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 "replicas.h"
  19
  20 #include <linux/prefetch.h>
  21 #include <linux/sort.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  25                                          const struct btree_insert_entry *r)
  26 {
  27         return   cmp_int(l->btree_id,   r->btree_id) ?:
  28                  -cmp_int(l->level,     r->level) ?:
  29                  bpos_cmp(l->k->k.p,    r->k->k.p);
  30 }
  31
  32 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  33                                      struct btree_insert_entry *i)
  34 {
  35         return i != trans->updates2 &&
  36                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  37 }
  38
  39 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  40                                             struct btree_iter *iter)
  41 {
  42         bch2_btree_node_lock_write(b, iter);
  43
  44         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
  45                 return;
  46
  47         if (unlikely(btree_node_just_written(b)) &&
  48             bch2_btree_post_write_cleanup(c, b))
  49                 bch2_btree_iter_reinit_node(iter, b);
  50
  51         /*
  52          * If the last bset has been written, or if it's gotten too big - start
  53          * a new bset to insert into:
  54          */
  55         if (want_new_bset(c, b))
  56                 bch2_btree_init_next(c, b, iter);
  57 }
  58
  59 /* Inserting into a given leaf node (last stage of insert): */
  60
  61 /* Handle overwrites and do insert, for non extents: */
  62 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  63                                 struct btree *b,
  64                                 struct btree_node_iter *node_iter,
  65                                 struct bkey_i *insert)
  66 {
  67         struct bkey_packed *k;
  68         unsigned clobber_u64s = 0, new_u64s = 0;
  69
  70         EBUG_ON(btree_node_just_written(b));
  71         EBUG_ON(bset_written(b, btree_bset_last(b)));
  72         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  73         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  74         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  75         EBUG_ON(insert->k.u64s >
  76                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  77         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  78
  79         k = bch2_btree_node_iter_peek_all(node_iter, b);
  80         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
  81                 k = NULL;
  82
  83         /* @k is the key being overwritten/deleted, if any: */
  84         EBUG_ON(k && bkey_deleted(k));
  85
  86         /* Deleting, but not found? nothing to do: */
  87         if (bkey_deleted(&insert->k) && !k)
  88                 return false;
  89
  90         if (bkey_deleted(&insert->k)) {
  91                 /* Deleting: */
  92                 btree_account_key_drop(b, k);
  93                 k->type = KEY_TYPE_deleted;
  94
  95                 if (k->needs_whiteout)
  96                         push_whiteout(iter->trans->c, b, insert->k.p);
  97                 k->needs_whiteout = false;
  98
  99                 if (k >= btree_bset_last(b)->start) {
 100                         clobber_u64s = k->u64s;
 101                         bch2_bset_delete(b, k, clobber_u64s);
 102                         goto fix_iter;
 103                 } else {
 104                         bch2_btree_iter_fix_key_modified(iter, b, k);
 105                 }
 106
 107                 return true;
 108         }
 109
 110         if (k) {
 111                 /* Overwriting: */
 112                 btree_account_key_drop(b, k);
 113                 k->type = KEY_TYPE_deleted;
 114
 115                 insert->k.needs_whiteout = k->needs_whiteout;
 116                 k->needs_whiteout = false;
 117
 118                 if (k >= btree_bset_last(b)->start) {
 119                         clobber_u64s = k->u64s;
 120                         goto overwrite;
 121                 } else {
 122                         bch2_btree_iter_fix_key_modified(iter, b, k);
 123                 }
 124         }
 125
 126         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 127 overwrite:
 128         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 129         new_u64s = k->u64s;
 130 fix_iter:
 131         if (clobber_u64s != new_u64s)
 132                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 133                                          clobber_u64s, new_u64s);
 134         return true;
 135 }
 136
 137 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 138                                unsigned i, u64 seq)
 139 {
 140         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 141         struct btree_write *w = container_of(pin, struct btree_write, journal);
 142         struct btree *b = container_of(w, struct btree, writes[i]);
 143
 144         btree_node_lock_type(c, b, SIX_LOCK_read);
 145         bch2_btree_node_write_cond(c, b,
 146                 (btree_current_write(b) == w && w->journal.seq == seq));
 147         six_unlock_read(&b->c.lock);
 148         return 0;
 149 }
 150
 151 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 152 {
 153         return __btree_node_flush(j, pin, 0, seq);
 154 }
 155
 156 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 157 {
 158         return __btree_node_flush(j, pin, 1, seq);
 159 }
 160
 161 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 162                                        struct btree *b, u64 seq)
 163 {
 164         struct btree_write *w = btree_current_write(b);
 165
 166         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 167                              btree_node_write_idx(b) == 0
 168                              ? btree_node_flush0
 169                              : btree_node_flush1);
 170 }
 171
 172 /**
 173  * btree_insert_key - insert a key one key into a leaf node
 174  */
 175 static bool btree_insert_key_leaf(struct btree_trans *trans,
 176                                   struct btree_iter *iter,
 177                                   struct bkey_i *insert)
 178 {
 179         struct bch_fs *c = trans->c;
 180         struct btree *b = iter_l(iter)->b;
 181         struct bset_tree *t = bset_tree_last(b);
 182         struct bset *i = bset(b, t);
 183         int old_u64s = bset_u64s(t);
 184         int old_live_u64s = b->nr.live_u64s;
 185         int live_u64s_added, u64s_added;
 186
 187         EBUG_ON(!iter->level &&
 188                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 189
 190         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 191                                         &iter_l(iter)->iter, insert)))
 192                 return false;
 193
 194         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 195                                          le64_to_cpu(i->journal_seq)));
 196
 197         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 198
 199         if (unlikely(!btree_node_dirty(b)))
 200                 set_btree_node_dirty(c, b);
 201
 202         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 203         u64s_added = (int) bset_u64s(t) - old_u64s;
 204
 205         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 206                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 207         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 208                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 209
 210         if (u64s_added > live_u64s_added &&
 211             bch2_maybe_compact_whiteouts(c, b))
 212                 bch2_btree_iter_reinit_node(iter, b);
 213
 214         trace_btree_insert_key(c, b, insert);
 215         return true;
 216 }
 217
 218 /* Cached btree updates: */
 219
 220 /* Normal update interface: */
 221
 222 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 223                                              struct btree_insert_entry *i)
 224 {
 225         BUG_ON(!i->is_extent && bpos_cmp(i->k->k.p, i->iter->real_pos));
 226         BUG_ON(i->level         != i->iter->level);
 227         BUG_ON(i->btree_id      != i->iter->btree_id);
 228 }
 229
 230 static noinline int
 231 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 232                                    unsigned long trace_ip)
 233 {
 234         struct bch_fs *c = trans->c;
 235         int ret;
 236
 237         bch2_trans_unlock(trans);
 238
 239         ret = bch2_journal_preres_get(&c->journal,
 240                         &trans->journal_preres, u64s, 0);
 241         if (ret)
 242                 return ret;
 243
 244         if (!bch2_trans_relock(trans)) {
 245                 trace_trans_restart_journal_preres_get(trans->ip, trace_ip);
 246                 return -EINTR;
 247         }
 248
 249         return 0;
 250 }
 251
 252 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 253                                              unsigned flags)
 254 {
 255         struct bch_fs *c = trans->c;
 256         int ret;
 257
 258         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 259                 flags |= JOURNAL_RES_GET_RESERVED;
 260
 261         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 262                                    trans->journal_u64s, flags);
 263
 264         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 265 }
 266
 267 static enum btree_insert_ret
 268 btree_key_can_insert(struct btree_trans *trans,
 269                      struct btree_iter *iter,
 270                      unsigned u64s)
 271 {
 272         struct bch_fs *c = trans->c;
 273         struct btree *b = iter_l(iter)->b;
 274
 275         if (!bch2_btree_node_insert_fits(c, b, u64s))
 276                 return BTREE_INSERT_BTREE_NODE_FULL;
 277
 278         return BTREE_INSERT_OK;
 279 }
 280
 281 static enum btree_insert_ret
 282 btree_key_can_insert_cached(struct btree_trans *trans,
 283                             struct btree_iter *iter,
 284                             unsigned u64s)
 285 {
 286         struct bkey_cached *ck = (void *) iter->l[0].b;
 287         unsigned new_u64s;
 288         struct bkey_i *new_k;
 289
 290         BUG_ON(iter->level);
 291
 292         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 293             bch2_btree_key_cache_must_wait(trans->c) &&
 294             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 295                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 296
 297         /*
 298          * bch2_varint_decode can read past the end of the buffer by at most 7
 299          * bytes (it won't be used):
 300          */
 301         u64s += 1;
 302
 303         if (u64s <= ck->u64s)
 304                 return BTREE_INSERT_OK;
 305
 306         new_u64s        = roundup_pow_of_two(u64s);
 307         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 308         if (!new_k)
 309                 return -ENOMEM;
 310
 311         ck->u64s        = new_u64s;
 312         ck->k           = new_k;
 313         return BTREE_INSERT_OK;
 314 }
 315
 316 static inline void do_btree_insert_one(struct btree_trans *trans,
 317                                        struct btree_insert_entry *i)
 318 {
 319         struct bch_fs *c = trans->c;
 320         struct journal *j = &c->journal;
 321         bool did_work;
 322
 323         EBUG_ON(trans->journal_res.ref !=
 324                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 325
 326         i->k->k.needs_whiteout = false;
 327
 328         did_work = (btree_iter_type(i->iter) != BTREE_ITER_CACHED)
 329                 ? btree_insert_key_leaf(trans, i->iter, i->k)
 330                 : bch2_btree_insert_key_cached(trans, i->iter, i->k);
 331         if (!did_work)
 332                 return;
 333
 334         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 335                 bch2_journal_add_keys(j, &trans->journal_res,
 336                                       i->btree_id,
 337                                       i->level,
 338                                       i->k);
 339
 340                 bch2_journal_set_has_inode(j, &trans->journal_res,
 341                                            i->k->k.p.inode);
 342
 343                 if (trans->journal_seq)
 344                         *trans->journal_seq = trans->journal_res.seq;
 345         }
 346 }
 347
 348 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 349 {
 350         __bch2_btree_iter_unlock(iter);
 351 }
 352
 353 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 354 {
 355         struct bch_fs *c = trans->c;
 356         struct btree_insert_entry *i;
 357
 358         trans_for_each_update(trans, i) {
 359                 /*
 360                  * XXX: synchronization of cached update triggers with gc
 361                  */
 362                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 363
 364                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 365                         bch2_mark_update(trans, i->iter, i->k, NULL,
 366                                          i->trigger_flags|BTREE_TRIGGER_GC);
 367         }
 368 }
 369
 370 static inline int
 371 bch2_trans_commit_write_locked(struct btree_trans *trans,
 372                                struct btree_insert_entry **stopped_at,
 373                                unsigned long trace_ip)
 374 {
 375         struct bch_fs *c = trans->c;
 376         struct btree_insert_entry *i;
 377         struct btree_trans_commit_hook *h;
 378         unsigned u64s = 0;
 379         bool marking = false;
 380         int ret;
 381
 382         if (race_fault()) {
 383                 trace_trans_restart_fault_inject(trans->ip, trace_ip);
 384                 return -EINTR;
 385         }
 386
 387         /*
 388          * Check if the insert will fit in the leaf node with the write lock
 389          * held, otherwise another thread could write the node changing the
 390          * amount of space available:
 391          */
 392
 393         prefetch(&trans->c->journal.flags);
 394
 395         h = trans->hooks;
 396         while (h) {
 397                 ret = h->fn(trans, h);
 398                 if (ret)
 399                         return ret;
 400                 h = h->next;
 401         }
 402
 403         trans_for_each_update2(trans, i) {
 404                 /* Multiple inserts might go to same leaf: */
 405                 if (!same_leaf_as_prev(trans, i))
 406                         u64s = 0;
 407
 408                 u64s += i->k->k.u64s;
 409                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 410                         ? btree_key_can_insert(trans, i->iter, u64s)
 411                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 412                 if (ret) {
 413                         *stopped_at = i;
 414                         return ret;
 415                 }
 416
 417                 if (btree_node_type_needs_gc(i->bkey_type))
 418                         marking = true;
 419         }
 420
 421         if (marking) {
 422                 percpu_down_read(&c->mark_lock);
 423         }
 424
 425         /* Must be called under mark_lock: */
 426         if (marking && trans->fs_usage_deltas &&
 427             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 428                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 429                 goto err;
 430         }
 431
 432         /*
 433          * Don't get journal reservation until after we know insert will
 434          * succeed:
 435          */
 436         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 437                 ret = bch2_trans_journal_res_get(trans,
 438                                 JOURNAL_RES_GET_NONBLOCK);
 439                 if (ret)
 440                         goto err;
 441         } else {
 442                 trans->journal_res.seq = c->journal.replay_journal_seq;
 443         }
 444
 445         if (unlikely(trans->extra_journal_entry_u64s)) {
 446                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 447                                   trans->extra_journal_entries,
 448                                   trans->extra_journal_entry_u64s);
 449
 450                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 451                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 452         }
 453
 454         /*
 455          * Not allowed to fail after we've gotten our journal reservation - we
 456          * have to use it:
 457          */
 458
 459         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 460                 if (bch2_journal_seq_verify)
 461                         trans_for_each_update2(trans, i)
 462                                 i->k->k.version.lo = trans->journal_res.seq;
 463                 else if (bch2_inject_invalid_keys)
 464                         trans_for_each_update2(trans, i)
 465                                 i->k->k.version = MAX_VERSION;
 466         }
 467
 468         trans_for_each_update(trans, i)
 469                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 470                         bch2_mark_update(trans, i->iter, i->k,
 471                                          NULL, i->trigger_flags);
 472
 473         if (marking && trans->fs_usage_deltas)
 474                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 475
 476         if (unlikely(c->gc_pos.phase))
 477                 bch2_trans_mark_gc(trans);
 478
 479         trans_for_each_update2(trans, i)
 480                 do_btree_insert_one(trans, i);
 481 err:
 482         if (marking) {
 483                 percpu_up_read(&c->mark_lock);
 484         }
 485
 486         return ret;
 487 }
 488
 489 static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 490 {
 491         struct btree_insert_entry *i;
 492         struct btree *b = iter_l(iter)->b;
 493         struct bkey_s_c old;
 494         int u64s_delta = 0;
 495         int ret;
 496
 497         /*
 498          * Inserting directly into interior nodes is an uncommon operation with
 499          * various weird edge cases: also, a lot of things about
 500          * BTREE_ITER_NODES iters need to be audited
 501          */
 502         if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
 503                 return 0;
 504
 505         BUG_ON(iter->level);
 506
 507         trans_for_each_update2(trans, i) {
 508                 if (iter_l(i->iter)->b != b)
 509                         continue;
 510
 511                 old = bch2_btree_iter_peek_slot(i->iter);
 512                 ret = bkey_err(old);
 513                 if (ret)
 514                         return ret;
 515
 516                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 517                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 518         }
 519
 520         return u64s_delta <= 0
 521                 ? (bch2_foreground_maybe_merge(trans->c, iter, iter->level,
 522                                 trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
 523                 : 0;
 524 }
 525
 526 /*
 527  * Get journal reservation, take write locks, and attempt to do btree update(s):
 528  */
 529 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 530                                        struct btree_insert_entry **stopped_at,
 531                                        unsigned long trace_ip)
 532 {
 533         struct bch_fs *c = trans->c;
 534         struct btree_insert_entry *i;
 535         struct btree_iter *iter;
 536         int ret;
 537
 538         trans_for_each_update2(trans, i) {
 539                 struct btree *b;
 540
 541                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 542
 543                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
 544                         continue;
 545
 546                 b = iter_l(i->iter)->b;
 547                 if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
 548                     b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
 549                         ret = maybe_do_btree_merge(trans, i->iter);
 550                         if (unlikely(ret))
 551                                 return ret;
 552                 }
 553         }
 554
 555         trans_for_each_update2(trans, i)
 556                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 557
 558         ret = bch2_journal_preres_get(&c->journal,
 559                         &trans->journal_preres, trans->journal_preres_u64s,
 560                         JOURNAL_RES_GET_NONBLOCK|
 561                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 562                          ? JOURNAL_RES_GET_RESERVED : 0));
 563         if (unlikely(ret == -EAGAIN))
 564                 ret = bch2_trans_journal_preres_get_cold(trans,
 565                                                 trans->journal_preres_u64s, trace_ip);
 566         if (unlikely(ret))
 567                 return ret;
 568
 569         /*
 570          * Can't be holding any read locks when we go to take write locks:
 571          * another thread could be holding an intent lock on the same node we
 572          * have a read lock on, and it'll block trying to take a write lock
 573          * (because we hold a read lock) and it could be blocking us by holding
 574          * its own read lock (while we're trying to to take write locks).
 575          *
 576          * note - this must be done after bch2_trans_journal_preres_get_cold()
 577          * or anything else that might call bch2_trans_relock(), since that
 578          * would just retake the read locks:
 579          */
 580         trans_for_each_iter(trans, iter) {
 581                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 582                         if (btree_iter_keep(trans, iter)) {
 583                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 584                                         trace_trans_restart_upgrade(trans->ip, trace_ip,
 585                                                                     iter->btree_id,
 586                                                                     &iter->real_pos);
 587                                         return -EINTR;
 588                                 }
 589                         } else {
 590                                 bch2_btree_iter_unlock_noinline(iter);
 591                         }
 592                 }
 593         }
 594
 595         trans_for_each_update2(trans, i) {
 596                 const char *invalid = bch2_bkey_invalid(c,
 597                                 bkey_i_to_s_c(i->k), i->bkey_type);
 598                 if (invalid) {
 599                         char buf[200];
 600
 601                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 602                         bch_err(c, "invalid bkey %s on insert: %s\n", buf, invalid);
 603                         bch2_fatal_error(c);
 604                 }
 605                 btree_insert_entry_checks(trans, i);
 606         }
 607         bch2_btree_trans_verify_locks(trans);
 608
 609         trans_for_each_update2(trans, i)
 610                 if (!same_leaf_as_prev(trans, i))
 611                         bch2_btree_node_lock_for_insert(c,
 612                                         iter_l(i->iter)->b, i->iter);
 613
 614         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 615
 616         trans_for_each_update2(trans, i)
 617                 if (!same_leaf_as_prev(trans, i))
 618                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 619                                                              i->iter);
 620
 621         if (!ret && trans->journal_pin)
 622                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 623                                      trans->journal_pin, NULL);
 624
 625         /*
 626          * Drop journal reservation after dropping write locks, since dropping
 627          * the journal reservation may kick off a journal write:
 628          */
 629         bch2_journal_res_put(&c->journal, &trans->journal_res);
 630
 631         if (unlikely(ret))
 632                 return ret;
 633
 634         bch2_trans_downgrade(trans);
 635
 636         return 0;
 637 }
 638
 639 static int journal_reclaim_wait_done(struct bch_fs *c)
 640 {
 641         int ret = bch2_journal_error(&c->journal) ?:
 642                 !bch2_btree_key_cache_must_wait(c);
 643
 644         if (!ret)
 645                 journal_reclaim_kick(&c->journal);
 646         return ret;
 647 }
 648
 649 static noinline
 650 int bch2_trans_commit_error(struct btree_trans *trans,
 651                             struct btree_insert_entry *i,
 652                             int ret, unsigned long trace_ip)
 653 {
 654         struct bch_fs *c = trans->c;
 655         unsigned flags = trans->flags;
 656
 657         /*
 658          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 659          * update; if we haven't done anything yet it doesn't apply
 660          */
 661         flags &= ~BTREE_INSERT_NOUNLOCK;
 662
 663         switch (ret) {
 664         case BTREE_INSERT_BTREE_NODE_FULL:
 665                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 666
 667                 /*
 668                  * if the split succeeded without dropping locks the insert will
 669                  * still be atomic (what the caller peeked() and is overwriting
 670                  * won't have changed)
 671                  */
 672 #if 0
 673                 /*
 674                  * XXX:
 675                  * split -> btree node merging (of parent node) might still drop
 676                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 677                  *
 678                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 679                  * will inhibit merging - but we don't have a reliable way yet
 680                  * (do we?) of checking if we dropped locks in this path
 681                  */
 682                 if (!ret)
 683                         goto retry;
 684 #endif
 685
 686                 /*
 687                  * don't care if we got ENOSPC because we told split it
 688                  * couldn't block:
 689                  */
 690                 if (!ret ||
 691                     ret == -EINTR ||
 692                     (flags & BTREE_INSERT_NOUNLOCK)) {
 693                         trace_trans_restart_btree_node_split(trans->ip, trace_ip,
 694                                                              i->iter->btree_id,
 695                                                              &i->iter->real_pos);
 696                         ret = -EINTR;
 697                 }
 698                 break;
 699         case BTREE_INSERT_ENOSPC:
 700                 BUG_ON(flags & BTREE_INSERT_NOFAIL);
 701                 ret = -ENOSPC;
 702                 break;
 703         case BTREE_INSERT_NEED_MARK_REPLICAS:
 704                 bch2_trans_unlock(trans);
 705
 706                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 707                 if (ret)
 708                         return ret;
 709
 710                 if (bch2_trans_relock(trans))
 711                         return 0;
 712
 713                 trace_trans_restart_mark_replicas(trans->ip, trace_ip);
 714                 ret = -EINTR;
 715                 break;
 716         case BTREE_INSERT_NEED_JOURNAL_RES:
 717                 bch2_trans_unlock(trans);
 718
 719                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 720                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
 721                         return -EAGAIN;
 722
 723                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 724                 if (ret)
 725                         return ret;
 726
 727                 if (bch2_trans_relock(trans))
 728                         return 0;
 729
 730                 trace_trans_restart_journal_res_get(trans->ip, trace_ip);
 731                 ret = -EINTR;
 732                 break;
 733         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 734                 bch2_trans_unlock(trans);
 735
 736                 trace_trans_blocked_journal_reclaim(trans->ip, trace_ip);
 737
 738                 wait_event_freezable(c->journal.reclaim_wait,
 739                                      (ret = journal_reclaim_wait_done(c)));
 740                 if (ret < 0)
 741                         return ret;
 742
 743                 if (bch2_trans_relock(trans))
 744                         return 0;
 745
 746                 trace_trans_restart_journal_reclaim(trans->ip, trace_ip);
 747                 ret = -EINTR;
 748                 break;
 749         default:
 750                 BUG_ON(ret >= 0);
 751                 break;
 752         }
 753
 754         BUG_ON(ret == -ENOSPC && (flags & BTREE_INSERT_NOFAIL));
 755
 756         return ret;
 757 }
 758
 759 static noinline int
 760 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 761 {
 762         struct bch_fs *c = trans->c;
 763         int ret;
 764
 765         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 766                 return -EROFS;
 767
 768         bch2_trans_unlock(trans);
 769
 770         ret = bch2_fs_read_write_early(c);
 771         if (ret)
 772                 return ret;
 773
 774         percpu_ref_get(&c->writes);
 775         return 0;
 776 }
 777
 778 static void __bch2_trans_update2(struct btree_trans *trans,
 779                                  struct btree_insert_entry n)
 780 {
 781         struct btree_insert_entry *i;
 782
 783         btree_insert_entry_checks(trans, &n);
 784
 785         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 786
 787         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 788
 789         trans_for_each_update2(trans, i)
 790                 if (btree_insert_entry_cmp(&n, i) <= 0)
 791                         break;
 792
 793         if (i < trans->updates2 + trans->nr_updates2 &&
 794             !btree_insert_entry_cmp(&n, i))
 795                 *i = n;
 796         else
 797                 array_insert_item(trans->updates2, trans->nr_updates2,
 798                                   i - trans->updates2, n);
 799 }
 800
 801 static void bch2_trans_update2(struct btree_trans *trans,
 802                                struct btree_iter *iter,
 803                                struct bkey_i *insert)
 804 {
 805         __bch2_trans_update2(trans, (struct btree_insert_entry) {
 806                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 807                 .btree_id       = iter->btree_id,
 808                 .level          = iter->level,
 809                 .iter           = iter,
 810                 .k              = insert,
 811         });
 812 }
 813
 814 static int extent_update_to_keys(struct btree_trans *trans,
 815                                  struct btree_insert_entry n)
 816 {
 817         int ret;
 818
 819         ret = bch2_extent_can_insert(trans, n.iter, n.k);
 820         if (ret)
 821                 return ret;
 822
 823         if (bkey_deleted(&n.k->k))
 824                 return 0;
 825
 826         n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p,
 827                                      BTREE_ITER_INTENT|
 828                                      BTREE_ITER_NOT_EXTENTS);
 829         n.is_extent = false;
 830
 831         __bch2_trans_update2(trans, n);
 832         bch2_trans_iter_put(trans, n.iter);
 833         return 0;
 834 }
 835
 836 static int extent_handle_overwrites(struct btree_trans *trans,
 837                                     enum btree_id btree_id,
 838                                     struct bkey_i *insert)
 839 {
 840         struct btree_iter *iter, *update_iter;
 841         struct bpos start = bkey_start_pos(&insert->k);
 842         struct bkey_i *update;
 843         struct bkey_s_c k;
 844         int ret = 0;
 845
 846         iter = bch2_trans_get_iter(trans, btree_id, start,
 847                                    BTREE_ITER_INTENT);
 848         k = bch2_btree_iter_peek_with_updates(iter);
 849
 850         while (k.k && !(ret = bkey_err(k))) {
 851                 if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
 852                         break;
 853
 854                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 855                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 856                         if ((ret = PTR_ERR_OR_ZERO(update)))
 857                                 break;
 858
 859                         bkey_reassemble(update, k);
 860
 861                         bch2_cut_back(start, update);
 862
 863                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 864                                                           BTREE_ITER_NOT_EXTENTS|
 865                                                           BTREE_ITER_INTENT);
 866                         bch2_trans_update2(trans, update_iter, update);
 867                         bch2_trans_iter_put(trans, update_iter);
 868                 }
 869
 870                 if (bkey_cmp(k.k->p, insert->k.p) < 0 ||
 871                     (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) {
 872                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 873                         if ((ret = PTR_ERR_OR_ZERO(update)))
 874                                 break;
 875
 876                         bkey_init(&update->k);
 877                         update->k.p = k.k->p;
 878
 879                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 880                                                           BTREE_ITER_NOT_EXTENTS|
 881                                                           BTREE_ITER_INTENT);
 882                         bch2_trans_update2(trans, update_iter, update);
 883                         bch2_trans_iter_put(trans, update_iter);
 884                 }
 885
 886                 if (bkey_cmp(k.k->p, insert->k.p) > 0) {
 887                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 888                         if ((ret = PTR_ERR_OR_ZERO(update)))
 889                                 break;
 890
 891                         bkey_reassemble(update, k);
 892                         bch2_cut_front(insert->k.p, update);
 893
 894                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 895                                                           BTREE_ITER_NOT_EXTENTS|
 896                                                           BTREE_ITER_INTENT);
 897                         bch2_trans_update2(trans, update_iter, update);
 898                         bch2_trans_iter_put(trans, update_iter);
 899                         break;
 900                 }
 901
 902                 k = bch2_btree_iter_next_with_updates(iter);
 903         }
 904         bch2_trans_iter_put(trans, iter);
 905
 906         return ret;
 907 }
 908
 909 int __bch2_trans_commit(struct btree_trans *trans)
 910 {
 911         struct btree_insert_entry *i = NULL;
 912         struct btree_iter *iter;
 913         bool trans_trigger_run;
 914         unsigned u64s, reset_flags = 0;
 915         int ret = 0;
 916
 917         if (!trans->nr_updates)
 918                 goto out_reset;
 919
 920         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 921                 lockdep_assert_held(&trans->c->gc_lock);
 922
 923         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 924
 925         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 926         trans->journal_preres_u64s      = 0;
 927
 928         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 929             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 930                 ret = bch2_trans_commit_get_rw_cold(trans);
 931                 if (ret)
 932                         goto out_reset;
 933         }
 934
 935 #ifdef CONFIG_BCACHEFS_DEBUG
 936         trans_for_each_update(trans, i)
 937                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 938                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 939                         bch2_btree_key_cache_verify_clean(trans,
 940                                         i->btree_id, i->k->k.p);
 941 #endif
 942
 943         /*
 944          * Running triggers will append more updates to the list of updates as
 945          * we're walking it:
 946          */
 947         do {
 948                 trans_trigger_run = false;
 949
 950                 trans_for_each_update(trans, i) {
 951                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 952                             !i->trans_triggers_run) {
 953                                 i->trans_triggers_run = true;
 954                                 trans_trigger_run = true;
 955
 956                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 957                                                              i->trigger_flags);
 958                                 if (unlikely(ret)) {
 959                                         if (ret == -EINTR)
 960                                                 trace_trans_restart_mark(trans->ip, _RET_IP_,
 961                                                                          i->iter->btree_id,
 962                                                                          &i->iter->pos);
 963                                         goto out;
 964                                 }
 965                         }
 966                 }
 967         } while (trans_trigger_run);
 968
 969         /* Turn extents updates into keys: */
 970         trans_for_each_update(trans, i)
 971                 if (i->is_extent) {
 972                         ret = extent_handle_overwrites(trans, i->btree_id, i->k);
 973                         if (unlikely(ret))
 974                                 goto out;
 975                 }
 976
 977         trans_for_each_update(trans, i) {
 978                 ret = i->is_extent
 979                         ? extent_update_to_keys(trans, *i)
 980                         : (__bch2_trans_update2(trans, *i), 0);
 981                 if (unlikely(ret))
 982                         goto out;
 983         }
 984
 985         trans_for_each_update2(trans, i) {
 986                 ret = bch2_btree_iter_traverse(i->iter);
 987                 if (unlikely(ret)) {
 988                         trace_trans_restart_traverse(trans->ip, _RET_IP_,
 989                                                      i->iter->btree_id,
 990                                                      &i->iter->pos);
 991                         goto out;
 992                 }
 993
 994                 if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
 995                         trace_trans_restart_upgrade(trans->ip, _RET_IP_,
 996                                                     i->iter->btree_id,
 997                                                     &i->iter->pos);
 998                         ret = -EINTR;
 999                         goto out;
1000                 }
1001
1002                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
1003
1004                 u64s = jset_u64s(i->k->k.u64s);
1005                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
1006                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1007                         trans->journal_preres_u64s += u64s;
1008                 trans->journal_u64s += u64s;
1009         }
1010 retry:
1011         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1012
1013         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1014
1015         /* make sure we didn't drop or screw up locks: */
1016         bch2_btree_trans_verify_locks(trans);
1017
1018         if (ret)
1019                 goto err;
1020
1021         trans_for_each_iter(trans, iter)
1022                 if (btree_iter_live(trans, iter) &&
1023                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
1024                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
1025 out:
1026         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
1027
1028         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1029                 percpu_ref_put(&trans->c->writes);
1030 out_reset:
1031         if (!ret)
1032                 reset_flags |= TRANS_RESET_NOTRAVERSE;
1033         if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
1034                 reset_flags |= TRANS_RESET_NOUNLOCK;
1035         bch2_trans_reset(trans, reset_flags);
1036
1037         return ret;
1038 err:
1039         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1040         if (ret)
1041                 goto out;
1042
1043         goto retry;
1044 }
1045
1046 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1047                       struct bkey_i *k, enum btree_trigger_flags flags)
1048 {
1049         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
1050                 .trigger_flags  = flags,
1051                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
1052                 .btree_id       = iter->btree_id,
1053                 .level          = iter->level,
1054                 .is_extent      = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0,
1055                 .iter           = iter,
1056                 .k              = k
1057         };
1058
1059         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1060
1061 #ifdef CONFIG_BCACHEFS_DEBUG
1062         BUG_ON(bkey_cmp(iter->pos,
1063                         n.is_extent ? bkey_start_pos(&k->k) : k->k.p));
1064
1065         trans_for_each_update(trans, i) {
1066                 BUG_ON(bkey_cmp(i->iter->pos,
1067                                 i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p));
1068
1069                 BUG_ON(i != trans->updates &&
1070                        btree_insert_entry_cmp(i - 1, i) >= 0);
1071         }
1072 #endif
1073
1074         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1075
1076         if (n.is_extent) {
1077                 iter->pos_after_commit = k->k.p;
1078                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1079         }
1080
1081         /*
1082          * Pending updates are kept sorted: first, find position of new update,
1083          * then delete/trim any updates the new update overwrites:
1084          */
1085         if (!n.is_extent) {
1086                 trans_for_each_update(trans, i)
1087                         if (btree_insert_entry_cmp(&n, i) <= 0)
1088                                 break;
1089
1090                 if (i < trans->updates + trans->nr_updates &&
1091                     !btree_insert_entry_cmp(&n, i))
1092                         *i = n;
1093                 else
1094                         array_insert_item(trans->updates, trans->nr_updates,
1095                                           i - trans->updates, n);
1096         } else {
1097                 trans_for_each_update(trans, i)
1098                         if (btree_insert_entry_cmp(&n, i) < 0)
1099                                 break;
1100
1101                 while (i > trans->updates &&
1102                        i[-1].btree_id == n.btree_id &&
1103                        bkey_cmp(bkey_start_pos(&n.k->k),
1104                                 bkey_start_pos(&i[-1].k->k)) <= 0) {
1105                         --i;
1106                         array_remove_item(trans->updates, trans->nr_updates,
1107                                           i - trans->updates);
1108                 }
1109
1110                 if (i > trans->updates &&
1111                     i[-1].btree_id == n.btree_id &&
1112                     bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0)
1113                         bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k);
1114
1115                 if (i < trans->updates + trans->nr_updates &&
1116                     i->btree_id == n.btree_id &&
1117                     bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) {
1118                         if (bkey_cmp(bkey_start_pos(&n.k->k),
1119                                      bkey_start_pos(&i->k->k)) > 0) {
1120                                 struct btree_insert_entry split = *i;
1121                                 int ret;
1122
1123                                 BUG_ON(trans->nr_updates + 1 >= BTREE_ITER_MAX);
1124
1125                                 split.k = bch2_trans_kmalloc(trans, bkey_bytes(&i->k->k));
1126                                 ret = PTR_ERR_OR_ZERO(split.k);
1127                                 if (ret)
1128                                         return ret;
1129
1130                                 bkey_copy(split.k, i->k);
1131                                 bch2_cut_back(bkey_start_pos(&n.k->k), split.k);
1132
1133                                 split.iter = bch2_trans_get_iter(trans, split.btree_id,
1134                                                                  bkey_start_pos(&split.k->k),
1135                                                                  BTREE_ITER_INTENT);
1136                                 split.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1137                                 bch2_trans_iter_put(trans, split.iter);
1138                                 array_insert_item(trans->updates, trans->nr_updates,
1139                                                   i - trans->updates, split);
1140                                 i++;
1141                         }
1142
1143                         /*
1144                          * When we have an extent that overwrites the start of another
1145                          * update, trimming that extent will mean the iterator's
1146                          * position has to change since the iterator position has to
1147                          * match the extent's start pos - but we don't want to change
1148                          * the iterator pos if some other code is using it, so we may
1149                          * need to clone it:
1150                          */
1151                         if (btree_iter_live(trans, i->iter)) {
1152                                 i->iter = bch2_trans_copy_iter(trans, i->iter);
1153
1154                                 i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1155                                 bch2_trans_iter_put(trans, i->iter);
1156                         }
1157
1158                         bch2_cut_front(n.k->k.p, i->k);
1159                         bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1160                 }
1161
1162                 array_insert_item(trans->updates, trans->nr_updates,
1163                                   i - trans->updates, n);
1164         }
1165
1166         return 0;
1167 }
1168
1169 void bch2_trans_commit_hook(struct btree_trans *trans,
1170                             struct btree_trans_commit_hook *h)
1171 {
1172         h->next = trans->hooks;
1173         trans->hooks = h;
1174 }
1175
1176 int __bch2_btree_insert(struct btree_trans *trans,
1177                         enum btree_id id, struct bkey_i *k)
1178 {
1179         struct btree_iter *iter;
1180         int ret;
1181
1182         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1183                                    BTREE_ITER_INTENT);
1184
1185         ret = bch2_trans_update(trans, iter, k, 0);
1186         bch2_trans_iter_put(trans, iter);
1187         return ret;
1188 }
1189
1190 /**
1191  * bch2_btree_insert - insert keys into the extent btree
1192  * @c:                  pointer to struct bch_fs
1193  * @id:                 btree to insert into
1194  * @insert_keys:        list of keys to insert
1195  * @hook:               insert callback
1196  */
1197 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1198                       struct bkey_i *k,
1199                       struct disk_reservation *disk_res,
1200                       u64 *journal_seq, int flags)
1201 {
1202         return bch2_trans_do(c, disk_res, journal_seq, flags,
1203                              __bch2_btree_insert(&trans, id, k));
1204 }
1205
1206 int bch2_btree_delete_at(struct btree_trans *trans,
1207                          struct btree_iter *iter, unsigned flags)
1208 {
1209         struct bkey_i k;
1210
1211         bkey_init(&k.k);
1212         k.k.p = iter->pos;
1213
1214         return  bch2_trans_update(trans, iter, &k, 0) ?:
1215                 bch2_trans_commit(trans, NULL, NULL,
1216                                   BTREE_INSERT_NOFAIL|flags);
1217 }
1218
1219 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1220                                   struct bpos start, struct bpos end,
1221                                   u64 *journal_seq)
1222 {
1223         struct btree_iter *iter;
1224         struct bkey_s_c k;
1225         int ret = 0;
1226
1227         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1228 retry:
1229         while ((k = bch2_btree_iter_peek(iter)).k &&
1230                !(ret = bkey_err(k)) &&
1231                bkey_cmp(iter->pos, end) < 0) {
1232                 struct bkey_i delete;
1233
1234                 bch2_trans_begin(trans);
1235
1236                 bkey_init(&delete.k);
1237
1238                 /*
1239                  * This could probably be more efficient for extents:
1240                  */
1241
1242                 /*
1243                  * For extents, iter.pos won't necessarily be the same as
1244                  * bkey_start_pos(k.k) (for non extents they always will be the
1245                  * same). It's important that we delete starting from iter.pos
1246                  * because the range we want to delete could start in the middle
1247                  * of k.
1248                  *
1249                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1250                  * bkey_start_pos(k.k)).
1251                  */
1252                 delete.k.p = iter->pos;
1253
1254                 if (btree_node_type_is_extents(iter->btree_id)) {
1255                         unsigned max_sectors =
1256                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1257
1258                         /* create the biggest key we can */
1259                         bch2_key_resize(&delete.k, max_sectors);
1260                         bch2_cut_back(end, &delete);
1261
1262                         ret = bch2_extent_trim_atomic(&delete, iter);
1263                         if (ret)
1264                                 break;
1265                 }
1266
1267                 ret   = bch2_trans_update(trans, iter, &delete, 0) ?:
1268                         bch2_trans_commit(trans, NULL, journal_seq,
1269                                         BTREE_INSERT_NOFAIL);
1270                 if (ret)
1271                         break;
1272
1273                 bch2_trans_cond_resched(trans);
1274         }
1275
1276         if (ret == -EINTR) {
1277                 ret = 0;
1278                 goto retry;
1279         }
1280
1281         bch2_trans_iter_free(trans, iter);
1282         return ret;
1283 }
1284
1285 /*
1286  * bch_btree_delete_range - delete everything within a given range
1287  *
1288  * Range is a half open interval - [start, end)
1289  */
1290 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1291                             struct bpos start, struct bpos end,
1292                             u64 *journal_seq)
1293 {
1294         return bch2_trans_do(c, NULL, journal_seq, 0,
1295                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1296 }