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 inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  27                                          const struct btree_insert_entry *r)
  28 {
  29         return   cmp_int(l->btree_id,   r->btree_id) ?:
  30                  -cmp_int(l->level,     r->level) ?:
  31                  bpos_cmp(l->k->k.p,    r->k->k.p);
  32 }
  33
  34 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  35 {
  36         return i->path->l + i->level;
  37 }
  38
  39 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  40                                      struct btree_insert_entry *i)
  41 {
  42         return i != trans->updates &&
  43                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  44 }
  45
  46 static inline bool same_leaf_as_next(struct btree_trans *trans,
  47                                      struct btree_insert_entry *i)
  48 {
  49         return i + 1 < trans->updates + trans->nr_updates &&
  50                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  51 }
  52
  53 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  54                                                   struct btree_path *path,
  55                                                   struct btree *b)
  56 {
  57         struct bch_fs *c = trans->c;
  58
  59         if (path->cached)
  60                 return;
  61
  62         if (unlikely(btree_node_just_written(b)) &&
  63             bch2_btree_post_write_cleanup(c, b))
  64                 bch2_trans_node_reinit_iter(trans, b);
  65
  66         /*
  67          * If the last bset has been written, or if it's gotten too big - start
  68          * a new bset to insert into:
  69          */
  70         if (want_new_bset(c, b))
  71                 bch2_btree_init_next(trans, b);
  72 }
  73
  74 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
  75                                      struct btree_path *path,
  76                                      struct btree *b)
  77 {
  78         bch2_btree_node_lock_write(trans, path, b);
  79         bch2_btree_node_prep_for_write(trans, path, b);
  80 }
  81
  82 /* Inserting into a given leaf node (last stage of insert): */
  83
  84 /* Handle overwrites and do insert, for non extents: */
  85 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  86                                 struct btree_path *path,
  87                                 struct btree *b,
  88                                 struct btree_node_iter *node_iter,
  89                                 struct bkey_i *insert)
  90 {
  91         struct bkey_packed *k;
  92         unsigned clobber_u64s = 0, new_u64s = 0;
  93
  94         EBUG_ON(btree_node_just_written(b));
  95         EBUG_ON(bset_written(b, btree_bset_last(b)));
  96         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  97         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  98         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  99         EBUG_ON(insert->k.u64s >
 100                 bch_btree_keys_u64s_remaining(trans->c, b));
 101
 102         k = bch2_btree_node_iter_peek_all(node_iter, b);
 103         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 104                 k = NULL;
 105
 106         /* @k is the key being overwritten/deleted, if any: */
 107         EBUG_ON(k && bkey_deleted(k));
 108
 109         /* Deleting, but not found? nothing to do: */
 110         if (bkey_deleted(&insert->k) && !k)
 111                 return false;
 112
 113         if (bkey_deleted(&insert->k)) {
 114                 /* Deleting: */
 115                 btree_account_key_drop(b, k);
 116                 k->type = KEY_TYPE_deleted;
 117
 118                 if (k->needs_whiteout)
 119                         push_whiteout(trans->c, b, insert->k.p);
 120                 k->needs_whiteout = false;
 121
 122                 if (k >= btree_bset_last(b)->start) {
 123                         clobber_u64s = k->u64s;
 124                         bch2_bset_delete(b, k, clobber_u64s);
 125                         goto fix_iter;
 126                 } else {
 127                         bch2_btree_path_fix_key_modified(trans, b, k);
 128                 }
 129
 130                 return true;
 131         }
 132
 133         if (k) {
 134                 /* Overwriting: */
 135                 btree_account_key_drop(b, k);
 136                 k->type = KEY_TYPE_deleted;
 137
 138                 insert->k.needs_whiteout = k->needs_whiteout;
 139                 k->needs_whiteout = false;
 140
 141                 if (k >= btree_bset_last(b)->start) {
 142                         clobber_u64s = k->u64s;
 143                         goto overwrite;
 144                 } else {
 145                         bch2_btree_path_fix_key_modified(trans, b, k);
 146                 }
 147         }
 148
 149         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 150 overwrite:
 151         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 152         new_u64s = k->u64s;
 153 fix_iter:
 154         if (clobber_u64s != new_u64s)
 155                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 156                                          clobber_u64s, new_u64s);
 157         return true;
 158 }
 159
 160 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 161                                unsigned i, u64 seq)
 162 {
 163         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 164         struct btree_write *w = container_of(pin, struct btree_write, journal);
 165         struct btree *b = container_of(w, struct btree, writes[i]);
 166
 167         btree_node_lock_type(c, b, SIX_LOCK_read);
 168         bch2_btree_node_write_cond(c, b,
 169                 (btree_current_write(b) == w && w->journal.seq == seq));
 170         six_unlock_read(&b->c.lock);
 171         return 0;
 172 }
 173
 174 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 175 {
 176         return __btree_node_flush(j, pin, 0, seq);
 177 }
 178
 179 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 180 {
 181         return __btree_node_flush(j, pin, 1, seq);
 182 }
 183
 184 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 185                                        struct btree *b, u64 seq)
 186 {
 187         struct btree_write *w = btree_current_write(b);
 188
 189         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 190                              btree_node_write_idx(b) == 0
 191                              ? btree_node_flush0
 192                              : btree_node_flush1);
 193 }
 194
 195 /**
 196  * btree_insert_key - insert a key one key into a leaf node
 197  */
 198 static bool btree_insert_key_leaf(struct btree_trans *trans,
 199                                   struct btree_insert_entry *insert)
 200 {
 201         struct bch_fs *c = trans->c;
 202         struct btree *b = insert_l(insert)->b;
 203         struct bset_tree *t = bset_tree_last(b);
 204         struct bset *i = bset(b, t);
 205         int old_u64s = bset_u64s(t);
 206         int old_live_u64s = b->nr.live_u64s;
 207         int live_u64s_added, u64s_added;
 208
 209         EBUG_ON(!insert->level &&
 210                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 211
 212         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 213                                         &insert_l(insert)->iter, insert->k)))
 214                 return false;
 215
 216         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 217                                          le64_to_cpu(i->journal_seq)));
 218
 219         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 220
 221         if (unlikely(!btree_node_dirty(b)))
 222                 set_btree_node_dirty(c, b);
 223
 224         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 225         u64s_added = (int) bset_u64s(t) - old_u64s;
 226
 227         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 228                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 229         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 230                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 231
 232         if (u64s_added > live_u64s_added &&
 233             bch2_maybe_compact_whiteouts(c, b))
 234                 bch2_trans_node_reinit_iter(trans, b);
 235
 236         return true;
 237 }
 238
 239 /* Cached btree updates: */
 240
 241 /* Normal update interface: */
 242
 243 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 244                                              struct btree_insert_entry *i)
 245 {
 246         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 247         BUG_ON(i->cached        != i->path->cached);
 248         BUG_ON(i->level         != i->path->level);
 249         BUG_ON(i->btree_id      != i->path->btree_id);
 250         EBUG_ON(!i->level &&
 251                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 252                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 253                 i->k->k.p.snapshot &&
 254                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 255 }
 256
 257 static noinline int
 258 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 259                                    unsigned long trace_ip)
 260 {
 261         struct bch_fs *c = trans->c;
 262         int ret;
 263
 264         bch2_trans_unlock(trans);
 265
 266         ret = bch2_journal_preres_get(&c->journal,
 267                         &trans->journal_preres, u64s, 0);
 268         if (ret)
 269                 return ret;
 270
 271         if (!bch2_trans_relock(trans)) {
 272                 trace_trans_restart_journal_preres_get(trans->ip, trace_ip);
 273                 return -EINTR;
 274         }
 275
 276         return 0;
 277 }
 278
 279 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 280                                              unsigned flags)
 281 {
 282         struct bch_fs *c = trans->c;
 283         int ret;
 284
 285         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 286                 flags |= JOURNAL_RES_GET_RESERVED;
 287
 288         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 289                                    trans->journal_u64s, flags);
 290
 291         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 292 }
 293
 294 #define JSET_ENTRY_LOG_U64s             4
 295
 296 static noinline void journal_transaction_name(struct btree_trans *trans)
 297 {
 298         struct bch_fs *c = trans->c;
 299         struct jset_entry *entry = journal_res_entry(&c->journal, &trans->journal_res);
 300         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 301         unsigned u64s = JSET_ENTRY_LOG_U64s - 1;
 302         unsigned b, buflen = u64s * sizeof(u64);
 303
 304         l->entry.u64s           = cpu_to_le16(u64s);
 305         l->entry.btree_id       = 0;
 306         l->entry.level          = 0;
 307         l->entry.type           = BCH_JSET_ENTRY_log;
 308         l->entry.pad[0]         = 0;
 309         l->entry.pad[1]         = 0;
 310         l->entry.pad[2]         = 0;
 311         b = snprintf(l->d, buflen, "%ps", (void *) trans->ip);
 312         while (b < buflen)
 313                 l->d[b++] = '\0';
 314
 315         trans->journal_res.offset       += JSET_ENTRY_LOG_U64s;
 316         trans->journal_res.u64s         -= JSET_ENTRY_LOG_U64s;
 317 }
 318
 319 static inline enum btree_insert_ret
 320 btree_key_can_insert(struct btree_trans *trans,
 321                      struct btree *b,
 322                      unsigned u64s)
 323 {
 324         struct bch_fs *c = trans->c;
 325
 326         if (!bch2_btree_node_insert_fits(c, b, u64s))
 327                 return BTREE_INSERT_BTREE_NODE_FULL;
 328
 329         return BTREE_INSERT_OK;
 330 }
 331
 332 static enum btree_insert_ret
 333 btree_key_can_insert_cached(struct btree_trans *trans,
 334                             struct btree_path *path,
 335                             unsigned u64s)
 336 {
 337         struct bch_fs *c = trans->c;
 338         struct bkey_cached *ck = (void *) path->l[0].b;
 339         unsigned new_u64s;
 340         struct bkey_i *new_k;
 341
 342         EBUG_ON(path->level);
 343
 344         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 345             bch2_btree_key_cache_must_wait(c) &&
 346             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 347                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 348
 349         /*
 350          * bch2_varint_decode can read past the end of the buffer by at most 7
 351          * bytes (it won't be used):
 352          */
 353         u64s += 1;
 354
 355         if (u64s <= ck->u64s)
 356                 return BTREE_INSERT_OK;
 357
 358         new_u64s        = roundup_pow_of_two(u64s);
 359         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 360         if (!new_k) {
 361                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 362                         bch2_btree_ids[path->btree_id], new_u64s);
 363                 return -ENOMEM;
 364         }
 365
 366         ck->u64s        = new_u64s;
 367         ck->k           = new_k;
 368         return BTREE_INSERT_OK;
 369 }
 370
 371 static inline void do_btree_insert_one(struct btree_trans *trans,
 372                                        struct btree_insert_entry *i)
 373 {
 374         struct bch_fs *c = trans->c;
 375         struct journal *j = &c->journal;
 376         bool did_work;
 377
 378         EBUG_ON(trans->journal_res.ref !=
 379                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 380
 381         i->k->k.needs_whiteout = false;
 382
 383         did_work = !i->cached
 384                 ? btree_insert_key_leaf(trans, i)
 385                 : bch2_btree_insert_key_cached(trans, i->path, i->k);
 386         if (!did_work)
 387                 return;
 388
 389         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 390                 bch2_journal_add_keys(j, &trans->journal_res,
 391                                       i->btree_id,
 392                                       i->level,
 393                                       i->k);
 394
 395                 if (trans->journal_seq)
 396                         *trans->journal_seq = trans->journal_res.seq;
 397         }
 398 }
 399
 400 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 401 {
 402         struct bch_fs *c = trans->c;
 403         struct btree_insert_entry *i;
 404
 405         trans_for_each_update(trans, i) {
 406                 /*
 407                  * XXX: synchronization of cached update triggers with gc
 408                  * XXX: synchronization of interior node updates with gc
 409                  */
 410                 BUG_ON(i->cached || i->level);
 411
 412                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b)))
 413                         bch2_mark_update(trans, i->path, i->k,
 414                                          i->flags|BTREE_TRIGGER_GC);
 415         }
 416 }
 417
 418 static inline int
 419 bch2_trans_commit_write_locked(struct btree_trans *trans,
 420                                struct btree_insert_entry **stopped_at,
 421                                unsigned long trace_ip)
 422 {
 423         struct bch_fs *c = trans->c;
 424         struct btree_insert_entry *i;
 425         struct btree_trans_commit_hook *h;
 426         unsigned u64s = 0;
 427         bool marking = false;
 428         int ret;
 429
 430         if (race_fault()) {
 431                 trace_trans_restart_fault_inject(trans->ip, trace_ip);
 432                 trans->restarted = true;
 433                 return -EINTR;
 434         }
 435
 436         /*
 437          * Check if the insert will fit in the leaf node with the write lock
 438          * held, otherwise another thread could write the node changing the
 439          * amount of space available:
 440          */
 441
 442         prefetch(&trans->c->journal.flags);
 443
 444         h = trans->hooks;
 445         while (h) {
 446                 ret = h->fn(trans, h);
 447                 if (ret)
 448                         return ret;
 449                 h = h->next;
 450         }
 451
 452         trans_for_each_update(trans, i) {
 453                 /* Multiple inserts might go to same leaf: */
 454                 if (!same_leaf_as_prev(trans, i))
 455                         u64s = 0;
 456
 457                 u64s += i->k->k.u64s;
 458                 ret = !i->cached
 459                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 460                         : btree_key_can_insert_cached(trans, i->path, u64s);
 461                 if (ret) {
 462                         *stopped_at = i;
 463                         return ret;
 464                 }
 465
 466                 if (btree_node_type_needs_gc(i->bkey_type))
 467                         marking = true;
 468         }
 469
 470         /*
 471          * Don't get journal reservation until after we know insert will
 472          * succeed:
 473          */
 474         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 475                 ret = bch2_trans_journal_res_get(trans,
 476                                 JOURNAL_RES_GET_NONBLOCK);
 477                 if (ret)
 478                         return ret;
 479         } else {
 480                 trans->journal_res.seq = c->journal.replay_journal_seq;
 481         }
 482
 483         if (unlikely(trans->journal_transaction_names))
 484                 journal_transaction_name(trans);
 485
 486         if (unlikely(trans->extra_journal_entry_u64s)) {
 487                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 488                                   trans->extra_journal_entries,
 489                                   trans->extra_journal_entry_u64s);
 490
 491                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 492                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 493         }
 494
 495         /*
 496          * Not allowed to fail after we've gotten our journal reservation - we
 497          * have to use it:
 498          */
 499
 500         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 501                 if (bch2_journal_seq_verify)
 502                         trans_for_each_update(trans, i)
 503                                 i->k->k.version.lo = trans->journal_res.seq;
 504                 else if (bch2_inject_invalid_keys)
 505                         trans_for_each_update(trans, i)
 506                                 i->k->k.version = MAX_VERSION;
 507         }
 508
 509         if (trans->fs_usage_deltas &&
 510             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 511                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 512
 513         trans_for_each_update(trans, i)
 514                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 515                         bch2_mark_update(trans, i->path, i->k, i->flags);
 516
 517         if (unlikely(c->gc_pos.phase))
 518                 bch2_trans_mark_gc(trans);
 519
 520         trans_for_each_update(trans, i)
 521                 do_btree_insert_one(trans, i);
 522
 523         return ret;
 524 }
 525
 526 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 527 {
 528         unsigned l;
 529
 530         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 531                 if (btree_node_read_locked(path, l))
 532                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 533 }
 534
 535 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 536 {
 537         struct btree *b = path_l(path)->b;
 538
 539         do {
 540                 if (path->nodes_locked &&
 541                     path->nodes_locked != path->nodes_intent_locked)
 542                         path_upgrade_readers(trans, path);
 543         } while ((path = prev_btree_path(trans, path)) &&
 544                  path_l(path)->b == b);
 545 }
 546
 547 /*
 548  * Check for nodes that we have both read and intent locks on, and upgrade the
 549  * readers to intent:
 550  */
 551 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 552 {
 553         struct btree_path *path;
 554         unsigned i, nr_read = 0, nr_intent = 0;
 555
 556         trans_for_each_path_inorder(trans, path, i) {
 557                 struct btree_path *next = i + 1 < trans->nr_sorted
 558                         ? trans->paths + trans->sorted[i + 1]
 559                         : NULL;
 560
 561                 if (path->nodes_locked) {
 562                         if (path->nodes_intent_locked)
 563                                 nr_intent++;
 564                         else
 565                                 nr_read++;
 566                 }
 567
 568                 if (!next || path_l(path)->b != path_l(next)->b) {
 569                         if (nr_read && nr_intent)
 570                                 upgrade_readers(trans, path);
 571
 572                         nr_read = nr_intent = 0;
 573                 }
 574         }
 575
 576         bch2_trans_verify_locks(trans);
 577 }
 578
 579 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 580 {
 581         struct btree_path *path;
 582         unsigned i;
 583
 584         trans_for_each_path_inorder(trans, path, i) {
 585                 //if (path == pos)
 586                 //      break;
 587
 588                 if (path->nodes_locked != path->nodes_intent_locked &&
 589                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 590                         return true;
 591         }
 592
 593         return false;
 594 }
 595
 596 static inline int trans_lock_write(struct btree_trans *trans)
 597 {
 598         struct btree_insert_entry *i;
 599
 600         trans_for_each_update(trans, i) {
 601                 if (same_leaf_as_prev(trans, i))
 602                         continue;
 603
 604                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 605                         if (have_conflicting_read_lock(trans, i->path))
 606                                 goto fail;
 607
 608                         __btree_node_lock_type(trans->c, insert_l(i)->b,
 609                                                SIX_LOCK_write);
 610                 }
 611
 612                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 613         }
 614
 615         return 0;
 616 fail:
 617         while (--i >= trans->updates) {
 618                 if (same_leaf_as_prev(trans, i))
 619                         continue;
 620
 621                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 622         }
 623
 624         trace_trans_restart_would_deadlock_write(trans->ip);
 625         return btree_trans_restart(trans);
 626 }
 627
 628 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 629 {
 630         struct btree_insert_entry *i;
 631
 632         trans_for_each_update(trans, i)
 633                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 634 }
 635
 636 /*
 637  * Get journal reservation, take write locks, and attempt to do btree update(s):
 638  */
 639 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 640                                        struct btree_insert_entry **stopped_at,
 641                                        unsigned long trace_ip)
 642 {
 643         struct bch_fs *c = trans->c;
 644         struct btree_insert_entry *i;
 645         struct bkey_s_c old;
 646         int ret, u64s_delta = 0;
 647
 648         trans_for_each_update(trans, i) {
 649                 const char *invalid = bch2_bkey_invalid(c,
 650                                 bkey_i_to_s_c(i->k), i->bkey_type);
 651                 if (invalid) {
 652                         char buf[200];
 653
 654                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 655                         bch_err(c, "invalid bkey %s on insert from %ps -> %ps: %s\n",
 656                                 buf, (void *) trans->ip,
 657                                 (void *) i->ip_allocated, invalid);
 658                         bch2_fatal_error(c);
 659                         return -EINVAL;
 660                 }
 661                 btree_insert_entry_checks(trans, i);
 662         }
 663
 664         trans_for_each_update(trans, i) {
 665                 struct bkey u;
 666
 667                 /*
 668                  * peek_slot() doesn't yet work on iterators that point to
 669                  * interior nodes:
 670                  */
 671                 if (i->cached || i->level)
 672                         continue;
 673
 674                 old = bch2_btree_path_peek_slot(i->path, &u);
 675                 ret = bkey_err(old);
 676                 if (unlikely(ret))
 677                         return ret;
 678
 679                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 680                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 681
 682                 if (!same_leaf_as_next(trans, i)) {
 683                         if (u64s_delta <= 0) {
 684                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 685                                                         i->level, trans->flags);
 686                                 if (unlikely(ret))
 687                                         return ret;
 688                         }
 689
 690                         u64s_delta = 0;
 691                 }
 692         }
 693
 694         ret = bch2_journal_preres_get(&c->journal,
 695                         &trans->journal_preres, trans->journal_preres_u64s,
 696                         JOURNAL_RES_GET_NONBLOCK|
 697                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 698                          ? JOURNAL_RES_GET_RESERVED : 0));
 699         if (unlikely(ret == -EAGAIN))
 700                 ret = bch2_trans_journal_preres_get_cold(trans,
 701                                                 trans->journal_preres_u64s, trace_ip);
 702         if (unlikely(ret))
 703                 return ret;
 704
 705         normalize_read_intent_locks(trans);
 706
 707         ret = trans_lock_write(trans);
 708         if (unlikely(ret))
 709                 return ret;
 710
 711         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 712
 713         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 714                 bch2_drop_overwrites_from_journal(trans);
 715
 716         trans_for_each_update(trans, i)
 717                 if (!same_leaf_as_prev(trans, i))
 718                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 719                                                         insert_l(i)->b);
 720
 721         if (!ret && trans->journal_pin)
 722                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 723                                      trans->journal_pin, NULL);
 724
 725         /*
 726          * Drop journal reservation after dropping write locks, since dropping
 727          * the journal reservation may kick off a journal write:
 728          */
 729         bch2_journal_res_put(&c->journal, &trans->journal_res);
 730
 731         if (unlikely(ret))
 732                 return ret;
 733
 734         bch2_trans_downgrade(trans);
 735
 736         return 0;
 737 }
 738
 739 static int journal_reclaim_wait_done(struct bch_fs *c)
 740 {
 741         int ret = bch2_journal_error(&c->journal) ?:
 742                 !bch2_btree_key_cache_must_wait(c);
 743
 744         if (!ret)
 745                 journal_reclaim_kick(&c->journal);
 746         return ret;
 747 }
 748
 749 static noinline
 750 int bch2_trans_commit_error(struct btree_trans *trans,
 751                             struct btree_insert_entry *i,
 752                             int ret, unsigned long trace_ip)
 753 {
 754         struct bch_fs *c = trans->c;
 755
 756         switch (ret) {
 757         case BTREE_INSERT_BTREE_NODE_FULL:
 758                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 759                 if (!ret)
 760                         return 0;
 761
 762                 if (ret == -EINTR)
 763                         trace_trans_restart_btree_node_split(trans->ip, trace_ip,
 764                                                 i->btree_id, &i->path->pos);
 765                 break;
 766         case BTREE_INSERT_NEED_MARK_REPLICAS:
 767                 bch2_trans_unlock(trans);
 768
 769                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 770                 if (ret)
 771                         break;
 772
 773                 if (bch2_trans_relock(trans))
 774                         return 0;
 775
 776                 trace_trans_restart_mark_replicas(trans->ip, trace_ip);
 777                 ret = -EINTR;
 778                 break;
 779         case BTREE_INSERT_NEED_JOURNAL_RES:
 780                 bch2_trans_unlock(trans);
 781
 782                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 783                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
 784                         trans->restarted = true;
 785                         ret = -EAGAIN;
 786                         break;
 787                 }
 788
 789                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 790                 if (ret)
 791                         break;
 792
 793                 if (bch2_trans_relock(trans))
 794                         return 0;
 795
 796                 trace_trans_restart_journal_res_get(trans->ip, trace_ip);
 797                 ret = -EINTR;
 798                 break;
 799         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 800                 bch2_trans_unlock(trans);
 801
 802                 trace_trans_blocked_journal_reclaim(trans->ip, trace_ip);
 803
 804                 wait_event_freezable(c->journal.reclaim_wait,
 805                                      (ret = journal_reclaim_wait_done(c)));
 806                 if (ret < 0)
 807                         break;
 808
 809                 if (bch2_trans_relock(trans))
 810                         return 0;
 811
 812                 trace_trans_restart_journal_reclaim(trans->ip, trace_ip);
 813                 ret = -EINTR;
 814                 break;
 815         default:
 816                 BUG_ON(ret >= 0);
 817                 break;
 818         }
 819
 820         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
 821         BUG_ON(ret == -ENOSPC && (trans->flags & BTREE_INSERT_NOFAIL));
 822
 823         return ret;
 824 }
 825
 826 static noinline int
 827 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 828 {
 829         struct bch_fs *c = trans->c;
 830         int ret;
 831
 832         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 833                 return -EROFS;
 834
 835         bch2_trans_unlock(trans);
 836
 837         ret = bch2_fs_read_write_early(c);
 838         if (ret)
 839                 return ret;
 840
 841         if (!bch2_trans_relock(trans))
 842                 return -EINTR;
 843
 844         percpu_ref_get(&c->writes);
 845         return 0;
 846 }
 847
 848 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 849 {
 850         struct bkey             _deleted = KEY(0, 0, 0);
 851         struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 852         struct bkey_s_c         old;
 853         struct bkey             unpacked;
 854         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 855         bool trans_trigger_run;
 856         unsigned btree_id = 0;
 857         int ret = 0;
 858
 859         /*
 860          *
 861          * For a given btree, this algorithm runs insert triggers before
 862          * overwrite triggers: this is so that when extents are being moved
 863          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 864          * they are re-added.
 865          */
 866         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 867                 while (btree_id_start < trans->updates + trans->nr_updates &&
 868                        btree_id_start->btree_id < btree_id)
 869                         btree_id_start++;
 870
 871                 /*
 872                  * Running triggers will append more updates to the list of updates as
 873                  * we're walking it:
 874                  */
 875                 do {
 876                         trans_trigger_run = false;
 877
 878                         for (i = btree_id_start;
 879                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 880                              i++) {
 881                                 if (i->insert_trigger_run ||
 882                                     (i->flags & BTREE_TRIGGER_NORUN) ||
 883                                     !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 884                                         continue;
 885
 886                                 BUG_ON(i->overwrite_trigger_run);
 887
 888                                 i->insert_trigger_run = true;
 889                                 trans_trigger_run = true;
 890
 891                                 old = bch2_btree_path_peek_slot(i->path, &unpacked);
 892                                 _deleted.p = i->path->pos;
 893
 894                                 if (old.k->type == i->k->k.type &&
 895                                     ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 896                                         i->overwrite_trigger_run = true;
 897                                         ret = bch2_trans_mark_key(trans, old, bkey_i_to_s_c(i->k),
 898                                                         BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|i->flags);
 899                                 } else {
 900                                         ret = bch2_trans_mark_key(trans, deleted, bkey_i_to_s_c(i->k),
 901                                                         BTREE_TRIGGER_INSERT|i->flags);
 902                                 }
 903
 904                                 if (ret == -EINTR)
 905                                         trace_trans_restart_mark(trans->ip, _RET_IP_,
 906                                                         i->btree_id, &i->path->pos);
 907                                 if (ret)
 908                                         return ret;
 909                         }
 910                 } while (trans_trigger_run);
 911
 912                 do {
 913                         trans_trigger_run = false;
 914
 915                         for (i = btree_id_start;
 916                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 917                              i++) {
 918                                 if (i->overwrite_trigger_run ||
 919                                     (i->flags & BTREE_TRIGGER_NORUN) ||
 920                                     !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 921                                         continue;
 922
 923                                 BUG_ON(!i->insert_trigger_run);
 924
 925                                 i->overwrite_trigger_run = true;
 926                                 trans_trigger_run = true;
 927
 928                                 old = bch2_btree_path_peek_slot(i->path, &unpacked);
 929                                 _deleted.p = i->path->pos;
 930
 931                                 ret = bch2_trans_mark_key(trans, old, deleted,
 932                                                 BTREE_TRIGGER_OVERWRITE|i->flags);
 933
 934                                 if (ret == -EINTR)
 935                                         trace_trans_restart_mark(trans->ip, _RET_IP_,
 936                                                         i->btree_id, &i->path->pos);
 937                                 if (ret)
 938                                         return ret;
 939                         }
 940                 } while (trans_trigger_run);
 941         }
 942
 943         trans_for_each_update(trans, i)
 944                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 945                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 946                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 947
 948         return 0;
 949 }
 950
 951 int __bch2_trans_commit(struct btree_trans *trans)
 952 {
 953         struct bch_fs *c = trans->c;
 954         struct btree_insert_entry *i = NULL;
 955         unsigned u64s;
 956         int ret = 0;
 957
 958         if (!trans->nr_updates &&
 959             !trans->extra_journal_entry_u64s)
 960                 goto out_reset;
 961
 962         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 963                 lockdep_assert_held(&c->gc_lock);
 964
 965         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 966
 967         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 968         trans->journal_preres_u64s      = 0;
 969
 970         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
 971
 972         if (trans->journal_transaction_names)
 973                 trans->journal_u64s += JSET_ENTRY_LOG_U64s;
 974
 975         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 976             unlikely(!percpu_ref_tryget(&c->writes))) {
 977                 ret = bch2_trans_commit_get_rw_cold(trans);
 978                 if (ret)
 979                         goto out_reset;
 980         }
 981
 982 #ifdef CONFIG_BCACHEFS_DEBUG
 983         /*
 984          * if BTREE_TRIGGER_NORUN is set, it means we're probably being called
 985          * from the key cache flush code:
 986          */
 987         trans_for_each_update(trans, i)
 988                 if (!i->cached &&
 989                     !(i->flags & BTREE_TRIGGER_NORUN))
 990                         bch2_btree_key_cache_verify_clean(trans,
 991                                         i->btree_id, i->k->k.p);
 992 #endif
 993
 994         ret = bch2_trans_commit_run_triggers(trans);
 995         if (ret)
 996                 goto out;
 997
 998         trans_for_each_update(trans, i) {
 999                 BUG_ON(!i->path->should_be_locked);
1000
1001                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1002                         trace_trans_restart_upgrade(trans->ip, _RET_IP_,
1003                                                     i->btree_id, &i->path->pos);
1004                         ret = btree_trans_restart(trans);
1005                         goto out;
1006                 }
1007
1008                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1009
1010                 u64s = jset_u64s(i->k->k.u64s);
1011                 if (i->cached &&
1012                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1013                         trans->journal_preres_u64s += u64s;
1014                 trans->journal_u64s += u64s;
1015         }
1016
1017         if (trans->extra_journal_res) {
1018                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1019                                 trans->extra_journal_res,
1020                                 (trans->flags & BTREE_INSERT_NOFAIL)
1021                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1022                 if (ret)
1023                         goto err;
1024         }
1025 retry:
1026         BUG_ON(trans->restarted);
1027         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1028
1029         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1030
1031         /* make sure we didn't drop or screw up locks: */
1032         bch2_trans_verify_locks(trans);
1033
1034         if (ret)
1035                 goto err;
1036 out:
1037         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1038
1039         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1040                 percpu_ref_put(&c->writes);
1041 out_reset:
1042         trans_for_each_update(trans, i)
1043                 bch2_path_put(trans, i->path, true);
1044
1045         trans->extra_journal_res        = 0;
1046         trans->nr_updates               = 0;
1047         trans->hooks                    = NULL;
1048         trans->extra_journal_entries    = NULL;
1049         trans->extra_journal_entry_u64s = 0;
1050
1051         if (trans->fs_usage_deltas) {
1052                 trans->fs_usage_deltas->used = 0;
1053                 memset(&trans->fs_usage_deltas->memset_start, 0,
1054                        (void *) &trans->fs_usage_deltas->memset_end -
1055                        (void *) &trans->fs_usage_deltas->memset_start);
1056         }
1057
1058         return ret;
1059 err:
1060         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1061         if (ret)
1062                 goto out;
1063
1064         goto retry;
1065 }
1066
1067 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1068                                           enum btree_id id,
1069                                           struct bpos pos)
1070 {
1071         struct bch_fs *c = trans->c;
1072         struct btree_iter iter;
1073         struct bkey_s_c k;
1074         int ret;
1075
1076         if (!snapshot_t(c, pos.snapshot)->children[0])
1077                 return 0;
1078
1079         bch2_trans_iter_init(trans, &iter, id, pos,
1080                              BTREE_ITER_NOT_EXTENTS|
1081                              BTREE_ITER_ALL_SNAPSHOTS);
1082         while (1) {
1083                 k = bch2_btree_iter_prev(&iter);
1084                 ret = bkey_err(k);
1085                 if (ret)
1086                         break;
1087
1088                 if (!k.k)
1089                         break;
1090
1091                 if (bkey_cmp(pos, k.k->p))
1092                         break;
1093
1094                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1095                         ret = 1;
1096                         break;
1097                 }
1098         }
1099         bch2_trans_iter_exit(trans, &iter);
1100
1101         return ret;
1102 }
1103
1104 static int bch2_trans_update_extent(struct btree_trans *trans,
1105                                     struct btree_iter *orig_iter,
1106                                     struct bkey_i *insert,
1107                                     enum btree_update_flags flags)
1108 {
1109         struct bch_fs *c = trans->c;
1110         struct btree_iter iter, update_iter;
1111         struct bpos start = bkey_start_pos(&insert->k);
1112         struct bkey_i *update;
1113         struct bkey_s_c k;
1114         enum btree_id btree_id = orig_iter->btree_id;
1115         int ret = 0, compressed_sectors;
1116
1117         bch2_trans_iter_init(trans, &iter, btree_id, start,
1118                              BTREE_ITER_INTENT|
1119                              BTREE_ITER_WITH_UPDATES|
1120                              BTREE_ITER_NOT_EXTENTS);
1121         k = bch2_btree_iter_peek(&iter);
1122         if ((ret = bkey_err(k)))
1123                 goto err;
1124         if (!k.k)
1125                 goto out;
1126
1127         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1128                 /*
1129                  * We can't merge extents if they belong to interior snapshot
1130                  * tree nodes, and there's a snapshot in which one extent is
1131                  * visible and the other is not - i.e. if visibility is
1132                  * different.
1133                  *
1134                  * Instead of checking if visibilitiy of the two extents is
1135                  * different, for now we just check if either has been
1136                  * overwritten:
1137                  */
1138                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1139                 if (ret < 0)
1140                         goto err;
1141                 if (ret)
1142                         goto nomerge1;
1143
1144                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1145                 if (ret < 0)
1146                         goto err;
1147                 if (ret)
1148                         goto nomerge1;
1149
1150                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1151                 if ((ret = PTR_ERR_OR_ZERO(update)))
1152                         goto err;
1153
1154                 bkey_reassemble(update, k);
1155
1156                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1157                         ret = bch2_btree_delete_at(trans, &iter, flags);
1158                         if (ret)
1159                                 goto err;
1160
1161                         insert = update;
1162                         goto next;
1163                 }
1164         }
1165 nomerge1:
1166         ret = 0;
1167         if (!bkey_cmp(k.k->p, start))
1168                 goto next;
1169
1170         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1171                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1172                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1173
1174                 /*
1175                  * If we're going to be splitting a compressed extent, note it
1176                  * so that __bch2_trans_commit() can increase our disk
1177                  * reservation:
1178                  */
1179                 if (((front_split && back_split) ||
1180                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1181                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1182                         trans->extra_journal_res += compressed_sectors;
1183
1184                 if (front_split) {
1185                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1186                         if ((ret = PTR_ERR_OR_ZERO(update)))
1187                                 goto err;
1188
1189                         bkey_reassemble(update, k);
1190
1191                         bch2_cut_back(start, update);
1192
1193                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1194                                              BTREE_ITER_NOT_EXTENTS|
1195                                              BTREE_ITER_ALL_SNAPSHOTS|
1196                                              BTREE_ITER_INTENT);
1197                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1198                                 bch2_trans_update(trans, &update_iter, update,
1199                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1200                                                   flags);
1201                         bch2_trans_iter_exit(trans, &update_iter);
1202
1203                         if (ret)
1204                                 goto err;
1205                 }
1206
1207                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1208                     (front_split || back_split)) {
1209                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1210                         if ((ret = PTR_ERR_OR_ZERO(update)))
1211                                 goto err;
1212
1213                         bkey_reassemble(update, k);
1214
1215                         bch2_cut_front(start, update);
1216                         bch2_cut_back(insert->k.p, update);
1217
1218                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1219                                              BTREE_ITER_NOT_EXTENTS|
1220                                              BTREE_ITER_ALL_SNAPSHOTS|
1221                                              BTREE_ITER_INTENT);
1222                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1223                                 bch2_trans_update(trans, &update_iter, update,
1224                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1225                                                   flags);
1226                         bch2_trans_iter_exit(trans, &update_iter);
1227                         if (ret)
1228                                 goto err;
1229                 }
1230
1231                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1232                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1233                         if ((ret = PTR_ERR_OR_ZERO(update)))
1234                                 goto err;
1235
1236                         bkey_init(&update->k);
1237                         update->k.p = k.k->p;
1238
1239                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1240                                 update->k.p.snapshot = insert->k.p.snapshot;
1241                                 update->k.type = KEY_TYPE_whiteout;
1242                         }
1243
1244                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1245                                              BTREE_ITER_NOT_EXTENTS|
1246                                              BTREE_ITER_INTENT);
1247                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1248                                 bch2_trans_update(trans, &update_iter, update,
1249                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1250                                                   flags);
1251                         bch2_trans_iter_exit(trans, &update_iter);
1252
1253                         if (ret)
1254                                 goto err;
1255                 }
1256
1257                 if (back_split) {
1258                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1259                         if ((ret = PTR_ERR_OR_ZERO(update)))
1260                                 goto err;
1261
1262                         bkey_reassemble(update, k);
1263                         bch2_cut_front(insert->k.p, update);
1264
1265                         bch2_trans_copy_iter(&update_iter, &iter);
1266                         update_iter.pos = update->k.p;
1267                         ret   = bch2_trans_update(trans, &update_iter, update,
1268                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1269                                                   flags);
1270                         bch2_trans_iter_exit(trans, &update_iter);
1271
1272                         if (ret)
1273                                 goto err;
1274                         goto out;
1275                 }
1276 next:
1277                 k = bch2_btree_iter_next(&iter);
1278                 if ((ret = bkey_err(k)))
1279                         goto err;
1280                 if (!k.k)
1281                         goto out;
1282         }
1283
1284         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1285                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1286                 if (ret < 0)
1287                         goto err;
1288                 if (ret)
1289                         goto nomerge2;
1290
1291                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1292                 if (ret < 0)
1293                         goto err;
1294                 if (ret)
1295                         goto nomerge2;
1296
1297                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1298         }
1299 nomerge2:
1300         ret = 0;
1301 out:
1302         if (!bkey_deleted(&insert->k)) {
1303                 /*
1304                  * Rewinding iterators is expensive: get a new one and the one
1305                  * that points to the start of insert will be cloned from:
1306                  */
1307                 bch2_trans_iter_exit(trans, &iter);
1308                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1309                                      BTREE_ITER_NOT_EXTENTS|
1310                                      BTREE_ITER_INTENT);
1311                 ret   = bch2_btree_iter_traverse(&iter) ?:
1312                         bch2_trans_update(trans, &iter, insert, flags);
1313         }
1314 err:
1315         bch2_trans_iter_exit(trans, &iter);
1316
1317         return ret;
1318 }
1319
1320 /*
1321  * When deleting, check if we need to emit a whiteout (because we're overwriting
1322  * something in an ancestor snapshot)
1323  */
1324 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1325                                       enum btree_id btree_id, struct bpos pos)
1326 {
1327         struct btree_iter iter;
1328         struct bkey_s_c k;
1329         u32 snapshot = pos.snapshot;
1330         int ret;
1331
1332         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1333                 return 0;
1334
1335         pos.snapshot++;
1336
1337         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1338                            BTREE_ITER_ALL_SNAPSHOTS|
1339                            BTREE_ITER_NOPRESERVE, k, ret) {
1340                 if (bkey_cmp(k.k->p, pos))
1341                         break;
1342
1343                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1344                                               k.k->p.snapshot)) {
1345                         ret = !bkey_whiteout(k.k);
1346                         break;
1347                 }
1348         }
1349         bch2_trans_iter_exit(trans, &iter);
1350
1351         return ret;
1352 }
1353
1354 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1355                                    struct bkey_i *k, enum btree_update_flags flags)
1356 {
1357         struct btree_insert_entry *i, n;
1358
1359         BUG_ON(!iter->path->should_be_locked);
1360
1361         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1362                 return bch2_trans_update_extent(trans, iter, k, flags);
1363
1364         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1365         BUG_ON(bpos_cmp(k->k.p, iter->path->pos));
1366
1367         n = (struct btree_insert_entry) {
1368                 .flags          = flags,
1369                 .bkey_type      = __btree_node_type(iter->path->level, iter->btree_id),
1370                 .btree_id       = iter->btree_id,
1371                 .level          = iter->path->level,
1372                 .cached         = iter->flags & BTREE_ITER_CACHED,
1373                 .path           = iter->path,
1374                 .k              = k,
1375                 .ip_allocated   = _RET_IP_,
1376         };
1377
1378 #ifdef CONFIG_BCACHEFS_DEBUG
1379         trans_for_each_update(trans, i)
1380                 BUG_ON(i != trans->updates &&
1381                        btree_insert_entry_cmp(i - 1, i) >= 0);
1382 #endif
1383
1384         if (bkey_deleted(&n.k->k) &&
1385             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1386                 int ret = need_whiteout_for_snapshot(trans, n.btree_id, n.k->k.p);
1387                 if (unlikely(ret < 0))
1388                         return ret;
1389
1390                 if (ret)
1391                         n.k->k.type = KEY_TYPE_whiteout;
1392         }
1393
1394         /*
1395          * Pending updates are kept sorted: first, find position of new update,
1396          * then delete/trim any updates the new update overwrites:
1397          */
1398         trans_for_each_update(trans, i)
1399                 if (btree_insert_entry_cmp(&n, i) <= 0)
1400                         break;
1401
1402         if (i < trans->updates + trans->nr_updates &&
1403             !btree_insert_entry_cmp(&n, i)) {
1404                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1405
1406                 /*
1407                  * This is a hack to ensure that inode creates update the btree,
1408                  * not the key cache, which helps with cache coherency issues in
1409                  * other areas:
1410                  */
1411                 if (n.cached && !i->cached) {
1412                         i->k = n.k;
1413                         i->flags = n.flags;
1414                         return 0;
1415                 }
1416
1417                 bch2_path_put(trans, i->path, true);
1418                 *i = n;
1419         } else
1420                 array_insert_item(trans->updates, trans->nr_updates,
1421                                   i - trans->updates, n);
1422
1423         __btree_path_get(n.path, true);
1424
1425         return 0;
1426 }
1427
1428 void bch2_trans_commit_hook(struct btree_trans *trans,
1429                             struct btree_trans_commit_hook *h)
1430 {
1431         h->next = trans->hooks;
1432         trans->hooks = h;
1433 }
1434
1435 int __bch2_btree_insert(struct btree_trans *trans,
1436                         enum btree_id id, struct bkey_i *k)
1437 {
1438         struct btree_iter iter;
1439         int ret;
1440
1441         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1442                              BTREE_ITER_INTENT);
1443         ret   = bch2_btree_iter_traverse(&iter) ?:
1444                 bch2_trans_update(trans, &iter, k, 0);
1445         bch2_trans_iter_exit(trans, &iter);
1446         return ret;
1447 }
1448
1449 /**
1450  * bch2_btree_insert - insert keys into the extent btree
1451  * @c:                  pointer to struct bch_fs
1452  * @id:                 btree to insert into
1453  * @insert_keys:        list of keys to insert
1454  * @hook:               insert callback
1455  */
1456 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1457                       struct bkey_i *k,
1458                       struct disk_reservation *disk_res,
1459                       u64 *journal_seq, int flags)
1460 {
1461         return bch2_trans_do(c, disk_res, journal_seq, flags,
1462                              __bch2_btree_insert(&trans, id, k));
1463 }
1464
1465 int bch2_btree_delete_at(struct btree_trans *trans,
1466                          struct btree_iter *iter, unsigned update_flags)
1467 {
1468         struct bkey_i *k;
1469
1470         k = bch2_trans_kmalloc(trans, sizeof(*k));
1471         if (IS_ERR(k))
1472                 return PTR_ERR(k);
1473
1474         bkey_init(&k->k);
1475         k->k.p = iter->pos;
1476         return bch2_trans_update(trans, iter, k, update_flags);
1477 }
1478
1479 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1480                                   struct bpos start, struct bpos end,
1481                                   unsigned iter_flags,
1482                                   u64 *journal_seq)
1483 {
1484         struct btree_iter iter;
1485         struct bkey_s_c k;
1486         int ret = 0;
1487
1488         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT|iter_flags);
1489 retry:
1490         while ((bch2_trans_begin(trans),
1491                (k = bch2_btree_iter_peek(&iter)).k) &&
1492                !(ret = bkey_err(k)) &&
1493                bkey_cmp(iter.pos, end) < 0) {
1494                 struct disk_reservation disk_res =
1495                         bch2_disk_reservation_init(trans->c, 0);
1496                 struct bkey_i delete;
1497
1498                 bkey_init(&delete.k);
1499
1500                 /*
1501                  * This could probably be more efficient for extents:
1502                  */
1503
1504                 /*
1505                  * For extents, iter.pos won't necessarily be the same as
1506                  * bkey_start_pos(k.k) (for non extents they always will be the
1507                  * same). It's important that we delete starting from iter.pos
1508                  * because the range we want to delete could start in the middle
1509                  * of k.
1510                  *
1511                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1512                  * bkey_start_pos(k.k)).
1513                  */
1514                 delete.k.p = iter.pos;
1515
1516                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1517                         unsigned max_sectors =
1518                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1519
1520                         /* create the biggest key we can */
1521                         bch2_key_resize(&delete.k, max_sectors);
1522                         bch2_cut_back(end, &delete);
1523
1524                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1525                         if (ret)
1526                                 break;
1527                 }
1528
1529                 ret   = bch2_trans_update(trans, &iter, &delete, 0) ?:
1530                         bch2_trans_commit(trans, &disk_res, journal_seq,
1531                                         BTREE_INSERT_NOFAIL);
1532                 bch2_disk_reservation_put(trans->c, &disk_res);
1533                 if (ret)
1534                         break;
1535         }
1536
1537         if (ret == -EINTR) {
1538                 ret = 0;
1539                 goto retry;
1540         }
1541
1542         bch2_trans_iter_exit(trans, &iter);
1543         return ret;
1544 }
1545
1546 /*
1547  * bch_btree_delete_range - delete everything within a given range
1548  *
1549  * Range is a half open interval - [start, end)
1550  */
1551 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1552                             struct bpos start, struct bpos end,
1553                             unsigned iter_flags,
1554                             u64 *journal_seq)
1555 {
1556         return bch2_trans_do(c, NULL, journal_seq, 0,
1557                              bch2_btree_delete_range_trans(&trans, id, start, end,
1558                                                            iter_flags, journal_seq));
1559 }