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 flags,
 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,
 320                         trans->journal_preres_u64s,
 321                         (flags & JOURNAL_WATERMARK_MASK));
 322         if (ret)
 323                 return ret;
 324
 325         ret = bch2_trans_relock(trans);
 326         if (ret) {
 327                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 328                 return ret;
 329         }
 330
 331         return 0;
 332 }
 333
 334 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 335                                                       unsigned flags)
 336 {
 337         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 338                                     trans->journal_u64s, flags);
 339 }
 340
 341 #define JSET_ENTRY_LOG_U64s             4
 342
 343 static noinline void journal_transaction_name(struct btree_trans *trans)
 344 {
 345         struct bch_fs *c = trans->c;
 346         struct journal *j = &c->journal;
 347         struct jset_entry *entry =
 348                 bch2_journal_add_entry(j, &trans->journal_res,
 349                                        BCH_JSET_ENTRY_log, 0, 0,
 350                                        JSET_ENTRY_LOG_U64s);
 351         struct jset_entry_log *l =
 352                 container_of(entry, struct jset_entry_log, entry);
 353
 354         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 355 }
 356
 357 static inline int btree_key_can_insert(struct btree_trans *trans,
 358                                        struct btree *b, unsigned u64s)
 359 {
 360         struct bch_fs *c = trans->c;
 361
 362         if (!bch2_btree_node_insert_fits(c, b, u64s))
 363                 return -BCH_ERR_btree_insert_btree_node_full;
 364
 365         return 0;
 366 }
 367
 368 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 369                                        struct btree_path *path, unsigned u64s)
 370 {
 371         struct bch_fs *c = trans->c;
 372         struct bkey_cached *ck = (void *) path->l[0].b;
 373         struct btree_insert_entry *i;
 374         unsigned new_u64s;
 375         struct bkey_i *new_k;
 376
 377         EBUG_ON(path->level);
 378
 379         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 380             bch2_btree_key_cache_must_wait(c) &&
 381             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 382                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 383
 384         /*
 385          * bch2_varint_decode can read past the end of the buffer by at most 7
 386          * bytes (it won't be used):
 387          */
 388         u64s += 1;
 389
 390         if (u64s <= ck->u64s)
 391                 return 0;
 392
 393         new_u64s        = roundup_pow_of_two(u64s);
 394         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 395         if (!new_k) {
 396                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 397                         bch2_btree_ids[path->btree_id], new_u64s);
 398                 return -ENOMEM;
 399         }
 400
 401         trans_for_each_update(trans, i)
 402                 if (i->old_v == &ck->k->v)
 403                         i->old_v = &new_k->v;
 404
 405         ck->u64s        = new_u64s;
 406         ck->k           = new_k;
 407         return 0;
 408 }
 409
 410 /* Triggers: */
 411
 412 static int run_one_mem_trigger(struct btree_trans *trans,
 413                                struct btree_insert_entry *i,
 414                                unsigned flags)
 415 {
 416         struct bkey_s_c old = { &i->old_k, i->old_v };
 417         struct bkey_i *new = i->k;
 418         int ret;
 419
 420         verify_update_old_key(trans, i);
 421
 422         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 423                 return 0;
 424
 425         if (!btree_node_type_needs_gc(i->btree_id))
 426                 return 0;
 427
 428         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 429             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 430             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 431                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 432                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 433         } else {
 434                 struct bkey             _deleted = KEY(0, 0, 0);
 435                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 436
 437                 _deleted.p = i->path->pos;
 438
 439                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 440                                 BTREE_TRIGGER_INSERT|flags) ?:
 441                         bch2_mark_key(trans, old, deleted,
 442                                 BTREE_TRIGGER_OVERWRITE|flags);
 443         }
 444
 445         return ret;
 446 }
 447
 448 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 449                                  bool overwrite)
 450 {
 451         /*
 452          * Transactional triggers create new btree_insert_entries, so we can't
 453          * pass them a pointer to a btree_insert_entry, that memory is going to
 454          * move:
 455          */
 456         struct bkey old_k = i->old_k;
 457         struct bkey_s_c old = { &old_k, i->old_v };
 458
 459         verify_update_old_key(trans, i);
 460
 461         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 462             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 463                 return 0;
 464
 465         if (!i->insert_trigger_run &&
 466             !i->overwrite_trigger_run &&
 467             bch2_bkey_ops[old.k->type].trans_trigger ==
 468             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 469             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 470                 i->overwrite_trigger_run = true;
 471                 i->insert_trigger_run = true;
 472                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 473                                            BTREE_TRIGGER_INSERT|
 474                                            BTREE_TRIGGER_OVERWRITE|
 475                                            i->flags) ?: 1;
 476         } else if (overwrite && !i->overwrite_trigger_run) {
 477                 i->overwrite_trigger_run = true;
 478                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 479         } else if (!overwrite && !i->insert_trigger_run) {
 480                 i->insert_trigger_run = true;
 481                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 482         } else {
 483                 return 0;
 484         }
 485 }
 486
 487 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 488                               struct btree_insert_entry *btree_id_start)
 489 {
 490         struct btree_insert_entry *i;
 491         bool trans_trigger_run;
 492         int ret, overwrite;
 493
 494         for (overwrite = 1; overwrite >= 0; --overwrite) {
 495
 496                 /*
 497                  * Running triggers will append more updates to the list of updates as
 498                  * we're walking it:
 499                  */
 500                 do {
 501                         trans_trigger_run = false;
 502
 503                         for (i = btree_id_start;
 504                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 505                              i++) {
 506                                 if (i->btree_id != btree_id)
 507                                         continue;
 508
 509                                 ret = run_one_trans_trigger(trans, i, overwrite);
 510                                 if (ret < 0)
 511                                         return ret;
 512                                 if (ret)
 513                                         trans_trigger_run = true;
 514                         }
 515                 } while (trans_trigger_run);
 516         }
 517
 518         return 0;
 519 }
 520
 521 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 522 {
 523         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 524         unsigned btree_id = 0;
 525         int ret = 0;
 526
 527         /*
 528          *
 529          * For a given btree, this algorithm runs insert triggers before
 530          * overwrite triggers: this is so that when extents are being moved
 531          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 532          * they are re-added.
 533          */
 534         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 535                 if (btree_id == BTREE_ID_alloc)
 536                         continue;
 537
 538                 while (btree_id_start < trans->updates + trans->nr_updates &&
 539                        btree_id_start->btree_id < btree_id)
 540                         btree_id_start++;
 541
 542                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 543                 if (ret)
 544                         return ret;
 545         }
 546
 547         trans_for_each_update(trans, i) {
 548                 if (i->btree_id > BTREE_ID_alloc)
 549                         break;
 550                 if (i->btree_id == BTREE_ID_alloc) {
 551                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 552                         if (ret)
 553                                 return ret;
 554                         break;
 555                 }
 556         }
 557
 558 #ifdef CONFIG_BCACHEFS_DEBUG
 559         trans_for_each_update(trans, i)
 560                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 561                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 562                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 563 #endif
 564         return 0;
 565 }
 566
 567 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 568 {
 569         struct bch_fs *c = trans->c;
 570         struct btree_insert_entry *i;
 571         int ret = 0;
 572
 573         trans_for_each_update(trans, i) {
 574                 /*
 575                  * XXX: synchronization of cached update triggers with gc
 576                  * XXX: synchronization of interior node updates with gc
 577                  */
 578                 BUG_ON(i->cached || i->level);
 579
 580                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 581                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 582                         if (ret)
 583                                 break;
 584                 }
 585         }
 586
 587         return ret;
 588 }
 589
 590 static inline int
 591 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 592                                struct btree_insert_entry **stopped_at,
 593                                unsigned long trace_ip)
 594 {
 595         struct bch_fs *c = trans->c;
 596         struct btree_insert_entry *i;
 597         struct btree_trans_commit_hook *h;
 598         unsigned u64s = 0;
 599         bool marking = false;
 600         int ret;
 601
 602         if (race_fault()) {
 603                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 604                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 605         }
 606
 607         /*
 608          * Check if the insert will fit in the leaf node with the write lock
 609          * held, otherwise another thread could write the node changing the
 610          * amount of space available:
 611          */
 612
 613         prefetch(&trans->c->journal.flags);
 614
 615         h = trans->hooks;
 616         while (h) {
 617                 ret = h->fn(trans, h);
 618                 if (ret)
 619                         return ret;
 620                 h = h->next;
 621         }
 622
 623         trans_for_each_update(trans, i) {
 624                 /* Multiple inserts might go to same leaf: */
 625                 if (!same_leaf_as_prev(trans, i))
 626                         u64s = 0;
 627
 628                 u64s += i->k->k.u64s;
 629                 ret = !i->cached
 630                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 631                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 632                 if (ret) {
 633                         *stopped_at = i;
 634                         return ret;
 635                 }
 636
 637                 if (btree_node_type_needs_gc(i->bkey_type))
 638                         marking = true;
 639         }
 640
 641         /*
 642          * Don't get journal reservation until after we know insert will
 643          * succeed:
 644          */
 645         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 646                 ret = bch2_trans_journal_res_get(trans,
 647                                 (flags & JOURNAL_WATERMARK_MASK)|
 648                                 JOURNAL_RES_GET_NONBLOCK);
 649                 if (ret)
 650                         return ret;
 651
 652                 if (unlikely(trans->journal_transaction_names))
 653                         journal_transaction_name(trans);
 654         } else {
 655                 trans->journal_res.seq = c->journal.replay_journal_seq;
 656         }
 657
 658         /*
 659          * Not allowed to fail after we've gotten our journal reservation - we
 660          * have to use it:
 661          */
 662
 663         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 664             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 665                 if (bch2_journal_seq_verify)
 666                         trans_for_each_update(trans, i)
 667                                 i->k->k.version.lo = trans->journal_res.seq;
 668                 else if (bch2_inject_invalid_keys)
 669                         trans_for_each_update(trans, i)
 670                                 i->k->k.version = MAX_VERSION;
 671         }
 672
 673         if (trans->fs_usage_deltas &&
 674             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 675                 return -BCH_ERR_btree_insert_need_mark_replicas;
 676
 677         trans_for_each_update(trans, i)
 678                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 679                         ret = run_one_mem_trigger(trans, i, i->flags);
 680                         if (ret)
 681                                 return ret;
 682                 }
 683
 684         if (unlikely(c->gc_pos.phase)) {
 685                 ret = bch2_trans_commit_run_gc_triggers(trans);
 686                 if  (ret)
 687                         return ret;
 688         }
 689
 690         if (unlikely(trans->extra_journal_entries.nr)) {
 691                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 692                                   trans->extra_journal_entries.data,
 693                                   trans->extra_journal_entries.nr);
 694
 695                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 696                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 697         }
 698
 699         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 700                 trans_for_each_update(trans, i) {
 701                         struct journal *j = &c->journal;
 702                         struct jset_entry *entry;
 703
 704                         if (i->key_cache_already_flushed)
 705                                 continue;
 706
 707                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 708                                 continue;
 709
 710                         verify_update_old_key(trans, i);
 711
 712                         if (trans->journal_transaction_names) {
 713                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 714                                                        BCH_JSET_ENTRY_overwrite,
 715                                                        i->btree_id, i->level,
 716                                                        i->old_k.u64s);
 717                                 bkey_reassemble(&entry->start[0],
 718                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 719                         }
 720
 721                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 722                                                BCH_JSET_ENTRY_btree_keys,
 723                                                i->btree_id, i->level,
 724                                                i->k->k.u64s);
 725                         bkey_copy(&entry->start[0], i->k);
 726                 }
 727
 728                 if (trans->journal_seq)
 729                         *trans->journal_seq = trans->journal_res.seq;
 730         }
 731
 732         trans_for_each_update(trans, i) {
 733                 i->k->k.needs_whiteout = false;
 734
 735                 if (!i->cached)
 736                         btree_insert_key_leaf(trans, i);
 737                 else if (!i->key_cache_already_flushed)
 738                         bch2_btree_insert_key_cached(trans, flags, i->path, i->k);
 739                 else {
 740                         bch2_btree_key_cache_drop(trans, i->path);
 741                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 742                 }
 743         }
 744
 745         return ret;
 746 }
 747
 748 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 749 {
 750         while (--i >= trans->updates) {
 751                 if (same_leaf_as_prev(trans, i))
 752                         continue;
 753
 754                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 755         }
 756
 757         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 758         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 759 }
 760
 761 static inline int trans_lock_write(struct btree_trans *trans)
 762 {
 763         struct btree_insert_entry *i;
 764
 765         trans_for_each_update(trans, i) {
 766                 if (same_leaf_as_prev(trans, i))
 767                         continue;
 768
 769                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 770                         return trans_lock_write_fail(trans, i);
 771
 772                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 773         }
 774
 775         return 0;
 776 }
 777
 778 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 779 {
 780         struct btree_insert_entry *i;
 781
 782         trans_for_each_update(trans, i)
 783                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 784 }
 785
 786 #ifdef CONFIG_BCACHEFS_DEBUG
 787 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 788                                                    struct btree_insert_entry *i,
 789                                                    struct printbuf *err)
 790 {
 791         struct bch_fs *c = trans->c;
 792         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 793
 794         printbuf_reset(err);
 795         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 796                    trans->fn, (void *) i->ip_allocated);
 797         prt_newline(err);
 798         printbuf_indent_add(err, 2);
 799
 800         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 801         prt_newline(err);
 802
 803         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 804                           i->bkey_type, rw, err);
 805         bch2_print_string_as_lines(KERN_ERR, err->buf);
 806
 807         bch2_inconsistent_error(c);
 808         bch2_dump_trans_updates(trans);
 809         printbuf_exit(err);
 810
 811         return -EINVAL;
 812 }
 813 #endif
 814
 815 /*
 816  * Get journal reservation, take write locks, and attempt to do btree update(s):
 817  */
 818 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 819                                        struct btree_insert_entry **stopped_at,
 820                                        unsigned long trace_ip)
 821 {
 822         struct bch_fs *c = trans->c;
 823         struct btree_insert_entry *i;
 824         struct printbuf buf = PRINTBUF;
 825         int ret, u64s_delta = 0;
 826
 827 #ifdef CONFIG_BCACHEFS_DEBUG
 828         trans_for_each_update(trans, i) {
 829                 int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 830
 831                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 832                                                i->bkey_type, rw, &buf)))
 833                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 834                 btree_insert_entry_checks(trans, i);
 835         }
 836 #endif
 837         printbuf_exit(&buf);
 838
 839         trans_for_each_update(trans, i) {
 840                 if (i->cached)
 841                         continue;
 842
 843                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 844                 u64s_delta -= i->old_btree_u64s;
 845
 846                 if (!same_leaf_as_next(trans, i)) {
 847                         if (u64s_delta <= 0) {
 848                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 849                                                         i->level, flags);
 850                                 if (unlikely(ret))
 851                                         return ret;
 852                         }
 853
 854                         u64s_delta = 0;
 855                 }
 856         }
 857
 858         ret = bch2_journal_preres_get(&c->journal,
 859                         &trans->journal_preres, trans->journal_preres_u64s,
 860                         (flags & JOURNAL_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 861         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 862                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 863         if (unlikely(ret))
 864                 return ret;
 865
 866         ret = trans_lock_write(trans);
 867         if (unlikely(ret))
 868                 return ret;
 869
 870         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 871
 872         if (!ret && unlikely(trans->journal_replay_not_finished))
 873                 bch2_drop_overwrites_from_journal(trans);
 874
 875         trans_for_each_update(trans, i)
 876                 if (!same_leaf_as_prev(trans, i))
 877                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 878                                                         insert_l(i)->b);
 879
 880         if (!ret && trans->journal_pin)
 881                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 882                                      trans->journal_pin, NULL);
 883
 884         /*
 885          * Drop journal reservation after dropping write locks, since dropping
 886          * the journal reservation may kick off a journal write:
 887          */
 888         bch2_journal_res_put(&c->journal, &trans->journal_res);
 889
 890         if (unlikely(ret))
 891                 return ret;
 892
 893         bch2_trans_downgrade(trans);
 894
 895         return 0;
 896 }
 897
 898 static int journal_reclaim_wait_done(struct bch_fs *c)
 899 {
 900         int ret = bch2_journal_error(&c->journal) ?:
 901                 !bch2_btree_key_cache_must_wait(c);
 902
 903         if (!ret)
 904                 journal_reclaim_kick(&c->journal);
 905         return ret;
 906 }
 907
 908 static noinline
 909 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 910                             struct btree_insert_entry *i,
 911                             int ret, unsigned long trace_ip)
 912 {
 913         struct bch_fs *c = trans->c;
 914
 915         switch (ret) {
 916         case -BCH_ERR_btree_insert_btree_node_full:
 917                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 918                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 919                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 920                 break;
 921         case -BCH_ERR_btree_insert_need_mark_replicas:
 922                 bch2_trans_unlock(trans);
 923
 924                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 925                 if (ret)
 926                         break;
 927
 928                 ret = bch2_trans_relock(trans);
 929                 if (ret)
 930                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 931                 break;
 932         case -BCH_ERR_journal_res_get_blocked:
 933                 bch2_trans_unlock(trans);
 934
 935                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 936                     !(flags & JOURNAL_WATERMARK_reserved)) {
 937                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 938                         break;
 939                 }
 940
 941                 ret = bch2_trans_journal_res_get(trans,
 942                                         (flags & JOURNAL_WATERMARK_MASK)|
 943                                         JOURNAL_RES_GET_CHECK);
 944                 if (ret)
 945                         break;
 946
 947                 ret = bch2_trans_relock(trans);
 948                 if (ret)
 949                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 950                 break;
 951         case -BCH_ERR_btree_insert_need_journal_reclaim:
 952                 bch2_trans_unlock(trans);
 953
 954                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 955
 956                 wait_event_freezable(c->journal.reclaim_wait,
 957                                      (ret = journal_reclaim_wait_done(c)));
 958                 if (ret < 0)
 959                         break;
 960
 961                 ret = bch2_trans_relock(trans);
 962                 if (ret)
 963                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 964                 break;
 965         default:
 966                 BUG_ON(ret >= 0);
 967                 break;
 968         }
 969
 970         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 971
 972         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 973                                 !(flags & BTREE_INSERT_NOWAIT) &&
 974                                 (flags & BTREE_INSERT_NOFAIL), c,
 975                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 976
 977         return ret;
 978 }
 979
 980 static noinline int
 981 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 982 {
 983         struct bch_fs *c = trans->c;
 984         int ret;
 985
 986         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
 987             test_bit(BCH_FS_STARTED, &c->flags))
 988                 return -BCH_ERR_erofs_trans_commit;
 989
 990         bch2_trans_unlock(trans);
 991
 992         ret =   bch2_fs_read_write_early(c) ?:
 993                 bch2_trans_relock(trans);
 994         if (ret)
 995                 return ret;
 996
 997         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 998         return 0;
 999 }
1000
1001 /*
1002  * This is for updates done in the early part of fsck - btree_gc - before we've
1003  * gone RW. we only add the new key to the list of keys for journal replay to
1004  * do.
1005  */
1006 static noinline int
1007 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1008 {
1009         struct bch_fs *c = trans->c;
1010         struct btree_insert_entry *i;
1011         int ret = 0;
1012
1013         trans_for_each_update(trans, i) {
1014                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1015                 if (ret)
1016                         break;
1017         }
1018
1019         return ret;
1020 }
1021
1022 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1023 {
1024         struct bch_fs *c = trans->c;
1025         struct btree_insert_entry *i = NULL;
1026         unsigned u64s;
1027         int ret = 0;
1028
1029         if (!trans->nr_updates &&
1030             !trans->extra_journal_entries.nr)
1031                 goto out_reset;
1032
1033         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1034                 lockdep_assert_held(&c->gc_lock);
1035
1036         ret = bch2_trans_commit_run_triggers(trans);
1037         if (ret)
1038                 goto out_reset;
1039
1040         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1041                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1042                 goto out_reset;
1043         }
1044
1045         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1046             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1047                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1048                 if (ret)
1049                         goto out_reset;
1050         }
1051
1052         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1053
1054         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1055
1056         trans->journal_u64s             = trans->extra_journal_entries.nr;
1057         trans->journal_preres_u64s      = 0;
1058
1059         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1060
1061         if (trans->journal_transaction_names)
1062                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1063
1064         trans_for_each_update(trans, i) {
1065                 EBUG_ON(!i->path->should_be_locked);
1066
1067                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1068                 if (unlikely(ret))
1069                         goto out;
1070
1071                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1072
1073                 if (i->key_cache_already_flushed)
1074                         continue;
1075
1076                 /* we're going to journal the key being updated: */
1077                 u64s = jset_u64s(i->k->k.u64s);
1078                 if (i->cached &&
1079                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1080                         trans->journal_preres_u64s += u64s;
1081
1082                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1083                         continue;
1084
1085                 trans->journal_u64s += u64s;
1086
1087                 /* and we're also going to log the overwrite: */
1088                 if (trans->journal_transaction_names)
1089                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1090         }
1091
1092         if (trans->extra_journal_res) {
1093                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1094                                 trans->extra_journal_res,
1095                                 (flags & BTREE_INSERT_NOFAIL)
1096                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1097                 if (ret)
1098                         goto err;
1099         }
1100 retry:
1101         bch2_trans_verify_not_in_restart(trans);
1102         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1103
1104         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1105
1106         /* make sure we didn't drop or screw up locks: */
1107         bch2_trans_verify_locks(trans);
1108
1109         if (ret)
1110                 goto err;
1111
1112         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1113 out:
1114         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1115
1116         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1117                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1118 out_reset:
1119         bch2_trans_reset_updates(trans);
1120
1121         return ret;
1122 err:
1123         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1124         if (ret)
1125                 goto out;
1126
1127         goto retry;
1128 }
1129
1130 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1131                                           enum btree_id id,
1132                                           struct bpos pos)
1133 {
1134         struct bch_fs *c = trans->c;
1135         struct btree_iter iter;
1136         struct bkey_s_c k;
1137         int ret;
1138
1139         bch2_trans_iter_init(trans, &iter, id, pos,
1140                              BTREE_ITER_NOT_EXTENTS|
1141                              BTREE_ITER_ALL_SNAPSHOTS);
1142         while (1) {
1143                 k = bch2_btree_iter_prev(&iter);
1144                 ret = bkey_err(k);
1145                 if (ret)
1146                         break;
1147
1148                 if (!k.k)
1149                         break;
1150
1151                 if (!bkey_eq(pos, k.k->p))
1152                         break;
1153
1154                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1155                         ret = 1;
1156                         break;
1157                 }
1158         }
1159         bch2_trans_iter_exit(trans, &iter);
1160
1161         return ret;
1162 }
1163
1164 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1165                                           enum btree_id id,
1166                                           struct bpos pos)
1167 {
1168         if (!btree_type_has_snapshots(id) ||
1169             pos.snapshot == U32_MAX ||
1170             !snapshot_t(trans->c, pos.snapshot)->children[0])
1171                 return 0;
1172
1173         return __check_pos_snapshot_overwritten(trans, id, pos);
1174 }
1175
1176 static noinline int extent_front_merge(struct btree_trans *trans,
1177                                        struct btree_iter *iter,
1178                                        struct bkey_s_c k,
1179                                        struct bkey_i **insert,
1180                                        enum btree_update_flags flags)
1181 {
1182         struct bch_fs *c = trans->c;
1183         struct bkey_i *update;
1184         int ret;
1185
1186         update = bch2_bkey_make_mut(trans, k);
1187         ret = PTR_ERR_OR_ZERO(update);
1188         if (ret)
1189                 return ret;
1190
1191         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1192                 return 0;
1193
1194         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1195                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1196         if (ret < 0)
1197                 return ret;
1198         if (ret)
1199                 return 0;
1200
1201         ret = bch2_btree_delete_at(trans, iter, flags);
1202         if (ret)
1203                 return ret;
1204
1205         *insert = update;
1206         return 0;
1207 }
1208
1209 static noinline int extent_back_merge(struct btree_trans *trans,
1210                                       struct btree_iter *iter,
1211                                       struct bkey_i *insert,
1212                                       struct bkey_s_c k)
1213 {
1214         struct bch_fs *c = trans->c;
1215         int ret;
1216
1217         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1218                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1219         if (ret < 0)
1220                 return ret;
1221         if (ret)
1222                 return 0;
1223
1224         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1225         return 0;
1226 }
1227
1228 int bch2_trans_update_extent(struct btree_trans *trans,
1229                              struct btree_iter *orig_iter,
1230                              struct bkey_i *insert,
1231                              enum btree_update_flags flags)
1232 {
1233         struct btree_iter iter, update_iter;
1234         struct bpos start = bkey_start_pos(&insert->k);
1235         struct bkey_i *update;
1236         struct bkey_s_c k;
1237         enum btree_id btree_id = orig_iter->btree_id;
1238         int ret = 0, compressed_sectors;
1239
1240         bch2_trans_iter_init(trans, &iter, btree_id, start,
1241                              BTREE_ITER_INTENT|
1242                              BTREE_ITER_WITH_UPDATES|
1243                              BTREE_ITER_NOT_EXTENTS);
1244         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1245         if ((ret = bkey_err(k)))
1246                 goto err;
1247         if (!k.k)
1248                 goto out;
1249
1250         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1251                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1252                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1253                         if (ret)
1254                                 goto err;
1255                 }
1256
1257                 goto next;
1258         }
1259
1260         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1261                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1262                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1263
1264                 /*
1265                  * If we're going to be splitting a compressed extent, note it
1266                  * so that __bch2_trans_commit() can increase our disk
1267                  * reservation:
1268                  */
1269                 if (((front_split && back_split) ||
1270                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1271                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1272                         trans->extra_journal_res += compressed_sectors;
1273
1274                 if (front_split) {
1275                         update = bch2_bkey_make_mut(trans, k);
1276                         if ((ret = PTR_ERR_OR_ZERO(update)))
1277                                 goto err;
1278
1279                         bch2_cut_back(start, update);
1280
1281                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1282                                              BTREE_ITER_NOT_EXTENTS|
1283                                              BTREE_ITER_ALL_SNAPSHOTS|
1284                                              BTREE_ITER_INTENT);
1285                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1286                                 bch2_trans_update(trans, &update_iter, update,
1287                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1288                                                   flags);
1289                         bch2_trans_iter_exit(trans, &update_iter);
1290
1291                         if (ret)
1292                                 goto err;
1293                 }
1294
1295                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1296                     (front_split || back_split)) {
1297                         update = bch2_bkey_make_mut(trans, k);
1298                         if ((ret = PTR_ERR_OR_ZERO(update)))
1299                                 goto err;
1300
1301                         bch2_cut_front(start, update);
1302                         bch2_cut_back(insert->k.p, update);
1303
1304                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1305                                              BTREE_ITER_NOT_EXTENTS|
1306                                              BTREE_ITER_ALL_SNAPSHOTS|
1307                                              BTREE_ITER_INTENT);
1308                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1309                                 bch2_trans_update(trans, &update_iter, update,
1310                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1311                                                   flags);
1312                         bch2_trans_iter_exit(trans, &update_iter);
1313                         if (ret)
1314                                 goto err;
1315                 }
1316
1317                 if (bkey_le(k.k->p, insert->k.p)) {
1318                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1319                         if ((ret = PTR_ERR_OR_ZERO(update)))
1320                                 goto err;
1321
1322                         bkey_init(&update->k);
1323                         update->k.p = k.k->p;
1324
1325                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1326                                 update->k.p.snapshot = insert->k.p.snapshot;
1327                                 update->k.type = KEY_TYPE_whiteout;
1328                         }
1329
1330                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1331                                              BTREE_ITER_NOT_EXTENTS|
1332                                              BTREE_ITER_INTENT);
1333                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1334                                 bch2_trans_update(trans, &update_iter, update,
1335                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1336                                                   flags);
1337                         bch2_trans_iter_exit(trans, &update_iter);
1338
1339                         if (ret)
1340                                 goto err;
1341                 }
1342
1343                 if (back_split) {
1344                         update = bch2_bkey_make_mut(trans, k);
1345                         if ((ret = PTR_ERR_OR_ZERO(update)))
1346                                 goto err;
1347
1348                         bch2_cut_front(insert->k.p, update);
1349
1350                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1351                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1352                                                   flags);
1353                         if (ret)
1354                                 goto err;
1355                         goto out;
1356                 }
1357 next:
1358                 bch2_btree_iter_advance(&iter);
1359                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1360                 if ((ret = bkey_err(k)))
1361                         goto err;
1362                 if (!k.k)
1363                         goto out;
1364         }
1365
1366         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1367                 ret = extent_back_merge(trans, &iter, insert, k);
1368                 if (ret)
1369                         goto err;
1370         }
1371 out:
1372         if (!bkey_deleted(&insert->k)) {
1373                 /*
1374                  * Rewinding iterators is expensive: get a new one and the one
1375                  * that points to the start of insert will be cloned from:
1376                  */
1377                 bch2_trans_iter_exit(trans, &iter);
1378                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1379                                      BTREE_ITER_NOT_EXTENTS|
1380                                      BTREE_ITER_INTENT);
1381                 ret   = bch2_btree_iter_traverse(&iter) ?:
1382                         bch2_trans_update(trans, &iter, insert, flags);
1383         }
1384 err:
1385         bch2_trans_iter_exit(trans, &iter);
1386
1387         return ret;
1388 }
1389
1390 /*
1391  * When deleting, check if we need to emit a whiteout (because we're overwriting
1392  * something in an ancestor snapshot)
1393  */
1394 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1395                                       enum btree_id btree_id, struct bpos pos)
1396 {
1397         struct btree_iter iter;
1398         struct bkey_s_c k;
1399         u32 snapshot = pos.snapshot;
1400         int ret;
1401
1402         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1403                 return 0;
1404
1405         pos.snapshot++;
1406
1407         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1408                            BTREE_ITER_ALL_SNAPSHOTS|
1409                            BTREE_ITER_NOPRESERVE, k, ret) {
1410                 if (!bkey_eq(k.k->p, pos))
1411                         break;
1412
1413                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1414                                               k.k->p.snapshot)) {
1415                         ret = !bkey_whiteout(k.k);
1416                         break;
1417                 }
1418         }
1419         bch2_trans_iter_exit(trans, &iter);
1420
1421         return ret;
1422 }
1423
1424 static int __must_check
1425 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1426                                 struct bkey_i *k, enum btree_update_flags flags,
1427                                 unsigned long ip);
1428
1429 static noinline int flush_new_cached_update(struct btree_trans *trans,
1430                                             struct btree_path *path,
1431                                             struct btree_insert_entry *i,
1432                                             enum btree_update_flags flags,
1433                                             unsigned long ip)
1434 {
1435         struct btree_path *btree_path;
1436         int ret;
1437
1438         i->key_cache_already_flushed = true;
1439         i->flags |= BTREE_TRIGGER_NORUN;
1440
1441         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1442                                    BTREE_ITER_INTENT, _THIS_IP_);
1443
1444         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1445         if (ret)
1446                 goto err;
1447
1448         btree_path_set_should_be_locked(btree_path);
1449         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1450 err:
1451         bch2_path_put(trans, btree_path, true);
1452         return ret;
1453 }
1454
1455 static int __must_check
1456 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1457                                 struct bkey_i *k, enum btree_update_flags flags,
1458                                 unsigned long ip)
1459 {
1460         struct bch_fs *c = trans->c;
1461         struct btree_insert_entry *i, n;
1462         int cmp;
1463
1464         EBUG_ON(!path->should_be_locked);
1465         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1466         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1467
1468         n = (struct btree_insert_entry) {
1469                 .flags          = flags,
1470                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1471                 .btree_id       = path->btree_id,
1472                 .level          = path->level,
1473                 .cached         = path->cached,
1474                 .path           = path,
1475                 .k              = k,
1476                 .ip_allocated   = ip,
1477         };
1478
1479 #ifdef CONFIG_BCACHEFS_DEBUG
1480         trans_for_each_update(trans, i)
1481                 BUG_ON(i != trans->updates &&
1482                        btree_insert_entry_cmp(i - 1, i) >= 0);
1483 #endif
1484
1485         /*
1486          * Pending updates are kept sorted: first, find position of new update,
1487          * then delete/trim any updates the new update overwrites:
1488          */
1489         trans_for_each_update(trans, i) {
1490                 cmp = btree_insert_entry_cmp(&n, i);
1491                 if (cmp <= 0)
1492                         break;
1493         }
1494
1495         if (!cmp && i < trans->updates + trans->nr_updates) {
1496                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1497
1498                 bch2_path_put(trans, i->path, true);
1499                 i->flags        = n.flags;
1500                 i->cached       = n.cached;
1501                 i->k            = n.k;
1502                 i->path         = n.path;
1503                 i->ip_allocated = n.ip_allocated;
1504         } else {
1505                 array_insert_item(trans->updates, trans->nr_updates,
1506                                   i - trans->updates, n);
1507
1508                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1509                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1510
1511                 if (unlikely(trans->journal_replay_not_finished)) {
1512                         struct bkey_i *j_k =
1513                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1514
1515                         if (j_k) {
1516                                 i->old_k = j_k->k;
1517                                 i->old_v = &j_k->v;
1518                         }
1519                 }
1520         }
1521
1522         __btree_path_get(i->path, true);
1523
1524         /*
1525          * If a key is present in the key cache, it must also exist in the
1526          * btree - this is necessary for cache coherency. When iterating over
1527          * a btree that's cached in the key cache, the btree iter code checks
1528          * the key cache - but the key has to exist in the btree for that to
1529          * work:
1530          */
1531         if (path->cached &&
1532             bkey_deleted(&i->old_k) &&
1533             !(flags & BTREE_UPDATE_NO_KEY_CACHE_COHERENCY))
1534                 return flush_new_cached_update(trans, path, i, flags, ip);
1535
1536         return 0;
1537 }
1538
1539 static inline int __must_check
1540 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1541                           struct bkey_i *k, enum btree_update_flags flags)
1542 {
1543         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1544 }
1545
1546 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1547                                    struct bkey_i *k, enum btree_update_flags flags)
1548 {
1549         struct btree_path *path = iter->update_path ?: iter->path;
1550         struct bkey_cached *ck;
1551         int ret;
1552
1553         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1554                 return bch2_trans_update_extent(trans, iter, k, flags);
1555
1556         if (bkey_deleted(&k->k) &&
1557             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1558             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1559                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1560                 if (unlikely(ret < 0))
1561                         return ret;
1562
1563                 if (ret)
1564                         k->k.type = KEY_TYPE_whiteout;
1565         }
1566
1567         /*
1568          * Ensure that updates to cached btrees go to the key cache:
1569          */
1570         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1571             !path->cached &&
1572             !path->level &&
1573             btree_id_cached(trans->c, path->btree_id)) {
1574                 if (!iter->key_cache_path ||
1575                     !iter->key_cache_path->should_be_locked ||
1576                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1577                         if (!iter->key_cache_path)
1578                                 iter->key_cache_path =
1579                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1580                                                       BTREE_ITER_INTENT|
1581                                                       BTREE_ITER_CACHED, _THIS_IP_);
1582
1583                         iter->key_cache_path =
1584                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1585                                                         iter->flags & BTREE_ITER_INTENT,
1586                                                         _THIS_IP_);
1587
1588                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1589                                                        BTREE_ITER_CACHED);
1590                         if (unlikely(ret))
1591                                 return ret;
1592
1593                         ck = (void *) iter->key_cache_path->l[0].b;
1594
1595                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1596                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1597                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1598                         }
1599
1600                         btree_path_set_should_be_locked(iter->key_cache_path);
1601                 }
1602
1603                 path = iter->key_cache_path;
1604         }
1605
1606         return bch2_trans_update_by_path(trans, path, k, flags);
1607 }
1608
1609 void bch2_trans_commit_hook(struct btree_trans *trans,
1610                             struct btree_trans_commit_hook *h)
1611 {
1612         h->next = trans->hooks;
1613         trans->hooks = h;
1614 }
1615
1616 int __bch2_btree_insert(struct btree_trans *trans,
1617                         enum btree_id id, struct bkey_i *k)
1618 {
1619         struct btree_iter iter;
1620         int ret;
1621
1622         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1623                              BTREE_ITER_INTENT);
1624         ret   = bch2_btree_iter_traverse(&iter) ?:
1625                 bch2_trans_update(trans, &iter, k, 0);
1626         bch2_trans_iter_exit(trans, &iter);
1627         return ret;
1628 }
1629
1630 /**
1631  * bch2_btree_insert - insert keys into the extent btree
1632  * @c:                  pointer to struct bch_fs
1633  * @id:                 btree to insert into
1634  * @insert_keys:        list of keys to insert
1635  * @hook:               insert callback
1636  */
1637 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1638                       struct bkey_i *k,
1639                       struct disk_reservation *disk_res,
1640                       u64 *journal_seq, int flags)
1641 {
1642         return bch2_trans_do(c, disk_res, journal_seq, flags,
1643                              __bch2_btree_insert(&trans, id, k));
1644 }
1645
1646 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1647                                 unsigned len, unsigned update_flags)
1648 {
1649         struct bkey_i *k;
1650
1651         k = bch2_trans_kmalloc(trans, sizeof(*k));
1652         if (IS_ERR(k))
1653                 return PTR_ERR(k);
1654
1655         bkey_init(&k->k);
1656         k->k.p = iter->pos;
1657         bch2_key_resize(&k->k, len);
1658         return bch2_trans_update(trans, iter, k, update_flags);
1659 }
1660
1661 int bch2_btree_delete_at(struct btree_trans *trans,
1662                          struct btree_iter *iter, unsigned update_flags)
1663 {
1664         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1665 }
1666
1667 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1668                                   struct bpos start, struct bpos end,
1669                                   unsigned update_flags,
1670                                   u64 *journal_seq)
1671 {
1672         u32 restart_count = trans->restart_count;
1673         struct btree_iter iter;
1674         struct bkey_s_c k;
1675         int ret = 0;
1676
1677         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1678         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1679                 struct disk_reservation disk_res =
1680                         bch2_disk_reservation_init(trans->c, 0);
1681                 struct bkey_i delete;
1682
1683                 ret = bkey_err(k);
1684                 if (ret)
1685                         goto err;
1686
1687                 bkey_init(&delete.k);
1688
1689                 /*
1690                  * This could probably be more efficient for extents:
1691                  */
1692
1693                 /*
1694                  * For extents, iter.pos won't necessarily be the same as
1695                  * bkey_start_pos(k.k) (for non extents they always will be the
1696                  * same). It's important that we delete starting from iter.pos
1697                  * because the range we want to delete could start in the middle
1698                  * of k.
1699                  *
1700                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1701                  * bkey_start_pos(k.k)).
1702                  */
1703                 delete.k.p = iter.pos;
1704
1705                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1706                         bch2_key_resize(&delete.k,
1707                                         bpos_min(end, k.k->p).offset -
1708                                         iter.pos.offset);
1709
1710                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1711                         bch2_trans_commit(trans, &disk_res, journal_seq,
1712                                           BTREE_INSERT_NOFAIL);
1713                 bch2_disk_reservation_put(trans->c, &disk_res);
1714 err:
1715                 /*
1716                  * the bch2_trans_begin() call is in a weird place because we
1717                  * need to call it after every transaction commit, to avoid path
1718                  * overflow, but don't want to call it if the delete operation
1719                  * is a no-op and we have no work to do:
1720                  */
1721                 bch2_trans_begin(trans);
1722
1723                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1724                         ret = 0;
1725                 if (ret)
1726                         break;
1727         }
1728         bch2_trans_iter_exit(trans, &iter);
1729
1730         if (!ret && trans_was_restarted(trans, restart_count))
1731                 ret = -BCH_ERR_transaction_restart_nested;
1732         return ret;
1733 }
1734
1735 /*
1736  * bch_btree_delete_range - delete everything within a given range
1737  *
1738  * Range is a half open interval - [start, end)
1739  */
1740 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1741                             struct bpos start, struct bpos end,
1742                             unsigned update_flags,
1743                             u64 *journal_seq)
1744 {
1745         int ret = bch2_trans_run(c,
1746                         bch2_btree_delete_range_trans(&trans, id, start, end,
1747                                                       update_flags, journal_seq));
1748         if (ret == -BCH_ERR_transaction_restart_nested)
1749                 ret = 0;
1750         return ret;
1751 }
1752
1753 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1754 {
1755         struct printbuf buf = PRINTBUF;
1756         struct jset_entry_log *l;
1757         unsigned u64s;
1758         int ret;
1759
1760         prt_vprintf(&buf, fmt, args);
1761         ret = buf.allocation_failure ? -ENOMEM : 0;
1762         if (ret)
1763                 goto err;
1764
1765         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
1766
1767         ret = darray_make_room(entries, jset_u64s(u64s));
1768         if (ret)
1769                 goto err;
1770
1771         l = (void *) &darray_top(*entries);
1772         l->entry.u64s           = cpu_to_le16(u64s);
1773         l->entry.btree_id       = 0;
1774         l->entry.level          = 1;
1775         l->entry.type           = BCH_JSET_ENTRY_log;
1776         l->entry.pad[0]         = 0;
1777         l->entry.pad[1]         = 0;
1778         l->entry.pad[2]         = 0;
1779         memcpy(l->d, buf.buf, buf.pos);
1780         while (buf.pos & 7)
1781                 l->d[buf.pos++] = '\0';
1782
1783         entries->nr += jset_u64s(u64s);
1784 err:
1785         printbuf_exit(&buf);
1786         return ret;
1787 }
1788
1789 int bch2_trans_log_msg(struct btree_trans *trans, const char *fmt, ...)
1790 {
1791         va_list args;
1792         int ret;
1793
1794         va_start(args, fmt);
1795         ret = __bch2_trans_log_msg(&trans->extra_journal_entries, fmt, args);
1796         va_end(args);
1797
1798         return ret;
1799 }
1800
1801 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
1802 {
1803         va_list args;
1804         int ret;
1805
1806         va_start(args, fmt);
1807
1808         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
1809                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
1810         } else {
1811                 ret = bch2_trans_do(c, NULL, NULL, BTREE_INSERT_LAZY_RW,
1812                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
1813         }
1814
1815         va_end(args);
1816
1817         return ret;
1818
1819 }