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          * another thread could be holding an intent lock on the same node we
 512          * have a read lock on, and it'll block trying to take a write lock
 513          * (because we hold a read lock) and it could be blocking us by holding
 514          * its own read lock (while we're trying to to take write locks).
 515          *
 516          * note - this must be done after bch2_trans_journal_preres_get_cold()
 517          * or anything else that might call bch2_trans_relock(), since that
 518          * would just retake the read locks:
 519          */
 520         trans_for_each_iter(trans, iter) {
 521                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 522                         if ((iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT) ||
 523                             (trans->iters_live & (1ULL << iter->idx))) {
 524                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 525                                         trace_trans_restart_upgrade(trans->ip);
 526                                         return -EINTR;
 527                                 }
 528                         } else {
 529                                 bch2_btree_iter_unlock_noinline(iter);
 530                         }
 531                 }
 532         }
 533
 534         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 535                 trans_for_each_update2(trans, i)
 536                         btree_insert_entry_checks(trans, i->iter, i->k);
 537         bch2_btree_trans_verify_locks(trans);
 538
 539         trans_for_each_update2(trans, i)
 540                 if (!same_leaf_as_prev(trans, i))
 541                         bch2_btree_node_lock_for_insert(trans->c,
 542                                         iter_l(i->iter)->b, i->iter);
 543
 544         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 545
 546         trans_for_each_update2(trans, i)
 547                 if (!same_leaf_as_prev(trans, i))
 548                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 549                                                              i->iter);
 550
 551         if (!ret && trans->journal_pin)
 552                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 553                                      trans->journal_pin, NULL);
 554
 555         /*
 556          * Drop journal reservation after dropping write locks, since dropping
 557          * the journal reservation may kick off a journal write:
 558          */
 559         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 560
 561         if (unlikely(ret))
 562                 return ret;
 563
 564         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 565                 trans->nounlock = true;
 566
 567         trans_for_each_update2(trans, i)
 568                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 569                     !same_leaf_as_prev(trans, i))
 570                         bch2_foreground_maybe_merge(trans->c, i->iter,
 571                                                     0, trans->flags);
 572
 573         trans->nounlock = false;
 574
 575         bch2_trans_downgrade(trans);
 576
 577         return 0;
 578 }
 579
 580 static noinline
 581 int bch2_trans_commit_error(struct btree_trans *trans,
 582                             struct btree_insert_entry *i,
 583                             int ret)
 584 {
 585         struct bch_fs *c = trans->c;
 586         unsigned flags = trans->flags;
 587
 588         /*
 589          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 590          * update; if we haven't done anything yet it doesn't apply
 591          */
 592         flags &= ~BTREE_INSERT_NOUNLOCK;
 593
 594         switch (ret) {
 595         case BTREE_INSERT_BTREE_NODE_FULL:
 596                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 597
 598                 /*
 599                  * if the split succeeded without dropping locks the insert will
 600                  * still be atomic (what the caller peeked() and is overwriting
 601                  * won't have changed)
 602                  */
 603 #if 0
 604                 /*
 605                  * XXX:
 606                  * split -> btree node merging (of parent node) might still drop
 607                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 608                  *
 609                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 610                  * will inhibit merging - but we don't have a reliable way yet
 611                  * (do we?) of checking if we dropped locks in this path
 612                  */
 613                 if (!ret)
 614                         goto retry;
 615 #endif
 616
 617                 /*
 618                  * don't care if we got ENOSPC because we told split it
 619                  * couldn't block:
 620                  */
 621                 if (!ret ||
 622                     ret == -EINTR ||
 623                     (flags & BTREE_INSERT_NOUNLOCK)) {
 624                         trace_trans_restart_btree_node_split(trans->ip);
 625                         ret = -EINTR;
 626                 }
 627                 break;
 628         case BTREE_INSERT_ENOSPC:
 629                 ret = -ENOSPC;
 630                 break;
 631         case BTREE_INSERT_NEED_MARK_REPLICAS:
 632                 bch2_trans_unlock(trans);
 633
 634                 trans_for_each_update(trans, i) {
 635                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 636                         if (ret)
 637                                 return ret;
 638                 }
 639
 640                 if (bch2_trans_relock(trans))
 641                         return 0;
 642
 643                 trace_trans_restart_mark_replicas(trans->ip);
 644                 ret = -EINTR;
 645                 break;
 646         case BTREE_INSERT_NEED_JOURNAL_RES:
 647                 bch2_trans_unlock(trans);
 648
 649                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 650                 if (ret)
 651                         return ret;
 652
 653                 if (bch2_trans_relock(trans))
 654                         return 0;
 655
 656                 trace_trans_restart_journal_res_get(trans->ip);
 657                 ret = -EINTR;
 658                 break;
 659         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 660                 bch2_trans_unlock(trans);
 661
 662                 do {
 663                         mutex_lock(&c->journal.reclaim_lock);
 664                         ret = bch2_journal_reclaim(&c->journal);
 665                         mutex_unlock(&c->journal.reclaim_lock);
 666                 } while (!ret && bch2_btree_key_cache_must_wait(c));
 667
 668                 if (!ret && bch2_trans_relock(trans))
 669                         return 0;
 670
 671                 trace_trans_restart_journal_reclaim(trans->ip);
 672                 ret = -EINTR;
 673                 break;
 674         default:
 675                 BUG_ON(ret >= 0);
 676                 break;
 677         }
 678
 679         return ret;
 680 }
 681
 682 static noinline int
 683 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 684 {
 685         struct bch_fs *c = trans->c;
 686         int ret;
 687
 688         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 689                 return -EROFS;
 690
 691         bch2_trans_unlock(trans);
 692
 693         ret = bch2_fs_read_write_early(c);
 694         if (ret)
 695                 return ret;
 696
 697         percpu_ref_get(&c->writes);
 698         return 0;
 699 }
 700
 701 static inline int btree_iter_pos_cmp(const struct btree_iter *l,
 702                                      const struct btree_iter *r)
 703 {
 704         return   cmp_int(l->btree_id, r->btree_id) ?:
 705                  bkey_cmp(l->pos, r->pos);
 706 }
 707
 708 static void bch2_trans_update2(struct btree_trans *trans,
 709                                struct btree_iter *iter,
 710                                struct bkey_i *insert)
 711 {
 712         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 713                 .iter = iter, .k = insert
 714         };
 715
 716         btree_insert_entry_checks(trans, n.iter, n.k);
 717
 718         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 719
 720         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 721
 722         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 723
 724         trans_for_each_update2(trans, i) {
 725                 if (btree_iter_pos_cmp(n.iter, i->iter) == 0) {
 726                         *i = n;
 727                         return;
 728                 }
 729
 730                 if (btree_iter_pos_cmp(n.iter, i->iter) <= 0)
 731                         break;
 732         }
 733
 734         array_insert_item(trans->updates2, trans->nr_updates2,
 735                           i - trans->updates2, n);
 736 }
 737
 738 static int extent_update_to_keys(struct btree_trans *trans,
 739                                  struct btree_iter *orig_iter,
 740                                  struct bkey_i *insert)
 741 {
 742         struct btree_iter *iter;
 743         int ret;
 744
 745         ret = bch2_extent_can_insert(trans, orig_iter, insert);
 746         if (ret)
 747                 return ret;
 748
 749         if (bkey_deleted(&insert->k))
 750                 return 0;
 751
 752         iter = bch2_trans_copy_iter(trans, orig_iter);
 753
 754         iter->flags |= BTREE_ITER_INTENT;
 755         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 756         bch2_trans_update2(trans, iter, insert);
 757         bch2_trans_iter_put(trans, iter);
 758         return 0;
 759 }
 760
 761 static int extent_handle_overwrites(struct btree_trans *trans,
 762                                     enum btree_id btree_id,
 763                                     struct bpos start, struct bpos end)
 764 {
 765         struct btree_iter *iter = NULL, *update_iter;
 766         struct bkey_i *update;
 767         struct bkey_s_c k;
 768         int ret = 0;
 769
 770         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 771         k = bch2_btree_iter_peek_with_updates(iter);
 772
 773         while (k.k && !(ret = bkey_err(k))) {
 774                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 775                         break;
 776
 777                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 778                         update_iter = bch2_trans_copy_iter(trans, iter);
 779
 780                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 781                         if ((ret = PTR_ERR_OR_ZERO(update)))
 782                                 goto err;
 783
 784                         bkey_reassemble(update, k);
 785                         bch2_cut_back(start, update);
 786
 787                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 788                         bch2_trans_update2(trans, update_iter, update);
 789                         bch2_trans_iter_put(trans, update_iter);
 790                 }
 791
 792                 if (bkey_cmp(k.k->p, end) > 0) {
 793                         update_iter = bch2_trans_copy_iter(trans, iter);
 794
 795                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 796                         if ((ret = PTR_ERR_OR_ZERO(update)))
 797                                 goto err;
 798
 799                         bkey_reassemble(update, k);
 800                         bch2_cut_front(end, update);
 801
 802                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 803                         bch2_trans_update2(trans, update_iter, update);
 804                         bch2_trans_iter_put(trans, update_iter);
 805                 } else {
 806                         update_iter = bch2_trans_copy_iter(trans, iter);
 807
 808                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 809                         if ((ret = PTR_ERR_OR_ZERO(update)))
 810                                 goto err;
 811
 812                         update->k = *k.k;
 813                         set_bkey_val_u64s(&update->k, 0);
 814                         update->k.type = KEY_TYPE_deleted;
 815                         update->k.size = 0;
 816
 817                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 818                         bch2_trans_update2(trans, update_iter, update);
 819                         bch2_trans_iter_put(trans, update_iter);
 820                 }
 821
 822                 k = bch2_btree_iter_next_with_updates(iter);
 823         }
 824 err:
 825         if (!IS_ERR_OR_NULL(iter))
 826                 bch2_trans_iter_put(trans, iter);
 827         return ret;
 828 }
 829
 830 int __bch2_trans_commit(struct btree_trans *trans)
 831 {
 832         struct btree_insert_entry *i = NULL;
 833         struct btree_iter *iter;
 834         bool trans_trigger_run;
 835         unsigned u64s;
 836         int ret = 0;
 837
 838         if (!trans->nr_updates)
 839                 goto out_reset;
 840
 841         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 842                 lockdep_assert_held(&trans->c->gc_lock);
 843
 844         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 845
 846         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 847         trans->journal_preres_u64s      = 0;
 848
 849         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 850             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 851                 ret = bch2_trans_commit_get_rw_cold(trans);
 852                 if (ret)
 853                         goto out_reset;
 854         }
 855
 856 #ifdef CONFIG_BCACHEFS_DEBUG
 857         trans_for_each_update(trans, i)
 858                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 859                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 860                         bch2_btree_key_cache_verify_clean(trans,
 861                                         i->iter->btree_id, i->iter->pos);
 862 #endif
 863
 864         /*
 865          * Running triggers will append more updates to the list of updates as
 866          * we're walking it:
 867          */
 868         do {
 869                 trans_trigger_run = false;
 870
 871                 trans_for_each_update(trans, i) {
 872                         ret = bch2_btree_iter_traverse(i->iter);
 873                         if (unlikely(ret)) {
 874                                 trace_trans_restart_traverse(trans->ip);
 875                                 goto out;
 876                         }
 877
 878                         /*
 879                          * We're not using bch2_btree_iter_upgrade here because
 880                          * we know trans->nounlock can't be set:
 881                          */
 882                         if (unlikely(!btree_node_intent_locked(i->iter, i->iter->level) &&
 883                                      !__bch2_btree_iter_upgrade(i->iter, i->iter->level + 1))) {
 884                                 trace_trans_restart_upgrade(trans->ip);
 885                                 ret = -EINTR;
 886                                 goto out;
 887                         }
 888
 889                         if (iter_has_trans_triggers(i->iter) &&
 890                             !i->trans_triggers_run) {
 891                                 i->trans_triggers_run = true;
 892                                 trans_trigger_run = true;
 893
 894                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 895                                                              i->trigger_flags);
 896                                 if (unlikely(ret)) {
 897                                         if (ret == -EINTR)
 898                                                 trace_trans_restart_mark(trans->ip);
 899                                         goto out;
 900                                 }
 901                         }
 902                 }
 903         } while (trans_trigger_run);
 904
 905         /* Turn extents updates into keys: */
 906         trans_for_each_update(trans, i)
 907                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 908                         struct bpos start = bkey_start_pos(&i->k->k);
 909
 910                         while (i + 1 < trans->updates + trans->nr_updates &&
 911                                i[0].iter->btree_id == i[1].iter->btree_id &&
 912                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 913                                 i++;
 914
 915                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 916                                                        start, i->k->k.p);
 917                         if (ret)
 918                                 goto out;
 919                 }
 920
 921         trans_for_each_update(trans, i) {
 922                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 923                         ret = extent_update_to_keys(trans, i->iter, i->k);
 924                         if (ret)
 925                                 goto out;
 926                 } else {
 927                         bch2_trans_update2(trans, i->iter, i->k);
 928                 }
 929         }
 930
 931         trans_for_each_update2(trans, i) {
 932                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 933                 BUG_ON(i->iter->locks_want < 1);
 934
 935                 u64s = jset_u64s(i->k->k.u64s);
 936                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 937                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 938                         trans->journal_preres_u64s += u64s;
 939                 trans->journal_u64s += u64s;
 940         }
 941 retry:
 942         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 943
 944         ret = do_bch2_trans_commit(trans, &i);
 945
 946         /* make sure we didn't drop or screw up locks: */
 947         bch2_btree_trans_verify_locks(trans);
 948
 949         if (ret)
 950                 goto err;
 951
 952         trans_for_each_iter(trans, iter)
 953                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 954                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 955                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 956                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 957                         else
 958                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 959                 }
 960 out:
 961         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 962
 963         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 964                 percpu_ref_put(&trans->c->writes);
 965 out_reset:
 966         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 967
 968         return ret;
 969 err:
 970         ret = bch2_trans_commit_error(trans, i, ret);
 971         if (ret)
 972                 goto out;
 973
 974         goto retry;
 975 }
 976
 977 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 978                       struct bkey_i *k, enum btree_trigger_flags flags)
 979 {
 980         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 981                 .trigger_flags = flags, .iter = iter, .k = k
 982         };
 983
 984 #ifdef CONFIG_BCACHEFS_DEBUG
 985         BUG_ON(bkey_cmp(iter->pos,
 986                         (iter->flags & BTREE_ITER_IS_EXTENTS)
 987                         ? bkey_start_pos(&k->k)
 988                         : k->k.p));
 989
 990         trans_for_each_update(trans, i) {
 991                 BUG_ON(bkey_cmp(i->iter->pos,
 992                                  (i->iter->flags & BTREE_ITER_IS_EXTENTS)
 993                                  ? bkey_start_pos(&i->k->k)
 994                                  : i->k->k.p));
 995
 996                 BUG_ON(i != trans->updates &&
 997                        btree_iter_pos_cmp(i[-1].iter, i[0].iter) >= 0);
 998         }
 999 #endif
1000
1001         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1002
1003         if (btree_node_type_is_extents(iter->btree_id)) {
1004                 iter->pos_after_commit = k->k.p;
1005                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1006         }
1007
1008         /*
1009          * Pending updates are kept sorted: first, find position of new update:
1010          */
1011         trans_for_each_update(trans, i)
1012                 if (btree_iter_pos_cmp(iter, i->iter) <= 0)
1013                         break;
1014
1015         /*
1016          * Now delete/trim any updates the new update overwrites:
1017          */
1018         if (i > trans->updates &&
1019             i[-1].iter->btree_id == iter->btree_id &&
1020             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
1021                 bch2_cut_back(n.iter->pos, i[-1].k);
1022
1023         while (i < trans->updates + trans->nr_updates &&
1024                iter->btree_id == i->iter->btree_id &&
1025                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1026                 array_remove_item(trans->updates, trans->nr_updates,
1027                                   i - trans->updates);
1028
1029         if (i < trans->updates + trans->nr_updates &&
1030             iter->btree_id == i->iter->btree_id &&
1031             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1032                 /*
1033                  * When we have an extent that overwrites the start of another
1034                  * update, trimming that extent will mean the iterator's
1035                  * position has to change since the iterator position has to
1036                  * match the extent's start pos - but we don't want to change
1037                  * the iterator pos if some other code is using it, so we may
1038                  * need to clone it:
1039                  */
1040                 if (trans->iters_live & (1ULL << i->iter->idx)) {
1041                         i->iter = bch2_trans_copy_iter(trans, i->iter);
1042
1043                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1044                         bch2_trans_iter_put(trans, i->iter);
1045                 }
1046
1047                 bch2_cut_front(n.k->k.p, i->k);
1048                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1049         }
1050
1051         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1052
1053         array_insert_item(trans->updates, trans->nr_updates,
1054                           i - trans->updates, n);
1055         return 0;
1056 }
1057
1058 int __bch2_btree_insert(struct btree_trans *trans,
1059                         enum btree_id id, struct bkey_i *k)
1060 {
1061         struct btree_iter *iter;
1062         int ret;
1063
1064         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1065                                    BTREE_ITER_INTENT);
1066
1067         ret   = bch2_btree_iter_traverse(iter) ?:
1068                 bch2_trans_update(trans, iter, k, 0);
1069         bch2_trans_iter_put(trans, iter);
1070         return ret;
1071 }
1072
1073 /**
1074  * bch2_btree_insert - insert keys into the extent btree
1075  * @c:                  pointer to struct bch_fs
1076  * @id:                 btree to insert into
1077  * @insert_keys:        list of keys to insert
1078  * @hook:               insert callback
1079  */
1080 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1081                       struct bkey_i *k,
1082                       struct disk_reservation *disk_res,
1083                       u64 *journal_seq, int flags)
1084 {
1085         return bch2_trans_do(c, disk_res, journal_seq, flags,
1086                              __bch2_btree_insert(&trans, id, k));
1087 }
1088
1089 int bch2_btree_delete_at(struct btree_trans *trans,
1090                          struct btree_iter *iter, unsigned flags)
1091 {
1092         struct bkey_i k;
1093
1094         bkey_init(&k.k);
1095         k.k.p = iter->pos;
1096
1097         bch2_trans_update(trans, iter, &k, 0);
1098         return bch2_trans_commit(trans, NULL, NULL,
1099                                  BTREE_INSERT_NOFAIL|flags);
1100 }
1101
1102 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1103                                   struct bpos start, struct bpos end,
1104                                   u64 *journal_seq)
1105 {
1106         struct btree_iter *iter;
1107         struct bkey_s_c k;
1108         int ret = 0;
1109
1110         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1111 retry:
1112         while ((k = bch2_btree_iter_peek(iter)).k &&
1113                !(ret = bkey_err(k)) &&
1114                bkey_cmp(iter->pos, end) < 0) {
1115                 struct bkey_i delete;
1116
1117                 bch2_trans_begin(trans);
1118
1119                 bkey_init(&delete.k);
1120
1121                 /*
1122                  * This could probably be more efficient for extents:
1123                  */
1124
1125                 /*
1126                  * For extents, iter.pos won't necessarily be the same as
1127                  * bkey_start_pos(k.k) (for non extents they always will be the
1128                  * same). It's important that we delete starting from iter.pos
1129                  * because the range we want to delete could start in the middle
1130                  * of k.
1131                  *
1132                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1133                  * bkey_start_pos(k.k)).
1134                  */
1135                 delete.k.p = iter->pos;
1136
1137                 if (btree_node_type_is_extents(iter->btree_id)) {
1138                         unsigned max_sectors =
1139                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1140
1141                         /* create the biggest key we can */
1142                         bch2_key_resize(&delete.k, max_sectors);
1143                         bch2_cut_back(end, &delete);
1144
1145                         ret = bch2_extent_trim_atomic(&delete, iter);
1146                         if (ret)
1147                                 break;
1148                 }
1149
1150                 bch2_trans_update(trans, iter, &delete, 0);
1151                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1152                                         BTREE_INSERT_NOFAIL);
1153                 if (ret)
1154                         break;
1155
1156                 bch2_trans_cond_resched(trans);
1157         }
1158
1159         if (ret == -EINTR) {
1160                 ret = 0;
1161                 goto retry;
1162         }
1163
1164         bch2_trans_iter_put(trans, iter);
1165         return ret;
1166 }
1167
1168 /*
1169  * bch_btree_delete_range - delete everything within a given range
1170  *
1171  * Range is a half open interval - [start, end)
1172  */
1173 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1174                             struct bpos start, struct bpos end,
1175                             u64 *journal_seq)
1176 {
1177         return bch2_trans_do(c, NULL, journal_seq, 0,
1178                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1179 }