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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "buckets.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "subvolume.h"
  19 #include "replicas.h"
  20
  21 #include <linux/prefetch.h>
  22 #include <linux/sort.h>
  23 #include <trace/events/bcachefs.h>
  24
  25 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  26                                          const struct btree_insert_entry *r)
  27 {
  28         return   cmp_int(l->btree_id,   r->btree_id) ?:
  29                  -cmp_int(l->level,     r->level) ?:
  30                  bpos_cmp(l->k->k.p,    r->k->k.p);
  31 }
  32
  33 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  34 {
  35         return i->path->l + i->level;
  36 }
  37
  38 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  39                                      struct btree_insert_entry *i)
  40 {
  41         return i != trans->updates &&
  42                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  43 }
  44
  45 static inline bool same_leaf_as_next(struct btree_trans *trans,
  46                                      struct btree_insert_entry *i)
  47 {
  48         return i + 1 < trans->updates + trans->nr_updates &&
  49                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  50 }
  51
  52 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  53                                                   struct btree_path *path,
  54                                                   struct btree *b)
  55 {
  56         struct bch_fs *c = trans->c;
  57
  58         if (path->cached)
  59                 return;
  60
  61         if (unlikely(btree_node_just_written(b)) &&
  62             bch2_btree_post_write_cleanup(c, b))
  63                 bch2_trans_node_reinit_iter(trans, b);
  64
  65         /*
  66          * If the last bset has been written, or if it's gotten too big - start
  67          * a new bset to insert into:
  68          */
  69         if (want_new_bset(c, b))
  70                 bch2_btree_init_next(trans, b);
  71 }
  72
  73 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
  74                                      struct btree_path *path,
  75                                      struct btree *b)
  76 {
  77         bch2_btree_node_lock_write(trans, path, b);
  78         bch2_btree_node_prep_for_write(trans, path, b);
  79 }
  80
  81 /* Inserting into a given leaf node (last stage of insert): */
  82
  83 /* Handle overwrites and do insert, for non extents: */
  84 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  85                                 struct btree_path *path,
  86                                 struct btree *b,
  87                                 struct btree_node_iter *node_iter,
  88                                 struct bkey_i *insert)
  89 {
  90         struct bkey_packed *k;
  91         unsigned clobber_u64s = 0, new_u64s = 0;
  92
  93         EBUG_ON(btree_node_just_written(b));
  94         EBUG_ON(bset_written(b, btree_bset_last(b)));
  95         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  96         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  97         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  98         EBUG_ON(insert->k.u64s >
  99                 bch_btree_keys_u64s_remaining(trans->c, b));
 100
 101         k = bch2_btree_node_iter_peek_all(node_iter, b);
 102         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 103                 k = NULL;
 104
 105         /* @k is the key being overwritten/deleted, if any: */
 106         EBUG_ON(k && bkey_deleted(k));
 107
 108         /* Deleting, but not found? nothing to do: */
 109         if (bkey_deleted(&insert->k) && !k)
 110                 return false;
 111
 112         if (bkey_deleted(&insert->k)) {
 113                 /* Deleting: */
 114                 btree_account_key_drop(b, k);
 115                 k->type = KEY_TYPE_deleted;
 116
 117                 if (k->needs_whiteout)
 118                         push_whiteout(trans->c, b, insert->k.p);
 119                 k->needs_whiteout = false;
 120
 121                 if (k >= btree_bset_last(b)->start) {
 122                         clobber_u64s = k->u64s;
 123                         bch2_bset_delete(b, k, clobber_u64s);
 124                         goto fix_iter;
 125                 } else {
 126                         bch2_btree_path_fix_key_modified(trans, b, k);
 127                 }
 128
 129                 return true;
 130         }
 131
 132         if (k) {
 133                 /* Overwriting: */
 134                 btree_account_key_drop(b, k);
 135                 k->type = KEY_TYPE_deleted;
 136
 137                 insert->k.needs_whiteout = k->needs_whiteout;
 138                 k->needs_whiteout = false;
 139
 140                 if (k >= btree_bset_last(b)->start) {
 141                         clobber_u64s = k->u64s;
 142                         goto overwrite;
 143                 } else {
 144                         bch2_btree_path_fix_key_modified(trans, b, k);
 145                 }
 146         }
 147
 148         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 149 overwrite:
 150         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 151         new_u64s = k->u64s;
 152 fix_iter:
 153         if (clobber_u64s != new_u64s)
 154                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 155                                          clobber_u64s, new_u64s);
 156         return true;
 157 }
 158
 159 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 160                                unsigned i, u64 seq)
 161 {
 162         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 163         struct btree_write *w = container_of(pin, struct btree_write, journal);
 164         struct btree *b = container_of(w, struct btree, writes[i]);
 165
 166         btree_node_lock_type(c, b, SIX_LOCK_read);
 167         bch2_btree_node_write_cond(c, b,
 168                 (btree_current_write(b) == w && w->journal.seq == seq));
 169         six_unlock_read(&b->c.lock);
 170         return 0;
 171 }
 172
 173 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 174 {
 175         return __btree_node_flush(j, pin, 0, seq);
 176 }
 177
 178 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 179 {
 180         return __btree_node_flush(j, pin, 1, seq);
 181 }
 182
 183 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 184                                        struct btree *b, u64 seq)
 185 {
 186         struct btree_write *w = btree_current_write(b);
 187
 188         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 189                              btree_node_write_idx(b) == 0
 190                              ? btree_node_flush0
 191                              : btree_node_flush1);
 192 }
 193
 194 /**
 195  * btree_insert_key - insert a key one key into a leaf node
 196  */
 197 static bool btree_insert_key_leaf(struct btree_trans *trans,
 198                                   struct btree_insert_entry *insert)
 199 {
 200         struct bch_fs *c = trans->c;
 201         struct btree *b = insert_l(insert)->b;
 202         struct bset_tree *t = bset_tree_last(b);
 203         struct bset *i = bset(b, t);
 204         int old_u64s = bset_u64s(t);
 205         int old_live_u64s = b->nr.live_u64s;
 206         int live_u64s_added, u64s_added;
 207
 208         EBUG_ON(!insert->level &&
 209                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 210
 211         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 212                                         &insert_l(insert)->iter, insert->k)))
 213                 return false;
 214
 215         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 216                                          le64_to_cpu(i->journal_seq)));
 217
 218         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 219
 220         if (unlikely(!btree_node_dirty(b)))
 221                 set_btree_node_dirty(c, b);
 222
 223         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 224         u64s_added = (int) bset_u64s(t) - old_u64s;
 225
 226         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 227                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 228         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 229                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 230
 231         if (u64s_added > live_u64s_added &&
 232             bch2_maybe_compact_whiteouts(c, b))
 233                 bch2_trans_node_reinit_iter(trans, b);
 234
 235         return true;
 236 }
 237
 238 /* Cached btree updates: */
 239
 240 /* Normal update interface: */
 241
 242 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 243                                              struct btree_insert_entry *i)
 244 {
 245         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 246         BUG_ON(i->cached        != i->path->cached);
 247         BUG_ON(i->level         != i->path->level);
 248         BUG_ON(i->btree_id      != i->path->btree_id);
 249         EBUG_ON(!i->level &&
 250                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 251                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 252                 i->k->k.p.snapshot &&
 253                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 254 }
 255
 256 static noinline int
 257 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 258                                    unsigned long trace_ip)
 259 {
 260         struct bch_fs *c = trans->c;
 261         int ret;
 262
 263         bch2_trans_unlock(trans);
 264
 265         ret = bch2_journal_preres_get(&c->journal,
 266                         &trans->journal_preres, u64s, 0);
 267         if (ret)
 268                 return ret;
 269
 270         if (!bch2_trans_relock(trans)) {
 271                 trace_trans_restart_journal_preres_get(trans->ip, trace_ip);
 272                 return -EINTR;
 273         }
 274
 275         return 0;
 276 }
 277
 278 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 279                                              unsigned flags)
 280 {
 281         struct bch_fs *c = trans->c;
 282         int ret;
 283
 284         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 285                 flags |= JOURNAL_RES_GET_RESERVED;
 286
 287         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 288                                    trans->journal_u64s, flags);
 289
 290         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 291 }
 292
 293 static inline enum btree_insert_ret
 294 btree_key_can_insert(struct btree_trans *trans,
 295                      struct btree *b,
 296                      unsigned u64s)
 297 {
 298         struct bch_fs *c = trans->c;
 299
 300         if (!bch2_btree_node_insert_fits(c, b, u64s))
 301                 return BTREE_INSERT_BTREE_NODE_FULL;
 302
 303         return BTREE_INSERT_OK;
 304 }
 305
 306 static enum btree_insert_ret
 307 btree_key_can_insert_cached(struct btree_trans *trans,
 308                             struct btree_path *path,
 309                             unsigned u64s)
 310 {
 311         struct bkey_cached *ck = (void *) path->l[0].b;
 312         unsigned new_u64s;
 313         struct bkey_i *new_k;
 314
 315         EBUG_ON(path->level);
 316
 317         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 318             bch2_btree_key_cache_must_wait(trans->c) &&
 319             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 320                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 321
 322         /*
 323          * bch2_varint_decode can read past the end of the buffer by at most 7
 324          * bytes (it won't be used):
 325          */
 326         u64s += 1;
 327
 328         if (u64s <= ck->u64s)
 329                 return BTREE_INSERT_OK;
 330
 331         new_u64s        = roundup_pow_of_two(u64s);
 332         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 333         if (!new_k)
 334                 return -ENOMEM;
 335
 336         ck->u64s        = new_u64s;
 337         ck->k           = new_k;
 338         return BTREE_INSERT_OK;
 339 }
 340
 341 static inline void do_btree_insert_one(struct btree_trans *trans,
 342                                        struct btree_insert_entry *i)
 343 {
 344         struct bch_fs *c = trans->c;
 345         struct journal *j = &c->journal;
 346         bool did_work;
 347
 348         EBUG_ON(trans->journal_res.ref !=
 349                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 350
 351         i->k->k.needs_whiteout = false;
 352
 353         did_work = !i->cached
 354                 ? btree_insert_key_leaf(trans, i)
 355                 : bch2_btree_insert_key_cached(trans, i->path, i->k);
 356         if (!did_work)
 357                 return;
 358
 359         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 360                 bch2_journal_add_keys(j, &trans->journal_res,
 361                                       i->btree_id,
 362                                       i->level,
 363                                       i->k);
 364
 365                 if (trans->journal_seq)
 366                         *trans->journal_seq = trans->journal_res.seq;
 367         }
 368 }
 369
 370 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 371 {
 372         struct bch_fs *c = trans->c;
 373         struct btree_insert_entry *i;
 374
 375         trans_for_each_update(trans, i) {
 376                 /*
 377                  * XXX: synchronization of cached update triggers with gc
 378                  * XXX: synchronization of interior node updates with gc
 379                  */
 380                 BUG_ON(i->cached || i->level);
 381
 382                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b)))
 383                         bch2_mark_update(trans, i->path, i->k,
 384                                          i->flags|BTREE_TRIGGER_GC);
 385         }
 386 }
 387
 388 static inline int
 389 bch2_trans_commit_write_locked(struct btree_trans *trans,
 390                                struct btree_insert_entry **stopped_at,
 391                                unsigned long trace_ip)
 392 {
 393         struct bch_fs *c = trans->c;
 394         struct btree_insert_entry *i;
 395         struct btree_trans_commit_hook *h;
 396         unsigned u64s = 0;
 397         bool marking = false;
 398         int ret;
 399
 400         if (race_fault()) {
 401                 trace_trans_restart_fault_inject(trans->ip, trace_ip);
 402                 trans->restarted = true;
 403                 return -EINTR;
 404         }
 405
 406         /*
 407          * Check if the insert will fit in the leaf node with the write lock
 408          * held, otherwise another thread could write the node changing the
 409          * amount of space available:
 410          */
 411
 412         prefetch(&trans->c->journal.flags);
 413
 414         h = trans->hooks;
 415         while (h) {
 416                 ret = h->fn(trans, h);
 417                 if (ret)
 418                         return ret;
 419                 h = h->next;
 420         }
 421
 422         trans_for_each_update(trans, i) {
 423                 /* Multiple inserts might go to same leaf: */
 424                 if (!same_leaf_as_prev(trans, i))
 425                         u64s = 0;
 426
 427                 u64s += i->k->k.u64s;
 428                 ret = !i->cached
 429                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 430                         : btree_key_can_insert_cached(trans, i->path, u64s);
 431                 if (ret) {
 432                         *stopped_at = i;
 433                         return ret;
 434                 }
 435
 436                 if (btree_node_type_needs_gc(i->bkey_type))
 437                         marking = true;
 438         }
 439
 440         /*
 441          * Don't get journal reservation until after we know insert will
 442          * succeed:
 443          */
 444         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 445                 ret = bch2_trans_journal_res_get(trans,
 446                                 JOURNAL_RES_GET_NONBLOCK);
 447                 if (ret)
 448                         return ret;
 449         } else {
 450                 trans->journal_res.seq = c->journal.replay_journal_seq;
 451         }
 452
 453         if (unlikely(trans->extra_journal_entry_u64s)) {
 454                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 455                                   trans->extra_journal_entries,
 456                                   trans->extra_journal_entry_u64s);
 457
 458                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 459                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 460         }
 461
 462         /*
 463          * Not allowed to fail after we've gotten our journal reservation - we
 464          * have to use it:
 465          */
 466
 467         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 468                 if (bch2_journal_seq_verify)
 469                         trans_for_each_update(trans, i)
 470                                 i->k->k.version.lo = trans->journal_res.seq;
 471                 else if (bch2_inject_invalid_keys)
 472                         trans_for_each_update(trans, i)
 473                                 i->k->k.version = MAX_VERSION;
 474         }
 475
 476         if (trans->fs_usage_deltas &&
 477             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 478                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 479
 480         trans_for_each_update(trans, i)
 481                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 482                         bch2_mark_update(trans, i->path, i->k, i->flags);
 483
 484         if (unlikely(c->gc_pos.phase))
 485                 bch2_trans_mark_gc(trans);
 486
 487         trans_for_each_update(trans, i)
 488                 do_btree_insert_one(trans, i);
 489
 490         return ret;
 491 }
 492
 493 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 494 {
 495         unsigned l;
 496
 497         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 498                 if (btree_node_read_locked(path, l))
 499                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 500 }
 501
 502 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 503 {
 504         struct btree *b = path_l(path)->b;
 505
 506         do {
 507                 if (path->nodes_locked &&
 508                     path->nodes_locked != path->nodes_intent_locked)
 509                         path_upgrade_readers(trans, path);
 510         } while ((path = prev_btree_path(trans, path)) &&
 511                  path_l(path)->b == b);
 512 }
 513
 514 /*
 515  * Check for nodes that we have both read and intent locks on, and upgrade the
 516  * readers to intent:
 517  */
 518 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 519 {
 520         struct btree_path *path;
 521         unsigned i, nr_read = 0, nr_intent = 0;
 522
 523         trans_for_each_path_inorder(trans, path, i) {
 524                 struct btree_path *next = i + 1 < trans->nr_sorted
 525                         ? trans->paths + trans->sorted[i + 1]
 526                         : NULL;
 527
 528                 if (path->nodes_locked) {
 529                         if (path->nodes_intent_locked)
 530                                 nr_intent++;
 531                         else
 532                                 nr_read++;
 533                 }
 534
 535                 if (!next || path_l(path)->b != path_l(next)->b) {
 536                         if (nr_read && nr_intent)
 537                                 upgrade_readers(trans, path);
 538
 539                         nr_read = nr_intent = 0;
 540                 }
 541         }
 542
 543         bch2_trans_verify_locks(trans);
 544 }
 545
 546 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 547 {
 548         struct btree_path *path;
 549         unsigned i;
 550
 551         trans_for_each_path_inorder(trans, path, i) {
 552                 //if (path == pos)
 553                 //      break;
 554
 555                 if (path->nodes_locked != path->nodes_intent_locked &&
 556                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 557                         return true;
 558         }
 559
 560         return false;
 561 }
 562
 563 static inline int trans_lock_write(struct btree_trans *trans)
 564 {
 565         struct btree_insert_entry *i;
 566
 567         trans_for_each_update(trans, i) {
 568                 if (same_leaf_as_prev(trans, i))
 569                         continue;
 570
 571                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 572                         if (have_conflicting_read_lock(trans, i->path))
 573                                 goto fail;
 574
 575                         __btree_node_lock_type(trans->c, insert_l(i)->b,
 576                                                SIX_LOCK_write);
 577                 }
 578
 579                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 580         }
 581
 582         return 0;
 583 fail:
 584         while (--i >= trans->updates) {
 585                 if (same_leaf_as_prev(trans, i))
 586                         continue;
 587
 588                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 589         }
 590
 591         trace_trans_restart_would_deadlock_write(trans->ip);
 592         return btree_trans_restart(trans);
 593 }
 594
 595 /*
 596  * Get journal reservation, take write locks, and attempt to do btree update(s):
 597  */
 598 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 599                                        struct btree_insert_entry **stopped_at,
 600                                        unsigned long trace_ip)
 601 {
 602         struct bch_fs *c = trans->c;
 603         struct btree_insert_entry *i;
 604         struct bkey_s_c old;
 605         int ret, u64s_delta = 0;
 606
 607         trans_for_each_update(trans, i) {
 608                 const char *invalid = bch2_bkey_invalid(c,
 609                                 bkey_i_to_s_c(i->k), i->bkey_type);
 610                 if (invalid) {
 611                         char buf[200];
 612
 613                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 614                         bch_err(c, "invalid bkey %s on insert from %ps -> %ps: %s\n",
 615                                 buf, (void *) trans->ip,
 616                                 (void *) i->ip_allocated, invalid);
 617                         bch2_fatal_error(c);
 618                         return -EINVAL;
 619                 }
 620                 btree_insert_entry_checks(trans, i);
 621         }
 622
 623         trans_for_each_update(trans, i) {
 624                 struct bkey u;
 625
 626                 /*
 627                  * peek_slot() doesn't yet work on iterators that point to
 628                  * interior nodes:
 629                  */
 630                 if (i->cached || i->level)
 631                         continue;
 632
 633                 old = bch2_btree_path_peek_slot(i->path, &u);
 634                 ret = bkey_err(old);
 635                 if (unlikely(ret))
 636                         return ret;
 637
 638                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 639                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 640
 641                 if (!same_leaf_as_next(trans, i)) {
 642                         if (u64s_delta <= 0) {
 643                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 644                                                         i->level, trans->flags);
 645                                 if (unlikely(ret))
 646                                         return ret;
 647                         }
 648
 649                         u64s_delta = 0;
 650                 }
 651         }
 652
 653         ret = bch2_journal_preres_get(&c->journal,
 654                         &trans->journal_preres, trans->journal_preres_u64s,
 655                         JOURNAL_RES_GET_NONBLOCK|
 656                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 657                          ? JOURNAL_RES_GET_RESERVED : 0));
 658         if (unlikely(ret == -EAGAIN))
 659                 ret = bch2_trans_journal_preres_get_cold(trans,
 660                                                 trans->journal_preres_u64s, trace_ip);
 661         if (unlikely(ret))
 662                 return ret;
 663
 664         normalize_read_intent_locks(trans);
 665
 666         ret = trans_lock_write(trans);
 667         if (unlikely(ret))
 668                 return ret;
 669
 670         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 671
 672         trans_for_each_update(trans, i)
 673                 if (!same_leaf_as_prev(trans, i))
 674                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 675                                                         insert_l(i)->b);
 676
 677         if (!ret && trans->journal_pin)
 678                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 679                                      trans->journal_pin, NULL);
 680
 681         /*
 682          * Drop journal reservation after dropping write locks, since dropping
 683          * the journal reservation may kick off a journal write:
 684          */
 685         bch2_journal_res_put(&c->journal, &trans->journal_res);
 686
 687         if (unlikely(ret))
 688                 return ret;
 689
 690         bch2_trans_downgrade(trans);
 691
 692         return 0;
 693 }
 694
 695 static int journal_reclaim_wait_done(struct bch_fs *c)
 696 {
 697         int ret = bch2_journal_error(&c->journal) ?:
 698                 !bch2_btree_key_cache_must_wait(c);
 699
 700         if (!ret)
 701                 journal_reclaim_kick(&c->journal);
 702         return ret;
 703 }
 704
 705 static noinline
 706 int bch2_trans_commit_error(struct btree_trans *trans,
 707                             struct btree_insert_entry *i,
 708                             int ret, unsigned long trace_ip)
 709 {
 710         struct bch_fs *c = trans->c;
 711
 712         switch (ret) {
 713         case BTREE_INSERT_BTREE_NODE_FULL:
 714                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 715                 if (!ret)
 716                         return 0;
 717
 718                 if (ret == -EINTR)
 719                         trace_trans_restart_btree_node_split(trans->ip, trace_ip,
 720                                                 i->btree_id, &i->path->pos);
 721                 break;
 722         case BTREE_INSERT_NEED_MARK_REPLICAS:
 723                 bch2_trans_unlock(trans);
 724
 725                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 726                 if (ret)
 727                         break;
 728
 729                 if (bch2_trans_relock(trans))
 730                         return 0;
 731
 732                 trace_trans_restart_mark_replicas(trans->ip, trace_ip);
 733                 ret = -EINTR;
 734                 break;
 735         case BTREE_INSERT_NEED_JOURNAL_RES:
 736                 bch2_trans_unlock(trans);
 737
 738                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 739                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
 740                         trans->restarted = true;
 741                         ret = -EAGAIN;
 742                         break;
 743                 }
 744
 745                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 746                 if (ret)
 747                         break;
 748
 749                 if (bch2_trans_relock(trans))
 750                         return 0;
 751
 752                 trace_trans_restart_journal_res_get(trans->ip, trace_ip);
 753                 ret = -EINTR;
 754                 break;
 755         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 756                 bch2_trans_unlock(trans);
 757
 758                 trace_trans_blocked_journal_reclaim(trans->ip, trace_ip);
 759
 760                 wait_event_freezable(c->journal.reclaim_wait,
 761                                      (ret = journal_reclaim_wait_done(c)));
 762                 if (ret < 0)
 763                         break;
 764
 765                 if (bch2_trans_relock(trans))
 766                         return 0;
 767
 768                 trace_trans_restart_journal_reclaim(trans->ip, trace_ip);
 769                 ret = -EINTR;
 770                 break;
 771         default:
 772                 BUG_ON(ret >= 0);
 773                 break;
 774         }
 775
 776         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
 777         BUG_ON(ret == -ENOSPC && (trans->flags & BTREE_INSERT_NOFAIL));
 778
 779         return ret;
 780 }
 781
 782 static noinline int
 783 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 784 {
 785         struct bch_fs *c = trans->c;
 786         int ret;
 787
 788         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 789                 return -EROFS;
 790
 791         bch2_trans_unlock(trans);
 792
 793         ret = bch2_fs_read_write_early(c);
 794         if (ret)
 795                 return ret;
 796
 797         if (!bch2_trans_relock(trans))
 798                 return -EINTR;
 799
 800         percpu_ref_get(&c->writes);
 801         return 0;
 802 }
 803
 804 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 805 {
 806         struct bkey             _deleted = KEY(0, 0, 0);
 807         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 808         struct bkey_s_c         old;
 809         struct bkey             unpacked;
 810         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 811         bool trans_trigger_run;
 812         unsigned btree_id = 0;
 813         int ret = 0;
 814
 815         /*
 816          *
 817          * For a given btree, this algorithm runs insert triggers before
 818          * overwrite triggers: this is so that when extents are being moved
 819          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 820          * they are re-added.
 821          */
 822         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 823                 while (btree_id_start < trans->updates + trans->nr_updates &&
 824                        btree_id_start->btree_id < btree_id)
 825                         btree_id_start++;
 826
 827                 /*
 828                  * Running triggers will append more updates to the list of updates as
 829                  * we're walking it:
 830                  */
 831                 do {
 832                         trans_trigger_run = false;
 833
 834                         for (i = btree_id_start;
 835                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 836                              i++) {
 837                                 if (i->insert_trigger_run ||
 838                                     (i->flags & BTREE_TRIGGER_NORUN) ||
 839                                     !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 840                                         continue;
 841
 842                                 BUG_ON(i->overwrite_trigger_run);
 843
 844                                 i->insert_trigger_run = true;
 845                                 trans_trigger_run = true;
 846
 847                                 old = bch2_btree_path_peek_slot(i->path, &unpacked);
 848                                 _deleted.p = i->path->pos;
 849
 850                                 if (old.k->type == i->k->k.type &&
 851                                     ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 852                                         i->overwrite_trigger_run = true;
 853                                         ret = bch2_trans_mark_key(trans, old, bkey_i_to_s_c(i->k),
 854                                                         BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|i->flags);
 855                                 } else {
 856                                         ret = bch2_trans_mark_key(trans, deleted, bkey_i_to_s_c(i->k),
 857                                                         BTREE_TRIGGER_INSERT|i->flags);
 858                                 }
 859
 860                                 if (ret == -EINTR)
 861                                         trace_trans_restart_mark(trans->ip, _RET_IP_,
 862                                                         i->btree_id, &i->path->pos);
 863                                 if (ret)
 864                                         return ret;
 865                         }
 866                 } while (trans_trigger_run);
 867
 868                 do {
 869                         trans_trigger_run = false;
 870
 871                         for (i = btree_id_start;
 872                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 873                              i++) {
 874                                 if (i->overwrite_trigger_run ||
 875                                     (i->flags & BTREE_TRIGGER_NORUN) ||
 876                                     !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 877                                         continue;
 878
 879                                 BUG_ON(!i->insert_trigger_run);
 880
 881                                 i->overwrite_trigger_run = true;
 882                                 trans_trigger_run = true;
 883
 884                                 old = bch2_btree_path_peek_slot(i->path, &unpacked);
 885                                 _deleted.p = i->path->pos;
 886
 887                                 ret = bch2_trans_mark_key(trans, old, deleted,
 888                                                 BTREE_TRIGGER_OVERWRITE|i->flags);
 889
 890                                 if (ret == -EINTR)
 891                                         trace_trans_restart_mark(trans->ip, _RET_IP_,
 892                                                         i->btree_id, &i->path->pos);
 893                                 if (ret)
 894                                         return ret;
 895                         }
 896                 } while (trans_trigger_run);
 897         }
 898
 899         trans_for_each_update(trans, i)
 900                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 901                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 902                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 903
 904         return 0;
 905 }
 906
 907 int __bch2_trans_commit(struct btree_trans *trans)
 908 {
 909         struct btree_insert_entry *i = NULL;
 910         unsigned u64s;
 911         int ret = 0;
 912
 913         if (!trans->nr_updates &&
 914             !trans->extra_journal_entry_u64s)
 915                 goto out_reset;
 916
 917         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 918                 lockdep_assert_held(&trans->c->gc_lock);
 919
 920         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 921
 922         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 923         trans->journal_preres_u64s      = 0;
 924
 925         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 926             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 927                 ret = bch2_trans_commit_get_rw_cold(trans);
 928                 if (ret)
 929                         goto out_reset;
 930         }
 931
 932 #ifdef CONFIG_BCACHEFS_DEBUG
 933         /*
 934          * if BTREE_TRIGGER_NORUN is set, it means we're probably being called
 935          * from the key cache flush code:
 936          */
 937         trans_for_each_update(trans, i)
 938                 if (!i->cached &&
 939                     !(i->flags & BTREE_TRIGGER_NORUN))
 940                         bch2_btree_key_cache_verify_clean(trans,
 941                                         i->btree_id, i->k->k.p);
 942 #endif
 943
 944         ret = bch2_trans_commit_run_triggers(trans);
 945         if (ret)
 946                 goto out;
 947
 948         trans_for_each_update(trans, i) {
 949                 BUG_ON(!i->path->should_be_locked);
 950
 951                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
 952                         trace_trans_restart_upgrade(trans->ip, _RET_IP_,
 953                                                     i->btree_id, &i->path->pos);
 954                         ret = btree_trans_restart(trans);
 955                         goto out;
 956                 }
 957
 958                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
 959
 960                 u64s = jset_u64s(i->k->k.u64s);
 961                 if (i->cached &&
 962                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 963                         trans->journal_preres_u64s += u64s;
 964                 trans->journal_u64s += u64s;
 965         }
 966
 967         if (trans->extra_journal_res) {
 968                 ret = bch2_disk_reservation_add(trans->c, trans->disk_res,
 969                                 trans->extra_journal_res,
 970                                 (trans->flags & BTREE_INSERT_NOFAIL)
 971                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 972                 if (ret)
 973                         goto err;
 974         }
 975 retry:
 976         BUG_ON(trans->restarted);
 977         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 978
 979         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
 980
 981         /* make sure we didn't drop or screw up locks: */
 982         bch2_trans_verify_locks(trans);
 983
 984         if (ret)
 985                 goto err;
 986 out:
 987         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 988
 989         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 990                 percpu_ref_put(&trans->c->writes);
 991 out_reset:
 992         trans_for_each_update(trans, i)
 993                 bch2_path_put(trans, i->path, true);
 994
 995         trans->extra_journal_res        = 0;
 996         trans->nr_updates               = 0;
 997         trans->hooks                    = NULL;
 998         trans->extra_journal_entries    = NULL;
 999         trans->extra_journal_entry_u64s = 0;
1000
1001         if (trans->fs_usage_deltas) {
1002                 trans->fs_usage_deltas->used = 0;
1003                 memset(&trans->fs_usage_deltas->memset_start, 0,
1004                        (void *) &trans->fs_usage_deltas->memset_end -
1005                        (void *) &trans->fs_usage_deltas->memset_start);
1006         }
1007
1008         return ret;
1009 err:
1010         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1011         if (ret)
1012                 goto out;
1013
1014         goto retry;
1015 }
1016
1017 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1018                                           enum btree_id id,
1019                                           struct bpos pos)
1020 {
1021         struct bch_fs *c = trans->c;
1022         struct btree_iter iter;
1023         struct bkey_s_c k;
1024         int ret;
1025
1026         if (!snapshot_t(c, pos.snapshot)->children[0])
1027                 return 0;
1028
1029         bch2_trans_iter_init(trans, &iter, id, pos,
1030                              BTREE_ITER_NOT_EXTENTS|
1031                              BTREE_ITER_ALL_SNAPSHOTS);
1032         while (1) {
1033                 k = bch2_btree_iter_prev(&iter);
1034                 ret = bkey_err(k);
1035                 if (ret)
1036                         break;
1037
1038                 if (!k.k)
1039                         break;
1040
1041                 if (bkey_cmp(pos, k.k->p))
1042                         break;
1043
1044                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1045                         ret = 1;
1046                         break;
1047                 }
1048         }
1049         bch2_trans_iter_exit(trans, &iter);
1050
1051         return ret;
1052 }
1053
1054 static int bch2_trans_update_extent(struct btree_trans *trans,
1055                                     struct btree_iter *orig_iter,
1056                                     struct bkey_i *insert,
1057                                     enum btree_update_flags flags)
1058 {
1059         struct bch_fs *c = trans->c;
1060         struct btree_iter iter, update_iter;
1061         struct bpos start = bkey_start_pos(&insert->k);
1062         struct bkey_i *update;
1063         struct bkey_s_c k;
1064         enum btree_id btree_id = orig_iter->btree_id;
1065         int ret = 0, compressed_sectors;
1066
1067         bch2_trans_iter_init(trans, &iter, btree_id, start,
1068                              BTREE_ITER_INTENT|
1069                              BTREE_ITER_WITH_UPDATES|
1070                              BTREE_ITER_NOT_EXTENTS);
1071         k = bch2_btree_iter_peek(&iter);
1072         if ((ret = bkey_err(k)))
1073                 goto err;
1074         if (!k.k)
1075                 goto out;
1076
1077         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1078                 /*
1079                  * We can't merge extents if they belong to interior snapshot
1080                  * tree nodes, and there's a snapshot in which one extent is
1081                  * visible and the other is not - i.e. if visibility is
1082                  * different.
1083                  *
1084                  * Instead of checking if visibilitiy of the two extents is
1085                  * different, for now we just check if either has been
1086                  * overwritten:
1087                  */
1088                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1089                 if (ret < 0)
1090                         goto err;
1091                 if (ret)
1092                         goto nomerge1;
1093
1094                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1095                 if (ret < 0)
1096                         goto err;
1097                 if (ret)
1098                         goto nomerge1;
1099
1100                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1101                 if ((ret = PTR_ERR_OR_ZERO(update)))
1102                         goto err;
1103
1104                 bkey_reassemble(update, k);
1105
1106                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1107                         ret = bch2_btree_delete_at(trans, &iter, flags);
1108                         if (ret)
1109                                 goto err;
1110
1111                         insert = update;
1112                         goto next;
1113                 }
1114         }
1115 nomerge1:
1116         ret = 0;
1117         if (!bkey_cmp(k.k->p, start))
1118                 goto next;
1119
1120         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1121                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1122                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1123
1124                 /*
1125                  * If we're going to be splitting a compressed extent, note it
1126                  * so that __bch2_trans_commit() can increase our disk
1127                  * reservation:
1128                  */
1129                 if (((front_split && back_split) ||
1130                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1131                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1132                         trans->extra_journal_res += compressed_sectors;
1133
1134                 if (front_split) {
1135                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1136                         if ((ret = PTR_ERR_OR_ZERO(update)))
1137                                 goto err;
1138
1139                         bkey_reassemble(update, k);
1140
1141                         bch2_cut_back(start, update);
1142
1143                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1144                                              BTREE_ITER_NOT_EXTENTS|
1145                                              BTREE_ITER_ALL_SNAPSHOTS|
1146                                              BTREE_ITER_INTENT);
1147                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1148                                 bch2_trans_update(trans, &update_iter, update,
1149                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1150                                                   flags);
1151                         bch2_trans_iter_exit(trans, &update_iter);
1152
1153                         if (ret)
1154                                 goto err;
1155                 }
1156
1157                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1158                     (front_split || back_split)) {
1159                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1160                         if ((ret = PTR_ERR_OR_ZERO(update)))
1161                                 goto err;
1162
1163                         bkey_reassemble(update, k);
1164
1165                         bch2_cut_front(start, update);
1166                         bch2_cut_back(insert->k.p, update);
1167
1168                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1169                                              BTREE_ITER_NOT_EXTENTS|
1170                                              BTREE_ITER_ALL_SNAPSHOTS|
1171                                              BTREE_ITER_INTENT);
1172                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1173                                 bch2_trans_update(trans, &update_iter, update,
1174                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1175                                                   flags);
1176                         bch2_trans_iter_exit(trans, &update_iter);
1177                         if (ret)
1178                                 goto err;
1179                 }
1180
1181                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1182                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1183                         if ((ret = PTR_ERR_OR_ZERO(update)))
1184                                 goto err;
1185
1186                         bkey_init(&update->k);
1187                         update->k.p = k.k->p;
1188
1189                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1190                                 update->k.p.snapshot = insert->k.p.snapshot;
1191                                 update->k.type = KEY_TYPE_whiteout;
1192                         }
1193
1194                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1195                                              BTREE_ITER_NOT_EXTENTS|
1196                                              BTREE_ITER_INTENT);
1197                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1198                                 bch2_trans_update(trans, &update_iter, update,
1199                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1200                                                   flags);
1201                         bch2_trans_iter_exit(trans, &update_iter);
1202
1203                         if (ret)
1204                                 goto err;
1205                 }
1206
1207                 if (back_split) {
1208                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1209                         if ((ret = PTR_ERR_OR_ZERO(update)))
1210                                 goto err;
1211
1212                         bkey_reassemble(update, k);
1213                         bch2_cut_front(insert->k.p, update);
1214
1215                         bch2_trans_copy_iter(&update_iter, &iter);
1216                         update_iter.pos = update->k.p;
1217                         ret   = bch2_trans_update(trans, &update_iter, update,
1218                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1219                                                   flags);
1220                         bch2_trans_iter_exit(trans, &update_iter);
1221
1222                         if (ret)
1223                                 goto err;
1224                         goto out;
1225                 }
1226 next:
1227                 k = bch2_btree_iter_next(&iter);
1228                 if ((ret = bkey_err(k)))
1229                         goto err;
1230                 if (!k.k)
1231                         goto out;
1232         }
1233
1234         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1235                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1236                 if (ret < 0)
1237                         goto err;
1238                 if (ret)
1239                         goto nomerge2;
1240
1241                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1242                 if (ret < 0)
1243                         goto err;
1244                 if (ret)
1245                         goto nomerge2;
1246
1247                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1248         }
1249 nomerge2:
1250         ret = 0;
1251 out:
1252         if (!bkey_deleted(&insert->k)) {
1253                 /*
1254                  * Rewinding iterators is expensive: get a new one and the one
1255                  * that points to the start of insert will be cloned from:
1256                  */
1257                 bch2_trans_iter_exit(trans, &iter);
1258                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1259                                      BTREE_ITER_NOT_EXTENTS|
1260                                      BTREE_ITER_INTENT);
1261                 ret   = bch2_btree_iter_traverse(&iter) ?:
1262                         bch2_trans_update(trans, &iter, insert, flags);
1263         }
1264 err:
1265         bch2_trans_iter_exit(trans, &iter);
1266
1267         return ret;
1268 }
1269
1270 /*
1271  * When deleting, check if we need to emit a whiteout (because we're overwriting
1272  * something in an ancestor snapshot)
1273  */
1274 static int need_whiteout_for_snapshot(struct btree_trans *trans, struct btree_iter *orig)
1275 {
1276         struct btree_iter iter;
1277         struct bkey_s_c k;
1278         u32 snapshot = orig->pos.snapshot;
1279         int ret;
1280
1281         if (!bch2_snapshot_parent(trans->c, snapshot))
1282                 return 0;
1283
1284         bch2_trans_copy_iter(&iter, orig);
1285         iter.flags &= BTREE_ITER_FILTER_SNAPSHOTS;
1286         iter.flags |= BTREE_ITER_ALL_SNAPSHOTS;
1287
1288         bch2_btree_iter_advance(&iter);
1289
1290         for_each_btree_key_continue_norestart(iter, 0, k, ret) {
1291                 if (bkey_cmp(k.k->p, orig->pos))
1292                         break;
1293
1294                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1295                                               k.k->p.snapshot)) {
1296                         ret = !bkey_whiteout(k.k);
1297                         break;
1298                 }
1299         }
1300         bch2_trans_iter_exit(trans, &iter);
1301
1302         return ret;
1303 }
1304
1305 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1306                                    struct bkey_i *k, enum btree_update_flags flags)
1307 {
1308         struct btree_insert_entry *i, n;
1309
1310         BUG_ON(!iter->path->should_be_locked);
1311
1312         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1313                 return bch2_trans_update_extent(trans, iter, k, flags);
1314
1315         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1316         BUG_ON(bpos_cmp(k->k.p, iter->path->pos));
1317         BUG_ON(bpos_cmp(k->k.p, iter->pos));
1318
1319         n = (struct btree_insert_entry) {
1320                 .flags          = flags,
1321                 .bkey_type      = __btree_node_type(iter->path->level, iter->btree_id),
1322                 .btree_id       = iter->btree_id,
1323                 .level          = iter->path->level,
1324                 .cached         = iter->flags & BTREE_ITER_CACHED,
1325                 .path           = iter->path,
1326                 .k              = k,
1327                 .ip_allocated   = _RET_IP_,
1328         };
1329
1330 #ifdef CONFIG_BCACHEFS_DEBUG
1331         trans_for_each_update(trans, i)
1332                 BUG_ON(i != trans->updates &&
1333                        btree_insert_entry_cmp(i - 1, i) >= 0);
1334 #endif
1335
1336         if (bkey_deleted(&n.k->k) &&
1337             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1338                 int ret = need_whiteout_for_snapshot(trans, iter);
1339                 if (unlikely(ret < 0))
1340                         return ret;
1341
1342                 if (ret)
1343                         n.k->k.type = KEY_TYPE_whiteout;
1344         }
1345
1346         /*
1347          * Pending updates are kept sorted: first, find position of new update,
1348          * then delete/trim any updates the new update overwrites:
1349          */
1350         trans_for_each_update(trans, i)
1351                 if (btree_insert_entry_cmp(&n, i) <= 0)
1352                         break;
1353
1354         if (i < trans->updates + trans->nr_updates &&
1355             !btree_insert_entry_cmp(&n, i)) {
1356                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1357
1358                 /*
1359                  * This is a hack to ensure that inode creates update the btree,
1360                  * not the key cache, which helps with cache coherency issues in
1361                  * other areas:
1362                  */
1363                 if (n.cached && !i->cached) {
1364                         i->k = n.k;
1365                         i->flags = n.flags;
1366                         return 0;
1367                 }
1368
1369                 bch2_path_put(trans, i->path, true);
1370                 *i = n;
1371         } else
1372                 array_insert_item(trans->updates, trans->nr_updates,
1373                                   i - trans->updates, n);
1374
1375         __btree_path_get(n.path, true);
1376
1377         return 0;
1378 }
1379
1380 void bch2_trans_commit_hook(struct btree_trans *trans,
1381                             struct btree_trans_commit_hook *h)
1382 {
1383         h->next = trans->hooks;
1384         trans->hooks = h;
1385 }
1386
1387 int __bch2_btree_insert(struct btree_trans *trans,
1388                         enum btree_id id, struct bkey_i *k)
1389 {
1390         struct btree_iter iter;
1391         int ret;
1392
1393         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1394                              BTREE_ITER_INTENT);
1395         ret   = bch2_btree_iter_traverse(&iter) ?:
1396                 bch2_trans_update(trans, &iter, k, 0);
1397         bch2_trans_iter_exit(trans, &iter);
1398         return ret;
1399 }
1400
1401 /**
1402  * bch2_btree_insert - insert keys into the extent btree
1403  * @c:                  pointer to struct bch_fs
1404  * @id:                 btree to insert into
1405  * @insert_keys:        list of keys to insert
1406  * @hook:               insert callback
1407  */
1408 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1409                       struct bkey_i *k,
1410                       struct disk_reservation *disk_res,
1411                       u64 *journal_seq, int flags)
1412 {
1413         return bch2_trans_do(c, disk_res, journal_seq, flags,
1414                              __bch2_btree_insert(&trans, id, k));
1415 }
1416
1417 int bch2_btree_delete_at(struct btree_trans *trans,
1418                          struct btree_iter *iter, unsigned update_flags)
1419 {
1420         struct bkey_i *k;
1421
1422         k = bch2_trans_kmalloc(trans, sizeof(*k));
1423         if (IS_ERR(k))
1424                 return PTR_ERR(k);
1425
1426         bkey_init(&k->k);
1427         k->k.p = iter->pos;
1428         return bch2_trans_update(trans, iter, k, update_flags);
1429 }
1430
1431 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1432                                   struct bpos start, struct bpos end,
1433                                   unsigned iter_flags,
1434                                   u64 *journal_seq)
1435 {
1436         struct btree_iter iter;
1437         struct bkey_s_c k;
1438         int ret = 0;
1439
1440         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT|iter_flags);
1441 retry:
1442         while ((bch2_trans_begin(trans),
1443                (k = bch2_btree_iter_peek(&iter)).k) &&
1444                !(ret = bkey_err(k)) &&
1445                bkey_cmp(iter.pos, end) < 0) {
1446                 struct disk_reservation disk_res =
1447                         bch2_disk_reservation_init(trans->c, 0);
1448                 struct bkey_i delete;
1449
1450                 bkey_init(&delete.k);
1451
1452                 /*
1453                  * This could probably be more efficient for extents:
1454                  */
1455
1456                 /*
1457                  * For extents, iter.pos won't necessarily be the same as
1458                  * bkey_start_pos(k.k) (for non extents they always will be the
1459                  * same). It's important that we delete starting from iter.pos
1460                  * because the range we want to delete could start in the middle
1461                  * of k.
1462                  *
1463                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1464                  * bkey_start_pos(k.k)).
1465                  */
1466                 delete.k.p = iter.pos;
1467
1468                 if (btree_node_type_is_extents(id)) {
1469                         unsigned max_sectors =
1470                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1471
1472                         /* create the biggest key we can */
1473                         bch2_key_resize(&delete.k, max_sectors);
1474                         bch2_cut_back(end, &delete);
1475
1476                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1477                         if (ret)
1478                                 break;
1479                 }
1480
1481                 ret   = bch2_trans_update(trans, &iter, &delete, 0) ?:
1482                         bch2_trans_commit(trans, &disk_res, journal_seq,
1483                                         BTREE_INSERT_NOFAIL);
1484                 bch2_disk_reservation_put(trans->c, &disk_res);
1485                 if (ret)
1486                         break;
1487         }
1488
1489         if (ret == -EINTR) {
1490                 ret = 0;
1491                 goto retry;
1492         }
1493
1494         bch2_trans_iter_exit(trans, &iter);
1495         return ret;
1496 }
1497
1498 /*
1499  * bch_btree_delete_range - delete everything within a given range
1500  *
1501  * Range is a half open interval - [start, end)
1502  */
1503 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1504                             struct bpos start, struct bpos end,
1505                             u64 *journal_seq)
1506 {
1507         return bch2_trans_do(c, NULL, journal_seq, 0,
1508                              bch2_btree_delete_range_trans(&trans, id, start, end, 0, journal_seq));
1509 }