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 "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "recovery.h"
  19 #include "subvolume.h"
  20 #include "replicas.h"
  21
  22 #include <linux/prefetch.h>
  23 #include <linux/sort.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static int __must_check
  27 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  28                           struct bkey_i *, enum btree_update_flags);
  29
  30 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  31                                          const struct btree_insert_entry *r)
  32 {
  33         return   cmp_int(l->btree_id,   r->btree_id) ?:
  34                  -cmp_int(l->level,     r->level) ?:
  35                  bpos_cmp(l->k->k.p,    r->k->k.p);
  36 }
  37
  38 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  39 {
  40         return i->path->l + i->level;
  41 }
  42
  43 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  44                                      struct btree_insert_entry *i)
  45 {
  46         return i != trans->updates &&
  47                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  48 }
  49
  50 static inline bool same_leaf_as_next(struct btree_trans *trans,
  51                                      struct btree_insert_entry *i)
  52 {
  53         return i + 1 < trans->updates + trans->nr_updates &&
  54                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  55 }
  56
  57 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  58                                                   struct btree_path *path,
  59                                                   struct btree *b)
  60 {
  61         struct bch_fs *c = trans->c;
  62
  63         if (path->cached)
  64                 return;
  65
  66         if (unlikely(btree_node_just_written(b)) &&
  67             bch2_btree_post_write_cleanup(c, b))
  68                 bch2_trans_node_reinit_iter(trans, b);
  69
  70         /*
  71          * If the last bset has been written, or if it's gotten too big - start
  72          * a new bset to insert into:
  73          */
  74         if (want_new_bset(c, b))
  75                 bch2_btree_init_next(trans, b);
  76 }
  77
  78 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
  79                                      struct btree_path *path,
  80                                      struct btree *b)
  81 {
  82         bch2_btree_node_lock_write(trans, path, b);
  83         bch2_btree_node_prep_for_write(trans, path, b);
  84 }
  85
  86 /* Inserting into a given leaf node (last stage of insert): */
  87
  88 /* Handle overwrites and do insert, for non extents: */
  89 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  90                                 struct btree_path *path,
  91                                 struct btree *b,
  92                                 struct btree_node_iter *node_iter,
  93                                 struct bkey_i *insert)
  94 {
  95         struct bkey_packed *k;
  96         unsigned clobber_u64s = 0, new_u64s = 0;
  97
  98         EBUG_ON(btree_node_just_written(b));
  99         EBUG_ON(bset_written(b, btree_bset_last(b)));
 100         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 101         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
 102         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
 103         EBUG_ON(insert->k.u64s >
 104                 bch_btree_keys_u64s_remaining(trans->c, b));
 105
 106         k = bch2_btree_node_iter_peek_all(node_iter, b);
 107         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 108                 k = NULL;
 109
 110         /* @k is the key being overwritten/deleted, if any: */
 111         EBUG_ON(k && bkey_deleted(k));
 112
 113         /* Deleting, but not found? nothing to do: */
 114         if (bkey_deleted(&insert->k) && !k)
 115                 return false;
 116
 117         if (bkey_deleted(&insert->k)) {
 118                 /* Deleting: */
 119                 btree_account_key_drop(b, k);
 120                 k->type = KEY_TYPE_deleted;
 121
 122                 if (k->needs_whiteout)
 123                         push_whiteout(trans->c, b, insert->k.p);
 124                 k->needs_whiteout = false;
 125
 126                 if (k >= btree_bset_last(b)->start) {
 127                         clobber_u64s = k->u64s;
 128                         bch2_bset_delete(b, k, clobber_u64s);
 129                         goto fix_iter;
 130                 } else {
 131                         bch2_btree_path_fix_key_modified(trans, b, k);
 132                 }
 133
 134                 return true;
 135         }
 136
 137         if (k) {
 138                 /* Overwriting: */
 139                 btree_account_key_drop(b, k);
 140                 k->type = KEY_TYPE_deleted;
 141
 142                 insert->k.needs_whiteout = k->needs_whiteout;
 143                 k->needs_whiteout = false;
 144
 145                 if (k >= btree_bset_last(b)->start) {
 146                         clobber_u64s = k->u64s;
 147                         goto overwrite;
 148                 } else {
 149                         bch2_btree_path_fix_key_modified(trans, b, k);
 150                 }
 151         }
 152
 153         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 154 overwrite:
 155         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 156         new_u64s = k->u64s;
 157 fix_iter:
 158         if (clobber_u64s != new_u64s)
 159                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 160                                          clobber_u64s, new_u64s);
 161         return true;
 162 }
 163
 164 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 165                                unsigned i, u64 seq)
 166 {
 167         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 168         struct btree_write *w = container_of(pin, struct btree_write, journal);
 169         struct btree *b = container_of(w, struct btree, writes[i]);
 170
 171         six_lock_read(&b->c.lock, NULL, NULL);
 172         bch2_btree_node_write_cond(c, b,
 173                 (btree_current_write(b) == w && w->journal.seq == seq));
 174         six_unlock_read(&b->c.lock);
 175         return 0;
 176 }
 177
 178 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 179 {
 180         return __btree_node_flush(j, pin, 0, seq);
 181 }
 182
 183 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 184 {
 185         return __btree_node_flush(j, pin, 1, seq);
 186 }
 187
 188 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 189                                        struct btree *b, u64 seq)
 190 {
 191         struct btree_write *w = btree_current_write(b);
 192
 193         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 194                              btree_node_write_idx(b) == 0
 195                              ? btree_node_flush0
 196                              : btree_node_flush1);
 197 }
 198
 199 /**
 200  * btree_insert_key - insert a key one key into a leaf node
 201  */
 202 static bool btree_insert_key_leaf(struct btree_trans *trans,
 203                                   struct btree_insert_entry *insert)
 204 {
 205         struct bch_fs *c = trans->c;
 206         struct btree *b = insert_l(insert)->b;
 207         struct bset_tree *t = bset_tree_last(b);
 208         struct bset *i = bset(b, t);
 209         int old_u64s = bset_u64s(t);
 210         int old_live_u64s = b->nr.live_u64s;
 211         int live_u64s_added, u64s_added;
 212
 213         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 214                                         &insert_l(insert)->iter, insert->k)))
 215                 return false;
 216
 217         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 218                                          le64_to_cpu(i->journal_seq)));
 219
 220         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 221
 222         if (unlikely(!btree_node_dirty(b)))
 223                 set_btree_node_dirty(c, b);
 224
 225         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 226         u64s_added = (int) bset_u64s(t) - old_u64s;
 227
 228         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 229                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 230         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 231                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 232
 233         if (u64s_added > live_u64s_added &&
 234             bch2_maybe_compact_whiteouts(c, b))
 235                 bch2_trans_node_reinit_iter(trans, b);
 236
 237         return true;
 238 }
 239
 240 /* Cached btree updates: */
 241
 242 /* Normal update interface: */
 243
 244 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 245                                              struct btree_insert_entry *i)
 246 {
 247         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 248         BUG_ON(i->cached        != i->path->cached);
 249         BUG_ON(i->level         != i->path->level);
 250         BUG_ON(i->btree_id      != i->path->btree_id);
 251         EBUG_ON(!i->level &&
 252                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 253                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 254                 i->k->k.p.snapshot &&
 255                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 256 }
 257
 258 static noinline int
 259 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 260                                    unsigned long trace_ip)
 261 {
 262         struct bch_fs *c = trans->c;
 263         int ret;
 264
 265         bch2_trans_unlock(trans);
 266
 267         ret = bch2_journal_preres_get(&c->journal,
 268                         &trans->journal_preres, u64s, 0);
 269         if (ret)
 270                 return ret;
 271
 272         if (!bch2_trans_relock(trans)) {
 273                 trace_trans_restart_journal_preres_get(trans->fn, trace_ip);
 274                 return -EINTR;
 275         }
 276
 277         return 0;
 278 }
 279
 280 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 281                                              unsigned flags)
 282 {
 283         struct bch_fs *c = trans->c;
 284         int ret;
 285
 286         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 287                 flags |= JOURNAL_RES_GET_RESERVED;
 288
 289         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 290                                    trans->journal_u64s, flags);
 291
 292         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 293 }
 294
 295 #define JSET_ENTRY_LOG_U64s             4
 296
 297 static noinline void journal_transaction_name(struct btree_trans *trans)
 298 {
 299         struct bch_fs *c = trans->c;
 300         struct jset_entry *entry = journal_res_entry(&c->journal, &trans->journal_res);
 301         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 302         unsigned u64s = JSET_ENTRY_LOG_U64s - 1;
 303         unsigned b, buflen = u64s * sizeof(u64);
 304
 305         l->entry.u64s           = cpu_to_le16(u64s);
 306         l->entry.btree_id       = 0;
 307         l->entry.level          = 0;
 308         l->entry.type           = BCH_JSET_ENTRY_log;
 309         l->entry.pad[0]         = 0;
 310         l->entry.pad[1]         = 0;
 311         l->entry.pad[2]         = 0;
 312         b = min_t(unsigned, strlen(trans->fn), buflen);
 313         memcpy(l->d, trans->fn, b);
 314         while (b < buflen)
 315                 l->d[b++] = '\0';
 316
 317         trans->journal_res.offset       += JSET_ENTRY_LOG_U64s;
 318         trans->journal_res.u64s         -= JSET_ENTRY_LOG_U64s;
 319 }
 320
 321 static inline enum btree_insert_ret
 322 btree_key_can_insert(struct btree_trans *trans,
 323                      struct btree *b,
 324                      unsigned u64s)
 325 {
 326         struct bch_fs *c = trans->c;
 327
 328         if (!bch2_btree_node_insert_fits(c, b, u64s))
 329                 return BTREE_INSERT_BTREE_NODE_FULL;
 330
 331         return BTREE_INSERT_OK;
 332 }
 333
 334 static enum btree_insert_ret
 335 btree_key_can_insert_cached(struct btree_trans *trans,
 336                             struct btree_path *path,
 337                             unsigned u64s)
 338 {
 339         struct bch_fs *c = trans->c;
 340         struct bkey_cached *ck = (void *) path->l[0].b;
 341         unsigned new_u64s;
 342         struct bkey_i *new_k;
 343
 344         EBUG_ON(path->level);
 345
 346         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 347             bch2_btree_key_cache_must_wait(c) &&
 348             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 349                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 350
 351         /*
 352          * bch2_varint_decode can read past the end of the buffer by at most 7
 353          * bytes (it won't be used):
 354          */
 355         u64s += 1;
 356
 357         if (u64s <= ck->u64s)
 358                 return BTREE_INSERT_OK;
 359
 360         new_u64s        = roundup_pow_of_two(u64s);
 361         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 362         if (!new_k) {
 363                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 364                         bch2_btree_ids[path->btree_id], new_u64s);
 365                 return -ENOMEM;
 366         }
 367
 368         ck->u64s        = new_u64s;
 369         ck->k           = new_k;
 370         return BTREE_INSERT_OK;
 371 }
 372
 373 static inline void do_btree_insert_one(struct btree_trans *trans,
 374                                        struct btree_insert_entry *i)
 375 {
 376         struct bch_fs *c = trans->c;
 377         struct journal *j = &c->journal;
 378         bool did_work;
 379
 380         EBUG_ON(trans->journal_res.ref !=
 381                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 382
 383         i->k->k.needs_whiteout = false;
 384
 385         did_work = !i->cached
 386                 ? btree_insert_key_leaf(trans, i)
 387                 : bch2_btree_insert_key_cached(trans, i->path, i->k);
 388         if (!did_work)
 389                 return;
 390
 391         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 392                 bch2_journal_add_keys(j, &trans->journal_res,
 393                                       i->btree_id,
 394                                       i->level,
 395                                       i->k);
 396
 397                 if (trans->journal_seq)
 398                         *trans->journal_seq = trans->journal_res.seq;
 399         }
 400 }
 401
 402 static noinline int bch2_trans_mark_gc(struct btree_trans *trans)
 403 {
 404         struct bch_fs *c = trans->c;
 405         struct btree_insert_entry *i;
 406         int ret = 0;
 407
 408         trans_for_each_update(trans, i) {
 409                 /*
 410                  * XXX: synchronization of cached update triggers with gc
 411                  * XXX: synchronization of interior node updates with gc
 412                  */
 413                 BUG_ON(i->cached || i->level);
 414
 415                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 416                         ret = bch2_mark_update(trans, i->path, i->k,
 417                                                i->flags|BTREE_TRIGGER_GC);
 418                         if (ret)
 419                                 break;
 420                 }
 421         }
 422
 423         return ret;
 424 }
 425
 426 static inline int
 427 bch2_trans_commit_write_locked(struct btree_trans *trans,
 428                                struct btree_insert_entry **stopped_at,
 429                                unsigned long trace_ip)
 430 {
 431         struct bch_fs *c = trans->c;
 432         struct btree_insert_entry *i;
 433         struct btree_trans_commit_hook *h;
 434         unsigned u64s = 0;
 435         bool marking = false;
 436         int ret;
 437
 438         if (race_fault()) {
 439                 trace_trans_restart_fault_inject(trans->fn, trace_ip);
 440                 trans->restarted = true;
 441                 return -EINTR;
 442         }
 443
 444         /*
 445          * Check if the insert will fit in the leaf node with the write lock
 446          * held, otherwise another thread could write the node changing the
 447          * amount of space available:
 448          */
 449
 450         prefetch(&trans->c->journal.flags);
 451
 452         h = trans->hooks;
 453         while (h) {
 454                 ret = h->fn(trans, h);
 455                 if (ret)
 456                         return ret;
 457                 h = h->next;
 458         }
 459
 460         trans_for_each_update(trans, i) {
 461                 /* Multiple inserts might go to same leaf: */
 462                 if (!same_leaf_as_prev(trans, i))
 463                         u64s = 0;
 464
 465                 u64s += i->k->k.u64s;
 466                 ret = !i->cached
 467                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 468                         : btree_key_can_insert_cached(trans, i->path, u64s);
 469                 if (ret) {
 470                         *stopped_at = i;
 471                         return ret;
 472                 }
 473
 474                 if (btree_node_type_needs_gc(i->bkey_type))
 475                         marking = true;
 476         }
 477
 478         /*
 479          * Don't get journal reservation until after we know insert will
 480          * succeed:
 481          */
 482         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 483                 ret = bch2_trans_journal_res_get(trans,
 484                                 JOURNAL_RES_GET_NONBLOCK);
 485                 if (ret)
 486                         return ret;
 487
 488                 if (unlikely(trans->journal_transaction_names))
 489                         journal_transaction_name(trans);
 490         } else {
 491                 trans->journal_res.seq = c->journal.replay_journal_seq;
 492         }
 493
 494         if (unlikely(trans->extra_journal_entry_u64s)) {
 495                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 496                                   trans->extra_journal_entries,
 497                                   trans->extra_journal_entry_u64s);
 498
 499                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 500                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 501         }
 502
 503         /*
 504          * Not allowed to fail after we've gotten our journal reservation - we
 505          * have to use it:
 506          */
 507
 508         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 509                 if (bch2_journal_seq_verify)
 510                         trans_for_each_update(trans, i)
 511                                 i->k->k.version.lo = trans->journal_res.seq;
 512                 else if (bch2_inject_invalid_keys)
 513                         trans_for_each_update(trans, i)
 514                                 i->k->k.version = MAX_VERSION;
 515         }
 516
 517         if (trans->fs_usage_deltas &&
 518             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 519                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 520
 521         trans_for_each_update(trans, i)
 522                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 523                         ret = bch2_mark_update(trans, i->path, i->k, i->flags);
 524                         if (ret)
 525                                 return ret;
 526                 }
 527
 528         if (unlikely(c->gc_pos.phase)) {
 529                 ret = bch2_trans_mark_gc(trans);
 530                 if  (ret)
 531                         return ret;
 532         }
 533
 534         trans_for_each_update(trans, i)
 535                 do_btree_insert_one(trans, i);
 536
 537         return ret;
 538 }
 539
 540 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 541 {
 542         unsigned l;
 543
 544         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 545                 if (btree_node_read_locked(path, l))
 546                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 547 }
 548
 549 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 550 {
 551         struct btree *b = path_l(path)->b;
 552
 553         do {
 554                 if (path->nodes_locked &&
 555                     path->nodes_locked != path->nodes_intent_locked)
 556                         path_upgrade_readers(trans, path);
 557         } while ((path = prev_btree_path(trans, path)) &&
 558                  path_l(path)->b == b);
 559 }
 560
 561 /*
 562  * Check for nodes that we have both read and intent locks on, and upgrade the
 563  * readers to intent:
 564  */
 565 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 566 {
 567         struct btree_path *path;
 568         unsigned i, nr_read = 0, nr_intent = 0;
 569
 570         trans_for_each_path_inorder(trans, path, i) {
 571                 struct btree_path *next = i + 1 < trans->nr_sorted
 572                         ? trans->paths + trans->sorted[i + 1]
 573                         : NULL;
 574
 575                 if (path->nodes_locked) {
 576                         if (path->nodes_intent_locked)
 577                                 nr_intent++;
 578                         else
 579                                 nr_read++;
 580                 }
 581
 582                 if (!next || path_l(path)->b != path_l(next)->b) {
 583                         if (nr_read && nr_intent)
 584                                 upgrade_readers(trans, path);
 585
 586                         nr_read = nr_intent = 0;
 587                 }
 588         }
 589
 590         bch2_trans_verify_locks(trans);
 591 }
 592
 593 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 594 {
 595         struct btree_path *path;
 596         unsigned i;
 597
 598         trans_for_each_path_inorder(trans, path, i) {
 599                 //if (path == pos)
 600                 //      break;
 601
 602                 if (path->nodes_locked != path->nodes_intent_locked &&
 603                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 604                         return true;
 605         }
 606
 607         return false;
 608 }
 609
 610 static inline int trans_lock_write(struct btree_trans *trans)
 611 {
 612         struct btree_insert_entry *i;
 613
 614         trans_for_each_update(trans, i) {
 615                 if (same_leaf_as_prev(trans, i))
 616                         continue;
 617
 618                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 619                         if (have_conflicting_read_lock(trans, i->path))
 620                                 goto fail;
 621
 622                         btree_node_lock_type(trans, i->path,
 623                                              insert_l(i)->b,
 624                                              i->path->pos, i->level,
 625                                              SIX_LOCK_write, NULL, NULL);
 626                 }
 627
 628                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 629         }
 630
 631         return 0;
 632 fail:
 633         while (--i >= trans->updates) {
 634                 if (same_leaf_as_prev(trans, i))
 635                         continue;
 636
 637                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 638         }
 639
 640         trace_trans_restart_would_deadlock_write(trans->fn);
 641         return btree_trans_restart(trans);
 642 }
 643
 644 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 645 {
 646         struct btree_insert_entry *i;
 647
 648         trans_for_each_update(trans, i)
 649                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 650 }
 651
 652 /*
 653  * Get journal reservation, take write locks, and attempt to do btree update(s):
 654  */
 655 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 656                                        struct btree_insert_entry **stopped_at,
 657                                        unsigned long trace_ip)
 658 {
 659         struct bch_fs *c = trans->c;
 660         struct btree_insert_entry *i;
 661         struct bkey_s_c old;
 662         int ret, u64s_delta = 0;
 663
 664         trans_for_each_update(trans, i) {
 665                 const char *invalid = bch2_bkey_invalid(c,
 666                                 bkey_i_to_s_c(i->k), i->bkey_type);
 667                 if (invalid) {
 668                         char buf[200];
 669
 670                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 671                         bch2_fs_fatal_error(c, "invalid bkey %s on insert from %s -> %ps: %s\n",
 672                                             buf, trans->fn, (void *) i->ip_allocated, invalid);
 673                         return -EINVAL;
 674                 }
 675                 btree_insert_entry_checks(trans, i);
 676         }
 677
 678         trans_for_each_update(trans, i) {
 679                 struct bkey u;
 680
 681                 /*
 682                  * peek_slot() doesn't yet work on iterators that point to
 683                  * interior nodes:
 684                  */
 685                 if (i->cached || i->level)
 686                         continue;
 687
 688                 old = bch2_btree_path_peek_slot(i->path, &u);
 689                 ret = bkey_err(old);
 690                 if (unlikely(ret))
 691                         return ret;
 692
 693                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 694                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 695
 696                 if (!same_leaf_as_next(trans, i)) {
 697                         if (u64s_delta <= 0) {
 698                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 699                                                         i->level, trans->flags);
 700                                 if (unlikely(ret))
 701                                         return ret;
 702                         }
 703
 704                         u64s_delta = 0;
 705                 }
 706         }
 707
 708         ret = bch2_journal_preres_get(&c->journal,
 709                         &trans->journal_preres, trans->journal_preres_u64s,
 710                         JOURNAL_RES_GET_NONBLOCK|
 711                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 712                          ? JOURNAL_RES_GET_RESERVED : 0));
 713         if (unlikely(ret == -EAGAIN))
 714                 ret = bch2_trans_journal_preres_get_cold(trans,
 715                                                 trans->journal_preres_u64s, trace_ip);
 716         if (unlikely(ret))
 717                 return ret;
 718
 719         normalize_read_intent_locks(trans);
 720
 721         ret = trans_lock_write(trans);
 722         if (unlikely(ret))
 723                 return ret;
 724
 725         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 726
 727         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 728                 bch2_drop_overwrites_from_journal(trans);
 729
 730         trans_for_each_update(trans, i)
 731                 if (!same_leaf_as_prev(trans, i))
 732                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 733                                                         insert_l(i)->b);
 734
 735         if (!ret && trans->journal_pin)
 736                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 737                                      trans->journal_pin, NULL);
 738
 739         /*
 740          * Drop journal reservation after dropping write locks, since dropping
 741          * the journal reservation may kick off a journal write:
 742          */
 743         bch2_journal_res_put(&c->journal, &trans->journal_res);
 744
 745         if (unlikely(ret))
 746                 return ret;
 747
 748         bch2_trans_downgrade(trans);
 749
 750         return 0;
 751 }
 752
 753 static int journal_reclaim_wait_done(struct bch_fs *c)
 754 {
 755         int ret = bch2_journal_error(&c->journal) ?:
 756                 !bch2_btree_key_cache_must_wait(c);
 757
 758         if (!ret)
 759                 journal_reclaim_kick(&c->journal);
 760         return ret;
 761 }
 762
 763 static noinline
 764 int bch2_trans_commit_error(struct btree_trans *trans,
 765                             struct btree_insert_entry *i,
 766                             int ret, unsigned long trace_ip)
 767 {
 768         struct bch_fs *c = trans->c;
 769
 770         switch (ret) {
 771         case BTREE_INSERT_BTREE_NODE_FULL:
 772                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 773                 if (!ret)
 774                         return 0;
 775
 776                 if (ret == -EINTR)
 777                         trace_trans_restart_btree_node_split(trans->fn, trace_ip,
 778                                                 i->btree_id, &i->path->pos);
 779                 break;
 780         case BTREE_INSERT_NEED_MARK_REPLICAS:
 781                 bch2_trans_unlock(trans);
 782
 783                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 784                 if (ret)
 785                         break;
 786
 787                 if (bch2_trans_relock(trans))
 788                         return 0;
 789
 790                 trace_trans_restart_mark_replicas(trans->fn, trace_ip);
 791                 ret = -EINTR;
 792                 break;
 793         case BTREE_INSERT_NEED_JOURNAL_RES:
 794                 bch2_trans_unlock(trans);
 795
 796                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 797                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
 798                         trans->restarted = true;
 799                         ret = -EAGAIN;
 800                         break;
 801                 }
 802
 803                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 804                 if (ret)
 805                         break;
 806
 807                 if (bch2_trans_relock(trans))
 808                         return 0;
 809
 810                 trace_trans_restart_journal_res_get(trans->fn, trace_ip);
 811                 ret = -EINTR;
 812                 break;
 813         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 814                 bch2_trans_unlock(trans);
 815
 816                 trace_trans_blocked_journal_reclaim(trans->fn, trace_ip);
 817
 818                 wait_event_freezable(c->journal.reclaim_wait,
 819                                      (ret = journal_reclaim_wait_done(c)));
 820                 if (ret < 0)
 821                         break;
 822
 823                 if (bch2_trans_relock(trans))
 824                         return 0;
 825
 826                 trace_trans_restart_journal_reclaim(trans->fn, trace_ip);
 827                 ret = -EINTR;
 828                 break;
 829         default:
 830                 BUG_ON(ret >= 0);
 831                 break;
 832         }
 833
 834         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
 835         BUG_ON(ret == -ENOSPC &&
 836                !(trans->flags & BTREE_INSERT_NOWAIT) &&
 837                (trans->flags & BTREE_INSERT_NOFAIL));
 838
 839         return ret;
 840 }
 841
 842 static noinline int
 843 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 844 {
 845         struct bch_fs *c = trans->c;
 846         int ret;
 847
 848         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 849             test_bit(BCH_FS_STARTED, &c->flags))
 850                 return -EROFS;
 851
 852         bch2_trans_unlock(trans);
 853
 854         ret = bch2_fs_read_write_early(c);
 855         if (ret)
 856                 return ret;
 857
 858         if (!bch2_trans_relock(trans))
 859                 return -EINTR;
 860
 861         percpu_ref_get(&c->writes);
 862         return 0;
 863 }
 864
 865 static int run_one_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 866                            bool overwrite)
 867 {
 868         struct bkey             _deleted = KEY(0, 0, 0);
 869         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 870         struct bkey_s_c         old;
 871         struct bkey             unpacked;
 872         int ret = 0;
 873
 874         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 875             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 876                 return 0;
 877
 878         if (!overwrite) {
 879                 if (i->insert_trigger_run)
 880                         return 0;
 881
 882                 BUG_ON(i->overwrite_trigger_run);
 883                 i->insert_trigger_run = true;
 884         } else {
 885                 if (i->overwrite_trigger_run)
 886                         return 0;
 887
 888                 BUG_ON(!i->insert_trigger_run);
 889                 i->overwrite_trigger_run = true;
 890         }
 891
 892         old = bch2_btree_path_peek_slot(i->path, &unpacked);
 893         _deleted.p = i->path->pos;
 894
 895         if (overwrite) {
 896                 ret = bch2_trans_mark_key(trans, old, deleted,
 897                                 BTREE_TRIGGER_OVERWRITE|i->flags);
 898         } else if (old.k->type == i->k->k.type &&
 899             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 900                 i->overwrite_trigger_run = true;
 901                 ret = bch2_trans_mark_key(trans, old, bkey_i_to_s_c(i->k),
 902                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|i->flags);
 903         } else {
 904                 ret = bch2_trans_mark_key(trans, deleted, bkey_i_to_s_c(i->k),
 905                                 BTREE_TRIGGER_INSERT|i->flags);
 906         }
 907
 908         if (ret == -EINTR)
 909                 trace_trans_restart_mark(trans->fn, _RET_IP_,
 910                                          i->btree_id, &i->path->pos);
 911         return ret ?: 1;
 912 }
 913
 914 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 915                               struct btree_insert_entry *btree_id_start)
 916 {
 917         struct btree_insert_entry *i;
 918         bool trans_trigger_run;
 919         int ret, overwrite;
 920
 921         for (overwrite = 0; overwrite < 2; overwrite++) {
 922
 923                 /*
 924                  * Running triggers will append more updates to the list of updates as
 925                  * we're walking it:
 926                  */
 927                 do {
 928                         trans_trigger_run = false;
 929
 930                         for (i = btree_id_start;
 931                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 932                              i++) {
 933                                 ret = run_one_trigger(trans, i, overwrite);
 934                                 if (ret < 0)
 935                                         return ret;
 936                                 if (ret)
 937                                         trans_trigger_run = true;
 938                         }
 939                 } while (trans_trigger_run);
 940         }
 941
 942         return 0;
 943 }
 944
 945 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 946 {
 947         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 948         unsigned btree_id = 0;
 949         int ret = 0;
 950
 951         /*
 952          *
 953          * For a given btree, this algorithm runs insert triggers before
 954          * overwrite triggers: this is so that when extents are being moved
 955          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 956          * they are re-added.
 957          */
 958         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 959                 while (btree_id_start < trans->updates + trans->nr_updates &&
 960                        btree_id_start->btree_id < btree_id)
 961                         btree_id_start++;
 962
 963                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 964                 if (ret)
 965                         return ret;
 966         }
 967
 968         trans_for_each_update(trans, i)
 969                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 970                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 971                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 972
 973         return 0;
 974 }
 975
 976 int __bch2_trans_commit(struct btree_trans *trans)
 977 {
 978         struct bch_fs *c = trans->c;
 979         struct btree_insert_entry *i = NULL;
 980         unsigned u64s;
 981         int ret = 0;
 982
 983         if (!trans->nr_updates &&
 984             !trans->extra_journal_entry_u64s)
 985                 goto out_reset;
 986
 987         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 988                 lockdep_assert_held(&c->gc_lock);
 989
 990         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 991
 992         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 993         trans->journal_preres_u64s      = 0;
 994
 995         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
 996
 997         if (trans->journal_transaction_names)
 998                 trans->journal_u64s += JSET_ENTRY_LOG_U64s;
 999
1000         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1001             unlikely(!percpu_ref_tryget(&c->writes))) {
1002                 ret = bch2_trans_commit_get_rw_cold(trans);
1003                 if (ret)
1004                         goto out_reset;
1005         }
1006
1007 #ifdef CONFIG_BCACHEFS_DEBUG
1008         /*
1009          * if BTREE_TRIGGER_NORUN is set, it means we're probably being called
1010          * from the key cache flush code:
1011          */
1012         trans_for_each_update(trans, i)
1013                 if (!i->cached &&
1014                     !(i->flags & BTREE_TRIGGER_NORUN))
1015                         bch2_btree_key_cache_verify_clean(trans,
1016                                         i->btree_id, i->k->k.p);
1017 #endif
1018
1019         ret = bch2_trans_commit_run_triggers(trans);
1020         if (ret)
1021                 goto out;
1022
1023         trans_for_each_update(trans, i) {
1024                 BUG_ON(!i->path->should_be_locked);
1025
1026                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1027                         trace_trans_restart_upgrade(trans->fn, _RET_IP_,
1028                                                     i->btree_id, &i->path->pos);
1029                         ret = btree_trans_restart(trans);
1030                         goto out;
1031                 }
1032
1033                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1034
1035                 u64s = jset_u64s(i->k->k.u64s);
1036                 if (i->cached &&
1037                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1038                         trans->journal_preres_u64s += u64s;
1039                 trans->journal_u64s += u64s;
1040         }
1041
1042         if (trans->extra_journal_res) {
1043                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1044                                 trans->extra_journal_res,
1045                                 (trans->flags & BTREE_INSERT_NOFAIL)
1046                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1047                 if (ret)
1048                         goto err;
1049         }
1050 retry:
1051         BUG_ON(trans->restarted);
1052         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1053
1054         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1055
1056         /* make sure we didn't drop or screw up locks: */
1057         bch2_trans_verify_locks(trans);
1058
1059         if (ret)
1060                 goto err;
1061 out:
1062         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1063
1064         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1065                 percpu_ref_put(&c->writes);
1066 out_reset:
1067         trans_for_each_update(trans, i)
1068                 bch2_path_put(trans, i->path, true);
1069
1070         trans->extra_journal_res        = 0;
1071         trans->nr_updates               = 0;
1072         trans->hooks                    = NULL;
1073         trans->extra_journal_entries    = NULL;
1074         trans->extra_journal_entry_u64s = 0;
1075
1076         if (trans->fs_usage_deltas) {
1077                 trans->fs_usage_deltas->used = 0;
1078                 memset(&trans->fs_usage_deltas->memset_start, 0,
1079                        (void *) &trans->fs_usage_deltas->memset_end -
1080                        (void *) &trans->fs_usage_deltas->memset_start);
1081         }
1082
1083         return ret;
1084 err:
1085         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1086         if (ret)
1087                 goto out;
1088
1089         goto retry;
1090 }
1091
1092 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1093                                           enum btree_id id,
1094                                           struct bpos pos)
1095 {
1096         struct bch_fs *c = trans->c;
1097         struct btree_iter iter;
1098         struct bkey_s_c k;
1099         int ret;
1100
1101         if (!btree_type_has_snapshots(id))
1102                 return 0;
1103
1104         if (!snapshot_t(c, pos.snapshot)->children[0])
1105                 return 0;
1106
1107         bch2_trans_iter_init(trans, &iter, id, pos,
1108                              BTREE_ITER_NOT_EXTENTS|
1109                              BTREE_ITER_ALL_SNAPSHOTS);
1110         while (1) {
1111                 k = bch2_btree_iter_prev(&iter);
1112                 ret = bkey_err(k);
1113                 if (ret)
1114                         break;
1115
1116                 if (!k.k)
1117                         break;
1118
1119                 if (bkey_cmp(pos, k.k->p))
1120                         break;
1121
1122                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1123                         ret = 1;
1124                         break;
1125                 }
1126         }
1127         bch2_trans_iter_exit(trans, &iter);
1128
1129         return ret;
1130 }
1131
1132 int bch2_trans_update_extent(struct btree_trans *trans,
1133                              struct btree_iter *orig_iter,
1134                              struct bkey_i *insert,
1135                              enum btree_update_flags flags)
1136 {
1137         struct bch_fs *c = trans->c;
1138         struct btree_iter iter, update_iter;
1139         struct bpos start = bkey_start_pos(&insert->k);
1140         struct bkey_i *update;
1141         struct bkey_s_c k;
1142         enum btree_id btree_id = orig_iter->btree_id;
1143         int ret = 0, compressed_sectors;
1144
1145         bch2_trans_iter_init(trans, &iter, btree_id, start,
1146                              BTREE_ITER_INTENT|
1147                              BTREE_ITER_WITH_UPDATES|
1148                              BTREE_ITER_NOT_EXTENTS);
1149         k = bch2_btree_iter_peek(&iter);
1150         if ((ret = bkey_err(k)))
1151                 goto err;
1152         if (!k.k)
1153                 goto out;
1154
1155         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1156                 /*
1157                  * We can't merge extents if they belong to interior snapshot
1158                  * tree nodes, and there's a snapshot in which one extent is
1159                  * visible and the other is not - i.e. if visibility is
1160                  * different.
1161                  *
1162                  * Instead of checking if visibilitiy of the two extents is
1163                  * different, for now we just check if either has been
1164                  * overwritten:
1165                  */
1166                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1167                 if (ret < 0)
1168                         goto err;
1169                 if (ret)
1170                         goto nomerge1;
1171
1172                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1173                 if (ret < 0)
1174                         goto err;
1175                 if (ret)
1176                         goto nomerge1;
1177
1178                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1179                 if ((ret = PTR_ERR_OR_ZERO(update)))
1180                         goto err;
1181
1182                 bkey_reassemble(update, k);
1183
1184                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1185                         ret = bch2_btree_delete_at(trans, &iter, flags);
1186                         if (ret)
1187                                 goto err;
1188
1189                         insert = update;
1190                         goto next;
1191                 }
1192         }
1193 nomerge1:
1194         ret = 0;
1195         if (!bkey_cmp(k.k->p, start))
1196                 goto next;
1197
1198         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1199                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1200                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1201
1202                 /*
1203                  * If we're going to be splitting a compressed extent, note it
1204                  * so that __bch2_trans_commit() can increase our disk
1205                  * reservation:
1206                  */
1207                 if (((front_split && back_split) ||
1208                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1209                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1210                         trans->extra_journal_res += compressed_sectors;
1211
1212                 if (front_split) {
1213                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1214                         if ((ret = PTR_ERR_OR_ZERO(update)))
1215                                 goto err;
1216
1217                         bkey_reassemble(update, k);
1218
1219                         bch2_cut_back(start, update);
1220
1221                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1222                                              BTREE_ITER_NOT_EXTENTS|
1223                                              BTREE_ITER_ALL_SNAPSHOTS|
1224                                              BTREE_ITER_INTENT);
1225                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1226                                 bch2_trans_update(trans, &update_iter, update,
1227                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1228                                                   flags);
1229                         bch2_trans_iter_exit(trans, &update_iter);
1230
1231                         if (ret)
1232                                 goto err;
1233                 }
1234
1235                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1236                     (front_split || back_split)) {
1237                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1238                         if ((ret = PTR_ERR_OR_ZERO(update)))
1239                                 goto err;
1240
1241                         bkey_reassemble(update, k);
1242
1243                         bch2_cut_front(start, update);
1244                         bch2_cut_back(insert->k.p, update);
1245
1246                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1247                                              BTREE_ITER_NOT_EXTENTS|
1248                                              BTREE_ITER_ALL_SNAPSHOTS|
1249                                              BTREE_ITER_INTENT);
1250                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1251                                 bch2_trans_update(trans, &update_iter, update,
1252                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1253                                                   flags);
1254                         bch2_trans_iter_exit(trans, &update_iter);
1255                         if (ret)
1256                                 goto err;
1257                 }
1258
1259                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1260                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1261                         if ((ret = PTR_ERR_OR_ZERO(update)))
1262                                 goto err;
1263
1264                         bkey_init(&update->k);
1265                         update->k.p = k.k->p;
1266
1267                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1268                                 update->k.p.snapshot = insert->k.p.snapshot;
1269                                 update->k.type = KEY_TYPE_whiteout;
1270                         }
1271
1272                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1273                                              BTREE_ITER_NOT_EXTENTS|
1274                                              BTREE_ITER_INTENT);
1275                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1276                                 bch2_trans_update(trans, &update_iter, update,
1277                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1278                                                   flags);
1279                         bch2_trans_iter_exit(trans, &update_iter);
1280
1281                         if (ret)
1282                                 goto err;
1283                 }
1284
1285                 if (back_split) {
1286                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1287                         if ((ret = PTR_ERR_OR_ZERO(update)))
1288                                 goto err;
1289
1290                         bkey_reassemble(update, k);
1291                         bch2_cut_front(insert->k.p, update);
1292
1293                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1294                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1295                                                   flags);
1296                         if (ret)
1297                                 goto err;
1298                         goto out;
1299                 }
1300 next:
1301                 k = bch2_btree_iter_next(&iter);
1302                 if ((ret = bkey_err(k)))
1303                         goto err;
1304                 if (!k.k)
1305                         goto out;
1306         }
1307
1308         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1309                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1310                 if (ret < 0)
1311                         goto err;
1312                 if (ret)
1313                         goto nomerge2;
1314
1315                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1316                 if (ret < 0)
1317                         goto err;
1318                 if (ret)
1319                         goto nomerge2;
1320
1321                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1322         }
1323 nomerge2:
1324         ret = 0;
1325 out:
1326         if (!bkey_deleted(&insert->k)) {
1327                 /*
1328                  * Rewinding iterators is expensive: get a new one and the one
1329                  * that points to the start of insert will be cloned from:
1330                  */
1331                 bch2_trans_iter_exit(trans, &iter);
1332                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1333                                      BTREE_ITER_NOT_EXTENTS|
1334                                      BTREE_ITER_INTENT);
1335                 ret   = bch2_btree_iter_traverse(&iter) ?:
1336                         bch2_trans_update(trans, &iter, insert, flags);
1337         }
1338 err:
1339         bch2_trans_iter_exit(trans, &iter);
1340
1341         return ret;
1342 }
1343
1344 /*
1345  * When deleting, check if we need to emit a whiteout (because we're overwriting
1346  * something in an ancestor snapshot)
1347  */
1348 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1349                                       enum btree_id btree_id, struct bpos pos)
1350 {
1351         struct btree_iter iter;
1352         struct bkey_s_c k;
1353         u32 snapshot = pos.snapshot;
1354         int ret;
1355
1356         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1357                 return 0;
1358
1359         pos.snapshot++;
1360
1361         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1362                            BTREE_ITER_ALL_SNAPSHOTS|
1363                            BTREE_ITER_NOPRESERVE, k, ret) {
1364                 if (bkey_cmp(k.k->p, pos))
1365                         break;
1366
1367                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1368                                               k.k->p.snapshot)) {
1369                         ret = !bkey_whiteout(k.k);
1370                         break;
1371                 }
1372         }
1373         bch2_trans_iter_exit(trans, &iter);
1374
1375         return ret;
1376 }
1377
1378 static int __must_check
1379 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1380                           struct bkey_i *k, enum btree_update_flags flags)
1381 {
1382         struct btree_insert_entry *i, n;
1383
1384         BUG_ON(!path->should_be_locked);
1385
1386         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1387         BUG_ON(bpos_cmp(k->k.p, path->pos));
1388
1389         n = (struct btree_insert_entry) {
1390                 .flags          = flags,
1391                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1392                 .btree_id       = path->btree_id,
1393                 .level          = path->level,
1394                 .cached         = path->cached,
1395                 .path           = path,
1396                 .k              = k,
1397                 .ip_allocated   = _RET_IP_,
1398         };
1399
1400 #ifdef CONFIG_BCACHEFS_DEBUG
1401         trans_for_each_update(trans, i)
1402                 BUG_ON(i != trans->updates &&
1403                        btree_insert_entry_cmp(i - 1, i) >= 0);
1404 #endif
1405
1406         /*
1407          * Pending updates are kept sorted: first, find position of new update,
1408          * then delete/trim any updates the new update overwrites:
1409          */
1410         trans_for_each_update(trans, i)
1411                 if (btree_insert_entry_cmp(&n, i) <= 0)
1412                         break;
1413
1414         if (i < trans->updates + trans->nr_updates &&
1415             !btree_insert_entry_cmp(&n, i)) {
1416                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1417
1418                 bch2_path_put(trans, i->path, true);
1419                 *i = n;
1420         } else
1421                 array_insert_item(trans->updates, trans->nr_updates,
1422                                   i - trans->updates, n);
1423
1424         __btree_path_get(n.path, true);
1425         return 0;
1426 }
1427
1428 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1429                                    struct bkey_i *k, enum btree_update_flags flags)
1430 {
1431         struct btree_path *path = iter->update_path ?: iter->path;
1432         struct bkey_cached *ck;
1433         int ret;
1434
1435         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1436                 return bch2_trans_update_extent(trans, iter, k, flags);
1437
1438         if (bkey_deleted(&k->k) &&
1439             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1440             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1441                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1442                 if (unlikely(ret < 0))
1443                         return ret;
1444
1445                 if (ret)
1446                         k->k.type = KEY_TYPE_whiteout;
1447         }
1448
1449         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1450             !path->cached &&
1451             !path->level &&
1452             btree_id_cached(trans->c, path->btree_id)) {
1453                 if (!iter->key_cache_path ||
1454                     !iter->key_cache_path->should_be_locked ||
1455                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1456                         if (!iter->key_cache_path)
1457                                 iter->key_cache_path =
1458                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1459                                                       BTREE_ITER_INTENT|
1460                                                       BTREE_ITER_CACHED, _THIS_IP_);
1461
1462                         iter->key_cache_path =
1463                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1464                                                         iter->flags & BTREE_ITER_INTENT,
1465                                                         _THIS_IP_);
1466
1467                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1468                                                        BTREE_ITER_CACHED|
1469                                                        BTREE_ITER_CACHED_NOFILL);
1470                         if (unlikely(ret))
1471                                 return ret;
1472
1473                         ck = (void *) iter->key_cache_path->l[0].b;
1474
1475                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1476                                 trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
1477                                 btree_trans_restart(trans);
1478                                 return -EINTR;
1479                         }
1480
1481                         iter->key_cache_path->should_be_locked = true;
1482                 }
1483
1484                 path = iter->key_cache_path;
1485         }
1486
1487         return bch2_trans_update_by_path(trans, path, k, flags);
1488 }
1489
1490 void bch2_trans_commit_hook(struct btree_trans *trans,
1491                             struct btree_trans_commit_hook *h)
1492 {
1493         h->next = trans->hooks;
1494         trans->hooks = h;
1495 }
1496
1497 int __bch2_btree_insert(struct btree_trans *trans,
1498                         enum btree_id id, struct bkey_i *k)
1499 {
1500         struct btree_iter iter;
1501         int ret;
1502
1503         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1504                              BTREE_ITER_INTENT);
1505         ret   = bch2_btree_iter_traverse(&iter) ?:
1506                 bch2_trans_update(trans, &iter, k, 0);
1507         bch2_trans_iter_exit(trans, &iter);
1508         return ret;
1509 }
1510
1511 /**
1512  * bch2_btree_insert - insert keys into the extent btree
1513  * @c:                  pointer to struct bch_fs
1514  * @id:                 btree to insert into
1515  * @insert_keys:        list of keys to insert
1516  * @hook:               insert callback
1517  */
1518 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1519                       struct bkey_i *k,
1520                       struct disk_reservation *disk_res,
1521                       u64 *journal_seq, int flags)
1522 {
1523         return bch2_trans_do(c, disk_res, journal_seq, flags,
1524                              __bch2_btree_insert(&trans, id, k));
1525 }
1526
1527 int bch2_btree_delete_at(struct btree_trans *trans,
1528                          struct btree_iter *iter, unsigned update_flags)
1529 {
1530         struct bkey_i *k;
1531
1532         k = bch2_trans_kmalloc(trans, sizeof(*k));
1533         if (IS_ERR(k))
1534                 return PTR_ERR(k);
1535
1536         bkey_init(&k->k);
1537         k->k.p = iter->pos;
1538         return bch2_trans_update(trans, iter, k, update_flags);
1539 }
1540
1541 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1542                                   struct bpos start, struct bpos end,
1543                                   unsigned iter_flags,
1544                                   u64 *journal_seq)
1545 {
1546         struct btree_iter iter;
1547         struct bkey_s_c k;
1548         int ret = 0;
1549
1550         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT|iter_flags);
1551 retry:
1552         while ((bch2_trans_begin(trans),
1553                (k = bch2_btree_iter_peek(&iter)).k) &&
1554                !(ret = bkey_err(k)) &&
1555                bkey_cmp(iter.pos, end) < 0) {
1556                 struct disk_reservation disk_res =
1557                         bch2_disk_reservation_init(trans->c, 0);
1558                 struct bkey_i delete;
1559
1560                 bkey_init(&delete.k);
1561
1562                 /*
1563                  * This could probably be more efficient for extents:
1564                  */
1565
1566                 /*
1567                  * For extents, iter.pos won't necessarily be the same as
1568                  * bkey_start_pos(k.k) (for non extents they always will be the
1569                  * same). It's important that we delete starting from iter.pos
1570                  * because the range we want to delete could start in the middle
1571                  * of k.
1572                  *
1573                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1574                  * bkey_start_pos(k.k)).
1575                  */
1576                 delete.k.p = iter.pos;
1577
1578                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1579                         unsigned max_sectors =
1580                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1581
1582                         /* create the biggest key we can */
1583                         bch2_key_resize(&delete.k, max_sectors);
1584                         bch2_cut_back(end, &delete);
1585
1586                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1587                         if (ret)
1588                                 break;
1589                 }
1590
1591                 ret   = bch2_trans_update(trans, &iter, &delete, 0) ?:
1592                         bch2_trans_commit(trans, &disk_res, journal_seq,
1593                                         BTREE_INSERT_NOFAIL);
1594                 bch2_disk_reservation_put(trans->c, &disk_res);
1595                 if (ret)
1596                         break;
1597         }
1598
1599         if (ret == -EINTR) {
1600                 ret = 0;
1601                 goto retry;
1602         }
1603
1604         bch2_trans_iter_exit(trans, &iter);
1605         return ret;
1606 }
1607
1608 /*
1609  * bch_btree_delete_range - delete everything within a given range
1610  *
1611  * Range is a half open interval - [start, end)
1612  */
1613 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1614                             struct bpos start, struct bpos end,
1615                             unsigned iter_flags,
1616                             u64 *journal_seq)
1617 {
1618         return bch2_trans_do(c, NULL, journal_seq, 0,
1619                              bch2_btree_delete_range_trans(&trans, id, start, end,
1620                                                            iter_flags, journal_seq));
1621 }