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         return drop_locks_do(trans,
 320                 bch2_journal_preres_get(&trans->c->journal,
 321                         &trans->journal_preres,
 322                         trans->journal_preres_u64s,
 323                         (flags & JOURNAL_WATERMARK_MASK)));
 324 }
 325
 326 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 327                                                       unsigned flags)
 328 {
 329         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 330                                     trans->journal_u64s, flags);
 331 }
 332
 333 #define JSET_ENTRY_LOG_U64s             4
 334
 335 static noinline void journal_transaction_name(struct btree_trans *trans)
 336 {
 337         struct bch_fs *c = trans->c;
 338         struct journal *j = &c->journal;
 339         struct jset_entry *entry =
 340                 bch2_journal_add_entry(j, &trans->journal_res,
 341                                        BCH_JSET_ENTRY_log, 0, 0,
 342                                        JSET_ENTRY_LOG_U64s);
 343         struct jset_entry_log *l =
 344                 container_of(entry, struct jset_entry_log, entry);
 345
 346         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 347 }
 348
 349 static inline int btree_key_can_insert(struct btree_trans *trans,
 350                                        struct btree *b, unsigned u64s)
 351 {
 352         struct bch_fs *c = trans->c;
 353
 354         if (!bch2_btree_node_insert_fits(c, b, u64s))
 355                 return -BCH_ERR_btree_insert_btree_node_full;
 356
 357         return 0;
 358 }
 359
 360 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 361                                        struct btree_path *path, unsigned u64s)
 362 {
 363         struct bch_fs *c = trans->c;
 364         struct bkey_cached *ck = (void *) path->l[0].b;
 365         struct btree_insert_entry *i;
 366         unsigned new_u64s;
 367         struct bkey_i *new_k;
 368
 369         EBUG_ON(path->level);
 370
 371         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 372             bch2_btree_key_cache_must_wait(c) &&
 373             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 374                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 375
 376         /*
 377          * bch2_varint_decode can read past the end of the buffer by at most 7
 378          * bytes (it won't be used):
 379          */
 380         u64s += 1;
 381
 382         if (u64s <= ck->u64s)
 383                 return 0;
 384
 385         new_u64s        = roundup_pow_of_two(u64s);
 386         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 387         if (!new_k) {
 388                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 389                         bch2_btree_ids[path->btree_id], new_u64s);
 390                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 391         }
 392
 393         trans_for_each_update(trans, i)
 394                 if (i->old_v == &ck->k->v)
 395                         i->old_v = &new_k->v;
 396
 397         ck->u64s        = new_u64s;
 398         ck->k           = new_k;
 399         return 0;
 400 }
 401
 402 /* Triggers: */
 403
 404 static int run_one_mem_trigger(struct btree_trans *trans,
 405                                struct btree_insert_entry *i,
 406                                unsigned flags)
 407 {
 408         struct bkey_s_c old = { &i->old_k, i->old_v };
 409         struct bkey_i *new = i->k;
 410         int ret;
 411
 412         verify_update_old_key(trans, i);
 413
 414         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 415                 return 0;
 416
 417         if (!btree_node_type_needs_gc(i->btree_id))
 418                 return 0;
 419
 420         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 421             bch2_bkey_ops[i->k->k.type].atomic_trigger) {
 422                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 423                                 old, bkey_i_to_s_c(new),
 424                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 425         } else {
 426                 struct bkey             _deleted = KEY(0, 0, 0);
 427                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 428
 429                 _deleted.p = i->path->pos;
 430
 431                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 432                                 deleted, bkey_i_to_s_c(new),
 433                                 BTREE_TRIGGER_INSERT|flags) ?:
 434                         bch2_mark_key(trans, i->btree_id, i->level,
 435                                 old, deleted,
 436                                 BTREE_TRIGGER_OVERWRITE|flags);
 437         }
 438
 439         return ret;
 440 }
 441
 442 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 443                                  bool overwrite)
 444 {
 445         /*
 446          * Transactional triggers create new btree_insert_entries, so we can't
 447          * pass them a pointer to a btree_insert_entry, that memory is going to
 448          * move:
 449          */
 450         struct bkey old_k = i->old_k;
 451         struct bkey_s_c old = { &old_k, i->old_v };
 452
 453         verify_update_old_key(trans, i);
 454
 455         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 456             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 457                 return 0;
 458
 459         if (!i->insert_trigger_run &&
 460             !i->overwrite_trigger_run &&
 461             bch2_bkey_ops[old.k->type].trans_trigger ==
 462             bch2_bkey_ops[i->k->k.type].trans_trigger) {
 463                 i->overwrite_trigger_run = true;
 464                 i->insert_trigger_run = true;
 465                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 466                                            BTREE_TRIGGER_INSERT|
 467                                            BTREE_TRIGGER_OVERWRITE|
 468                                            i->flags) ?: 1;
 469         } else if (overwrite && !i->overwrite_trigger_run) {
 470                 i->overwrite_trigger_run = true;
 471                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 472         } else if (!overwrite && !i->insert_trigger_run) {
 473                 i->insert_trigger_run = true;
 474                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 475         } else {
 476                 return 0;
 477         }
 478 }
 479
 480 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 481                               struct btree_insert_entry *btree_id_start)
 482 {
 483         struct btree_insert_entry *i;
 484         bool trans_trigger_run;
 485         int ret, overwrite;
 486
 487         for (overwrite = 1; overwrite >= 0; --overwrite) {
 488
 489                 /*
 490                  * Running triggers will append more updates to the list of updates as
 491                  * we're walking it:
 492                  */
 493                 do {
 494                         trans_trigger_run = false;
 495
 496                         for (i = btree_id_start;
 497                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 498                              i++) {
 499                                 if (i->btree_id != btree_id)
 500                                         continue;
 501
 502                                 ret = run_one_trans_trigger(trans, i, overwrite);
 503                                 if (ret < 0)
 504                                         return ret;
 505                                 if (ret)
 506                                         trans_trigger_run = true;
 507                         }
 508                 } while (trans_trigger_run);
 509         }
 510
 511         return 0;
 512 }
 513
 514 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 515 {
 516         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 517         unsigned btree_id = 0;
 518         int ret = 0;
 519
 520         /*
 521          *
 522          * For a given btree, this algorithm runs insert triggers before
 523          * overwrite triggers: this is so that when extents are being moved
 524          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 525          * they are re-added.
 526          */
 527         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 528                 if (btree_id == BTREE_ID_alloc)
 529                         continue;
 530
 531                 while (btree_id_start < trans->updates + trans->nr_updates &&
 532                        btree_id_start->btree_id < btree_id)
 533                         btree_id_start++;
 534
 535                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 536                 if (ret)
 537                         return ret;
 538         }
 539
 540         trans_for_each_update(trans, i) {
 541                 if (i->btree_id > BTREE_ID_alloc)
 542                         break;
 543                 if (i->btree_id == BTREE_ID_alloc) {
 544                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 545                         if (ret)
 546                                 return ret;
 547                         break;
 548                 }
 549         }
 550
 551 #ifdef CONFIG_BCACHEFS_DEBUG
 552         trans_for_each_update(trans, i)
 553                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 554                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 555                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 556 #endif
 557         return 0;
 558 }
 559
 560 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 561 {
 562         struct bch_fs *c = trans->c;
 563         struct btree_insert_entry *i;
 564         int ret = 0;
 565
 566         trans_for_each_update(trans, i) {
 567                 /*
 568                  * XXX: synchronization of cached update triggers with gc
 569                  * XXX: synchronization of interior node updates with gc
 570                  */
 571                 BUG_ON(i->cached || i->level);
 572
 573                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 574                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 575                         if (ret)
 576                                 break;
 577                 }
 578         }
 579
 580         return ret;
 581 }
 582
 583 static inline int
 584 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 585                                struct btree_insert_entry **stopped_at,
 586                                unsigned long trace_ip)
 587 {
 588         struct bch_fs *c = trans->c;
 589         struct btree_insert_entry *i;
 590         struct btree_write_buffered_key *wb;
 591         struct btree_trans_commit_hook *h;
 592         unsigned u64s = 0;
 593         bool marking = false;
 594         int ret;
 595
 596         if (race_fault()) {
 597                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 598                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 599         }
 600
 601         /*
 602          * Check if the insert will fit in the leaf node with the write lock
 603          * held, otherwise another thread could write the node changing the
 604          * amount of space available:
 605          */
 606
 607         prefetch(&trans->c->journal.flags);
 608
 609         trans_for_each_update(trans, i) {
 610                 /* Multiple inserts might go to same leaf: */
 611                 if (!same_leaf_as_prev(trans, i))
 612                         u64s = 0;
 613
 614                 u64s += i->k->k.u64s;
 615                 ret = !i->cached
 616                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 617                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 618                 if (ret) {
 619                         *stopped_at = i;
 620                         return ret;
 621                 }
 622
 623                 if (btree_node_type_needs_gc(i->bkey_type))
 624                         marking = true;
 625         }
 626
 627         if (trans->nr_wb_updates &&
 628             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 629                 return -BCH_ERR_btree_insert_need_flush_buffer;
 630
 631         /*
 632          * Don't get journal reservation until after we know insert will
 633          * succeed:
 634          */
 635         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 636                 ret = bch2_trans_journal_res_get(trans,
 637                                 (flags & JOURNAL_WATERMARK_MASK)|
 638                                 JOURNAL_RES_GET_NONBLOCK);
 639                 if (ret)
 640                         return ret;
 641
 642                 if (unlikely(trans->journal_transaction_names))
 643                         journal_transaction_name(trans);
 644         } else {
 645                 trans->journal_res.seq = c->journal.replay_journal_seq;
 646         }
 647
 648         /*
 649          * Not allowed to fail after we've gotten our journal reservation - we
 650          * have to use it:
 651          */
 652
 653         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 654             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 655                 if (bch2_journal_seq_verify)
 656                         trans_for_each_update(trans, i)
 657                                 i->k->k.version.lo = trans->journal_res.seq;
 658                 else if (bch2_inject_invalid_keys)
 659                         trans_for_each_update(trans, i)
 660                                 i->k->k.version = MAX_VERSION;
 661         }
 662
 663         if (trans->fs_usage_deltas &&
 664             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 665                 return -BCH_ERR_btree_insert_need_mark_replicas;
 666
 667         if (trans->nr_wb_updates) {
 668                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 669
 670                 ret = bch2_btree_insert_keys_write_buffer(trans);
 671                 if (ret)
 672                         goto revert_fs_usage;
 673         }
 674
 675         h = trans->hooks;
 676         while (h) {
 677                 ret = h->fn(trans, h);
 678                 if (ret)
 679                         goto revert_fs_usage;
 680                 h = h->next;
 681         }
 682
 683         trans_for_each_update(trans, i)
 684                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 685                         ret = run_one_mem_trigger(trans, i, i->flags);
 686                         if (ret)
 687                                 goto fatal_err;
 688                 }
 689
 690         if (unlikely(c->gc_pos.phase)) {
 691                 ret = bch2_trans_commit_run_gc_triggers(trans);
 692                 if  (ret)
 693                         goto fatal_err;
 694         }
 695
 696         if (unlikely(trans->extra_journal_entries.nr)) {
 697                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 698                                   trans->extra_journal_entries.data,
 699                                   trans->extra_journal_entries.nr);
 700
 701                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 702                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 703         }
 704
 705         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 706                 struct journal *j = &c->journal;
 707                 struct jset_entry *entry;
 708
 709                 trans_for_each_update(trans, i) {
 710                         if (i->key_cache_already_flushed)
 711                                 continue;
 712
 713                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 714                                 continue;
 715
 716                         verify_update_old_key(trans, i);
 717
 718                         if (trans->journal_transaction_names) {
 719                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 720                                                        BCH_JSET_ENTRY_overwrite,
 721                                                        i->btree_id, i->level,
 722                                                        i->old_k.u64s);
 723                                 bkey_reassemble(&entry->start[0],
 724                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 725                         }
 726
 727                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 728                                                BCH_JSET_ENTRY_btree_keys,
 729                                                i->btree_id, i->level,
 730                                                i->k->k.u64s);
 731                         bkey_copy(&entry->start[0], i->k);
 732                 }
 733
 734                 trans_for_each_wb_update(trans, wb) {
 735                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 736                                                BCH_JSET_ENTRY_btree_keys,
 737                                                wb->btree, 0,
 738                                                wb->k.k.u64s);
 739                         bkey_copy(&entry->start[0], &wb->k);
 740                 }
 741
 742                 if (trans->journal_seq)
 743                         *trans->journal_seq = trans->journal_res.seq;
 744         }
 745
 746         trans_for_each_update(trans, i) {
 747                 i->k->k.needs_whiteout = false;
 748
 749                 if (!i->cached)
 750                         btree_insert_key_leaf(trans, i);
 751                 else if (!i->key_cache_already_flushed)
 752                         bch2_btree_insert_key_cached(trans, flags, i);
 753                 else {
 754                         bch2_btree_key_cache_drop(trans, i->path);
 755                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 756                 }
 757         }
 758
 759         return 0;
 760 fatal_err:
 761         bch2_fatal_error(c);
 762 revert_fs_usage:
 763         if (trans->fs_usage_deltas)
 764                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 765         return ret;
 766 }
 767
 768 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 769 {
 770         while (--i >= trans->updates) {
 771                 if (same_leaf_as_prev(trans, i))
 772                         continue;
 773
 774                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 775         }
 776
 777         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 778         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 779 }
 780
 781 static inline int trans_lock_write(struct btree_trans *trans)
 782 {
 783         struct btree_insert_entry *i;
 784
 785         trans_for_each_update(trans, i) {
 786                 if (same_leaf_as_prev(trans, i))
 787                         continue;
 788
 789                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 790                         return trans_lock_write_fail(trans, i);
 791
 792                 if (!i->cached)
 793                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 794         }
 795
 796         return 0;
 797 }
 798
 799 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 800 {
 801         struct btree_insert_entry *i;
 802         struct btree_write_buffered_key *wb;
 803
 804         trans_for_each_update(trans, i)
 805                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 806
 807         trans_for_each_wb_update(trans, wb)
 808                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 809 }
 810
 811 #ifdef CONFIG_BCACHEFS_DEBUG
 812 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 813                                                    struct btree_insert_entry *i,
 814                                                    struct printbuf *err)
 815 {
 816         struct bch_fs *c = trans->c;
 817         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 818
 819         printbuf_reset(err);
 820         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 821                    trans->fn, (void *) i->ip_allocated);
 822         prt_newline(err);
 823         printbuf_indent_add(err, 2);
 824
 825         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 826         prt_newline(err);
 827
 828         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 829                           i->bkey_type, rw, err);
 830         bch2_print_string_as_lines(KERN_ERR, err->buf);
 831
 832         bch2_inconsistent_error(c);
 833         bch2_dump_trans_updates(trans);
 834         printbuf_exit(err);
 835
 836         return -EINVAL;
 837 }
 838 #endif
 839
 840 /*
 841  * Get journal reservation, take write locks, and attempt to do btree update(s):
 842  */
 843 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 844                                        struct btree_insert_entry **stopped_at,
 845                                        unsigned long trace_ip)
 846 {
 847         struct bch_fs *c = trans->c;
 848         struct btree_insert_entry *i;
 849         int ret, u64s_delta = 0;
 850
 851 #ifdef CONFIG_BCACHEFS_DEBUG
 852         struct printbuf buf = PRINTBUF;
 853
 854         trans_for_each_update(trans, i) {
 855                 int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 856
 857                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 858                                                i->bkey_type, rw, &buf)))
 859                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 860                 btree_insert_entry_checks(trans, i);
 861         }
 862         printbuf_exit(&buf);
 863 #endif
 864
 865         trans_for_each_update(trans, i) {
 866                 if (i->cached)
 867                         continue;
 868
 869                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 870                 u64s_delta -= i->old_btree_u64s;
 871
 872                 if (!same_leaf_as_next(trans, i)) {
 873                         if (u64s_delta <= 0) {
 874                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 875                                                         i->level, flags);
 876                                 if (unlikely(ret))
 877                                         return ret;
 878                         }
 879
 880                         u64s_delta = 0;
 881                 }
 882         }
 883
 884         ret = bch2_journal_preres_get(&c->journal,
 885                         &trans->journal_preres, trans->journal_preres_u64s,
 886                         (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 887         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 888                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 889         if (unlikely(ret))
 890                 return ret;
 891
 892         ret = trans_lock_write(trans);
 893         if (unlikely(ret))
 894                 return ret;
 895
 896         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 897
 898         if (!ret && unlikely(trans->journal_replay_not_finished))
 899                 bch2_drop_overwrites_from_journal(trans);
 900
 901         trans_for_each_update(trans, i)
 902                 if (!same_leaf_as_prev(trans, i))
 903                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 904                                                         insert_l(i)->b);
 905
 906         if (!ret && trans->journal_pin)
 907                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 908                                      trans->journal_pin, NULL);
 909
 910         /*
 911          * Drop journal reservation after dropping write locks, since dropping
 912          * the journal reservation may kick off a journal write:
 913          */
 914         bch2_journal_res_put(&c->journal, &trans->journal_res);
 915
 916         if (unlikely(ret))
 917                 return ret;
 918
 919         bch2_trans_downgrade(trans);
 920
 921         return 0;
 922 }
 923
 924 static int journal_reclaim_wait_done(struct bch_fs *c)
 925 {
 926         int ret = bch2_journal_error(&c->journal) ?:
 927                 !bch2_btree_key_cache_must_wait(c);
 928
 929         if (!ret)
 930                 journal_reclaim_kick(&c->journal);
 931         return ret;
 932 }
 933
 934 static noinline
 935 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 936                             struct btree_insert_entry *i,
 937                             int ret, unsigned long trace_ip)
 938 {
 939         struct bch_fs *c = trans->c;
 940
 941         switch (ret) {
 942         case -BCH_ERR_btree_insert_btree_node_full:
 943                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 944                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 945                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 946                 break;
 947         case -BCH_ERR_btree_insert_need_mark_replicas:
 948                 ret = drop_locks_do(trans,
 949                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 950                 break;
 951         case -BCH_ERR_journal_res_get_blocked:
 952                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 953                     !(flags & JOURNAL_WATERMARK_reserved)) {
 954                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 955                         break;
 956                 }
 957
 958                 ret = drop_locks_do(trans,
 959                         bch2_trans_journal_res_get(trans,
 960                                         (flags & JOURNAL_WATERMARK_MASK)|
 961                                         JOURNAL_RES_GET_CHECK));
 962                 break;
 963         case -BCH_ERR_btree_insert_need_journal_reclaim:
 964                 bch2_trans_unlock(trans);
 965
 966                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 967
 968                 wait_event_freezable(c->journal.reclaim_wait,
 969                                      (ret = journal_reclaim_wait_done(c)));
 970                 if (ret < 0)
 971                         break;
 972
 973                 ret = bch2_trans_relock(trans);
 974                 break;
 975         case -BCH_ERR_btree_insert_need_flush_buffer: {
 976                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 977
 978                 ret = 0;
 979
 980                 if (wb->state.nr > wb->size * 3 / 4) {
 981                         bch2_trans_unlock(trans);
 982                         mutex_lock(&wb->flush_lock);
 983
 984                         if (wb->state.nr > wb->size * 3 / 4) {
 985                                 bch2_trans_begin(trans);
 986                                 ret = __bch2_btree_write_buffer_flush(trans,
 987                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
 988                                 if (!ret) {
 989                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
 990                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
 991                                 }
 992                         } else {
 993                                 mutex_unlock(&wb->flush_lock);
 994                                 ret = bch2_trans_relock(trans);
 995                         }
 996                 }
 997                 break;
 998         }
 999         default:
1000                 BUG_ON(ret >= 0);
1001                 break;
1002         }
1003
1004         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1005
1006         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1007                                 !(flags & BTREE_INSERT_NOWAIT) &&
1008                                 (flags & BTREE_INSERT_NOFAIL), c,
1009                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1010
1011         return ret;
1012 }
1013
1014 static noinline int
1015 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1016 {
1017         struct bch_fs *c = trans->c;
1018         int ret;
1019
1020         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1021             test_bit(BCH_FS_STARTED, &c->flags))
1022                 return -BCH_ERR_erofs_trans_commit;
1023
1024         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1025         if (ret)
1026                 return ret;
1027
1028         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1029         return 0;
1030 }
1031
1032 /*
1033  * This is for updates done in the early part of fsck - btree_gc - before we've
1034  * gone RW. we only add the new key to the list of keys for journal replay to
1035  * do.
1036  */
1037 static noinline int
1038 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1039 {
1040         struct bch_fs *c = trans->c;
1041         struct btree_insert_entry *i;
1042         int ret = 0;
1043
1044         trans_for_each_update(trans, i) {
1045                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1046                 if (ret)
1047                         break;
1048         }
1049
1050         return ret;
1051 }
1052
1053 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1054 {
1055         struct bch_fs *c = trans->c;
1056         struct btree_insert_entry *i = NULL;
1057         struct btree_write_buffered_key *wb;
1058         unsigned u64s;
1059         int ret = 0;
1060
1061         if (!trans->nr_updates &&
1062             !trans->nr_wb_updates &&
1063             !trans->extra_journal_entries.nr)
1064                 goto out_reset;
1065
1066         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1067                 lockdep_assert_held(&c->gc_lock);
1068
1069         ret = bch2_trans_commit_run_triggers(trans);
1070         if (ret)
1071                 goto out_reset;
1072
1073         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1074                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1075                 goto out_reset;
1076         }
1077
1078         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1079             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1080                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1081                 if (ret)
1082                         goto out_reset;
1083         }
1084
1085         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1086             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1087                 bch2_trans_begin(trans);
1088                 bch2_trans_unlock(trans);
1089
1090                 ret = __bch2_btree_write_buffer_flush(trans,
1091                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1092                 if (!ret) {
1093                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1094                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1095                 }
1096                 goto out;
1097         }
1098
1099         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1100
1101         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1102
1103         trans->journal_u64s             = trans->extra_journal_entries.nr;
1104         trans->journal_preres_u64s      = 0;
1105
1106         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1107
1108         if (trans->journal_transaction_names)
1109                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1110
1111         trans_for_each_update(trans, i) {
1112                 EBUG_ON(!i->path->should_be_locked);
1113
1114                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1115                 if (unlikely(ret))
1116                         goto out;
1117
1118                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1119
1120                 if (i->key_cache_already_flushed)
1121                         continue;
1122
1123                 /* we're going to journal the key being updated: */
1124                 u64s = jset_u64s(i->k->k.u64s);
1125                 if (i->cached &&
1126                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1127                         trans->journal_preres_u64s += u64s;
1128
1129                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1130                         continue;
1131
1132                 trans->journal_u64s += u64s;
1133
1134                 /* and we're also going to log the overwrite: */
1135                 if (trans->journal_transaction_names)
1136                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1137         }
1138
1139         trans_for_each_wb_update(trans, wb)
1140                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1141
1142         if (trans->extra_journal_res) {
1143                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1144                                 trans->extra_journal_res,
1145                                 (flags & BTREE_INSERT_NOFAIL)
1146                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1147                 if (ret)
1148                         goto err;
1149         }
1150 retry:
1151         bch2_trans_verify_not_in_restart(trans);
1152         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1153
1154         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1155
1156         /* make sure we didn't drop or screw up locks: */
1157         bch2_trans_verify_locks(trans);
1158
1159         if (ret)
1160                 goto err;
1161
1162         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1163 out:
1164         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1165
1166         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1167                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1168 out_reset:
1169         bch2_trans_reset_updates(trans);
1170
1171         return ret;
1172 err:
1173         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1174         if (ret)
1175                 goto out;
1176
1177         goto retry;
1178 }
1179
1180 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1181                                           enum btree_id id,
1182                                           struct bpos pos)
1183 {
1184         struct bch_fs *c = trans->c;
1185         struct btree_iter iter;
1186         struct bkey_s_c k;
1187         int ret;
1188
1189         bch2_trans_iter_init(trans, &iter, id, pos,
1190                              BTREE_ITER_NOT_EXTENTS|
1191                              BTREE_ITER_ALL_SNAPSHOTS);
1192         while (1) {
1193                 k = bch2_btree_iter_prev(&iter);
1194                 ret = bkey_err(k);
1195                 if (ret)
1196                         break;
1197
1198                 if (!k.k)
1199                         break;
1200
1201                 if (!bkey_eq(pos, k.k->p))
1202                         break;
1203
1204                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1205                         ret = 1;
1206                         break;
1207                 }
1208         }
1209         bch2_trans_iter_exit(trans, &iter);
1210
1211         return ret;
1212 }
1213
1214 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1215                                           enum btree_id id,
1216                                           struct bpos pos)
1217 {
1218         if (!btree_type_has_snapshots(id) ||
1219             !snapshot_t(trans->c, pos.snapshot)->children[0])
1220                 return 0;
1221
1222         return __check_pos_snapshot_overwritten(trans, id, pos);
1223 }
1224
1225 static noinline int extent_front_merge(struct btree_trans *trans,
1226                                        struct btree_iter *iter,
1227                                        struct bkey_s_c k,
1228                                        struct bkey_i **insert,
1229                                        enum btree_update_flags flags)
1230 {
1231         struct bch_fs *c = trans->c;
1232         struct bkey_i *update;
1233         int ret;
1234
1235         update = bch2_bkey_make_mut_noupdate(trans, k);
1236         ret = PTR_ERR_OR_ZERO(update);
1237         if (ret)
1238                 return ret;
1239
1240         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1241                 return 0;
1242
1243         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1244                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1245         if (ret < 0)
1246                 return ret;
1247         if (ret)
1248                 return 0;
1249
1250         ret = bch2_btree_delete_at(trans, iter, flags);
1251         if (ret)
1252                 return ret;
1253
1254         *insert = update;
1255         return 0;
1256 }
1257
1258 static noinline int extent_back_merge(struct btree_trans *trans,
1259                                       struct btree_iter *iter,
1260                                       struct bkey_i *insert,
1261                                       struct bkey_s_c k)
1262 {
1263         struct bch_fs *c = trans->c;
1264         int ret;
1265
1266         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1267                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1268         if (ret < 0)
1269                 return ret;
1270         if (ret)
1271                 return 0;
1272
1273         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1274         return 0;
1275 }
1276
1277 /*
1278  * When deleting, check if we need to emit a whiteout (because we're overwriting
1279  * something in an ancestor snapshot)
1280  */
1281 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1282                                       enum btree_id btree_id, struct bpos pos)
1283 {
1284         struct btree_iter iter;
1285         struct bkey_s_c k;
1286         u32 snapshot = pos.snapshot;
1287         int ret;
1288
1289         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1290                 return 0;
1291
1292         pos.snapshot++;
1293
1294         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1295                            BTREE_ITER_ALL_SNAPSHOTS|
1296                            BTREE_ITER_NOPRESERVE, k, ret) {
1297                 if (!bkey_eq(k.k->p, pos))
1298                         break;
1299
1300                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1301                                               k.k->p.snapshot)) {
1302                         ret = !bkey_whiteout(k.k);
1303                         break;
1304                 }
1305         }
1306         bch2_trans_iter_exit(trans, &iter);
1307
1308         return ret;
1309 }
1310
1311 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
1312                                    enum btree_id id,
1313                                    struct bpos old_pos,
1314                                    struct bpos new_pos)
1315 {
1316         struct bch_fs *c = trans->c;
1317         struct btree_iter old_iter, new_iter;
1318         struct bkey_s_c old_k, new_k;
1319         snapshot_id_list s;
1320         struct bkey_i *update;
1321         int ret;
1322
1323         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
1324                 return 0;
1325
1326         darray_init(&s);
1327
1328         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
1329                              BTREE_ITER_NOT_EXTENTS|
1330                              BTREE_ITER_ALL_SNAPSHOTS);
1331         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
1332                !(ret = bkey_err(old_k)) &&
1333                bkey_eq(old_pos, old_k.k->p)) {
1334                 struct bpos whiteout_pos =
1335                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
1336
1337                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
1338                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
1339                         continue;
1340
1341                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
1342                                            BTREE_ITER_NOT_EXTENTS|
1343                                            BTREE_ITER_INTENT);
1344                 ret = bkey_err(new_k);
1345                 if (ret)
1346                         break;
1347
1348                 if (new_k.k->type == KEY_TYPE_deleted) {
1349                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
1350                         ret = PTR_ERR_OR_ZERO(update);
1351                         if (ret)
1352                                 break;
1353
1354                         bkey_init(&update->k);
1355                         update->k.p             = whiteout_pos;
1356                         update->k.type          = KEY_TYPE_whiteout;
1357
1358                         ret = bch2_trans_update(trans, &new_iter, update,
1359                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1360                 }
1361                 bch2_trans_iter_exit(trans, &new_iter);
1362
1363                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
1364                 if (ret)
1365                         break;
1366         }
1367         bch2_trans_iter_exit(trans, &old_iter);
1368         darray_exit(&s);
1369
1370         return ret;
1371 }
1372
1373 int bch2_trans_update_extent(struct btree_trans *trans,
1374                              struct btree_iter *orig_iter,
1375                              struct bkey_i *insert,
1376                              enum btree_update_flags flags)
1377 {
1378         struct btree_iter iter;
1379         struct bpos start = bkey_start_pos(&insert->k);
1380         struct bkey_i *update;
1381         struct bkey_s_c k;
1382         enum btree_id btree_id = orig_iter->btree_id;
1383         int ret = 0, compressed_sectors;
1384
1385         bch2_trans_iter_init(trans, &iter, btree_id, start,
1386                              BTREE_ITER_INTENT|
1387                              BTREE_ITER_WITH_UPDATES|
1388                              BTREE_ITER_NOT_EXTENTS);
1389         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1390         if ((ret = bkey_err(k)))
1391                 goto err;
1392         if (!k.k)
1393                 goto out;
1394
1395         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1396                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1397                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1398                         if (ret)
1399                                 goto err;
1400                 }
1401
1402                 goto next;
1403         }
1404
1405         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1406                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1407                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1408
1409                 /*
1410                  * If we're going to be splitting a compressed extent, note it
1411                  * so that __bch2_trans_commit() can increase our disk
1412                  * reservation:
1413                  */
1414                 if (((front_split && back_split) ||
1415                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1416                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1417                         trans->extra_journal_res += compressed_sectors;
1418
1419                 if (front_split) {
1420                         update = bch2_bkey_make_mut_noupdate(trans, k);
1421                         if ((ret = PTR_ERR_OR_ZERO(update)))
1422                                 goto err;
1423
1424                         bch2_cut_back(start, update);
1425
1426                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1427                                                 k.k->p, update->k.p) ?:
1428                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1429                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1430                         if (ret)
1431                                 goto err;
1432                 }
1433
1434                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1435                     (front_split || back_split)) {
1436                         update = bch2_bkey_make_mut_noupdate(trans, k);
1437                         if ((ret = PTR_ERR_OR_ZERO(update)))
1438                                 goto err;
1439
1440                         bch2_cut_front(start, update);
1441                         bch2_cut_back(insert->k.p, update);
1442
1443                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1444                                                 k.k->p, update->k.p) ?:
1445                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1446                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1447                         if (ret)
1448                                 goto err;
1449                 }
1450
1451                 if (bkey_le(k.k->p, insert->k.p)) {
1452                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1453                         if ((ret = PTR_ERR_OR_ZERO(update)))
1454                                 goto err;
1455
1456                         bkey_init(&update->k);
1457                         update->k.p = k.k->p;
1458                         update->k.p.snapshot = insert->k.p.snapshot;
1459
1460                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1461                                 update->k.type = KEY_TYPE_whiteout;
1462                         } else if (btree_type_has_snapshots(btree_id)) {
1463                                 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1464                                 if (ret < 0)
1465                                         goto err;
1466                                 if (ret)
1467                                         update->k.type = KEY_TYPE_whiteout;
1468                         }
1469
1470                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1471                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1472                         if (ret)
1473                                 goto err;
1474                 }
1475
1476                 if (back_split) {
1477                         update = bch2_bkey_make_mut_noupdate(trans, k);
1478                         if ((ret = PTR_ERR_OR_ZERO(update)))
1479                                 goto err;
1480
1481                         bch2_cut_front(insert->k.p, update);
1482
1483                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1484                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1485                                                   flags);
1486                         if (ret)
1487                                 goto err;
1488                         goto out;
1489                 }
1490 next:
1491                 bch2_btree_iter_advance(&iter);
1492                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1493                 if ((ret = bkey_err(k)))
1494                         goto err;
1495                 if (!k.k)
1496                         goto out;
1497         }
1498
1499         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1500                 ret = extent_back_merge(trans, &iter, insert, k);
1501                 if (ret)
1502                         goto err;
1503         }
1504 out:
1505         if (!bkey_deleted(&insert->k)) {
1506                 /*
1507                  * Rewinding iterators is expensive: get a new one and the one
1508                  * that points to the start of insert will be cloned from:
1509                  */
1510                 bch2_trans_iter_exit(trans, &iter);
1511                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1512                                      BTREE_ITER_NOT_EXTENTS|
1513                                      BTREE_ITER_INTENT);
1514                 ret   = bch2_btree_iter_traverse(&iter) ?:
1515                         bch2_trans_update(trans, &iter, insert, flags);
1516         }
1517 err:
1518         bch2_trans_iter_exit(trans, &iter);
1519
1520         return ret;
1521 }
1522
1523 static int __must_check
1524 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1525                                 struct bkey_i *k, enum btree_update_flags flags,
1526                                 unsigned long ip);
1527
1528 static noinline int flush_new_cached_update(struct btree_trans *trans,
1529                                             struct btree_path *path,
1530                                             struct btree_insert_entry *i,
1531                                             enum btree_update_flags flags,
1532                                             unsigned long ip)
1533 {
1534         struct btree_path *btree_path;
1535         struct bkey k;
1536         int ret;
1537
1538         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1539                                    BTREE_ITER_INTENT, _THIS_IP_);
1540         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1541         if (ret)
1542                 goto out;
1543
1544         /*
1545          * The old key in the insert entry might actually refer to an existing
1546          * key in the btree that has been deleted from cache and not yet
1547          * flushed. Check for this and skip the flush so we don't run triggers
1548          * against a stale key.
1549          */
1550         bch2_btree_path_peek_slot_exact(btree_path, &k);
1551         if (!bkey_deleted(&k))
1552                 goto out;
1553
1554         i->key_cache_already_flushed = true;
1555         i->flags |= BTREE_TRIGGER_NORUN;
1556
1557         btree_path_set_should_be_locked(btree_path);
1558         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1559 out:
1560         bch2_path_put(trans, btree_path, true);
1561         return ret;
1562 }
1563
1564 static int __must_check
1565 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1566                                 struct bkey_i *k, enum btree_update_flags flags,
1567                                 unsigned long ip)
1568 {
1569         struct bch_fs *c = trans->c;
1570         struct btree_insert_entry *i, n;
1571         int cmp;
1572
1573         EBUG_ON(!path->should_be_locked);
1574         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1575         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1576
1577         n = (struct btree_insert_entry) {
1578                 .flags          = flags,
1579                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1580                 .btree_id       = path->btree_id,
1581                 .level          = path->level,
1582                 .cached         = path->cached,
1583                 .path           = path,
1584                 .k              = k,
1585                 .ip_allocated   = ip,
1586         };
1587
1588 #ifdef CONFIG_BCACHEFS_DEBUG
1589         trans_for_each_update(trans, i)
1590                 BUG_ON(i != trans->updates &&
1591                        btree_insert_entry_cmp(i - 1, i) >= 0);
1592 #endif
1593
1594         /*
1595          * Pending updates are kept sorted: first, find position of new update,
1596          * then delete/trim any updates the new update overwrites:
1597          */
1598         trans_for_each_update(trans, i) {
1599                 cmp = btree_insert_entry_cmp(&n, i);
1600                 if (cmp <= 0)
1601                         break;
1602         }
1603
1604         if (!cmp && i < trans->updates + trans->nr_updates) {
1605                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1606
1607                 bch2_path_put(trans, i->path, true);
1608                 i->flags        = n.flags;
1609                 i->cached       = n.cached;
1610                 i->k            = n.k;
1611                 i->path         = n.path;
1612                 i->ip_allocated = n.ip_allocated;
1613         } else {
1614                 array_insert_item(trans->updates, trans->nr_updates,
1615                                   i - trans->updates, n);
1616
1617                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1618                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1619
1620                 if (unlikely(trans->journal_replay_not_finished)) {
1621                         struct bkey_i *j_k =
1622                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1623
1624                         if (j_k) {
1625                                 i->old_k = j_k->k;
1626                                 i->old_v = &j_k->v;
1627                         }
1628                 }
1629         }
1630
1631         __btree_path_get(i->path, true);
1632
1633         /*
1634          * If a key is present in the key cache, it must also exist in the
1635          * btree - this is necessary for cache coherency. When iterating over
1636          * a btree that's cached in the key cache, the btree iter code checks
1637          * the key cache - but the key has to exist in the btree for that to
1638          * work:
1639          */
1640         if (path->cached && bkey_deleted(&i->old_k))
1641                 return flush_new_cached_update(trans, path, i, flags, ip);
1642
1643         return 0;
1644 }
1645
1646 static inline int __must_check
1647 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1648                           struct bkey_i *k, enum btree_update_flags flags)
1649 {
1650         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1651 }
1652
1653 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1654                                    struct bkey_i *k, enum btree_update_flags flags)
1655 {
1656         struct btree_path *path = iter->update_path ?: iter->path;
1657         struct bkey_cached *ck;
1658         int ret;
1659
1660         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1661                 return bch2_trans_update_extent(trans, iter, k, flags);
1662
1663         if (bkey_deleted(&k->k) &&
1664             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1665             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1666                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1667                 if (unlikely(ret < 0))
1668                         return ret;
1669
1670                 if (ret)
1671                         k->k.type = KEY_TYPE_whiteout;
1672         }
1673
1674         /*
1675          * Ensure that updates to cached btrees go to the key cache:
1676          */
1677         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1678             !path->cached &&
1679             !path->level &&
1680             btree_id_cached(trans->c, path->btree_id)) {
1681                 if (!iter->key_cache_path ||
1682                     !iter->key_cache_path->should_be_locked ||
1683                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1684                         if (!iter->key_cache_path)
1685                                 iter->key_cache_path =
1686                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1687                                                       BTREE_ITER_INTENT|
1688                                                       BTREE_ITER_CACHED, _THIS_IP_);
1689
1690                         iter->key_cache_path =
1691                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1692                                                         iter->flags & BTREE_ITER_INTENT,
1693                                                         _THIS_IP_);
1694
1695                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1696                                                        BTREE_ITER_CACHED);
1697                         if (unlikely(ret))
1698                                 return ret;
1699
1700                         ck = (void *) iter->key_cache_path->l[0].b;
1701
1702                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1703                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1704                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1705                         }
1706
1707                         btree_path_set_should_be_locked(iter->key_cache_path);
1708                 }
1709
1710                 path = iter->key_cache_path;
1711         }
1712
1713         return bch2_trans_update_by_path(trans, path, k, flags);
1714 }
1715
1716 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1717                                             enum btree_id btree,
1718                                             struct bkey_i *k,
1719                                             bool head)
1720 {
1721         int ret, pos;
1722
1723         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1724         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1725
1726         if (!trans->wb_updates ||
1727             trans->nr_wb_updates == trans->wb_updates_size) {
1728                 struct btree_write_buffered_key *u;
1729
1730                 if (trans->nr_wb_updates == trans->wb_updates_size) {
1731                         struct btree_transaction_stats *s = btree_trans_stats(trans);
1732
1733                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1734                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1735                         if (s)
1736                                 s->wb_updates_size = trans->wb_updates_size;
1737                 }
1738
1739                 u = bch2_trans_kmalloc_nomemzero(trans,
1740                                         trans->wb_updates_size *
1741                                         sizeof(struct btree_write_buffered_key));
1742                 ret = PTR_ERR_OR_ZERO(u);
1743                 if (ret)
1744                         return ret;
1745
1746                 if (trans->nr_wb_updates)
1747                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1748                                sizeof(struct btree_write_buffered_key));
1749                 trans->wb_updates = u;
1750         }
1751
1752         if (head) {
1753                 memmove(&trans->wb_updates[1],
1754                         &trans->wb_updates[0],
1755                         sizeof(trans->wb_updates[0]) * trans->nr_wb_updates);
1756                 pos = 0;
1757         } else {
1758                 pos = trans->nr_wb_updates;
1759         }
1760
1761         trans->wb_updates[pos] = (struct btree_write_buffered_key) { .btree = btree, };
1762         bkey_copy(&trans->wb_updates[pos].k, k);
1763         trans->nr_wb_updates++;
1764         return 0;
1765 }
1766
1767 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1768                              enum btree_id btree, struct bpos end)
1769 {
1770         struct bkey_s_c k;
1771         int ret = 0;
1772
1773         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1774         k = bch2_btree_iter_prev(iter);
1775         ret = bkey_err(k);
1776         if (ret)
1777                 goto err;
1778
1779         bch2_btree_iter_advance(iter);
1780         k = bch2_btree_iter_peek_slot(iter);
1781         ret = bkey_err(k);
1782         if (ret)
1783                 goto err;
1784
1785         BUG_ON(k.k->type != KEY_TYPE_deleted);
1786
1787         if (bkey_gt(k.k->p, end)) {
1788                 ret = -BCH_ERR_ENOSPC_btree_slot;
1789                 goto err;
1790         }
1791
1792         return 0;
1793 err:
1794         bch2_trans_iter_exit(trans, iter);
1795         return ret;
1796 }
1797
1798 void bch2_trans_commit_hook(struct btree_trans *trans,
1799                             struct btree_trans_commit_hook *h)
1800 {
1801         h->next = trans->hooks;
1802         trans->hooks = h;
1803 }
1804
1805 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1806                                 enum btree_id btree, struct bkey_i *k,
1807                                 enum btree_update_flags flags)
1808 {
1809         struct btree_iter iter;
1810         int ret;
1811
1812         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1813                              BTREE_ITER_NOT_EXTENTS|
1814                              BTREE_ITER_INTENT);
1815         ret   = bch2_btree_iter_traverse(&iter) ?:
1816                 bch2_trans_update(trans, &iter, k, flags);
1817         bch2_trans_iter_exit(trans, &iter);
1818         return ret;
1819 }
1820
1821 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1822                         struct bkey_i *k, enum btree_update_flags flags)
1823 {
1824         struct btree_iter iter;
1825         int ret;
1826
1827         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1828                              BTREE_ITER_CACHED|
1829                              BTREE_ITER_INTENT);
1830         ret   = bch2_btree_iter_traverse(&iter) ?:
1831                 bch2_trans_update(trans, &iter, k, flags);
1832         bch2_trans_iter_exit(trans, &iter);
1833         return ret;
1834 }
1835
1836 /**
1837  * bch2_btree_insert - insert keys into the extent btree
1838  * @c:                  pointer to struct bch_fs
1839  * @id:                 btree to insert into
1840  * @insert_keys:        list of keys to insert
1841  * @hook:               insert callback
1842  */
1843 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1844                       struct bkey_i *k,
1845                       struct disk_reservation *disk_res,
1846                       u64 *journal_seq, int flags)
1847 {
1848         return bch2_trans_do(c, disk_res, journal_seq, flags,
1849                              __bch2_btree_insert(&trans, id, k, 0));
1850 }
1851
1852 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1853                                 unsigned len, unsigned update_flags)
1854 {
1855         struct bkey_i *k;
1856
1857         k = bch2_trans_kmalloc(trans, sizeof(*k));
1858         if (IS_ERR(k))
1859                 return PTR_ERR(k);
1860
1861         bkey_init(&k->k);
1862         k->k.p = iter->pos;
1863         bch2_key_resize(&k->k, len);
1864         return bch2_trans_update(trans, iter, k, update_flags);
1865 }
1866
1867 int bch2_btree_delete_at(struct btree_trans *trans,
1868                          struct btree_iter *iter, unsigned update_flags)
1869 {
1870         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1871 }
1872
1873 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1874                                   enum btree_id btree, struct bpos pos)
1875 {
1876         struct bkey_i *k;
1877
1878         k = bch2_trans_kmalloc(trans, sizeof(*k));
1879         if (IS_ERR(k))
1880                 return PTR_ERR(k);
1881
1882         bkey_init(&k->k);
1883         k->k.p = pos;
1884         return bch2_trans_update_buffered(trans, btree, k, false);
1885 }
1886
1887 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1888                                   struct bpos start, struct bpos end,
1889                                   unsigned update_flags,
1890                                   u64 *journal_seq)
1891 {
1892         u32 restart_count = trans->restart_count;
1893         struct btree_iter iter;
1894         struct bkey_s_c k;
1895         int ret = 0;
1896
1897         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1898         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1899                 struct disk_reservation disk_res =
1900                         bch2_disk_reservation_init(trans->c, 0);
1901                 struct bkey_i delete;
1902
1903                 ret = bkey_err(k);
1904                 if (ret)
1905                         goto err;
1906
1907                 bkey_init(&delete.k);
1908
1909                 /*
1910                  * This could probably be more efficient for extents:
1911                  */
1912
1913                 /*
1914                  * For extents, iter.pos won't necessarily be the same as
1915                  * bkey_start_pos(k.k) (for non extents they always will be the
1916                  * same). It's important that we delete starting from iter.pos
1917                  * because the range we want to delete could start in the middle
1918                  * of k.
1919                  *
1920                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1921                  * bkey_start_pos(k.k)).
1922                  */
1923                 delete.k.p = iter.pos;
1924
1925                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1926                         bch2_key_resize(&delete.k,
1927                                         bpos_min(end, k.k->p).offset -
1928                                         iter.pos.offset);
1929
1930                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1931                         bch2_trans_commit(trans, &disk_res, journal_seq,
1932                                           BTREE_INSERT_NOFAIL);
1933                 bch2_disk_reservation_put(trans->c, &disk_res);
1934 err:
1935                 /*
1936                  * the bch2_trans_begin() call is in a weird place because we
1937                  * need to call it after every transaction commit, to avoid path
1938                  * overflow, but don't want to call it if the delete operation
1939                  * is a no-op and we have no work to do:
1940                  */
1941                 bch2_trans_begin(trans);
1942
1943                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1944                         ret = 0;
1945                 if (ret)
1946                         break;
1947         }
1948         bch2_trans_iter_exit(trans, &iter);
1949
1950         if (!ret && trans_was_restarted(trans, restart_count))
1951                 ret = -BCH_ERR_transaction_restart_nested;
1952         return ret;
1953 }
1954
1955 /*
1956  * bch_btree_delete_range - delete everything within a given range
1957  *
1958  * Range is a half open interval - [start, end)
1959  */
1960 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1961                             struct bpos start, struct bpos end,
1962                             unsigned update_flags,
1963                             u64 *journal_seq)
1964 {
1965         int ret = bch2_trans_run(c,
1966                         bch2_btree_delete_range_trans(&trans, id, start, end,
1967                                                       update_flags, journal_seq));
1968         if (ret == -BCH_ERR_transaction_restart_nested)
1969                 ret = 0;
1970         return ret;
1971 }
1972
1973 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1974 {
1975         struct printbuf buf = PRINTBUF;
1976         struct jset_entry_log *l;
1977         unsigned u64s;
1978         int ret;
1979
1980         prt_vprintf(&buf, fmt, args);
1981         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
1982         if (ret)
1983                 goto err;
1984
1985         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
1986
1987         ret = darray_make_room(entries, jset_u64s(u64s));
1988         if (ret)
1989                 goto err;
1990
1991         l = (void *) &darray_top(*entries);
1992         l->entry.u64s           = cpu_to_le16(u64s);
1993         l->entry.btree_id       = 0;
1994         l->entry.level          = 1;
1995         l->entry.type           = BCH_JSET_ENTRY_log;
1996         l->entry.pad[0]         = 0;
1997         l->entry.pad[1]         = 0;
1998         l->entry.pad[2]         = 0;
1999         memcpy(l->d, buf.buf, buf.pos);
2000         while (buf.pos & 7)
2001                 l->d[buf.pos++] = '\0';
2002
2003         entries->nr += jset_u64s(u64s);
2004 err:
2005         printbuf_exit(&buf);
2006         return ret;
2007 }
2008
2009 static int
2010 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
2011                   va_list args)
2012 {
2013         int ret;
2014
2015         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
2016                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
2017         } else {
2018                 ret = bch2_trans_do(c, NULL, NULL,
2019                         BTREE_INSERT_LAZY_RW|commit_flags,
2020                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
2021         }
2022
2023         return ret;
2024 }
2025
2026 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
2027 {
2028         va_list args;
2029         int ret;
2030
2031         va_start(args, fmt);
2032         ret = __bch2_fs_log_msg(c, 0, fmt, args);
2033         va_end(args);
2034         return ret;
2035 }
2036
2037 /*
2038  * Use for logging messages during recovery to enable reserved space and avoid
2039  * blocking.
2040  */
2041 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2042 {
2043         va_list args;
2044         int ret;
2045
2046         va_start(args, fmt);
2047         ret = __bch2_fs_log_msg(c, JOURNAL_WATERMARK_reserved, fmt, args);
2048         va_end(args);
2049         return ret;
2050 }