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 "btree_write_buffer.h"
  12 #include "buckets.h"
  13 #include "debug.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "extent_update.h"
  17 #include "journal.h"
  18 #include "journal_reclaim.h"
  19 #include "keylist.h"
  20 #include "recovery.h"
  21 #include "subvolume.h"
  22 #include "replicas.h"
  23 #include "trace.h"
  24
  25 #include <linux/prefetch.h>
  26 #include <linux/sort.h>
  27
  28 /*
  29  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
  30  * different snapshot:
  31  */
  32 struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
  33 {
  34         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
  35
  36         if (k.k && bpos_eq(path->pos, k.k->p))
  37                 return k;
  38
  39         bkey_init(u);
  40         u->p = path->pos;
  41         return (struct bkey_s_c) { u, NULL };
  42 }
  43
  44 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  45 {
  46 #ifdef CONFIG_BCACHEFS_DEBUG
  47         struct bch_fs *c = trans->c;
  48         struct bkey u;
  49         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
  50
  51         if (unlikely(trans->journal_replay_not_finished)) {
  52                 struct bkey_i *j_k =
  53                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  54
  55                 if (j_k)
  56                         k = bkey_i_to_s_c(j_k);
  57         }
  58
  59         u = *k.k;
  60         u.needs_whiteout = i->old_k.needs_whiteout;
  61
  62         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  63         BUG_ON(i->old_v != k.v);
  64 #endif
  65 }
  66
  67 static int __must_check
  68 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  69                           struct bkey_i *, enum btree_update_flags);
  70
  71 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  72                                          const struct btree_insert_entry *r)
  73 {
  74         return   cmp_int(l->btree_id,   r->btree_id) ?:
  75                  cmp_int(l->cached,     r->cached) ?:
  76                  -cmp_int(l->level,     r->level) ?:
  77                  bpos_cmp(l->k->k.p,    r->k->k.p);
  78 }
  79
  80 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  81 {
  82         return i->path->l + i->level;
  83 }
  84
  85 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  86                                      struct btree_insert_entry *i)
  87 {
  88         return i != trans->updates &&
  89                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  90 }
  91
  92 static inline bool same_leaf_as_next(struct btree_trans *trans,
  93                                      struct btree_insert_entry *i)
  94 {
  95         return i + 1 < trans->updates + trans->nr_updates &&
  96                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  97 }
  98
  99 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
 100                                            struct btree_path *path,
 101                                            struct btree *b)
 102 {
 103         struct bch_fs *c = trans->c;
 104
 105         if (unlikely(btree_node_just_written(b)) &&
 106             bch2_btree_post_write_cleanup(c, b))
 107                 bch2_trans_node_reinit_iter(trans, b);
 108
 109         /*
 110          * If the last bset has been written, or if it's gotten too big - start
 111          * a new bset to insert into:
 112          */
 113         if (want_new_bset(c, b))
 114                 bch2_btree_init_next(trans, b);
 115 }
 116
 117 /* Inserting into a given leaf node (last stage of insert): */
 118
 119 /* Handle overwrites and do insert, for non extents: */
 120 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 121                                 struct btree_path *path,
 122                                 struct btree *b,
 123                                 struct btree_node_iter *node_iter,
 124                                 struct bkey_i *insert)
 125 {
 126         struct bkey_packed *k;
 127         unsigned clobber_u64s = 0, new_u64s = 0;
 128
 129         EBUG_ON(btree_node_just_written(b));
 130         EBUG_ON(bset_written(b, btree_bset_last(b)));
 131         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 132         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 133         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 134         EBUG_ON(insert->k.u64s >
 135                 bch_btree_keys_u64s_remaining(trans->c, b));
 136
 137         k = bch2_btree_node_iter_peek_all(node_iter, b);
 138         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 139                 k = NULL;
 140
 141         /* @k is the key being overwritten/deleted, if any: */
 142         EBUG_ON(k && bkey_deleted(k));
 143
 144         /* Deleting, but not found? nothing to do: */
 145         if (bkey_deleted(&insert->k) && !k)
 146                 return false;
 147
 148         if (bkey_deleted(&insert->k)) {
 149                 /* Deleting: */
 150                 btree_account_key_drop(b, k);
 151                 k->type = KEY_TYPE_deleted;
 152
 153                 if (k->needs_whiteout)
 154                         push_whiteout(trans->c, b, insert->k.p);
 155                 k->needs_whiteout = false;
 156
 157                 if (k >= btree_bset_last(b)->start) {
 158                         clobber_u64s = k->u64s;
 159                         bch2_bset_delete(b, k, clobber_u64s);
 160                         goto fix_iter;
 161                 } else {
 162                         bch2_btree_path_fix_key_modified(trans, b, k);
 163                 }
 164
 165                 return true;
 166         }
 167
 168         if (k) {
 169                 /* Overwriting: */
 170                 btree_account_key_drop(b, k);
 171                 k->type = KEY_TYPE_deleted;
 172
 173                 insert->k.needs_whiteout = k->needs_whiteout;
 174                 k->needs_whiteout = false;
 175
 176                 if (k >= btree_bset_last(b)->start) {
 177                         clobber_u64s = k->u64s;
 178                         goto overwrite;
 179                 } else {
 180                         bch2_btree_path_fix_key_modified(trans, b, k);
 181                 }
 182         }
 183
 184         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 185 overwrite:
 186         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 187         new_u64s = k->u64s;
 188 fix_iter:
 189         if (clobber_u64s != new_u64s)
 190                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 191                                          clobber_u64s, new_u64s);
 192         return true;
 193 }
 194
 195 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 196                                unsigned i, u64 seq)
 197 {
 198         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 199         struct btree_write *w = container_of(pin, struct btree_write, journal);
 200         struct btree *b = container_of(w, struct btree, writes[i]);
 201         struct btree_trans trans;
 202         unsigned long old, new, v;
 203         unsigned idx = w - b->writes;
 204
 205         bch2_trans_init(&trans, c, 0, 0);
 206
 207         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 208         v = READ_ONCE(b->flags);
 209
 210         do {
 211                 old = new = v;
 212
 213                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 214                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 215                     w->journal.seq != seq)
 216                         break;
 217
 218                 new &= ~BTREE_WRITE_TYPE_MASK;
 219                 new |= BTREE_WRITE_journal_reclaim;
 220                 new |= 1 << BTREE_NODE_need_write;
 221         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 222
 223         btree_node_write_if_need(c, b, SIX_LOCK_read);
 224         six_unlock_read(&b->c.lock);
 225
 226         bch2_trans_exit(&trans);
 227         return 0;
 228 }
 229
 230 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 231 {
 232         return __btree_node_flush(j, pin, 0, seq);
 233 }
 234
 235 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 236 {
 237         return __btree_node_flush(j, pin, 1, seq);
 238 }
 239
 240 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 241                                        struct btree *b, u64 seq)
 242 {
 243         struct btree_write *w = btree_current_write(b);
 244
 245         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 246                              btree_node_write_idx(b) == 0
 247                              ? bch2_btree_node_flush0
 248                              : bch2_btree_node_flush1);
 249 }
 250
 251 /**
 252  * btree_insert_key - insert a key one key into a leaf node
 253  */
 254 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 255                                        struct btree_path *path,
 256                                        struct bkey_i *insert,
 257                                        u64 journal_seq)
 258 {
 259         struct bch_fs *c = trans->c;
 260         struct btree *b = path_l(path)->b;
 261         struct bset_tree *t = bset_tree_last(b);
 262         struct bset *i = bset(b, t);
 263         int old_u64s = bset_u64s(t);
 264         int old_live_u64s = b->nr.live_u64s;
 265         int live_u64s_added, u64s_added;
 266
 267         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 268                                         &path_l(path)->iter, insert)))
 269                 return;
 270
 271         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 272
 273         bch2_btree_add_journal_pin(c, b, journal_seq);
 274
 275         if (unlikely(!btree_node_dirty(b)))
 276                 set_btree_node_dirty_acct(c, b);
 277
 278         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 279         u64s_added = (int) bset_u64s(t) - old_u64s;
 280
 281         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 282                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 283         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 284                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 285
 286         if (u64s_added > live_u64s_added &&
 287             bch2_maybe_compact_whiteouts(c, b))
 288                 bch2_trans_node_reinit_iter(trans, b);
 289 }
 290
 291 static void btree_insert_key_leaf(struct btree_trans *trans,
 292                                   struct btree_insert_entry *insert)
 293 {
 294         bch2_btree_insert_key_leaf(trans, insert->path, insert->k, trans->journal_res.seq);
 295 }
 296
 297 /* Cached btree updates: */
 298
 299 /* Normal update interface: */
 300
 301 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 302                                              struct btree_insert_entry *i)
 303 {
 304         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 305         BUG_ON(i->cached        != i->path->cached);
 306         BUG_ON(i->level         != i->path->level);
 307         BUG_ON(i->btree_id      != i->path->btree_id);
 308         EBUG_ON(!i->level &&
 309                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 310                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 311                 i->k->k.p.snapshot &&
 312                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 313 }
 314
 315 static noinline int
 316 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
 317                                    unsigned long trace_ip)
 318 {
 319         struct bch_fs *c = trans->c;
 320         int ret;
 321
 322         bch2_trans_unlock(trans);
 323
 324         ret = bch2_journal_preres_get(&c->journal,
 325                         &trans->journal_preres,
 326                         trans->journal_preres_u64s,
 327                         (flags & JOURNAL_WATERMARK_MASK));
 328         if (ret)
 329                 return ret;
 330
 331         ret = bch2_trans_relock(trans);
 332         if (ret) {
 333                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 334                 return ret;
 335         }
 336
 337         return 0;
 338 }
 339
 340 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 341                                                       unsigned flags)
 342 {
 343         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 344                                     trans->journal_u64s, flags);
 345 }
 346
 347 #define JSET_ENTRY_LOG_U64s             4
 348
 349 static noinline void journal_transaction_name(struct btree_trans *trans)
 350 {
 351         struct bch_fs *c = trans->c;
 352         struct journal *j = &c->journal;
 353         struct jset_entry *entry =
 354                 bch2_journal_add_entry(j, &trans->journal_res,
 355                                        BCH_JSET_ENTRY_log, 0, 0,
 356                                        JSET_ENTRY_LOG_U64s);
 357         struct jset_entry_log *l =
 358                 container_of(entry, struct jset_entry_log, entry);
 359
 360         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 361 }
 362
 363 static inline int btree_key_can_insert(struct btree_trans *trans,
 364                                        struct btree *b, unsigned u64s)
 365 {
 366         struct bch_fs *c = trans->c;
 367
 368         if (!bch2_btree_node_insert_fits(c, b, u64s))
 369                 return -BCH_ERR_btree_insert_btree_node_full;
 370
 371         return 0;
 372 }
 373
 374 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 375                                        struct btree_path *path, unsigned u64s)
 376 {
 377         struct bch_fs *c = trans->c;
 378         struct bkey_cached *ck = (void *) path->l[0].b;
 379         struct btree_insert_entry *i;
 380         unsigned new_u64s;
 381         struct bkey_i *new_k;
 382
 383         EBUG_ON(path->level);
 384
 385         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 386             bch2_btree_key_cache_must_wait(c) &&
 387             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 388                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 389
 390         /*
 391          * bch2_varint_decode can read past the end of the buffer by at most 7
 392          * bytes (it won't be used):
 393          */
 394         u64s += 1;
 395
 396         if (u64s <= ck->u64s)
 397                 return 0;
 398
 399         new_u64s        = roundup_pow_of_two(u64s);
 400         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 401         if (!new_k) {
 402                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 403                         bch2_btree_ids[path->btree_id], new_u64s);
 404                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 405         }
 406
 407         trans_for_each_update(trans, i)
 408                 if (i->old_v == &ck->k->v)
 409                         i->old_v = &new_k->v;
 410
 411         ck->u64s        = new_u64s;
 412         ck->k           = new_k;
 413         return 0;
 414 }
 415
 416 /* Triggers: */
 417
 418 static int run_one_mem_trigger(struct btree_trans *trans,
 419                                struct btree_insert_entry *i,
 420                                unsigned flags)
 421 {
 422         struct bkey_s_c old = { &i->old_k, i->old_v };
 423         struct bkey_i *new = i->k;
 424         int ret;
 425
 426         verify_update_old_key(trans, i);
 427
 428         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 429                 return 0;
 430
 431         if (!btree_node_type_needs_gc(i->btree_id))
 432                 return 0;
 433
 434         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 435             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 436             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 437                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 438                                 old, bkey_i_to_s_c(new),
 439                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 440         } else {
 441                 struct bkey             _deleted = KEY(0, 0, 0);
 442                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 443
 444                 _deleted.p = i->path->pos;
 445
 446                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 447                                 deleted, bkey_i_to_s_c(new),
 448                                 BTREE_TRIGGER_INSERT|flags) ?:
 449                         bch2_mark_key(trans, i->btree_id, i->level,
 450                                 old, deleted,
 451                                 BTREE_TRIGGER_OVERWRITE|flags);
 452         }
 453
 454         return ret;
 455 }
 456
 457 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 458                                  bool overwrite)
 459 {
 460         /*
 461          * Transactional triggers create new btree_insert_entries, so we can't
 462          * pass them a pointer to a btree_insert_entry, that memory is going to
 463          * move:
 464          */
 465         struct bkey old_k = i->old_k;
 466         struct bkey_s_c old = { &old_k, i->old_v };
 467
 468         verify_update_old_key(trans, i);
 469
 470         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 471             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 472                 return 0;
 473
 474         if (!i->insert_trigger_run &&
 475             !i->overwrite_trigger_run &&
 476             bch2_bkey_ops[old.k->type].trans_trigger ==
 477             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 478             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 479                 i->overwrite_trigger_run = true;
 480                 i->insert_trigger_run = true;
 481                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 482                                            BTREE_TRIGGER_INSERT|
 483                                            BTREE_TRIGGER_OVERWRITE|
 484                                            i->flags) ?: 1;
 485         } else if (overwrite && !i->overwrite_trigger_run) {
 486                 i->overwrite_trigger_run = true;
 487                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 488         } else if (!overwrite && !i->insert_trigger_run) {
 489                 i->insert_trigger_run = true;
 490                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 491         } else {
 492                 return 0;
 493         }
 494 }
 495
 496 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 497                               struct btree_insert_entry *btree_id_start)
 498 {
 499         struct btree_insert_entry *i;
 500         bool trans_trigger_run;
 501         int ret, overwrite;
 502
 503         for (overwrite = 1; overwrite >= 0; --overwrite) {
 504
 505                 /*
 506                  * Running triggers will append more updates to the list of updates as
 507                  * we're walking it:
 508                  */
 509                 do {
 510                         trans_trigger_run = false;
 511
 512                         for (i = btree_id_start;
 513                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 514                              i++) {
 515                                 if (i->btree_id != btree_id)
 516                                         continue;
 517
 518                                 ret = run_one_trans_trigger(trans, i, overwrite);
 519                                 if (ret < 0)
 520                                         return ret;
 521                                 if (ret)
 522                                         trans_trigger_run = true;
 523                         }
 524                 } while (trans_trigger_run);
 525         }
 526
 527         return 0;
 528 }
 529
 530 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 531 {
 532         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 533         unsigned btree_id = 0;
 534         int ret = 0;
 535
 536         /*
 537          *
 538          * For a given btree, this algorithm runs insert triggers before
 539          * overwrite triggers: this is so that when extents are being moved
 540          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 541          * they are re-added.
 542          */
 543         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 544                 if (btree_id == BTREE_ID_alloc)
 545                         continue;
 546
 547                 while (btree_id_start < trans->updates + trans->nr_updates &&
 548                        btree_id_start->btree_id < btree_id)
 549                         btree_id_start++;
 550
 551                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 552                 if (ret)
 553                         return ret;
 554         }
 555
 556         trans_for_each_update(trans, i) {
 557                 if (i->btree_id > BTREE_ID_alloc)
 558                         break;
 559                 if (i->btree_id == BTREE_ID_alloc) {
 560                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 561                         if (ret)
 562                                 return ret;
 563                         break;
 564                 }
 565         }
 566
 567 #ifdef CONFIG_BCACHEFS_DEBUG
 568         trans_for_each_update(trans, i)
 569                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 570                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 571                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 572 #endif
 573         return 0;
 574 }
 575
 576 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 577 {
 578         struct bch_fs *c = trans->c;
 579         struct btree_insert_entry *i;
 580         int ret = 0;
 581
 582         trans_for_each_update(trans, i) {
 583                 /*
 584                  * XXX: synchronization of cached update triggers with gc
 585                  * XXX: synchronization of interior node updates with gc
 586                  */
 587                 BUG_ON(i->cached || i->level);
 588
 589                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 590                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 591                         if (ret)
 592                                 break;
 593                 }
 594         }
 595
 596         return ret;
 597 }
 598
 599 static inline int
 600 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 601                                struct btree_insert_entry **stopped_at,
 602                                unsigned long trace_ip)
 603 {
 604         struct bch_fs *c = trans->c;
 605         struct btree_insert_entry *i;
 606         struct btree_write_buffered_key *wb;
 607         struct btree_trans_commit_hook *h;
 608         unsigned u64s = 0;
 609         bool marking = false;
 610         int ret;
 611
 612         if (race_fault()) {
 613                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 614                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 615         }
 616
 617         /*
 618          * Check if the insert will fit in the leaf node with the write lock
 619          * held, otherwise another thread could write the node changing the
 620          * amount of space available:
 621          */
 622
 623         prefetch(&trans->c->journal.flags);
 624
 625         trans_for_each_update(trans, i) {
 626                 /* Multiple inserts might go to same leaf: */
 627                 if (!same_leaf_as_prev(trans, i))
 628                         u64s = 0;
 629
 630                 u64s += i->k->k.u64s;
 631                 ret = !i->cached
 632                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 633                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 634                 if (ret) {
 635                         *stopped_at = i;
 636                         return ret;
 637                 }
 638
 639                 if (btree_node_type_needs_gc(i->bkey_type))
 640                         marking = true;
 641         }
 642
 643         if (trans->nr_wb_updates &&
 644             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 645                 return -BCH_ERR_btree_insert_need_flush_buffer;
 646
 647         /*
 648          * Don't get journal reservation until after we know insert will
 649          * succeed:
 650          */
 651         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 652                 ret = bch2_trans_journal_res_get(trans,
 653                                 (flags & JOURNAL_WATERMARK_MASK)|
 654                                 JOURNAL_RES_GET_NONBLOCK);
 655                 if (ret)
 656                         return ret;
 657
 658                 if (unlikely(trans->journal_transaction_names))
 659                         journal_transaction_name(trans);
 660         } else {
 661                 trans->journal_res.seq = c->journal.replay_journal_seq;
 662         }
 663
 664         /*
 665          * Not allowed to fail after we've gotten our journal reservation - we
 666          * have to use it:
 667          */
 668
 669         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 670             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 671                 if (bch2_journal_seq_verify)
 672                         trans_for_each_update(trans, i)
 673                                 i->k->k.version.lo = trans->journal_res.seq;
 674                 else if (bch2_inject_invalid_keys)
 675                         trans_for_each_update(trans, i)
 676                                 i->k->k.version = MAX_VERSION;
 677         }
 678
 679         if (trans->fs_usage_deltas &&
 680             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 681                 return -BCH_ERR_btree_insert_need_mark_replicas;
 682
 683         if (trans->nr_wb_updates) {
 684                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 685
 686                 ret = bch2_btree_insert_keys_write_buffer(trans);
 687                 if (ret)
 688                         goto revert_fs_usage;
 689         }
 690
 691         h = trans->hooks;
 692         while (h) {
 693                 ret = h->fn(trans, h);
 694                 if (ret)
 695                         goto revert_fs_usage;
 696                 h = h->next;
 697         }
 698
 699         trans_for_each_update(trans, i)
 700                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 701                         ret = run_one_mem_trigger(trans, i, i->flags);
 702                         if (ret)
 703                                 goto fatal_err;
 704                 }
 705
 706         if (unlikely(c->gc_pos.phase)) {
 707                 ret = bch2_trans_commit_run_gc_triggers(trans);
 708                 if  (ret)
 709                         goto fatal_err;
 710         }
 711
 712         if (unlikely(trans->extra_journal_entries.nr)) {
 713                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 714                                   trans->extra_journal_entries.data,
 715                                   trans->extra_journal_entries.nr);
 716
 717                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 718                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 719         }
 720
 721         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 722                 struct journal *j = &c->journal;
 723                 struct jset_entry *entry;
 724
 725                 trans_for_each_update(trans, i) {
 726                         if (i->key_cache_already_flushed)
 727                                 continue;
 728
 729                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 730                                 continue;
 731
 732                         verify_update_old_key(trans, i);
 733
 734                         if (trans->journal_transaction_names) {
 735                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 736                                                        BCH_JSET_ENTRY_overwrite,
 737                                                        i->btree_id, i->level,
 738                                                        i->old_k.u64s);
 739                                 bkey_reassemble(&entry->start[0],
 740                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 741                         }
 742
 743                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 744                                                BCH_JSET_ENTRY_btree_keys,
 745                                                i->btree_id, i->level,
 746                                                i->k->k.u64s);
 747                         bkey_copy(&entry->start[0], i->k);
 748                 }
 749
 750                 trans_for_each_wb_update(trans, wb) {
 751                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 752                                                BCH_JSET_ENTRY_btree_keys,
 753                                                wb->btree, 0,
 754                                                wb->k.k.u64s);
 755                         bkey_copy(&entry->start[0], &wb->k);
 756                 }
 757
 758                 if (trans->journal_seq)
 759                         *trans->journal_seq = trans->journal_res.seq;
 760         }
 761
 762         trans_for_each_update(trans, i) {
 763                 i->k->k.needs_whiteout = false;
 764
 765                 if (!i->cached)
 766                         btree_insert_key_leaf(trans, i);
 767                 else if (!i->key_cache_already_flushed)
 768                         bch2_btree_insert_key_cached(trans, flags, i);
 769                 else {
 770                         bch2_btree_key_cache_drop(trans, i->path);
 771                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 772                 }
 773         }
 774
 775         return 0;
 776 fatal_err:
 777         bch2_fatal_error(c);
 778 revert_fs_usage:
 779         if (trans->fs_usage_deltas)
 780                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 781         return ret;
 782 }
 783
 784 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 785 {
 786         while (--i >= trans->updates) {
 787                 if (same_leaf_as_prev(trans, i))
 788                         continue;
 789
 790                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 791         }
 792
 793         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 794         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 795 }
 796
 797 static inline int trans_lock_write(struct btree_trans *trans)
 798 {
 799         struct btree_insert_entry *i;
 800
 801         trans_for_each_update(trans, i) {
 802                 if (same_leaf_as_prev(trans, i))
 803                         continue;
 804
 805                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 806                         return trans_lock_write_fail(trans, i);
 807
 808                 if (!i->cached)
 809                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 810         }
 811
 812         return 0;
 813 }
 814
 815 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 816 {
 817         struct btree_insert_entry *i;
 818         struct btree_write_buffered_key *wb;
 819
 820         trans_for_each_update(trans, i)
 821                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 822
 823         trans_for_each_wb_update(trans, wb)
 824                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 825 }
 826
 827 #ifdef CONFIG_BCACHEFS_DEBUG
 828 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 829                                                    struct btree_insert_entry *i,
 830                                                    struct printbuf *err)
 831 {
 832         struct bch_fs *c = trans->c;
 833         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 834
 835         printbuf_reset(err);
 836         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 837                    trans->fn, (void *) i->ip_allocated);
 838         prt_newline(err);
 839         printbuf_indent_add(err, 2);
 840
 841         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 842         prt_newline(err);
 843
 844         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 845                           i->bkey_type, rw, err);
 846         bch2_print_string_as_lines(KERN_ERR, err->buf);
 847
 848         bch2_inconsistent_error(c);
 849         bch2_dump_trans_updates(trans);
 850         printbuf_exit(err);
 851
 852         return -EINVAL;
 853 }
 854 #endif
 855
 856 /*
 857  * Get journal reservation, take write locks, and attempt to do btree update(s):
 858  */
 859 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 860                                        struct btree_insert_entry **stopped_at,
 861                                        unsigned long trace_ip)
 862 {
 863         struct bch_fs *c = trans->c;
 864         struct btree_insert_entry *i;
 865         int ret, u64s_delta = 0;
 866
 867 #ifdef CONFIG_BCACHEFS_DEBUG
 868         struct printbuf buf = PRINTBUF;
 869
 870         trans_for_each_update(trans, i) {
 871                 int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 872
 873                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 874                                                i->bkey_type, rw, &buf)))
 875                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 876                 btree_insert_entry_checks(trans, i);
 877         }
 878         printbuf_exit(&buf);
 879 #endif
 880
 881         trans_for_each_update(trans, i) {
 882                 if (i->cached)
 883                         continue;
 884
 885                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 886                 u64s_delta -= i->old_btree_u64s;
 887
 888                 if (!same_leaf_as_next(trans, i)) {
 889                         if (u64s_delta <= 0) {
 890                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 891                                                         i->level, flags);
 892                                 if (unlikely(ret))
 893                                         return ret;
 894                         }
 895
 896                         u64s_delta = 0;
 897                 }
 898         }
 899
 900         ret = bch2_journal_preres_get(&c->journal,
 901                         &trans->journal_preres, trans->journal_preres_u64s,
 902                         (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 903         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 904                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 905         if (unlikely(ret))
 906                 return ret;
 907
 908         ret = trans_lock_write(trans);
 909         if (unlikely(ret))
 910                 return ret;
 911
 912         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 913
 914         if (!ret && unlikely(trans->journal_replay_not_finished))
 915                 bch2_drop_overwrites_from_journal(trans);
 916
 917         trans_for_each_update(trans, i)
 918                 if (!same_leaf_as_prev(trans, i))
 919                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 920                                                         insert_l(i)->b);
 921
 922         if (!ret && trans->journal_pin)
 923                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 924                                      trans->journal_pin, NULL);
 925
 926         /*
 927          * Drop journal reservation after dropping write locks, since dropping
 928          * the journal reservation may kick off a journal write:
 929          */
 930         bch2_journal_res_put(&c->journal, &trans->journal_res);
 931
 932         if (unlikely(ret))
 933                 return ret;
 934
 935         bch2_trans_downgrade(trans);
 936
 937         return 0;
 938 }
 939
 940 static int journal_reclaim_wait_done(struct bch_fs *c)
 941 {
 942         int ret = bch2_journal_error(&c->journal) ?:
 943                 !bch2_btree_key_cache_must_wait(c);
 944
 945         if (!ret)
 946                 journal_reclaim_kick(&c->journal);
 947         return ret;
 948 }
 949
 950 static noinline
 951 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 952                             struct btree_insert_entry *i,
 953                             int ret, unsigned long trace_ip)
 954 {
 955         struct bch_fs *c = trans->c;
 956
 957         switch (ret) {
 958         case -BCH_ERR_btree_insert_btree_node_full:
 959                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 960                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 961                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 962                 break;
 963         case -BCH_ERR_btree_insert_need_mark_replicas:
 964                 bch2_trans_unlock(trans);
 965
 966                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 967                 if (ret)
 968                         break;
 969
 970                 ret = bch2_trans_relock(trans);
 971                 if (ret)
 972                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 973                 break;
 974         case -BCH_ERR_journal_res_get_blocked:
 975                 bch2_trans_unlock(trans);
 976
 977                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 978                     !(flags & JOURNAL_WATERMARK_reserved)) {
 979                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 980                         break;
 981                 }
 982
 983                 ret = bch2_trans_journal_res_get(trans,
 984                                         (flags & JOURNAL_WATERMARK_MASK)|
 985                                         JOURNAL_RES_GET_CHECK);
 986                 if (ret)
 987                         break;
 988
 989                 ret = bch2_trans_relock(trans);
 990                 if (ret)
 991                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 992                 break;
 993         case -BCH_ERR_btree_insert_need_journal_reclaim:
 994                 bch2_trans_unlock(trans);
 995
 996                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 997
 998                 wait_event_freezable(c->journal.reclaim_wait,
 999                                      (ret = journal_reclaim_wait_done(c)));
1000                 if (ret < 0)
1001                         break;
1002
1003                 ret = bch2_trans_relock(trans);
1004                 if (ret)
1005                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
1006                 break;
1007         case -BCH_ERR_btree_insert_need_flush_buffer: {
1008                 struct btree_write_buffer *wb = &c->btree_write_buffer;
1009
1010                 ret = 0;
1011
1012                 if (wb->state.nr > wb->size * 3 / 4) {
1013                         bch2_trans_reset_updates(trans);
1014                         bch2_trans_unlock(trans);
1015
1016                         mutex_lock(&wb->flush_lock);
1017
1018                         if (wb->state.nr > wb->size * 3 / 4)
1019                                 ret = __bch2_btree_write_buffer_flush(trans,
1020                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
1021                         else
1022                                 mutex_unlock(&wb->flush_lock);
1023
1024                         if (!ret) {
1025                                 trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1026                                 ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1027                         }
1028                 }
1029                 break;
1030         }
1031         default:
1032                 BUG_ON(ret >= 0);
1033                 break;
1034         }
1035
1036         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1037
1038         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1039                                 !(flags & BTREE_INSERT_NOWAIT) &&
1040                                 (flags & BTREE_INSERT_NOFAIL), c,
1041                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1042
1043         return ret;
1044 }
1045
1046 static noinline int
1047 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1048 {
1049         struct bch_fs *c = trans->c;
1050         int ret;
1051
1052         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1053             test_bit(BCH_FS_STARTED, &c->flags))
1054                 return -BCH_ERR_erofs_trans_commit;
1055
1056         bch2_trans_unlock(trans);
1057
1058         ret =   bch2_fs_read_write_early(c) ?:
1059                 bch2_trans_relock(trans);
1060         if (ret)
1061                 return ret;
1062
1063         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1064         return 0;
1065 }
1066
1067 /*
1068  * This is for updates done in the early part of fsck - btree_gc - before we've
1069  * gone RW. we only add the new key to the list of keys for journal replay to
1070  * do.
1071  */
1072 static noinline int
1073 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1074 {
1075         struct bch_fs *c = trans->c;
1076         struct btree_insert_entry *i;
1077         int ret = 0;
1078
1079         trans_for_each_update(trans, i) {
1080                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1081                 if (ret)
1082                         break;
1083         }
1084
1085         return ret;
1086 }
1087
1088 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1089 {
1090         struct bch_fs *c = trans->c;
1091         struct btree_insert_entry *i = NULL;
1092         struct btree_write_buffered_key *wb;
1093         unsigned u64s;
1094         int ret = 0;
1095
1096         if (!trans->nr_updates &&
1097             !trans->nr_wb_updates &&
1098             !trans->extra_journal_entries.nr)
1099                 goto out_reset;
1100
1101         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1102                 lockdep_assert_held(&c->gc_lock);
1103
1104         ret = bch2_trans_commit_run_triggers(trans);
1105         if (ret)
1106                 goto out_reset;
1107
1108         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1109                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1110                 goto out_reset;
1111         }
1112
1113         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1114             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1115                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1116                 if (ret)
1117                         goto out_reset;
1118         }
1119
1120         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1121             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1122                 bch2_trans_begin(trans);
1123                 bch2_trans_unlock(trans);
1124
1125                 ret = __bch2_btree_write_buffer_flush(trans,
1126                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1127                 if (!ret) {
1128                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1129                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1130                 }
1131                 goto out;
1132         }
1133
1134         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1135
1136         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1137
1138         trans->journal_u64s             = trans->extra_journal_entries.nr;
1139         trans->journal_preres_u64s      = 0;
1140
1141         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1142
1143         if (trans->journal_transaction_names)
1144                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1145
1146         trans_for_each_update(trans, i) {
1147                 EBUG_ON(!i->path->should_be_locked);
1148
1149                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1150                 if (unlikely(ret))
1151                         goto out;
1152
1153                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1154
1155                 if (i->key_cache_already_flushed)
1156                         continue;
1157
1158                 /* we're going to journal the key being updated: */
1159                 u64s = jset_u64s(i->k->k.u64s);
1160                 if (i->cached &&
1161                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1162                         trans->journal_preres_u64s += u64s;
1163
1164                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1165                         continue;
1166
1167                 trans->journal_u64s += u64s;
1168
1169                 /* and we're also going to log the overwrite: */
1170                 if (trans->journal_transaction_names)
1171                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1172         }
1173
1174         trans_for_each_wb_update(trans, wb)
1175                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1176
1177         if (trans->extra_journal_res) {
1178                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1179                                 trans->extra_journal_res,
1180                                 (flags & BTREE_INSERT_NOFAIL)
1181                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1182                 if (ret)
1183                         goto err;
1184         }
1185 retry:
1186         bch2_trans_verify_not_in_restart(trans);
1187         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1188
1189         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1190
1191         /* make sure we didn't drop or screw up locks: */
1192         bch2_trans_verify_locks(trans);
1193
1194         if (ret)
1195                 goto err;
1196
1197         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1198 out:
1199         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1200
1201         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1202                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1203 out_reset:
1204         bch2_trans_reset_updates(trans);
1205
1206         return ret;
1207 err:
1208         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1209         if (ret)
1210                 goto out;
1211
1212         goto retry;
1213 }
1214
1215 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1216                                           enum btree_id id,
1217                                           struct bpos pos)
1218 {
1219         struct bch_fs *c = trans->c;
1220         struct btree_iter iter;
1221         struct bkey_s_c k;
1222         int ret;
1223
1224         bch2_trans_iter_init(trans, &iter, id, pos,
1225                              BTREE_ITER_NOT_EXTENTS|
1226                              BTREE_ITER_ALL_SNAPSHOTS);
1227         while (1) {
1228                 k = bch2_btree_iter_prev(&iter);
1229                 ret = bkey_err(k);
1230                 if (ret)
1231                         break;
1232
1233                 if (!k.k)
1234                         break;
1235
1236                 if (!bkey_eq(pos, k.k->p))
1237                         break;
1238
1239                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1240                         ret = 1;
1241                         break;
1242                 }
1243         }
1244         bch2_trans_iter_exit(trans, &iter);
1245
1246         return ret;
1247 }
1248
1249 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1250                                           enum btree_id id,
1251                                           struct bpos pos)
1252 {
1253         if (!btree_type_has_snapshots(id) ||
1254             pos.snapshot == U32_MAX ||
1255             !snapshot_t(trans->c, pos.snapshot)->children[0])
1256                 return 0;
1257
1258         return __check_pos_snapshot_overwritten(trans, id, pos);
1259 }
1260
1261 static noinline int extent_front_merge(struct btree_trans *trans,
1262                                        struct btree_iter *iter,
1263                                        struct bkey_s_c k,
1264                                        struct bkey_i **insert,
1265                                        enum btree_update_flags flags)
1266 {
1267         struct bch_fs *c = trans->c;
1268         struct bkey_i *update;
1269         int ret;
1270
1271         update = bch2_bkey_make_mut_noupdate(trans, k);
1272         ret = PTR_ERR_OR_ZERO(update);
1273         if (ret)
1274                 return ret;
1275
1276         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1277                 return 0;
1278
1279         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1280                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1281         if (ret < 0)
1282                 return ret;
1283         if (ret)
1284                 return 0;
1285
1286         ret = bch2_btree_delete_at(trans, iter, flags);
1287         if (ret)
1288                 return ret;
1289
1290         *insert = update;
1291         return 0;
1292 }
1293
1294 static noinline int extent_back_merge(struct btree_trans *trans,
1295                                       struct btree_iter *iter,
1296                                       struct bkey_i *insert,
1297                                       struct bkey_s_c k)
1298 {
1299         struct bch_fs *c = trans->c;
1300         int ret;
1301
1302         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1303                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1304         if (ret < 0)
1305                 return ret;
1306         if (ret)
1307                 return 0;
1308
1309         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1310         return 0;
1311 }
1312
1313 /*
1314  * When deleting, check if we need to emit a whiteout (because we're overwriting
1315  * something in an ancestor snapshot)
1316  */
1317 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1318                                       enum btree_id btree_id, struct bpos pos)
1319 {
1320         struct btree_iter iter;
1321         struct bkey_s_c k;
1322         u32 snapshot = pos.snapshot;
1323         int ret;
1324
1325         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1326                 return 0;
1327
1328         pos.snapshot++;
1329
1330         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1331                            BTREE_ITER_ALL_SNAPSHOTS|
1332                            BTREE_ITER_NOPRESERVE, k, ret) {
1333                 if (!bkey_eq(k.k->p, pos))
1334                         break;
1335
1336                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1337                                               k.k->p.snapshot)) {
1338                         ret = !bkey_whiteout(k.k);
1339                         break;
1340                 }
1341         }
1342         bch2_trans_iter_exit(trans, &iter);
1343
1344         return ret;
1345 }
1346 int bch2_trans_update_extent(struct btree_trans *trans,
1347                              struct btree_iter *orig_iter,
1348                              struct bkey_i *insert,
1349                              enum btree_update_flags flags)
1350 {
1351         struct btree_iter iter;
1352         struct bpos start = bkey_start_pos(&insert->k);
1353         struct bkey_i *update;
1354         struct bkey_s_c k;
1355         enum btree_id btree_id = orig_iter->btree_id;
1356         int ret = 0, compressed_sectors;
1357
1358         bch2_trans_iter_init(trans, &iter, btree_id, start,
1359                              BTREE_ITER_INTENT|
1360                              BTREE_ITER_WITH_UPDATES|
1361                              BTREE_ITER_NOT_EXTENTS);
1362         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1363         if ((ret = bkey_err(k)))
1364                 goto err;
1365         if (!k.k)
1366                 goto out;
1367
1368         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1369                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1370                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1371                         if (ret)
1372                                 goto err;
1373                 }
1374
1375                 goto next;
1376         }
1377
1378         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1379                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1380                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1381
1382                 /*
1383                  * If we're going to be splitting a compressed extent, note it
1384                  * so that __bch2_trans_commit() can increase our disk
1385                  * reservation:
1386                  */
1387                 if (((front_split && back_split) ||
1388                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1389                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1390                         trans->extra_journal_res += compressed_sectors;
1391
1392                 if (front_split) {
1393                         update = bch2_bkey_make_mut_noupdate(trans, k);
1394                         if ((ret = PTR_ERR_OR_ZERO(update)))
1395                                 goto err;
1396
1397                         bch2_cut_back(start, update);
1398
1399                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1400                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1401                         if (ret)
1402                                 goto err;
1403                 }
1404
1405                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1406                     (front_split || back_split)) {
1407                         update = bch2_bkey_make_mut_noupdate(trans, k);
1408                         if ((ret = PTR_ERR_OR_ZERO(update)))
1409                                 goto err;
1410
1411                         bch2_cut_front(start, update);
1412                         bch2_cut_back(insert->k.p, update);
1413
1414                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1415                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1416                         if (ret)
1417                                 goto err;
1418                 }
1419
1420                 if (bkey_le(k.k->p, insert->k.p)) {
1421                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1422                         if ((ret = PTR_ERR_OR_ZERO(update)))
1423                                 goto err;
1424
1425                         bkey_init(&update->k);
1426                         update->k.p = k.k->p;
1427                         update->k.p.snapshot = insert->k.p.snapshot;
1428
1429                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1430                                 update->k.type = KEY_TYPE_whiteout;
1431                         } else if (btree_type_has_snapshots(btree_id)) {
1432                                 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1433                                 if (ret < 0)
1434                                         goto err;
1435                                 if (ret)
1436                                         update->k.type = KEY_TYPE_whiteout;
1437                         }
1438
1439                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1440                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1441                         if (ret)
1442                                 goto err;
1443                 }
1444
1445                 if (back_split) {
1446                         update = bch2_bkey_make_mut_noupdate(trans, k);
1447                         if ((ret = PTR_ERR_OR_ZERO(update)))
1448                                 goto err;
1449
1450                         bch2_cut_front(insert->k.p, update);
1451
1452                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1453                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1454                                                   flags);
1455                         if (ret)
1456                                 goto err;
1457                         goto out;
1458                 }
1459 next:
1460                 bch2_btree_iter_advance(&iter);
1461                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1462                 if ((ret = bkey_err(k)))
1463                         goto err;
1464                 if (!k.k)
1465                         goto out;
1466         }
1467
1468         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1469                 ret = extent_back_merge(trans, &iter, insert, k);
1470                 if (ret)
1471                         goto err;
1472         }
1473 out:
1474         if (!bkey_deleted(&insert->k)) {
1475                 /*
1476                  * Rewinding iterators is expensive: get a new one and the one
1477                  * that points to the start of insert will be cloned from:
1478                  */
1479                 bch2_trans_iter_exit(trans, &iter);
1480                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1481                                      BTREE_ITER_NOT_EXTENTS|
1482                                      BTREE_ITER_INTENT);
1483                 ret   = bch2_btree_iter_traverse(&iter) ?:
1484                         bch2_trans_update(trans, &iter, insert, flags);
1485         }
1486 err:
1487         bch2_trans_iter_exit(trans, &iter);
1488
1489         return ret;
1490 }
1491
1492 static int __must_check
1493 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1494                                 struct bkey_i *k, enum btree_update_flags flags,
1495                                 unsigned long ip);
1496
1497 static noinline int flush_new_cached_update(struct btree_trans *trans,
1498                                             struct btree_path *path,
1499                                             struct btree_insert_entry *i,
1500                                             enum btree_update_flags flags,
1501                                             unsigned long ip)
1502 {
1503         struct btree_path *btree_path;
1504         struct bkey k;
1505         int ret;
1506
1507         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1508                                    BTREE_ITER_INTENT, _THIS_IP_);
1509         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1510         if (ret)
1511                 goto out;
1512
1513         /*
1514          * The old key in the insert entry might actually refer to an existing
1515          * key in the btree that has been deleted from cache and not yet
1516          * flushed. Check for this and skip the flush so we don't run triggers
1517          * against a stale key.
1518          */
1519         bch2_btree_path_peek_slot_exact(btree_path, &k);
1520         if (!bkey_deleted(&k))
1521                 goto out;
1522
1523         i->key_cache_already_flushed = true;
1524         i->flags |= BTREE_TRIGGER_NORUN;
1525
1526         btree_path_set_should_be_locked(btree_path);
1527         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1528 out:
1529         bch2_path_put(trans, btree_path, true);
1530         return ret;
1531 }
1532
1533 static int __must_check
1534 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1535                                 struct bkey_i *k, enum btree_update_flags flags,
1536                                 unsigned long ip)
1537 {
1538         struct bch_fs *c = trans->c;
1539         struct btree_insert_entry *i, n;
1540         int cmp;
1541
1542         EBUG_ON(!path->should_be_locked);
1543         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1544         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1545
1546         n = (struct btree_insert_entry) {
1547                 .flags          = flags,
1548                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1549                 .btree_id       = path->btree_id,
1550                 .level          = path->level,
1551                 .cached         = path->cached,
1552                 .path           = path,
1553                 .k              = k,
1554                 .ip_allocated   = ip,
1555         };
1556
1557 #ifdef CONFIG_BCACHEFS_DEBUG
1558         trans_for_each_update(trans, i)
1559                 BUG_ON(i != trans->updates &&
1560                        btree_insert_entry_cmp(i - 1, i) >= 0);
1561 #endif
1562
1563         /*
1564          * Pending updates are kept sorted: first, find position of new update,
1565          * then delete/trim any updates the new update overwrites:
1566          */
1567         trans_for_each_update(trans, i) {
1568                 cmp = btree_insert_entry_cmp(&n, i);
1569                 if (cmp <= 0)
1570                         break;
1571         }
1572
1573         if (!cmp && i < trans->updates + trans->nr_updates) {
1574                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1575
1576                 bch2_path_put(trans, i->path, true);
1577                 i->flags        = n.flags;
1578                 i->cached       = n.cached;
1579                 i->k            = n.k;
1580                 i->path         = n.path;
1581                 i->ip_allocated = n.ip_allocated;
1582         } else {
1583                 array_insert_item(trans->updates, trans->nr_updates,
1584                                   i - trans->updates, n);
1585
1586                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1587                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1588
1589                 if (unlikely(trans->journal_replay_not_finished)) {
1590                         struct bkey_i *j_k =
1591                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1592
1593                         if (j_k) {
1594                                 i->old_k = j_k->k;
1595                                 i->old_v = &j_k->v;
1596                         }
1597                 }
1598         }
1599
1600         __btree_path_get(i->path, true);
1601
1602         /*
1603          * If a key is present in the key cache, it must also exist in the
1604          * btree - this is necessary for cache coherency. When iterating over
1605          * a btree that's cached in the key cache, the btree iter code checks
1606          * the key cache - but the key has to exist in the btree for that to
1607          * work:
1608          */
1609         if (path->cached && bkey_deleted(&i->old_k))
1610                 return flush_new_cached_update(trans, path, i, flags, ip);
1611
1612         return 0;
1613 }
1614
1615 static inline int __must_check
1616 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1617                           struct bkey_i *k, enum btree_update_flags flags)
1618 {
1619         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1620 }
1621
1622 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1623                                    struct bkey_i *k, enum btree_update_flags flags)
1624 {
1625         struct btree_path *path = iter->update_path ?: iter->path;
1626         struct bkey_cached *ck;
1627         int ret;
1628
1629         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1630                 return bch2_trans_update_extent(trans, iter, k, flags);
1631
1632         if (bkey_deleted(&k->k) &&
1633             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1634             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1635                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1636                 if (unlikely(ret < 0))
1637                         return ret;
1638
1639                 if (ret)
1640                         k->k.type = KEY_TYPE_whiteout;
1641         }
1642
1643         /*
1644          * Ensure that updates to cached btrees go to the key cache:
1645          */
1646         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1647             !path->cached &&
1648             !path->level &&
1649             btree_id_cached(trans->c, path->btree_id)) {
1650                 if (!iter->key_cache_path ||
1651                     !iter->key_cache_path->should_be_locked ||
1652                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1653                         if (!iter->key_cache_path)
1654                                 iter->key_cache_path =
1655                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1656                                                       BTREE_ITER_INTENT|
1657                                                       BTREE_ITER_CACHED, _THIS_IP_);
1658
1659                         iter->key_cache_path =
1660                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1661                                                         iter->flags & BTREE_ITER_INTENT,
1662                                                         _THIS_IP_);
1663
1664                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1665                                                        BTREE_ITER_CACHED);
1666                         if (unlikely(ret))
1667                                 return ret;
1668
1669                         ck = (void *) iter->key_cache_path->l[0].b;
1670
1671                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1672                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1673                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1674                         }
1675
1676                         btree_path_set_should_be_locked(iter->key_cache_path);
1677                 }
1678
1679                 path = iter->key_cache_path;
1680         }
1681
1682         return bch2_trans_update_by_path(trans, path, k, flags);
1683 }
1684
1685 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1686                                             enum btree_id btree,
1687                                             struct bkey_i *k)
1688 {
1689         struct btree_write_buffered_key *i;
1690         int ret;
1691
1692         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1693         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1694
1695         trans_for_each_wb_update(trans, i) {
1696                 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
1697                         bkey_copy(&i->k, k);
1698                         return 0;
1699                 }
1700         }
1701
1702         if (!trans->wb_updates ||
1703             trans->nr_wb_updates == trans->wb_updates_size) {
1704                 struct btree_write_buffered_key *u;
1705
1706                 if (trans->nr_wb_updates == trans->wb_updates_size) {
1707                         struct btree_transaction_stats *s = btree_trans_stats(trans);
1708
1709                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1710                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1711                         if (s)
1712                                 s->wb_updates_size = trans->wb_updates_size;
1713                 }
1714
1715                 u = bch2_trans_kmalloc_nomemzero(trans,
1716                                         trans->wb_updates_size *
1717                                         sizeof(struct btree_write_buffered_key));
1718                 ret = PTR_ERR_OR_ZERO(u);
1719                 if (ret)
1720                         return ret;
1721
1722                 if (trans->nr_wb_updates)
1723                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1724                                sizeof(struct btree_write_buffered_key));
1725                 trans->wb_updates = u;
1726         }
1727
1728         trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
1729                 .btree  = btree,
1730         };
1731
1732         bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
1733         trans->nr_wb_updates++;
1734
1735         return 0;
1736 }
1737
1738 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1739                              enum btree_id btree, struct bpos end)
1740 {
1741         struct bkey_s_c k;
1742         int ret = 0;
1743
1744         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1745         k = bch2_btree_iter_prev(iter);
1746         ret = bkey_err(k);
1747         if (ret)
1748                 goto err;
1749
1750         bch2_btree_iter_advance(iter);
1751         k = bch2_btree_iter_peek_slot(iter);
1752         ret = bkey_err(k);
1753         if (ret)
1754                 goto err;
1755
1756         BUG_ON(k.k->type != KEY_TYPE_deleted);
1757
1758         if (bkey_gt(k.k->p, end)) {
1759                 ret = -BCH_ERR_ENOSPC_btree_slot;
1760                 goto err;
1761         }
1762
1763         return 0;
1764 err:
1765         bch2_trans_iter_exit(trans, iter);
1766         return ret;
1767 }
1768
1769 void bch2_trans_commit_hook(struct btree_trans *trans,
1770                             struct btree_trans_commit_hook *h)
1771 {
1772         h->next = trans->hooks;
1773         trans->hooks = h;
1774 }
1775
1776 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1777                                 enum btree_id btree, struct bkey_i *k,
1778                                 enum btree_update_flags flags)
1779 {
1780         struct btree_iter iter;
1781         int ret;
1782
1783         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1784                              BTREE_ITER_NOT_EXTENTS|
1785                              BTREE_ITER_INTENT);
1786         ret   = bch2_btree_iter_traverse(&iter) ?:
1787                 bch2_trans_update(trans, &iter, k, flags);
1788         bch2_trans_iter_exit(trans, &iter);
1789         return ret;
1790 }
1791
1792 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1793                         struct bkey_i *k, enum btree_update_flags flags)
1794 {
1795         struct btree_iter iter;
1796         int ret;
1797
1798         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1799                              BTREE_ITER_CACHED|
1800                              BTREE_ITER_INTENT);
1801         ret   = bch2_btree_iter_traverse(&iter) ?:
1802                 bch2_trans_update(trans, &iter, k, flags);
1803         bch2_trans_iter_exit(trans, &iter);
1804         return ret;
1805 }
1806
1807 /**
1808  * bch2_btree_insert - insert keys into the extent btree
1809  * @c:                  pointer to struct bch_fs
1810  * @id:                 btree to insert into
1811  * @insert_keys:        list of keys to insert
1812  * @hook:               insert callback
1813  */
1814 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1815                       struct bkey_i *k,
1816                       struct disk_reservation *disk_res,
1817                       u64 *journal_seq, int flags)
1818 {
1819         return bch2_trans_do(c, disk_res, journal_seq, flags,
1820                              __bch2_btree_insert(&trans, id, k, 0));
1821 }
1822
1823 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1824                                 unsigned len, unsigned update_flags)
1825 {
1826         struct bkey_i *k;
1827
1828         k = bch2_trans_kmalloc(trans, sizeof(*k));
1829         if (IS_ERR(k))
1830                 return PTR_ERR(k);
1831
1832         bkey_init(&k->k);
1833         k->k.p = iter->pos;
1834         bch2_key_resize(&k->k, len);
1835         return bch2_trans_update(trans, iter, k, update_flags);
1836 }
1837
1838 int bch2_btree_delete_at(struct btree_trans *trans,
1839                          struct btree_iter *iter, unsigned update_flags)
1840 {
1841         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1842 }
1843
1844 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1845                                   enum btree_id btree, struct bpos pos)
1846 {
1847         struct bkey_i *k;
1848
1849         k = bch2_trans_kmalloc(trans, sizeof(*k));
1850         if (IS_ERR(k))
1851                 return PTR_ERR(k);
1852
1853         bkey_init(&k->k);
1854         k->k.p = pos;
1855         return bch2_trans_update_buffered(trans, btree, k);
1856 }
1857
1858 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1859                                   struct bpos start, struct bpos end,
1860                                   unsigned update_flags,
1861                                   u64 *journal_seq)
1862 {
1863         u32 restart_count = trans->restart_count;
1864         struct btree_iter iter;
1865         struct bkey_s_c k;
1866         int ret = 0;
1867
1868         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1869         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1870                 struct disk_reservation disk_res =
1871                         bch2_disk_reservation_init(trans->c, 0);
1872                 struct bkey_i delete;
1873
1874                 ret = bkey_err(k);
1875                 if (ret)
1876                         goto err;
1877
1878                 bkey_init(&delete.k);
1879
1880                 /*
1881                  * This could probably be more efficient for extents:
1882                  */
1883
1884                 /*
1885                  * For extents, iter.pos won't necessarily be the same as
1886                  * bkey_start_pos(k.k) (for non extents they always will be the
1887                  * same). It's important that we delete starting from iter.pos
1888                  * because the range we want to delete could start in the middle
1889                  * of k.
1890                  *
1891                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1892                  * bkey_start_pos(k.k)).
1893                  */
1894                 delete.k.p = iter.pos;
1895
1896                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1897                         bch2_key_resize(&delete.k,
1898                                         bpos_min(end, k.k->p).offset -
1899                                         iter.pos.offset);
1900
1901                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1902                         bch2_trans_commit(trans, &disk_res, journal_seq,
1903                                           BTREE_INSERT_NOFAIL);
1904                 bch2_disk_reservation_put(trans->c, &disk_res);
1905 err:
1906                 /*
1907                  * the bch2_trans_begin() call is in a weird place because we
1908                  * need to call it after every transaction commit, to avoid path
1909                  * overflow, but don't want to call it if the delete operation
1910                  * is a no-op and we have no work to do:
1911                  */
1912                 bch2_trans_begin(trans);
1913
1914                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1915                         ret = 0;
1916                 if (ret)
1917                         break;
1918         }
1919         bch2_trans_iter_exit(trans, &iter);
1920
1921         if (!ret && trans_was_restarted(trans, restart_count))
1922                 ret = -BCH_ERR_transaction_restart_nested;
1923         return ret;
1924 }
1925
1926 /*
1927  * bch_btree_delete_range - delete everything within a given range
1928  *
1929  * Range is a half open interval - [start, end)
1930  */
1931 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1932                             struct bpos start, struct bpos end,
1933                             unsigned update_flags,
1934                             u64 *journal_seq)
1935 {
1936         int ret = bch2_trans_run(c,
1937                         bch2_btree_delete_range_trans(&trans, id, start, end,
1938                                                       update_flags, journal_seq));
1939         if (ret == -BCH_ERR_transaction_restart_nested)
1940                 ret = 0;
1941         return ret;
1942 }
1943
1944 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1945 {
1946         struct printbuf buf = PRINTBUF;
1947         struct jset_entry_log *l;
1948         unsigned u64s;
1949         int ret;
1950
1951         prt_vprintf(&buf, fmt, args);
1952         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
1953         if (ret)
1954                 goto err;
1955
1956         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
1957
1958         ret = darray_make_room(entries, jset_u64s(u64s));
1959         if (ret)
1960                 goto err;
1961
1962         l = (void *) &darray_top(*entries);
1963         l->entry.u64s           = cpu_to_le16(u64s);
1964         l->entry.btree_id       = 0;
1965         l->entry.level          = 1;
1966         l->entry.type           = BCH_JSET_ENTRY_log;
1967         l->entry.pad[0]         = 0;
1968         l->entry.pad[1]         = 0;
1969         l->entry.pad[2]         = 0;
1970         memcpy(l->d, buf.buf, buf.pos);
1971         while (buf.pos & 7)
1972                 l->d[buf.pos++] = '\0';
1973
1974         entries->nr += jset_u64s(u64s);
1975 err:
1976         printbuf_exit(&buf);
1977         return ret;
1978 }
1979
1980 static int
1981 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
1982                   va_list args)
1983 {
1984         int ret;
1985
1986         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
1987                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
1988         } else {
1989                 ret = bch2_trans_do(c, NULL, NULL,
1990                         BTREE_INSERT_LAZY_RW|commit_flags,
1991                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
1992         }
1993
1994         return ret;
1995 }
1996
1997 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
1998 {
1999         va_list args;
2000         int ret;
2001
2002         va_start(args, fmt);
2003         ret = __bch2_fs_log_msg(c, 0, fmt, args);
2004         va_end(args);
2005         return ret;
2006 }
2007
2008 /*
2009  * Use for logging messages during recovery to enable reserved space and avoid
2010  * blocking.
2011  */
2012 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2013 {
2014         va_list args;
2015         int ret;
2016
2017         va_start(args, fmt);
2018         ret = __bch2_fs_log_msg(c, JOURNAL_WATERMARK_reserved, fmt, args);
2019         va_end(args);
2020         return ret;
2021 }