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_locking.h"
  10 #include "buckets.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extent_update.h"
  14 #include "journal.h"
  15 #include "journal_reclaim.h"
  16 #include "keylist.h"
  17 #include "replicas.h"
  18
  19 #include <linux/prefetch.h>
  20 #include <linux/sort.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  24                                      struct btree_insert_entry *i)
  25 {
  26         return i != trans->updates2 &&
  27                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  28 }
  29
  30 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  31                                             struct btree_iter *iter)
  32 {
  33         bch2_btree_node_lock_write(b, iter);
  34
  35         if (unlikely(btree_node_just_written(b)) &&
  36             bch2_btree_post_write_cleanup(c, b))
  37                 bch2_btree_iter_reinit_node(iter, b);
  38
  39         /*
  40          * If the last bset has been written, or if it's gotten too big - start
  41          * a new bset to insert into:
  42          */
  43         if (want_new_bset(c, b))
  44                 bch2_btree_init_next(c, b, iter);
  45 }
  46
  47 /* Inserting into a given leaf node (last stage of insert): */
  48
  49 /* Handle overwrites and do insert, for non extents: */
  50 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  51                                 struct btree *b,
  52                                 struct btree_node_iter *node_iter,
  53                                 struct bkey_i *insert)
  54 {
  55         struct bkey_packed *k;
  56         unsigned clobber_u64s = 0, new_u64s = 0;
  57
  58         EBUG_ON(btree_node_just_written(b));
  59         EBUG_ON(bset_written(b, btree_bset_last(b)));
  60         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  61         EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
  62                 bkey_cmp(bkey_start_pos(&insert->k),
  63                          bkey_predecessor(b->data->min_key)) < 0);
  64         EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
  65         EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
  66         EBUG_ON(insert->k.u64s >
  67                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  68         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  69
  70         k = bch2_btree_node_iter_peek_all(node_iter, b);
  71         if (k && bkey_cmp_packed(b, k, &insert->k))
  72                 k = NULL;
  73
  74         /* @k is the key being overwritten/deleted, if any: */
  75         EBUG_ON(k && bkey_whiteout(k));
  76
  77         /* Deleting, but not found? nothing to do: */
  78         if (bkey_whiteout(&insert->k) && !k)
  79                 return false;
  80
  81         if (bkey_whiteout(&insert->k)) {
  82                 /* Deleting: */
  83                 btree_account_key_drop(b, k);
  84                 k->type = KEY_TYPE_deleted;
  85
  86                 if (k->needs_whiteout)
  87                         push_whiteout(iter->trans->c, b, insert->k.p);
  88                 k->needs_whiteout = false;
  89
  90                 if (k >= btree_bset_last(b)->start) {
  91                         clobber_u64s = k->u64s;
  92                         bch2_bset_delete(b, k, clobber_u64s);
  93                         goto fix_iter;
  94                 } else {
  95                         bch2_btree_iter_fix_key_modified(iter, b, k);
  96                 }
  97
  98                 return true;
  99         }
 100
 101         if (k) {
 102                 /* Overwriting: */
 103                 btree_account_key_drop(b, k);
 104                 k->type = KEY_TYPE_deleted;
 105
 106                 insert->k.needs_whiteout = k->needs_whiteout;
 107                 k->needs_whiteout = false;
 108
 109                 if (k >= btree_bset_last(b)->start) {
 110                         clobber_u64s = k->u64s;
 111                         goto overwrite;
 112                 } else {
 113                         bch2_btree_iter_fix_key_modified(iter, b, k);
 114                 }
 115         }
 116
 117         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 118 overwrite:
 119         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 120         new_u64s = k->u64s;
 121 fix_iter:
 122         if (clobber_u64s != new_u64s)
 123                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 124                                          clobber_u64s, new_u64s);
 125         return true;
 126 }
 127
 128 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 129                                unsigned i, u64 seq)
 130 {
 131         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 132         struct btree_write *w = container_of(pin, struct btree_write, journal);
 133         struct btree *b = container_of(w, struct btree, writes[i]);
 134
 135         btree_node_lock_type(c, b, SIX_LOCK_read);
 136         bch2_btree_node_write_cond(c, b,
 137                 (btree_current_write(b) == w && w->journal.seq == seq));
 138         six_unlock_read(&b->lock);
 139 }
 140
 141 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 142 {
 143         return __btree_node_flush(j, pin, 0, seq);
 144 }
 145
 146 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 147 {
 148         return __btree_node_flush(j, pin, 1, seq);
 149 }
 150
 151 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 152                                        struct btree *b, u64 seq)
 153 {
 154         struct btree_write *w = btree_current_write(b);
 155
 156         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 157                              btree_node_write_idx(b) == 0
 158                              ? btree_node_flush0
 159                              : btree_node_flush1);
 160 }
 161
 162 static inline void __btree_journal_key(struct btree_trans *trans,
 163                                        enum btree_id btree_id,
 164                                        struct bkey_i *insert)
 165 {
 166         struct journal *j = &trans->c->journal;
 167         u64 seq = trans->journal_res.seq;
 168         bool needs_whiteout = insert->k.needs_whiteout;
 169
 170         /* ick */
 171         insert->k.needs_whiteout = false;
 172         bch2_journal_add_keys(j, &trans->journal_res,
 173                               btree_id, insert);
 174         insert->k.needs_whiteout = needs_whiteout;
 175
 176         bch2_journal_set_has_inode(j, &trans->journal_res,
 177                                    insert->k.p.inode);
 178
 179         if (trans->journal_seq)
 180                 *trans->journal_seq = seq;
 181 }
 182
 183 static void bch2_btree_journal_key(struct btree_trans *trans,
 184                                    struct btree_iter *iter,
 185                                    struct bkey_i *insert)
 186 {
 187         struct bch_fs *c = trans->c;
 188         struct journal *j = &c->journal;
 189         struct btree *b = iter_l(iter)->b;
 190
 191         EBUG_ON(trans->journal_res.ref !=
 192                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 193
 194         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 195                 __btree_journal_key(trans, iter->btree_id, insert);
 196                 btree_bset_last(b)->journal_seq =
 197                         cpu_to_le64(trans->journal_res.seq);
 198         }
 199
 200         bch2_btree_add_journal_pin(c, b,
 201                 likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 202                         ? trans->journal_res.seq
 203                         : j->replay_journal_seq);
 204
 205         if (unlikely(!btree_node_dirty(b)))
 206                 set_btree_node_dirty(b);
 207 }
 208
 209 /**
 210  * btree_insert_key - insert a key one key into a leaf node
 211  */
 212 static void btree_insert_key_leaf(struct btree_trans *trans,
 213                                   struct btree_iter *iter,
 214                                   struct bkey_i *insert)
 215 {
 216         struct bch_fs *c = trans->c;
 217         struct btree *b = iter_l(iter)->b;
 218         struct bset_tree *t = bset_tree_last(b);
 219         int old_u64s = bset_u64s(t);
 220         int old_live_u64s = b->nr.live_u64s;
 221         int live_u64s_added, u64s_added;
 222
 223         insert->k.needs_whiteout = false;
 224
 225         if (likely(bch2_btree_bset_insert_key(iter, b, &iter_l(iter)->iter, insert)))
 226                 bch2_btree_journal_key(trans, iter, insert);
 227
 228         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 229         u64s_added = (int) bset_u64s(t) - old_u64s;
 230
 231         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 232                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 233         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 234                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 235
 236         if (u64s_added > live_u64s_added &&
 237             bch2_maybe_compact_whiteouts(c, b))
 238                 bch2_btree_iter_reinit_node(iter, b);
 239
 240         trace_btree_insert_key(c, b, insert);
 241 }
 242
 243 /* Normal update interface: */
 244
 245 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 246                                              struct btree_iter *iter,
 247                                              struct bkey_i *insert)
 248 {
 249         struct bch_fs *c = trans->c;
 250
 251         BUG_ON(bkey_cmp(insert->k.p, iter->pos));
 252         BUG_ON(debug_check_bkeys(c) &&
 253                bch2_bkey_invalid(c, bkey_i_to_s_c(insert),
 254                                  __btree_node_type(iter->level, iter->btree_id)));
 255 }
 256
 257 static noinline int
 258 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 259 {
 260         struct bch_fs *c = trans->c;
 261         int ret;
 262
 263         bch2_trans_unlock(trans);
 264
 265         ret = bch2_journal_preres_get(&c->journal,
 266                         &trans->journal_preres, u64s, 0);
 267         if (ret)
 268                 return ret;
 269
 270         if (!bch2_trans_relock(trans)) {
 271                 trace_trans_restart_journal_preres_get(trans->ip);
 272                 return -EINTR;
 273         }
 274
 275         return 0;
 276 }
 277
 278 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 279                                              unsigned flags)
 280 {
 281         struct bch_fs *c = trans->c;
 282         int ret;
 283
 284         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 285                 flags |= JOURNAL_RES_GET_RESERVED;
 286
 287         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 288                                    trans->journal_u64s, flags);
 289
 290         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 291 }
 292
 293 static enum btree_insert_ret
 294 btree_key_can_insert(struct btree_trans *trans,
 295                      struct btree_iter *iter,
 296                      struct bkey_i *insert,
 297                      unsigned *u64s)
 298 {
 299         struct bch_fs *c = trans->c;
 300         struct btree *b = iter_l(iter)->b;
 301         static enum btree_insert_ret ret;
 302
 303         if (unlikely(btree_node_fake(b)))
 304                 return BTREE_INSERT_BTREE_NODE_FULL;
 305
 306         /*
 307          * old bch2_extent_sort_fix_overlapping() algorithm won't work with new
 308          * style extent updates:
 309          */
 310         if (unlikely(btree_node_old_extent_overwrite(b)))
 311                 return BTREE_INSERT_BTREE_NODE_FULL;
 312
 313         ret = !(iter->flags & BTREE_ITER_IS_EXTENTS)
 314                 ? BTREE_INSERT_OK
 315                 : bch2_extent_can_insert(trans, iter, insert);
 316         if (ret)
 317                 return ret;
 318
 319         if (*u64s > bch_btree_keys_u64s_remaining(c, b))
 320                 return BTREE_INSERT_BTREE_NODE_FULL;
 321
 322         return BTREE_INSERT_OK;
 323 }
 324
 325 static inline void do_btree_insert_one(struct btree_trans *trans,
 326                                        struct btree_iter *iter,
 327                                        struct bkey_i *insert)
 328 {
 329         btree_insert_key_leaf(trans, iter, insert);
 330 }
 331
 332 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 333 {
 334         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 335 }
 336
 337 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 338 {
 339         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 340                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 341                 (1U << iter->btree_id);
 342 }
 343
 344 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 345 {
 346         __bch2_btree_iter_unlock(iter);
 347 }
 348
 349 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 350 {
 351         struct bch_fs *c = trans->c;
 352         struct btree_insert_entry *i;
 353
 354         trans_for_each_update(trans, i)
 355                 if (gc_visited(c, gc_pos_btree_node(iter_l(i->iter)->b)))
 356                         bch2_mark_update(trans, i->iter, i->k, NULL,
 357                                          i->trigger_flags|BTREE_TRIGGER_GC);
 358 }
 359
 360 static inline int
 361 bch2_trans_commit_write_locked(struct btree_trans *trans,
 362                                struct btree_insert_entry **stopped_at)
 363 {
 364         struct bch_fs *c = trans->c;
 365         struct bch_fs_usage *fs_usage = NULL;
 366         struct btree_insert_entry *i;
 367         unsigned u64s = 0;
 368         bool marking = false;
 369         int ret;
 370
 371         if (race_fault()) {
 372                 trace_trans_restart_fault_inject(trans->ip);
 373                 return -EINTR;
 374         }
 375
 376         /*
 377          * Check if the insert will fit in the leaf node with the write lock
 378          * held, otherwise another thread could write the node changing the
 379          * amount of space available:
 380          */
 381
 382         prefetch(&trans->c->journal.flags);
 383
 384         trans_for_each_update2(trans, i) {
 385                 /* Multiple inserts might go to same leaf: */
 386                 if (!same_leaf_as_prev(trans, i))
 387                         u64s = 0;
 388
 389                 u64s += i->k->k.u64s;
 390                 ret = btree_key_can_insert(trans, i->iter, i->k, &u64s);
 391                 if (ret) {
 392                         *stopped_at = i;
 393                         return ret;
 394                 }
 395
 396                 if (btree_node_type_needs_gc(i->iter->btree_id))
 397                         marking = true;
 398         }
 399
 400         if (marking) {
 401                 percpu_down_read(&c->mark_lock);
 402                 fs_usage = bch2_fs_usage_scratch_get(c);
 403         }
 404
 405         /*
 406          * Don't get journal reservation until after we know insert will
 407          * succeed:
 408          */
 409         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 410                 ret = bch2_trans_journal_res_get(trans,
 411                                 JOURNAL_RES_GET_NONBLOCK);
 412                 if (ret)
 413                         goto err;
 414         }
 415
 416         /*
 417          * Not allowed to fail after we've gotten our journal reservation - we
 418          * have to use it:
 419          */
 420
 421         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 422                 if (journal_seq_verify(c))
 423                         trans_for_each_update2(trans, i)
 424                                 i->k->k.version.lo = trans->journal_res.seq;
 425                 else if (inject_invalid_keys(c))
 426                         trans_for_each_update2(trans, i)
 427                                 i->k->k.version = MAX_VERSION;
 428         }
 429
 430         /* Must be called under mark_lock: */
 431         if (marking && trans->fs_usage_deltas &&
 432             bch2_replicas_delta_list_apply(c, fs_usage,
 433                                            trans->fs_usage_deltas)) {
 434                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 435                 goto err;
 436         }
 437
 438         trans_for_each_update(trans, i)
 439                 if (iter_has_nontrans_triggers(i->iter))
 440                         bch2_mark_update(trans, i->iter, i->k,
 441                                          fs_usage, i->trigger_flags);
 442
 443         if (marking)
 444                 bch2_trans_fs_usage_apply(trans, fs_usage);
 445
 446         if (unlikely(c->gc_pos.phase))
 447                 bch2_trans_mark_gc(trans);
 448
 449         trans_for_each_update2(trans, i)
 450                 do_btree_insert_one(trans, i->iter, i->k);
 451 err:
 452         if (marking) {
 453                 bch2_fs_usage_scratch_put(c, fs_usage);
 454                 percpu_up_read(&c->mark_lock);
 455         }
 456
 457         return ret;
 458 }
 459
 460 /*
 461  * Get journal reservation, take write locks, and attempt to do btree update(s):
 462  */
 463 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 464                                        struct btree_insert_entry **stopped_at)
 465 {
 466         struct btree_insert_entry *i;
 467         struct btree_iter *iter;
 468         int ret;
 469
 470         trans_for_each_update2(trans, i)
 471                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 472
 473         ret = bch2_journal_preres_get(&trans->c->journal,
 474                         &trans->journal_preres, trans->journal_preres_u64s,
 475                         JOURNAL_RES_GET_NONBLOCK);
 476         if (unlikely(ret == -EAGAIN))
 477                 ret = bch2_trans_journal_preres_get_cold(trans,
 478                                                 trans->journal_preres_u64s);
 479         if (unlikely(ret))
 480                 return ret;
 481
 482         /*
 483          * Can't be holding any read locks when we go to take write locks:
 484          *
 485          * note - this must be done after bch2_trans_journal_preres_get_cold()
 486          * or anything else that might call bch2_trans_relock(), since that
 487          * would just retake the read locks:
 488          */
 489         trans_for_each_iter_all(trans, iter) {
 490                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 491                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 492                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 493                         bch2_btree_iter_unlock_noinline(iter);
 494                 }
 495         }
 496
 497         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 498                 trans_for_each_update2(trans, i)
 499                         btree_insert_entry_checks(trans, i->iter, i->k);
 500         bch2_btree_trans_verify_locks(trans);
 501
 502         trans_for_each_update2(trans, i)
 503                 if (!same_leaf_as_prev(trans, i))
 504                         bch2_btree_node_lock_for_insert(trans->c,
 505                                         iter_l(i->iter)->b, i->iter);
 506
 507         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 508
 509         trans_for_each_update2(trans, i)
 510                 if (!same_leaf_as_prev(trans, i))
 511                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 512                                                              i->iter);
 513
 514         /*
 515          * Drop journal reservation after dropping write locks, since dropping
 516          * the journal reservation may kick off a journal write:
 517          */
 518         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 519
 520         if (unlikely(ret))
 521                 return ret;
 522
 523         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 524                 trans->nounlock = true;
 525
 526         trans_for_each_update2(trans, i)
 527                 if (!same_leaf_as_prev(trans, i))
 528                         bch2_foreground_maybe_merge(trans->c, i->iter,
 529                                                     0, trans->flags);
 530
 531         trans->nounlock = false;
 532
 533         trans_for_each_update2(trans, i)
 534                 bch2_btree_iter_downgrade(i->iter);
 535
 536         return 0;
 537 }
 538
 539 static noinline
 540 int bch2_trans_commit_error(struct btree_trans *trans,
 541                             struct btree_insert_entry *i,
 542                             int ret)
 543 {
 544         struct bch_fs *c = trans->c;
 545         unsigned flags = trans->flags;
 546
 547         /*
 548          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 549          * update; if we haven't done anything yet it doesn't apply
 550          */
 551         flags &= ~BTREE_INSERT_NOUNLOCK;
 552
 553         switch (ret) {
 554         case BTREE_INSERT_BTREE_NODE_FULL:
 555                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 556
 557                 /*
 558                  * if the split succeeded without dropping locks the insert will
 559                  * still be atomic (what the caller peeked() and is overwriting
 560                  * won't have changed)
 561                  */
 562 #if 0
 563                 /*
 564                  * XXX:
 565                  * split -> btree node merging (of parent node) might still drop
 566                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 567                  *
 568                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 569                  * will inhibit merging - but we don't have a reliable way yet
 570                  * (do we?) of checking if we dropped locks in this path
 571                  */
 572                 if (!ret)
 573                         goto retry;
 574 #endif
 575
 576                 /*
 577                  * don't care if we got ENOSPC because we told split it
 578                  * couldn't block:
 579                  */
 580                 if (!ret ||
 581                     ret == -EINTR ||
 582                     (flags & BTREE_INSERT_NOUNLOCK)) {
 583                         trace_trans_restart_btree_node_split(trans->ip);
 584                         ret = -EINTR;
 585                 }
 586                 break;
 587         case BTREE_INSERT_ENOSPC:
 588                 ret = -ENOSPC;
 589                 break;
 590         case BTREE_INSERT_NEED_MARK_REPLICAS:
 591                 bch2_trans_unlock(trans);
 592
 593                 trans_for_each_update(trans, i) {
 594                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 595                         if (ret)
 596                                 return ret;
 597                 }
 598
 599                 if (bch2_trans_relock(trans))
 600                         return 0;
 601
 602                 trace_trans_restart_mark_replicas(trans->ip);
 603                 ret = -EINTR;
 604                 break;
 605         case BTREE_INSERT_NEED_JOURNAL_RES:
 606                 bch2_trans_unlock(trans);
 607
 608                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 609                 if (ret)
 610                         return ret;
 611
 612                 if (bch2_trans_relock(trans))
 613                         return 0;
 614
 615                 trace_trans_restart_journal_res_get(trans->ip);
 616                 ret = -EINTR;
 617                 break;
 618         default:
 619                 BUG_ON(ret >= 0);
 620                 break;
 621         }
 622
 623         if (ret == -EINTR) {
 624                 int ret2 = bch2_btree_iter_traverse_all(trans);
 625
 626                 if (ret2) {
 627                         trace_trans_restart_traverse(trans->ip);
 628                         return ret2;
 629                 }
 630
 631                 trace_trans_restart_atomic(trans->ip);
 632         }
 633
 634         return ret;
 635 }
 636
 637 static noinline int
 638 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 639 {
 640         struct bch_fs *c = trans->c;
 641         int ret;
 642
 643         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 644                 return -EROFS;
 645
 646         bch2_trans_unlock(trans);
 647
 648         ret = bch2_fs_read_write_early(c);
 649         if (ret)
 650                 return ret;
 651
 652         percpu_ref_get(&c->writes);
 653         return 0;
 654 }
 655
 656 static void bch2_trans_update2(struct btree_trans *trans,
 657                                struct btree_iter *iter,
 658                                struct bkey_i *insert)
 659 {
 660         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 661                 .iter = iter, .k = insert
 662         };
 663
 664         btree_insert_entry_checks(trans, n.iter, n.k);
 665
 666         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 667
 668         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 669
 670         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 671
 672         trans_for_each_update2(trans, i) {
 673                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 674                         *i = n;
 675                         return;
 676                 }
 677
 678                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 679                         break;
 680         }
 681
 682         array_insert_item(trans->updates2, trans->nr_updates2,
 683                           i - trans->updates2, n);
 684 }
 685
 686 static int extent_update_to_keys(struct btree_trans *trans,
 687                                  struct btree_iter *orig_iter,
 688                                  struct bkey_i *insert)
 689 {
 690         struct btree_iter *iter;
 691
 692         if (bkey_deleted(&insert->k))
 693                 return 0;
 694
 695         iter = bch2_trans_copy_iter(trans, orig_iter);
 696         if (IS_ERR(iter))
 697                 return PTR_ERR(iter);
 698
 699         iter->flags |= BTREE_ITER_INTENT;
 700         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 701         bch2_trans_update2(trans, iter, insert);
 702         bch2_trans_iter_put(trans, iter);
 703         return 0;
 704 }
 705
 706 static int extent_handle_overwrites(struct btree_trans *trans,
 707                                     enum btree_id btree_id,
 708                                     struct bpos start, struct bpos end)
 709 {
 710         struct btree_iter *iter = NULL, *update_iter;
 711         struct bkey_i *update;
 712         struct bkey_s_c k;
 713         int ret = 0;
 714
 715         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 716         ret = PTR_ERR_OR_ZERO(iter);
 717         if (ret)
 718                 return ret;
 719
 720         k = bch2_btree_iter_peek_with_updates(iter);
 721
 722         while (k.k && !(ret = bkey_err(k))) {
 723                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 724                         break;
 725
 726                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 727                         update_iter = bch2_trans_copy_iter(trans, iter);
 728                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 729                                 goto err;
 730
 731                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 732                         if ((ret = PTR_ERR_OR_ZERO(update)))
 733                                 goto err;
 734
 735                         bkey_reassemble(update, k);
 736                         bch2_cut_back(start, update);
 737
 738                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 739                         bch2_trans_update2(trans, update_iter, update);
 740                         bch2_trans_iter_put(trans, update_iter);
 741                 }
 742
 743                 if (bkey_cmp(k.k->p, end) > 0) {
 744                         update_iter = bch2_trans_copy_iter(trans, iter);
 745                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 746                                 goto err;
 747
 748                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 749                         if ((ret = PTR_ERR_OR_ZERO(update)))
 750                                 goto err;
 751
 752                         bkey_reassemble(update, k);
 753                         bch2_cut_front(end, update);
 754
 755                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 756                         bch2_trans_update2(trans, update_iter, update);
 757                         bch2_trans_iter_put(trans, update_iter);
 758                 } else {
 759                         update_iter = bch2_trans_copy_iter(trans, iter);
 760                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 761                                 goto err;
 762
 763                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 764                         if ((ret = PTR_ERR_OR_ZERO(update)))
 765                                 goto err;
 766
 767                         update->k = *k.k;
 768                         set_bkey_val_u64s(&update->k, 0);
 769                         update->k.type = KEY_TYPE_deleted;
 770                         update->k.size = 0;
 771
 772                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 773                         bch2_trans_update2(trans, update_iter, update);
 774                         bch2_trans_iter_put(trans, update_iter);
 775                 }
 776
 777                 k = bch2_btree_iter_next_with_updates(iter);
 778         }
 779 err:
 780         if (!IS_ERR_OR_NULL(iter))
 781                 bch2_trans_iter_put(trans, iter);
 782         return ret;
 783 }
 784
 785 int __bch2_trans_commit(struct btree_trans *trans)
 786 {
 787         struct btree_insert_entry *i = NULL;
 788         struct btree_iter *iter;
 789         bool trans_trigger_run;
 790         unsigned u64s;
 791         int ret = 0;
 792
 793         BUG_ON(trans->need_reset);
 794
 795         if (!trans->nr_updates)
 796                 goto out_noupdates;
 797
 798         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 799                 lockdep_assert_held(&trans->c->gc_lock);
 800
 801         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 802
 803         trans->journal_u64s             = 0;
 804         trans->journal_preres_u64s      = 0;
 805
 806         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 807             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 808                 ret = bch2_trans_commit_get_rw_cold(trans);
 809                 if (ret)
 810                         return ret;
 811         }
 812
 813         /*
 814          * Running triggers will append more updates to the list of updates as
 815          * we're walking it:
 816          */
 817         do {
 818                 trans_trigger_run = false;
 819
 820                 trans_for_each_update(trans, i) {
 821                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK)) {
 822                                 trace_trans_restart_traverse(trans->ip);
 823                                 ret = -EINTR;
 824                                 goto out;
 825                         }
 826
 827                         /*
 828                          * We're not using bch2_btree_iter_upgrade here because
 829                          * we know trans->nounlock can't be set:
 830                          */
 831                         if (unlikely(i->iter->locks_want < 1 &&
 832                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 833                                 trace_trans_restart_upgrade(trans->ip);
 834                                 ret = -EINTR;
 835                                 goto out;
 836                         }
 837
 838                         if (iter_has_trans_triggers(i->iter) &&
 839                             !i->trans_triggers_run) {
 840                                 i->trans_triggers_run = true;
 841                                 trans_trigger_run = true;
 842
 843                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 844                                                              i->trigger_flags);
 845                                 if (unlikely(ret)) {
 846                                         if (ret == -EINTR)
 847                                                 trace_trans_restart_mark(trans->ip);
 848                                         goto out;
 849                                 }
 850                         }
 851                 }
 852         } while (trans_trigger_run);
 853
 854         /* Turn extents updates into keys: */
 855         trans_for_each_update(trans, i)
 856                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 857                         struct bpos start = bkey_start_pos(&i->k->k);
 858
 859                         while (i + 1 < trans->updates + trans->nr_updates &&
 860                                i[0].iter->btree_id == i[1].iter->btree_id &&
 861                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 862                                 i++;
 863
 864                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 865                                                        start, i->k->k.p);
 866                         if (ret)
 867                                 goto out;
 868                 }
 869
 870         trans_for_each_update(trans, i) {
 871                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 872                         ret = extent_update_to_keys(trans, i->iter, i->k);
 873                         if (ret)
 874                                 goto out;
 875                 } else {
 876                         bch2_trans_update2(trans, i->iter, i->k);
 877                 }
 878         }
 879
 880         trans_for_each_update2(trans, i) {
 881                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 882                 BUG_ON(i->iter->locks_want < 1);
 883
 884                 u64s = jset_u64s(i->k->k.u64s);
 885                 if (0)
 886                         trans->journal_preres_u64s += u64s;
 887                 trans->journal_u64s += u64s;
 888         }
 889 retry:
 890         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 891
 892         ret = do_bch2_trans_commit(trans, &i);
 893
 894         /* make sure we didn't drop or screw up locks: */
 895         bch2_btree_trans_verify_locks(trans);
 896
 897         if (ret)
 898                 goto err;
 899
 900         trans_for_each_iter(trans, iter)
 901                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 902                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 903                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 904                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 905                         else
 906                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 907                 }
 908 out:
 909         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 910
 911         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 912                 percpu_ref_put(&trans->c->writes);
 913 out_noupdates:
 914         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 915
 916         return ret;
 917 err:
 918         ret = bch2_trans_commit_error(trans, i, ret);
 919         if (ret)
 920                 goto out;
 921
 922         goto retry;
 923 }
 924
 925 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 926                       struct bkey_i *k, enum btree_trigger_flags flags)
 927 {
 928         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 929                 .trigger_flags = flags, .iter = iter, .k = k
 930         };
 931
 932         EBUG_ON(bkey_cmp(iter->pos,
 933                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 934                          ? bkey_start_pos(&k->k)
 935                          : k->k.p));
 936
 937         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 938
 939         if (btree_node_type_is_extents(iter->btree_id)) {
 940                 iter->pos_after_commit = k->k.p;
 941                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 942         }
 943
 944         /*
 945          * Pending updates are kept sorted: first, find position of new update:
 946          */
 947         trans_for_each_update(trans, i)
 948                 if (btree_iter_cmp(iter, i->iter) <= 0)
 949                         break;
 950
 951         /*
 952          * Now delete/trim any updates the new update overwrites:
 953          */
 954         if (i > trans->updates &&
 955             i[-1].iter->btree_id == iter->btree_id &&
 956             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 957                 bch2_cut_back(n.iter->pos, i[-1].k);
 958
 959         while (i < trans->updates + trans->nr_updates &&
 960                iter->btree_id == i->iter->btree_id &&
 961                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
 962                 array_remove_item(trans->updates, trans->nr_updates,
 963                                   i - trans->updates);
 964
 965         if (i < trans->updates + trans->nr_updates &&
 966             iter->btree_id == i->iter->btree_id &&
 967             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
 968                 /*
 969                  * When we have an extent that overwrites the start of another
 970                  * update, trimming that extent will mean the iterator's
 971                  * position has to change since the iterator position has to
 972                  * match the extent's start pos - but we don't want to change
 973                  * the iterator pos if some other code is using it, so we may
 974                  * need to clone it:
 975                  */
 976                 if (trans->iters_live & (1ULL << i->iter->idx)) {
 977                         i->iter = bch2_trans_copy_iter(trans, i->iter);
 978                         if (IS_ERR(i->iter)) {
 979                                 trans->need_reset = true;
 980                                 return PTR_ERR(i->iter);
 981                         }
 982
 983                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 984                         bch2_trans_iter_put(trans, i->iter);
 985                 }
 986
 987                 bch2_cut_front(n.k->k.p, i->k);
 988                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
 989         }
 990
 991         EBUG_ON(trans->nr_updates >= trans->nr_iters);
 992
 993         array_insert_item(trans->updates, trans->nr_updates,
 994                           i - trans->updates, n);
 995         return 0;
 996 }
 997
 998 int __bch2_btree_insert(struct btree_trans *trans,
 999                         enum btree_id id, struct bkey_i *k)
1000 {
1001         struct btree_iter *iter;
1002         int ret;
1003
1004         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1005                                    BTREE_ITER_INTENT);
1006         if (IS_ERR(iter))
1007                 return PTR_ERR(iter);
1008
1009         ret   = bch2_btree_iter_traverse(iter) ?:
1010                 bch2_trans_update(trans, iter, k, 0);
1011         bch2_trans_iter_put(trans, iter);
1012         return ret;
1013 }
1014
1015 /**
1016  * bch2_btree_insert - insert keys into the extent btree
1017  * @c:                  pointer to struct bch_fs
1018  * @id:                 btree to insert into
1019  * @insert_keys:        list of keys to insert
1020  * @hook:               insert callback
1021  */
1022 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1023                       struct bkey_i *k,
1024                       struct disk_reservation *disk_res,
1025                       u64 *journal_seq, int flags)
1026 {
1027         return bch2_trans_do(c, disk_res, journal_seq, flags,
1028                              __bch2_btree_insert(&trans, id, k));
1029 }
1030
1031 int bch2_btree_delete_at_range(struct btree_trans *trans,
1032                                struct btree_iter *iter,
1033                                struct bpos end,
1034                                u64 *journal_seq)
1035 {
1036         struct bkey_s_c k;
1037         int ret = 0;
1038 retry:
1039         while ((k = bch2_btree_iter_peek(iter)).k &&
1040                !(ret = bkey_err(k)) &&
1041                bkey_cmp(iter->pos, end) < 0) {
1042                 struct bkey_i delete;
1043
1044                 bch2_trans_begin(trans);
1045
1046                 bkey_init(&delete.k);
1047
1048                 /*
1049                  * For extents, iter.pos won't necessarily be the same as
1050                  * bkey_start_pos(k.k) (for non extents they always will be the
1051                  * same). It's important that we delete starting from iter.pos
1052                  * because the range we want to delete could start in the middle
1053                  * of k.
1054                  *
1055                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1056                  * bkey_start_pos(k.k)).
1057                  */
1058                 delete.k.p = iter->pos;
1059
1060                 if (btree_node_type_is_extents(iter->btree_id)) {
1061                         unsigned max_sectors =
1062                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1063
1064                         /* create the biggest key we can */
1065                         bch2_key_resize(&delete.k, max_sectors);
1066                         bch2_cut_back(end, &delete);
1067
1068                         ret = bch2_extent_trim_atomic(&delete, iter);
1069                         if (ret)
1070                                 break;
1071                 }
1072
1073                 bch2_trans_update(trans, iter, &delete, 0);
1074                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1075                                         BTREE_INSERT_NOFAIL);
1076                 if (ret)
1077                         break;
1078
1079                 bch2_trans_cond_resched(trans);
1080         }
1081
1082         if (ret == -EINTR) {
1083                 ret = 0;
1084                 goto retry;
1085         }
1086
1087         return ret;
1088
1089 }
1090
1091 int bch2_btree_delete_at(struct btree_trans *trans,
1092                          struct btree_iter *iter, unsigned flags)
1093 {
1094         struct bkey_i k;
1095
1096         bkey_init(&k.k);
1097         k.k.p = iter->pos;
1098
1099         bch2_trans_update(trans, iter, &k, 0);
1100         return bch2_trans_commit(trans, NULL, NULL,
1101                                  BTREE_INSERT_NOFAIL|
1102                                  BTREE_INSERT_USE_RESERVE|flags);
1103 }
1104
1105 /*
1106  * bch_btree_delete_range - delete everything within a given range
1107  *
1108  * Range is a half open interval - [start, end)
1109  */
1110 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1111                             struct bpos start, struct bpos end,
1112                             u64 *journal_seq)
1113 {
1114         struct btree_trans trans;
1115         struct btree_iter *iter;
1116         int ret = 0;
1117
1118         /*
1119          * XXX: whether we need mem/more iters depends on whether this btree id
1120          * has triggers
1121          */
1122         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1123
1124         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1125
1126         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1127         ret = bch2_trans_exit(&trans) ?: ret;
1128
1129         BUG_ON(ret == -EINTR);
1130         return ret;
1131 }