compare per level), then
  91                  * backtrack to find where to insert the target element.
  92                  *
  93                  * Because elements tend to sift down close to the leaves,
  94                  * this uses fewer compares than doing two per level
  95                  * on the way down.  (A bit more than half as many on
  96                  * average, 3/4 worst-case.)
  97                  */
  98                 for (b = a; c = 2*b + 1, (d = c + 1) < n;)
  99                         b = wb_key_ref_cmp(base + c, base + d) ? c : d;
 100                 if (d == n)             /* Special case last leaf with no sibling */
 101                         b = c;
 102
 103                 /* Now backtrack from "b" to the correct location for "a" */
 104                 while (b != a && wb_key_ref_cmp(base + a, base + b))
 105                         b = (b - 1) / 2;
 106                 c = b;                  /* Where "a" belongs */
 107                 while (b != a) {        /* Shift it into place */
 108                         b = (b - 1) / 2;
 109                         swap(base[b], base[c]);
 110                 }
 111         }
 112 }
 113
 114 static noinline int wb_flush_one_slowpath(struct btree_trans *trans,
 115                                           struct btree_iter *iter,
 116                                           struct btree_write_buffered_key *wb)
 117 {
 118         struct btree_path *path = btree_iter_path(trans, iter);
 119
 120         bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 121
 122         trans->journal_res.seq = wb->journal_seq;
 123
 124         return bch2_trans_update(trans, iter, &wb->k,
 125                                  BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) ?:
 126                 bch2_trans_commit(trans, NULL, NULL,
 127                                   BCH_TRANS_COMMIT_no_enospc|
 128                                   BCH_TRANS_COMMIT_no_check_rw|
 129                                   BCH_TRANS_COMMIT_no_journal_res|
 130                                   BCH_TRANS_COMMIT_journal_reclaim);
 131 }
 132
 133 static inline int wb_flush_one(struct btree_trans *trans, struct btree_iter *iter,
 134                                struct btree_write_buffered_key *wb,
 135                                bool *write_locked, size_t *fast)
 136 {
 137         struct btree_path *path;
 138         int ret;
 139
 140         EBUG_ON(!wb->journal_seq);
 141         EBUG_ON(!trans->c->btree_write_buffer.flushing.pin.seq);
 142         EBUG_ON(trans->c->btree_write_buffer.flushing.pin.seq > wb->journal_seq);
 143
 144         ret = bch2_btree_iter_traverse(iter);
 145         if (ret)
 146                 return ret;
 147
 148         /*
 149          * We can't clone a path that has write locks: unshare it now, before
 150          * set_pos and traverse():
 151          */
 152         if (btree_iter_path(trans, iter)->ref > 1)
 153                 iter->path = __bch2_btree_path_make_mut(trans, iter->path, true, _THIS_IP_);
 154
 155         path = btree_iter_path(trans, iter);
 156
 157         if (!*write_locked) {
 158                 ret = bch2_btree_node_lock_write(trans, path, &path->l[0].b->c);
 159                 if (ret)
 160                         return ret;
 161
 162                 bch2_btree_node_prep_for_write(trans, path, path->l[0].b);
 163                 *write_locked = true;
 164         }
 165
 166         if (unlikely(!bch2_btree_node_insert_fits(path->l[0].b, wb->k.k.u64s))) {
 167                 *write_locked = false;
 168                 return wb_flush_one_slowpath(trans, iter, wb);
 169         }
 170
 171         bch2_btree_insert_key_leaf(trans, path, &wb->k, wb->journal_seq);
 172         (*fast)++;
 173         return 0;
 174 }
 175
 176 /*
 177  * Update a btree with a write buffered key using the journal seq of the
 178  * original write buffer insert.
 179  *
 180  * It is not safe to rejournal the key once it has been inserted into the write
 181  * buffer because that may break recovery ordering. For example, the key may
 182  * have already been modified in the active write buffer in a seq that comes
 183  * before the current transaction. If we were to journal this key again and
 184  * crash, recovery would process updates in the wrong order.
 185  */
 186 static int
 187 btree_write_buffered_insert(struct btree_trans *trans,
 188                           struct btree_write_buffered_key *wb)
 189 {
 190         struct btree_iter iter;
 191         int ret;
 192
 193         bch2_trans_iter_init(trans, &iter, wb->btree, bkey_start_pos(&wb->k.k),
 194                              BTREE_ITER_CACHED|BTREE_ITER_INTENT);
 195
 196         trans->journal_res.seq = wb->journal_seq;
 197
 198         ret   = bch2_btree_iter_traverse(&iter) ?:
 199                 bch2_trans_update(trans, &iter, &wb->k,
 200                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 201         bch2_trans_iter_exit(trans, &iter);
 202         return ret;
 203 }
 204
 205 static void move_keys_from_inc_to_flushing(struct btree_write_buffer *wb)
 206 {
 207         struct bch_fs *c = container_of(wb, struct bch_fs, btree_write_buffer);
 208         struct journal *j = &c->journal;
 209
 210         if (!wb->inc.keys.nr)
 211                 return;
 212
 213         bch2_journal_pin_add(j, wb->inc.keys.data[0].journal_seq, &wb->flushing.pin,
 214                              bch2_btree_write_buffer_journal_flush);
 215
 216         darray_resize(&wb->flushing.keys, min_t(size_t, 1U << 20, wb->flushing.keys.nr + wb->inc.keys.nr));
 217         darray_resize(&wb->sorted, wb->flushing.keys.size);
 218
 219         if (!wb->flushing.keys.nr && wb->sorted.size >= wb->inc.keys.nr) {
 220                 swap(wb->flushing.keys, wb->inc.keys);
 221                 goto out;
 222         }
 223
 224         size_t nr = min(darray_room(wb->flushing.keys),
 225                         wb->sorted.size - wb->flushing.keys.nr);
 226         nr = min(nr, wb->inc.keys.nr);
 227
 228         memcpy(&darray_top(wb->flushing.keys),
 229                wb->inc.keys.data,
 230                sizeof(wb->inc.keys.data[0]) * nr);
 231
 232         memmove(wb->inc.keys.data,
 233                 wb->inc.keys.data + nr,
 234                sizeof(wb->inc.keys.data[0]) * (wb->inc.keys.nr - nr));
 235
 236         wb->flushing.keys.nr    += nr;
 237         wb->inc.keys.nr         -= nr;
 238 out:
 239         if (!wb->inc.keys.nr)
 240                 bch2_journal_pin_drop(j, &wb->inc.pin);
 241         else
 242                 bch2_journal_pin_update(j, wb->inc.keys.data[0].journal_seq, &wb->inc.pin,
 243                                         bch2_btree_write_buffer_journal_flush);
 244
 245         if (j->watermark) {
 246                 spin_lock(&j->lock);
 247                 bch2_journal_set_watermark(j);
 248                 spin_unlock(&j->lock);
 249         }
 250
 251         BUG_ON(wb->sorted.size < wb->flushing.keys.nr);
 252 }
 253
 254 static int bch2_btree_write_buffer_flush_locked(struct btree_trans *trans)
 255 {
 256         struct bch_fs *c = trans->c;
 257         struct journal *j = &c->journal;
 258         struct btree_write_buffer *wb = &c->btree_write_buffer;
 259         struct btree_iter iter = { NULL };
 260         size_t skipped = 0, fast = 0, slowpath = 0;
 261         bool write_locked = false;
 262         int ret = 0;
 263
 264         bch2_trans_unlock(trans);
 265         bch2_trans_begin(trans);
 266
 267         mutex_lock(&wb->inc.lock);
 268         move_keys_from_inc_to_flushing(wb);
 269         mutex_unlock(&wb->inc.lock);
 270
 271         for (size_t i = 0; i < wb->flushing.keys.nr; i++) {
 272                 wb->sorted.data[i].idx = i;
 273                 wb->sorted.data[i].btree = wb->flushing.keys.data[i].btree;
 274                 memcpy(&wb->sorted.data[i].pos, &wb->flushing.keys.data[i].k.k.p, sizeof(struct bpos));
 275         }
 276         wb->sorted.nr = wb->flushing.keys.nr;
 277
 278         /*
 279          * We first sort so that we can detect and skip redundant updates, and
 280          * then we attempt to flush in sorted btree order, as this is most
 281          * efficient.
 282          *
 283          * However, since we're not flushing in the order they appear in the
 284          * journal we won't be able to drop our journal pin until everything is
 285          * flushed - which means this could deadlock the journal if we weren't
 286          * passing BCH_TRANS_COMMIT_journal_reclaim. This causes the update to fail
 287          * if it would block taking a journal reservation.
 288          *
 289          * If that happens, simply skip the key so we can optimistically insert
 290          * as many keys as possible in the fast path.
 291          */
 292         wb_sort(wb->sorted.data, wb->sorted.nr);
 293
 294         darray_for_each(wb->sorted, i) {
 295                 struct btree_write_buffered_key *k = &wb->flushing.keys.data[i->idx];
 296
 297                 for (struct wb_key_ref *n = i + 1; n < min(i + 4, &darray_top(wb->sorted)); n++)
 298                         prefetch(&wb->flushing.keys.data[n->idx]);
 299
 300                 BUG_ON(!k->journal_seq);
 301
 302                 if (i + 1 < &darray_top(wb->sorted) &&
 303                     wb_key_eq(i, i + 1)) {
 304                         struct btree_write_buffered_key *n = &wb->flushing.keys.data[i[1].idx];
 305
 306                         skipped++;
 307                         n->journal_seq = min_t(u64, n->journal_seq, k->journal_seq);
 308                         k->journal_seq = 0;
 309                         continue;
 310                 }
 311
 312                 if (write_locked) {
 313                         struct btree_path *path = btree_iter_path(trans, &iter);
 314
 315                         if (path->btree_id != i->btree ||
 316                             bpos_gt(k->k.k.p, path->l[0].b->key.k.p)) {
 317                                 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 318                                 write_locked = false;
 319                         }
 320                 }
 321
 322                 if (!iter.path || iter.btree_id != k->btree) {
 323                         bch2_trans_iter_exit(trans, &iter);
 324                         bch2_trans_iter_init(trans, &iter, k->btree, k->k.k.p,
 325                                              BTREE_ITER_INTENT|BTREE_ITER_ALL_SNAPSHOTS);
 326                 }
 327
 328                 bch2_btree_iter_set_pos(&iter, k->k.k.p);
 329                 btree_iter_path(trans, &iter)->preserve = false;
 330
 331                 do {
 332                         if (race_fault()) {
 333                                 ret = -BCH_ERR_journal_reclaim_would_deadlock;
 334                                 break;
 335                         }
 336
 337                         ret = wb_flush_one(trans, &iter, k, &write_locked, &fast);
 338                         if (!write_locked)
 339                                 bch2_trans_begin(trans);
 340                 } while (bch2_err_matches(ret, BCH_ERR_transaction_restart));
 341
 342                 if (!ret) {
 343                         k->journal_seq = 0;
 344                 } else if (ret == -BCH_ERR_journal_reclaim_would_deadlock) {
 345                         slowpath++;
 346                         ret = 0;
 347                 } else
 348                         break;
 349         }
 350
 351         if (write_locked) {
 352                 struct btree_path *path = btree_iter_path(trans, &iter);
 353                 bch2_btree_node_unlock_write(trans, path, path->l[0].b);
 354         }
 355         bch2_trans_iter_exit(trans, &iter);
 356
 357         if (ret)
 358                 goto err;
 359
 360         if (slowpath) {
 361                 /*
 362                  * Flush in the order they were present in the journal, so that
 363                  * we can release journal pins:
 364                  * The fastpath zapped the seq of keys that were successfully flushed so
 365                  * we can skip those here.
 366                  */
 367                 trace_and_count(c, write_buffer_flush_slowpath, trans, slowpath, wb->flushing.keys.nr);
 368
 369                 sort(wb->flushing.keys.data,
 370                      wb->flushing.keys.nr,
 371                      sizeof(wb->flushing.keys.data[0]),
 372                      wb_key_seq_cmp, NULL);
 373
 374                 darray_for_each(wb->flushing.keys, i) {
 375                         if (!i->journal_seq)
 376                                 continue;
 377
 378                         bch2_journal_pin_update(j, i->journal_seq, &wb->flushing.pin,
 379                                                 bch2_btree_write_buffer_journal_flush);
 380
 381                         bch2_trans_begin(trans);
 382
 383                         ret = commit_do(trans, NULL, NULL,
 384                                         BCH_WATERMARK_reclaim|
 385                                         BCH_TRANS_COMMIT_no_check_rw|
 386                                         BCH_TRANS_COMMIT_no_enospc|
 387                                         BCH_TRANS_COMMIT_no_journal_res|
 388                                         BCH_TRANS_COMMIT_journal_reclaim,
 389                                         btree_write_buffered_insert(trans, i));
 390                         if (ret)
 391                                 goto err;
 392                 }
 393         }
 394 err:
 395         bch2_fs_fatal_err_on(ret, c, "%s", bch2_err_str(ret));
 396         trace_write_buffer_flush(trans, wb->flushing.keys.nr, skipped, fast, 0);
 397         bch2_journal_pin_drop(j, &wb->flushing.pin);
 398         wb->flushing.keys.nr = 0;
 399         return ret;
 400 }
 401
 402 static int fetch_wb_keys_from_journal(struct bch_fs *c, u64 seq)
 403 {
 404         struct journal *j = &c->journal;
 405         struct journal_buf *buf;
 406         int ret = 0;
 407
 408         while (!ret && (buf = bch2_next_write_buffer_flush_journal_buf(j, seq))) {
 409                 ret = bch2_journal_keys_to_write_buffer(c, buf);
 410                 mutex_unlock(&j->buf_lock);
 411         }
 412
 413         return ret;
 414 }
 415
 416 static int btree_write_buffer_flush_seq(struct btree_trans *trans, u64 seq)
 417 {
 418         struct bch_fs *c = trans->c;
 419         struct btree_write_buffer *wb = &c->btree_write_buffer;
 420         int ret = 0, fetch_from_journal_err;
 421
 422         do {
 423                 bch2_trans_unlock(trans);
 424
 425                 fetch_from_journal_err = fetch_wb_keys_from_journal(c, seq);
 426
 427                 /*
 428                  * On memory allocation failure, bch2_btree_write_buffer_flush_locked()
 429                  * is not guaranteed to empty wb->inc:
 430                  */
 431                 mutex_lock(&wb->flushing.lock);
 432                 ret = bch2_btree_write_buffer_flush_locked(trans);
 433                 mutex_unlock(&wb->flushing.lock);
 434         } while (!ret &&
 435                  (fetch_from_journal_err ||
 436                   (wb->inc.pin.seq && wb->inc.pin.seq <= seq) ||
 437                   (wb->flushing.pin.seq && wb->flushing.pin.seq <= seq)));
 438
 439         return ret;
 440 }
 441
 442 static int bch2_btree_write_buffer_journal_flush(struct journal *j,
 443                                 struct journal_entry_pin *_pin, u64 seq)
 444 {
 445         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 446
 447         return bch2_trans_run(c, btree_write_buffer_flush_seq(trans, seq));
 448 }
 449
 450 int bch2_btree_write_buffer_flush_sync(struct btree_trans *trans)
 451 {
 452         struct bch_fs *c = trans->c;
 453
 454         trace_and_count(c, write_buffer_flush_sync, trans, _RET_IP_);
 455
 456         return btree_write_buffer_flush_seq(trans, journal_cur_seq(&c->journal));
 457 }
 458
 459 int bch2_btree_write_buffer_flush_nocheck_rw(struct btree_trans *trans)
 460 {
 461         struct bch_fs *c = trans->c;
 462         struct btree_write_buffer *wb = &c->btree_write_buffer;
 463         int ret = 0;
 464
 465         if (mutex_trylock(&wb->flushing.lock)) {
 466                 ret = bch2_btree_write_buffer_flush_locked(trans);
 467                 mutex_unlock(&wb->flushing.lock);
 468         }
 469
 470         return ret;
 471 }
 472
 473 int bch2_btree_write_buffer_tryflush(struct btree_trans *trans)
 474 {
 475         struct bch_fs *c = trans->c;
 476
 477         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer))
 478                 return -BCH_ERR_erofs_no_writes;
 479
 480         int ret = bch2_btree_write_buffer_flush_nocheck_rw(trans);
 481         bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 482         return ret;
 483 }
 484
 485 static void bch2_btree_write_buffer_flush_work(struct work_struct *work)
 486 {
 487         struct bch_fs *c = container_of(work, struct bch_fs, btree_write_buffer.flush_work);
 488         struct btree_write_buffer *wb = &c->btree_write_buffer;
 489         int ret;
 490
 491         mutex_lock(&wb->flushing.lock);
 492         do {
 493                 ret = bch2_trans_run(c, bch2_btree_write_buffer_flush_locked(trans));
 494         } while (!ret && bch2_btree_write_buffer_should_flush(c));
 495         mutex_unlock(&wb->flushing.lock);
 496
 497         bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 498 }
 499
 500 int bch2_journal_key_to_wb_slowpath(struct bch_fs *c,
 501                              struct journal_keys_to_wb *dst,
 502                              enum btree_id btree, struct bkey_i *k)
 503 {
 504         struct btree_write_buffer *wb = &c->btree_write_buffer;
 505         int ret;
 506 retry:
 507         ret = darray_make_room_gfp(&dst->wb->keys, 1, GFP_KERNEL);
 508         if (!ret && dst->wb == &wb->flushing)
 509                 ret = darray_resize(&wb->sorted, wb->flushing.keys.size);
 510
 511         if (unlikely(ret)) {
 512                 if (dst->wb == &c->btree_write_buffer.flushing) {
 513                         mutex_unlock(&dst->wb->lock);
 514                         dst->wb = &c->btree_write_buffer.inc;
 515                         bch2_journal_pin_add(&c->journal, dst->seq, &dst->wb->pin,
 516                                              bch2_btree_write_buffer_journal_flush);
 517                         goto retry;
 518                 }
 519
 520                 return ret;
 521         }
 522
 523         dst->room = darray_room(dst->wb->keys);
 524         if (dst->wb == &wb->flushing)
 525                 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 526         BUG_ON(!dst->room);
 527         BUG_ON(!dst->seq);
 528
 529         struct btree_write_buffered_key *wb_k = &darray_top(dst->wb->keys);
 530         wb_k->journal_seq       = dst->seq;
 531         wb_k->btree             = btree;
 532         bkey_copy(&wb_k->k, k);
 533         dst->wb->keys.nr++;
 534         dst->room--;
 535         return 0;
 536 }
 537
 538 void bch2_journal_keys_to_write_buffer_start(struct bch_fs *c, struct journal_keys_to_wb *dst, u64 seq)
 539 {
 540         struct btree_write_buffer *wb = &c->btree_write_buffer;
 541
 542         if (mutex_trylock(&wb->flushing.lock)) {
 543                 mutex_lock(&wb->inc.lock);
 544                 move_keys_from_inc_to_flushing(wb);
 545
 546                 /*
 547                  * Attempt to skip wb->inc, and add keys directly to
 548                  * wb->flushing, saving us a copy later:
 549                  */
 550
 551                 if (!wb->inc.keys.nr) {
 552                         dst->wb = &wb->flushing;
 553                 } else {
 554                         mutex_unlock(&wb->flushing.lock);
 555                         dst->wb = &wb->inc;
 556                 }
 557         } else {
 558                 mutex_lock(&wb->inc.lock);
 559                 dst->wb = &wb->inc;
 560         }
 561
 562         dst->room = darray_room(dst->wb->keys);
 563         if (dst->wb == &wb->flushing)
 564                 dst->room = min(dst->room, wb->sorted.size - wb->flushing.keys.nr);
 565         dst->seq = seq;
 566
 567         bch2_journal_pin_add(&c->journal, seq, &dst->wb->pin,
 568                              bch2_btree_write_buffer_journal_flush);
 569 }
 570
 571 void bch2_journal_keys_to_write_buffer_end(struct bch_fs *c, struct journal_keys_to_wb *dst)
 572 {
 573         struct btree_write_buffer *wb = &c->btree_write_buffer;
 574
 575         if (!dst->wb->keys.nr)
 576                 bch2_journal_pin_drop(&c->journal, &dst->wb->pin);
 577
 578         if (bch2_btree_write_buffer_should_flush(c) &&
 579             __bch2_write_ref_tryget(c, BCH_WRITE_REF_btree_write_buffer) &&
 580             !queue_work(system_unbound_wq, &c->btree_write_buffer.flush_work))
 581                 bch2_write_ref_put(c, BCH_WRITE_REF_btree_write_buffer);
 582
 583         if (dst->wb == &wb->flushing)
 584                 mutex_unlock(&wb->flushing.lock);
 585         mutex_unlock(&wb->inc.lock);
 586 }
 587
 588 static int bch2_journal_keys_to_write_buffer(struct bch_fs *c, struct journal_buf *buf)
 589 {
 590         struct journal_keys_to_wb dst;
 591         int ret = 0;
 592
 593         bch2_journal_keys_to_write_buffer_start(c, &dst, le64_to_cpu(buf->data->seq));
 594
 595         for_each_jset_entry_type(entry, buf->data, BCH_JSET_ENTRY_write_buffer_keys) {
 596                 jset_entry_for_each_key(entry, k) {
 597                         ret = bch2_journal_key_to_wb(c, &dst, entry->btree_id, k);
 598                         if (ret)
 599                                 goto out;
 600                 }
 601
 602                 entry->type = BCH_JSET_ENTRY_btree_keys;
 603         }
 604
 605         spin_lock(&c->journal.lock);
 606         buf->need_flush_to_write_buffer = false;
 607         spin_unlock(&c->journal.lock);
 608 out:
 609         bch2_journal_keys_to_write_buffer_end(c, &dst);
 610         return ret;
 611 }
 612
 613 static int wb_keys_resize(struct btree_write_buffer_keys *wb, size_t new_size)
 614 {
 615         if (wb->keys.size >= new_size)
 616                 return 0;
 617
 618         if (!mutex_trylock(&wb->lock))
 619                 return -EINTR;
 620
 621         int ret = darray_resize(&wb->keys, new_size);
 622         mutex_unlock(&wb->lock);
 623         return ret;
 624 }
 625
 626 int bch2_btree_write_buffer_resize(struct bch_fs *c, size_t new_size)
 627 {
 628         struct btree_write_buffer *wb = &c->btree_write_buffer;
 629
 630         return wb_keys_resize(&wb->flushing, new_size) ?:
 631                 wb_keys_resize(&wb->inc, new_size);
 632 }
 633
 634 void bch2_fs_btree_write_buffer_exit(struct bch_fs *c)
 635 {
 636         struct btree_write_buffer *wb = &c->btree_write_buffer;
 637
 638         BUG_ON((wb->inc.keys.nr || wb->flushing.keys.nr) &&
 639                !bch2_journal_error(&c->journal));
 640
 641         darray_exit(&wb->sorted);
 642         darray_exit(&wb->flushing.keys);
 643         darray_exit(&wb->inc.keys);
 644 }
 645
 646 int bch2_fs_btree_write_buffer_init(struct bch_fs *c)
 647 {
 648         struct btree_write_buffer *wb = &c->btree_write_buffer;
 649
 650         mutex_init(&wb->inc.lock);
 651         mutex_init(&wb->flushing.lock);
 652         INIT_WORK(&wb->flush_work, bch2_btree_write_buffer_flush_work);
 653
 654         /* Will be resized by journal as needed: */
 655         unsigned initial_size = 1 << 16;
 656
 657         return  darray_make_room(&wb->inc.keys, initial_size) ?:
 658                 darray_make_room(&wb->flushing.keys, initial_size) ?:
 659                 darray_make_room(&wb->sorted, initial_size);
 660 }