git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_leaf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "buckets.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "replicas.h"
  19
  20 #include <linux/prefetch.h>
  21 #include <linux/sort.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  25                                      struct btree_insert_entry *i)
  26 {
  27         return i != trans->updates2 &&
  28                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  29 }
  30
  31 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  32                                             struct btree_iter *iter)
  33 {
  34         bch2_btree_node_lock_write(b, iter);
  35
  36         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
  37                 return;
  38
  39         if (unlikely(btree_node_just_written(b)) &&
  40             bch2_btree_post_write_cleanup(c, b))
  41                 bch2_btree_iter_reinit_node(iter, b);
  42
  43         /*
  44          * If the last bset has been written, or if it's gotten too big - start
  45          * a new bset to insert into:
  46          */
  47         if (want_new_bset(c, b))
  48                 bch2_btree_init_next(c, b, iter);
  49 }
  50
  51 /* Inserting into a given leaf node (last stage of insert): */
  52
  53 /* Handle overwrites and do insert, for non extents: */
  54 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  55                                 struct btree *b,
  56                                 struct btree_node_iter *node_iter,
  57                                 struct bkey_i *insert)
  58 {
  59         struct bkey_packed *k;
  60         unsigned clobber_u64s = 0, new_u64s = 0;
  61
  62         EBUG_ON(btree_node_just_written(b));
  63         EBUG_ON(bset_written(b, btree_bset_last(b)));
  64         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  65         EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
  66                 bkey_cmp(bkey_start_pos(&insert->k),
  67                          bkey_predecessor(b->data->min_key)) < 0);
  68         EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
  69         EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
  70         EBUG_ON(insert->k.u64s >
  71                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  72         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  73
  74         k = bch2_btree_node_iter_peek_all(node_iter, b);
  75         if (k && bkey_cmp_packed(b, k, &insert->k))
  76                 k = NULL;
  77
  78         /* @k is the key being overwritten/deleted, if any: */
  79         EBUG_ON(k && bkey_whiteout(k));
  80
  81         /* Deleting, but not found? nothing to do: */
  82         if (bkey_whiteout(&insert->k) && !k)
  83                 return false;
  84
  85         if (bkey_whiteout(&insert->k)) {
  86                 /* Deleting: */
  87                 btree_account_key_drop(b, k);
  88                 k->type = KEY_TYPE_deleted;
  89
  90                 if (k->needs_whiteout)
  91                         push_whiteout(iter->trans->c, b, insert->k.p);
  92                 k->needs_whiteout = false;
  93
  94                 if (k >= btree_bset_last(b)->start) {
  95                         clobber_u64s = k->u64s;
  96                         bch2_bset_delete(b, k, clobber_u64s);
  97                         goto fix_iter;
  98                 } else {
  99                         bch2_btree_iter_fix_key_modified(iter, b, k);
 100                 }
 101
 102                 return true;
 103         }
 104
 105         if (k) {
 106                 /* Overwriting: */
 107                 btree_account_key_drop(b, k);
 108                 k->type = KEY_TYPE_deleted;
 109
 110                 insert->k.needs_whiteout = k->needs_whiteout;
 111                 k->needs_whiteout = false;
 112
 113                 if (k >= btree_bset_last(b)->start) {
 114                         clobber_u64s = k->u64s;
 115                         goto overwrite;
 116                 } else {
 117                         bch2_btree_iter_fix_key_modified(iter, b, k);
 118                 }
 119         }
 120
 121         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 122 overwrite:
 123         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 124         new_u64s = k->u64s;
 125 fix_iter:
 126         if (clobber_u64s != new_u64s)
 127                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 128                                          clobber_u64s, new_u64s);
 129         return true;
 130 }
 131
 132 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 133                                unsigned i, u64 seq)
 134 {
 135         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 136         struct btree_write *w = container_of(pin, struct btree_write, journal);
 137         struct btree *b = container_of(w, struct btree, writes[i]);
 138
 139         btree_node_lock_type(c, b, SIX_LOCK_read);
 140         bch2_btree_node_write_cond(c, b,
 141                 (btree_current_write(b) == w && w->journal.seq == seq));
 142         six_unlock_read(&b->c.lock);
 143 }
 144
 145 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 146 {
 147         return __btree_node_flush(j, pin, 0, seq);
 148 }
 149
 150 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 151 {
 152         return __btree_node_flush(j, pin, 1, seq);
 153 }
 154
 155 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 156                                        struct btree *b, u64 seq)
 157 {
 158         struct btree_write *w = btree_current_write(b);
 159
 160         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 161                              btree_node_write_idx(b) == 0
 162                              ? btree_node_flush0
 163                              : btree_node_flush1);
 164 }
 165
 166 /**
 167  * btree_insert_key - insert a key one key into a leaf node
 168  */
 169 static bool btree_insert_key_leaf(struct btree_trans *trans,
 170                                   struct btree_iter *iter,
 171                                   struct bkey_i *insert)
 172 {
 173         struct bch_fs *c = trans->c;
 174         struct btree *b = iter_l(iter)->b;
 175         struct bset_tree *t = bset_tree_last(b);
 176         struct bset *i = bset(b, t);
 177         int old_u64s = bset_u64s(t);
 178         int old_live_u64s = b->nr.live_u64s;
 179         int live_u64s_added, u64s_added;
 180
 181         EBUG_ON(!iter->level &&
 182                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 183
 184         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 185                                         &iter_l(iter)->iter, insert)))
 186                 return false;
 187
 188         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 189                                          le64_to_cpu(i->journal_seq)));
 190
 191         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 192
 193         if (unlikely(!btree_node_dirty(b)))
 194                 set_btree_node_dirty(b);
 195
 196         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 197         u64s_added = (int) bset_u64s(t) - old_u64s;
 198
 199         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 200                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 201         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 202                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 203
 204         if (u64s_added > live_u64s_added &&
 205             bch2_maybe_compact_whiteouts(c, b))
 206                 bch2_btree_iter_reinit_node(iter, b);
 207
 208         trace_btree_insert_key(c, b, insert);
 209         return true;
 210 }
 211
 212 /* Cached btree updates: */
 213
 214 /* Normal update interface: */
 215
 216 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 217                                              struct btree_iter *iter,
 218                                              struct bkey_i *insert)
 219 {
 220         struct bch_fs *c = trans->c;
 221
 222         BUG_ON(bkey_cmp(insert->k.p, iter->pos));
 223         BUG_ON(debug_check_bkeys(c) &&
 224                bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
 225                                  __btree_node_type(iter->level, iter->btree_id)));
 226 }
 227
 228 static noinline int
 229 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 230 {
 231         struct bch_fs *c = trans->c;
 232         int ret;
 233
 234         bch2_trans_unlock(trans);
 235
 236         ret = bch2_journal_preres_get(&c->journal,
 237                         &trans->journal_preres, u64s, 0);
 238         if (ret)
 239                 return ret;
 240
 241         if (!bch2_trans_relock(trans)) {
 242                 trace_trans_restart_journal_preres_get(trans->ip);
 243                 return -EINTR;
 244         }
 245
 246         return 0;
 247 }
 248
 249 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 250                                              unsigned flags)
 251 {
 252         struct bch_fs *c = trans->c;
 253         int ret;
 254
 255         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 256                 flags |= JOURNAL_RES_GET_RESERVED;
 257
 258         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 259                                    trans->journal_u64s, flags);
 260
 261         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 262 }
 263
 264 static enum btree_insert_ret
 265 btree_key_can_insert(struct btree_trans *trans,
 266                      struct btree_iter *iter,
 267                      struct bkey_i *insert,
 268                      unsigned *u64s)
 269 {
 270         struct bch_fs *c = trans->c;
 271         struct btree *b = iter_l(iter)->b;
 272         static enum btree_insert_ret ret;
 273
 274         if (unlikely(btree_node_fake(b)))
 275                 return BTREE_INSERT_BTREE_NODE_FULL;
 276
 277         /*
 278          * old bch2_extent_sort_fix_overlapping() algorithm won't work with new
 279          * style extent updates:
 280          */
 281         if (unlikely(btree_node_old_extent_overwrite(b)))
 282                 return BTREE_INSERT_BTREE_NODE_FULL;
 283
 284         ret = !(iter->flags & BTREE_ITER_IS_EXTENTS)
 285                 ? BTREE_INSERT_OK
 286                 : bch2_extent_can_insert(trans, iter, insert);
 287         if (ret)
 288                 return ret;
 289
 290         if (*u64s > bch_btree_keys_u64s_remaining(c, b))
 291                 return BTREE_INSERT_BTREE_NODE_FULL;
 292
 293         return BTREE_INSERT_OK;
 294 }
 295
 296 static enum btree_insert_ret
 297 btree_key_can_insert_cached(struct btree_trans *trans,
 298                             struct btree_iter *iter,
 299                             struct bkey_i *insert,
 300                             unsigned *u64s)
 301 {
 302         struct bkey_cached *ck = (void *) iter->l[0].b;
 303         unsigned new_u64s;
 304         struct bkey_i *new_k;
 305
 306         BUG_ON(iter->level);
 307
 308         if (*u64s <= ck->u64s)
 309                 return BTREE_INSERT_OK;
 310
 311         new_u64s        = roundup_pow_of_two(*u64s);
 312         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 313         if (!new_k)
 314                 return -ENOMEM;
 315
 316         ck->u64s        = new_u64s;
 317         ck->k           = new_k;
 318         return BTREE_INSERT_OK;
 319 }
 320
 321 static inline void do_btree_insert_one(struct btree_trans *trans,
 322                                        struct btree_iter *iter,
 323                                        struct bkey_i *insert)
 324 {
 325         struct bch_fs *c = trans->c;
 326         struct journal *j = &c->journal;
 327         bool did_work;
 328
 329         EBUG_ON(trans->journal_res.ref !=
 330                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 331
 332         insert->k.needs_whiteout = false;
 333
 334         did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
 335                 ? btree_insert_key_leaf(trans, iter, insert)
 336                 : bch2_btree_insert_key_cached(trans, iter, insert);
 337         if (!did_work)
 338                 return;
 339
 340         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 341                 bch2_journal_add_keys(j, &trans->journal_res,
 342                                       iter->btree_id, insert);
 343
 344                 bch2_journal_set_has_inode(j, &trans->journal_res,
 345                                            insert->k.p.inode);
 346
 347                 if (trans->journal_seq)
 348                         *trans->journal_seq = trans->journal_res.seq;
 349         }
 350 }
 351
 352 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 353 {
 354         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 355 }
 356
 357 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 358 {
 359         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 360                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 361                 (1U << iter->btree_id);
 362 }
 363
 364 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 365 {
 366         __bch2_btree_iter_unlock(iter);
 367 }
 368
 369 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 370 {
 371         struct bch_fs *c = trans->c;
 372         struct btree_insert_entry *i;
 373
 374         trans_for_each_update(trans, i) {
 375                 /*
 376                  * XXX: synchronization of cached update triggers with gc
 377                  */
 378                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 379
 380                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 381                         bch2_mark_update(trans, i->iter, i->k, NULL,
 382                                          i->trigger_flags|BTREE_TRIGGER_GC);
 383         }
 384 }
 385
 386 static inline int
 387 bch2_trans_commit_write_locked(struct btree_trans *trans,
 388                                struct btree_insert_entry **stopped_at)
 389 {
 390         struct bch_fs *c = trans->c;
 391         struct bch_fs_usage *fs_usage = NULL;
 392         struct btree_insert_entry *i;
 393         unsigned u64s = 0;
 394         bool marking = false;
 395         int ret;
 396
 397         if (race_fault()) {
 398                 trace_trans_restart_fault_inject(trans->ip);
 399                 return -EINTR;
 400         }
 401
 402         /*
 403          * Check if the insert will fit in the leaf node with the write lock
 404          * held, otherwise another thread could write the node changing the
 405          * amount of space available:
 406          */
 407
 408         prefetch(&trans->c->journal.flags);
 409
 410         trans_for_each_update2(trans, i) {
 411                 /* Multiple inserts might go to same leaf: */
 412                 if (!same_leaf_as_prev(trans, i))
 413                         u64s = 0;
 414
 415                 u64s += i->k->k.u64s;
 416                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 417                         ? btree_key_can_insert(trans, i->iter, i->k, &u64s)
 418                         : btree_key_can_insert_cached(trans, i->iter, i->k, &u64s);
 419                 if (ret) {
 420                         *stopped_at = i;
 421                         return ret;
 422                 }
 423
 424                 if (btree_node_type_needs_gc(i->iter->btree_id))
 425                         marking = true;
 426         }
 427
 428         if (marking) {
 429                 percpu_down_read(&c->mark_lock);
 430                 fs_usage = bch2_fs_usage_scratch_get(c);
 431         }
 432
 433         /*
 434          * Don't get journal reservation until after we know insert will
 435          * succeed:
 436          */
 437         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 438                 ret = bch2_trans_journal_res_get(trans,
 439                                 JOURNAL_RES_GET_NONBLOCK);
 440                 if (ret)
 441                         goto err;
 442         } else {
 443                 trans->journal_res.seq = c->journal.replay_journal_seq;
 444         }
 445
 446         if (unlikely(trans->extra_journal_entry_u64s)) {
 447                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 448                                   trans->extra_journal_entries,
 449                                   trans->extra_journal_entry_u64s);
 450
 451                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 452                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 453         }
 454
 455         /*
 456          * Not allowed to fail after we've gotten our journal reservation - we
 457          * have to use it:
 458          */
 459
 460         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 461                 if (journal_seq_verify(c))
 462                         trans_for_each_update2(trans, i)
 463                                 i->k->k.version.lo = trans->journal_res.seq;
 464                 else if (inject_invalid_keys(c))
 465                         trans_for_each_update2(trans, i)
 466                                 i->k->k.version = MAX_VERSION;
 467         }
 468
 469         /* Must be called under mark_lock: */
 470         if (marking && trans->fs_usage_deltas &&
 471             bch2_replicas_delta_list_apply(c, fs_usage,
 472                                            trans->fs_usage_deltas)) {
 473                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 474                 goto err;
 475         }
 476
 477         trans_for_each_update(trans, i)
 478                 if (iter_has_nontrans_triggers(i->iter))
 479                         bch2_mark_update(trans, i->iter, i->k,
 480                                          fs_usage, i->trigger_flags);
 481
 482         if (marking)
 483                 bch2_trans_fs_usage_apply(trans, fs_usage);
 484
 485         if (unlikely(c->gc_pos.phase))
 486                 bch2_trans_mark_gc(trans);
 487
 488         trans_for_each_update2(trans, i)
 489                 do_btree_insert_one(trans, i->iter, i->k);
 490 err:
 491         if (marking) {
 492                 bch2_fs_usage_scratch_put(c, fs_usage);
 493                 percpu_up_read(&c->mark_lock);
 494         }
 495
 496         return ret;
 497 }
 498
 499 /*
 500  * Get journal reservation, take write locks, and attempt to do btree update(s):
 501  */
 502 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 503                                        struct btree_insert_entry **stopped_at)
 504 {
 505         struct btree_insert_entry *i;
 506         struct btree_iter *iter;
 507         int ret;
 508
 509         trans_for_each_update2(trans, i)
 510                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 511
 512         ret = bch2_journal_preres_get(&trans->c->journal,
 513                         &trans->journal_preres, trans->journal_preres_u64s,
 514                         JOURNAL_RES_GET_NONBLOCK|
 515                         ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
 516                          ? JOURNAL_RES_GET_RECLAIM : 0));
 517         if (unlikely(ret == -EAGAIN))
 518                 ret = bch2_trans_journal_preres_get_cold(trans,
 519                                                 trans->journal_preres_u64s);
 520         if (unlikely(ret))
 521                 return ret;
 522
 523         /*
 524          * Can't be holding any read locks when we go to take write locks:
 525          *
 526          * note - this must be done after bch2_trans_journal_preres_get_cold()
 527          * or anything else that might call bch2_trans_relock(), since that
 528          * would just retake the read locks:
 529          */
 530         trans_for_each_iter(trans, iter) {
 531                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 532                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 533                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 534                         bch2_btree_iter_unlock_noinline(iter);
 535                 }
 536         }
 537
 538         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 539                 trans_for_each_update2(trans, i)
 540                         btree_insert_entry_checks(trans, i->iter, i->k);
 541         bch2_btree_trans_verify_locks(trans);
 542
 543         trans_for_each_update2(trans, i)
 544                 if (!same_leaf_as_prev(trans, i))
 545                         bch2_btree_node_lock_for_insert(trans->c,
 546                                         iter_l(i->iter)->b, i->iter);
 547
 548         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 549
 550         trans_for_each_update2(trans, i)
 551                 if (!same_leaf_as_prev(trans, i))
 552                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 553                                                              i->iter);
 554
 555         if (!ret && trans->journal_pin)
 556                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 557                                      trans->journal_pin, NULL);
 558
 559         /*
 560          * Drop journal reservation after dropping write locks, since dropping
 561          * the journal reservation may kick off a journal write:
 562          */
 563         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 564
 565         if (unlikely(ret))
 566                 return ret;
 567
 568         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 569                 trans->nounlock = true;
 570
 571         trans_for_each_update2(trans, i)
 572                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 573                     !same_leaf_as_prev(trans, i))
 574                         bch2_foreground_maybe_merge(trans->c, i->iter,
 575                                                     0, trans->flags);
 576
 577         trans->nounlock = false;
 578
 579         bch2_trans_downgrade(trans);
 580
 581         return 0;
 582 }
 583
 584 static noinline
 585 int bch2_trans_commit_error(struct btree_trans *trans,
 586                             struct btree_insert_entry *i,
 587                             int ret)
 588 {
 589         struct bch_fs *c = trans->c;
 590         unsigned flags = trans->flags;
 591
 592         /*
 593          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 594          * update; if we haven't done anything yet it doesn't apply
 595          */
 596         flags &= ~BTREE_INSERT_NOUNLOCK;
 597
 598         switch (ret) {
 599         case BTREE_INSERT_BTREE_NODE_FULL:
 600                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 601
 602                 /*
 603                  * if the split succeeded without dropping locks the insert will
 604                  * still be atomic (what the caller peeked() and is overwriting
 605                  * won't have changed)
 606                  */
 607 #if 0
 608                 /*
 609                  * XXX:
 610                  * split -> btree node merging (of parent node) might still drop
 611                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 612                  *
 613                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 614                  * will inhibit merging - but we don't have a reliable way yet
 615                  * (do we?) of checking if we dropped locks in this path
 616                  */
 617                 if (!ret)
 618                         goto retry;
 619 #endif
 620
 621                 /*
 622                  * don't care if we got ENOSPC because we told split it
 623                  * couldn't block:
 624                  */
 625                 if (!ret ||
 626                     ret == -EINTR ||
 627                     (flags & BTREE_INSERT_NOUNLOCK)) {
 628                         trace_trans_restart_btree_node_split(trans->ip);
 629                         ret = -EINTR;
 630                 }
 631                 break;
 632         case BTREE_INSERT_ENOSPC:
 633                 ret = -ENOSPC;
 634                 break;
 635         case BTREE_INSERT_NEED_MARK_REPLICAS:
 636                 bch2_trans_unlock(trans);
 637
 638                 trans_for_each_update(trans, i) {
 639                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 640                         if (ret)
 641                                 return ret;
 642                 }
 643
 644                 if (bch2_trans_relock(trans))
 645                         return 0;
 646
 647                 trace_trans_restart_mark_replicas(trans->ip);
 648                 ret = -EINTR;
 649                 break;
 650         case BTREE_INSERT_NEED_JOURNAL_RES:
 651                 bch2_trans_unlock(trans);
 652
 653                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 654                 if (ret)
 655                         return ret;
 656
 657                 if (bch2_trans_relock(trans))
 658                         return 0;
 659
 660                 trace_trans_restart_journal_res_get(trans->ip);
 661                 ret = -EINTR;
 662                 break;
 663         default:
 664                 BUG_ON(ret >= 0);
 665                 break;
 666         }
 667
 668         if (ret == -EINTR) {
 669                 int ret2 = bch2_btree_iter_traverse_all(trans);
 670
 671                 if (ret2) {
 672                         trace_trans_restart_traverse(trans->ip);
 673                         return ret2;
 674                 }
 675
 676                 trace_trans_restart_atomic(trans->ip);
 677         }
 678
 679         return ret;
 680 }
 681
 682 static noinline int
 683 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 684 {
 685         struct bch_fs *c = trans->c;
 686         int ret;
 687
 688         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 689                 return -EROFS;
 690
 691         bch2_trans_unlock(trans);
 692
 693         ret = bch2_fs_read_write_early(c);
 694         if (ret)
 695                 return ret;
 696
 697         percpu_ref_get(&c->writes);
 698         return 0;
 699 }
 700
 701 static void bch2_trans_update2(struct btree_trans *trans,
 702                                struct btree_iter *iter,
 703                                struct bkey_i *insert)
 704 {
 705         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 706                 .iter = iter, .k = insert
 707         };
 708
 709         btree_insert_entry_checks(trans, n.iter, n.k);
 710
 711         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 712
 713         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 714
 715         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 716
 717         trans_for_each_update2(trans, i) {
 718                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 719                         *i = n;
 720                         return;
 721                 }
 722
 723                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 724                         break;
 725         }
 726
 727         array_insert_item(trans->updates2, trans->nr_updates2,
 728                           i - trans->updates2, n);
 729 }
 730
 731 static int extent_update_to_keys(struct btree_trans *trans,
 732                                  struct btree_iter *orig_iter,
 733                                  struct bkey_i *insert)
 734 {
 735         struct btree_iter *iter;
 736
 737         if (bkey_deleted(&insert->k))
 738                 return 0;
 739
 740         iter = bch2_trans_copy_iter(trans, orig_iter);
 741         if (IS_ERR(iter))
 742                 return PTR_ERR(iter);
 743
 744         iter->flags |= BTREE_ITER_INTENT;
 745         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 746         bch2_trans_update2(trans, iter, insert);
 747         bch2_trans_iter_put(trans, iter);
 748         return 0;
 749 }
 750
 751 static int extent_handle_overwrites(struct btree_trans *trans,
 752                                     enum btree_id btree_id,
 753                                     struct bpos start, struct bpos end)
 754 {
 755         struct btree_iter *iter = NULL, *update_iter;
 756         struct bkey_i *update;
 757         struct bkey_s_c k;
 758         int ret = 0;
 759
 760         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 761         ret = PTR_ERR_OR_ZERO(iter);
 762         if (ret)
 763                 return ret;
 764
 765         k = bch2_btree_iter_peek_with_updates(iter);
 766
 767         while (k.k && !(ret = bkey_err(k))) {
 768                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 769                         break;
 770
 771                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 772                         update_iter = bch2_trans_copy_iter(trans, iter);
 773                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 774                                 goto err;
 775
 776                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 777                         if ((ret = PTR_ERR_OR_ZERO(update)))
 778                                 goto err;
 779
 780                         bkey_reassemble(update, k);
 781                         bch2_cut_back(start, update);
 782
 783                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 784                         bch2_trans_update2(trans, update_iter, update);
 785                         bch2_trans_iter_put(trans, update_iter);
 786                 }
 787
 788                 if (bkey_cmp(k.k->p, end) > 0) {
 789                         update_iter = bch2_trans_copy_iter(trans, iter);
 790                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 791                                 goto err;
 792
 793                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 794                         if ((ret = PTR_ERR_OR_ZERO(update)))
 795                                 goto err;
 796
 797                         bkey_reassemble(update, k);
 798                         bch2_cut_front(end, update);
 799
 800                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 801                         bch2_trans_update2(trans, update_iter, update);
 802                         bch2_trans_iter_put(trans, update_iter);
 803                 } else {
 804                         update_iter = bch2_trans_copy_iter(trans, iter);
 805                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 806                                 goto err;
 807
 808                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 809                         if ((ret = PTR_ERR_OR_ZERO(update)))
 810                                 goto err;
 811
 812                         update->k = *k.k;
 813                         set_bkey_val_u64s(&update->k, 0);
 814                         update->k.type = KEY_TYPE_deleted;
 815                         update->k.size = 0;
 816
 817                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 818                         bch2_trans_update2(trans, update_iter, update);
 819                         bch2_trans_iter_put(trans, update_iter);
 820                 }
 821
 822                 k = bch2_btree_iter_next_with_updates(iter);
 823         }
 824 err:
 825         if (!IS_ERR_OR_NULL(iter))
 826                 bch2_trans_iter_put(trans, iter);
 827         return ret;
 828 }
 829
 830 int __bch2_trans_commit(struct btree_trans *trans)
 831 {
 832         struct btree_insert_entry *i = NULL;
 833         struct btree_iter *iter;
 834         bool trans_trigger_run;
 835         unsigned u64s;
 836         int ret = 0;
 837
 838         BUG_ON(trans->need_reset);
 839
 840         if (!trans->nr_updates)
 841                 goto out_noupdates;
 842
 843         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 844                 lockdep_assert_held(&trans->c->gc_lock);
 845
 846         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 847
 848         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 849         trans->journal_preres_u64s      = 0;
 850
 851         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 852             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 853                 ret = bch2_trans_commit_get_rw_cold(trans);
 854                 if (ret)
 855                         return ret;
 856         }
 857
 858 #ifdef CONFIG_BCACHEFS_DEBUG
 859         trans_for_each_update(trans, i)
 860                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 861                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 862                         bch2_btree_key_cache_verify_clean(trans,
 863                                         i->iter->btree_id, i->iter->pos);
 864 #endif
 865
 866         /*
 867          * Running triggers will append more updates to the list of updates as
 868          * we're walking it:
 869          */
 870         do {
 871                 trans_trigger_run = false;
 872
 873                 trans_for_each_update(trans, i) {
 874                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK &&
 875                                      (ret = bch2_btree_iter_traverse(i->iter)))) {
 876                                 trace_trans_restart_traverse(trans->ip);
 877                                 goto out;
 878                         }
 879
 880                         /*
 881                          * We're not using bch2_btree_iter_upgrade here because
 882                          * we know trans->nounlock can't be set:
 883                          */
 884                         if (unlikely(i->iter->locks_want < 1 &&
 885                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 886                                 trace_trans_restart_upgrade(trans->ip);
 887                                 ret = -EINTR;
 888                                 goto out;
 889                         }
 890
 891                         if (iter_has_trans_triggers(i->iter) &&
 892                             !i->trans_triggers_run) {
 893                                 i->trans_triggers_run = true;
 894                                 trans_trigger_run = true;
 895
 896                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 897                                                              i->trigger_flags);
 898                                 if (unlikely(ret)) {
 899                                         if (ret == -EINTR)
 900                                                 trace_trans_restart_mark(trans->ip);
 901                                         goto out;
 902                                 }
 903                         }
 904                 }
 905         } while (trans_trigger_run);
 906
 907         /* Turn extents updates into keys: */
 908         trans_for_each_update(trans, i)
 909                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 910                         struct bpos start = bkey_start_pos(&i->k->k);
 911
 912                         while (i + 1 < trans->updates + trans->nr_updates &&
 913                                i[0].iter->btree_id == i[1].iter->btree_id &&
 914                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 915                                 i++;
 916
 917                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 918                                                        start, i->k->k.p);
 919                         if (ret)
 920                                 goto out;
 921                 }
 922
 923         trans_for_each_update(trans, i) {
 924                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 925                         ret = extent_update_to_keys(trans, i->iter, i->k);
 926                         if (ret)
 927                                 goto out;
 928                 } else {
 929                         bch2_trans_update2(trans, i->iter, i->k);
 930                 }
 931         }
 932
 933         trans_for_each_update2(trans, i) {
 934                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 935                 BUG_ON(i->iter->locks_want < 1);
 936
 937                 u64s = jset_u64s(i->k->k.u64s);
 938                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 939                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 940                         trans->journal_preres_u64s += u64s;
 941                 trans->journal_u64s += u64s;
 942         }
 943 retry:
 944         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 945
 946         ret = do_bch2_trans_commit(trans, &i);
 947
 948         /* make sure we didn't drop or screw up locks: */
 949         bch2_btree_trans_verify_locks(trans);
 950
 951         if (ret)
 952                 goto err;
 953
 954         trans_for_each_iter(trans, iter)
 955                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 956                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 957                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 958                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 959                         else
 960                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 961                 }
 962 out:
 963         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 964
 965         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 966                 percpu_ref_put(&trans->c->writes);
 967 out_noupdates:
 968         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 969
 970         return ret;
 971 err:
 972         ret = bch2_trans_commit_error(trans, i, ret);
 973         if (ret)
 974                 goto out;
 975
 976         goto retry;
 977 }
 978
 979 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 980                       struct bkey_i *k, enum btree_trigger_flags flags)
 981 {
 982         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 983                 .trigger_flags = flags, .iter = iter, .k = k
 984         };
 985
 986         EBUG_ON(bkey_cmp(iter->pos,
 987                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 988                          ? bkey_start_pos(&k->k)
 989                          : k->k.p));
 990
 991         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 992
 993         if (btree_node_type_is_extents(iter->btree_id)) {
 994                 iter->pos_after_commit = k->k.p;
 995                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 996         }
 997
 998         /*
 999          * Pending updates are kept sorted: first, find position of new update:
1000          */
1001         trans_for_each_update(trans, i)
1002                 if (btree_iter_cmp(iter, i->iter) <= 0)
1003                         break;
1004
1005         /*
1006          * Now delete/trim any updates the new update overwrites:
1007          */
1008         if (i > trans->updates &&
1009             i[-1].iter->btree_id == iter->btree_id &&
1010             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
1011                 bch2_cut_back(n.iter->pos, i[-1].k);
1012
1013         while (i < trans->updates + trans->nr_updates &&
1014                iter->btree_id == i->iter->btree_id &&
1015                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
1016                 array_remove_item(trans->updates, trans->nr_updates,
1017                                   i - trans->updates);
1018
1019         if (i < trans->updates + trans->nr_updates &&
1020             iter->btree_id == i->iter->btree_id &&
1021             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
1022                 /*
1023                  * When we have an extent that overwrites the start of another
1024                  * update, trimming that extent will mean the iterator's
1025                  * position has to change since the iterator position has to
1026                  * match the extent's start pos - but we don't want to change
1027                  * the iterator pos if some other code is using it, so we may
1028                  * need to clone it:
1029                  */
1030                 if (trans->iters_live & (1ULL << i->iter->idx)) {
1031                         i->iter = bch2_trans_copy_iter(trans, i->iter);
1032                         if (IS_ERR(i->iter)) {
1033                                 trans->need_reset = true;
1034                                 return PTR_ERR(i->iter);
1035                         }
1036
1037                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1038                         bch2_trans_iter_put(trans, i->iter);
1039                 }
1040
1041                 bch2_cut_front(n.k->k.p, i->k);
1042                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1043         }
1044
1045         EBUG_ON(trans->nr_updates >= trans->nr_iters);
1046
1047         array_insert_item(trans->updates, trans->nr_updates,
1048                           i - trans->updates, n);
1049         return 0;
1050 }
1051
1052 int __bch2_btree_insert(struct btree_trans *trans,
1053                         enum btree_id id, struct bkey_i *k)
1054 {
1055         struct btree_iter *iter;
1056         int ret;
1057
1058         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1059                                    BTREE_ITER_INTENT);
1060         if (IS_ERR(iter))
1061                 return PTR_ERR(iter);
1062
1063         ret   = bch2_btree_iter_traverse(iter) ?:
1064                 bch2_trans_update(trans, iter, k, 0);
1065         bch2_trans_iter_put(trans, iter);
1066         return ret;
1067 }
1068
1069 /**
1070  * bch2_btree_insert - insert keys into the extent btree
1071  * @c:                  pointer to struct bch_fs
1072  * @id:                 btree to insert into
1073  * @insert_keys:        list of keys to insert
1074  * @hook:               insert callback
1075  */
1076 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1077                       struct bkey_i *k,
1078                       struct disk_reservation *disk_res,
1079                       u64 *journal_seq, int flags)
1080 {
1081         return bch2_trans_do(c, disk_res, journal_seq, flags,
1082                              __bch2_btree_insert(&trans, id, k));
1083 }
1084
1085 int bch2_btree_delete_at_range(struct btree_trans *trans,
1086                                struct btree_iter *iter,
1087                                struct bpos end,
1088                                u64 *journal_seq)
1089 {
1090         struct bkey_s_c k;
1091         int ret = 0;
1092 retry:
1093         while ((k = bch2_btree_iter_peek(iter)).k &&
1094                !(ret = bkey_err(k)) &&
1095                bkey_cmp(iter->pos, end) < 0) {
1096                 struct bkey_i delete;
1097
1098                 bch2_trans_begin(trans);
1099
1100                 bkey_init(&delete.k);
1101
1102                 /*
1103                  * For extents, iter.pos won't necessarily be the same as
1104                  * bkey_start_pos(k.k) (for non extents they always will be the
1105                  * same). It's important that we delete starting from iter.pos
1106                  * because the range we want to delete could start in the middle
1107                  * of k.
1108                  *
1109                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1110                  * bkey_start_pos(k.k)).
1111                  */
1112                 delete.k.p = iter->pos;
1113
1114                 if (btree_node_type_is_extents(iter->btree_id)) {
1115                         unsigned max_sectors =
1116                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1117
1118                         /* create the biggest key we can */
1119                         bch2_key_resize(&delete.k, max_sectors);
1120                         bch2_cut_back(end, &delete);
1121
1122                         ret = bch2_extent_trim_atomic(&delete, iter);
1123                         if (ret)
1124                                 break;
1125                 }
1126
1127                 bch2_trans_update(trans, iter, &delete, 0);
1128                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1129                                         BTREE_INSERT_NOFAIL);
1130                 if (ret)
1131                         break;
1132
1133                 bch2_trans_cond_resched(trans);
1134         }
1135
1136         if (ret == -EINTR) {
1137                 ret = 0;
1138                 goto retry;
1139         }
1140
1141         return ret;
1142
1143 }
1144
1145 int bch2_btree_delete_at(struct btree_trans *trans,
1146                          struct btree_iter *iter, unsigned flags)
1147 {
1148         struct bkey_i k;
1149
1150         bkey_init(&k.k);
1151         k.k.p = iter->pos;
1152
1153         bch2_trans_update(trans, iter, &k, 0);
1154         return bch2_trans_commit(trans, NULL, NULL,
1155                                  BTREE_INSERT_NOFAIL|
1156                                  BTREE_INSERT_USE_RESERVE|flags);
1157 }
1158
1159 /*
1160  * bch_btree_delete_range - delete everything within a given range
1161  *
1162  * Range is a half open interval - [start, end)
1163  */
1164 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1165                             struct bpos start, struct bpos end,
1166                             u64 *journal_seq)
1167 {
1168         struct btree_trans trans;
1169         struct btree_iter *iter;
1170         int ret = 0;
1171
1172         /*
1173          * XXX: whether we need mem/more iters depends on whether this btree id
1174          * has triggers
1175          */
1176         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1177
1178         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1179
1180         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1181         ret = bch2_trans_exit(&trans) ?: ret;
1182
1183         BUG_ON(ret == -EINTR);
1184         return ret;
1185 }