git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_leaf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "buckets.h"
  12 #include "debug.h"
  13 #include "errcode.h"
  14 #include "error.h"
  15 #include "extent_update.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "subvolume.h"
  21 #include "replicas.h"
  22
  23 #include <linux/prefetch.h>
  24 #include <linux/sort.h>
  25 #include <trace/events/bcachefs.h>
  26
  27 /*
  28  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
  29  * different snapshot:
  30  */
  31 struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
  32 {
  33         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
  34
  35         if (k.k && bpos_eq(path->pos, k.k->p))
  36                 return k;
  37
  38         bkey_init(u);
  39         u->p = path->pos;
  40         return (struct bkey_s_c) { u, NULL };
  41 }
  42
  43 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  44 {
  45 #ifdef CONFIG_BCACHEFS_DEBUG
  46         struct bch_fs *c = trans->c;
  47         struct bkey u;
  48         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
  49
  50         if (unlikely(trans->journal_replay_not_finished)) {
  51                 struct bkey_i *j_k =
  52                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  53
  54                 if (j_k)
  55                         k = bkey_i_to_s_c(j_k);
  56         }
  57
  58         i->old_k.needs_whiteout = k.k->needs_whiteout;
  59
  60         BUG_ON(memcmp(&i->old_k, k.k, sizeof(struct bkey)));
  61         BUG_ON(i->old_v != k.v);
  62 #endif
  63 }
  64
  65 static int __must_check
  66 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  67                           struct bkey_i *, enum btree_update_flags);
  68
  69 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  70                                          const struct btree_insert_entry *r)
  71 {
  72         return   cmp_int(l->btree_id,   r->btree_id) ?:
  73                  cmp_int(l->cached,     r->cached) ?:
  74                  -cmp_int(l->level,     r->level) ?:
  75                  bpos_cmp(l->k->k.p,    r->k->k.p);
  76 }
  77
  78 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  79 {
  80         return i->path->l + i->level;
  81 }
  82
  83 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  84                                      struct btree_insert_entry *i)
  85 {
  86         return i != trans->updates &&
  87                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  88 }
  89
  90 static inline bool same_leaf_as_next(struct btree_trans *trans,
  91                                      struct btree_insert_entry *i)
  92 {
  93         return i + 1 < trans->updates + trans->nr_updates &&
  94                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  95 }
  96
  97 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  98                                            struct btree_path *path,
  99                                            struct btree *b)
 100 {
 101         struct bch_fs *c = trans->c;
 102
 103         if (path->cached)
 104                 return;
 105
 106         if (unlikely(btree_node_just_written(b)) &&
 107             bch2_btree_post_write_cleanup(c, b))
 108                 bch2_trans_node_reinit_iter(trans, b);
 109
 110         /*
 111          * If the last bset has been written, or if it's gotten too big - start
 112          * a new bset to insert into:
 113          */
 114         if (want_new_bset(c, b))
 115                 bch2_btree_init_next(trans, b);
 116 }
 117
 118 /* Inserting into a given leaf node (last stage of insert): */
 119
 120 /* Handle overwrites and do insert, for non extents: */
 121 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 122                                 struct btree_path *path,
 123                                 struct btree *b,
 124                                 struct btree_node_iter *node_iter,
 125                                 struct bkey_i *insert)
 126 {
 127         struct bkey_packed *k;
 128         unsigned clobber_u64s = 0, new_u64s = 0;
 129
 130         EBUG_ON(btree_node_just_written(b));
 131         EBUG_ON(bset_written(b, btree_bset_last(b)));
 132         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 133         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 134         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 135         EBUG_ON(insert->k.u64s >
 136                 bch_btree_keys_u64s_remaining(trans->c, b));
 137
 138         k = bch2_btree_node_iter_peek_all(node_iter, b);
 139         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 140                 k = NULL;
 141
 142         /* @k is the key being overwritten/deleted, if any: */
 143         EBUG_ON(k && bkey_deleted(k));
 144
 145         /* Deleting, but not found? nothing to do: */
 146         if (bkey_deleted(&insert->k) && !k)
 147                 return false;
 148
 149         if (bkey_deleted(&insert->k)) {
 150                 /* Deleting: */
 151                 btree_account_key_drop(b, k);
 152                 k->type = KEY_TYPE_deleted;
 153
 154                 if (k->needs_whiteout)
 155                         push_whiteout(trans->c, b, insert->k.p);
 156                 k->needs_whiteout = false;
 157
 158                 if (k >= btree_bset_last(b)->start) {
 159                         clobber_u64s = k->u64s;
 160                         bch2_bset_delete(b, k, clobber_u64s);
 161                         goto fix_iter;
 162                 } else {
 163                         bch2_btree_path_fix_key_modified(trans, b, k);
 164                 }
 165
 166                 return true;
 167         }
 168
 169         if (k) {
 170                 /* Overwriting: */
 171                 btree_account_key_drop(b, k);
 172                 k->type = KEY_TYPE_deleted;
 173
 174                 insert->k.needs_whiteout = k->needs_whiteout;
 175                 k->needs_whiteout = false;
 176
 177                 if (k >= btree_bset_last(b)->start) {
 178                         clobber_u64s = k->u64s;
 179                         goto overwrite;
 180                 } else {
 181                         bch2_btree_path_fix_key_modified(trans, b, k);
 182                 }
 183         }
 184
 185         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 186 overwrite:
 187         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 188         new_u64s = k->u64s;
 189 fix_iter:
 190         if (clobber_u64s != new_u64s)
 191                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 192                                          clobber_u64s, new_u64s);
 193         return true;
 194 }
 195
 196 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 197                                unsigned i, u64 seq)
 198 {
 199         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 200         struct btree_write *w = container_of(pin, struct btree_write, journal);
 201         struct btree *b = container_of(w, struct btree, writes[i]);
 202         struct btree_trans trans;
 203         unsigned long old, new, v;
 204         unsigned idx = w - b->writes;
 205
 206         bch2_trans_init(&trans, c, 0, 0);
 207
 208         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 209         v = READ_ONCE(b->flags);
 210
 211         do {
 212                 old = new = v;
 213
 214                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 215                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 216                     w->journal.seq != seq)
 217                         break;
 218
 219                 new &= ~BTREE_WRITE_TYPE_MASK;
 220                 new |= BTREE_WRITE_journal_reclaim;
 221                 new |= 1 << BTREE_NODE_need_write;
 222         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 223
 224         btree_node_write_if_need(c, b, SIX_LOCK_read);
 225         six_unlock_read(&b->c.lock);
 226
 227         bch2_trans_exit(&trans);
 228         return 0;
 229 }
 230
 231 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 232 {
 233         return __btree_node_flush(j, pin, 0, seq);
 234 }
 235
 236 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 237 {
 238         return __btree_node_flush(j, pin, 1, seq);
 239 }
 240
 241 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 242                                        struct btree *b, u64 seq)
 243 {
 244         struct btree_write *w = btree_current_write(b);
 245
 246         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 247                              btree_node_write_idx(b) == 0
 248                              ? btree_node_flush0
 249                              : btree_node_flush1);
 250 }
 251
 252 /**
 253  * btree_insert_key - insert a key one key into a leaf node
 254  */
 255 static void btree_insert_key_leaf(struct btree_trans *trans,
 256                                   struct btree_insert_entry *insert)
 257 {
 258         struct bch_fs *c = trans->c;
 259         struct btree *b = insert_l(insert)->b;
 260         struct bset_tree *t = bset_tree_last(b);
 261         struct bset *i = bset(b, t);
 262         int old_u64s = bset_u64s(t);
 263         int old_live_u64s = b->nr.live_u64s;
 264         int live_u64s_added, u64s_added;
 265
 266         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 267                                         &insert_l(insert)->iter, insert->k)))
 268                 return;
 269
 270         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 271                                          le64_to_cpu(i->journal_seq)));
 272
 273         bch2_btree_add_journal_pin(c, b, trans->journal_res.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 /* Cached btree updates: */
 292
 293 /* Normal update interface: */
 294
 295 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 296                                              struct btree_insert_entry *i)
 297 {
 298         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 299         BUG_ON(i->cached        != i->path->cached);
 300         BUG_ON(i->level         != i->path->level);
 301         BUG_ON(i->btree_id      != i->path->btree_id);
 302         EBUG_ON(!i->level &&
 303                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 304                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 305                 i->k->k.p.snapshot &&
 306                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 307 }
 308
 309 static noinline int
 310 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 311                                    unsigned long trace_ip)
 312 {
 313         struct bch_fs *c = trans->c;
 314         int ret;
 315
 316         bch2_trans_unlock(trans);
 317
 318         ret = bch2_journal_preres_get(&c->journal,
 319                         &trans->journal_preres, u64s, 0);
 320         if (ret)
 321                 return ret;
 322
 323         ret = bch2_trans_relock(trans);
 324         if (ret) {
 325                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 326                 return ret;
 327         }
 328
 329         return 0;
 330 }
 331
 332 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 333                                                       unsigned flags)
 334 {
 335         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 336                                     trans->journal_u64s, flags);
 337 }
 338
 339 #define JSET_ENTRY_LOG_U64s             4
 340
 341 static noinline void journal_transaction_name(struct btree_trans *trans)
 342 {
 343         struct bch_fs *c = trans->c;
 344         struct journal *j = &c->journal;
 345         struct jset_entry *entry =
 346                 bch2_journal_add_entry(j, &trans->journal_res,
 347                                        BCH_JSET_ENTRY_log, 0, 0,
 348                                        JSET_ENTRY_LOG_U64s);
 349         struct jset_entry_log *l =
 350                 container_of(entry, struct jset_entry_log, entry);
 351
 352         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 353 }
 354
 355 static inline int btree_key_can_insert(struct btree_trans *trans,
 356                                        struct btree *b, unsigned u64s)
 357 {
 358         struct bch_fs *c = trans->c;
 359
 360         if (!bch2_btree_node_insert_fits(c, b, u64s))
 361                 return -BCH_ERR_btree_insert_btree_node_full;
 362
 363         return 0;
 364 }
 365
 366 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 367                                        struct btree_path *path, unsigned u64s)
 368 {
 369         struct bch_fs *c = trans->c;
 370         struct bkey_cached *ck = (void *) path->l[0].b;
 371         struct btree_insert_entry *i;
 372         unsigned new_u64s;
 373         struct bkey_i *new_k;
 374
 375         EBUG_ON(path->level);
 376
 377         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 378             bch2_btree_key_cache_must_wait(c) &&
 379             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 380                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 381
 382         /*
 383          * bch2_varint_decode can read past the end of the buffer by at most 7
 384          * bytes (it won't be used):
 385          */
 386         u64s += 1;
 387
 388         if (u64s <= ck->u64s)
 389                 return 0;
 390
 391         new_u64s        = roundup_pow_of_two(u64s);
 392         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 393         if (!new_k) {
 394                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 395                         bch2_btree_ids[path->btree_id], new_u64s);
 396                 return -ENOMEM;
 397         }
 398
 399         trans_for_each_update(trans, i)
 400                 if (i->old_v == &ck->k->v)
 401                         i->old_v = &new_k->v;
 402
 403         ck->u64s        = new_u64s;
 404         ck->k           = new_k;
 405         return 0;
 406 }
 407
 408 /* Triggers: */
 409
 410 static int run_one_mem_trigger(struct btree_trans *trans,
 411                                struct btree_insert_entry *i,
 412                                unsigned flags)
 413 {
 414         struct bkey_s_c old = { &i->old_k, i->old_v };
 415         struct bkey_i *new = i->k;
 416         int ret;
 417
 418         verify_update_old_key(trans, i);
 419
 420         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 421                 return 0;
 422
 423         if (!btree_node_type_needs_gc(i->btree_id))
 424                 return 0;
 425
 426         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 427             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 428             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 429                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 430                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 431         } else {
 432                 struct bkey             _deleted = KEY(0, 0, 0);
 433                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 434
 435                 _deleted.p = i->path->pos;
 436
 437                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 438                                 BTREE_TRIGGER_INSERT|flags) ?:
 439                         bch2_mark_key(trans, old, deleted,
 440                                 BTREE_TRIGGER_OVERWRITE|flags);
 441         }
 442
 443         return ret;
 444 }
 445
 446 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 447                                  bool overwrite)
 448 {
 449         /*
 450          * Transactional triggers create new btree_insert_entries, so we can't
 451          * pass them a pointer to a btree_insert_entry, that memory is going to
 452          * move:
 453          */
 454         struct bkey old_k = i->old_k;
 455         struct bkey_s_c old = { &old_k, i->old_v };
 456
 457         verify_update_old_key(trans, i);
 458
 459         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 460             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 461                 return 0;
 462
 463         if (!i->insert_trigger_run &&
 464             !i->overwrite_trigger_run &&
 465             bch2_bkey_ops[old.k->type].trans_trigger ==
 466             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 467             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 468                 i->overwrite_trigger_run = true;
 469                 i->insert_trigger_run = true;
 470                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 471                                            BTREE_TRIGGER_INSERT|
 472                                            BTREE_TRIGGER_OVERWRITE|
 473                                            i->flags) ?: 1;
 474         } else if (overwrite && !i->overwrite_trigger_run) {
 475                 i->overwrite_trigger_run = true;
 476                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 477         } else if (!overwrite && !i->insert_trigger_run) {
 478                 i->insert_trigger_run = true;
 479                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 480         } else {
 481                 return 0;
 482         }
 483 }
 484
 485 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 486                               struct btree_insert_entry *btree_id_start)
 487 {
 488         struct btree_insert_entry *i;
 489         bool trans_trigger_run;
 490         int ret, overwrite;
 491
 492         for (overwrite = 1; overwrite >= 0; --overwrite) {
 493
 494                 /*
 495                  * Running triggers will append more updates to the list of updates as
 496                  * we're walking it:
 497                  */
 498                 do {
 499                         trans_trigger_run = false;
 500
 501                         for (i = btree_id_start;
 502                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 503                              i++) {
 504                                 if (i->btree_id != btree_id)
 505                                         continue;
 506
 507                                 ret = run_one_trans_trigger(trans, i, overwrite);
 508                                 if (ret < 0)
 509                                         return ret;
 510                                 if (ret)
 511                                         trans_trigger_run = true;
 512                         }
 513                 } while (trans_trigger_run);
 514         }
 515
 516         return 0;
 517 }
 518
 519 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 520 {
 521         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 522         unsigned btree_id = 0;
 523         int ret = 0;
 524
 525         /*
 526          *
 527          * For a given btree, this algorithm runs insert triggers before
 528          * overwrite triggers: this is so that when extents are being moved
 529          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 530          * they are re-added.
 531          */
 532         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 533                 if (btree_id == BTREE_ID_alloc)
 534                         continue;
 535
 536                 while (btree_id_start < trans->updates + trans->nr_updates &&
 537                        btree_id_start->btree_id < btree_id)
 538                         btree_id_start++;
 539
 540                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 541                 if (ret)
 542                         return ret;
 543         }
 544
 545         trans_for_each_update(trans, i) {
 546                 if (i->btree_id > BTREE_ID_alloc)
 547                         break;
 548                 if (i->btree_id == BTREE_ID_alloc) {
 549                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 550                         if (ret)
 551                                 return ret;
 552                         break;
 553                 }
 554         }
 555
 556 #ifdef CONFIG_BCACHEFS_DEBUG
 557         trans_for_each_update(trans, i)
 558                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 559                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 560                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 561 #endif
 562         return 0;
 563 }
 564
 565 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 566 {
 567         struct bch_fs *c = trans->c;
 568         struct btree_insert_entry *i;
 569         int ret = 0;
 570
 571         trans_for_each_update(trans, i) {
 572                 /*
 573                  * XXX: synchronization of cached update triggers with gc
 574                  * XXX: synchronization of interior node updates with gc
 575                  */
 576                 BUG_ON(i->cached || i->level);
 577
 578                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 579                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 580                         if (ret)
 581                                 break;
 582                 }
 583         }
 584
 585         return ret;
 586 }
 587
 588 static inline int
 589 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 590                                struct btree_insert_entry **stopped_at,
 591                                unsigned long trace_ip)
 592 {
 593         struct bch_fs *c = trans->c;
 594         struct btree_insert_entry *i;
 595         struct btree_trans_commit_hook *h;
 596         unsigned u64s = 0;
 597         bool marking = false;
 598         int ret;
 599
 600         if (race_fault()) {
 601                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 602                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 603         }
 604
 605         /*
 606          * Check if the insert will fit in the leaf node with the write lock
 607          * held, otherwise another thread could write the node changing the
 608          * amount of space available:
 609          */
 610
 611         prefetch(&trans->c->journal.flags);
 612
 613         h = trans->hooks;
 614         while (h) {
 615                 ret = h->fn(trans, h);
 616                 if (ret)
 617                         return ret;
 618                 h = h->next;
 619         }
 620
 621         trans_for_each_update(trans, i) {
 622                 /* Multiple inserts might go to same leaf: */
 623                 if (!same_leaf_as_prev(trans, i))
 624                         u64s = 0;
 625
 626                 u64s += i->k->k.u64s;
 627                 ret = !i->cached
 628                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 629                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 630                 if (ret) {
 631                         *stopped_at = i;
 632                         return ret;
 633                 }
 634
 635                 if (btree_node_type_needs_gc(i->bkey_type))
 636                         marking = true;
 637         }
 638
 639         /*
 640          * Don't get journal reservation until after we know insert will
 641          * succeed:
 642          */
 643         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 644                 ret = bch2_trans_journal_res_get(trans,
 645                                 (flags & JOURNAL_WATERMARK_MASK)|
 646                                 JOURNAL_RES_GET_NONBLOCK);
 647                 if (ret)
 648                         return ret;
 649
 650                 if (unlikely(trans->journal_transaction_names))
 651                         journal_transaction_name(trans);
 652         } else {
 653                 trans->journal_res.seq = c->journal.replay_journal_seq;
 654         }
 655
 656         /*
 657          * Not allowed to fail after we've gotten our journal reservation - we
 658          * have to use it:
 659          */
 660
 661         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 662             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 663                 if (bch2_journal_seq_verify)
 664                         trans_for_each_update(trans, i)
 665                                 i->k->k.version.lo = trans->journal_res.seq;
 666                 else if (bch2_inject_invalid_keys)
 667                         trans_for_each_update(trans, i)
 668                                 i->k->k.version = MAX_VERSION;
 669         }
 670
 671         if (trans->fs_usage_deltas &&
 672             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 673                 return -BCH_ERR_btree_insert_need_mark_replicas;
 674
 675         trans_for_each_update(trans, i)
 676                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 677                         ret = run_one_mem_trigger(trans, i, i->flags);
 678                         if (ret)
 679                                 return ret;
 680                 }
 681
 682         if (unlikely(c->gc_pos.phase)) {
 683                 ret = bch2_trans_commit_run_gc_triggers(trans);
 684                 if  (ret)
 685                         return ret;
 686         }
 687
 688         if (unlikely(trans->extra_journal_entries.nr)) {
 689                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 690                                   trans->extra_journal_entries.data,
 691                                   trans->extra_journal_entries.nr);
 692
 693                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 694                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 695         }
 696
 697         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 698                 trans_for_each_update(trans, i) {
 699                         struct journal *j = &c->journal;
 700                         struct jset_entry *entry;
 701
 702                         if (i->key_cache_already_flushed)
 703                                 continue;
 704
 705                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 706                                 continue;
 707
 708                         verify_update_old_key(trans, i);
 709
 710                         if (trans->journal_transaction_names) {
 711                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 712                                                        BCH_JSET_ENTRY_overwrite,
 713                                                        i->btree_id, i->level,
 714                                                        i->old_k.u64s);
 715                                 bkey_reassemble(&entry->start[0],
 716                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 717                         }
 718
 719                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 720                                                BCH_JSET_ENTRY_btree_keys,
 721                                                i->btree_id, i->level,
 722                                                i->k->k.u64s);
 723                         bkey_copy(&entry->start[0], i->k);
 724                 }
 725
 726                 if (trans->journal_seq)
 727                         *trans->journal_seq = trans->journal_res.seq;
 728         }
 729
 730         trans_for_each_update(trans, i) {
 731                 i->k->k.needs_whiteout = false;
 732
 733                 if (!i->cached)
 734                         btree_insert_key_leaf(trans, i);
 735                 else if (!i->key_cache_already_flushed)
 736                         bch2_btree_insert_key_cached(trans, flags, i->path, i->k);
 737                 else {
 738                         bch2_btree_key_cache_drop(trans, i->path);
 739                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 740                 }
 741         }
 742
 743         return ret;
 744 }
 745
 746 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 747 {
 748         while (--i >= trans->updates) {
 749                 if (same_leaf_as_prev(trans, i))
 750                         continue;
 751
 752                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 753         }
 754
 755         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 756         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 757 }
 758
 759 static inline int trans_lock_write(struct btree_trans *trans)
 760 {
 761         struct btree_insert_entry *i;
 762
 763         trans_for_each_update(trans, i) {
 764                 if (same_leaf_as_prev(trans, i))
 765                         continue;
 766
 767                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 768                         return trans_lock_write_fail(trans, i);
 769
 770                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 771         }
 772
 773         return 0;
 774 }
 775
 776 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 777 {
 778         struct btree_insert_entry *i;
 779
 780         trans_for_each_update(trans, i)
 781                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 782 }
 783
 784 #ifdef CONFIG_BCACHEFS_DEBUG
 785 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 786                                                    struct btree_insert_entry *i,
 787                                                    struct printbuf *err)
 788 {
 789         struct bch_fs *c = trans->c;
 790         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 791
 792         printbuf_reset(err);
 793         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 794                    trans->fn, (void *) i->ip_allocated);
 795         prt_newline(err);
 796         printbuf_indent_add(err, 2);
 797
 798         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 799         prt_newline(err);
 800
 801         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 802                           i->bkey_type, rw, err);
 803         bch2_print_string_as_lines(KERN_ERR, err->buf);
 804
 805         bch2_inconsistent_error(c);
 806         bch2_dump_trans_updates(trans);
 807         printbuf_exit(err);
 808
 809         return -EINVAL;
 810 }
 811 #endif
 812
 813 /*
 814  * Get journal reservation, take write locks, and attempt to do btree update(s):
 815  */
 816 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 817                                        struct btree_insert_entry **stopped_at,
 818                                        unsigned long trace_ip)
 819 {
 820         struct bch_fs *c = trans->c;
 821         struct btree_insert_entry *i;
 822         struct printbuf buf = PRINTBUF;
 823         int ret, u64s_delta = 0;
 824
 825 #ifdef CONFIG_BCACHEFS_DEBUG
 826         trans_for_each_update(trans, i) {
 827                 int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 828
 829                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 830                                                i->bkey_type, rw, &buf)))
 831                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 832                 btree_insert_entry_checks(trans, i);
 833         }
 834 #endif
 835         printbuf_exit(&buf);
 836
 837         trans_for_each_update(trans, i) {
 838                 if (i->cached)
 839                         continue;
 840
 841                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 842                 u64s_delta -= i->old_btree_u64s;
 843
 844                 if (!same_leaf_as_next(trans, i)) {
 845                         if (u64s_delta <= 0) {
 846                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 847                                                         i->level, flags);
 848                                 if (unlikely(ret))
 849                                         return ret;
 850                         }
 851
 852                         u64s_delta = 0;
 853                 }
 854         }
 855
 856         ret = bch2_journal_preres_get(&c->journal,
 857                         &trans->journal_preres, trans->journal_preres_u64s,
 858                         (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 859         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 860                 ret = bch2_trans_journal_preres_get_cold(trans,
 861                                                 trans->journal_preres_u64s, trace_ip);
 862         if (unlikely(ret))
 863                 return ret;
 864
 865         ret = trans_lock_write(trans);
 866         if (unlikely(ret))
 867                 return ret;
 868
 869         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 870
 871         if (!ret && unlikely(trans->journal_replay_not_finished))
 872                 bch2_drop_overwrites_from_journal(trans);
 873
 874         trans_for_each_update(trans, i)
 875                 if (!same_leaf_as_prev(trans, i))
 876                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 877                                                         insert_l(i)->b);
 878
 879         if (!ret && trans->journal_pin)
 880                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 881                                      trans->journal_pin, NULL);
 882
 883         /*
 884          * Drop journal reservation after dropping write locks, since dropping
 885          * the journal reservation may kick off a journal write:
 886          */
 887         bch2_journal_res_put(&c->journal, &trans->journal_res);
 888
 889         if (unlikely(ret))
 890                 return ret;
 891
 892         bch2_trans_downgrade(trans);
 893
 894         return 0;
 895 }
 896
 897 static int journal_reclaim_wait_done(struct bch_fs *c)
 898 {
 899         int ret = bch2_journal_error(&c->journal) ?:
 900                 !bch2_btree_key_cache_must_wait(c);
 901
 902         if (!ret)
 903                 journal_reclaim_kick(&c->journal);
 904         return ret;
 905 }
 906
 907 static noinline
 908 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 909                             struct btree_insert_entry *i,
 910                             int ret, unsigned long trace_ip)
 911 {
 912         struct bch_fs *c = trans->c;
 913
 914         switch (ret) {
 915         case -BCH_ERR_btree_insert_btree_node_full:
 916                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 917                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 918                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 919                 break;
 920         case -BCH_ERR_btree_insert_need_mark_replicas:
 921                 bch2_trans_unlock(trans);
 922
 923                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 924                 if (ret)
 925                         break;
 926
 927                 ret = bch2_trans_relock(trans);
 928                 if (ret)
 929                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 930                 break;
 931         case -BCH_ERR_journal_res_get_blocked:
 932                 bch2_trans_unlock(trans);
 933
 934                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 935                     !(flags & JOURNAL_WATERMARK_reserved)) {
 936                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 937                         break;
 938                 }
 939
 940                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 941                 if (ret)
 942                         break;
 943
 944                 ret = bch2_trans_relock(trans);
 945                 if (ret)
 946                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 947                 break;
 948         case -BCH_ERR_btree_insert_need_journal_reclaim:
 949                 bch2_trans_unlock(trans);
 950
 951                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 952
 953                 wait_event_freezable(c->journal.reclaim_wait,
 954                                      (ret = journal_reclaim_wait_done(c)));
 955                 if (ret < 0)
 956                         break;
 957
 958                 ret = bch2_trans_relock(trans);
 959                 if (ret)
 960                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 961                 break;
 962         default:
 963                 BUG_ON(ret >= 0);
 964                 break;
 965         }
 966
 967         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 968
 969         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 970                                 !(flags & BTREE_INSERT_NOWAIT) &&
 971                                 (flags & BTREE_INSERT_NOFAIL), c,
 972                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 973
 974         return ret;
 975 }
 976
 977 static noinline int
 978 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 979 {
 980         struct bch_fs *c = trans->c;
 981         int ret;
 982
 983         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
 984             test_bit(BCH_FS_STARTED, &c->flags))
 985                 return -BCH_ERR_erofs_trans_commit;
 986
 987         bch2_trans_unlock(trans);
 988
 989         ret =   bch2_fs_read_write_early(c) ?:
 990                 bch2_trans_relock(trans);
 991         if (ret)
 992                 return ret;
 993
 994         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 995         return 0;
 996 }
 997
 998 /*
 999  * This is for updates done in the early part of fsck - btree_gc - before we've
1000  * gone RW. we only add the new key to the list of keys for journal replay to
1001  * do.
1002  */
1003 static noinline int
1004 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1005 {
1006         struct bch_fs *c = trans->c;
1007         struct btree_insert_entry *i;
1008         int ret = 0;
1009
1010         trans_for_each_update(trans, i) {
1011                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1012                 if (ret)
1013                         break;
1014         }
1015
1016         return ret;
1017 }
1018
1019 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1020 {
1021         struct bch_fs *c = trans->c;
1022         struct btree_insert_entry *i = NULL;
1023         unsigned u64s;
1024         int ret = 0;
1025
1026         if (!trans->nr_updates &&
1027             !trans->extra_journal_entries.nr)
1028                 goto out_reset;
1029
1030         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1031                 lockdep_assert_held(&c->gc_lock);
1032
1033         ret = bch2_trans_commit_run_triggers(trans);
1034         if (ret)
1035                 goto out_reset;
1036
1037         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1038                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1039                 goto out_reset;
1040         }
1041
1042         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1043             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1044                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1045                 if (ret)
1046                         goto out_reset;
1047         }
1048
1049         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1050
1051         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1052
1053         trans->journal_u64s             = trans->extra_journal_entries.nr;
1054         trans->journal_preres_u64s      = 0;
1055
1056         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1057
1058         if (trans->journal_transaction_names)
1059                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1060
1061         trans_for_each_update(trans, i) {
1062                 EBUG_ON(!i->path->should_be_locked);
1063
1064                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1065                 if (unlikely(ret))
1066                         goto out;
1067
1068                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1069
1070                 if (i->key_cache_already_flushed)
1071                         continue;
1072
1073                 /* we're going to journal the key being updated: */
1074                 u64s = jset_u64s(i->k->k.u64s);
1075                 if (i->cached &&
1076                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1077                         trans->journal_preres_u64s += u64s;
1078
1079                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1080                         continue;
1081
1082                 trans->journal_u64s += u64s;
1083
1084                 /* and we're also going to log the overwrite: */
1085                 if (trans->journal_transaction_names)
1086                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1087         }
1088
1089         if (trans->extra_journal_res) {
1090                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1091                                 trans->extra_journal_res,
1092                                 (flags & BTREE_INSERT_NOFAIL)
1093                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1094                 if (ret)
1095                         goto err;
1096         }
1097 retry:
1098         bch2_trans_verify_not_in_restart(trans);
1099         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1100
1101         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1102
1103         /* make sure we didn't drop or screw up locks: */
1104         bch2_trans_verify_locks(trans);
1105
1106         if (ret)
1107                 goto err;
1108
1109         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1110 out:
1111         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1112
1113         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1114                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1115 out_reset:
1116         bch2_trans_reset_updates(trans);
1117
1118         return ret;
1119 err:
1120         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1121         if (ret)
1122                 goto out;
1123
1124         goto retry;
1125 }
1126
1127 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1128                                           enum btree_id id,
1129                                           struct bpos pos)
1130 {
1131         struct bch_fs *c = trans->c;
1132         struct btree_iter iter;
1133         struct bkey_s_c k;
1134         int ret;
1135
1136         bch2_trans_iter_init(trans, &iter, id, pos,
1137                              BTREE_ITER_NOT_EXTENTS|
1138                              BTREE_ITER_ALL_SNAPSHOTS);
1139         while (1) {
1140                 k = bch2_btree_iter_prev(&iter);
1141                 ret = bkey_err(k);
1142                 if (ret)
1143                         break;
1144
1145                 if (!k.k)
1146                         break;
1147
1148                 if (!bkey_eq(pos, k.k->p))
1149                         break;
1150
1151                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1152                         ret = 1;
1153                         break;
1154                 }
1155         }
1156         bch2_trans_iter_exit(trans, &iter);
1157
1158         return ret;
1159 }
1160
1161 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1162                                           enum btree_id id,
1163                                           struct bpos pos)
1164 {
1165         if (!btree_type_has_snapshots(id) ||
1166             pos.snapshot == U32_MAX ||
1167             !snapshot_t(trans->c, pos.snapshot)->children[0])
1168                 return 0;
1169
1170         return __check_pos_snapshot_overwritten(trans, id, pos);
1171 }
1172
1173 static noinline int extent_front_merge(struct btree_trans *trans,
1174                                        struct btree_iter *iter,
1175                                        struct bkey_s_c k,
1176                                        struct bkey_i **insert,
1177                                        enum btree_update_flags flags)
1178 {
1179         struct bch_fs *c = trans->c;
1180         struct bkey_i *update;
1181         int ret;
1182
1183         update = bch2_bkey_make_mut(trans, k);
1184         ret = PTR_ERR_OR_ZERO(update);
1185         if (ret)
1186                 return ret;
1187
1188         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1189                 return 0;
1190
1191         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1192                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1193         if (ret < 0)
1194                 return ret;
1195         if (ret)
1196                 return 0;
1197
1198         ret = bch2_btree_delete_at(trans, iter, flags);
1199         if (ret)
1200                 return ret;
1201
1202         *insert = update;
1203         return 0;
1204 }
1205
1206 static noinline int extent_back_merge(struct btree_trans *trans,
1207                                       struct btree_iter *iter,
1208                                       struct bkey_i *insert,
1209                                       struct bkey_s_c k)
1210 {
1211         struct bch_fs *c = trans->c;
1212         int ret;
1213
1214         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1215                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1216         if (ret < 0)
1217                 return ret;
1218         if (ret)
1219                 return 0;
1220
1221         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1222         return 0;
1223 }
1224
1225 int bch2_trans_update_extent(struct btree_trans *trans,
1226                              struct btree_iter *orig_iter,
1227                              struct bkey_i *insert,
1228                              enum btree_update_flags flags)
1229 {
1230         struct btree_iter iter, update_iter;
1231         struct bpos start = bkey_start_pos(&insert->k);
1232         struct bkey_i *update;
1233         struct bkey_s_c k;
1234         enum btree_id btree_id = orig_iter->btree_id;
1235         int ret = 0, compressed_sectors;
1236
1237         bch2_trans_iter_init(trans, &iter, btree_id, start,
1238                              BTREE_ITER_INTENT|
1239                              BTREE_ITER_WITH_UPDATES|
1240                              BTREE_ITER_NOT_EXTENTS);
1241         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1242         if ((ret = bkey_err(k)))
1243                 goto err;
1244         if (!k.k)
1245                 goto out;
1246
1247         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1248                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1249                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1250                         if (ret)
1251                                 goto err;
1252                 }
1253
1254                 goto next;
1255         }
1256
1257         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1258                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1259                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1260
1261                 /*
1262                  * If we're going to be splitting a compressed extent, note it
1263                  * so that __bch2_trans_commit() can increase our disk
1264                  * reservation:
1265                  */
1266                 if (((front_split && back_split) ||
1267                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1268                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1269                         trans->extra_journal_res += compressed_sectors;
1270
1271                 if (front_split) {
1272                         update = bch2_bkey_make_mut(trans, k);
1273                         if ((ret = PTR_ERR_OR_ZERO(update)))
1274                                 goto err;
1275
1276                         bch2_cut_back(start, update);
1277
1278                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1279                                              BTREE_ITER_NOT_EXTENTS|
1280                                              BTREE_ITER_ALL_SNAPSHOTS|
1281                                              BTREE_ITER_INTENT);
1282                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1283                                 bch2_trans_update(trans, &update_iter, update,
1284                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1285                                                   flags);
1286                         bch2_trans_iter_exit(trans, &update_iter);
1287
1288                         if (ret)
1289                                 goto err;
1290                 }
1291
1292                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1293                     (front_split || back_split)) {
1294                         update = bch2_bkey_make_mut(trans, k);
1295                         if ((ret = PTR_ERR_OR_ZERO(update)))
1296                                 goto err;
1297
1298                         bch2_cut_front(start, update);
1299                         bch2_cut_back(insert->k.p, update);
1300
1301                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1302                                              BTREE_ITER_NOT_EXTENTS|
1303                                              BTREE_ITER_ALL_SNAPSHOTS|
1304                                              BTREE_ITER_INTENT);
1305                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1306                                 bch2_trans_update(trans, &update_iter, update,
1307                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1308                                                   flags);
1309                         bch2_trans_iter_exit(trans, &update_iter);
1310                         if (ret)
1311                                 goto err;
1312                 }
1313
1314                 if (bkey_le(k.k->p, insert->k.p)) {
1315                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1316                         if ((ret = PTR_ERR_OR_ZERO(update)))
1317                                 goto err;
1318
1319                         bkey_init(&update->k);
1320                         update->k.p = k.k->p;
1321
1322                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1323                                 update->k.p.snapshot = insert->k.p.snapshot;
1324                                 update->k.type = KEY_TYPE_whiteout;
1325                         }
1326
1327                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1328                                              BTREE_ITER_NOT_EXTENTS|
1329                                              BTREE_ITER_INTENT);
1330                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1331                                 bch2_trans_update(trans, &update_iter, update,
1332                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1333                                                   flags);
1334                         bch2_trans_iter_exit(trans, &update_iter);
1335
1336                         if (ret)
1337                                 goto err;
1338                 }
1339
1340                 if (back_split) {
1341                         update = bch2_bkey_make_mut(trans, k);
1342                         if ((ret = PTR_ERR_OR_ZERO(update)))
1343                                 goto err;
1344
1345                         bch2_cut_front(insert->k.p, update);
1346
1347                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1348                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1349                                                   flags);
1350                         if (ret)
1351                                 goto err;
1352                         goto out;
1353                 }
1354 next:
1355                 bch2_btree_iter_advance(&iter);
1356                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1357                 if ((ret = bkey_err(k)))
1358                         goto err;
1359                 if (!k.k)
1360                         goto out;
1361         }
1362
1363         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1364                 ret = extent_back_merge(trans, &iter, insert, k);
1365                 if (ret)
1366                         goto err;
1367         }
1368 out:
1369         if (!bkey_deleted(&insert->k)) {
1370                 /*
1371                  * Rewinding iterators is expensive: get a new one and the one
1372                  * that points to the start of insert will be cloned from:
1373                  */
1374                 bch2_trans_iter_exit(trans, &iter);
1375                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1376                                      BTREE_ITER_NOT_EXTENTS|
1377                                      BTREE_ITER_INTENT);
1378                 ret   = bch2_btree_iter_traverse(&iter) ?:
1379                         bch2_trans_update(trans, &iter, insert, flags);
1380         }
1381 err:
1382         bch2_trans_iter_exit(trans, &iter);
1383
1384         return ret;
1385 }
1386
1387 /*
1388  * When deleting, check if we need to emit a whiteout (because we're overwriting
1389  * something in an ancestor snapshot)
1390  */
1391 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1392                                       enum btree_id btree_id, struct bpos pos)
1393 {
1394         struct btree_iter iter;
1395         struct bkey_s_c k;
1396         u32 snapshot = pos.snapshot;
1397         int ret;
1398
1399         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1400                 return 0;
1401
1402         pos.snapshot++;
1403
1404         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1405                            BTREE_ITER_ALL_SNAPSHOTS|
1406                            BTREE_ITER_NOPRESERVE, k, ret) {
1407                 if (!bkey_eq(k.k->p, pos))
1408                         break;
1409
1410                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1411                                               k.k->p.snapshot)) {
1412                         ret = !bkey_whiteout(k.k);
1413                         break;
1414                 }
1415         }
1416         bch2_trans_iter_exit(trans, &iter);
1417
1418         return ret;
1419 }
1420
1421 static int __must_check
1422 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1423                                 struct bkey_i *k, enum btree_update_flags flags,
1424                                 unsigned long ip);
1425
1426 static noinline int flush_new_cached_update(struct btree_trans *trans,
1427                                             struct btree_path *path,
1428                                             struct btree_insert_entry *i,
1429                                             enum btree_update_flags flags,
1430                                             unsigned long ip)
1431 {
1432         struct btree_path *btree_path;
1433         int ret;
1434
1435         i->key_cache_already_flushed = true;
1436         i->flags |= BTREE_TRIGGER_NORUN;
1437
1438         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1439                                    BTREE_ITER_INTENT, _THIS_IP_);
1440
1441         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1442         if (ret)
1443                 goto err;
1444
1445         btree_path_set_should_be_locked(btree_path);
1446         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1447 err:
1448         bch2_path_put(trans, btree_path, true);
1449         return ret;
1450 }
1451
1452 static int __must_check
1453 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1454                                 struct bkey_i *k, enum btree_update_flags flags,
1455                                 unsigned long ip)
1456 {
1457         struct bch_fs *c = trans->c;
1458         struct btree_insert_entry *i, n;
1459         int cmp;
1460
1461         EBUG_ON(!path->should_be_locked);
1462         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1463         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1464
1465         n = (struct btree_insert_entry) {
1466                 .flags          = flags,
1467                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1468                 .btree_id       = path->btree_id,
1469                 .level          = path->level,
1470                 .cached         = path->cached,
1471                 .path           = path,
1472                 .k              = k,
1473                 .ip_allocated   = ip,
1474         };
1475
1476 #ifdef CONFIG_BCACHEFS_DEBUG
1477         trans_for_each_update(trans, i)
1478                 BUG_ON(i != trans->updates &&
1479                        btree_insert_entry_cmp(i - 1, i) >= 0);
1480 #endif
1481
1482         /*
1483          * Pending updates are kept sorted: first, find position of new update,
1484          * then delete/trim any updates the new update overwrites:
1485          */
1486         trans_for_each_update(trans, i) {
1487                 cmp = btree_insert_entry_cmp(&n, i);
1488                 if (cmp <= 0)
1489                         break;
1490         }
1491
1492         if (!cmp && i < trans->updates + trans->nr_updates) {
1493                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1494
1495                 bch2_path_put(trans, i->path, true);
1496                 i->flags        = n.flags;
1497                 i->cached       = n.cached;
1498                 i->k            = n.k;
1499                 i->path         = n.path;
1500                 i->ip_allocated = n.ip_allocated;
1501         } else {
1502                 array_insert_item(trans->updates, trans->nr_updates,
1503                                   i - trans->updates, n);
1504
1505                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1506                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1507
1508                 if (unlikely(trans->journal_replay_not_finished)) {
1509                         struct bkey_i *j_k =
1510                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1511
1512                         if (j_k) {
1513                                 i->old_k = j_k->k;
1514                                 i->old_v = &j_k->v;
1515                         }
1516                 }
1517         }
1518
1519         __btree_path_get(i->path, true);
1520
1521         /*
1522          * If a key is present in the key cache, it must also exist in the
1523          * btree - this is necessary for cache coherency. When iterating over
1524          * a btree that's cached in the key cache, the btree iter code checks
1525          * the key cache - but the key has to exist in the btree for that to
1526          * work:
1527          */
1528         if (path->cached &&
1529             bkey_deleted(&i->old_k) &&
1530             !(flags & BTREE_UPDATE_NO_KEY_CACHE_COHERENCY))
1531                 return flush_new_cached_update(trans, path, i, flags, ip);
1532
1533         return 0;
1534 }
1535
1536 static inline int __must_check
1537 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1538                           struct bkey_i *k, enum btree_update_flags flags)
1539 {
1540         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1541 }
1542
1543 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1544                                    struct bkey_i *k, enum btree_update_flags flags)
1545 {
1546         struct btree_path *path = iter->update_path ?: iter->path;
1547         struct bkey_cached *ck;
1548         int ret;
1549
1550         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1551                 return bch2_trans_update_extent(trans, iter, k, flags);
1552
1553         if (bkey_deleted(&k->k) &&
1554             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1555             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1556                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1557                 if (unlikely(ret < 0))
1558                         return ret;
1559
1560                 if (ret)
1561                         k->k.type = KEY_TYPE_whiteout;
1562         }
1563
1564         /*
1565          * Ensure that updates to cached btrees go to the key cache:
1566          */
1567         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1568             !path->cached &&
1569             !path->level &&
1570             btree_id_cached(trans->c, path->btree_id)) {
1571                 if (!iter->key_cache_path ||
1572                     !iter->key_cache_path->should_be_locked ||
1573                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1574                         if (!iter->key_cache_path)
1575                                 iter->key_cache_path =
1576                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1577                                                       BTREE_ITER_INTENT|
1578                                                       BTREE_ITER_CACHED, _THIS_IP_);
1579
1580                         iter->key_cache_path =
1581                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1582                                                         iter->flags & BTREE_ITER_INTENT,
1583                                                         _THIS_IP_);
1584
1585                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1586                                                        BTREE_ITER_CACHED);
1587                         if (unlikely(ret))
1588                                 return ret;
1589
1590                         ck = (void *) iter->key_cache_path->l[0].b;
1591
1592                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1593                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1594                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1595                         }
1596
1597                         btree_path_set_should_be_locked(iter->key_cache_path);
1598                 }
1599
1600                 path = iter->key_cache_path;
1601         }
1602
1603         return bch2_trans_update_by_path(trans, path, k, flags);
1604 }
1605
1606 void bch2_trans_commit_hook(struct btree_trans *trans,
1607                             struct btree_trans_commit_hook *h)
1608 {
1609         h->next = trans->hooks;
1610         trans->hooks = h;
1611 }
1612
1613 int __bch2_btree_insert(struct btree_trans *trans,
1614                         enum btree_id id, struct bkey_i *k)
1615 {
1616         struct btree_iter iter;
1617         int ret;
1618
1619         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1620                              BTREE_ITER_INTENT);
1621         ret   = bch2_btree_iter_traverse(&iter) ?:
1622                 bch2_trans_update(trans, &iter, k, 0);
1623         bch2_trans_iter_exit(trans, &iter);
1624         return ret;
1625 }
1626
1627 /**
1628  * bch2_btree_insert - insert keys into the extent btree
1629  * @c:                  pointer to struct bch_fs
1630  * @id:                 btree to insert into
1631  * @insert_keys:        list of keys to insert
1632  * @hook:               insert callback
1633  */
1634 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1635                       struct bkey_i *k,
1636                       struct disk_reservation *disk_res,
1637                       u64 *journal_seq, int flags)
1638 {
1639         return bch2_trans_do(c, disk_res, journal_seq, flags,
1640                              __bch2_btree_insert(&trans, id, k));
1641 }
1642
1643 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1644                                 unsigned len, unsigned update_flags)
1645 {
1646         struct bkey_i *k;
1647
1648         k = bch2_trans_kmalloc(trans, sizeof(*k));
1649         if (IS_ERR(k))
1650                 return PTR_ERR(k);
1651
1652         bkey_init(&k->k);
1653         k->k.p = iter->pos;
1654         bch2_key_resize(&k->k, len);
1655         return bch2_trans_update(trans, iter, k, update_flags);
1656 }
1657
1658 int bch2_btree_delete_at(struct btree_trans *trans,
1659                          struct btree_iter *iter, unsigned update_flags)
1660 {
1661         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1662 }
1663
1664 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1665                                   struct bpos start, struct bpos end,
1666                                   unsigned update_flags,
1667                                   u64 *journal_seq)
1668 {
1669         u32 restart_count = trans->restart_count;
1670         struct btree_iter iter;
1671         struct bkey_s_c k;
1672         int ret = 0;
1673
1674         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1675         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1676                 struct disk_reservation disk_res =
1677                         bch2_disk_reservation_init(trans->c, 0);
1678                 struct bkey_i delete;
1679
1680                 ret = bkey_err(k);
1681                 if (ret)
1682                         goto err;
1683
1684                 bkey_init(&delete.k);
1685
1686                 /*
1687                  * This could probably be more efficient for extents:
1688                  */
1689
1690                 /*
1691                  * For extents, iter.pos won't necessarily be the same as
1692                  * bkey_start_pos(k.k) (for non extents they always will be the
1693                  * same). It's important that we delete starting from iter.pos
1694                  * because the range we want to delete could start in the middle
1695                  * of k.
1696                  *
1697                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1698                  * bkey_start_pos(k.k)).
1699                  */
1700                 delete.k.p = iter.pos;
1701
1702                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1703                         bch2_key_resize(&delete.k,
1704                                         bpos_min(end, k.k->p).offset -
1705                                         iter.pos.offset);
1706
1707                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1708                         bch2_trans_commit(trans, &disk_res, journal_seq,
1709                                           BTREE_INSERT_NOFAIL);
1710                 bch2_disk_reservation_put(trans->c, &disk_res);
1711 err:
1712                 /*
1713                  * the bch2_trans_begin() call is in a weird place because we
1714                  * need to call it after every transaction commit, to avoid path
1715                  * overflow, but don't want to call it if the delete operation
1716                  * is a no-op and we have no work to do:
1717                  */
1718                 bch2_trans_begin(trans);
1719
1720                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1721                         ret = 0;
1722                 if (ret)
1723                         break;
1724         }
1725         bch2_trans_iter_exit(trans, &iter);
1726
1727         if (!ret && trans_was_restarted(trans, restart_count))
1728                 ret = -BCH_ERR_transaction_restart_nested;
1729         return ret;
1730 }
1731
1732 /*
1733  * bch_btree_delete_range - delete everything within a given range
1734  *
1735  * Range is a half open interval - [start, end)
1736  */
1737 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1738                             struct bpos start, struct bpos end,
1739                             unsigned update_flags,
1740                             u64 *journal_seq)
1741 {
1742         int ret = bch2_trans_run(c,
1743                         bch2_btree_delete_range_trans(&trans, id, start, end,
1744                                                       update_flags, journal_seq));
1745         if (ret == -BCH_ERR_transaction_restart_nested)
1746                 ret = 0;
1747         return ret;
1748 }
1749
1750 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1751 {
1752         struct printbuf buf = PRINTBUF;
1753         struct jset_entry_log *l;
1754         unsigned u64s;
1755         int ret;
1756
1757         prt_vprintf(&buf, fmt, args);
1758         ret = buf.allocation_failure ? -ENOMEM : 0;
1759         if (ret)
1760                 goto err;
1761
1762         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
1763
1764         ret = darray_make_room(entries, jset_u64s(u64s));
1765         if (ret)
1766                 goto err;
1767
1768         l = (void *) &darray_top(*entries);
1769         l->entry.u64s           = cpu_to_le16(u64s);
1770         l->entry.btree_id       = 0;
1771         l->entry.level          = 1;
1772         l->entry.type           = BCH_JSET_ENTRY_log;
1773         l->entry.pad[0]         = 0;
1774         l->entry.pad[1]         = 0;
1775         l->entry.pad[2]         = 0;
1776         memcpy(l->d, buf.buf, buf.pos);
1777         while (buf.pos & 7)
1778                 l->d[buf.pos++] = '\0';
1779
1780         entries->nr += jset_u64s(u64s);
1781 err:
1782         printbuf_exit(&buf);
1783         return ret;
1784 }
1785
1786 int bch2_trans_log_msg(struct btree_trans *trans, const char *fmt, ...)
1787 {
1788         va_list args;
1789         int ret;
1790
1791         va_start(args, fmt);
1792         ret = __bch2_trans_log_msg(&trans->extra_journal_entries, fmt, args);
1793         va_end(args);
1794
1795         return ret;
1796 }
1797
1798 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
1799 {
1800         va_list args;
1801         int ret;
1802
1803         va_start(args, fmt);
1804
1805         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
1806                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
1807         } else {
1808                 ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1809                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
1810         }
1811
1812         va_end(args);
1813
1814         return ret;
1815
1816 }