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 static int __must_check
  28 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  29                           struct bkey_i *, enum btree_update_flags);
  30
  31 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  32                                          const struct btree_insert_entry *r)
  33 {
  34         return   cmp_int(l->btree_id,   r->btree_id) ?:
  35                  cmp_int(l->cached,     r->cached) ?:
  36                  -cmp_int(l->level,     r->level) ?:
  37                  bpos_cmp(l->k->k.p,    r->k->k.p);
  38 }
  39
  40 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  41 {
  42         return i->path->l + i->level;
  43 }
  44
  45 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  46                                      struct btree_insert_entry *i)
  47 {
  48         return i != trans->updates &&
  49                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  50 }
  51
  52 static inline bool same_leaf_as_next(struct btree_trans *trans,
  53                                      struct btree_insert_entry *i)
  54 {
  55         return i + 1 < trans->updates + trans->nr_updates &&
  56                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  57 }
  58
  59 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  60                                            struct btree_path *path,
  61                                            struct btree *b)
  62 {
  63         struct bch_fs *c = trans->c;
  64
  65         if (path->cached)
  66                 return;
  67
  68         if (unlikely(btree_node_just_written(b)) &&
  69             bch2_btree_post_write_cleanup(c, b))
  70                 bch2_trans_node_reinit_iter(trans, b);
  71
  72         /*
  73          * If the last bset has been written, or if it's gotten too big - start
  74          * a new bset to insert into:
  75          */
  76         if (want_new_bset(c, b))
  77                 bch2_btree_init_next(trans, b);
  78 }
  79
  80 /* Inserting into a given leaf node (last stage of insert): */
  81
  82 /* Handle overwrites and do insert, for non extents: */
  83 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  84                                 struct btree_path *path,
  85                                 struct btree *b,
  86                                 struct btree_node_iter *node_iter,
  87                                 struct bkey_i *insert)
  88 {
  89         struct bkey_packed *k;
  90         unsigned clobber_u64s = 0, new_u64s = 0;
  91
  92         EBUG_ON(btree_node_just_written(b));
  93         EBUG_ON(bset_written(b, btree_bset_last(b)));
  94         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  95         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  96         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  97         EBUG_ON(insert->k.u64s >
  98                 bch_btree_keys_u64s_remaining(trans->c, b));
  99
 100         k = bch2_btree_node_iter_peek_all(node_iter, b);
 101         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 102                 k = NULL;
 103
 104         /* @k is the key being overwritten/deleted, if any: */
 105         EBUG_ON(k && bkey_deleted(k));
 106
 107         /* Deleting, but not found? nothing to do: */
 108         if (bkey_deleted(&insert->k) && !k)
 109                 return false;
 110
 111         if (bkey_deleted(&insert->k)) {
 112                 /* Deleting: */
 113                 btree_account_key_drop(b, k);
 114                 k->type = KEY_TYPE_deleted;
 115
 116                 if (k->needs_whiteout)
 117                         push_whiteout(trans->c, b, insert->k.p);
 118                 k->needs_whiteout = false;
 119
 120                 if (k >= btree_bset_last(b)->start) {
 121                         clobber_u64s = k->u64s;
 122                         bch2_bset_delete(b, k, clobber_u64s);
 123                         goto fix_iter;
 124                 } else {
 125                         bch2_btree_path_fix_key_modified(trans, b, k);
 126                 }
 127
 128                 return true;
 129         }
 130
 131         if (k) {
 132                 /* Overwriting: */
 133                 btree_account_key_drop(b, k);
 134                 k->type = KEY_TYPE_deleted;
 135
 136                 insert->k.needs_whiteout = k->needs_whiteout;
 137                 k->needs_whiteout = false;
 138
 139                 if (k >= btree_bset_last(b)->start) {
 140                         clobber_u64s = k->u64s;
 141                         goto overwrite;
 142                 } else {
 143                         bch2_btree_path_fix_key_modified(trans, b, k);
 144                 }
 145         }
 146
 147         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 148 overwrite:
 149         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 150         new_u64s = k->u64s;
 151 fix_iter:
 152         if (clobber_u64s != new_u64s)
 153                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 154                                          clobber_u64s, new_u64s);
 155         return true;
 156 }
 157
 158 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 159                                unsigned i, u64 seq)
 160 {
 161         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 162         struct btree_write *w = container_of(pin, struct btree_write, journal);
 163         struct btree *b = container_of(w, struct btree, writes[i]);
 164         struct btree_trans trans;
 165         unsigned long old, new, v;
 166         unsigned idx = w - b->writes;
 167
 168         bch2_trans_init(&trans, c, 0, 0);
 169
 170         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 171         v = READ_ONCE(b->flags);
 172
 173         do {
 174                 old = new = v;
 175
 176                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 177                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 178                     w->journal.seq != seq)
 179                         break;
 180
 181                 new |= 1 << BTREE_NODE_need_write;
 182         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 183
 184         btree_node_write_if_need(c, b, SIX_LOCK_read);
 185         six_unlock_read(&b->c.lock);
 186
 187         bch2_trans_exit(&trans);
 188         return 0;
 189 }
 190
 191 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 192 {
 193         return __btree_node_flush(j, pin, 0, seq);
 194 }
 195
 196 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 197 {
 198         return __btree_node_flush(j, pin, 1, seq);
 199 }
 200
 201 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 202                                        struct btree *b, u64 seq)
 203 {
 204         struct btree_write *w = btree_current_write(b);
 205
 206         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 207                              btree_node_write_idx(b) == 0
 208                              ? btree_node_flush0
 209                              : btree_node_flush1);
 210 }
 211
 212 /**
 213  * btree_insert_key - insert a key one key into a leaf node
 214  */
 215 static void btree_insert_key_leaf(struct btree_trans *trans,
 216                                   struct btree_insert_entry *insert)
 217 {
 218         struct bch_fs *c = trans->c;
 219         struct btree *b = insert_l(insert)->b;
 220         struct bset_tree *t = bset_tree_last(b);
 221         struct bset *i = bset(b, t);
 222         int old_u64s = bset_u64s(t);
 223         int old_live_u64s = b->nr.live_u64s;
 224         int live_u64s_added, u64s_added;
 225
 226         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 227                                         &insert_l(insert)->iter, insert->k)))
 228                 return;
 229
 230         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 231                                          le64_to_cpu(i->journal_seq)));
 232
 233         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 234
 235         if (unlikely(!btree_node_dirty(b)))
 236                 set_btree_node_dirty_acct(c, b);
 237
 238         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 239         u64s_added = (int) bset_u64s(t) - old_u64s;
 240
 241         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 242                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 243         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 244                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 245
 246         if (u64s_added > live_u64s_added &&
 247             bch2_maybe_compact_whiteouts(c, b))
 248                 bch2_trans_node_reinit_iter(trans, b);
 249 }
 250
 251 /* Cached btree updates: */
 252
 253 /* Normal update interface: */
 254
 255 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 256                                              struct btree_insert_entry *i)
 257 {
 258         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 259         BUG_ON(i->cached        != i->path->cached);
 260         BUG_ON(i->level         != i->path->level);
 261         BUG_ON(i->btree_id      != i->path->btree_id);
 262         EBUG_ON(!i->level &&
 263                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 264                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 265                 i->k->k.p.snapshot &&
 266                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 267 }
 268
 269 static noinline int
 270 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 271                                    unsigned long trace_ip)
 272 {
 273         struct bch_fs *c = trans->c;
 274         int ret;
 275
 276         bch2_trans_unlock(trans);
 277
 278         ret = bch2_journal_preres_get(&c->journal,
 279                         &trans->journal_preres, u64s, 0);
 280         if (ret)
 281                 return ret;
 282
 283         ret = bch2_trans_relock(trans);
 284         if (ret) {
 285                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 286                 return ret;
 287         }
 288
 289         return 0;
 290 }
 291
 292 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 293                                              unsigned flags)
 294 {
 295         struct bch_fs *c = trans->c;
 296         int ret;
 297
 298         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 299                                    trans->journal_u64s,
 300                                    flags|
 301                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 302
 303         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 304 }
 305
 306 #define JSET_ENTRY_LOG_U64s             4
 307
 308 static void journal_transaction_name(struct btree_trans *trans)
 309 {
 310         struct bch_fs *c = trans->c;
 311         struct journal *j = &c->journal;
 312         struct jset_entry *entry =
 313                 bch2_journal_add_entry(j, &trans->journal_res,
 314                                        BCH_JSET_ENTRY_log, 0, 0,
 315                                        JSET_ENTRY_LOG_U64s);
 316         struct jset_entry_log *l =
 317                 container_of(entry, struct jset_entry_log, entry);
 318
 319         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 320 }
 321
 322 static inline enum btree_insert_ret
 323 btree_key_can_insert(struct btree_trans *trans,
 324                      struct btree *b,
 325                      unsigned u64s)
 326 {
 327         struct bch_fs *c = trans->c;
 328
 329         if (!bch2_btree_node_insert_fits(c, b, u64s))
 330                 return BTREE_INSERT_BTREE_NODE_FULL;
 331
 332         return BTREE_INSERT_OK;
 333 }
 334
 335 static enum btree_insert_ret
 336 btree_key_can_insert_cached(struct btree_trans *trans,
 337                             struct btree_path *path,
 338                             unsigned u64s)
 339 {
 340         struct bch_fs *c = trans->c;
 341         struct bkey_cached *ck = (void *) path->l[0].b;
 342         unsigned old_u64s = ck->u64s, new_u64s;
 343         struct bkey_i *new_k;
 344
 345         EBUG_ON(path->level);
 346
 347         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 348             bch2_btree_key_cache_must_wait(c) &&
 349             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 350                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 351
 352         /*
 353          * bch2_varint_decode can read past the end of the buffer by at most 7
 354          * bytes (it won't be used):
 355          */
 356         u64s += 1;
 357
 358         if (u64s <= ck->u64s)
 359                 return BTREE_INSERT_OK;
 360
 361         new_u64s        = roundup_pow_of_two(u64s);
 362         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 363         if (!new_k) {
 364                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 365                         bch2_btree_ids[path->btree_id], new_u64s);
 366                 return -ENOMEM;
 367         }
 368
 369         ck->u64s        = new_u64s;
 370         ck->k           = new_k;
 371         /*
 372          * Keys returned by peek() are no longer valid pointers, so we need a
 373          * transaction restart:
 374          */
 375         trace_and_count(c, trans_restart_key_cache_key_realloced, trans, _RET_IP_, path, old_u64s, new_u64s);
 376         return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_key_cache_realloced);
 377 }
 378
 379 /* Triggers: */
 380
 381 static int run_one_mem_trigger(struct btree_trans *trans,
 382                                struct btree_insert_entry *i,
 383                                unsigned flags)
 384 {
 385         struct bkey_s_c old = { &i->old_k, i->old_v };
 386         struct bkey_i *new = i->k;
 387         int ret;
 388
 389         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 390                 return 0;
 391
 392         if (!btree_node_type_needs_gc(i->btree_id))
 393                 return 0;
 394
 395         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 396             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 397             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 398                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 399                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 400         } else {
 401                 struct bkey             _deleted = KEY(0, 0, 0);
 402                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 403
 404                 _deleted.p = i->path->pos;
 405
 406                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 407                                 BTREE_TRIGGER_INSERT|flags) ?:
 408                         bch2_mark_key(trans, old, deleted,
 409                                 BTREE_TRIGGER_OVERWRITE|flags);
 410         }
 411
 412         return ret;
 413 }
 414
 415 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 416                                  bool overwrite)
 417 {
 418         /*
 419          * Transactional triggers create new btree_insert_entries, so we can't
 420          * pass them a pointer to a btree_insert_entry, that memory is going to
 421          * move:
 422          */
 423         struct bkey old_k = i->old_k;
 424         struct bkey_s_c old = { &old_k, i->old_v };
 425
 426         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 427             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 428                 return 0;
 429
 430         if (!i->insert_trigger_run &&
 431             !i->overwrite_trigger_run &&
 432             bch2_bkey_ops[old.k->type].trans_trigger ==
 433             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 434             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 435                 i->overwrite_trigger_run = true;
 436                 i->insert_trigger_run = true;
 437                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 438                                            BTREE_TRIGGER_INSERT|
 439                                            BTREE_TRIGGER_OVERWRITE|
 440                                            i->flags) ?: 1;
 441         } else if (overwrite && !i->overwrite_trigger_run) {
 442                 i->overwrite_trigger_run = true;
 443                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 444         } else if (!overwrite && !i->insert_trigger_run) {
 445                 i->insert_trigger_run = true;
 446                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 447         } else {
 448                 return 0;
 449         }
 450 }
 451
 452 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 453                               struct btree_insert_entry *btree_id_start)
 454 {
 455         struct btree_insert_entry *i;
 456         bool trans_trigger_run;
 457         int ret, overwrite;
 458
 459         for (overwrite = 1; overwrite >= 0; --overwrite) {
 460
 461                 /*
 462                  * Running triggers will append more updates to the list of updates as
 463                  * we're walking it:
 464                  */
 465                 do {
 466                         trans_trigger_run = false;
 467
 468                         for (i = btree_id_start;
 469                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 470                              i++) {
 471                                 if (i->btree_id != btree_id)
 472                                         continue;
 473
 474                                 ret = run_one_trans_trigger(trans, i, overwrite);
 475                                 if (ret < 0)
 476                                         return ret;
 477                                 if (ret)
 478                                         trans_trigger_run = true;
 479                         }
 480                 } while (trans_trigger_run);
 481         }
 482
 483         return 0;
 484 }
 485
 486 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 487 {
 488         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 489         unsigned btree_id = 0;
 490         int ret = 0;
 491
 492         /*
 493          *
 494          * For a given btree, this algorithm runs insert triggers before
 495          * overwrite triggers: this is so that when extents are being moved
 496          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 497          * they are re-added.
 498          */
 499         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 500                 if (btree_id == BTREE_ID_alloc)
 501                         continue;
 502
 503                 while (btree_id_start < trans->updates + trans->nr_updates &&
 504                        btree_id_start->btree_id < btree_id)
 505                         btree_id_start++;
 506
 507                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 508                 if (ret)
 509                         return ret;
 510         }
 511
 512         trans_for_each_update(trans, i) {
 513                 if (i->btree_id > BTREE_ID_alloc)
 514                         break;
 515                 if (i->btree_id == BTREE_ID_alloc) {
 516                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 517                         if (ret)
 518                                 return ret;
 519                         break;
 520                 }
 521         }
 522
 523         trans_for_each_update(trans, i)
 524                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 525                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 526                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 527
 528         return 0;
 529 }
 530
 531 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 532 {
 533         struct bch_fs *c = trans->c;
 534         struct btree_insert_entry *i;
 535         int ret = 0;
 536
 537         trans_for_each_update(trans, i) {
 538                 /*
 539                  * XXX: synchronization of cached update triggers with gc
 540                  * XXX: synchronization of interior node updates with gc
 541                  */
 542                 BUG_ON(i->cached || i->level);
 543
 544                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 545                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 546                         if (ret)
 547                                 break;
 548                 }
 549         }
 550
 551         return ret;
 552 }
 553
 554 static inline int
 555 bch2_trans_commit_write_locked(struct btree_trans *trans,
 556                                struct btree_insert_entry **stopped_at,
 557                                unsigned long trace_ip)
 558 {
 559         struct bch_fs *c = trans->c;
 560         struct btree_insert_entry *i;
 561         struct btree_trans_commit_hook *h;
 562         unsigned u64s = 0;
 563         bool marking = false;
 564         int ret;
 565
 566         if (race_fault()) {
 567                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 568                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 569         }
 570
 571         /*
 572          * Check if the insert will fit in the leaf node with the write lock
 573          * held, otherwise another thread could write the node changing the
 574          * amount of space available:
 575          */
 576
 577         prefetch(&trans->c->journal.flags);
 578
 579         h = trans->hooks;
 580         while (h) {
 581                 ret = h->fn(trans, h);
 582                 if (ret)
 583                         return ret;
 584                 h = h->next;
 585         }
 586
 587         trans_for_each_update(trans, i) {
 588                 /* Multiple inserts might go to same leaf: */
 589                 if (!same_leaf_as_prev(trans, i))
 590                         u64s = 0;
 591
 592                 u64s += i->k->k.u64s;
 593                 ret = !i->cached
 594                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 595                         : btree_key_can_insert_cached(trans, i->path, u64s);
 596                 if (ret) {
 597                         *stopped_at = i;
 598                         return ret;
 599                 }
 600
 601                 if (btree_node_type_needs_gc(i->bkey_type))
 602                         marking = true;
 603
 604                 /*
 605                  * Revalidate before calling mem triggers - XXX, ugly:
 606                  *
 607                  * - successful btree node splits don't cause transaction
 608                  *   restarts and will have invalidated the pointer to the bkey
 609                  *   value
 610                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 611                  *   when it has to resort
 612                  * - btree_key_can_insert_cached() when it has to reallocate
 613                  *
 614                  *   Ugly because we currently have no way to tell if the
 615                  *   pointer's been invalidated, which means it's debatabale
 616                  *   whether we should be stashing the old key at all.
 617                  */
 618                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 619
 620                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 621                         struct bkey_i *j_k =
 622                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 623                                                             i->k->k.p);
 624
 625                         if (j_k) {
 626                                 i->old_k = j_k->k;
 627                                 i->old_v = &j_k->v;
 628                         }
 629                 }
 630         }
 631
 632         /*
 633          * Don't get journal reservation until after we know insert will
 634          * succeed:
 635          */
 636         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 637                 ret = bch2_trans_journal_res_get(trans,
 638                                 JOURNAL_RES_GET_NONBLOCK);
 639                 if (ret)
 640                         return ret;
 641
 642                 journal_transaction_name(trans);
 643         } else {
 644                 trans->journal_res.seq = c->journal.replay_journal_seq;
 645         }
 646
 647         if (unlikely(trans->extra_journal_entries.nr)) {
 648                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 649                                   trans->extra_journal_entries.data,
 650                                   trans->extra_journal_entries.nr);
 651
 652                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 653                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 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 (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 662                 if (bch2_journal_seq_verify)
 663                         trans_for_each_update(trans, i)
 664                                 i->k->k.version.lo = trans->journal_res.seq;
 665                 else if (bch2_inject_invalid_keys)
 666                         trans_for_each_update(trans, i)
 667                                 i->k->k.version = MAX_VERSION;
 668         }
 669
 670         if (trans->fs_usage_deltas &&
 671             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 672                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 673
 674         trans_for_each_update(trans, i)
 675                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 676                         ret = run_one_mem_trigger(trans, i, i->flags);
 677                         if (ret)
 678                                 return ret;
 679                 }
 680
 681         if (unlikely(c->gc_pos.phase)) {
 682                 ret = bch2_trans_commit_run_gc_triggers(trans);
 683                 if  (ret)
 684                         return ret;
 685         }
 686
 687         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 688                 trans_for_each_update(trans, i) {
 689                         struct journal *j = &c->journal;
 690                         struct jset_entry *entry;
 691
 692                         if (i->key_cache_already_flushed)
 693                                 continue;
 694
 695                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 696                                                BCH_JSET_ENTRY_overwrite,
 697                                                i->btree_id, i->level,
 698                                                i->old_k.u64s);
 699                         bkey_reassemble(&entry->start[0],
 700                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 701
 702                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 703                                                BCH_JSET_ENTRY_btree_keys,
 704                                                i->btree_id, i->level,
 705                                                i->k->k.u64s);
 706                         bkey_copy(&entry->start[0], i->k);
 707                 }
 708
 709                 if (trans->journal_seq)
 710                         *trans->journal_seq = trans->journal_res.seq;
 711         }
 712
 713         trans_for_each_update(trans, i) {
 714                 i->k->k.needs_whiteout = false;
 715
 716                 if (!i->cached)
 717                         btree_insert_key_leaf(trans, i);
 718                 else if (!i->key_cache_already_flushed)
 719                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 720                 else {
 721                         bch2_btree_key_cache_drop(trans, i->path);
 722                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 723                 }
 724         }
 725
 726         return ret;
 727 }
 728
 729 static inline int trans_lock_write(struct btree_trans *trans)
 730 {
 731         struct btree_insert_entry *i;
 732         int ret;
 733
 734         trans_for_each_update(trans, i) {
 735                 if (same_leaf_as_prev(trans, i))
 736                         continue;
 737
 738                 ret = bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c);
 739                 if (ret)
 740                         goto fail;
 741
 742                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 743         }
 744
 745         return 0;
 746 fail:
 747         while (--i >= trans->updates) {
 748                 if (same_leaf_as_prev(trans, i))
 749                         continue;
 750
 751                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 752         }
 753
 754         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 755         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 756 }
 757
 758 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 759 {
 760         struct btree_insert_entry *i;
 761
 762         trans_for_each_update(trans, i)
 763                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 764 }
 765
 766 /*
 767  * Get journal reservation, take write locks, and attempt to do btree update(s):
 768  */
 769 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 770                                        struct btree_insert_entry **stopped_at,
 771                                        unsigned long trace_ip)
 772 {
 773         struct bch_fs *c = trans->c;
 774         struct btree_insert_entry *i;
 775         struct printbuf buf = PRINTBUF;
 776         int ret, u64s_delta = 0;
 777         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 778
 779         trans_for_each_update(trans, i) {
 780                 if (bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 781                                       i->bkey_type, rw, &buf)) {
 782                         printbuf_reset(&buf);
 783                         prt_printf(&buf, "invalid bkey on insert from %s -> %ps",
 784                                trans->fn, (void *) i->ip_allocated);
 785                         prt_newline(&buf);
 786                         printbuf_indent_add(&buf, 2);
 787
 788                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
 789                         prt_newline(&buf);
 790
 791                         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 792                                           i->bkey_type, rw, &buf);
 793
 794                         bch2_trans_inconsistent(trans, "%s", buf.buf);
 795                         printbuf_exit(&buf);
 796                         return -EINVAL;
 797                 }
 798                 btree_insert_entry_checks(trans, i);
 799         }
 800
 801         printbuf_exit(&buf);
 802
 803         trans_for_each_update(trans, i) {
 804                 if (i->cached)
 805                         continue;
 806
 807                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 808                 u64s_delta -= i->old_btree_u64s;
 809
 810                 if (!same_leaf_as_next(trans, i)) {
 811                         if (u64s_delta <= 0) {
 812                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 813                                                         i->level, trans->flags);
 814                                 if (unlikely(ret))
 815                                         return ret;
 816                         }
 817
 818                         u64s_delta = 0;
 819                 }
 820         }
 821
 822         ret = bch2_journal_preres_get(&c->journal,
 823                         &trans->journal_preres, trans->journal_preres_u64s,
 824                         JOURNAL_RES_GET_NONBLOCK|
 825                         (trans->flags & JOURNAL_WATERMARK_MASK));
 826         if (unlikely(ret == -EAGAIN))
 827                 ret = bch2_trans_journal_preres_get_cold(trans,
 828                                                 trans->journal_preres_u64s, trace_ip);
 829         if (unlikely(ret))
 830                 return ret;
 831
 832         ret = trans_lock_write(trans);
 833         if (unlikely(ret))
 834                 return ret;
 835
 836         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 837
 838         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 839                 bch2_drop_overwrites_from_journal(trans);
 840
 841         trans_for_each_update(trans, i)
 842                 if (!same_leaf_as_prev(trans, i))
 843                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 844                                                         insert_l(i)->b);
 845
 846         if (!ret && trans->journal_pin)
 847                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 848                                      trans->journal_pin, NULL);
 849
 850         /*
 851          * Drop journal reservation after dropping write locks, since dropping
 852          * the journal reservation may kick off a journal write:
 853          */
 854         bch2_journal_res_put(&c->journal, &trans->journal_res);
 855
 856         if (unlikely(ret))
 857                 return ret;
 858
 859         bch2_trans_downgrade(trans);
 860
 861         return 0;
 862 }
 863
 864 static int journal_reclaim_wait_done(struct bch_fs *c)
 865 {
 866         int ret = bch2_journal_error(&c->journal) ?:
 867                 !bch2_btree_key_cache_must_wait(c);
 868
 869         if (!ret)
 870                 journal_reclaim_kick(&c->journal);
 871         return ret;
 872 }
 873
 874 static noinline
 875 int bch2_trans_commit_error(struct btree_trans *trans,
 876                             struct btree_insert_entry *i,
 877                             int ret, unsigned long trace_ip)
 878 {
 879         struct bch_fs *c = trans->c;
 880
 881         switch (ret) {
 882         case BTREE_INSERT_BTREE_NODE_FULL:
 883                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 884                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 885                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 886                 break;
 887         case BTREE_INSERT_NEED_MARK_REPLICAS:
 888                 bch2_trans_unlock(trans);
 889
 890                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 891                 if (ret)
 892                         break;
 893
 894                 ret = bch2_trans_relock(trans);
 895                 if (ret)
 896                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 897                 break;
 898         case BTREE_INSERT_NEED_JOURNAL_RES:
 899                 bch2_trans_unlock(trans);
 900
 901                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 902                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 903                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 904                         break;
 905                 }
 906
 907                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 908                 if (ret)
 909                         break;
 910
 911                 ret = bch2_trans_relock(trans);
 912                 if (ret)
 913                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 914                 break;
 915         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 916                 bch2_trans_unlock(trans);
 917
 918                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 919
 920                 wait_event_freezable(c->journal.reclaim_wait,
 921                                      (ret = journal_reclaim_wait_done(c)));
 922                 if (ret < 0)
 923                         break;
 924
 925                 ret = bch2_trans_relock(trans);
 926                 if (ret)
 927                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 928                 break;
 929         default:
 930                 BUG_ON(ret >= 0);
 931                 break;
 932         }
 933
 934         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 935
 936         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 937                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 938                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 939                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 940
 941         return ret;
 942 }
 943
 944 static noinline int
 945 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 946 {
 947         struct bch_fs *c = trans->c;
 948         int ret;
 949
 950         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 951             test_bit(BCH_FS_STARTED, &c->flags))
 952                 return -EROFS;
 953
 954         bch2_trans_unlock(trans);
 955
 956         ret =   bch2_fs_read_write_early(c) ?:
 957                 bch2_trans_relock(trans);
 958         if (ret)
 959                 return ret;
 960
 961         percpu_ref_get(&c->writes);
 962         return 0;
 963 }
 964
 965 /*
 966  * This is for updates done in the early part of fsck - btree_gc - before we've
 967  * gone RW. we only add the new key to the list of keys for journal replay to
 968  * do.
 969  */
 970 static noinline int
 971 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 972 {
 973         struct bch_fs *c = trans->c;
 974         struct btree_insert_entry *i;
 975         int ret = 0;
 976
 977         trans_for_each_update(trans, i) {
 978                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 979                 if (ret)
 980                         break;
 981         }
 982
 983         return ret;
 984 }
 985
 986 int __bch2_trans_commit(struct btree_trans *trans)
 987 {
 988         struct bch_fs *c = trans->c;
 989         struct btree_insert_entry *i = NULL;
 990         unsigned u64s;
 991         int ret = 0;
 992
 993         if (!trans->nr_updates &&
 994             !trans->extra_journal_entries.nr)
 995                 goto out_reset;
 996
 997         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 998                 lockdep_assert_held(&c->gc_lock);
 999
1000         ret = bch2_trans_commit_run_triggers(trans);
1001         if (ret)
1002                 goto out_reset;
1003
1004         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1005                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1006                 goto out_reset;
1007         }
1008
1009         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1010             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1011                 ret = bch2_trans_commit_get_rw_cold(trans);
1012                 if (ret)
1013                         goto out_reset;
1014         }
1015
1016         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1017
1018         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1019
1020         trans->journal_u64s             = trans->extra_journal_entries.nr;
1021         trans->journal_preres_u64s      = 0;
1022
1023         /* For journalling transaction name: */
1024         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1025
1026         trans_for_each_update(trans, i) {
1027                 BUG_ON(!i->path->should_be_locked);
1028
1029                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1030                 if (unlikely(ret))
1031                         goto out;
1032
1033                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1034
1035                 if (i->key_cache_already_flushed)
1036                         continue;
1037
1038                 /* we're going to journal the key being updated: */
1039                 u64s = jset_u64s(i->k->k.u64s);
1040                 if (i->cached &&
1041                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1042                         trans->journal_preres_u64s += u64s;
1043                 trans->journal_u64s += u64s;
1044
1045                 /* and we're also going to log the overwrite: */
1046                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1047         }
1048
1049         if (trans->extra_journal_res) {
1050                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1051                                 trans->extra_journal_res,
1052                                 (trans->flags & BTREE_INSERT_NOFAIL)
1053                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1054                 if (ret)
1055                         goto err;
1056         }
1057 retry:
1058         BUG_ON(trans->restarted);
1059         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1060
1061         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1062
1063         /* make sure we didn't drop or screw up locks: */
1064         bch2_trans_verify_locks(trans);
1065
1066         if (ret)
1067                 goto err;
1068
1069         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1070 out:
1071         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1072
1073         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1074                 percpu_ref_put(&c->writes);
1075 out_reset:
1076         bch2_trans_reset_updates(trans);
1077
1078         if (trans->fs_usage_deltas) {
1079                 trans->fs_usage_deltas->used = 0;
1080                 memset((void *) trans->fs_usage_deltas +
1081                        offsetof(struct replicas_delta_list, memset_start), 0,
1082                        (void *) &trans->fs_usage_deltas->memset_end -
1083                        (void *) &trans->fs_usage_deltas->memset_start);
1084         }
1085
1086         return ret;
1087 err:
1088         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1089         if (ret)
1090                 goto out;
1091
1092         goto retry;
1093 }
1094
1095 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1096                                           enum btree_id id,
1097                                           struct bpos pos)
1098 {
1099         struct bch_fs *c = trans->c;
1100         struct btree_iter iter;
1101         struct bkey_s_c k;
1102         int ret;
1103
1104         bch2_trans_iter_init(trans, &iter, id, pos,
1105                              BTREE_ITER_NOT_EXTENTS|
1106                              BTREE_ITER_ALL_SNAPSHOTS);
1107         while (1) {
1108                 k = bch2_btree_iter_prev(&iter);
1109                 ret = bkey_err(k);
1110                 if (ret)
1111                         break;
1112
1113                 if (!k.k)
1114                         break;
1115
1116                 if (bkey_cmp(pos, k.k->p))
1117                         break;
1118
1119                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1120                         ret = 1;
1121                         break;
1122                 }
1123         }
1124         bch2_trans_iter_exit(trans, &iter);
1125
1126         return ret;
1127 }
1128
1129 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1130                                           enum btree_id id,
1131                                           struct bpos pos)
1132 {
1133         if (!btree_type_has_snapshots(id) ||
1134             pos.snapshot == U32_MAX ||
1135             !snapshot_t(trans->c, pos.snapshot)->children[0])
1136                 return 0;
1137
1138         return __check_pos_snapshot_overwritten(trans, id, pos);
1139 }
1140
1141 int bch2_trans_update_extent(struct btree_trans *trans,
1142                              struct btree_iter *orig_iter,
1143                              struct bkey_i *insert,
1144                              enum btree_update_flags flags)
1145 {
1146         struct bch_fs *c = trans->c;
1147         struct btree_iter iter, update_iter;
1148         struct bpos start = bkey_start_pos(&insert->k);
1149         struct bkey_i *update;
1150         struct bkey_s_c k;
1151         enum btree_id btree_id = orig_iter->btree_id;
1152         int ret = 0, compressed_sectors;
1153
1154         bch2_trans_iter_init(trans, &iter, btree_id, start,
1155                              BTREE_ITER_INTENT|
1156                              BTREE_ITER_WITH_UPDATES|
1157                              BTREE_ITER_NOT_EXTENTS);
1158         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1159         if ((ret = bkey_err(k)))
1160                 goto err;
1161         if (!k.k)
1162                 goto out;
1163
1164         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1165                 /*
1166                  * We can't merge extents if they belong to interior snapshot
1167                  * tree nodes, and there's a snapshot in which one extent is
1168                  * visible and the other is not - i.e. if visibility is
1169                  * different.
1170                  *
1171                  * Instead of checking if visibilitiy of the two extents is
1172                  * different, for now we just check if either has been
1173                  * overwritten:
1174                  */
1175                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1176                 if (ret < 0)
1177                         goto err;
1178                 if (ret)
1179                         goto nomerge1;
1180
1181                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1182                 if (ret < 0)
1183                         goto err;
1184                 if (ret)
1185                         goto nomerge1;
1186
1187                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1188                 if ((ret = PTR_ERR_OR_ZERO(update)))
1189                         goto err;
1190
1191                 bkey_reassemble(update, k);
1192
1193                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1194                         ret = bch2_btree_delete_at(trans, &iter, flags);
1195                         if (ret)
1196                                 goto err;
1197
1198                         insert = update;
1199                         goto next;
1200                 }
1201         }
1202 nomerge1:
1203         ret = 0;
1204         if (!bkey_cmp(k.k->p, start))
1205                 goto next;
1206
1207         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1208                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1209                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1210
1211                 /*
1212                  * If we're going to be splitting a compressed extent, note it
1213                  * so that __bch2_trans_commit() can increase our disk
1214                  * reservation:
1215                  */
1216                 if (((front_split && back_split) ||
1217                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1218                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1219                         trans->extra_journal_res += compressed_sectors;
1220
1221                 if (front_split) {
1222                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1223                         if ((ret = PTR_ERR_OR_ZERO(update)))
1224                                 goto err;
1225
1226                         bkey_reassemble(update, k);
1227
1228                         bch2_cut_back(start, update);
1229
1230                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1231                                              BTREE_ITER_NOT_EXTENTS|
1232                                              BTREE_ITER_ALL_SNAPSHOTS|
1233                                              BTREE_ITER_INTENT);
1234                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1235                                 bch2_trans_update(trans, &update_iter, update,
1236                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1237                                                   flags);
1238                         bch2_trans_iter_exit(trans, &update_iter);
1239
1240                         if (ret)
1241                                 goto err;
1242                 }
1243
1244                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1245                     (front_split || back_split)) {
1246                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1247                         if ((ret = PTR_ERR_OR_ZERO(update)))
1248                                 goto err;
1249
1250                         bkey_reassemble(update, k);
1251
1252                         bch2_cut_front(start, update);
1253                         bch2_cut_back(insert->k.p, update);
1254
1255                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1256                                              BTREE_ITER_NOT_EXTENTS|
1257                                              BTREE_ITER_ALL_SNAPSHOTS|
1258                                              BTREE_ITER_INTENT);
1259                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1260                                 bch2_trans_update(trans, &update_iter, update,
1261                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1262                                                   flags);
1263                         bch2_trans_iter_exit(trans, &update_iter);
1264                         if (ret)
1265                                 goto err;
1266                 }
1267
1268                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1269                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1270                         if ((ret = PTR_ERR_OR_ZERO(update)))
1271                                 goto err;
1272
1273                         bkey_init(&update->k);
1274                         update->k.p = k.k->p;
1275
1276                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1277                                 update->k.p.snapshot = insert->k.p.snapshot;
1278                                 update->k.type = KEY_TYPE_whiteout;
1279                         }
1280
1281                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1282                                              BTREE_ITER_NOT_EXTENTS|
1283                                              BTREE_ITER_INTENT);
1284                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1285                                 bch2_trans_update(trans, &update_iter, update,
1286                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1287                                                   flags);
1288                         bch2_trans_iter_exit(trans, &update_iter);
1289
1290                         if (ret)
1291                                 goto err;
1292                 }
1293
1294                 if (back_split) {
1295                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1296                         if ((ret = PTR_ERR_OR_ZERO(update)))
1297                                 goto err;
1298
1299                         bkey_reassemble(update, k);
1300                         bch2_cut_front(insert->k.p, update);
1301
1302                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1303                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1304                                                   flags);
1305                         if (ret)
1306                                 goto err;
1307                         goto out;
1308                 }
1309 next:
1310                 bch2_btree_iter_advance(&iter);
1311                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1312                 if ((ret = bkey_err(k)))
1313                         goto err;
1314                 if (!k.k)
1315                         goto out;
1316         }
1317
1318         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1319                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1320                 if (ret < 0)
1321                         goto err;
1322                 if (ret)
1323                         goto nomerge2;
1324
1325                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1326                 if (ret < 0)
1327                         goto err;
1328                 if (ret)
1329                         goto nomerge2;
1330
1331                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1332         }
1333 nomerge2:
1334         ret = 0;
1335 out:
1336         if (!bkey_deleted(&insert->k)) {
1337                 /*
1338                  * Rewinding iterators is expensive: get a new one and the one
1339                  * that points to the start of insert will be cloned from:
1340                  */
1341                 bch2_trans_iter_exit(trans, &iter);
1342                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1343                                      BTREE_ITER_NOT_EXTENTS|
1344                                      BTREE_ITER_INTENT);
1345                 ret   = bch2_btree_iter_traverse(&iter) ?:
1346                         bch2_trans_update(trans, &iter, insert, flags);
1347         }
1348 err:
1349         bch2_trans_iter_exit(trans, &iter);
1350
1351         return ret;
1352 }
1353
1354 /*
1355  * When deleting, check if we need to emit a whiteout (because we're overwriting
1356  * something in an ancestor snapshot)
1357  */
1358 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1359                                       enum btree_id btree_id, struct bpos pos)
1360 {
1361         struct btree_iter iter;
1362         struct bkey_s_c k;
1363         u32 snapshot = pos.snapshot;
1364         int ret;
1365
1366         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1367                 return 0;
1368
1369         pos.snapshot++;
1370
1371         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1372                            BTREE_ITER_ALL_SNAPSHOTS|
1373                            BTREE_ITER_NOPRESERVE, k, ret) {
1374                 if (bkey_cmp(k.k->p, pos))
1375                         break;
1376
1377                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1378                                               k.k->p.snapshot)) {
1379                         ret = !bkey_whiteout(k.k);
1380                         break;
1381                 }
1382         }
1383         bch2_trans_iter_exit(trans, &iter);
1384
1385         return ret;
1386 }
1387
1388 static int __must_check
1389 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1390                                 struct bkey_i *k, enum btree_update_flags flags,
1391                                 unsigned long ip)
1392 {
1393         struct bch_fs *c = trans->c;
1394         struct btree_insert_entry *i, n;
1395         int ret = 0;
1396
1397         BUG_ON(!path->should_be_locked);
1398
1399         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1400         BUG_ON(bpos_cmp(k->k.p, path->pos));
1401
1402         n = (struct btree_insert_entry) {
1403                 .flags          = flags,
1404                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1405                 .btree_id       = path->btree_id,
1406                 .level          = path->level,
1407                 .cached         = path->cached,
1408                 .path           = path,
1409                 .k              = k,
1410                 .ip_allocated   = ip,
1411         };
1412
1413 #ifdef CONFIG_BCACHEFS_DEBUG
1414         trans_for_each_update(trans, i)
1415                 BUG_ON(i != trans->updates &&
1416                        btree_insert_entry_cmp(i - 1, i) >= 0);
1417 #endif
1418
1419         /*
1420          * Pending updates are kept sorted: first, find position of new update,
1421          * then delete/trim any updates the new update overwrites:
1422          */
1423         trans_for_each_update(trans, i)
1424                 if (btree_insert_entry_cmp(&n, i) <= 0)
1425                         break;
1426
1427         if (i < trans->updates + trans->nr_updates &&
1428             !btree_insert_entry_cmp(&n, i)) {
1429                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1430
1431                 bch2_path_put(trans, i->path, true);
1432                 i->flags        = n.flags;
1433                 i->cached       = n.cached;
1434                 i->k            = n.k;
1435                 i->path         = n.path;
1436                 i->ip_allocated = n.ip_allocated;
1437         } else {
1438                 array_insert_item(trans->updates, trans->nr_updates,
1439                                   i - trans->updates, n);
1440
1441                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1442                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1443
1444                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1445                         struct bkey_i *j_k =
1446                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1447
1448                         if (j_k) {
1449                                 i->old_k = j_k->k;
1450                                 i->old_v = &j_k->v;
1451                         }
1452                 }
1453         }
1454
1455         __btree_path_get(i->path, true);
1456
1457         /*
1458          * If a key is present in the key cache, it must also exist in the
1459          * btree - this is necessary for cache coherency. When iterating over
1460          * a btree that's cached in the key cache, the btree iter code checks
1461          * the key cache - but the key has to exist in the btree for that to
1462          * work:
1463          */
1464         if (path->cached &&
1465             bkey_deleted(&i->old_k)) {
1466                 struct btree_path *btree_path;
1467
1468                 i->key_cache_already_flushed = true;
1469                 i->flags |= BTREE_TRIGGER_NORUN;
1470
1471                 btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1472                                            BTREE_ITER_INTENT, _THIS_IP_);
1473
1474                 ret = bch2_btree_path_traverse(trans, btree_path, 0);
1475                 if (ret)
1476                         goto err;
1477
1478                 btree_path_set_should_be_locked(btree_path);
1479                 ret = bch2_trans_update_by_path_trace(trans, btree_path, k, flags, ip);
1480 err:
1481                 bch2_path_put(trans, btree_path, true);
1482         }
1483
1484         return ret;
1485 }
1486
1487 static int __must_check
1488 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1489                           struct bkey_i *k, enum btree_update_flags flags)
1490 {
1491         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1492 }
1493
1494 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1495                                    struct bkey_i *k, enum btree_update_flags flags)
1496 {
1497         struct btree_path *path = iter->update_path ?: iter->path;
1498         struct bkey_cached *ck;
1499         int ret;
1500
1501         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1502                 return bch2_trans_update_extent(trans, iter, k, flags);
1503
1504         if (bkey_deleted(&k->k) &&
1505             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1506             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1507                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1508                 if (unlikely(ret < 0))
1509                         return ret;
1510
1511                 if (ret)
1512                         k->k.type = KEY_TYPE_whiteout;
1513         }
1514
1515         /*
1516          * Ensure that updates to cached btrees go to the key cache:
1517          */
1518         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1519             !path->cached &&
1520             !path->level &&
1521             btree_id_cached(trans->c, path->btree_id)) {
1522                 if (!iter->key_cache_path ||
1523                     !iter->key_cache_path->should_be_locked ||
1524                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1525                         if (!iter->key_cache_path)
1526                                 iter->key_cache_path =
1527                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1528                                                       BTREE_ITER_INTENT|
1529                                                       BTREE_ITER_CACHED, _THIS_IP_);
1530
1531                         iter->key_cache_path =
1532                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1533                                                         iter->flags & BTREE_ITER_INTENT,
1534                                                         _THIS_IP_);
1535
1536                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1537                                                        BTREE_ITER_CACHED);
1538                         if (unlikely(ret))
1539                                 return ret;
1540
1541                         ck = (void *) iter->key_cache_path->l[0].b;
1542
1543                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1544                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1545                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1546                         }
1547
1548                         btree_path_set_should_be_locked(iter->key_cache_path);
1549                 }
1550
1551                 path = iter->key_cache_path;
1552         }
1553
1554         return bch2_trans_update_by_path(trans, path, k, flags);
1555 }
1556
1557 void bch2_trans_commit_hook(struct btree_trans *trans,
1558                             struct btree_trans_commit_hook *h)
1559 {
1560         h->next = trans->hooks;
1561         trans->hooks = h;
1562 }
1563
1564 int __bch2_btree_insert(struct btree_trans *trans,
1565                         enum btree_id id, struct bkey_i *k)
1566 {
1567         struct btree_iter iter;
1568         int ret;
1569
1570         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1571                              BTREE_ITER_INTENT);
1572         ret   = bch2_btree_iter_traverse(&iter) ?:
1573                 bch2_trans_update(trans, &iter, k, 0);
1574         bch2_trans_iter_exit(trans, &iter);
1575         return ret;
1576 }
1577
1578 /**
1579  * bch2_btree_insert - insert keys into the extent btree
1580  * @c:                  pointer to struct bch_fs
1581  * @id:                 btree to insert into
1582  * @insert_keys:        list of keys to insert
1583  * @hook:               insert callback
1584  */
1585 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1586                       struct bkey_i *k,
1587                       struct disk_reservation *disk_res,
1588                       u64 *journal_seq, int flags)
1589 {
1590         return bch2_trans_do(c, disk_res, journal_seq, flags,
1591                              __bch2_btree_insert(&trans, id, k));
1592 }
1593
1594 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1595                                 unsigned len, unsigned update_flags)
1596 {
1597         struct bkey_i *k;
1598
1599         k = bch2_trans_kmalloc(trans, sizeof(*k));
1600         if (IS_ERR(k))
1601                 return PTR_ERR(k);
1602
1603         bkey_init(&k->k);
1604         k->k.p = iter->pos;
1605         bch2_key_resize(&k->k, len);
1606         return bch2_trans_update(trans, iter, k, update_flags);
1607 }
1608
1609 int bch2_btree_delete_at(struct btree_trans *trans,
1610                          struct btree_iter *iter, unsigned update_flags)
1611 {
1612         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1613 }
1614
1615 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1616                                   struct bpos start, struct bpos end,
1617                                   unsigned update_flags,
1618                                   u64 *journal_seq)
1619 {
1620         u32 restart_count = trans->restart_count;
1621         struct btree_iter iter;
1622         struct bkey_s_c k;
1623         int ret = 0;
1624
1625         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1626         while ((k = bch2_btree_iter_peek(&iter)).k) {
1627                 struct disk_reservation disk_res =
1628                         bch2_disk_reservation_init(trans->c, 0);
1629                 struct bkey_i delete;
1630
1631                 ret = bkey_err(k);
1632                 if (ret)
1633                         goto err;
1634
1635                 if (bkey_cmp(iter.pos, end) >= 0)
1636                         break;
1637
1638                 bkey_init(&delete.k);
1639
1640                 /*
1641                  * This could probably be more efficient for extents:
1642                  */
1643
1644                 /*
1645                  * For extents, iter.pos won't necessarily be the same as
1646                  * bkey_start_pos(k.k) (for non extents they always will be the
1647                  * same). It's important that we delete starting from iter.pos
1648                  * because the range we want to delete could start in the middle
1649                  * of k.
1650                  *
1651                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1652                  * bkey_start_pos(k.k)).
1653                  */
1654                 delete.k.p = iter.pos;
1655
1656                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1657                         unsigned max_sectors =
1658                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1659
1660                         /* create the biggest key we can */
1661                         bch2_key_resize(&delete.k, max_sectors);
1662                         bch2_cut_back(end, &delete);
1663
1664                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1665                         if (ret)
1666                                 goto err;
1667                 }
1668
1669                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1670                         bch2_trans_commit(trans, &disk_res, journal_seq,
1671                                           BTREE_INSERT_NOFAIL);
1672                 bch2_disk_reservation_put(trans->c, &disk_res);
1673 err:
1674                 /*
1675                  * the bch2_trans_begin() call is in a weird place because we
1676                  * need to call it after every transaction commit, to avoid path
1677                  * overflow, but don't want to call it if the delete operation
1678                  * is a no-op and we have no work to do:
1679                  */
1680                 bch2_trans_begin(trans);
1681
1682                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1683                         ret = 0;
1684                 if (ret)
1685                         break;
1686         }
1687         bch2_trans_iter_exit(trans, &iter);
1688
1689         if (!ret && trans_was_restarted(trans, restart_count))
1690                 ret = -BCH_ERR_transaction_restart_nested;
1691         return ret;
1692 }
1693
1694 /*
1695  * bch_btree_delete_range - delete everything within a given range
1696  *
1697  * Range is a half open interval - [start, end)
1698  */
1699 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1700                             struct bpos start, struct bpos end,
1701                             unsigned update_flags,
1702                             u64 *journal_seq)
1703 {
1704         int ret = bch2_trans_run(c,
1705                         bch2_btree_delete_range_trans(&trans, id, start, end,
1706                                                       update_flags, journal_seq));
1707         if (ret == -BCH_ERR_transaction_restart_nested)
1708                 ret = 0;
1709         return ret;
1710 }
1711
1712 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1713 {
1714         unsigned len = strlen(msg);
1715         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1716         struct jset_entry_log *l;
1717         int ret;
1718
1719         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1720         if (ret)
1721                 return ret;
1722
1723         l = (void *) &darray_top(trans->extra_journal_entries);
1724         l->entry.u64s           = cpu_to_le16(u64s);
1725         l->entry.btree_id       = 0;
1726         l->entry.level          = 1;
1727         l->entry.type           = BCH_JSET_ENTRY_log;
1728         l->entry.pad[0]         = 0;
1729         l->entry.pad[1]         = 0;
1730         l->entry.pad[2]         = 0;
1731         memcpy(l->d, msg, len);
1732         while (len & 7)
1733                 l->d[len++] = '\0';
1734
1735         trans->extra_journal_entries.nr += jset_u64s(u64s);
1736         return 0;
1737 }