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                 bch2_journal_set_has_inode(j, &trans->journal_res,
 366                                            i->k->k.p.inode);
 367
 368                 if (trans->journal_seq)
 369                         *trans->journal_seq = trans->journal_res.seq;
 370         }
 371 }
 372
 373 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 374 {
 375         struct bch_fs *c = trans->c;
 376         struct btree_insert_entry *i;
 377
 378         trans_for_each_update(trans, i) {
 379                 /*
 380                  * XXX: synchronization of cached update triggers with gc
 381                  * XXX: synchronization of interior node updates with gc
 382                  */
 383                 BUG_ON(i->cached || i->level);
 384
 385                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b)))
 386                         bch2_mark_update(trans, i->path, i->k,
 387                                          i->flags|BTREE_TRIGGER_GC);
 388         }
 389 }
 390
 391 static inline int
 392 bch2_trans_commit_write_locked(struct btree_trans *trans,
 393                                struct btree_insert_entry **stopped_at,
 394                                unsigned long trace_ip)
 395 {
 396         struct bch_fs *c = trans->c;
 397         struct btree_insert_entry *i;
 398         struct btree_trans_commit_hook *h;
 399         unsigned u64s = 0;
 400         bool marking = false;
 401         int ret;
 402
 403         if (race_fault()) {
 404                 trace_trans_restart_fault_inject(trans->ip, trace_ip);
 405                 trans->restarted = true;
 406                 return -EINTR;
 407         }
 408
 409         /*
 410          * Check if the insert will fit in the leaf node with the write lock
 411          * held, otherwise another thread could write the node changing the
 412          * amount of space available:
 413          */
 414
 415         prefetch(&trans->c->journal.flags);
 416
 417         h = trans->hooks;
 418         while (h) {
 419                 ret = h->fn(trans, h);
 420                 if (ret)
 421                         return ret;
 422                 h = h->next;
 423         }
 424
 425         trans_for_each_update(trans, i) {
 426                 /* Multiple inserts might go to same leaf: */
 427                 if (!same_leaf_as_prev(trans, i))
 428                         u64s = 0;
 429
 430                 u64s += i->k->k.u64s;
 431                 ret = !i->cached
 432                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 433                         : btree_key_can_insert_cached(trans, i->path, u64s);
 434                 if (ret) {
 435                         *stopped_at = i;
 436                         return ret;
 437                 }
 438
 439                 if (btree_node_type_needs_gc(i->bkey_type))
 440                         marking = true;
 441         }
 442
 443         if (marking) {
 444                 percpu_down_read(&c->mark_lock);
 445         }
 446
 447         /* Must be called under mark_lock: */
 448         if (marking && trans->fs_usage_deltas &&
 449             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 450                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 451                 goto err;
 452         }
 453
 454         /*
 455          * Don't get journal reservation until after we know insert will
 456          * succeed:
 457          */
 458         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 459                 ret = bch2_trans_journal_res_get(trans,
 460                                 JOURNAL_RES_GET_NONBLOCK);
 461                 if (ret)
 462                         goto err;
 463         } else {
 464                 trans->journal_res.seq = c->journal.replay_journal_seq;
 465         }
 466
 467         if (unlikely(trans->extra_journal_entry_u64s)) {
 468                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 469                                   trans->extra_journal_entries,
 470                                   trans->extra_journal_entry_u64s);
 471
 472                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 473                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 474         }
 475
 476         /*
 477          * Not allowed to fail after we've gotten our journal reservation - we
 478          * have to use it:
 479          */
 480
 481         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 482                 if (bch2_journal_seq_verify)
 483                         trans_for_each_update(trans, i)
 484                                 i->k->k.version.lo = trans->journal_res.seq;
 485                 else if (bch2_inject_invalid_keys)
 486                         trans_for_each_update(trans, i)
 487                                 i->k->k.version = MAX_VERSION;
 488         }
 489
 490         trans_for_each_update(trans, i)
 491                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 492                         bch2_mark_update(trans, i->path, i->k, i->flags);
 493
 494         if (marking && trans->fs_usage_deltas)
 495                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 496
 497         if (unlikely(c->gc_pos.phase))
 498                 bch2_trans_mark_gc(trans);
 499
 500         trans_for_each_update(trans, i)
 501                 do_btree_insert_one(trans, i);
 502 err:
 503         if (marking) {
 504                 percpu_up_read(&c->mark_lock);
 505         }
 506
 507         return ret;
 508 }
 509
 510 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 511 {
 512         struct btree *b = path_l(path)->b;
 513
 514         do {
 515                 if (path->nodes_locked &&
 516                     path->nodes_locked != path->nodes_intent_locked)
 517                         BUG_ON(!bch2_btree_path_upgrade(trans, path, path->level + 1));
 518         } while ((path = prev_btree_path(trans, path)) &&
 519                  path_l(path)->b == b);
 520 }
 521
 522 /*
 523  * Check for nodes that we have both read and intent locks on, and upgrade the
 524  * readers to intent:
 525  */
 526 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 527 {
 528         struct btree_path *path;
 529         unsigned i, nr_read = 0, nr_intent = 0;
 530
 531         trans_for_each_path_inorder(trans, path, i) {
 532                 struct btree_path *next = i + 1 < trans->nr_sorted
 533                         ? trans->paths + trans->sorted[i + 1]
 534                         : NULL;
 535
 536                 if (path->nodes_locked) {
 537                         if (path->nodes_intent_locked)
 538                                 nr_intent++;
 539                         else
 540                                 nr_read++;
 541                 }
 542
 543                 if (!next || path_l(path)->b != path_l(next)->b) {
 544                         if (nr_read && nr_intent)
 545                                 upgrade_readers(trans, path);
 546
 547                         nr_read = nr_intent = 0;
 548                 }
 549         }
 550
 551         bch2_trans_verify_locks(trans);
 552 }
 553
 554 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 555 {
 556         struct btree_path *path;
 557         unsigned i;
 558
 559         trans_for_each_path_inorder(trans, path, i) {
 560                 //if (path == pos)
 561                 //      break;
 562
 563                 if (path->nodes_locked != path->nodes_intent_locked)
 564                         return true;
 565         }
 566
 567         return false;
 568 }
 569
 570 static inline int trans_lock_write(struct btree_trans *trans)
 571 {
 572         struct btree_insert_entry *i;
 573
 574         trans_for_each_update(trans, i) {
 575                 if (same_leaf_as_prev(trans, i))
 576                         continue;
 577
 578                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 579                         if (have_conflicting_read_lock(trans, i->path))
 580                                 goto fail;
 581
 582                         __btree_node_lock_type(trans->c, insert_l(i)->b,
 583                                                SIX_LOCK_write);
 584                 }
 585
 586                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 587         }
 588
 589         return 0;
 590 fail:
 591         while (--i >= trans->updates) {
 592                 if (same_leaf_as_prev(trans, i))
 593                         continue;
 594
 595                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 596         }
 597
 598         trace_trans_restart_would_deadlock_write(trans->ip);
 599         return btree_trans_restart(trans);
 600 }
 601
 602 /*
 603  * Get journal reservation, take write locks, and attempt to do btree update(s):
 604  */
 605 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 606                                        struct btree_insert_entry **stopped_at,
 607                                        unsigned long trace_ip)
 608 {
 609         struct bch_fs *c = trans->c;
 610         struct btree_insert_entry *i;
 611         struct bkey_s_c old;
 612         int ret, u64s_delta = 0;
 613
 614         trans_for_each_update(trans, i) {
 615                 const char *invalid = bch2_bkey_invalid(c,
 616                                 bkey_i_to_s_c(i->k), i->bkey_type);
 617                 if (invalid) {
 618                         char buf[200];
 619
 620                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 621                         bch_err(c, "invalid bkey %s on insert from %ps -> %ps: %s\n",
 622                                 buf, (void *) trans->ip,
 623                                 (void *) i->ip_allocated, invalid);
 624                         bch2_fatal_error(c);
 625                         return -EINVAL;
 626                 }
 627                 btree_insert_entry_checks(trans, i);
 628         }
 629
 630         trans_for_each_update(trans, i) {
 631                 struct bkey u;
 632
 633                 /*
 634                  * peek_slot() doesn't yet work on iterators that point to
 635                  * interior nodes:
 636                  */
 637                 if (i->cached || i->level)
 638                         continue;
 639
 640                 old = bch2_btree_path_peek_slot(i->path, &u);
 641                 ret = bkey_err(old);
 642                 if (unlikely(ret))
 643                         return ret;
 644
 645                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 646                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 647
 648                 if (!same_leaf_as_next(trans, i)) {
 649                         if (u64s_delta <= 0) {
 650                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 651                                                         i->level, trans->flags);
 652                                 if (unlikely(ret))
 653                                         return ret;
 654                         }
 655
 656                         u64s_delta = 0;
 657                 }
 658         }
 659
 660         ret = bch2_journal_preres_get(&c->journal,
 661                         &trans->journal_preres, trans->journal_preres_u64s,
 662                         JOURNAL_RES_GET_NONBLOCK|
 663                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 664                          ? JOURNAL_RES_GET_RESERVED : 0));
 665         if (unlikely(ret == -EAGAIN))
 666                 ret = bch2_trans_journal_preres_get_cold(trans,
 667                                                 trans->journal_preres_u64s, trace_ip);
 668         if (unlikely(ret))
 669                 return ret;
 670
 671         normalize_read_intent_locks(trans);
 672
 673         ret = trans_lock_write(trans);
 674         if (unlikely(ret))
 675                 return ret;
 676
 677         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 678
 679         trans_for_each_update(trans, i)
 680                 if (!same_leaf_as_prev(trans, i))
 681                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 682                                                         insert_l(i)->b);
 683
 684         if (!ret && trans->journal_pin)
 685                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 686                                      trans->journal_pin, NULL);
 687
 688         /*
 689          * Drop journal reservation after dropping write locks, since dropping
 690          * the journal reservation may kick off a journal write:
 691          */
 692         bch2_journal_res_put(&c->journal, &trans->journal_res);
 693
 694         if (unlikely(ret))
 695                 return ret;
 696
 697         bch2_trans_downgrade(trans);
 698
 699         return 0;
 700 }
 701
 702 static int journal_reclaim_wait_done(struct bch_fs *c)
 703 {
 704         int ret = bch2_journal_error(&c->journal) ?:
 705                 !bch2_btree_key_cache_must_wait(c);
 706
 707         if (!ret)
 708                 journal_reclaim_kick(&c->journal);
 709         return ret;
 710 }
 711
 712 static noinline
 713 int bch2_trans_commit_error(struct btree_trans *trans,
 714                             struct btree_insert_entry *i,
 715                             int ret, unsigned long trace_ip)
 716 {
 717         struct bch_fs *c = trans->c;
 718
 719         switch (ret) {
 720         case BTREE_INSERT_BTREE_NODE_FULL:
 721                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 722                 if (!ret)
 723                         return 0;
 724
 725                 if (ret == -EINTR)
 726                         trace_trans_restart_btree_node_split(trans->ip, trace_ip,
 727                                                 i->btree_id, &i->path->pos);
 728                 break;
 729         case BTREE_INSERT_NEED_MARK_REPLICAS:
 730                 bch2_trans_unlock(trans);
 731
 732                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 733                 if (ret)
 734                         break;
 735
 736                 if (bch2_trans_relock(trans))
 737                         return 0;
 738
 739                 trace_trans_restart_mark_replicas(trans->ip, trace_ip);
 740                 ret = -EINTR;
 741                 break;
 742         case BTREE_INSERT_NEED_JOURNAL_RES:
 743                 bch2_trans_unlock(trans);
 744
 745                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 746                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
 747                         trans->restarted = true;
 748                         ret = -EAGAIN;
 749                         break;
 750                 }
 751
 752                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 753                 if (ret)
 754                         break;
 755
 756                 if (bch2_trans_relock(trans))
 757                         return 0;
 758
 759                 trace_trans_restart_journal_res_get(trans->ip, trace_ip);
 760                 ret = -EINTR;
 761                 break;
 762         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 763                 bch2_trans_unlock(trans);
 764
 765                 trace_trans_blocked_journal_reclaim(trans->ip, trace_ip);
 766
 767                 wait_event_freezable(c->journal.reclaim_wait,
 768                                      (ret = journal_reclaim_wait_done(c)));
 769                 if (ret < 0)
 770                         break;
 771
 772                 if (bch2_trans_relock(trans))
 773                         return 0;
 774
 775                 trace_trans_restart_journal_reclaim(trans->ip, trace_ip);
 776                 ret = -EINTR;
 777                 break;
 778         default:
 779                 BUG_ON(ret >= 0);
 780                 break;
 781         }
 782
 783         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
 784         BUG_ON(ret == -ENOSPC && (trans->flags & BTREE_INSERT_NOFAIL));
 785
 786         return ret;
 787 }
 788
 789 static noinline int
 790 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 791 {
 792         struct bch_fs *c = trans->c;
 793         int ret;
 794
 795         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 796                 return -EROFS;
 797
 798         bch2_trans_unlock(trans);
 799
 800         ret = bch2_fs_read_write_early(c);
 801         if (ret)
 802                 return ret;
 803
 804         if (!bch2_trans_relock(trans))
 805                 return -EINTR;
 806
 807         percpu_ref_get(&c->writes);
 808         return 0;
 809 }
 810
 811 int __bch2_trans_commit(struct btree_trans *trans)
 812 {
 813         struct btree_insert_entry *i = NULL;
 814         bool trans_trigger_run;
 815         unsigned u64s;
 816         int ret = 0;
 817
 818         if (!trans->nr_updates &&
 819             !trans->extra_journal_entry_u64s)
 820                 goto out_reset;
 821
 822         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 823                 lockdep_assert_held(&trans->c->gc_lock);
 824
 825         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 826
 827         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 828         trans->journal_preres_u64s      = 0;
 829
 830         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 831             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 832                 ret = bch2_trans_commit_get_rw_cold(trans);
 833                 if (ret)
 834                         goto out_reset;
 835         }
 836
 837 #ifdef CONFIG_BCACHEFS_DEBUG
 838         /*
 839          * if BTREE_TRIGGER_NORUN is set, it means we're probably being called
 840          * from the key cache flush code:
 841          */
 842         trans_for_each_update(trans, i)
 843                 if (!i->cached &&
 844                     !(i->flags & BTREE_TRIGGER_NORUN))
 845                         bch2_btree_key_cache_verify_clean(trans,
 846                                         i->btree_id, i->k->k.p);
 847 #endif
 848
 849         /*
 850          * Running triggers will append more updates to the list of updates as
 851          * we're walking it:
 852          */
 853         do {
 854                 trans_trigger_run = false;
 855
 856                 trans_for_each_update(trans, i) {
 857                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 858                             !i->trans_triggers_run) {
 859                                 i->trans_triggers_run = true;
 860                                 trans_trigger_run = true;
 861
 862                                 ret = bch2_trans_mark_update(trans, i->path,
 863                                                              i->k, i->flags);
 864                                 if (unlikely(ret)) {
 865                                         if (ret == -EINTR)
 866                                                 trace_trans_restart_mark(trans->ip, _RET_IP_,
 867                                                                 i->btree_id, &i->path->pos);
 868                                         goto out;
 869                                 }
 870                         }
 871                 }
 872         } while (trans_trigger_run);
 873
 874         trans_for_each_update(trans, i) {
 875                 BUG_ON(!i->path->should_be_locked);
 876
 877                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
 878                         trace_trans_restart_upgrade(trans->ip, _RET_IP_,
 879                                                     i->btree_id, &i->path->pos);
 880                         ret = btree_trans_restart(trans);
 881                         goto out;
 882                 }
 883
 884                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
 885
 886                 u64s = jset_u64s(i->k->k.u64s);
 887                 if (i->cached &&
 888                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 889                         trans->journal_preres_u64s += u64s;
 890                 trans->journal_u64s += u64s;
 891         }
 892
 893         if (trans->extra_journal_res) {
 894                 ret = bch2_disk_reservation_add(trans->c, trans->disk_res,
 895                                 trans->extra_journal_res,
 896                                 (trans->flags & BTREE_INSERT_NOFAIL)
 897                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
 898                 if (ret)
 899                         goto err;
 900         }
 901 retry:
 902         BUG_ON(trans->restarted);
 903         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 904
 905         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
 906
 907         /* make sure we didn't drop or screw up locks: */
 908         bch2_trans_verify_locks(trans);
 909
 910         if (ret)
 911                 goto err;
 912 out:
 913         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 914
 915         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 916                 percpu_ref_put(&trans->c->writes);
 917 out_reset:
 918         trans_for_each_update(trans, i)
 919                 bch2_path_put(trans, i->path, true);
 920
 921         trans->extra_journal_res        = 0;
 922         trans->nr_updates               = 0;
 923         trans->hooks                    = NULL;
 924         trans->extra_journal_entries    = NULL;
 925         trans->extra_journal_entry_u64s = 0;
 926
 927         if (trans->fs_usage_deltas) {
 928                 trans->fs_usage_deltas->used = 0;
 929                 memset(&trans->fs_usage_deltas->memset_start, 0,
 930                        (void *) &trans->fs_usage_deltas->memset_end -
 931                        (void *) &trans->fs_usage_deltas->memset_start);
 932         }
 933
 934         return ret;
 935 err:
 936         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
 937         if (ret)
 938                 goto out;
 939
 940         goto retry;
 941 }
 942
 943 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
 944                                           enum btree_id id,
 945                                           struct bpos pos)
 946 {
 947         struct bch_fs *c = trans->c;
 948         struct btree_iter iter;
 949         struct bkey_s_c k;
 950         int ret;
 951
 952         if (!snapshot_t(c, pos.snapshot)->children[0])
 953                 return 0;
 954
 955         bch2_trans_iter_init(trans, &iter, id, pos,
 956                              BTREE_ITER_NOT_EXTENTS|
 957                              BTREE_ITER_ALL_SNAPSHOTS);
 958         while (1) {
 959                 k = bch2_btree_iter_prev(&iter);
 960                 ret = bkey_err(k);
 961                 if (ret)
 962                         break;
 963
 964                 if (!k.k)
 965                         break;
 966
 967                 if (bkey_cmp(pos, k.k->p))
 968                         break;
 969
 970                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
 971                         ret = 1;
 972                         break;
 973                 }
 974         }
 975         bch2_trans_iter_exit(trans, &iter);
 976
 977         return ret;
 978 }
 979
 980 static int bch2_trans_update_extent(struct btree_trans *trans,
 981                                     struct btree_iter *orig_iter,
 982                                     struct bkey_i *insert,
 983                                     enum btree_update_flags flags)
 984 {
 985         struct bch_fs *c = trans->c;
 986         struct btree_iter iter, update_iter;
 987         struct bpos start = bkey_start_pos(&insert->k);
 988         struct bkey_i *update;
 989         struct bkey_s_c k;
 990         enum btree_id btree_id = orig_iter->btree_id;
 991         int ret = 0, compressed_sectors;
 992
 993         bch2_trans_iter_init(trans, &iter, btree_id, start,
 994                              BTREE_ITER_INTENT|
 995                              BTREE_ITER_WITH_UPDATES|
 996                              BTREE_ITER_NOT_EXTENTS);
 997         k = bch2_btree_iter_peek(&iter);
 998         if ((ret = bkey_err(k)))
 999                 goto err;
1000         if (!k.k)
1001                 goto out;
1002
1003         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1004                 /*
1005                  * We can't merge extents if they belong to interior snapshot
1006                  * tree nodes, and there's a snapshot in which one extent is
1007                  * visible and the other is not - i.e. if visibility is
1008                  * different.
1009                  *
1010                  * Instead of checking if visibilitiy of the two extents is
1011                  * different, for now we just check if either has been
1012                  * overwritten:
1013                  */
1014                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1015                 if (ret < 0)
1016                         goto err;
1017                 if (ret)
1018                         goto nomerge1;
1019
1020                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1021                 if (ret < 0)
1022                         goto err;
1023                 if (ret)
1024                         goto nomerge1;
1025
1026                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1027                 if ((ret = PTR_ERR_OR_ZERO(update)))
1028                         goto err;
1029
1030                 bkey_reassemble(update, k);
1031
1032                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1033                         ret = bch2_btree_delete_at(trans, &iter, flags);
1034                         if (ret)
1035                                 goto err;
1036
1037                         insert = update;
1038                         goto next;
1039                 }
1040         }
1041 nomerge1:
1042         ret = 0;
1043         if (!bkey_cmp(k.k->p, start))
1044                 goto next;
1045
1046         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1047                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1048                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1049
1050                 /*
1051                  * If we're going to be splitting a compressed extent, note it
1052                  * so that __bch2_trans_commit() can increase our disk
1053                  * reservation:
1054                  */
1055                 if (((front_split && back_split) ||
1056                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1057                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1058                         trans->extra_journal_res += compressed_sectors;
1059
1060                 if (front_split) {
1061                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1062                         if ((ret = PTR_ERR_OR_ZERO(update)))
1063                                 goto err;
1064
1065                         bkey_reassemble(update, k);
1066
1067                         bch2_cut_back(start, update);
1068
1069                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1070                                              BTREE_ITER_NOT_EXTENTS|
1071                                              BTREE_ITER_ALL_SNAPSHOTS|
1072                                              BTREE_ITER_INTENT);
1073                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1074                                 bch2_trans_update(trans, &update_iter, update,
1075                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1076                                                   flags);
1077                         bch2_trans_iter_exit(trans, &update_iter);
1078
1079                         if (ret)
1080                                 goto err;
1081                 }
1082
1083                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1084                     (front_split || back_split)) {
1085                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1086                         if ((ret = PTR_ERR_OR_ZERO(update)))
1087                                 goto err;
1088
1089                         bkey_reassemble(update, k);
1090
1091                         bch2_cut_front(start, update);
1092                         bch2_cut_back(insert->k.p, update);
1093
1094                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1095                                              BTREE_ITER_NOT_EXTENTS|
1096                                              BTREE_ITER_ALL_SNAPSHOTS|
1097                                              BTREE_ITER_INTENT);
1098                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1099                                 bch2_trans_update(trans, &update_iter, update,
1100                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1101                                                   flags);
1102                         bch2_trans_iter_exit(trans, &update_iter);
1103                         if (ret)
1104                                 goto err;
1105                 }
1106
1107                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1108                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1109                         if ((ret = PTR_ERR_OR_ZERO(update)))
1110                                 goto err;
1111
1112                         bkey_init(&update->k);
1113                         update->k.p = k.k->p;
1114
1115                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1116                                 update->k.p.snapshot = insert->k.p.snapshot;
1117                                 update->k.type = KEY_TYPE_whiteout;
1118                         }
1119
1120                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1121                                              BTREE_ITER_NOT_EXTENTS|
1122                                              BTREE_ITER_INTENT);
1123                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1124                                 bch2_trans_update(trans, &update_iter, update,
1125                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1126                                                   flags);
1127                         bch2_trans_iter_exit(trans, &update_iter);
1128
1129                         if (ret)
1130                                 goto err;
1131                 }
1132
1133                 if (back_split) {
1134                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1135                         if ((ret = PTR_ERR_OR_ZERO(update)))
1136                                 goto err;
1137
1138                         bkey_reassemble(update, k);
1139                         bch2_cut_front(insert->k.p, update);
1140
1141                         bch2_trans_copy_iter(&update_iter, &iter);
1142                         update_iter.pos = update->k.p;
1143                         ret   = bch2_trans_update(trans, &update_iter, update,
1144                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1145                                                   flags);
1146                         bch2_trans_iter_exit(trans, &update_iter);
1147
1148                         if (ret)
1149                                 goto err;
1150                         goto out;
1151                 }
1152 next:
1153                 k = bch2_btree_iter_next(&iter);
1154                 if ((ret = bkey_err(k)))
1155                         goto err;
1156                 if (!k.k)
1157                         goto out;
1158         }
1159
1160         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1161                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1162                 if (ret < 0)
1163                         goto out;
1164                 if (ret)
1165                         goto nomerge2;
1166
1167                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1168                 if (ret < 0)
1169                         goto out;
1170                 if (ret)
1171                         goto nomerge2;
1172
1173                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1174         }
1175 nomerge2:
1176         ret = 0;
1177 out:
1178         if (!bkey_deleted(&insert->k)) {
1179                 /*
1180                  * Rewinding iterators is expensive: get a new one and the one
1181                  * that points to the start of insert will be cloned from:
1182                  */
1183                 bch2_trans_iter_exit(trans, &iter);
1184                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1185                                      BTREE_ITER_NOT_EXTENTS|
1186                                      BTREE_ITER_INTENT);
1187                 ret   = bch2_btree_iter_traverse(&iter) ?:
1188                         bch2_trans_update(trans, &iter, insert, flags);
1189         }
1190 err:
1191         bch2_trans_iter_exit(trans, &iter);
1192
1193         return ret;
1194 }
1195
1196 /*
1197  * When deleting, check if we need to emit a whiteout (because we're overwriting
1198  * something in an ancestor snapshot)
1199  */
1200 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1201                                       enum btree_id btree_id, struct bpos pos)
1202 {
1203         struct btree_iter iter;
1204         struct bkey_s_c k;
1205         u32 snapshot = pos.snapshot;
1206         int ret;
1207
1208         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1209                 return 0;
1210
1211         pos.snapshot++;
1212
1213         for_each_btree_key(trans, iter, btree_id, pos,
1214                            BTREE_ITER_ALL_SNAPSHOTS, k, ret) {
1215                 if (bkey_cmp(k.k->p, pos))
1216                         break;
1217
1218                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1219                                               k.k->p.snapshot)) {
1220                         ret = !bkey_whiteout(k.k);
1221                         break;
1222                 }
1223         }
1224         bch2_trans_iter_exit(trans, &iter);
1225
1226         return ret;
1227 }
1228
1229 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1230                       struct bkey_i *k, enum btree_update_flags flags)
1231 {
1232         struct btree_insert_entry *i, n;
1233
1234         BUG_ON(!iter->path->should_be_locked);
1235
1236         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1237                 return bch2_trans_update_extent(trans, iter, k, flags);
1238
1239         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1240         BUG_ON(bpos_cmp(k->k.p, iter->path->pos));
1241
1242         n = (struct btree_insert_entry) {
1243                 .flags          = flags,
1244                 .bkey_type      = __btree_node_type(iter->path->level, iter->btree_id),
1245                 .btree_id       = iter->btree_id,
1246                 .level          = iter->path->level,
1247                 .cached         = iter->flags & BTREE_ITER_CACHED,
1248                 .path           = iter->path,
1249                 .k              = k,
1250                 .ip_allocated   = _RET_IP_,
1251         };
1252
1253         __btree_path_get(n.path, true);
1254
1255 #ifdef CONFIG_BCACHEFS_DEBUG
1256         trans_for_each_update(trans, i)
1257                 BUG_ON(i != trans->updates &&
1258                        btree_insert_entry_cmp(i - 1, i) >= 0);
1259 #endif
1260
1261         if (bkey_deleted(&n.k->k) &&
1262             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1263                 int ret = need_whiteout_for_snapshot(trans, n.btree_id, n.k->k.p);
1264                 if (unlikely(ret < 0))
1265                         return ret;
1266
1267                 if (ret)
1268                         n.k->k.type = KEY_TYPE_whiteout;
1269         }
1270
1271         /*
1272          * Pending updates are kept sorted: first, find position of new update,
1273          * then delete/trim any updates the new update overwrites:
1274          */
1275         trans_for_each_update(trans, i)
1276                 if (btree_insert_entry_cmp(&n, i) <= 0)
1277                         break;
1278
1279         if (i < trans->updates + trans->nr_updates &&
1280             !btree_insert_entry_cmp(&n, i)) {
1281                 BUG_ON(i->trans_triggers_run);
1282
1283                 /*
1284                  * This is a hack to ensure that inode creates update the btree,
1285                  * not the key cache, which helps with cache coherency issues in
1286                  * other areas:
1287                  */
1288                 if (n.cached && !i->cached) {
1289                         i->k = n.k;
1290                         i->flags = n.flags;
1291
1292                         __btree_path_get(n.path, false);
1293                 } else {
1294                         bch2_path_put(trans, i->path, true);
1295                         *i = n;
1296                 }
1297         } else
1298                 array_insert_item(trans->updates, trans->nr_updates,
1299                                   i - trans->updates, n);
1300
1301         return 0;
1302 }
1303
1304 void bch2_trans_commit_hook(struct btree_trans *trans,
1305                             struct btree_trans_commit_hook *h)
1306 {
1307         h->next = trans->hooks;
1308         trans->hooks = h;
1309 }
1310
1311 int __bch2_btree_insert(struct btree_trans *trans,
1312                         enum btree_id id, struct bkey_i *k)
1313 {
1314         struct btree_iter iter;
1315         int ret;
1316
1317         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1318                              BTREE_ITER_INTENT);
1319         ret   = bch2_btree_iter_traverse(&iter) ?:
1320                 bch2_trans_update(trans, &iter, k, 0);
1321         bch2_trans_iter_exit(trans, &iter);
1322         return ret;
1323 }
1324
1325 /**
1326  * bch2_btree_insert - insert keys into the extent btree
1327  * @c:                  pointer to struct bch_fs
1328  * @id:                 btree to insert into
1329  * @insert_keys:        list of keys to insert
1330  * @hook:               insert callback
1331  */
1332 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1333                       struct bkey_i *k,
1334                       struct disk_reservation *disk_res,
1335                       u64 *journal_seq, int flags)
1336 {
1337         return bch2_trans_do(c, disk_res, journal_seq, flags,
1338                              __bch2_btree_insert(&trans, id, k));
1339 }
1340
1341 int bch2_btree_delete_at(struct btree_trans *trans,
1342                          struct btree_iter *iter, unsigned update_flags)
1343 {
1344         struct bkey_i *k;
1345
1346         k = bch2_trans_kmalloc(trans, sizeof(*k));
1347         if (IS_ERR(k))
1348                 return PTR_ERR(k);
1349
1350         bkey_init(&k->k);
1351         k->k.p = iter->pos;
1352         return bch2_trans_update(trans, iter, k, update_flags);
1353 }
1354
1355 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1356                                   struct bpos start, struct bpos end,
1357                                   unsigned iter_flags,
1358                                   u64 *journal_seq)
1359 {
1360         struct btree_iter iter;
1361         struct bkey_s_c k;
1362         int ret = 0;
1363
1364         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT|iter_flags);
1365 retry:
1366         while ((bch2_trans_begin(trans),
1367                (k = bch2_btree_iter_peek(&iter)).k) &&
1368                !(ret = bkey_err(k)) &&
1369                bkey_cmp(iter.pos, end) < 0) {
1370                 struct bkey_i delete;
1371
1372                 bkey_init(&delete.k);
1373
1374                 /*
1375                  * This could probably be more efficient for extents:
1376                  */
1377
1378                 /*
1379                  * For extents, iter.pos won't necessarily be the same as
1380                  * bkey_start_pos(k.k) (for non extents they always will be the
1381                  * same). It's important that we delete starting from iter.pos
1382                  * because the range we want to delete could start in the middle
1383                  * of k.
1384                  *
1385                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1386                  * bkey_start_pos(k.k)).
1387                  */
1388                 delete.k.p = iter.pos;
1389
1390                 if (btree_node_type_is_extents(id)) {
1391                         unsigned max_sectors =
1392                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1393
1394                         /* create the biggest key we can */
1395                         bch2_key_resize(&delete.k, max_sectors);
1396                         bch2_cut_back(end, &delete);
1397
1398                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1399                         if (ret)
1400                                 break;
1401                 }
1402
1403                 ret   = bch2_trans_update(trans, &iter, &delete, 0) ?:
1404                         bch2_trans_commit(trans, NULL, journal_seq,
1405                                         BTREE_INSERT_NOFAIL);
1406                 if (ret)
1407                         break;
1408
1409                 bch2_trans_cond_resched(trans);
1410         }
1411
1412         if (ret == -EINTR) {
1413                 ret = 0;
1414                 goto retry;
1415         }
1416
1417         bch2_trans_iter_exit(trans, &iter);
1418         return ret;
1419 }
1420
1421 /*
1422  * bch_btree_delete_range - delete everything within a given range
1423  *
1424  * Range is a half open interval - [start, end)
1425  */
1426 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1427                             struct bpos start, struct bpos end,
1428                             u64 *journal_seq)
1429 {
1430         return bch2_trans_do(c, NULL, journal_seq, 0,
1431                              bch2_btree_delete_range_trans(&trans, id, start, end, 0, journal_seq));
1432 }