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 "replicas.h"
  19
  20 #include <linux/prefetch.h>
  21 #include <linux/sort.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  25                                      struct btree_insert_entry *i)
  26 {
  27         return i != trans->updates2 &&
  28                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  29 }
  30
  31 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  32                                             struct btree_iter *iter)
  33 {
  34         bch2_btree_node_lock_write(b, iter);
  35
  36         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
  37                 return;
  38
  39         if (unlikely(btree_node_just_written(b)) &&
  40             bch2_btree_post_write_cleanup(c, b))
  41                 bch2_btree_iter_reinit_node(iter, b);
  42
  43         /*
  44          * If the last bset has been written, or if it's gotten too big - start
  45          * a new bset to insert into:
  46          */
  47         if (want_new_bset(c, b))
  48                 bch2_btree_init_next(c, b, iter);
  49 }
  50
  51 /* Inserting into a given leaf node (last stage of insert): */
  52
  53 /* Handle overwrites and do insert, for non extents: */
  54 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  55                                 struct btree *b,
  56                                 struct btree_node_iter *node_iter,
  57                                 struct bkey_i *insert)
  58 {
  59         struct bkey_packed *k;
  60         unsigned clobber_u64s = 0, new_u64s = 0;
  61
  62         EBUG_ON(btree_node_just_written(b));
  63         EBUG_ON(bset_written(b, btree_bset_last(b)));
  64         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  65         EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
  66                 bkey_cmp(bkey_start_pos(&insert->k),
  67                          bkey_predecessor(b->data->min_key)) < 0);
  68         EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
  69         EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
  70         EBUG_ON(insert->k.u64s >
  71                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  72         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  73
  74         k = bch2_btree_node_iter_peek_all(node_iter, b);
  75         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
  76                 k = NULL;
  77
  78         /* @k is the key being overwritten/deleted, if any: */
  79         EBUG_ON(k && bkey_whiteout(k));
  80
  81         /* Deleting, but not found? nothing to do: */
  82         if (bkey_whiteout(&insert->k) && !k)
  83                 return false;
  84
  85         if (bkey_whiteout(&insert->k)) {
  86                 /* Deleting: */
  87                 btree_account_key_drop(b, k);
  88                 k->type = KEY_TYPE_deleted;
  89
  90                 if (k->needs_whiteout)
  91                         push_whiteout(iter->trans->c, b, insert->k.p);
  92                 k->needs_whiteout = false;
  93
  94                 if (k >= btree_bset_last(b)->start) {
  95                         clobber_u64s = k->u64s;
  96                         bch2_bset_delete(b, k, clobber_u64s);
  97                         goto fix_iter;
  98                 } else {
  99                         bch2_btree_iter_fix_key_modified(iter, b, k);
 100                 }
 101
 102                 return true;
 103         }
 104
 105         if (k) {
 106                 /* Overwriting: */
 107                 btree_account_key_drop(b, k);
 108                 k->type = KEY_TYPE_deleted;
 109
 110                 insert->k.needs_whiteout = k->needs_whiteout;
 111                 k->needs_whiteout = false;
 112
 113                 if (k >= btree_bset_last(b)->start) {
 114                         clobber_u64s = k->u64s;
 115                         goto overwrite;
 116                 } else {
 117                         bch2_btree_iter_fix_key_modified(iter, b, k);
 118                 }
 119         }
 120
 121         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 122 overwrite:
 123         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 124         new_u64s = k->u64s;
 125 fix_iter:
 126         if (clobber_u64s != new_u64s)
 127                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 128                                          clobber_u64s, new_u64s);
 129         return true;
 130 }
 131
 132 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 133                                unsigned i, u64 seq)
 134 {
 135         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 136         struct btree_write *w = container_of(pin, struct btree_write, journal);
 137         struct btree *b = container_of(w, struct btree, writes[i]);
 138
 139         btree_node_lock_type(c, b, SIX_LOCK_read);
 140         bch2_btree_node_write_cond(c, b,
 141                 (btree_current_write(b) == w && w->journal.seq == seq));
 142         six_unlock_read(&b->c.lock);
 143 }
 144
 145 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 146 {
 147         return __btree_node_flush(j, pin, 0, seq);
 148 }
 149
 150 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 151 {
 152         return __btree_node_flush(j, pin, 1, seq);
 153 }
 154
 155 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 156                                        struct btree *b, u64 seq)
 157 {
 158         struct btree_write *w = btree_current_write(b);
 159
 160         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 161                              btree_node_write_idx(b) == 0
 162                              ? btree_node_flush0
 163                              : btree_node_flush1);
 164 }
 165
 166 /**
 167  * btree_insert_key - insert a key one key into a leaf node
 168  */
 169 static bool btree_insert_key_leaf(struct btree_trans *trans,
 170                                   struct btree_iter *iter,
 171                                   struct bkey_i *insert)
 172 {
 173         struct bch_fs *c = trans->c;
 174         struct btree *b = iter_l(iter)->b;
 175         struct bset_tree *t = bset_tree_last(b);
 176         struct bset *i = bset(b, t);
 177         int old_u64s = bset_u64s(t);
 178         int old_live_u64s = b->nr.live_u64s;
 179         int live_u64s_added, u64s_added;
 180
 181         EBUG_ON(!iter->level &&
 182                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 183
 184         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 185                                         &iter_l(iter)->iter, insert)))
 186                 return false;
 187
 188         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 189                                          le64_to_cpu(i->journal_seq)));
 190
 191         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 192
 193         if (unlikely(!btree_node_dirty(b)))
 194                 set_btree_node_dirty(c, b);
 195
 196         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 197         u64s_added = (int) bset_u64s(t) - old_u64s;
 198
 199         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 200                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 201         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 202                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 203
 204         if (u64s_added > live_u64s_added &&
 205             bch2_maybe_compact_whiteouts(c, b))
 206                 bch2_btree_iter_reinit_node(iter, b);
 207
 208         trace_btree_insert_key(c, b, insert);
 209         return true;
 210 }
 211
 212 /* Cached btree updates: */
 213
 214 /* Normal update interface: */
 215
 216 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 217                                              struct btree_iter *iter,
 218                                              struct bkey_i *insert)
 219 {
 220         struct bch_fs *c = trans->c;
 221
 222         BUG_ON(bkey_cmp(insert->k.p, iter->pos));
 223         BUG_ON(bch2_debug_check_bkeys &&
 224                bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
 225                                  __btree_node_type(iter->level, iter->btree_id)));
 226 }
 227
 228 static noinline int
 229 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 230 {
 231         struct bch_fs *c = trans->c;
 232         int ret;
 233
 234         bch2_trans_unlock(trans);
 235
 236         ret = bch2_journal_preres_get(&c->journal,
 237                         &trans->journal_preres, u64s, 0);
 238         if (ret)
 239                 return ret;
 240
 241         if (!bch2_trans_relock(trans)) {
 242                 trace_trans_restart_journal_preres_get(trans->ip);
 243                 return -EINTR;
 244         }
 245
 246         return 0;
 247 }
 248
 249 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 250                                              unsigned flags)
 251 {
 252         struct bch_fs *c = trans->c;
 253         int ret;
 254
 255         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 256                 flags |= JOURNAL_RES_GET_RESERVED;
 257
 258         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 259                                    trans->journal_u64s, flags);
 260
 261         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 262 }
 263
 264 static enum btree_insert_ret
 265 btree_key_can_insert(struct btree_trans *trans,
 266                      struct btree_iter *iter,
 267                      unsigned u64s)
 268 {
 269         struct bch_fs *c = trans->c;
 270         struct btree *b = iter_l(iter)->b;
 271
 272         if (!bch2_btree_node_insert_fits(c, b, u64s))
 273                 return BTREE_INSERT_BTREE_NODE_FULL;
 274
 275         return BTREE_INSERT_OK;
 276 }
 277
 278 static enum btree_insert_ret
 279 btree_key_can_insert_cached(struct btree_trans *trans,
 280                             struct btree_iter *iter,
 281                             unsigned u64s)
 282 {
 283         struct bkey_cached *ck = (void *) iter->l[0].b;
 284         unsigned new_u64s;
 285         struct bkey_i *new_k;
 286
 287         BUG_ON(iter->level);
 288
 289         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 290             bch2_btree_key_cache_must_wait(trans->c))
 291                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 292
 293         if (u64s <= ck->u64s)
 294                 return BTREE_INSERT_OK;
 295
 296         new_u64s        = roundup_pow_of_two(u64s);
 297         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 298         if (!new_k)
 299                 return -ENOMEM;
 300
 301         ck->u64s        = new_u64s;
 302         ck->k           = new_k;
 303         return BTREE_INSERT_OK;
 304 }
 305
 306 static inline void do_btree_insert_one(struct btree_trans *trans,
 307                                        struct btree_iter *iter,
 308                                        struct bkey_i *insert)
 309 {
 310         struct bch_fs *c = trans->c;
 311         struct journal *j = &c->journal;
 312         bool did_work;
 313
 314         EBUG_ON(trans->journal_res.ref !=
 315                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 316
 317         insert->k.needs_whiteout = false;
 318
 319         did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
 320                 ? btree_insert_key_leaf(trans, iter, insert)
 321                 : bch2_btree_insert_key_cached(trans, iter, insert);
 322         if (!did_work)
 323                 return;
 324
 325         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 326                 bch2_journal_add_keys(j, &trans->journal_res,
 327                                       iter->btree_id, insert);
 328
 329                 bch2_journal_set_has_inode(j, &trans->journal_res,
 330                                            insert->k.p.inode);
 331
 332                 if (trans->journal_seq)
 333                         *trans->journal_seq = trans->journal_res.seq;
 334         }
 335 }
 336
 337 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 338 {
 339         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 340 }
 341
 342 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 343 {
 344         return (((BTREE_NODE_TYPE_HAS_TRIGGERS &
 345                   ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS)) |
 346                 (1U << BTREE_ID_EC)) &
 347                 (1U << iter->btree_id);
 348 }
 349
 350 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 351 {
 352         __bch2_btree_iter_unlock(iter);
 353 }
 354
 355 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 356 {
 357         struct bch_fs *c = trans->c;
 358         struct btree_insert_entry *i;
 359
 360         trans_for_each_update(trans, i) {
 361                 /*
 362                  * XXX: synchronization of cached update triggers with gc
 363                  */
 364                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 365
 366                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 367                         bch2_mark_update(trans, i->iter, i->k, NULL,
 368                                          i->trigger_flags|BTREE_TRIGGER_GC);
 369         }
 370 }
 371
 372 static inline int
 373 bch2_trans_commit_write_locked(struct btree_trans *trans,
 374                                struct btree_insert_entry **stopped_at)
 375 {
 376         struct bch_fs *c = trans->c;
 377         struct bch_fs_usage *fs_usage = NULL;
 378         struct btree_insert_entry *i;
 379         unsigned u64s = 0;
 380         bool marking = false;
 381         int ret;
 382
 383         if (race_fault()) {
 384                 trace_trans_restart_fault_inject(trans->ip);
 385                 return -EINTR;
 386         }
 387
 388         /*
 389          * Check if the insert will fit in the leaf node with the write lock
 390          * held, otherwise another thread could write the node changing the
 391          * amount of space available:
 392          */
 393
 394         prefetch(&trans->c->journal.flags);
 395
 396         trans_for_each_update2(trans, i) {
 397                 /* Multiple inserts might go to same leaf: */
 398                 if (!same_leaf_as_prev(trans, i))
 399                         u64s = 0;
 400
 401                 u64s += i->k->k.u64s;
 402                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 403                         ? btree_key_can_insert(trans, i->iter, u64s)
 404                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 405                 if (ret) {
 406                         *stopped_at = i;
 407                         return ret;
 408                 }
 409
 410                 if (btree_node_type_needs_gc(i->iter->btree_id))
 411                         marking = true;
 412         }
 413
 414         if (marking) {
 415                 percpu_down_read(&c->mark_lock);
 416                 fs_usage = bch2_fs_usage_scratch_get(c);
 417         }
 418
 419         /*
 420          * Don't get journal reservation until after we know insert will
 421          * succeed:
 422          */
 423         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 424                 ret = bch2_trans_journal_res_get(trans,
 425                                 JOURNAL_RES_GET_NONBLOCK);
 426                 if (ret)
 427                         goto err;
 428         } else {
 429                 trans->journal_res.seq = c->journal.replay_journal_seq;
 430         }
 431
 432         if (unlikely(trans->extra_journal_entry_u64s)) {
 433                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 434                                   trans->extra_journal_entries,
 435                                   trans->extra_journal_entry_u64s);
 436
 437                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 438                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 439         }
 440
 441         /*
 442          * Not allowed to fail after we've gotten our journal reservation - we
 443          * have to use it:
 444          */
 445
 446         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 447                 if (bch2_journal_seq_verify)
 448                         trans_for_each_update2(trans, i)
 449                                 i->k->k.version.lo = trans->journal_res.seq;
 450                 else if (bch2_inject_invalid_keys)
 451                         trans_for_each_update2(trans, i)
 452                                 i->k->k.version = MAX_VERSION;
 453         }
 454
 455         /* Must be called under mark_lock: */
 456         if (marking && trans->fs_usage_deltas &&
 457             bch2_replicas_delta_list_apply(c, fs_usage,
 458                                            trans->fs_usage_deltas)) {
 459                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 460                 goto err;
 461         }
 462
 463         trans_for_each_update(trans, i)
 464                 if (iter_has_nontrans_triggers(i->iter))
 465                         bch2_mark_update(trans, i->iter, i->k,
 466                                          fs_usage, i->trigger_flags);
 467
 468         if (marking)
 469                 bch2_trans_fs_usage_apply(trans, fs_usage);
 470
 471         if (unlikely(c->gc_pos.phase))
 472                 bch2_trans_mark_gc(trans);
 473
 474         trans_for_each_update2(trans, i)
 475                 do_btree_insert_one(trans, i->iter, i->k);
 476 err:
 477         if (marking) {
 478                 bch2_fs_usage_scratch_put(c, fs_usage);
 479                 percpu_up_read(&c->mark_lock);
 480         }
 481
 482         return ret;
 483 }
 484
 485 /*
 486  * Get journal reservation, take write locks, and attempt to do btree update(s):
 487  */
 488 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 489                                        struct btree_insert_entry **stopped_at)
 490 {
 491         struct btree_insert_entry *i;
 492         struct btree_iter *iter;
 493         int ret;
 494
 495         trans_for_each_update2(trans, i)
 496                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 497
 498         ret = bch2_journal_preres_get(&trans->c->journal,
 499                         &trans->journal_preres, trans->journal_preres_u64s,
 500                         JOURNAL_RES_GET_NONBLOCK|
 501                         ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
 502                          ? JOURNAL_RES_GET_RECLAIM : 0));
 503         if (unlikely(ret == -EAGAIN))
 504                 ret = bch2_trans_journal_preres_get_cold(trans,
 505                                                 trans->journal_preres_u64s);
 506         if (unlikely(ret))
 507                 return ret;
 508
 509         /*
 510          * Can't be holding any read locks when we go to take write locks:
 511          *
 512          * note - this must be done after bch2_trans_journal_preres_get_cold()
 513          * or anything else that might call bch2_trans_relock(), since that
 514          * would just retake the read locks:
 515          */
 516         trans_for_each_iter(trans, iter) {
 517                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 518                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 519                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 520                         bch2_btree_iter_unlock_noinline(iter);
 521                 }
 522         }
 523
 524         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 525                 trans_for_each_update2(trans, i)
 526                         btree_insert_entry_checks(trans, i->iter, i->k);
 527         bch2_btree_trans_verify_locks(trans);
 528
 529         trans_for_each_update2(trans, i)
 530                 if (!same_leaf_as_prev(trans, i))
 531                         bch2_btree_node_lock_for_insert(trans->c,
 532                                         iter_l(i->iter)->b, i->iter);
 533
 534         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 535
 536         trans_for_each_update2(trans, i)
 537                 if (!same_leaf_as_prev(trans, i))
 538                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 539                                                              i->iter);
 540
 541         if (!ret && trans->journal_pin)
 542                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 543                                      trans->journal_pin, NULL);
 544
 545         /*
 546          * Drop journal reservation after dropping write locks, since dropping
 547          * the journal reservation may kick off a journal write:
 548          */
 549         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 550
 551         if (unlikely(ret))
 552                 return ret;
 553
 554         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 555                 trans->nounlock = true;
 556
 557         trans_for_each_update2(trans, i)
 558                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 559                     !same_leaf_as_prev(trans, i))
 560                         bch2_foreground_maybe_merge(trans->c, i->iter,
 561                                                     0, trans->flags);
 562
 563         trans->nounlock = false;
 564
 565         bch2_trans_downgrade(trans);
 566
 567         return 0;
 568 }
 569
 570 static noinline
 571 int bch2_trans_commit_error(struct btree_trans *trans,
 572                             struct btree_insert_entry *i,
 573                             int ret)
 574 {
 575         struct bch_fs *c = trans->c;
 576         unsigned flags = trans->flags;
 577
 578         /*
 579          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 580          * update; if we haven't done anything yet it doesn't apply
 581          */
 582         flags &= ~BTREE_INSERT_NOUNLOCK;
 583
 584         switch (ret) {
 585         case BTREE_INSERT_BTREE_NODE_FULL:
 586                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 587
 588                 /*
 589                  * if the split succeeded without dropping locks the insert will
 590                  * still be atomic (what the caller peeked() and is overwriting
 591                  * won't have changed)
 592                  */
 593 #if 0
 594                 /*
 595                  * XXX:
 596                  * split -> btree node merging (of parent node) might still drop
 597                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 598                  *
 599                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 600                  * will inhibit merging - but we don't have a reliable way yet
 601                  * (do we?) of checking if we dropped locks in this path
 602                  */
 603                 if (!ret)
 604                         goto retry;
 605 #endif
 606
 607                 /*
 608                  * don't care if we got ENOSPC because we told split it
 609                  * couldn't block:
 610                  */
 611                 if (!ret ||
 612                     ret == -EINTR ||
 613                     (flags & BTREE_INSERT_NOUNLOCK)) {
 614                         trace_trans_restart_btree_node_split(trans->ip);
 615                         ret = -EINTR;
 616                 }
 617                 break;
 618         case BTREE_INSERT_ENOSPC:
 619                 ret = -ENOSPC;
 620                 break;
 621         case BTREE_INSERT_NEED_MARK_REPLICAS:
 622                 bch2_trans_unlock(trans);
 623
 624                 trans_for_each_update(trans, i) {
 625                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 626                         if (ret)
 627                                 return ret;
 628                 }
 629
 630                 if (bch2_trans_relock(trans))
 631                         return 0;
 632
 633                 trace_trans_restart_mark_replicas(trans->ip);
 634                 ret = -EINTR;
 635                 break;
 636         case BTREE_INSERT_NEED_JOURNAL_RES:
 637                 bch2_trans_unlock(trans);
 638
 639                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 640                 if (ret)
 641                         return ret;
 642
 643                 if (bch2_trans_relock(trans))
 644                         return 0;
 645
 646                 trace_trans_restart_journal_res_get(trans->ip);
 647                 ret = -EINTR;
 648                 break;
 649         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 650                 bch2_trans_unlock(trans);
 651
 652                 while (bch2_btree_key_cache_must_wait(c)) {
 653                         mutex_lock(&c->journal.reclaim_lock);
 654                         bch2_journal_reclaim(&c->journal);
 655                         mutex_unlock(&c->journal.reclaim_lock);
 656                 }
 657
 658                 if (bch2_trans_relock(trans))
 659                         return 0;
 660
 661                 trace_trans_restart_journal_reclaim(trans->ip);
 662                 ret = -EINTR;
 663                 break;
 664         default:
 665                 BUG_ON(ret >= 0);
 666                 break;
 667         }
 668
 669         return ret;
 670 }
 671
 672 static noinline int
 673 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 674 {
 675         struct bch_fs *c = trans->c;
 676         int ret;
 677
 678         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 679                 return -EROFS;
 680
 681         bch2_trans_unlock(trans);
 682
 683         ret = bch2_fs_read_write_early(c);
 684         if (ret)
 685                 return ret;
 686
 687         percpu_ref_get(&c->writes);
 688         return 0;
 689 }
 690
 691 static inline int btree_iter_pos_cmp(const struct btree_iter *l,
 692                                      const struct btree_iter *r)
 693 {
 694         return   cmp_int(l->btree_id, r->btree_id) ?:
 695                  bkey_cmp(l->pos, r->pos);
 696 }
 697
 698 static void bch2_trans_update2(struct btree_trans *trans,
 699                                struct btree_iter *iter,
 700                                struct bkey_i *insert)
 701 {
 702         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 703                 .iter = iter, .k = insert
 704         };
 705
 706         btree_insert_entry_checks(trans, n.iter, n.k);
 707
 708         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 709
 710         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 711
 712         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 713
 714         trans_for_each_update2(trans, i) {
 715                 if (btree_iter_pos_cmp(n.iter, i->iter) == 0) {
 716                         *i = n;
 717                         return;
 718                 }
 719
 720                 if (btree_iter_pos_cmp(n.iter, i->iter) <= 0)
 721                         break;
 722         }
 723
 724         array_insert_item(trans->updates2, trans->nr_updates2,
 725                           i - trans->updates2, n);
 726 }
 727
 728 static int extent_update_to_keys(struct btree_trans *trans,
 729                                  struct btree_iter *orig_iter,
 730                                  struct bkey_i *insert)
 731 {
 732         struct btree_iter *iter;
 733         int ret;
 734
 735         ret = bch2_extent_can_insert(trans, orig_iter, insert);
 736         if (ret)
 737                 return ret;
 738
 739         if (bkey_deleted(&insert->k))
 740                 return 0;
 741
 742         iter = bch2_trans_copy_iter(trans, orig_iter);
 743
 744         iter->flags |= BTREE_ITER_INTENT;
 745         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 746         bch2_trans_update2(trans, iter, insert);
 747         bch2_trans_iter_put(trans, iter);
 748         return 0;
 749 }
 750
 751 static int extent_handle_overwrites(struct btree_trans *trans,
 752                                     enum btree_id btree_id,
 753                                     struct bpos start, struct bpos end)
 754 {
 755         struct btree_iter *iter = NULL, *update_iter;
 756         struct bkey_i *update;
 757         struct bkey_s_c k;
 758         int ret = 0;
 759
 760         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 761         k = bch2_btree_iter_peek_with_updates(iter);
 762
 763         while (k.k && !(ret = bkey_err(k))) {
 764                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 765                         break;
 766
 767                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 768                         update_iter = bch2_trans_copy_iter(trans, iter);
 769
 770                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 771                         if ((ret = PTR_ERR_OR_ZERO(update)))
 772                                 goto err;
 773
 774                         bkey_reassemble(update, k);
 775                         bch2_cut_back(start, update);
 776
 777                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 778                         bch2_trans_update2(trans, update_iter, update);
 779                         bch2_trans_iter_put(trans, update_iter);
 780                 }
 781
 782                 if (bkey_cmp(k.k->p, end) > 0) {
 783                         update_iter = bch2_trans_copy_iter(trans, iter);
 784
 785                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 786                         if ((ret = PTR_ERR_OR_ZERO(update)))
 787                                 goto err;
 788
 789                         bkey_reassemble(update, k);
 790                         bch2_cut_front(end, update);
 791
 792                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 793                         bch2_trans_update2(trans, update_iter, update);
 794                         bch2_trans_iter_put(trans, update_iter);
 795                 } else {
 796                         update_iter = bch2_trans_copy_iter(trans, iter);
 797
 798                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 799                         if ((ret = PTR_ERR_OR_ZERO(update)))
 800                                 goto err;
 801
 802                         update->k = *k.k;
 803                         set_bkey_val_u64s(&update->k, 0);
 804                         update->k.type = KEY_TYPE_deleted;
 805                         update->k.size = 0;
 806
 807                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 808                         bch2_trans_update2(trans, update_iter, update);
 809                         bch2_trans_iter_put(trans, update_iter);
 810                 }
 811
 812                 k = bch2_btree_iter_next_with_updates(iter);
 813         }
 814 err:
 815         if (!IS_ERR_OR_NULL(iter))
 816                 bch2_trans_iter_put(trans, iter);
 817         return ret;
 818 }
 819
 820 int __bch2_trans_commit(struct btree_trans *trans)
 821 {
 822         struct btree_insert_entry *i = NULL;
 823         struct btree_iter *iter;
 824         bool trans_trigger_run;
 825         unsigned u64s;
 826         int ret = 0;
 827
 828         if (!trans->nr_updates)
 829                 goto out_noupdates;
 830
 831         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 832                 lockdep_assert_held(&trans->c->gc_lock);
 833
 834         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 835
 836         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 837         trans->journal_preres_u64s      = 0;
 838
 839         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 840             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 841                 ret = bch2_trans_commit_get_rw_cold(trans);
 842                 if (ret)
 843                         return ret;
 844         }
 845
 846 #ifdef CONFIG_BCACHEFS_DEBUG
 847         trans_for_each_update(trans, i)
 848                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 849                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 850                         bch2_btree_key_cache_verify_clean(trans,
 851                                         i->iter->btree_id, i->iter->pos);
 852 #endif
 853
 854         /*
 855          * Running triggers will append more updates to the list of updates as
 856          * we're walking it:
 857          */
 858         do {
 859                 trans_trigger_run = false;
 860
 861                 trans_for_each_update(trans, i) {
 862                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
 863                                      (ret = bch2_btree_iter_traverse(i->iter)))) {
 864                                 trace_trans_restart_traverse(trans->ip);
 865                                 goto out;
 866                         }
 867
 868                         /*
 869                          * We're not using bch2_btree_iter_upgrade here because
 870                          * we know trans->nounlock can't be set:
 871                          */
 872                         if (unlikely(i->iter->locks_want < 1 &&
 873                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 874                                 trace_trans_restart_upgrade(trans->ip);
 875                                 ret = -EINTR;
 876                                 goto out;
 877                         }
 878
 879                         if (iter_has_trans_triggers(i->iter) &&
 880                             !i->trans_triggers_run) {
 881                                 i->trans_triggers_run = true;
 882                                 trans_trigger_run = true;
 883
 884                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 885                                                              i->trigger_flags);
 886                                 if (unlikely(ret)) {
 887                                         if (ret == -EINTR)
 888                                                 trace_trans_restart_mark(trans->ip);
 889                                         goto out;
 890                                 }
 891                         }
 892                 }
 893         } while (trans_trigger_run);
 894
 895         /* Turn extents updates into keys: */
 896         trans_for_each_update(trans, i)
 897                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 898                         struct bpos start = bkey_start_pos(&i->k->k);
 899
 900                         while (i + 1 < trans->updates + trans->nr_updates &&
 901                                i[0].iter->btree_id == i[1].iter->btree_id &&
 902                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 903                                 i++;
 904
 905                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 906                                                        start, i->k->k.p);
 907                         if (ret)
 908                                 goto out;
 909                 }
 910
 911         trans_for_each_update(trans, i) {
 912                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 913                         ret = extent_update_to_keys(trans, i->iter, i->k);
 914                         if (ret)
 915                                 goto out;
 916                 } else {
 917                         bch2_trans_update2(trans, i->iter, i->k);
 918                 }
 919         }
 920
 921         trans_for_each_update2(trans, i) {
 922                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 923                 BUG_ON(i->iter->locks_want < 1);
 924
 925                 u64s = jset_u64s(i->k->k.u64s);
 926                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 927                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 928                         trans->journal_preres_u64s += u64s;
 929                 trans->journal_u64s += u64s;
 930         }
 931 retry:
 932         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 933
 934         ret = do_bch2_trans_commit(trans, &i);
 935
 936         /* make sure we didn't drop or screw up locks: */
 937         bch2_btree_trans_verify_locks(trans);
 938
 939         if (ret)
 940                 goto err;
 941
 942         trans_for_each_iter(trans, iter)
 943                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 944                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 945                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 946                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 947                         else
 948                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 949                 }
 950 out:
 951         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 952
 953         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 954                 percpu_ref_put(&trans->c->writes);
 955 out_noupdates:
 956         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 957
 958         return ret;
 959 err:
 960         ret = bch2_trans_commit_error(trans, i, ret);
 961         if (ret)
 962                 goto out;
 963
 964         goto retry;
 965 }
 966
 967 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 968                       struct bkey_i *k, enum btree_trigger_flags flags)
 969 {
 970         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 971                 .trigger_flags = flags, .iter = iter, .k = k
 972         };
 973
 974         EBUG_ON(bkey_cmp(iter->pos,
 975                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 976                          ? bkey_start_pos(&k->k)
 977                          : k->k.p));
 978
 979         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 980
 981         if (btree_node_type_is_extents(iter->btree_id)) {
 982                 iter->pos_after_commit = k->k.p;
 983                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 984         }
 985
 986         /*
 987          * Pending updates are kept sorted: first, find position of new update:
 988          */
 989         trans_for_each_update(trans, i)
 990                 if (btree_iter_pos_cmp(iter, i->iter) <= 0)
 991                         break;
 992
 993         /*
 994          * Now delete/trim any updates the new update overwrites:
 995          */
 996         if (i > trans->updates &&
 997             i[-1].iter->btree_id == iter->btree_id &&
 998             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 999                 bch2_cut_back(n.iter->pos, i[-1].k);
1000
1001         while (i < trans->updates + trans->nr_updates &&
1002                iter->btree_id == i->iter->btree_id &&
1003                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1004                 array_remove_item(trans->updates, trans->nr_updates,
1005                                   i - trans->updates);
1006
1007         if (i < trans->updates + trans->nr_updates &&
1008             iter->btree_id == i->iter->btree_id &&
1009             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1010                 /*
1011                  * When we have an extent that overwrites the start of another
1012                  * update, trimming that extent will mean the iterator's
1013                  * position has to change since the iterator position has to
1014                  * match the extent's start pos - but we don't want to change
1015                  * the iterator pos if some other code is using it, so we may
1016                  * need to clone it:
1017                  */
1018                 if (trans->iters_live & (1ULL << i->iter->idx)) {
1019                         i->iter = bch2_trans_copy_iter(trans, i->iter);
1020
1021                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1022                         bch2_trans_iter_put(trans, i->iter);
1023                 }
1024
1025                 bch2_cut_front(n.k->k.p, i->k);
1026                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1027         }
1028
1029         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1030
1031         array_insert_item(trans->updates, trans->nr_updates,
1032                           i - trans->updates, n);
1033         return 0;
1034 }
1035
1036 int __bch2_btree_insert(struct btree_trans *trans,
1037                         enum btree_id id, struct bkey_i *k)
1038 {
1039         struct btree_iter *iter;
1040         int ret;
1041
1042         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1043                                    BTREE_ITER_INTENT);
1044
1045         ret   = bch2_btree_iter_traverse(iter) ?:
1046                 bch2_trans_update(trans, iter, k, 0);
1047         bch2_trans_iter_put(trans, iter);
1048         return ret;
1049 }
1050
1051 /**
1052  * bch2_btree_insert - insert keys into the extent btree
1053  * @c:                  pointer to struct bch_fs
1054  * @id:                 btree to insert into
1055  * @insert_keys:        list of keys to insert
1056  * @hook:               insert callback
1057  */
1058 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1059                       struct bkey_i *k,
1060                       struct disk_reservation *disk_res,
1061                       u64 *journal_seq, int flags)
1062 {
1063         return bch2_trans_do(c, disk_res, journal_seq, flags,
1064                              __bch2_btree_insert(&trans, id, k));
1065 }
1066
1067 int bch2_btree_delete_at(struct btree_trans *trans,
1068                          struct btree_iter *iter, unsigned flags)
1069 {
1070         struct bkey_i k;
1071
1072         bkey_init(&k.k);
1073         k.k.p = iter->pos;
1074
1075         bch2_trans_update(trans, iter, &k, 0);
1076         return bch2_trans_commit(trans, NULL, NULL,
1077                                  BTREE_INSERT_NOFAIL|
1078                                  BTREE_INSERT_USE_RESERVE|flags);
1079 }
1080
1081 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1082                                   struct bpos start, struct bpos end,
1083                                   u64 *journal_seq)
1084 {
1085         struct btree_iter *iter;
1086         struct bkey_s_c k;
1087         int ret = 0;
1088
1089         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1090 retry:
1091         while ((k = bch2_btree_iter_peek(iter)).k &&
1092                !(ret = bkey_err(k)) &&
1093                bkey_cmp(iter->pos, end) < 0) {
1094                 struct bkey_i delete;
1095
1096                 bch2_trans_begin(trans);
1097
1098                 bkey_init(&delete.k);
1099
1100                 /*
1101                  * This could probably be more efficient for extents:
1102                  */
1103
1104                 /*
1105                  * For extents, iter.pos won't necessarily be the same as
1106                  * bkey_start_pos(k.k) (for non extents they always will be the
1107                  * same). It's important that we delete starting from iter.pos
1108                  * because the range we want to delete could start in the middle
1109                  * of k.
1110                  *
1111                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1112                  * bkey_start_pos(k.k)).
1113                  */
1114                 delete.k.p = iter->pos;
1115
1116                 if (btree_node_type_is_extents(iter->btree_id)) {
1117                         unsigned max_sectors =
1118                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1119
1120                         /* create the biggest key we can */
1121                         bch2_key_resize(&delete.k, max_sectors);
1122                         bch2_cut_back(end, &delete);
1123
1124                         ret = bch2_extent_trim_atomic(&delete, iter);
1125                         if (ret)
1126                                 break;
1127                 }
1128
1129                 bch2_trans_update(trans, iter, &delete, 0);
1130                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1131                                         BTREE_INSERT_NOFAIL);
1132                 if (ret)
1133                         break;
1134
1135                 bch2_trans_cond_resched(trans);
1136         }
1137
1138         if (ret == -EINTR) {
1139                 ret = 0;
1140                 goto retry;
1141         }
1142
1143         bch2_trans_iter_put(trans, iter);
1144         return ret;
1145 }
1146
1147 /*
1148  * bch_btree_delete_range - delete everything within a given range
1149  *
1150  * Range is a half open interval - [start, end)
1151  */
1152 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1153                             struct bpos start, struct bpos end,
1154                             u64 *journal_seq)
1155 {
1156         return bch2_trans_do(c, NULL, journal_seq, 0,
1157                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1158 }