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             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 423                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 424                                 old, bkey_i_to_s_c(new),
 425                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 426         } else {
 427                 struct bkey             _deleted = KEY(0, 0, 0);
 428                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 429
 430                 _deleted.p = i->path->pos;
 431
 432                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 433                                 deleted, bkey_i_to_s_c(new),
 434                                 BTREE_TRIGGER_INSERT|flags) ?:
 435                         bch2_mark_key(trans, i->btree_id, i->level,
 436                                 old, deleted,
 437                                 BTREE_TRIGGER_OVERWRITE|flags);
 438         }
 439
 440         return ret;
 441 }
 442
 443 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 444                                  bool overwrite)
 445 {
 446         /*
 447          * Transactional triggers create new btree_insert_entries, so we can't
 448          * pass them a pointer to a btree_insert_entry, that memory is going to
 449          * move:
 450          */
 451         struct bkey old_k = i->old_k;
 452         struct bkey_s_c old = { &old_k, i->old_v };
 453
 454         verify_update_old_key(trans, i);
 455
 456         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 457             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 458                 return 0;
 459
 460         if (!i->insert_trigger_run &&
 461             !i->overwrite_trigger_run &&
 462             bch2_bkey_ops[old.k->type].trans_trigger ==
 463             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 464             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 465                 i->overwrite_trigger_run = true;
 466                 i->insert_trigger_run = true;
 467                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 468                                            BTREE_TRIGGER_INSERT|
 469                                            BTREE_TRIGGER_OVERWRITE|
 470                                            i->flags) ?: 1;
 471         } else if (overwrite && !i->overwrite_trigger_run) {
 472                 i->overwrite_trigger_run = true;
 473                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 474         } else if (!overwrite && !i->insert_trigger_run) {
 475                 i->insert_trigger_run = true;
 476                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 477         } else {
 478                 return 0;
 479         }
 480 }
 481
 482 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 483                               struct btree_insert_entry *btree_id_start)
 484 {
 485         struct btree_insert_entry *i;
 486         bool trans_trigger_run;
 487         int ret, overwrite;
 488
 489         for (overwrite = 1; overwrite >= 0; --overwrite) {
 490
 491                 /*
 492                  * Running triggers will append more updates to the list of updates as
 493                  * we're walking it:
 494                  */
 495                 do {
 496                         trans_trigger_run = false;
 497
 498                         for (i = btree_id_start;
 499                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 500                              i++) {
 501                                 if (i->btree_id != btree_id)
 502                                         continue;
 503
 504                                 ret = run_one_trans_trigger(trans, i, overwrite);
 505                                 if (ret < 0)
 506                                         return ret;
 507                                 if (ret)
 508                                         trans_trigger_run = true;
 509                         }
 510                 } while (trans_trigger_run);
 511         }
 512
 513         return 0;
 514 }
 515
 516 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 517 {
 518         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 519         unsigned btree_id = 0;
 520         int ret = 0;
 521
 522         /*
 523          *
 524          * For a given btree, this algorithm runs insert triggers before
 525          * overwrite triggers: this is so that when extents are being moved
 526          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 527          * they are re-added.
 528          */
 529         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 530                 if (btree_id == BTREE_ID_alloc)
 531                         continue;
 532
 533                 while (btree_id_start < trans->updates + trans->nr_updates &&
 534                        btree_id_start->btree_id < btree_id)
 535                         btree_id_start++;
 536
 537                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 538                 if (ret)
 539                         return ret;
 540         }
 541
 542         trans_for_each_update(trans, i) {
 543                 if (i->btree_id > BTREE_ID_alloc)
 544                         break;
 545                 if (i->btree_id == BTREE_ID_alloc) {
 546                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 547                         if (ret)
 548                                 return ret;
 549                         break;
 550                 }
 551         }
 552
 553 #ifdef CONFIG_BCACHEFS_DEBUG
 554         trans_for_each_update(trans, i)
 555                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 556                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 557                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 558 #endif
 559         return 0;
 560 }
 561
 562 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 563 {
 564         struct bch_fs *c = trans->c;
 565         struct btree_insert_entry *i;
 566         int ret = 0;
 567
 568         trans_for_each_update(trans, i) {
 569                 /*
 570                  * XXX: synchronization of cached update triggers with gc
 571                  * XXX: synchronization of interior node updates with gc
 572                  */
 573                 BUG_ON(i->cached || i->level);
 574
 575                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 576                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 577                         if (ret)
 578                                 break;
 579                 }
 580         }
 581
 582         return ret;
 583 }
 584
 585 static inline int
 586 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 587                                struct btree_insert_entry **stopped_at,
 588                                unsigned long trace_ip)
 589 {
 590         struct bch_fs *c = trans->c;
 591         struct btree_insert_entry *i;
 592         struct btree_write_buffered_key *wb;
 593         struct btree_trans_commit_hook *h;
 594         unsigned u64s = 0;
 595         bool marking = false;
 596         int ret;
 597
 598         if (race_fault()) {
 599                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 600                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 601         }
 602
 603         /*
 604          * Check if the insert will fit in the leaf node with the write lock
 605          * held, otherwise another thread could write the node changing the
 606          * amount of space available:
 607          */
 608
 609         prefetch(&trans->c->journal.flags);
 610
 611         trans_for_each_update(trans, i) {
 612                 /* Multiple inserts might go to same leaf: */
 613                 if (!same_leaf_as_prev(trans, i))
 614                         u64s = 0;
 615
 616                 u64s += i->k->k.u64s;
 617                 ret = !i->cached
 618                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 619                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 620                 if (ret) {
 621                         *stopped_at = i;
 622                         return ret;
 623                 }
 624
 625                 if (btree_node_type_needs_gc(i->bkey_type))
 626                         marking = true;
 627         }
 628
 629         if (trans->nr_wb_updates &&
 630             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 631                 return -BCH_ERR_btree_insert_need_flush_buffer;
 632
 633         /*
 634          * Don't get journal reservation until after we know insert will
 635          * succeed:
 636          */
 637         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 638                 ret = bch2_trans_journal_res_get(trans,
 639                                 (flags & JOURNAL_WATERMARK_MASK)|
 640                                 JOURNAL_RES_GET_NONBLOCK);
 641                 if (ret)
 642                         return ret;
 643
 644                 if (unlikely(trans->journal_transaction_names))
 645                         journal_transaction_name(trans);
 646         } else {
 647                 trans->journal_res.seq = c->journal.replay_journal_seq;
 648         }
 649
 650         /*
 651          * Not allowed to fail after we've gotten our journal reservation - we
 652          * have to use it:
 653          */
 654
 655         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 656             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 657                 if (bch2_journal_seq_verify)
 658                         trans_for_each_update(trans, i)
 659                                 i->k->k.version.lo = trans->journal_res.seq;
 660                 else if (bch2_inject_invalid_keys)
 661                         trans_for_each_update(trans, i)
 662                                 i->k->k.version = MAX_VERSION;
 663         }
 664
 665         if (trans->fs_usage_deltas &&
 666             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 667                 return -BCH_ERR_btree_insert_need_mark_replicas;
 668
 669         if (trans->nr_wb_updates) {
 670                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 671
 672                 ret = bch2_btree_insert_keys_write_buffer(trans);
 673                 if (ret)
 674                         goto revert_fs_usage;
 675         }
 676
 677         h = trans->hooks;
 678         while (h) {
 679                 ret = h->fn(trans, h);
 680                 if (ret)
 681                         goto revert_fs_usage;
 682                 h = h->next;
 683         }
 684
 685         trans_for_each_update(trans, i)
 686                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 687                         ret = run_one_mem_trigger(trans, i, i->flags);
 688                         if (ret)
 689                                 goto fatal_err;
 690                 }
 691
 692         if (unlikely(c->gc_pos.phase)) {
 693                 ret = bch2_trans_commit_run_gc_triggers(trans);
 694                 if  (ret)
 695                         goto fatal_err;
 696         }
 697
 698         if (unlikely(trans->extra_journal_entries.nr)) {
 699                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 700                                   trans->extra_journal_entries.data,
 701                                   trans->extra_journal_entries.nr);
 702
 703                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 704                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 705         }
 706
 707         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 708                 struct journal *j = &c->journal;
 709                 struct jset_entry *entry;
 710
 711                 trans_for_each_update(trans, i) {
 712                         if (i->key_cache_already_flushed)
 713                                 continue;
 714
 715                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 716                                 continue;
 717
 718                         verify_update_old_key(trans, i);
 719
 720                         if (trans->journal_transaction_names) {
 721                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 722                                                        BCH_JSET_ENTRY_overwrite,
 723                                                        i->btree_id, i->level,
 724                                                        i->old_k.u64s);
 725                                 bkey_reassemble(&entry->start[0],
 726                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 727                         }
 728
 729                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 730                                                BCH_JSET_ENTRY_btree_keys,
 731                                                i->btree_id, i->level,
 732                                                i->k->k.u64s);
 733                         bkey_copy(&entry->start[0], i->k);
 734                 }
 735
 736                 trans_for_each_wb_update(trans, wb) {
 737                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 738                                                BCH_JSET_ENTRY_btree_keys,
 739                                                wb->btree, 0,
 740                                                wb->k.k.u64s);
 741                         bkey_copy(&entry->start[0], &wb->k);
 742                 }
 743
 744                 if (trans->journal_seq)
 745                         *trans->journal_seq = trans->journal_res.seq;
 746         }
 747
 748         trans_for_each_update(trans, i) {
 749                 i->k->k.needs_whiteout = false;
 750
 751                 if (!i->cached)
 752                         btree_insert_key_leaf(trans, i);
 753                 else if (!i->key_cache_already_flushed)
 754                         bch2_btree_insert_key_cached(trans, flags, i);
 755                 else {
 756                         bch2_btree_key_cache_drop(trans, i->path);
 757                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 758                 }
 759         }
 760
 761         return 0;
 762 fatal_err:
 763         bch2_fatal_error(c);
 764 revert_fs_usage:
 765         if (trans->fs_usage_deltas)
 766                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 767         return ret;
 768 }
 769
 770 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 771 {
 772         while (--i >= trans->updates) {
 773                 if (same_leaf_as_prev(trans, i))
 774                         continue;
 775
 776                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 777         }
 778
 779         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 780         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 781 }
 782
 783 static inline int trans_lock_write(struct btree_trans *trans)
 784 {
 785         struct btree_insert_entry *i;
 786
 787         trans_for_each_update(trans, i) {
 788                 if (same_leaf_as_prev(trans, i))
 789                         continue;
 790
 791                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 792                         return trans_lock_write_fail(trans, i);
 793
 794                 if (!i->cached)
 795                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 796         }
 797
 798         return 0;
 799 }
 800
 801 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 802 {
 803         struct btree_insert_entry *i;
 804         struct btree_write_buffered_key *wb;
 805
 806         trans_for_each_update(trans, i)
 807                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 808
 809         trans_for_each_wb_update(trans, wb)
 810                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 811 }
 812
 813 #ifdef CONFIG_BCACHEFS_DEBUG
 814 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 815                                                    struct btree_insert_entry *i,
 816                                                    struct printbuf *err)
 817 {
 818         struct bch_fs *c = trans->c;
 819         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 820
 821         printbuf_reset(err);
 822         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 823                    trans->fn, (void *) i->ip_allocated);
 824         prt_newline(err);
 825         printbuf_indent_add(err, 2);
 826
 827         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 828         prt_newline(err);
 829
 830         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 831                           i->bkey_type, rw, err);
 832         bch2_print_string_as_lines(KERN_ERR, err->buf);
 833
 834         bch2_inconsistent_error(c);
 835         bch2_dump_trans_updates(trans);
 836         printbuf_exit(err);
 837
 838         return -EINVAL;
 839 }
 840 #endif
 841
 842 /*
 843  * Get journal reservation, take write locks, and attempt to do btree update(s):
 844  */
 845 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 846                                        struct btree_insert_entry **stopped_at,
 847                                        unsigned long trace_ip)
 848 {
 849         struct bch_fs *c = trans->c;
 850         struct btree_insert_entry *i;
 851         int ret, u64s_delta = 0;
 852
 853 #ifdef CONFIG_BCACHEFS_DEBUG
 854         struct printbuf buf = PRINTBUF;
 855
 856         trans_for_each_update(trans, i) {
 857                 int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 858
 859                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 860                                                i->bkey_type, rw, &buf)))
 861                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 862                 btree_insert_entry_checks(trans, i);
 863         }
 864         printbuf_exit(&buf);
 865 #endif
 866
 867         trans_for_each_update(trans, i) {
 868                 if (i->cached)
 869                         continue;
 870
 871                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 872                 u64s_delta -= i->old_btree_u64s;
 873
 874                 if (!same_leaf_as_next(trans, i)) {
 875                         if (u64s_delta <= 0) {
 876                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 877                                                         i->level, flags);
 878                                 if (unlikely(ret))
 879                                         return ret;
 880                         }
 881
 882                         u64s_delta = 0;
 883                 }
 884         }
 885
 886         ret = bch2_journal_preres_get(&c->journal,
 887                         &trans->journal_preres, trans->journal_preres_u64s,
 888                         (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 889         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 890                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 891         if (unlikely(ret))
 892                 return ret;
 893
 894         ret = trans_lock_write(trans);
 895         if (unlikely(ret))
 896                 return ret;
 897
 898         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 899
 900         if (!ret && unlikely(trans->journal_replay_not_finished))
 901                 bch2_drop_overwrites_from_journal(trans);
 902
 903         trans_for_each_update(trans, i)
 904                 if (!same_leaf_as_prev(trans, i))
 905                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 906                                                         insert_l(i)->b);
 907
 908         if (!ret && trans->journal_pin)
 909                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 910                                      trans->journal_pin, NULL);
 911
 912         /*
 913          * Drop journal reservation after dropping write locks, since dropping
 914          * the journal reservation may kick off a journal write:
 915          */
 916         bch2_journal_res_put(&c->journal, &trans->journal_res);
 917
 918         if (unlikely(ret))
 919                 return ret;
 920
 921         bch2_trans_downgrade(trans);
 922
 923         return 0;
 924 }
 925
 926 static int journal_reclaim_wait_done(struct bch_fs *c)
 927 {
 928         int ret = bch2_journal_error(&c->journal) ?:
 929                 !bch2_btree_key_cache_must_wait(c);
 930
 931         if (!ret)
 932                 journal_reclaim_kick(&c->journal);
 933         return ret;
 934 }
 935
 936 static noinline
 937 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 938                             struct btree_insert_entry *i,
 939                             int ret, unsigned long trace_ip)
 940 {
 941         struct bch_fs *c = trans->c;
 942
 943         switch (ret) {
 944         case -BCH_ERR_btree_insert_btree_node_full:
 945                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 946                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 947                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 948                 break;
 949         case -BCH_ERR_btree_insert_need_mark_replicas:
 950                 ret = drop_locks_do(trans,
 951                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 952                 break;
 953         case -BCH_ERR_journal_res_get_blocked:
 954                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 955                     !(flags & JOURNAL_WATERMARK_reserved)) {
 956                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 957                         break;
 958                 }
 959
 960                 ret = drop_locks_do(trans,
 961                         bch2_trans_journal_res_get(trans,
 962                                         (flags & JOURNAL_WATERMARK_MASK)|
 963                                         JOURNAL_RES_GET_CHECK));
 964                 break;
 965         case -BCH_ERR_btree_insert_need_journal_reclaim:
 966                 bch2_trans_unlock(trans);
 967
 968                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 969
 970                 wait_event_freezable(c->journal.reclaim_wait,
 971                                      (ret = journal_reclaim_wait_done(c)));
 972                 if (ret < 0)
 973                         break;
 974
 975                 ret = bch2_trans_relock(trans);
 976                 break;
 977         case -BCH_ERR_btree_insert_need_flush_buffer: {
 978                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 979
 980                 ret = 0;
 981
 982                 if (wb->state.nr > wb->size * 3 / 4) {
 983                         bch2_trans_unlock(trans);
 984                         mutex_lock(&wb->flush_lock);
 985
 986                         if (wb->state.nr > wb->size * 3 / 4) {
 987                                 bch2_trans_begin(trans);
 988                                 ret = __bch2_btree_write_buffer_flush(trans,
 989                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
 990                                 if (!ret) {
 991                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
 992                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
 993                                 }
 994                         } else {
 995                                 mutex_unlock(&wb->flush_lock);
 996                                 ret = bch2_trans_relock(trans);
 997                         }
 998                 }
 999                 break;
1000         }
1001         default:
1002                 BUG_ON(ret >= 0);
1003                 break;
1004         }
1005
1006         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1007
1008         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1009                                 !(flags & BTREE_INSERT_NOWAIT) &&
1010                                 (flags & BTREE_INSERT_NOFAIL), c,
1011                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1012
1013         return ret;
1014 }
1015
1016 static noinline int
1017 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1018 {
1019         struct bch_fs *c = trans->c;
1020         int ret;
1021
1022         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1023             test_bit(BCH_FS_STARTED, &c->flags))
1024                 return -BCH_ERR_erofs_trans_commit;
1025
1026         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1027         if (ret)
1028                 return ret;
1029
1030         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1031         return 0;
1032 }
1033
1034 /*
1035  * This is for updates done in the early part of fsck - btree_gc - before we've
1036  * gone RW. we only add the new key to the list of keys for journal replay to
1037  * do.
1038  */
1039 static noinline int
1040 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1041 {
1042         struct bch_fs *c = trans->c;
1043         struct btree_insert_entry *i;
1044         int ret = 0;
1045
1046         trans_for_each_update(trans, i) {
1047                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1048                 if (ret)
1049                         break;
1050         }
1051
1052         return ret;
1053 }
1054
1055 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1056 {
1057         struct bch_fs *c = trans->c;
1058         struct btree_insert_entry *i = NULL;
1059         struct btree_write_buffered_key *wb;
1060         unsigned u64s;
1061         int ret = 0;
1062
1063         if (!trans->nr_updates &&
1064             !trans->nr_wb_updates &&
1065             !trans->extra_journal_entries.nr)
1066                 goto out_reset;
1067
1068         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1069                 lockdep_assert_held(&c->gc_lock);
1070
1071         ret = bch2_trans_commit_run_triggers(trans);
1072         if (ret)
1073                 goto out_reset;
1074
1075         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1076                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1077                 goto out_reset;
1078         }
1079
1080         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1081             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1082                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1083                 if (ret)
1084                         goto out_reset;
1085         }
1086
1087         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1088             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1089                 bch2_trans_begin(trans);
1090                 bch2_trans_unlock(trans);
1091
1092                 ret = __bch2_btree_write_buffer_flush(trans,
1093                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1094                 if (!ret) {
1095                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1096                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1097                 }
1098                 goto out;
1099         }
1100
1101         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1102
1103         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1104
1105         trans->journal_u64s             = trans->extra_journal_entries.nr;
1106         trans->journal_preres_u64s      = 0;
1107
1108         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1109
1110         if (trans->journal_transaction_names)
1111                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1112
1113         trans_for_each_update(trans, i) {
1114                 EBUG_ON(!i->path->should_be_locked);
1115
1116                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1117                 if (unlikely(ret))
1118                         goto out;
1119
1120                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1121
1122                 if (i->key_cache_already_flushed)
1123                         continue;
1124
1125                 /* we're going to journal the key being updated: */
1126                 u64s = jset_u64s(i->k->k.u64s);
1127                 if (i->cached &&
1128                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1129                         trans->journal_preres_u64s += u64s;
1130
1131                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1132                         continue;
1133
1134                 trans->journal_u64s += u64s;
1135
1136                 /* and we're also going to log the overwrite: */
1137                 if (trans->journal_transaction_names)
1138                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1139         }
1140
1141         trans_for_each_wb_update(trans, wb)
1142                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1143
1144         if (trans->extra_journal_res) {
1145                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1146                                 trans->extra_journal_res,
1147                                 (flags & BTREE_INSERT_NOFAIL)
1148                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1149                 if (ret)
1150                         goto err;
1151         }
1152 retry:
1153         bch2_trans_verify_not_in_restart(trans);
1154         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1155
1156         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1157
1158         /* make sure we didn't drop or screw up locks: */
1159         bch2_trans_verify_locks(trans);
1160
1161         if (ret)
1162                 goto err;
1163
1164         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1165 out:
1166         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1167
1168         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1169                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1170 out_reset:
1171         bch2_trans_reset_updates(trans);
1172
1173         return ret;
1174 err:
1175         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1176         if (ret)
1177                 goto out;
1178
1179         goto retry;
1180 }
1181
1182 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1183                                           enum btree_id id,
1184                                           struct bpos pos)
1185 {
1186         struct bch_fs *c = trans->c;
1187         struct btree_iter iter;
1188         struct bkey_s_c k;
1189         int ret;
1190
1191         bch2_trans_iter_init(trans, &iter, id, pos,
1192                              BTREE_ITER_NOT_EXTENTS|
1193                              BTREE_ITER_ALL_SNAPSHOTS);
1194         while (1) {
1195                 k = bch2_btree_iter_prev(&iter);
1196                 ret = bkey_err(k);
1197                 if (ret)
1198                         break;
1199
1200                 if (!k.k)
1201                         break;
1202
1203                 if (!bkey_eq(pos, k.k->p))
1204                         break;
1205
1206                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1207                         ret = 1;
1208                         break;
1209                 }
1210         }
1211         bch2_trans_iter_exit(trans, &iter);
1212
1213         return ret;
1214 }
1215
1216 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1217                                           enum btree_id id,
1218                                           struct bpos pos)
1219 {
1220         if (!btree_type_has_snapshots(id) ||
1221             pos.snapshot == U32_MAX ||
1222             !snapshot_t(trans->c, pos.snapshot)->children[0])
1223                 return 0;
1224
1225         return __check_pos_snapshot_overwritten(trans, id, pos);
1226 }
1227
1228 static noinline int extent_front_merge(struct btree_trans *trans,
1229                                        struct btree_iter *iter,
1230                                        struct bkey_s_c k,
1231                                        struct bkey_i **insert,
1232                                        enum btree_update_flags flags)
1233 {
1234         struct bch_fs *c = trans->c;
1235         struct bkey_i *update;
1236         int ret;
1237
1238         update = bch2_bkey_make_mut_noupdate(trans, k);
1239         ret = PTR_ERR_OR_ZERO(update);
1240         if (ret)
1241                 return ret;
1242
1243         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1244                 return 0;
1245
1246         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1247                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1248         if (ret < 0)
1249                 return ret;
1250         if (ret)
1251                 return 0;
1252
1253         ret = bch2_btree_delete_at(trans, iter, flags);
1254         if (ret)
1255                 return ret;
1256
1257         *insert = update;
1258         return 0;
1259 }
1260
1261 static noinline int extent_back_merge(struct btree_trans *trans,
1262                                       struct btree_iter *iter,
1263                                       struct bkey_i *insert,
1264                                       struct bkey_s_c k)
1265 {
1266         struct bch_fs *c = trans->c;
1267         int ret;
1268
1269         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1270                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1271         if (ret < 0)
1272                 return ret;
1273         if (ret)
1274                 return 0;
1275
1276         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1277         return 0;
1278 }
1279
1280 /*
1281  * When deleting, check if we need to emit a whiteout (because we're overwriting
1282  * something in an ancestor snapshot)
1283  */
1284 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1285                                       enum btree_id btree_id, struct bpos pos)
1286 {
1287         struct btree_iter iter;
1288         struct bkey_s_c k;
1289         u32 snapshot = pos.snapshot;
1290         int ret;
1291
1292         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1293                 return 0;
1294
1295         pos.snapshot++;
1296
1297         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1298                            BTREE_ITER_ALL_SNAPSHOTS|
1299                            BTREE_ITER_NOPRESERVE, k, ret) {
1300                 if (!bkey_eq(k.k->p, pos))
1301                         break;
1302
1303                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1304                                               k.k->p.snapshot)) {
1305                         ret = !bkey_whiteout(k.k);
1306                         break;
1307                 }
1308         }
1309         bch2_trans_iter_exit(trans, &iter);
1310
1311         return ret;
1312 }
1313
1314 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
1315                                    enum btree_id id,
1316                                    struct bpos old_pos,
1317                                    struct bpos new_pos)
1318 {
1319         struct bch_fs *c = trans->c;
1320         struct btree_iter old_iter, new_iter;
1321         struct bkey_s_c old_k, new_k;
1322         snapshot_id_list s;
1323         struct bkey_i *update;
1324         int ret;
1325
1326         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
1327                 return 0;
1328
1329         darray_init(&s);
1330
1331         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
1332                              BTREE_ITER_NOT_EXTENTS|
1333                              BTREE_ITER_ALL_SNAPSHOTS);
1334         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
1335                !(ret = bkey_err(old_k)) &&
1336                bkey_eq(old_pos, old_k.k->p)) {
1337                 struct bpos whiteout_pos =
1338                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
1339
1340                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
1341                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
1342                         continue;
1343
1344                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
1345                                            BTREE_ITER_NOT_EXTENTS|
1346                                            BTREE_ITER_INTENT);
1347                 ret = bkey_err(new_k);
1348                 if (ret)
1349                         break;
1350
1351                 if (new_k.k->type == KEY_TYPE_deleted) {
1352                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
1353                         ret = PTR_ERR_OR_ZERO(update);
1354                         if (ret)
1355                                 break;
1356
1357                         bkey_init(&update->k);
1358                         update->k.p             = whiteout_pos;
1359                         update->k.type          = KEY_TYPE_whiteout;
1360
1361                         ret = bch2_trans_update(trans, &new_iter, update,
1362                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1363                 }
1364                 bch2_trans_iter_exit(trans, &new_iter);
1365
1366                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
1367                 if (ret)
1368                         break;
1369         }
1370         bch2_trans_iter_exit(trans, &old_iter);
1371         darray_exit(&s);
1372
1373         return ret;
1374 }
1375
1376 int bch2_trans_update_extent(struct btree_trans *trans,
1377                              struct btree_iter *orig_iter,
1378                              struct bkey_i *insert,
1379                              enum btree_update_flags flags)
1380 {
1381         struct btree_iter iter;
1382         struct bpos start = bkey_start_pos(&insert->k);
1383         struct bkey_i *update;
1384         struct bkey_s_c k;
1385         enum btree_id btree_id = orig_iter->btree_id;
1386         int ret = 0, compressed_sectors;
1387
1388         bch2_trans_iter_init(trans, &iter, btree_id, start,
1389                              BTREE_ITER_INTENT|
1390                              BTREE_ITER_WITH_UPDATES|
1391                              BTREE_ITER_NOT_EXTENTS);
1392         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1393         if ((ret = bkey_err(k)))
1394                 goto err;
1395         if (!k.k)
1396                 goto out;
1397
1398         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1399                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1400                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1401                         if (ret)
1402                                 goto err;
1403                 }
1404
1405                 goto next;
1406         }
1407
1408         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1409                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1410                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1411
1412                 /*
1413                  * If we're going to be splitting a compressed extent, note it
1414                  * so that __bch2_trans_commit() can increase our disk
1415                  * reservation:
1416                  */
1417                 if (((front_split && back_split) ||
1418                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1419                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1420                         trans->extra_journal_res += compressed_sectors;
1421
1422                 if (front_split) {
1423                         update = bch2_bkey_make_mut_noupdate(trans, k);
1424                         if ((ret = PTR_ERR_OR_ZERO(update)))
1425                                 goto err;
1426
1427                         bch2_cut_back(start, update);
1428
1429                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1430                                                 k.k->p, update->k.p) ?:
1431                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1432                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1433                         if (ret)
1434                                 goto err;
1435                 }
1436
1437                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1438                     (front_split || back_split)) {
1439                         update = bch2_bkey_make_mut_noupdate(trans, k);
1440                         if ((ret = PTR_ERR_OR_ZERO(update)))
1441                                 goto err;
1442
1443                         bch2_cut_front(start, update);
1444                         bch2_cut_back(insert->k.p, update);
1445
1446                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1447                                                 k.k->p, update->k.p) ?:
1448                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1449                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1450                         if (ret)
1451                                 goto err;
1452                 }
1453
1454                 if (bkey_le(k.k->p, insert->k.p)) {
1455                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1456                         if ((ret = PTR_ERR_OR_ZERO(update)))
1457                                 goto err;
1458
1459                         bkey_init(&update->k);
1460                         update->k.p = k.k->p;
1461                         update->k.p.snapshot = insert->k.p.snapshot;
1462
1463                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1464                                 update->k.type = KEY_TYPE_whiteout;
1465                         } else if (btree_type_has_snapshots(btree_id)) {
1466                                 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1467                                 if (ret < 0)
1468                                         goto err;
1469                                 if (ret)
1470                                         update->k.type = KEY_TYPE_whiteout;
1471                         }
1472
1473                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1474                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1475                         if (ret)
1476                                 goto err;
1477                 }
1478
1479                 if (back_split) {
1480                         update = bch2_bkey_make_mut_noupdate(trans, k);
1481                         if ((ret = PTR_ERR_OR_ZERO(update)))
1482                                 goto err;
1483
1484                         bch2_cut_front(insert->k.p, update);
1485
1486                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1487                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1488                                                   flags);
1489                         if (ret)
1490                                 goto err;
1491                         goto out;
1492                 }
1493 next:
1494                 bch2_btree_iter_advance(&iter);
1495                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1496                 if ((ret = bkey_err(k)))
1497                         goto err;
1498                 if (!k.k)
1499                         goto out;
1500         }
1501
1502         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1503                 ret = extent_back_merge(trans, &iter, insert, k);
1504                 if (ret)
1505                         goto err;
1506         }
1507 out:
1508         if (!bkey_deleted(&insert->k)) {
1509                 /*
1510                  * Rewinding iterators is expensive: get a new one and the one
1511                  * that points to the start of insert will be cloned from:
1512                  */
1513                 bch2_trans_iter_exit(trans, &iter);
1514                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1515                                      BTREE_ITER_NOT_EXTENTS|
1516                                      BTREE_ITER_INTENT);
1517                 ret   = bch2_btree_iter_traverse(&iter) ?:
1518                         bch2_trans_update(trans, &iter, insert, flags);
1519         }
1520 err:
1521         bch2_trans_iter_exit(trans, &iter);
1522
1523         return ret;
1524 }
1525
1526 static int __must_check
1527 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1528                                 struct bkey_i *k, enum btree_update_flags flags,
1529                                 unsigned long ip);
1530
1531 static noinline int flush_new_cached_update(struct btree_trans *trans,
1532                                             struct btree_path *path,
1533                                             struct btree_insert_entry *i,
1534                                             enum btree_update_flags flags,
1535                                             unsigned long ip)
1536 {
1537         struct btree_path *btree_path;
1538         struct bkey k;
1539         int ret;
1540
1541         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1542                                    BTREE_ITER_INTENT, _THIS_IP_);
1543         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1544         if (ret)
1545                 goto out;
1546
1547         /*
1548          * The old key in the insert entry might actually refer to an existing
1549          * key in the btree that has been deleted from cache and not yet
1550          * flushed. Check for this and skip the flush so we don't run triggers
1551          * against a stale key.
1552          */
1553         bch2_btree_path_peek_slot_exact(btree_path, &k);
1554         if (!bkey_deleted(&k))
1555                 goto out;
1556
1557         i->key_cache_already_flushed = true;
1558         i->flags |= BTREE_TRIGGER_NORUN;
1559
1560         btree_path_set_should_be_locked(btree_path);
1561         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1562 out:
1563         bch2_path_put(trans, btree_path, true);
1564         return ret;
1565 }
1566
1567 static int __must_check
1568 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1569                                 struct bkey_i *k, enum btree_update_flags flags,
1570                                 unsigned long ip)
1571 {
1572         struct bch_fs *c = trans->c;
1573         struct btree_insert_entry *i, n;
1574         int cmp;
1575
1576         EBUG_ON(!path->should_be_locked);
1577         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1578         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1579
1580         n = (struct btree_insert_entry) {
1581                 .flags          = flags,
1582                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1583                 .btree_id       = path->btree_id,
1584                 .level          = path->level,
1585                 .cached         = path->cached,
1586                 .path           = path,
1587                 .k              = k,
1588                 .ip_allocated   = ip,
1589         };
1590
1591 #ifdef CONFIG_BCACHEFS_DEBUG
1592         trans_for_each_update(trans, i)
1593                 BUG_ON(i != trans->updates &&
1594                        btree_insert_entry_cmp(i - 1, i) >= 0);
1595 #endif
1596
1597         /*
1598          * Pending updates are kept sorted: first, find position of new update,
1599          * then delete/trim any updates the new update overwrites:
1600          */
1601         trans_for_each_update(trans, i) {
1602                 cmp = btree_insert_entry_cmp(&n, i);
1603                 if (cmp <= 0)
1604                         break;
1605         }
1606
1607         if (!cmp && i < trans->updates + trans->nr_updates) {
1608                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1609
1610                 bch2_path_put(trans, i->path, true);
1611                 i->flags        = n.flags;
1612                 i->cached       = n.cached;
1613                 i->k            = n.k;
1614                 i->path         = n.path;
1615                 i->ip_allocated = n.ip_allocated;
1616         } else {
1617                 array_insert_item(trans->updates, trans->nr_updates,
1618                                   i - trans->updates, n);
1619
1620                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1621                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1622
1623                 if (unlikely(trans->journal_replay_not_finished)) {
1624                         struct bkey_i *j_k =
1625                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1626
1627                         if (j_k) {
1628                                 i->old_k = j_k->k;
1629                                 i->old_v = &j_k->v;
1630                         }
1631                 }
1632         }
1633
1634         __btree_path_get(i->path, true);
1635
1636         /*
1637          * If a key is present in the key cache, it must also exist in the
1638          * btree - this is necessary for cache coherency. When iterating over
1639          * a btree that's cached in the key cache, the btree iter code checks
1640          * the key cache - but the key has to exist in the btree for that to
1641          * work:
1642          */
1643         if (path->cached && bkey_deleted(&i->old_k))
1644                 return flush_new_cached_update(trans, path, i, flags, ip);
1645
1646         return 0;
1647 }
1648
1649 static inline int __must_check
1650 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1651                           struct bkey_i *k, enum btree_update_flags flags)
1652 {
1653         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1654 }
1655
1656 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1657                                    struct bkey_i *k, enum btree_update_flags flags)
1658 {
1659         struct btree_path *path = iter->update_path ?: iter->path;
1660         struct bkey_cached *ck;
1661         int ret;
1662
1663         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1664                 return bch2_trans_update_extent(trans, iter, k, flags);
1665
1666         if (bkey_deleted(&k->k) &&
1667             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1668             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1669                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1670                 if (unlikely(ret < 0))
1671                         return ret;
1672
1673                 if (ret)
1674                         k->k.type = KEY_TYPE_whiteout;
1675         }
1676
1677         /*
1678          * Ensure that updates to cached btrees go to the key cache:
1679          */
1680         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1681             !path->cached &&
1682             !path->level &&
1683             btree_id_cached(trans->c, path->btree_id)) {
1684                 if (!iter->key_cache_path ||
1685                     !iter->key_cache_path->should_be_locked ||
1686                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1687                         if (!iter->key_cache_path)
1688                                 iter->key_cache_path =
1689                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1690                                                       BTREE_ITER_INTENT|
1691                                                       BTREE_ITER_CACHED, _THIS_IP_);
1692
1693                         iter->key_cache_path =
1694                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1695                                                         iter->flags & BTREE_ITER_INTENT,
1696                                                         _THIS_IP_);
1697
1698                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1699                                                        BTREE_ITER_CACHED);
1700                         if (unlikely(ret))
1701                                 return ret;
1702
1703                         ck = (void *) iter->key_cache_path->l[0].b;
1704
1705                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1706                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1707                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1708                         }
1709
1710                         btree_path_set_should_be_locked(iter->key_cache_path);
1711                 }
1712
1713                 path = iter->key_cache_path;
1714         }
1715
1716         return bch2_trans_update_by_path(trans, path, k, flags);
1717 }
1718
1719 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1720                                             enum btree_id btree,
1721                                             struct bkey_i *k)
1722 {
1723         struct btree_write_buffered_key *i;
1724         int ret;
1725
1726         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1727         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1728
1729         trans_for_each_wb_update(trans, i) {
1730                 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
1731                         bkey_copy(&i->k, k);
1732                         return 0;
1733                 }
1734         }
1735
1736         if (!trans->wb_updates ||
1737             trans->nr_wb_updates == trans->wb_updates_size) {
1738                 struct btree_write_buffered_key *u;
1739
1740                 if (trans->nr_wb_updates == trans->wb_updates_size) {
1741                         struct btree_transaction_stats *s = btree_trans_stats(trans);
1742
1743                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1744                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1745                         if (s)
1746                                 s->wb_updates_size = trans->wb_updates_size;
1747                 }
1748
1749                 u = bch2_trans_kmalloc_nomemzero(trans,
1750                                         trans->wb_updates_size *
1751                                         sizeof(struct btree_write_buffered_key));
1752                 ret = PTR_ERR_OR_ZERO(u);
1753                 if (ret)
1754                         return ret;
1755
1756                 if (trans->nr_wb_updates)
1757                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1758                                sizeof(struct btree_write_buffered_key));
1759                 trans->wb_updates = u;
1760         }
1761
1762         trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
1763                 .btree  = btree,
1764         };
1765
1766         bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
1767         trans->nr_wb_updates++;
1768
1769         return 0;
1770 }
1771
1772 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1773                              enum btree_id btree, struct bpos end)
1774 {
1775         struct bkey_s_c k;
1776         int ret = 0;
1777
1778         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1779         k = bch2_btree_iter_prev(iter);
1780         ret = bkey_err(k);
1781         if (ret)
1782                 goto err;
1783
1784         bch2_btree_iter_advance(iter);
1785         k = bch2_btree_iter_peek_slot(iter);
1786         ret = bkey_err(k);
1787         if (ret)
1788                 goto err;
1789
1790         BUG_ON(k.k->type != KEY_TYPE_deleted);
1791
1792         if (bkey_gt(k.k->p, end)) {
1793                 ret = -BCH_ERR_ENOSPC_btree_slot;
1794                 goto err;
1795         }
1796
1797         return 0;
1798 err:
1799         bch2_trans_iter_exit(trans, iter);
1800         return ret;
1801 }
1802
1803 void bch2_trans_commit_hook(struct btree_trans *trans,
1804                             struct btree_trans_commit_hook *h)
1805 {
1806         h->next = trans->hooks;
1807         trans->hooks = h;
1808 }
1809
1810 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1811                                 enum btree_id btree, struct bkey_i *k,
1812                                 enum btree_update_flags flags)
1813 {
1814         struct btree_iter iter;
1815         int ret;
1816
1817         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1818                              BTREE_ITER_NOT_EXTENTS|
1819                              BTREE_ITER_INTENT);
1820         ret   = bch2_btree_iter_traverse(&iter) ?:
1821                 bch2_trans_update(trans, &iter, k, flags);
1822         bch2_trans_iter_exit(trans, &iter);
1823         return ret;
1824 }
1825
1826 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1827                         struct bkey_i *k, enum btree_update_flags flags)
1828 {
1829         struct btree_iter iter;
1830         int ret;
1831
1832         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1833                              BTREE_ITER_CACHED|
1834                              BTREE_ITER_INTENT);
1835         ret   = bch2_btree_iter_traverse(&iter) ?:
1836                 bch2_trans_update(trans, &iter, k, flags);
1837         bch2_trans_iter_exit(trans, &iter);
1838         return ret;
1839 }
1840
1841 /**
1842  * bch2_btree_insert - insert keys into the extent btree
1843  * @c:                  pointer to struct bch_fs
1844  * @id:                 btree to insert into
1845  * @insert_keys:        list of keys to insert
1846  * @hook:               insert callback
1847  */
1848 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1849                       struct bkey_i *k,
1850                       struct disk_reservation *disk_res,
1851                       u64 *journal_seq, int flags)
1852 {
1853         return bch2_trans_do(c, disk_res, journal_seq, flags,
1854                              __bch2_btree_insert(&trans, id, k, 0));
1855 }
1856
1857 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1858                                 unsigned len, unsigned update_flags)
1859 {
1860         struct bkey_i *k;
1861
1862         k = bch2_trans_kmalloc(trans, sizeof(*k));
1863         if (IS_ERR(k))
1864                 return PTR_ERR(k);
1865
1866         bkey_init(&k->k);
1867         k->k.p = iter->pos;
1868         bch2_key_resize(&k->k, len);
1869         return bch2_trans_update(trans, iter, k, update_flags);
1870 }
1871
1872 int bch2_btree_delete_at(struct btree_trans *trans,
1873                          struct btree_iter *iter, unsigned update_flags)
1874 {
1875         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1876 }
1877
1878 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1879                                   enum btree_id btree, struct bpos pos)
1880 {
1881         struct bkey_i *k;
1882
1883         k = bch2_trans_kmalloc(trans, sizeof(*k));
1884         if (IS_ERR(k))
1885                 return PTR_ERR(k);
1886
1887         bkey_init(&k->k);
1888         k->k.p = pos;
1889         return bch2_trans_update_buffered(trans, btree, k);
1890 }
1891
1892 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1893                                   struct bpos start, struct bpos end,
1894                                   unsigned update_flags,
1895                                   u64 *journal_seq)
1896 {
1897         u32 restart_count = trans->restart_count;
1898         struct btree_iter iter;
1899         struct bkey_s_c k;
1900         int ret = 0;
1901
1902         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1903         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1904                 struct disk_reservation disk_res =
1905                         bch2_disk_reservation_init(trans->c, 0);
1906                 struct bkey_i delete;
1907
1908                 ret = bkey_err(k);
1909                 if (ret)
1910                         goto err;
1911
1912                 bkey_init(&delete.k);
1913
1914                 /*
1915                  * This could probably be more efficient for extents:
1916                  */
1917
1918                 /*
1919                  * For extents, iter.pos won't necessarily be the same as
1920                  * bkey_start_pos(k.k) (for non extents they always will be the
1921                  * same). It's important that we delete starting from iter.pos
1922                  * because the range we want to delete could start in the middle
1923                  * of k.
1924                  *
1925                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1926                  * bkey_start_pos(k.k)).
1927                  */
1928                 delete.k.p = iter.pos;
1929
1930                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1931                         bch2_key_resize(&delete.k,
1932                                         bpos_min(end, k.k->p).offset -
1933                                         iter.pos.offset);
1934
1935                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1936                         bch2_trans_commit(trans, &disk_res, journal_seq,
1937                                           BTREE_INSERT_NOFAIL);
1938                 bch2_disk_reservation_put(trans->c, &disk_res);
1939 err:
1940                 /*
1941                  * the bch2_trans_begin() call is in a weird place because we
1942                  * need to call it after every transaction commit, to avoid path
1943                  * overflow, but don't want to call it if the delete operation
1944                  * is a no-op and we have no work to do:
1945                  */
1946                 bch2_trans_begin(trans);
1947
1948                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1949                         ret = 0;
1950                 if (ret)
1951                         break;
1952         }
1953         bch2_trans_iter_exit(trans, &iter);
1954
1955         if (!ret && trans_was_restarted(trans, restart_count))
1956                 ret = -BCH_ERR_transaction_restart_nested;
1957         return ret;
1958 }
1959
1960 /*
1961  * bch_btree_delete_range - delete everything within a given range
1962  *
1963  * Range is a half open interval - [start, end)
1964  */
1965 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1966                             struct bpos start, struct bpos end,
1967                             unsigned update_flags,
1968                             u64 *journal_seq)
1969 {
1970         int ret = bch2_trans_run(c,
1971                         bch2_btree_delete_range_trans(&trans, id, start, end,
1972                                                       update_flags, journal_seq));
1973         if (ret == -BCH_ERR_transaction_restart_nested)
1974                 ret = 0;
1975         return ret;
1976 }
1977
1978 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1979 {
1980         struct printbuf buf = PRINTBUF;
1981         struct jset_entry_log *l;
1982         unsigned u64s;
1983         int ret;
1984
1985         prt_vprintf(&buf, fmt, args);
1986         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
1987         if (ret)
1988                 goto err;
1989
1990         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
1991
1992         ret = darray_make_room(entries, jset_u64s(u64s));
1993         if (ret)
1994                 goto err;
1995
1996         l = (void *) &darray_top(*entries);
1997         l->entry.u64s           = cpu_to_le16(u64s);
1998         l->entry.btree_id       = 0;
1999         l->entry.level          = 1;
2000         l->entry.type           = BCH_JSET_ENTRY_log;
2001         l->entry.pad[0]         = 0;
2002         l->entry.pad[1]         = 0;
2003         l->entry.pad[2]         = 0;
2004         memcpy(l->d, buf.buf, buf.pos);
2005         while (buf.pos & 7)
2006                 l->d[buf.pos++] = '\0';
2007
2008         entries->nr += jset_u64s(u64s);
2009 err:
2010         printbuf_exit(&buf);
2011         return ret;
2012 }
2013
2014 static int
2015 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
2016                   va_list args)
2017 {
2018         int ret;
2019
2020         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
2021                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
2022         } else {
2023                 ret = bch2_trans_do(c, NULL, NULL,
2024                         BTREE_INSERT_LAZY_RW|commit_flags,
2025                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
2026         }
2027
2028         return ret;
2029 }
2030
2031 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
2032 {
2033         va_list args;
2034         int ret;
2035
2036         va_start(args, fmt);
2037         ret = __bch2_fs_log_msg(c, 0, fmt, args);
2038         va_end(args);
2039         return ret;
2040 }
2041
2042 /*
2043  * Use for logging messages during recovery to enable reserved space and avoid
2044  * blocking.
2045  */
2046 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2047 {
2048         va_list args;
2049         int ret;
2050
2051         va_start(args, fmt);
2052         ret = __bch2_fs_log_msg(c, JOURNAL_WATERMARK_reserved, fmt, args);
2053         va_end(args);
2054         return ret;
2055 }