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                      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 (u64s <= ck->u64s)
 290                 return BTREE_INSERT_OK;
 291
 292         new_u64s        = roundup_pow_of_two(u64s);
 293         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 294         if (!new_k)
 295                 return -ENOMEM;
 296
 297         ck->u64s        = new_u64s;
 298         ck->k           = new_k;
 299         return BTREE_INSERT_OK;
 300 }
 301
 302 static inline void do_btree_insert_one(struct btree_trans *trans,
 303                                        struct btree_iter *iter,
 304                                        struct bkey_i *insert)
 305 {
 306         struct bch_fs *c = trans->c;
 307         struct journal *j = &c->journal;
 308         bool did_work;
 309
 310         EBUG_ON(trans->journal_res.ref !=
 311                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 312
 313         insert->k.needs_whiteout = false;
 314
 315         did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
 316                 ? btree_insert_key_leaf(trans, iter, insert)
 317                 : bch2_btree_insert_key_cached(trans, iter, insert);
 318         if (!did_work)
 319                 return;
 320
 321         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 322                 bch2_journal_add_keys(j, &trans->journal_res,
 323                                       iter->btree_id, insert);
 324
 325                 bch2_journal_set_has_inode(j, &trans->journal_res,
 326                                            insert->k.p.inode);
 327
 328                 if (trans->journal_seq)
 329                         *trans->journal_seq = trans->journal_res.seq;
 330         }
 331 }
 332
 333 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 334 {
 335         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 336 }
 337
 338 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 339 {
 340         return (((BTREE_NODE_TYPE_HAS_TRIGGERS &
 341                   ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS)) |
 342                 (1U << BTREE_ID_EC)) &
 343                 (1U << iter->btree_id);
 344 }
 345
 346 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 347 {
 348         __bch2_btree_iter_unlock(iter);
 349 }
 350
 351 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 352 {
 353         struct bch_fs *c = trans->c;
 354         struct btree_insert_entry *i;
 355
 356         trans_for_each_update(trans, i) {
 357                 /*
 358                  * XXX: synchronization of cached update triggers with gc
 359                  */
 360                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 361
 362                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 363                         bch2_mark_update(trans, i->iter, i->k, NULL,
 364                                          i->trigger_flags|BTREE_TRIGGER_GC);
 365         }
 366 }
 367
 368 static inline int
 369 bch2_trans_commit_write_locked(struct btree_trans *trans,
 370                                struct btree_insert_entry **stopped_at)
 371 {
 372         struct bch_fs *c = trans->c;
 373         struct bch_fs_usage *fs_usage = NULL;
 374         struct btree_insert_entry *i;
 375         unsigned u64s = 0;
 376         bool marking = false;
 377         int ret;
 378
 379         if (race_fault()) {
 380                 trace_trans_restart_fault_inject(trans->ip);
 381                 return -EINTR;
 382         }
 383
 384         /*
 385          * Check if the insert will fit in the leaf node with the write lock
 386          * held, otherwise another thread could write the node changing the
 387          * amount of space available:
 388          */
 389
 390         prefetch(&trans->c->journal.flags);
 391
 392         trans_for_each_update2(trans, i) {
 393                 /* Multiple inserts might go to same leaf: */
 394                 if (!same_leaf_as_prev(trans, i))
 395                         u64s = 0;
 396
 397                 u64s += i->k->k.u64s;
 398                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 399                         ? btree_key_can_insert(trans, i->iter, u64s)
 400                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 401                 if (ret) {
 402                         *stopped_at = i;
 403                         return ret;
 404                 }
 405
 406                 if (btree_node_type_needs_gc(i->iter->btree_id))
 407                         marking = true;
 408         }
 409
 410         if (marking) {
 411                 percpu_down_read(&c->mark_lock);
 412                 fs_usage = bch2_fs_usage_scratch_get(c);
 413         }
 414
 415         /*
 416          * Don't get journal reservation until after we know insert will
 417          * succeed:
 418          */
 419         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 420                 ret = bch2_trans_journal_res_get(trans,
 421                                 JOURNAL_RES_GET_NONBLOCK);
 422                 if (ret)
 423                         goto err;
 424         } else {
 425                 trans->journal_res.seq = c->journal.replay_journal_seq;
 426         }
 427
 428         if (unlikely(trans->extra_journal_entry_u64s)) {
 429                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 430                                   trans->extra_journal_entries,
 431                                   trans->extra_journal_entry_u64s);
 432
 433                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 434                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 435         }
 436
 437         /*
 438          * Not allowed to fail after we've gotten our journal reservation - we
 439          * have to use it:
 440          */
 441
 442         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 443                 if (journal_seq_verify(c))
 444                         trans_for_each_update2(trans, i)
 445                                 i->k->k.version.lo = trans->journal_res.seq;
 446                 else if (inject_invalid_keys(c))
 447                         trans_for_each_update2(trans, i)
 448                                 i->k->k.version = MAX_VERSION;
 449         }
 450
 451         /* Must be called under mark_lock: */
 452         if (marking && trans->fs_usage_deltas &&
 453             bch2_replicas_delta_list_apply(c, fs_usage,
 454                                            trans->fs_usage_deltas)) {
 455                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 456                 goto err;
 457         }
 458
 459         trans_for_each_update(trans, i)
 460                 if (iter_has_nontrans_triggers(i->iter))
 461                         bch2_mark_update(trans, i->iter, i->k,
 462                                          fs_usage, i->trigger_flags);
 463
 464         if (marking)
 465                 bch2_trans_fs_usage_apply(trans, fs_usage);
 466
 467         if (unlikely(c->gc_pos.phase))
 468                 bch2_trans_mark_gc(trans);
 469
 470         trans_for_each_update2(trans, i)
 471                 do_btree_insert_one(trans, i->iter, i->k);
 472 err:
 473         if (marking) {
 474                 bch2_fs_usage_scratch_put(c, fs_usage);
 475                 percpu_up_read(&c->mark_lock);
 476         }
 477
 478         return ret;
 479 }
 480
 481 /*
 482  * Get journal reservation, take write locks, and attempt to do btree update(s):
 483  */
 484 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 485                                        struct btree_insert_entry **stopped_at)
 486 {
 487         struct btree_insert_entry *i;
 488         struct btree_iter *iter;
 489         int ret;
 490
 491         trans_for_each_update2(trans, i)
 492                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 493
 494         ret = bch2_journal_preres_get(&trans->c->journal,
 495                         &trans->journal_preres, trans->journal_preres_u64s,
 496                         JOURNAL_RES_GET_NONBLOCK|
 497                         ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
 498                          ? JOURNAL_RES_GET_RECLAIM : 0));
 499         if (unlikely(ret == -EAGAIN))
 500                 ret = bch2_trans_journal_preres_get_cold(trans,
 501                                                 trans->journal_preres_u64s);
 502         if (unlikely(ret))
 503                 return ret;
 504
 505         /*
 506          * Can't be holding any read locks when we go to take write locks:
 507          *
 508          * note - this must be done after bch2_trans_journal_preres_get_cold()
 509          * or anything else that might call bch2_trans_relock(), since that
 510          * would just retake the read locks:
 511          */
 512         trans_for_each_iter(trans, iter) {
 513                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 514                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 515                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 516                         bch2_btree_iter_unlock_noinline(iter);
 517                 }
 518         }
 519
 520         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 521                 trans_for_each_update2(trans, i)
 522                         btree_insert_entry_checks(trans, i->iter, i->k);
 523         bch2_btree_trans_verify_locks(trans);
 524
 525         trans_for_each_update2(trans, i)
 526                 if (!same_leaf_as_prev(trans, i))
 527                         bch2_btree_node_lock_for_insert(trans->c,
 528                                         iter_l(i->iter)->b, i->iter);
 529
 530         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 531
 532         trans_for_each_update2(trans, i)
 533                 if (!same_leaf_as_prev(trans, i))
 534                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 535                                                              i->iter);
 536
 537         if (!ret && trans->journal_pin)
 538                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 539                                      trans->journal_pin, NULL);
 540
 541         /*
 542          * Drop journal reservation after dropping write locks, since dropping
 543          * the journal reservation may kick off a journal write:
 544          */
 545         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 546
 547         if (unlikely(ret))
 548                 return ret;
 549
 550         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 551                 trans->nounlock = true;
 552
 553         trans_for_each_update2(trans, i)
 554                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 555                     !same_leaf_as_prev(trans, i))
 556                         bch2_foreground_maybe_merge(trans->c, i->iter,
 557                                                     0, trans->flags);
 558
 559         trans->nounlock = false;
 560
 561         bch2_trans_downgrade(trans);
 562
 563         return 0;
 564 }
 565
 566 static noinline
 567 int bch2_trans_commit_error(struct btree_trans *trans,
 568                             struct btree_insert_entry *i,
 569                             int ret)
 570 {
 571         struct bch_fs *c = trans->c;
 572         unsigned flags = trans->flags;
 573
 574         /*
 575          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 576          * update; if we haven't done anything yet it doesn't apply
 577          */
 578         flags &= ~BTREE_INSERT_NOUNLOCK;
 579
 580         switch (ret) {
 581         case BTREE_INSERT_BTREE_NODE_FULL:
 582                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 583
 584                 /*
 585                  * if the split succeeded without dropping locks the insert will
 586                  * still be atomic (what the caller peeked() and is overwriting
 587                  * won't have changed)
 588                  */
 589 #if 0
 590                 /*
 591                  * XXX:
 592                  * split -> btree node merging (of parent node) might still drop
 593                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 594                  *
 595                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 596                  * will inhibit merging - but we don't have a reliable way yet
 597                  * (do we?) of checking if we dropped locks in this path
 598                  */
 599                 if (!ret)
 600                         goto retry;
 601 #endif
 602
 603                 /*
 604                  * don't care if we got ENOSPC because we told split it
 605                  * couldn't block:
 606                  */
 607                 if (!ret ||
 608                     ret == -EINTR ||
 609                     (flags & BTREE_INSERT_NOUNLOCK)) {
 610                         trace_trans_restart_btree_node_split(trans->ip);
 611                         ret = -EINTR;
 612                 }
 613                 break;
 614         case BTREE_INSERT_ENOSPC:
 615                 ret = -ENOSPC;
 616                 break;
 617         case BTREE_INSERT_NEED_MARK_REPLICAS:
 618                 bch2_trans_unlock(trans);
 619
 620                 trans_for_each_update(trans, i) {
 621                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 622                         if (ret)
 623                                 return ret;
 624                 }
 625
 626                 if (bch2_trans_relock(trans))
 627                         return 0;
 628
 629                 trace_trans_restart_mark_replicas(trans->ip);
 630                 ret = -EINTR;
 631                 break;
 632         case BTREE_INSERT_NEED_JOURNAL_RES:
 633                 bch2_trans_unlock(trans);
 634
 635                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 636                 if (ret)
 637                         return ret;
 638
 639                 if (bch2_trans_relock(trans))
 640                         return 0;
 641
 642                 trace_trans_restart_journal_res_get(trans->ip);
 643                 ret = -EINTR;
 644                 break;
 645         default:
 646                 BUG_ON(ret >= 0);
 647                 break;
 648         }
 649
 650         if (ret == -EINTR) {
 651                 int ret2 = bch2_btree_iter_traverse_all(trans);
 652
 653                 if (ret2) {
 654                         trace_trans_restart_traverse(trans->ip);
 655                         return ret2;
 656                 }
 657
 658                 trace_trans_restart_atomic(trans->ip);
 659         }
 660
 661         return ret;
 662 }
 663
 664 static noinline int
 665 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 666 {
 667         struct bch_fs *c = trans->c;
 668         int ret;
 669
 670         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 671                 return -EROFS;
 672
 673         bch2_trans_unlock(trans);
 674
 675         ret = bch2_fs_read_write_early(c);
 676         if (ret)
 677                 return ret;
 678
 679         percpu_ref_get(&c->writes);
 680         return 0;
 681 }
 682
 683 static void bch2_trans_update2(struct btree_trans *trans,
 684                                struct btree_iter *iter,
 685                                struct bkey_i *insert)
 686 {
 687         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 688                 .iter = iter, .k = insert
 689         };
 690
 691         btree_insert_entry_checks(trans, n.iter, n.k);
 692
 693         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 694
 695         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 696
 697         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 698
 699         trans_for_each_update2(trans, i) {
 700                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 701                         *i = n;
 702                         return;
 703                 }
 704
 705                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 706                         break;
 707         }
 708
 709         array_insert_item(trans->updates2, trans->nr_updates2,
 710                           i - trans->updates2, n);
 711 }
 712
 713 static int extent_update_to_keys(struct btree_trans *trans,
 714                                  struct btree_iter *orig_iter,
 715                                  struct bkey_i *insert)
 716 {
 717         struct btree_iter *iter;
 718         int ret;
 719
 720         ret = bch2_extent_can_insert(trans, orig_iter, insert);
 721         if (ret)
 722                 return ret;
 723
 724         if (bkey_deleted(&insert->k))
 725                 return 0;
 726
 727         iter = bch2_trans_copy_iter(trans, orig_iter);
 728         if (IS_ERR(iter))
 729                 return PTR_ERR(iter);
 730
 731         iter->flags |= BTREE_ITER_INTENT;
 732         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 733         bch2_trans_update2(trans, iter, insert);
 734         bch2_trans_iter_put(trans, iter);
 735         return 0;
 736 }
 737
 738 static int extent_handle_overwrites(struct btree_trans *trans,
 739                                     enum btree_id btree_id,
 740                                     struct bpos start, struct bpos end)
 741 {
 742         struct btree_iter *iter = NULL, *update_iter;
 743         struct bkey_i *update;
 744         struct bkey_s_c k;
 745         int ret = 0;
 746
 747         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 748         ret = PTR_ERR_OR_ZERO(iter);
 749         if (ret)
 750                 return ret;
 751
 752         k = bch2_btree_iter_peek_with_updates(iter);
 753
 754         while (k.k && !(ret = bkey_err(k))) {
 755                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 756                         break;
 757
 758                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 759                         update_iter = bch2_trans_copy_iter(trans, iter);
 760                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 761                                 goto err;
 762
 763                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 764                         if ((ret = PTR_ERR_OR_ZERO(update)))
 765                                 goto err;
 766
 767                         bkey_reassemble(update, k);
 768                         bch2_cut_back(start, update);
 769
 770                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 771                         bch2_trans_update2(trans, update_iter, update);
 772                         bch2_trans_iter_put(trans, update_iter);
 773                 }
 774
 775                 if (bkey_cmp(k.k->p, end) > 0) {
 776                         update_iter = bch2_trans_copy_iter(trans, iter);
 777                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 778                                 goto err;
 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_front(end, 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                 } else {
 791                         update_iter = bch2_trans_copy_iter(trans, iter);
 792                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 793                                 goto err;
 794
 795                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 796                         if ((ret = PTR_ERR_OR_ZERO(update)))
 797                                 goto err;
 798
 799                         update->k = *k.k;
 800                         set_bkey_val_u64s(&update->k, 0);
 801                         update->k.type = KEY_TYPE_deleted;
 802                         update->k.size = 0;
 803
 804                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 805                         bch2_trans_update2(trans, update_iter, update);
 806                         bch2_trans_iter_put(trans, update_iter);
 807                 }
 808
 809                 k = bch2_btree_iter_next_with_updates(iter);
 810         }
 811 err:
 812         if (!IS_ERR_OR_NULL(iter))
 813                 bch2_trans_iter_put(trans, iter);
 814         return ret;
 815 }
 816
 817 int __bch2_trans_commit(struct btree_trans *trans)
 818 {
 819         struct btree_insert_entry *i = NULL;
 820         struct btree_iter *iter;
 821         bool trans_trigger_run;
 822         unsigned u64s;
 823         int ret = 0;
 824
 825         BUG_ON(trans->need_reset);
 826
 827         if (!trans->nr_updates)
 828                 goto out_noupdates;
 829
 830         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 831                 lockdep_assert_held(&trans->c->gc_lock);
 832
 833         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 834
 835         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 836         trans->journal_preres_u64s      = 0;
 837
 838         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 839             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 840                 ret = bch2_trans_commit_get_rw_cold(trans);
 841                 if (ret)
 842                         return ret;
 843         }
 844
 845 #ifdef CONFIG_BCACHEFS_DEBUG
 846         trans_for_each_update(trans, i)
 847                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 848                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 849                         bch2_btree_key_cache_verify_clean(trans,
 850                                         i->iter->btree_id, i->iter->pos);
 851 #endif
 852
 853         /*
 854          * Running triggers will append more updates to the list of updates as
 855          * we're walking it:
 856          */
 857         do {
 858                 trans_trigger_run = false;
 859
 860                 trans_for_each_update(trans, i) {
 861                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
 862                                      (ret = bch2_btree_iter_traverse(i->iter)))) {
 863                                 trace_trans_restart_traverse(trans->ip);
 864                                 goto out;
 865                         }
 866
 867                         /*
 868                          * We're not using bch2_btree_iter_upgrade here because
 869                          * we know trans->nounlock can't be set:
 870                          */
 871                         if (unlikely(i->iter->locks_want < 1 &&
 872                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 873                                 trace_trans_restart_upgrade(trans->ip);
 874                                 ret = -EINTR;
 875                                 goto out;
 876                         }
 877
 878                         if (iter_has_trans_triggers(i->iter) &&
 879                             !i->trans_triggers_run) {
 880                                 i->trans_triggers_run = true;
 881                                 trans_trigger_run = true;
 882
 883                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 884                                                              i->trigger_flags);
 885                                 if (unlikely(ret)) {
 886                                         if (ret == -EINTR)
 887                                                 trace_trans_restart_mark(trans->ip);
 888                                         goto out;
 889                                 }
 890                         }
 891                 }
 892         } while (trans_trigger_run);
 893
 894         /* Turn extents updates into keys: */
 895         trans_for_each_update(trans, i)
 896                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 897                         struct bpos start = bkey_start_pos(&i->k->k);
 898
 899                         while (i + 1 < trans->updates + trans->nr_updates &&
 900                                i[0].iter->btree_id == i[1].iter->btree_id &&
 901                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 902                                 i++;
 903
 904                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 905                                                        start, i->k->k.p);
 906                         if (ret)
 907                                 goto out;
 908                 }
 909
 910         trans_for_each_update(trans, i) {
 911                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 912                         ret = extent_update_to_keys(trans, i->iter, i->k);
 913                         if (ret)
 914                                 goto out;
 915                 } else {
 916                         bch2_trans_update2(trans, i->iter, i->k);
 917                 }
 918         }
 919
 920         trans_for_each_update2(trans, i) {
 921                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 922                 BUG_ON(i->iter->locks_want < 1);
 923
 924                 u64s = jset_u64s(i->k->k.u64s);
 925                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 926                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 927                         trans->journal_preres_u64s += u64s;
 928                 trans->journal_u64s += u64s;
 929         }
 930 retry:
 931         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 932
 933         ret = do_bch2_trans_commit(trans, &i);
 934
 935         /* make sure we didn't drop or screw up locks: */
 936         bch2_btree_trans_verify_locks(trans);
 937
 938         if (ret)
 939                 goto err;
 940
 941         trans_for_each_iter(trans, iter)
 942                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 943                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 944                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 945                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 946                         else
 947                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 948                 }
 949 out:
 950         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 951
 952         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 953                 percpu_ref_put(&trans->c->writes);
 954 out_noupdates:
 955         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 956
 957         return ret;
 958 err:
 959         ret = bch2_trans_commit_error(trans, i, ret);
 960         if (ret)
 961                 goto out;
 962
 963         goto retry;
 964 }
 965
 966 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 967                       struct bkey_i *k, enum btree_trigger_flags flags)
 968 {
 969         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 970                 .trigger_flags = flags, .iter = iter, .k = k
 971         };
 972
 973         EBUG_ON(bkey_cmp(iter->pos,
 974                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 975                          ? bkey_start_pos(&k->k)
 976                          : k->k.p));
 977
 978         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 979
 980         if (btree_node_type_is_extents(iter->btree_id)) {
 981                 iter->pos_after_commit = k->k.p;
 982                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 983         }
 984
 985         /*
 986          * Pending updates are kept sorted: first, find position of new update:
 987          */
 988         trans_for_each_update(trans, i)
 989                 if (btree_iter_cmp(iter, i->iter) <= 0)
 990                         break;
 991
 992         /*
 993          * Now delete/trim any updates the new update overwrites:
 994          */
 995         if (i > trans->updates &&
 996             i[-1].iter->btree_id == iter->btree_id &&
 997             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 998                 bch2_cut_back(n.iter->pos, i[-1].k);
 999
1000         while (i < trans->updates + trans->nr_updates &&
1001                iter->btree_id == i->iter->btree_id &&
1002                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1003                 array_remove_item(trans->updates, trans->nr_updates,
1004                                   i - trans->updates);
1005
1006         if (i < trans->updates + trans->nr_updates &&
1007             iter->btree_id == i->iter->btree_id &&
1008             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1009                 /*
1010                  * When we have an extent that overwrites the start of another
1011                  * update, trimming that extent will mean the iterator's
1012                  * position has to change since the iterator position has to
1013                  * match the extent's start pos - but we don't want to change
1014                  * the iterator pos if some other code is using it, so we may
1015                  * need to clone it:
1016                  */
1017                 if (trans->iters_live & (1ULL << i->iter->idx)) {
1018                         i->iter = bch2_trans_copy_iter(trans, i->iter);
1019                         if (IS_ERR(i->iter)) {
1020                                 trans->need_reset = true;
1021                                 return PTR_ERR(i->iter);
1022                         }
1023
1024                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1025                         bch2_trans_iter_put(trans, i->iter);
1026                 }
1027
1028                 bch2_cut_front(n.k->k.p, i->k);
1029                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1030         }
1031
1032         EBUG_ON(trans->nr_updates >= trans->nr_iters);
1033
1034         array_insert_item(trans->updates, trans->nr_updates,
1035                           i - trans->updates, n);
1036         return 0;
1037 }
1038
1039 int __bch2_btree_insert(struct btree_trans *trans,
1040                         enum btree_id id, struct bkey_i *k)
1041 {
1042         struct btree_iter *iter;
1043         int ret;
1044
1045         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1046                                    BTREE_ITER_INTENT);
1047         if (IS_ERR(iter))
1048                 return PTR_ERR(iter);
1049
1050         ret   = bch2_btree_iter_traverse(iter) ?:
1051                 bch2_trans_update(trans, iter, k, 0);
1052         bch2_trans_iter_put(trans, iter);
1053         return ret;
1054 }
1055
1056 /**
1057  * bch2_btree_insert - insert keys into the extent btree
1058  * @c:                  pointer to struct bch_fs
1059  * @id:                 btree to insert into
1060  * @insert_keys:        list of keys to insert
1061  * @hook:               insert callback
1062  */
1063 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1064                       struct bkey_i *k,
1065                       struct disk_reservation *disk_res,
1066                       u64 *journal_seq, int flags)
1067 {
1068         return bch2_trans_do(c, disk_res, journal_seq, flags,
1069                              __bch2_btree_insert(&trans, id, k));
1070 }
1071
1072 int bch2_btree_delete_at_range(struct btree_trans *trans,
1073                                struct btree_iter *iter,
1074                                struct bpos end,
1075                                u64 *journal_seq)
1076 {
1077         struct bkey_s_c k;
1078         int ret = 0;
1079 retry:
1080         while ((k = bch2_btree_iter_peek(iter)).k &&
1081                !(ret = bkey_err(k)) &&
1082                bkey_cmp(iter->pos, end) < 0) {
1083                 struct bkey_i delete;
1084
1085                 bch2_trans_begin(trans);
1086
1087                 bkey_init(&delete.k);
1088
1089                 /*
1090                  * For extents, iter.pos won't necessarily be the same as
1091                  * bkey_start_pos(k.k) (for non extents they always will be the
1092                  * same). It's important that we delete starting from iter.pos
1093                  * because the range we want to delete could start in the middle
1094                  * of k.
1095                  *
1096                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1097                  * bkey_start_pos(k.k)).
1098                  */
1099                 delete.k.p = iter->pos;
1100
1101                 if (btree_node_type_is_extents(iter->btree_id)) {
1102                         unsigned max_sectors =
1103                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1104
1105                         /* create the biggest key we can */
1106                         bch2_key_resize(&delete.k, max_sectors);
1107                         bch2_cut_back(end, &delete);
1108
1109                         ret = bch2_extent_trim_atomic(&delete, iter);
1110                         if (ret)
1111                                 break;
1112                 }
1113
1114                 bch2_trans_update(trans, iter, &delete, 0);
1115                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1116                                         BTREE_INSERT_NOFAIL);
1117                 if (ret)
1118                         break;
1119
1120                 bch2_trans_cond_resched(trans);
1121         }
1122
1123         if (ret == -EINTR) {
1124                 ret = 0;
1125                 goto retry;
1126         }
1127
1128         return ret;
1129
1130 }
1131
1132 int bch2_btree_delete_at(struct btree_trans *trans,
1133                          struct btree_iter *iter, unsigned flags)
1134 {
1135         struct bkey_i k;
1136
1137         bkey_init(&k.k);
1138         k.k.p = iter->pos;
1139
1140         bch2_trans_update(trans, iter, &k, 0);
1141         return bch2_trans_commit(trans, NULL, NULL,
1142                                  BTREE_INSERT_NOFAIL|
1143                                  BTREE_INSERT_USE_RESERVE|flags);
1144 }
1145
1146 /*
1147  * bch_btree_delete_range - delete everything within a given range
1148  *
1149  * Range is a half open interval - [start, end)
1150  */
1151 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1152                             struct bpos start, struct bpos end,
1153                             u64 *journal_seq)
1154 {
1155         struct btree_trans trans;
1156         struct btree_iter *iter;
1157         int ret = 0;
1158
1159         /*
1160          * XXX: whether we need mem/more iters depends on whether this btree id
1161          * has triggers
1162          */
1163         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1164
1165         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1166
1167         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1168         ret = bch2_trans_exit(&trans) ?: ret;
1169
1170         BUG_ON(ret == -EINTR);
1171         return ret;
1172 }