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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_iter.h"
   6 #include "btree_journal_iter.h"
   7 #include "btree_locking.h"
   8 #include "buckets.h"
   9 #include "debug.h"
  10 #include "errcode.h"
  11 #include "error.h"
  12 #include "extents.h"
  13 #include "keylist.h"
  14 #include "subvolume.h"
  15 #include "trace.h"
  16
  17 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  18                                          const struct btree_insert_entry *r)
  19 {
  20         return   cmp_int(l->btree_id,   r->btree_id) ?:
  21                  cmp_int(l->cached,     r->cached) ?:
  22                  -cmp_int(l->level,     r->level) ?:
  23                  bpos_cmp(l->k->k.p,    r->k->k.p);
  24 }
  25
  26 static int __must_check
  27 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  28                           struct bkey_i *, enum btree_update_flags,
  29                           unsigned long ip);
  30
  31 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
  32                                           enum btree_id id,
  33                                           struct bpos pos)
  34 {
  35         struct bch_fs *c = trans->c;
  36         struct btree_iter iter;
  37         struct bkey_s_c k;
  38         int ret;
  39
  40         bch2_trans_iter_init(trans, &iter, id, pos,
  41                              BTREE_ITER_NOT_EXTENTS|
  42                              BTREE_ITER_ALL_SNAPSHOTS);
  43         while (1) {
  44                 k = bch2_btree_iter_prev(&iter);
  45                 ret = bkey_err(k);
  46                 if (ret)
  47                         break;
  48
  49                 if (!k.k)
  50                         break;
  51
  52                 if (!bkey_eq(pos, k.k->p))
  53                         break;
  54
  55                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
  56                         ret = 1;
  57                         break;
  58                 }
  59         }
  60         bch2_trans_iter_exit(trans, &iter);
  61
  62         return ret;
  63 }
  64
  65 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
  66                                           enum btree_id id,
  67                                           struct bpos pos)
  68 {
  69         if (!btree_type_has_snapshots(id) ||
  70             bch2_snapshot_is_leaf(trans->c, pos.snapshot))
  71                 return 0;
  72
  73         return __check_pos_snapshot_overwritten(trans, id, pos);
  74 }
  75
  76 static noinline int extent_front_merge(struct btree_trans *trans,
  77                                        struct btree_iter *iter,
  78                                        struct bkey_s_c k,
  79                                        struct bkey_i **insert,
  80                                        enum btree_update_flags flags)
  81 {
  82         struct bch_fs *c = trans->c;
  83         struct bkey_i *update;
  84         int ret;
  85
  86         update = bch2_bkey_make_mut_noupdate(trans, k);
  87         ret = PTR_ERR_OR_ZERO(update);
  88         if (ret)
  89                 return ret;
  90
  91         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
  92                 return 0;
  93
  94         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
  95                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
  96         if (ret < 0)
  97                 return ret;
  98         if (ret)
  99                 return 0;
 100
 101         ret = bch2_btree_delete_at(trans, iter, flags);
 102         if (ret)
 103                 return ret;
 104
 105         *insert = update;
 106         return 0;
 107 }
 108
 109 static noinline int extent_back_merge(struct btree_trans *trans,
 110                                       struct btree_iter *iter,
 111                                       struct bkey_i *insert,
 112                                       struct bkey_s_c k)
 113 {
 114         struct bch_fs *c = trans->c;
 115         int ret;
 116
 117         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
 118                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
 119         if (ret < 0)
 120                 return ret;
 121         if (ret)
 122                 return 0;
 123
 124         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
 125         return 0;
 126 }
 127
 128 /*
 129  * When deleting, check if we need to emit a whiteout (because we're overwriting
 130  * something in an ancestor snapshot)
 131  */
 132 static int need_whiteout_for_snapshot(struct btree_trans *trans,
 133                                       enum btree_id btree_id, struct bpos pos)
 134 {
 135         struct btree_iter iter;
 136         struct bkey_s_c k;
 137         u32 snapshot = pos.snapshot;
 138         int ret;
 139
 140         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
 141                 return 0;
 142
 143         pos.snapshot++;
 144
 145         for_each_btree_key_norestart(trans, iter, btree_id, pos,
 146                            BTREE_ITER_ALL_SNAPSHOTS|
 147                            BTREE_ITER_NOPRESERVE, k, ret) {
 148                 if (!bkey_eq(k.k->p, pos))
 149                         break;
 150
 151                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
 152                                               k.k->p.snapshot)) {
 153                         ret = !bkey_whiteout(k.k);
 154                         break;
 155                 }
 156         }
 157         bch2_trans_iter_exit(trans, &iter);
 158
 159         return ret;
 160 }
 161
 162 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
 163                                    enum btree_id id,
 164                                    struct bpos old_pos,
 165                                    struct bpos new_pos)
 166 {
 167         struct bch_fs *c = trans->c;
 168         struct btree_iter old_iter, new_iter = { NULL };
 169         struct bkey_s_c old_k, new_k;
 170         snapshot_id_list s;
 171         struct bkey_i *update;
 172         int ret;
 173
 174         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
 175                 return 0;
 176
 177         darray_init(&s);
 178
 179         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
 180                              BTREE_ITER_NOT_EXTENTS|
 181                              BTREE_ITER_ALL_SNAPSHOTS);
 182         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
 183                !(ret = bkey_err(old_k)) &&
 184                bkey_eq(old_pos, old_k.k->p)) {
 185                 struct bpos whiteout_pos =
 186                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
 187
 188                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
 189                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
 190                         continue;
 191
 192                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
 193                                            BTREE_ITER_NOT_EXTENTS|
 194                                            BTREE_ITER_INTENT);
 195                 ret = bkey_err(new_k);
 196                 if (ret)
 197                         break;
 198
 199                 if (new_k.k->type == KEY_TYPE_deleted) {
 200                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 201                         ret = PTR_ERR_OR_ZERO(update);
 202                         if (ret)
 203                                 break;
 204
 205                         bkey_init(&update->k);
 206                         update->k.p             = whiteout_pos;
 207                         update->k.type          = KEY_TYPE_whiteout;
 208
 209                         ret = bch2_trans_update(trans, &new_iter, update,
 210                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 211                 }
 212                 bch2_trans_iter_exit(trans, &new_iter);
 213
 214                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
 215                 if (ret)
 216                         break;
 217         }
 218         bch2_trans_iter_exit(trans, &new_iter);
 219         bch2_trans_iter_exit(trans, &old_iter);
 220         darray_exit(&s);
 221
 222         return ret;
 223 }
 224
 225 int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
 226                                        struct btree_iter *iter,
 227                                        enum btree_update_flags flags,
 228                                        struct bkey_s_c old,
 229                                        struct bkey_s_c new)
 230 {
 231         enum btree_id btree_id = iter->btree_id;
 232         struct bkey_i *update;
 233         struct bpos new_start = bkey_start_pos(new.k);
 234         bool front_split = bkey_lt(bkey_start_pos(old.k), new_start);
 235         bool back_split  = bkey_gt(old.k->p, new.k->p);
 236         int ret = 0, compressed_sectors;
 237
 238         /*
 239          * If we're going to be splitting a compressed extent, note it
 240          * so that __bch2_trans_commit() can increase our disk
 241          * reservation:
 242          */
 243         if (((front_split && back_split) ||
 244              ((front_split || back_split) && old.k->p.snapshot != new.k->p.snapshot)) &&
 245             (compressed_sectors = bch2_bkey_sectors_compressed(old)))
 246                 trans->extra_journal_res += compressed_sectors;
 247
 248         if (front_split) {
 249                 update = bch2_bkey_make_mut_noupdate(trans, old);
 250                 if ((ret = PTR_ERR_OR_ZERO(update)))
 251                         return ret;
 252
 253                 bch2_cut_back(new_start, update);
 254
 255                 ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
 256                                         old.k->p, update->k.p) ?:
 257                         bch2_btree_insert_nonextent(trans, btree_id, update,
 258                                         BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 259                 if (ret)
 260                         return ret;
 261         }
 262
 263         /* If we're overwriting in a different snapshot - middle split: */
 264         if (old.k->p.snapshot != new.k->p.snapshot &&
 265             (front_split || back_split)) {
 266                 update = bch2_bkey_make_mut_noupdate(trans, old);
 267                 if ((ret = PTR_ERR_OR_ZERO(update)))
 268                         return ret;
 269
 270                 bch2_cut_front(new_start, update);
 271                 bch2_cut_back(new.k->p, update);
 272
 273                 ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
 274                                         old.k->p, update->k.p) ?:
 275                         bch2_btree_insert_nonextent(trans, btree_id, update,
 276                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 277                 if (ret)
 278                         return ret;
 279         }
 280
 281         if (bkey_le(old.k->p, new.k->p)) {
 282                 update = bch2_trans_kmalloc(trans, sizeof(*update));
 283                 if ((ret = PTR_ERR_OR_ZERO(update)))
 284                         return ret;
 285
 286                 bkey_init(&update->k);
 287                 update->k.p = old.k->p;
 288                 update->k.p.snapshot = new.k->p.snapshot;
 289
 290                 if (new.k->p.snapshot != old.k->p.snapshot) {
 291                         update->k.type = KEY_TYPE_whiteout;
 292                 } else if (btree_type_has_snapshots(btree_id)) {
 293                         ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
 294                         if (ret < 0)
 295                                 return ret;
 296                         if (ret)
 297                                 update->k.type = KEY_TYPE_whiteout;
 298                 }
 299
 300                 ret = bch2_btree_insert_nonextent(trans, btree_id, update,
 301                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 302                 if (ret)
 303                         return ret;
 304         }
 305
 306         if (back_split) {
 307                 update = bch2_bkey_make_mut_noupdate(trans, old);
 308                 if ((ret = PTR_ERR_OR_ZERO(update)))
 309                         return ret;
 310
 311                 bch2_cut_front(new.k->p, update);
 312
 313                 ret = bch2_trans_update_by_path(trans, iter->path, update,
 314                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 315                                           flags, _RET_IP_);
 316                 if (ret)
 317                         return ret;
 318         }
 319
 320         return 0;
 321 }
 322
 323 static int bch2_trans_update_extent(struct btree_trans *trans,
 324                                     struct btree_iter *orig_iter,
 325                                     struct bkey_i *insert,
 326                                     enum btree_update_flags flags)
 327 {
 328         struct btree_iter iter;
 329         struct bkey_s_c k;
 330         enum btree_id btree_id = orig_iter->btree_id;
 331         int ret = 0;
 332
 333         bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
 334                              BTREE_ITER_INTENT|
 335                              BTREE_ITER_WITH_UPDATES|
 336                              BTREE_ITER_NOT_EXTENTS);
 337         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
 338         if ((ret = bkey_err(k)))
 339                 goto err;
 340         if (!k.k)
 341                 goto out;
 342
 343         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
 344                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
 345                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
 346                         if (ret)
 347                                 goto err;
 348                 }
 349
 350                 goto next;
 351         }
 352
 353         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
 354                 bool done = bkey_lt(insert->k.p, k.k->p);
 355
 356                 ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
 357                 if (ret)
 358                         goto err;
 359
 360                 if (done)
 361                         goto out;
 362 next:
 363                 bch2_btree_iter_advance(&iter);
 364                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
 365                 if ((ret = bkey_err(k)))
 366                         goto err;
 367                 if (!k.k)
 368                         goto out;
 369         }
 370
 371         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
 372                 ret = extent_back_merge(trans, &iter, insert, k);
 373                 if (ret)
 374                         goto err;
 375         }
 376 out:
 377         if (!bkey_deleted(&insert->k))
 378                 ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
 379 err:
 380         bch2_trans_iter_exit(trans, &iter);
 381
 382         return ret;
 383 }
 384
 385 static noinline int flush_new_cached_update(struct btree_trans *trans,
 386                                             struct btree_path *path,
 387                                             struct btree_insert_entry *i,
 388                                             enum btree_update_flags flags,
 389                                             unsigned long ip)
 390 {
 391         struct btree_path *btree_path;
 392         struct bkey k;
 393         int ret;
 394
 395         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
 396                                    BTREE_ITER_INTENT, _THIS_IP_);
 397         ret = bch2_btree_path_traverse(trans, btree_path, 0);
 398         if (ret)
 399                 goto out;
 400
 401         /*
 402          * The old key in the insert entry might actually refer to an existing
 403          * key in the btree that has been deleted from cache and not yet
 404          * flushed. Check for this and skip the flush so we don't run triggers
 405          * against a stale key.
 406          */
 407         bch2_btree_path_peek_slot_exact(btree_path, &k);
 408         if (!bkey_deleted(&k))
 409                 goto out;
 410
 411         i->key_cache_already_flushed = true;
 412         i->flags |= BTREE_TRIGGER_NORUN;
 413
 414         btree_path_set_should_be_locked(btree_path);
 415         ret = bch2_trans_update_by_path(trans, btree_path, i->k, flags, ip);
 416 out:
 417         bch2_path_put(trans, btree_path, true);
 418         return ret;
 419 }
 420
 421 static int __must_check
 422 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
 423                           struct bkey_i *k, enum btree_update_flags flags,
 424                           unsigned long ip)
 425 {
 426         struct bch_fs *c = trans->c;
 427         struct btree_insert_entry *i, n;
 428         u64 seq = 0;
 429         int cmp;
 430
 431         EBUG_ON(!path->should_be_locked);
 432         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
 433         EBUG_ON(!bpos_eq(k->k.p, path->pos));
 434
 435         /*
 436          * The transaction journal res hasn't been allocated at this point.
 437          * That occurs at commit time. Reuse the seq field to pass in the seq
 438          * of a prejournaled key.
 439          */
 440         if (flags & BTREE_UPDATE_PREJOURNAL)
 441                 seq = trans->journal_res.seq;
 442
 443         n = (struct btree_insert_entry) {
 444                 .flags          = flags,
 445                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
 446                 .btree_id       = path->btree_id,
 447                 .level          = path->level,
 448                 .cached         = path->cached,
 449                 .path           = path,
 450                 .k              = k,
 451                 .seq            = seq,
 452                 .ip_allocated   = ip,
 453         };
 454
 455 #ifdef CONFIG_BCACHEFS_DEBUG
 456         trans_for_each_update(trans, i)
 457                 BUG_ON(i != trans->updates &&
 458                        btree_insert_entry_cmp(i - 1, i) >= 0);
 459 #endif
 460
 461         /*
 462          * Pending updates are kept sorted: first, find position of new update,
 463          * then delete/trim any updates the new update overwrites:
 464          */
 465         trans_for_each_update(trans, i) {
 466                 cmp = btree_insert_entry_cmp(&n, i);
 467                 if (cmp <= 0)
 468                         break;
 469         }
 470
 471         if (!cmp && i < trans->updates + trans->nr_updates) {
 472                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
 473
 474                 bch2_path_put(trans, i->path, true);
 475                 i->flags        = n.flags;
 476                 i->cached       = n.cached;
 477                 i->k            = n.k;
 478                 i->path         = n.path;
 479                 i->seq          = n.seq;
 480                 i->ip_allocated = n.ip_allocated;
 481         } else {
 482                 array_insert_item(trans->updates, trans->nr_updates,
 483                                   i - trans->updates, n);
 484
 485                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
 486                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
 487
 488                 if (unlikely(trans->journal_replay_not_finished)) {
 489                         struct bkey_i *j_k =
 490                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
 491
 492                         if (j_k) {
 493                                 i->old_k = j_k->k;
 494                                 i->old_v = &j_k->v;
 495                         }
 496                 }
 497         }
 498
 499         __btree_path_get(i->path, true);
 500
 501         /*
 502          * If a key is present in the key cache, it must also exist in the
 503          * btree - this is necessary for cache coherency. When iterating over
 504          * a btree that's cached in the key cache, the btree iter code checks
 505          * the key cache - but the key has to exist in the btree for that to
 506          * work:
 507          */
 508         if (path->cached && bkey_deleted(&i->old_k))
 509                 return flush_new_cached_update(trans, path, i, flags, ip);
 510
 511         return 0;
 512 }
 513
 514 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 515                                    struct bkey_i *k, enum btree_update_flags flags)
 516 {
 517         struct btree_path *path = iter->update_path ?: iter->path;
 518         struct bkey_cached *ck;
 519         int ret;
 520
 521         if (iter->flags & BTREE_ITER_IS_EXTENTS)
 522                 return bch2_trans_update_extent(trans, iter, k, flags);
 523
 524         if (bkey_deleted(&k->k) &&
 525             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
 526             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
 527                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
 528                 if (unlikely(ret < 0))
 529                         return ret;
 530
 531                 if (ret)
 532                         k->k.type = KEY_TYPE_whiteout;
 533         }
 534
 535         /*
 536          * Ensure that updates to cached btrees go to the key cache:
 537          */
 538         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
 539             !path->cached &&
 540             !path->level &&
 541             btree_id_cached(trans->c, path->btree_id)) {
 542                 if (!iter->key_cache_path ||
 543                     !iter->key_cache_path->should_be_locked ||
 544                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
 545                         if (!iter->key_cache_path)
 546                                 iter->key_cache_path =
 547                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
 548                                                       BTREE_ITER_INTENT|
 549                                                       BTREE_ITER_CACHED, _THIS_IP_);
 550
 551                         iter->key_cache_path =
 552                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
 553                                                         iter->flags & BTREE_ITER_INTENT,
 554                                                         _THIS_IP_);
 555
 556                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
 557                                                        BTREE_ITER_CACHED);
 558                         if (unlikely(ret))
 559                                 return ret;
 560
 561                         ck = (void *) iter->key_cache_path->l[0].b;
 562
 563                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 564                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
 565                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 566                         }
 567
 568                         btree_path_set_should_be_locked(iter->key_cache_path);
 569                 }
 570
 571                 path = iter->key_cache_path;
 572         }
 573
 574         return bch2_trans_update_by_path(trans, path, k, flags, _RET_IP_);
 575 }
 576
 577 /*
 578  * Add a transaction update for a key that has already been journaled.
 579  */
 580 int __must_check bch2_trans_update_seq(struct btree_trans *trans, u64 seq,
 581                                        struct btree_iter *iter, struct bkey_i *k,
 582                                        enum btree_update_flags flags)
 583 {
 584         trans->journal_res.seq = seq;
 585         return bch2_trans_update(trans, iter, k, flags|BTREE_UPDATE_NOJOURNAL|
 586                                                  BTREE_UPDATE_PREJOURNAL);
 587 }
 588
 589 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
 590                                             enum btree_id btree,
 591                                             struct bkey_i *k)
 592 {
 593         struct btree_write_buffered_key *i;
 594         int ret;
 595
 596         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
 597         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
 598
 599         trans_for_each_wb_update(trans, i) {
 600                 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
 601                         bkey_copy(&i->k, k);
 602                         return 0;
 603                 }
 604         }
 605
 606         if (!trans->wb_updates ||
 607             trans->nr_wb_updates == trans->wb_updates_size) {
 608                 struct btree_write_buffered_key *u;
 609
 610                 if (trans->nr_wb_updates == trans->wb_updates_size) {
 611                         struct btree_transaction_stats *s = btree_trans_stats(trans);
 612
 613                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
 614                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
 615                         if (s)
 616                                 s->wb_updates_size = trans->wb_updates_size;
 617                 }
 618
 619                 u = bch2_trans_kmalloc_nomemzero(trans,
 620                                         trans->wb_updates_size *
 621                                         sizeof(struct btree_write_buffered_key));
 622                 ret = PTR_ERR_OR_ZERO(u);
 623                 if (ret)
 624                         return ret;
 625
 626                 if (trans->nr_wb_updates)
 627                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
 628                                sizeof(struct btree_write_buffered_key));
 629                 trans->wb_updates = u;
 630         }
 631
 632         trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
 633                 .btree  = btree,
 634         };
 635
 636         bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
 637         trans->nr_wb_updates++;
 638
 639         return 0;
 640 }
 641
 642 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
 643                              enum btree_id btree, struct bpos end)
 644 {
 645         struct bkey_s_c k;
 646         int ret = 0;
 647
 648         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
 649         k = bch2_btree_iter_prev(iter);
 650         ret = bkey_err(k);
 651         if (ret)
 652                 goto err;
 653
 654         bch2_btree_iter_advance(iter);
 655         k = bch2_btree_iter_peek_slot(iter);
 656         ret = bkey_err(k);
 657         if (ret)
 658                 goto err;
 659
 660         BUG_ON(k.k->type != KEY_TYPE_deleted);
 661
 662         if (bkey_gt(k.k->p, end)) {
 663                 ret = -BCH_ERR_ENOSPC_btree_slot;
 664                 goto err;
 665         }
 666
 667         return 0;
 668 err:
 669         bch2_trans_iter_exit(trans, iter);
 670         return ret;
 671 }
 672
 673 void bch2_trans_commit_hook(struct btree_trans *trans,
 674                             struct btree_trans_commit_hook *h)
 675 {
 676         h->next = trans->hooks;
 677         trans->hooks = h;
 678 }
 679
 680 int bch2_btree_insert_nonextent(struct btree_trans *trans,
 681                                 enum btree_id btree, struct bkey_i *k,
 682                                 enum btree_update_flags flags)
 683 {
 684         struct btree_iter iter;
 685         int ret;
 686
 687         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
 688                              BTREE_ITER_NOT_EXTENTS|
 689                              BTREE_ITER_INTENT);
 690         ret   = bch2_btree_iter_traverse(&iter) ?:
 691                 bch2_trans_update(trans, &iter, k, flags);
 692         bch2_trans_iter_exit(trans, &iter);
 693         return ret;
 694 }
 695
 696 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
 697                         struct bkey_i *k, enum btree_update_flags flags)
 698 {
 699         struct btree_iter iter;
 700         int ret;
 701
 702         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
 703                              BTREE_ITER_CACHED|
 704                              BTREE_ITER_INTENT);
 705         ret   = bch2_btree_iter_traverse(&iter) ?:
 706                 bch2_trans_update(trans, &iter, k, flags);
 707         bch2_trans_iter_exit(trans, &iter);
 708         return ret;
 709 }
 710
 711 /**
 712  * bch2_btree_insert - insert keys into the extent btree
 713  * @c:                  pointer to struct bch_fs
 714  * @id:                 btree to insert into
 715  * @insert_keys:        list of keys to insert
 716  * @hook:               insert callback
 717  */
 718 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 719                       struct bkey_i *k,
 720                       struct disk_reservation *disk_res,
 721                       u64 *journal_seq, int flags)
 722 {
 723         return bch2_trans_do(c, disk_res, journal_seq, flags,
 724                              __bch2_btree_insert(&trans, id, k, 0));
 725 }
 726
 727 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
 728                                 unsigned len, unsigned update_flags)
 729 {
 730         struct bkey_i *k;
 731
 732         k = bch2_trans_kmalloc(trans, sizeof(*k));
 733         if (IS_ERR(k))
 734                 return PTR_ERR(k);
 735
 736         bkey_init(&k->k);
 737         k->k.p = iter->pos;
 738         bch2_key_resize(&k->k, len);
 739         return bch2_trans_update(trans, iter, k, update_flags);
 740 }
 741
 742 int bch2_btree_delete_at(struct btree_trans *trans,
 743                          struct btree_iter *iter, unsigned update_flags)
 744 {
 745         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
 746 }
 747
 748 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
 749                                   enum btree_id btree, struct bpos pos)
 750 {
 751         struct bkey_i *k;
 752
 753         k = bch2_trans_kmalloc(trans, sizeof(*k));
 754         if (IS_ERR(k))
 755                 return PTR_ERR(k);
 756
 757         bkey_init(&k->k);
 758         k->k.p = pos;
 759         return bch2_trans_update_buffered(trans, btree, k);
 760 }
 761
 762 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
 763                                   struct bpos start, struct bpos end,
 764                                   unsigned update_flags,
 765                                   u64 *journal_seq)
 766 {
 767         u32 restart_count = trans->restart_count;
 768         struct btree_iter iter;
 769         struct bkey_s_c k;
 770         int ret = 0;
 771
 772         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
 773         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
 774                 struct disk_reservation disk_res =
 775                         bch2_disk_reservation_init(trans->c, 0);
 776                 struct bkey_i delete;
 777
 778                 ret = bkey_err(k);
 779                 if (ret)
 780                         goto err;
 781
 782                 bkey_init(&delete.k);
 783
 784                 /*
 785                  * This could probably be more efficient for extents:
 786                  */
 787
 788                 /*
 789                  * For extents, iter.pos won't necessarily be the same as
 790                  * bkey_start_pos(k.k) (for non extents they always will be the
 791                  * same). It's important that we delete starting from iter.pos
 792                  * because the range we want to delete could start in the middle
 793                  * of k.
 794                  *
 795                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 796                  * bkey_start_pos(k.k)).
 797                  */
 798                 delete.k.p = iter.pos;
 799
 800                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
 801                         bch2_key_resize(&delete.k,
 802                                         bpos_min(end, k.k->p).offset -
 803                                         iter.pos.offset);
 804
 805                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
 806                         bch2_trans_commit(trans, &disk_res, journal_seq,
 807                                           BTREE_INSERT_NOFAIL);
 808                 bch2_disk_reservation_put(trans->c, &disk_res);
 809 err:
 810                 /*
 811                  * the bch2_trans_begin() call is in a weird place because we
 812                  * need to call it after every transaction commit, to avoid path
 813                  * overflow, but don't want to call it if the delete operation
 814                  * is a no-op and we have no work to do:
 815                  */
 816                 bch2_trans_begin(trans);
 817
 818                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 819                         ret = 0;
 820                 if (ret)
 821                         break;
 822         }
 823         bch2_trans_iter_exit(trans, &iter);
 824
 825         if (!ret && trans_was_restarted(trans, restart_count))
 826                 ret = -BCH_ERR_transaction_restart_nested;
 827         return ret;
 828 }
 829
 830 /*
 831  * bch_btree_delete_range - delete everything within a given range
 832  *
 833  * Range is a half open interval - [start, end)
 834  */
 835 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 836                             struct bpos start, struct bpos end,
 837                             unsigned update_flags,
 838                             u64 *journal_seq)
 839 {
 840         int ret = bch2_trans_run(c,
 841                         bch2_btree_delete_range_trans(&trans, id, start, end,
 842                                                       update_flags, journal_seq));
 843         if (ret == -BCH_ERR_transaction_restart_nested)
 844                 ret = 0;
 845         return ret;
 846 }
 847
 848 int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
 849                        struct bpos pos, bool set)
 850 {
 851         struct bkey_i *k;
 852         int ret = 0;
 853
 854         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
 855         ret = PTR_ERR_OR_ZERO(k);
 856         if (unlikely(ret))
 857                 return ret;
 858
 859         bkey_init(&k->k);
 860         k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
 861         k->k.p = pos;
 862
 863         return bch2_trans_update_buffered(trans, btree, k);
 864 }
 865
 866 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
 867 {
 868         struct printbuf buf = PRINTBUF;
 869         struct jset_entry_log *l;
 870         unsigned u64s;
 871         int ret;
 872
 873         prt_vprintf(&buf, fmt, args);
 874         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
 875         if (ret)
 876                 goto err;
 877
 878         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
 879
 880         ret = darray_make_room(entries, jset_u64s(u64s));
 881         if (ret)
 882                 goto err;
 883
 884         l = (void *) &darray_top(*entries);
 885         l->entry.u64s           = cpu_to_le16(u64s);
 886         l->entry.btree_id       = 0;
 887         l->entry.level          = 1;
 888         l->entry.type           = BCH_JSET_ENTRY_log;
 889         l->entry.pad[0]         = 0;
 890         l->entry.pad[1]         = 0;
 891         l->entry.pad[2]         = 0;
 892         memcpy(l->d, buf.buf, buf.pos);
 893         while (buf.pos & 7)
 894                 l->d[buf.pos++] = '\0';
 895
 896         entries->nr += jset_u64s(u64s);
 897 err:
 898         printbuf_exit(&buf);
 899         return ret;
 900 }
 901
 902 static int
 903 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
 904                   va_list args)
 905 {
 906         int ret;
 907
 908         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
 909                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
 910         } else {
 911                 ret = bch2_trans_do(c, NULL, NULL,
 912                         BTREE_INSERT_LAZY_RW|commit_flags,
 913                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
 914         }
 915
 916         return ret;
 917 }
 918
 919 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
 920 {
 921         va_list args;
 922         int ret;
 923
 924         va_start(args, fmt);
 925         ret = __bch2_fs_log_msg(c, 0, fmt, args);
 926         va_end(args);
 927         return ret;
 928 }
 929
 930 /*
 931  * Use for logging messages during recovery to enable reserved space and avoid
 932  * blocking.
 933  */
 934 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
 935 {
 936         va_list args;
 937         int ret;
 938
 939         va_start(args, fmt);
 940         ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
 941         va_end(args);
 942         return ret;
 943 }