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 << iter->btree_id);
 343 }
 344
 345 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 346 {
 347         __bch2_btree_iter_unlock(iter);
 348 }
 349
 350 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 351 {
 352         struct bch_fs *c = trans->c;
 353         struct btree_insert_entry *i;
 354
 355         trans_for_each_update(trans, i) {
 356                 /*
 357                  * XXX: synchronization of cached update triggers with gc
 358                  */
 359                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 360
 361                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 362                         bch2_mark_update(trans, i->iter, i->k, NULL,
 363                                          i->trigger_flags|BTREE_TRIGGER_GC);
 364         }
 365 }
 366
 367 static inline int
 368 bch2_trans_commit_write_locked(struct btree_trans *trans,
 369                                struct btree_insert_entry **stopped_at)
 370 {
 371         struct bch_fs *c = trans->c;
 372         struct bch_fs_usage *fs_usage = NULL;
 373         struct btree_insert_entry *i;
 374         unsigned u64s = 0;
 375         bool marking = false;
 376         int ret;
 377
 378         if (race_fault()) {
 379                 trace_trans_restart_fault_inject(trans->ip);
 380                 return -EINTR;
 381         }
 382
 383         /*
 384          * Check if the insert will fit in the leaf node with the write lock
 385          * held, otherwise another thread could write the node changing the
 386          * amount of space available:
 387          */
 388
 389         prefetch(&trans->c->journal.flags);
 390
 391         trans_for_each_update2(trans, i) {
 392                 /* Multiple inserts might go to same leaf: */
 393                 if (!same_leaf_as_prev(trans, i))
 394                         u64s = 0;
 395
 396                 u64s += i->k->k.u64s;
 397                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 398                         ? btree_key_can_insert(trans, i->iter, u64s)
 399                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 400                 if (ret) {
 401                         *stopped_at = i;
 402                         return ret;
 403                 }
 404
 405                 if (btree_node_type_needs_gc(i->iter->btree_id))
 406                         marking = true;
 407         }
 408
 409         if (marking) {
 410                 percpu_down_read(&c->mark_lock);
 411                 fs_usage = bch2_fs_usage_scratch_get(c);
 412         }
 413
 414         /*
 415          * Don't get journal reservation until after we know insert will
 416          * succeed:
 417          */
 418         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 419                 ret = bch2_trans_journal_res_get(trans,
 420                                 JOURNAL_RES_GET_NONBLOCK);
 421                 if (ret)
 422                         goto err;
 423         } else {
 424                 trans->journal_res.seq = c->journal.replay_journal_seq;
 425         }
 426
 427         if (unlikely(trans->extra_journal_entry_u64s)) {
 428                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 429                                   trans->extra_journal_entries,
 430                                   trans->extra_journal_entry_u64s);
 431
 432                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 433                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 434         }
 435
 436         /*
 437          * Not allowed to fail after we've gotten our journal reservation - we
 438          * have to use it:
 439          */
 440
 441         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 442                 if (journal_seq_verify(c))
 443                         trans_for_each_update2(trans, i)
 444                                 i->k->k.version.lo = trans->journal_res.seq;
 445                 else if (inject_invalid_keys(c))
 446                         trans_for_each_update2(trans, i)
 447                                 i->k->k.version = MAX_VERSION;
 448         }
 449
 450         /* Must be called under mark_lock: */
 451         if (marking && trans->fs_usage_deltas &&
 452             bch2_replicas_delta_list_apply(c, fs_usage,
 453                                            trans->fs_usage_deltas)) {
 454                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 455                 goto err;
 456         }
 457
 458         trans_for_each_update(trans, i)
 459                 if (iter_has_nontrans_triggers(i->iter))
 460                         bch2_mark_update(trans, i->iter, i->k,
 461                                          fs_usage, i->trigger_flags);
 462
 463         if (marking)
 464                 bch2_trans_fs_usage_apply(trans, fs_usage);
 465
 466         if (unlikely(c->gc_pos.phase))
 467                 bch2_trans_mark_gc(trans);
 468
 469         trans_for_each_update2(trans, i)
 470                 do_btree_insert_one(trans, i->iter, i->k);
 471 err:
 472         if (marking) {
 473                 bch2_fs_usage_scratch_put(c, fs_usage);
 474                 percpu_up_read(&c->mark_lock);
 475         }
 476
 477         return ret;
 478 }
 479
 480 /*
 481  * Get journal reservation, take write locks, and attempt to do btree update(s):
 482  */
 483 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 484                                        struct btree_insert_entry **stopped_at)
 485 {
 486         struct btree_insert_entry *i;
 487         struct btree_iter *iter;
 488         int ret;
 489
 490         trans_for_each_update2(trans, i)
 491                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 492
 493         ret = bch2_journal_preres_get(&trans->c->journal,
 494                         &trans->journal_preres, trans->journal_preres_u64s,
 495                         JOURNAL_RES_GET_NONBLOCK|
 496                         ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
 497                          ? JOURNAL_RES_GET_RECLAIM : 0));
 498         if (unlikely(ret == -EAGAIN))
 499                 ret = bch2_trans_journal_preres_get_cold(trans,
 500                                                 trans->journal_preres_u64s);
 501         if (unlikely(ret))
 502                 return ret;
 503
 504         /*
 505          * Can't be holding any read locks when we go to take write locks:
 506          *
 507          * note - this must be done after bch2_trans_journal_preres_get_cold()
 508          * or anything else that might call bch2_trans_relock(), since that
 509          * would just retake the read locks:
 510          */
 511         trans_for_each_iter(trans, iter) {
 512                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 513                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 514                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 515                         bch2_btree_iter_unlock_noinline(iter);
 516                 }
 517         }
 518
 519         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 520                 trans_for_each_update2(trans, i)
 521                         btree_insert_entry_checks(trans, i->iter, i->k);
 522         bch2_btree_trans_verify_locks(trans);
 523
 524         trans_for_each_update2(trans, i)
 525                 if (!same_leaf_as_prev(trans, i))
 526                         bch2_btree_node_lock_for_insert(trans->c,
 527                                         iter_l(i->iter)->b, i->iter);
 528
 529         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 530
 531         trans_for_each_update2(trans, i)
 532                 if (!same_leaf_as_prev(trans, i))
 533                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 534                                                              i->iter);
 535
 536         if (!ret && trans->journal_pin)
 537                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 538                                      trans->journal_pin, NULL);
 539
 540         /*
 541          * Drop journal reservation after dropping write locks, since dropping
 542          * the journal reservation may kick off a journal write:
 543          */
 544         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 545
 546         if (unlikely(ret))
 547                 return ret;
 548
 549         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 550                 trans->nounlock = true;
 551
 552         trans_for_each_update2(trans, i)
 553                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 554                     !same_leaf_as_prev(trans, i))
 555                         bch2_foreground_maybe_merge(trans->c, i->iter,
 556                                                     0, trans->flags);
 557
 558         trans->nounlock = false;
 559
 560         bch2_trans_downgrade(trans);
 561
 562         return 0;
 563 }
 564
 565 static noinline
 566 int bch2_trans_commit_error(struct btree_trans *trans,
 567                             struct btree_insert_entry *i,
 568                             int ret)
 569 {
 570         struct bch_fs *c = trans->c;
 571         unsigned flags = trans->flags;
 572
 573         /*
 574          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 575          * update; if we haven't done anything yet it doesn't apply
 576          */
 577         flags &= ~BTREE_INSERT_NOUNLOCK;
 578
 579         switch (ret) {
 580         case BTREE_INSERT_BTREE_NODE_FULL:
 581                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 582
 583                 /*
 584                  * if the split succeeded without dropping locks the insert will
 585                  * still be atomic (what the caller peeked() and is overwriting
 586                  * won't have changed)
 587                  */
 588 #if 0
 589                 /*
 590                  * XXX:
 591                  * split -> btree node merging (of parent node) might still drop
 592                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 593                  *
 594                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 595                  * will inhibit merging - but we don't have a reliable way yet
 596                  * (do we?) of checking if we dropped locks in this path
 597                  */
 598                 if (!ret)
 599                         goto retry;
 600 #endif
 601
 602                 /*
 603                  * don't care if we got ENOSPC because we told split it
 604                  * couldn't block:
 605                  */
 606                 if (!ret ||
 607                     ret == -EINTR ||
 608                     (flags & BTREE_INSERT_NOUNLOCK)) {
 609                         trace_trans_restart_btree_node_split(trans->ip);
 610                         ret = -EINTR;
 611                 }
 612                 break;
 613         case BTREE_INSERT_ENOSPC:
 614                 ret = -ENOSPC;
 615                 break;
 616         case BTREE_INSERT_NEED_MARK_REPLICAS:
 617                 bch2_trans_unlock(trans);
 618
 619                 trans_for_each_update(trans, i) {
 620                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 621                         if (ret)
 622                                 return ret;
 623                 }
 624
 625                 if (bch2_trans_relock(trans))
 626                         return 0;
 627
 628                 trace_trans_restart_mark_replicas(trans->ip);
 629                 ret = -EINTR;
 630                 break;
 631         case BTREE_INSERT_NEED_JOURNAL_RES:
 632                 bch2_trans_unlock(trans);
 633
 634                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 635                 if (ret)
 636                         return ret;
 637
 638                 if (bch2_trans_relock(trans))
 639                         return 0;
 640
 641                 trace_trans_restart_journal_res_get(trans->ip);
 642                 ret = -EINTR;
 643                 break;
 644         default:
 645                 BUG_ON(ret >= 0);
 646                 break;
 647         }
 648
 649         if (ret == -EINTR) {
 650                 int ret2 = bch2_btree_iter_traverse_all(trans);
 651
 652                 if (ret2) {
 653                         trace_trans_restart_traverse(trans->ip);
 654                         return ret2;
 655                 }
 656
 657                 trace_trans_restart_atomic(trans->ip);
 658         }
 659
 660         return ret;
 661 }
 662
 663 static noinline int
 664 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 665 {
 666         struct bch_fs *c = trans->c;
 667         int ret;
 668
 669         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 670                 return -EROFS;
 671
 672         bch2_trans_unlock(trans);
 673
 674         ret = bch2_fs_read_write_early(c);
 675         if (ret)
 676                 return ret;
 677
 678         percpu_ref_get(&c->writes);
 679         return 0;
 680 }
 681
 682 static void bch2_trans_update2(struct btree_trans *trans,
 683                                struct btree_iter *iter,
 684                                struct bkey_i *insert)
 685 {
 686         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 687                 .iter = iter, .k = insert
 688         };
 689
 690         btree_insert_entry_checks(trans, n.iter, n.k);
 691
 692         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 693
 694         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 695
 696         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 697
 698         trans_for_each_update2(trans, i) {
 699                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 700                         *i = n;
 701                         return;
 702                 }
 703
 704                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 705                         break;
 706         }
 707
 708         array_insert_item(trans->updates2, trans->nr_updates2,
 709                           i - trans->updates2, n);
 710 }
 711
 712 static int extent_update_to_keys(struct btree_trans *trans,
 713                                  struct btree_iter *orig_iter,
 714                                  struct bkey_i *insert)
 715 {
 716         struct btree_iter *iter;
 717         int ret;
 718
 719         ret = bch2_extent_can_insert(trans, orig_iter, insert);
 720         if (ret)
 721                 return ret;
 722
 723         if (bkey_deleted(&insert->k))
 724                 return 0;
 725
 726         iter = bch2_trans_copy_iter(trans, orig_iter);
 727         if (IS_ERR(iter))
 728                 return PTR_ERR(iter);
 729
 730         iter->flags |= BTREE_ITER_INTENT;
 731         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 732         bch2_trans_update2(trans, iter, insert);
 733         bch2_trans_iter_put(trans, iter);
 734         return 0;
 735 }
 736
 737 static int extent_handle_overwrites(struct btree_trans *trans,
 738                                     enum btree_id btree_id,
 739                                     struct bpos start, struct bpos end)
 740 {
 741         struct btree_iter *iter = NULL, *update_iter;
 742         struct bkey_i *update;
 743         struct bkey_s_c k;
 744         int ret = 0;
 745
 746         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 747         ret = PTR_ERR_OR_ZERO(iter);
 748         if (ret)
 749                 return ret;
 750
 751         k = bch2_btree_iter_peek_with_updates(iter);
 752
 753         while (k.k && !(ret = bkey_err(k))) {
 754                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 755                         break;
 756
 757                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 758                         update_iter = bch2_trans_copy_iter(trans, iter);
 759                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 760                                 goto err;
 761
 762                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 763                         if ((ret = PTR_ERR_OR_ZERO(update)))
 764                                 goto err;
 765
 766                         bkey_reassemble(update, k);
 767                         bch2_cut_back(start, update);
 768
 769                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 770                         bch2_trans_update2(trans, update_iter, update);
 771                         bch2_trans_iter_put(trans, update_iter);
 772                 }
 773
 774                 if (bkey_cmp(k.k->p, end) > 0) {
 775                         update_iter = bch2_trans_copy_iter(trans, iter);
 776                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 777                                 goto err;
 778
 779                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 780                         if ((ret = PTR_ERR_OR_ZERO(update)))
 781                                 goto err;
 782
 783                         bkey_reassemble(update, k);
 784                         bch2_cut_front(end, update);
 785
 786                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 787                         bch2_trans_update2(trans, update_iter, update);
 788                         bch2_trans_iter_put(trans, update_iter);
 789                 } else {
 790                         update_iter = bch2_trans_copy_iter(trans, iter);
 791                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 792                                 goto err;
 793
 794                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 795                         if ((ret = PTR_ERR_OR_ZERO(update)))
 796                                 goto err;
 797
 798                         update->k = *k.k;
 799                         set_bkey_val_u64s(&update->k, 0);
 800                         update->k.type = KEY_TYPE_deleted;
 801                         update->k.size = 0;
 802
 803                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 804                         bch2_trans_update2(trans, update_iter, update);
 805                         bch2_trans_iter_put(trans, update_iter);
 806                 }
 807
 808                 k = bch2_btree_iter_next_with_updates(iter);
 809         }
 810 err:
 811         if (!IS_ERR_OR_NULL(iter))
 812                 bch2_trans_iter_put(trans, iter);
 813         return ret;
 814 }
 815
 816 int __bch2_trans_commit(struct btree_trans *trans)
 817 {
 818         struct btree_insert_entry *i = NULL;
 819         struct btree_iter *iter;
 820         bool trans_trigger_run;
 821         unsigned u64s;
 822         int ret = 0;
 823
 824         BUG_ON(trans->need_reset);
 825
 826         if (!trans->nr_updates)
 827                 goto out_noupdates;
 828
 829         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 830                 lockdep_assert_held(&trans->c->gc_lock);
 831
 832         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 833
 834         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 835         trans->journal_preres_u64s      = 0;
 836
 837         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 838             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 839                 ret = bch2_trans_commit_get_rw_cold(trans);
 840                 if (ret)
 841                         return ret;
 842         }
 843
 844 #ifdef CONFIG_BCACHEFS_DEBUG
 845         trans_for_each_update(trans, i)
 846                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 847                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 848                         bch2_btree_key_cache_verify_clean(trans,
 849                                         i->iter->btree_id, i->iter->pos);
 850 #endif
 851
 852         /*
 853          * Running triggers will append more updates to the list of updates as
 854          * we're walking it:
 855          */
 856         do {
 857                 trans_trigger_run = false;
 858
 859                 trans_for_each_update(trans, i) {
 860                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
 861                                      (ret = bch2_btree_iter_traverse(i->iter)))) {
 862                                 trace_trans_restart_traverse(trans->ip);
 863                                 goto out;
 864                         }
 865
 866                         /*
 867                          * We're not using bch2_btree_iter_upgrade here because
 868                          * we know trans->nounlock can't be set:
 869                          */
 870                         if (unlikely(i->iter->locks_want < 1 &&
 871                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 872                                 trace_trans_restart_upgrade(trans->ip);
 873                                 ret = -EINTR;
 874                                 goto out;
 875                         }
 876
 877                         if (iter_has_trans_triggers(i->iter) &&
 878                             !i->trans_triggers_run) {
 879                                 i->trans_triggers_run = true;
 880                                 trans_trigger_run = true;
 881
 882                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 883                                                              i->trigger_flags);
 884                                 if (unlikely(ret)) {
 885                                         if (ret == -EINTR)
 886                                                 trace_trans_restart_mark(trans->ip);
 887                                         goto out;
 888                                 }
 889                         }
 890                 }
 891         } while (trans_trigger_run);
 892
 893         /* Turn extents updates into keys: */
 894         trans_for_each_update(trans, i)
 895                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 896                         struct bpos start = bkey_start_pos(&i->k->k);
 897
 898                         while (i + 1 < trans->updates + trans->nr_updates &&
 899                                i[0].iter->btree_id == i[1].iter->btree_id &&
 900                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 901                                 i++;
 902
 903                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 904                                                        start, i->k->k.p);
 905                         if (ret)
 906                                 goto out;
 907                 }
 908
 909         trans_for_each_update(trans, i) {
 910                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 911                         ret = extent_update_to_keys(trans, i->iter, i->k);
 912                         if (ret)
 913                                 goto out;
 914                 } else {
 915                         bch2_trans_update2(trans, i->iter, i->k);
 916                 }
 917         }
 918
 919         trans_for_each_update2(trans, i) {
 920                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 921                 BUG_ON(i->iter->locks_want < 1);
 922
 923                 u64s = jset_u64s(i->k->k.u64s);
 924                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 925                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 926                         trans->journal_preres_u64s += u64s;
 927                 trans->journal_u64s += u64s;
 928         }
 929 retry:
 930         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 931
 932         ret = do_bch2_trans_commit(trans, &i);
 933
 934         /* make sure we didn't drop or screw up locks: */
 935         bch2_btree_trans_verify_locks(trans);
 936
 937         if (ret)
 938                 goto err;
 939
 940         trans_for_each_iter(trans, iter)
 941                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 942                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 943                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 944                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 945                         else
 946                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 947                 }
 948 out:
 949         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 950
 951         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 952                 percpu_ref_put(&trans->c->writes);
 953 out_noupdates:
 954         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 955
 956         return ret;
 957 err:
 958         ret = bch2_trans_commit_error(trans, i, ret);
 959         if (ret)
 960                 goto out;
 961
 962         goto retry;
 963 }
 964
 965 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 966                       struct bkey_i *k, enum btree_trigger_flags flags)
 967 {
 968         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 969                 .trigger_flags = flags, .iter = iter, .k = k
 970         };
 971
 972         EBUG_ON(bkey_cmp(iter->pos,
 973                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 974                          ? bkey_start_pos(&k->k)
 975                          : k->k.p));
 976
 977         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 978
 979         if (btree_node_type_is_extents(iter->btree_id)) {
 980                 iter->pos_after_commit = k->k.p;
 981                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 982         }
 983
 984         /*
 985          * Pending updates are kept sorted: first, find position of new update:
 986          */
 987         trans_for_each_update(trans, i)
 988                 if (btree_iter_cmp(iter, i->iter) <= 0)
 989                         break;
 990
 991         /*
 992          * Now delete/trim any updates the new update overwrites:
 993          */
 994         if (i > trans->updates &&
 995             i[-1].iter->btree_id == iter->btree_id &&
 996             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 997                 bch2_cut_back(n.iter->pos, i[-1].k);
 998
 999         while (i < trans->updates + trans->nr_updates &&
1000                iter->btree_id == i->iter->btree_id &&
1001                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1002                 array_remove_item(trans->updates, trans->nr_updates,
1003                                   i - trans->updates);
1004
1005         if (i < trans->updates + trans->nr_updates &&
1006             iter->btree_id == i->iter->btree_id &&
1007             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1008                 /*
1009                  * When we have an extent that overwrites the start of another
1010                  * update, trimming that extent will mean the iterator's
1011                  * position has to change since the iterator position has to
1012                  * match the extent's start pos - but we don't want to change
1013                  * the iterator pos if some other code is using it, so we may
1014                  * need to clone it:
1015                  */
1016                 if (trans->iters_live & (1ULL << i->iter->idx)) {
1017                         i->iter = bch2_trans_copy_iter(trans, i->iter);
1018                         if (IS_ERR(i->iter)) {
1019                                 trans->need_reset = true;
1020                                 return PTR_ERR(i->iter);
1021                         }
1022
1023                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1024                         bch2_trans_iter_put(trans, i->iter);
1025                 }
1026
1027                 bch2_cut_front(n.k->k.p, i->k);
1028                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1029         }
1030
1031         EBUG_ON(trans->nr_updates >= trans->nr_iters);
1032
1033         array_insert_item(trans->updates, trans->nr_updates,
1034                           i - trans->updates, n);
1035         return 0;
1036 }
1037
1038 int __bch2_btree_insert(struct btree_trans *trans,
1039                         enum btree_id id, struct bkey_i *k)
1040 {
1041         struct btree_iter *iter;
1042         int ret;
1043
1044         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1045                                    BTREE_ITER_INTENT);
1046         if (IS_ERR(iter))
1047                 return PTR_ERR(iter);
1048
1049         ret   = bch2_btree_iter_traverse(iter) ?:
1050                 bch2_trans_update(trans, iter, k, 0);
1051         bch2_trans_iter_put(trans, iter);
1052         return ret;
1053 }
1054
1055 /**
1056  * bch2_btree_insert - insert keys into the extent btree
1057  * @c:                  pointer to struct bch_fs
1058  * @id:                 btree to insert into
1059  * @insert_keys:        list of keys to insert
1060  * @hook:               insert callback
1061  */
1062 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1063                       struct bkey_i *k,
1064                       struct disk_reservation *disk_res,
1065                       u64 *journal_seq, int flags)
1066 {
1067         return bch2_trans_do(c, disk_res, journal_seq, flags,
1068                              __bch2_btree_insert(&trans, id, k));
1069 }
1070
1071 int bch2_btree_delete_at_range(struct btree_trans *trans,
1072                                struct btree_iter *iter,
1073                                struct bpos end,
1074                                u64 *journal_seq)
1075 {
1076         struct bkey_s_c k;
1077         int ret = 0;
1078 retry:
1079         while ((k = bch2_btree_iter_peek(iter)).k &&
1080                !(ret = bkey_err(k)) &&
1081                bkey_cmp(iter->pos, end) < 0) {
1082                 struct bkey_i delete;
1083
1084                 bch2_trans_begin(trans);
1085
1086                 bkey_init(&delete.k);
1087
1088                 /*
1089                  * For extents, iter.pos won't necessarily be the same as
1090                  * bkey_start_pos(k.k) (for non extents they always will be the
1091                  * same). It's important that we delete starting from iter.pos
1092                  * because the range we want to delete could start in the middle
1093                  * of k.
1094                  *
1095                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1096                  * bkey_start_pos(k.k)).
1097                  */
1098                 delete.k.p = iter->pos;
1099
1100                 if (btree_node_type_is_extents(iter->btree_id)) {
1101                         unsigned max_sectors =
1102                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1103
1104                         /* create the biggest key we can */
1105                         bch2_key_resize(&delete.k, max_sectors);
1106                         bch2_cut_back(end, &delete);
1107
1108                         ret = bch2_extent_trim_atomic(&delete, iter);
1109                         if (ret)
1110                                 break;
1111                 }
1112
1113                 bch2_trans_update(trans, iter, &delete, 0);
1114                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1115                                         BTREE_INSERT_NOFAIL);
1116                 if (ret)
1117                         break;
1118
1119                 bch2_trans_cond_resched(trans);
1120         }
1121
1122         if (ret == -EINTR) {
1123                 ret = 0;
1124                 goto retry;
1125         }
1126
1127         return ret;
1128
1129 }
1130
1131 int bch2_btree_delete_at(struct btree_trans *trans,
1132                          struct btree_iter *iter, unsigned flags)
1133 {
1134         struct bkey_i k;
1135
1136         bkey_init(&k.k);
1137         k.k.p = iter->pos;
1138
1139         bch2_trans_update(trans, iter, &k, 0);
1140         return bch2_trans_commit(trans, NULL, NULL,
1141                                  BTREE_INSERT_NOFAIL|
1142                                  BTREE_INSERT_USE_RESERVE|flags);
1143 }
1144
1145 /*
1146  * bch_btree_delete_range - delete everything within a given range
1147  *
1148  * Range is a half open interval - [start, end)
1149  */
1150 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1151                             struct bpos start, struct bpos end,
1152                             u64 *journal_seq)
1153 {
1154         struct btree_trans trans;
1155         struct btree_iter *iter;
1156         int ret = 0;
1157
1158         /*
1159          * XXX: whether we need mem/more iters depends on whether this btree id
1160          * has triggers
1161          */
1162         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1163
1164         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1165
1166         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1167         ret = bch2_trans_exit(&trans) ?: ret;
1168
1169         BUG_ON(ret == -EINTR);
1170         return ret;
1171 }