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_packed(b, k, &insert->k))
  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(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(debug_check_bkeys(c) &&
 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                      struct bkey_i *insert,
 268                      unsigned u64s)
 269 {
 270         struct bch_fs *c = trans->c;
 271         struct btree *b = iter_l(iter)->b;
 272
 273         if (unlikely(btree_node_fake(b)))
 274                 return BTREE_INSERT_BTREE_NODE_FULL;
 275
 276         /*
 277          * old bch2_extent_sort_fix_overlapping() algorithm won't work with new
 278          * style extent updates:
 279          */
 280         if (unlikely(btree_node_old_extent_overwrite(b)))
 281                 return BTREE_INSERT_BTREE_NODE_FULL;
 282
 283         if (unlikely(u64s > bch_btree_keys_u64s_remaining(c, b)))
 284                 return BTREE_INSERT_BTREE_NODE_FULL;
 285
 286         return BTREE_INSERT_OK;
 287 }
 288
 289 static enum btree_insert_ret
 290 btree_key_can_insert_cached(struct btree_trans *trans,
 291                             struct btree_iter *iter,
 292                             struct bkey_i *insert,
 293                             unsigned u64s)
 294 {
 295         struct bkey_cached *ck = (void *) iter->l[0].b;
 296         unsigned new_u64s;
 297         struct bkey_i *new_k;
 298
 299         BUG_ON(iter->level);
 300
 301         if (u64s <= ck->u64s)
 302                 return BTREE_INSERT_OK;
 303
 304         new_u64s        = roundup_pow_of_two(u64s);
 305         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 306         if (!new_k)
 307                 return -ENOMEM;
 308
 309         ck->u64s        = new_u64s;
 310         ck->k           = new_k;
 311         return BTREE_INSERT_OK;
 312 }
 313
 314 static inline void do_btree_insert_one(struct btree_trans *trans,
 315                                        struct btree_iter *iter,
 316                                        struct bkey_i *insert)
 317 {
 318         struct bch_fs *c = trans->c;
 319         struct journal *j = &c->journal;
 320         bool did_work;
 321
 322         EBUG_ON(trans->journal_res.ref !=
 323                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 324
 325         insert->k.needs_whiteout = false;
 326
 327         did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
 328                 ? btree_insert_key_leaf(trans, iter, insert)
 329                 : bch2_btree_insert_key_cached(trans, iter, insert);
 330         if (!did_work)
 331                 return;
 332
 333         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 334                 bch2_journal_add_keys(j, &trans->journal_res,
 335                                       iter->btree_id, insert);
 336
 337                 bch2_journal_set_has_inode(j, &trans->journal_res,
 338                                            insert->k.p.inode);
 339
 340                 if (trans->journal_seq)
 341                         *trans->journal_seq = trans->journal_res.seq;
 342         }
 343 }
 344
 345 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 346 {
 347         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 348 }
 349
 350 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 351 {
 352         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 353                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 354                 (1U << iter->btree_id);
 355 }
 356
 357 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 358 {
 359         __bch2_btree_iter_unlock(iter);
 360 }
 361
 362 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 363 {
 364         struct bch_fs *c = trans->c;
 365         struct btree_insert_entry *i;
 366
 367         trans_for_each_update(trans, i) {
 368                 /*
 369                  * XXX: synchronization of cached update triggers with gc
 370                  */
 371                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 372
 373                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 374                         bch2_mark_update(trans, i->iter, i->k, NULL,
 375                                          i->trigger_flags|BTREE_TRIGGER_GC);
 376         }
 377 }
 378
 379 static inline int
 380 bch2_trans_commit_write_locked(struct btree_trans *trans,
 381                                struct btree_insert_entry **stopped_at)
 382 {
 383         struct bch_fs *c = trans->c;
 384         struct bch_fs_usage *fs_usage = NULL;
 385         struct btree_insert_entry *i;
 386         unsigned u64s = 0;
 387         bool marking = false;
 388         int ret;
 389
 390         if (race_fault()) {
 391                 trace_trans_restart_fault_inject(trans->ip);
 392                 return -EINTR;
 393         }
 394
 395         /*
 396          * Check if the insert will fit in the leaf node with the write lock
 397          * held, otherwise another thread could write the node changing the
 398          * amount of space available:
 399          */
 400
 401         prefetch(&trans->c->journal.flags);
 402
 403         trans_for_each_update2(trans, i) {
 404                 /* Multiple inserts might go to same leaf: */
 405                 if (!same_leaf_as_prev(trans, i))
 406                         u64s = 0;
 407
 408                 u64s += i->k->k.u64s;
 409                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 410                         ? btree_key_can_insert(trans, i->iter, i->k, u64s)
 411                         : btree_key_can_insert_cached(trans, i->iter, i->k, u64s);
 412                 if (ret) {
 413                         *stopped_at = i;
 414                         return ret;
 415                 }
 416
 417                 if (btree_node_type_needs_gc(i->iter->btree_id))
 418                         marking = true;
 419         }
 420
 421         if (marking) {
 422                 percpu_down_read(&c->mark_lock);
 423                 fs_usage = bch2_fs_usage_scratch_get(c);
 424         }
 425
 426         /*
 427          * Don't get journal reservation until after we know insert will
 428          * succeed:
 429          */
 430         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 431                 ret = bch2_trans_journal_res_get(trans,
 432                                 JOURNAL_RES_GET_NONBLOCK);
 433                 if (ret)
 434                         goto err;
 435         } else {
 436                 trans->journal_res.seq = c->journal.replay_journal_seq;
 437         }
 438
 439         if (unlikely(trans->extra_journal_entry_u64s)) {
 440                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 441                                   trans->extra_journal_entries,
 442                                   trans->extra_journal_entry_u64s);
 443
 444                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 445                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 446         }
 447
 448         /*
 449          * Not allowed to fail after we've gotten our journal reservation - we
 450          * have to use it:
 451          */
 452
 453         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 454                 if (journal_seq_verify(c))
 455                         trans_for_each_update2(trans, i)
 456                                 i->k->k.version.lo = trans->journal_res.seq;
 457                 else if (inject_invalid_keys(c))
 458                         trans_for_each_update2(trans, i)
 459                                 i->k->k.version = MAX_VERSION;
 460         }
 461
 462         /* Must be called under mark_lock: */
 463         if (marking && trans->fs_usage_deltas &&
 464             bch2_replicas_delta_list_apply(c, fs_usage,
 465                                            trans->fs_usage_deltas)) {
 466                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 467                 goto err;
 468         }
 469
 470         trans_for_each_update(trans, i)
 471                 if (iter_has_nontrans_triggers(i->iter))
 472                         bch2_mark_update(trans, i->iter, i->k,
 473                                          fs_usage, i->trigger_flags);
 474
 475         if (marking)
 476                 bch2_trans_fs_usage_apply(trans, fs_usage);
 477
 478         if (unlikely(c->gc_pos.phase))
 479                 bch2_trans_mark_gc(trans);
 480
 481         trans_for_each_update2(trans, i)
 482                 do_btree_insert_one(trans, i->iter, i->k);
 483 err:
 484         if (marking) {
 485                 bch2_fs_usage_scratch_put(c, fs_usage);
 486                 percpu_up_read(&c->mark_lock);
 487         }
 488
 489         return ret;
 490 }
 491
 492 /*
 493  * Get journal reservation, take write locks, and attempt to do btree update(s):
 494  */
 495 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 496                                        struct btree_insert_entry **stopped_at)
 497 {
 498         struct btree_insert_entry *i;
 499         struct btree_iter *iter;
 500         int ret;
 501
 502         trans_for_each_update2(trans, i)
 503                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 504
 505         ret = bch2_journal_preres_get(&trans->c->journal,
 506                         &trans->journal_preres, trans->journal_preres_u64s,
 507                         JOURNAL_RES_GET_NONBLOCK|
 508                         ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
 509                          ? JOURNAL_RES_GET_RECLAIM : 0));
 510         if (unlikely(ret == -EAGAIN))
 511                 ret = bch2_trans_journal_preres_get_cold(trans,
 512                                                 trans->journal_preres_u64s);
 513         if (unlikely(ret))
 514                 return ret;
 515
 516         /*
 517          * Can't be holding any read locks when we go to take write locks:
 518          *
 519          * note - this must be done after bch2_trans_journal_preres_get_cold()
 520          * or anything else that might call bch2_trans_relock(), since that
 521          * would just retake the read locks:
 522          */
 523         trans_for_each_iter(trans, iter) {
 524                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 525                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 526                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 527                         bch2_btree_iter_unlock_noinline(iter);
 528                 }
 529         }
 530
 531         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 532                 trans_for_each_update2(trans, i)
 533                         btree_insert_entry_checks(trans, i->iter, i->k);
 534         bch2_btree_trans_verify_locks(trans);
 535
 536         trans_for_each_update2(trans, i)
 537                 if (!same_leaf_as_prev(trans, i))
 538                         bch2_btree_node_lock_for_insert(trans->c,
 539                                         iter_l(i->iter)->b, i->iter);
 540
 541         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 542
 543         trans_for_each_update2(trans, i)
 544                 if (!same_leaf_as_prev(trans, i))
 545                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 546                                                              i->iter);
 547
 548         if (!ret && trans->journal_pin)
 549                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 550                                      trans->journal_pin, NULL);
 551
 552         /*
 553          * Drop journal reservation after dropping write locks, since dropping
 554          * the journal reservation may kick off a journal write:
 555          */
 556         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 557
 558         if (unlikely(ret))
 559                 return ret;
 560
 561         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 562                 trans->nounlock = true;
 563
 564         trans_for_each_update2(trans, i)
 565                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 566                     !same_leaf_as_prev(trans, i))
 567                         bch2_foreground_maybe_merge(trans->c, i->iter,
 568                                                     0, trans->flags);
 569
 570         trans->nounlock = false;
 571
 572         bch2_trans_downgrade(trans);
 573
 574         return 0;
 575 }
 576
 577 static noinline
 578 int bch2_trans_commit_error(struct btree_trans *trans,
 579                             struct btree_insert_entry *i,
 580                             int ret)
 581 {
 582         struct bch_fs *c = trans->c;
 583         unsigned flags = trans->flags;
 584
 585         /*
 586          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 587          * update; if we haven't done anything yet it doesn't apply
 588          */
 589         flags &= ~BTREE_INSERT_NOUNLOCK;
 590
 591         switch (ret) {
 592         case BTREE_INSERT_BTREE_NODE_FULL:
 593                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 594
 595                 /*
 596                  * if the split succeeded without dropping locks the insert will
 597                  * still be atomic (what the caller peeked() and is overwriting
 598                  * won't have changed)
 599                  */
 600 #if 0
 601                 /*
 602                  * XXX:
 603                  * split -> btree node merging (of parent node) might still drop
 604                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 605                  *
 606                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 607                  * will inhibit merging - but we don't have a reliable way yet
 608                  * (do we?) of checking if we dropped locks in this path
 609                  */
 610                 if (!ret)
 611                         goto retry;
 612 #endif
 613
 614                 /*
 615                  * don't care if we got ENOSPC because we told split it
 616                  * couldn't block:
 617                  */
 618                 if (!ret ||
 619                     ret == -EINTR ||
 620                     (flags & BTREE_INSERT_NOUNLOCK)) {
 621                         trace_trans_restart_btree_node_split(trans->ip);
 622                         ret = -EINTR;
 623                 }
 624                 break;
 625         case BTREE_INSERT_ENOSPC:
 626                 ret = -ENOSPC;
 627                 break;
 628         case BTREE_INSERT_NEED_MARK_REPLICAS:
 629                 bch2_trans_unlock(trans);
 630
 631                 trans_for_each_update(trans, i) {
 632                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 633                         if (ret)
 634                                 return ret;
 635                 }
 636
 637                 if (bch2_trans_relock(trans))
 638                         return 0;
 639
 640                 trace_trans_restart_mark_replicas(trans->ip);
 641                 ret = -EINTR;
 642                 break;
 643         case BTREE_INSERT_NEED_JOURNAL_RES:
 644                 bch2_trans_unlock(trans);
 645
 646                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 647                 if (ret)
 648                         return ret;
 649
 650                 if (bch2_trans_relock(trans))
 651                         return 0;
 652
 653                 trace_trans_restart_journal_res_get(trans->ip);
 654                 ret = -EINTR;
 655                 break;
 656         default:
 657                 BUG_ON(ret >= 0);
 658                 break;
 659         }
 660
 661         if (ret == -EINTR) {
 662                 int ret2 = bch2_btree_iter_traverse_all(trans);
 663
 664                 if (ret2) {
 665                         trace_trans_restart_traverse(trans->ip);
 666                         return ret2;
 667                 }
 668
 669                 trace_trans_restart_atomic(trans->ip);
 670         }
 671
 672         return ret;
 673 }
 674
 675 static noinline int
 676 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 677 {
 678         struct bch_fs *c = trans->c;
 679         int ret;
 680
 681         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 682                 return -EROFS;
 683
 684         bch2_trans_unlock(trans);
 685
 686         ret = bch2_fs_read_write_early(c);
 687         if (ret)
 688                 return ret;
 689
 690         percpu_ref_get(&c->writes);
 691         return 0;
 692 }
 693
 694 static void bch2_trans_update2(struct btree_trans *trans,
 695                                struct btree_iter *iter,
 696                                struct bkey_i *insert)
 697 {
 698         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 699                 .iter = iter, .k = insert
 700         };
 701
 702         btree_insert_entry_checks(trans, n.iter, n.k);
 703
 704         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 705
 706         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 707
 708         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 709
 710         trans_for_each_update2(trans, i) {
 711                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 712                         *i = n;
 713                         return;
 714                 }
 715
 716                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 717                         break;
 718         }
 719
 720         array_insert_item(trans->updates2, trans->nr_updates2,
 721                           i - trans->updates2, n);
 722 }
 723
 724 static int extent_update_to_keys(struct btree_trans *trans,
 725                                  struct btree_iter *orig_iter,
 726                                  struct bkey_i *insert)
 727 {
 728         struct btree_iter *iter;
 729         int ret;
 730
 731         ret = bch2_extent_can_insert(trans, orig_iter, insert);
 732         if (ret)
 733                 return ret;
 734
 735         if (bkey_deleted(&insert->k))
 736                 return 0;
 737
 738         iter = bch2_trans_copy_iter(trans, orig_iter);
 739         if (IS_ERR(iter))
 740                 return PTR_ERR(iter);
 741
 742         iter->flags |= BTREE_ITER_INTENT;
 743         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 744         bch2_trans_update2(trans, iter, insert);
 745         bch2_trans_iter_put(trans, iter);
 746         return 0;
 747 }
 748
 749 static int extent_handle_overwrites(struct btree_trans *trans,
 750                                     enum btree_id btree_id,
 751                                     struct bpos start, struct bpos end)
 752 {
 753         struct btree_iter *iter = NULL, *update_iter;
 754         struct bkey_i *update;
 755         struct bkey_s_c k;
 756         int ret = 0;
 757
 758         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 759         ret = PTR_ERR_OR_ZERO(iter);
 760         if (ret)
 761                 return ret;
 762
 763         k = bch2_btree_iter_peek_with_updates(iter);
 764
 765         while (k.k && !(ret = bkey_err(k))) {
 766                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 767                         break;
 768
 769                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 770                         update_iter = bch2_trans_copy_iter(trans, iter);
 771                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 772                                 goto err;
 773
 774                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 775                         if ((ret = PTR_ERR_OR_ZERO(update)))
 776                                 goto err;
 777
 778                         bkey_reassemble(update, k);
 779                         bch2_cut_back(start, update);
 780
 781                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 782                         bch2_trans_update2(trans, update_iter, update);
 783                         bch2_trans_iter_put(trans, update_iter);
 784                 }
 785
 786                 if (bkey_cmp(k.k->p, end) > 0) {
 787                         update_iter = bch2_trans_copy_iter(trans, iter);
 788                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 789                                 goto err;
 790
 791                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 792                         if ((ret = PTR_ERR_OR_ZERO(update)))
 793                                 goto err;
 794
 795                         bkey_reassemble(update, k);
 796                         bch2_cut_front(end, update);
 797
 798                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 799                         bch2_trans_update2(trans, update_iter, update);
 800                         bch2_trans_iter_put(trans, update_iter);
 801                 } else {
 802                         update_iter = bch2_trans_copy_iter(trans, iter);
 803                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 804                                 goto err;
 805
 806                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 807                         if ((ret = PTR_ERR_OR_ZERO(update)))
 808                                 goto err;
 809
 810                         update->k = *k.k;
 811                         set_bkey_val_u64s(&update->k, 0);
 812                         update->k.type = KEY_TYPE_deleted;
 813                         update->k.size = 0;
 814
 815                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 816                         bch2_trans_update2(trans, update_iter, update);
 817                         bch2_trans_iter_put(trans, update_iter);
 818                 }
 819
 820                 k = bch2_btree_iter_next_with_updates(iter);
 821         }
 822 err:
 823         if (!IS_ERR_OR_NULL(iter))
 824                 bch2_trans_iter_put(trans, iter);
 825         return ret;
 826 }
 827
 828 int __bch2_trans_commit(struct btree_trans *trans)
 829 {
 830         struct btree_insert_entry *i = NULL;
 831         struct btree_iter *iter;
 832         bool trans_trigger_run;
 833         unsigned u64s;
 834         int ret = 0;
 835
 836         BUG_ON(trans->need_reset);
 837
 838         if (!trans->nr_updates)
 839                 goto out_noupdates;
 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                         return ret;
 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                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
 873                                      (ret = bch2_btree_iter_traverse(i->iter)))) {
 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(i->iter->locks_want < 1 &&
 883                                      !__bch2_btree_iter_upgrade(i->iter, 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_noupdates:
 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         EBUG_ON(bkey_cmp(iter->pos,
 985                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 986                          ? bkey_start_pos(&k->k)
 987                          : k->k.p));
 988
 989         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 990
 991         if (btree_node_type_is_extents(iter->btree_id)) {
 992                 iter->pos_after_commit = k->k.p;
 993                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 994         }
 995
 996         /*
 997          * Pending updates are kept sorted: first, find position of new update:
 998          */
 999         trans_for_each_update(trans, i)
1000                 if (btree_iter_cmp(iter, i->iter) <= 0)
1001                         break;
1002
1003         /*
1004          * Now delete/trim any updates the new update overwrites:
1005          */
1006         if (i > trans->updates &&
1007             i[-1].iter->btree_id == iter->btree_id &&
1008             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
1009                 bch2_cut_back(n.iter->pos, i[-1].k);
1010
1011         while (i < trans->updates + trans->nr_updates &&
1012                iter->btree_id == i->iter->btree_id &&
1013                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1014                 array_remove_item(trans->updates, trans->nr_updates,
1015                                   i - trans->updates);
1016
1017         if (i < trans->updates + trans->nr_updates &&
1018             iter->btree_id == i->iter->btree_id &&
1019             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1020                 /*
1021                  * When we have an extent that overwrites the start of another
1022                  * update, trimming that extent will mean the iterator's
1023                  * position has to change since the iterator position has to
1024                  * match the extent's start pos - but we don't want to change
1025                  * the iterator pos if some other code is using it, so we may
1026                  * need to clone it:
1027                  */
1028                 if (trans->iters_live & (1ULL << i->iter->idx)) {
1029                         i->iter = bch2_trans_copy_iter(trans, i->iter);
1030                         if (IS_ERR(i->iter)) {
1031                                 trans->need_reset = true;
1032                                 return PTR_ERR(i->iter);
1033                         }
1034
1035                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1036                         bch2_trans_iter_put(trans, i->iter);
1037                 }
1038
1039                 bch2_cut_front(n.k->k.p, i->k);
1040                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1041         }
1042
1043         EBUG_ON(trans->nr_updates >= trans->nr_iters);
1044
1045         array_insert_item(trans->updates, trans->nr_updates,
1046                           i - trans->updates, n);
1047         return 0;
1048 }
1049
1050 int __bch2_btree_insert(struct btree_trans *trans,
1051                         enum btree_id id, struct bkey_i *k)
1052 {
1053         struct btree_iter *iter;
1054         int ret;
1055
1056         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1057                                    BTREE_ITER_INTENT);
1058         if (IS_ERR(iter))
1059                 return PTR_ERR(iter);
1060
1061         ret   = bch2_btree_iter_traverse(iter) ?:
1062                 bch2_trans_update(trans, iter, k, 0);
1063         bch2_trans_iter_put(trans, iter);
1064         return ret;
1065 }
1066
1067 /**
1068  * bch2_btree_insert - insert keys into the extent btree
1069  * @c:                  pointer to struct bch_fs
1070  * @id:                 btree to insert into
1071  * @insert_keys:        list of keys to insert
1072  * @hook:               insert callback
1073  */
1074 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1075                       struct bkey_i *k,
1076                       struct disk_reservation *disk_res,
1077                       u64 *journal_seq, int flags)
1078 {
1079         return bch2_trans_do(c, disk_res, journal_seq, flags,
1080                              __bch2_btree_insert(&trans, id, k));
1081 }
1082
1083 int bch2_btree_delete_at_range(struct btree_trans *trans,
1084                                struct btree_iter *iter,
1085                                struct bpos end,
1086                                u64 *journal_seq)
1087 {
1088         struct bkey_s_c k;
1089         int ret = 0;
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                  * For extents, iter.pos won't necessarily be the same as
1102                  * bkey_start_pos(k.k) (for non extents they always will be the
1103                  * same). It's important that we delete starting from iter.pos
1104                  * because the range we want to delete could start in the middle
1105                  * of k.
1106                  *
1107                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1108                  * bkey_start_pos(k.k)).
1109                  */
1110                 delete.k.p = iter->pos;
1111
1112                 if (btree_node_type_is_extents(iter->btree_id)) {
1113                         unsigned max_sectors =
1114                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1115
1116                         /* create the biggest key we can */
1117                         bch2_key_resize(&delete.k, max_sectors);
1118                         bch2_cut_back(end, &delete);
1119
1120                         ret = bch2_extent_trim_atomic(&delete, iter);
1121                         if (ret)
1122                                 break;
1123                 }
1124
1125                 bch2_trans_update(trans, iter, &delete, 0);
1126                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1127                                         BTREE_INSERT_NOFAIL);
1128                 if (ret)
1129                         break;
1130
1131                 bch2_trans_cond_resched(trans);
1132         }
1133
1134         if (ret == -EINTR) {
1135                 ret = 0;
1136                 goto retry;
1137         }
1138
1139         return ret;
1140
1141 }
1142
1143 int bch2_btree_delete_at(struct btree_trans *trans,
1144                          struct btree_iter *iter, unsigned flags)
1145 {
1146         struct bkey_i k;
1147
1148         bkey_init(&k.k);
1149         k.k.p = iter->pos;
1150
1151         bch2_trans_update(trans, iter, &k, 0);
1152         return bch2_trans_commit(trans, NULL, NULL,
1153                                  BTREE_INSERT_NOFAIL|
1154                                  BTREE_INSERT_USE_RESERVE|flags);
1155 }
1156
1157 /*
1158  * bch_btree_delete_range - delete everything within a given range
1159  *
1160  * Range is a half open interval - [start, end)
1161  */
1162 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1163                             struct bpos start, struct bpos end,
1164                             u64 *journal_seq)
1165 {
1166         struct btree_trans trans;
1167         struct btree_iter *iter;
1168         int ret = 0;
1169
1170         /*
1171          * XXX: whether we need mem/more iters depends on whether this btree id
1172          * has triggers
1173          */
1174         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1175
1176         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1177
1178         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1179         ret = bch2_trans_exit(&trans) ?: ret;
1180
1181         BUG_ON(ret == -EINTR);
1182         return ret;
1183 }