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         if (unlikely(trans->extra_journal_entry_u64s)) {
 417                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 418                                   trans->extra_journal_entries,
 419                                   trans->extra_journal_entry_u64s);
 420
 421                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 422                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 423         }
 424
 425         /*
 426          * Not allowed to fail after we've gotten our journal reservation - we
 427          * have to use it:
 428          */
 429
 430         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 431                 if (journal_seq_verify(c))
 432                         trans_for_each_update2(trans, i)
 433                                 i->k->k.version.lo = trans->journal_res.seq;
 434                 else if (inject_invalid_keys(c))
 435                         trans_for_each_update2(trans, i)
 436                                 i->k->k.version = MAX_VERSION;
 437         }
 438
 439         /* Must be called under mark_lock: */
 440         if (marking && trans->fs_usage_deltas &&
 441             bch2_replicas_delta_list_apply(c, fs_usage,
 442                                            trans->fs_usage_deltas)) {
 443                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 444                 goto err;
 445         }
 446
 447         trans_for_each_update(trans, i)
 448                 if (iter_has_nontrans_triggers(i->iter))
 449                         bch2_mark_update(trans, i->iter, i->k,
 450                                          fs_usage, i->trigger_flags);
 451
 452         if (marking)
 453                 bch2_trans_fs_usage_apply(trans, fs_usage);
 454
 455         if (unlikely(c->gc_pos.phase))
 456                 bch2_trans_mark_gc(trans);
 457
 458         trans_for_each_update2(trans, i)
 459                 do_btree_insert_one(trans, i->iter, i->k);
 460 err:
 461         if (marking) {
 462                 bch2_fs_usage_scratch_put(c, fs_usage);
 463                 percpu_up_read(&c->mark_lock);
 464         }
 465
 466         return ret;
 467 }
 468
 469 /*
 470  * Get journal reservation, take write locks, and attempt to do btree update(s):
 471  */
 472 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 473                                        struct btree_insert_entry **stopped_at)
 474 {
 475         struct btree_insert_entry *i;
 476         struct btree_iter *iter;
 477         int ret;
 478
 479         trans_for_each_update2(trans, i)
 480                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 481
 482         ret = bch2_journal_preres_get(&trans->c->journal,
 483                         &trans->journal_preres, trans->journal_preres_u64s,
 484                         JOURNAL_RES_GET_NONBLOCK);
 485         if (unlikely(ret == -EAGAIN))
 486                 ret = bch2_trans_journal_preres_get_cold(trans,
 487                                                 trans->journal_preres_u64s);
 488         if (unlikely(ret))
 489                 return ret;
 490
 491         /*
 492          * Can't be holding any read locks when we go to take write locks:
 493          *
 494          * note - this must be done after bch2_trans_journal_preres_get_cold()
 495          * or anything else that might call bch2_trans_relock(), since that
 496          * would just retake the read locks:
 497          */
 498         trans_for_each_iter_all(trans, iter) {
 499                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 500                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 501                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 502                         bch2_btree_iter_unlock_noinline(iter);
 503                 }
 504         }
 505
 506         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 507                 trans_for_each_update2(trans, i)
 508                         btree_insert_entry_checks(trans, i->iter, i->k);
 509         bch2_btree_trans_verify_locks(trans);
 510
 511         trans_for_each_update2(trans, i)
 512                 if (!same_leaf_as_prev(trans, i))
 513                         bch2_btree_node_lock_for_insert(trans->c,
 514                                         iter_l(i->iter)->b, i->iter);
 515
 516         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 517
 518         trans_for_each_update2(trans, i)
 519                 if (!same_leaf_as_prev(trans, i))
 520                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 521                                                              i->iter);
 522
 523         if (!ret && trans->journal_pin)
 524                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 525                                      trans->journal_pin, NULL);
 526
 527         /*
 528          * Drop journal reservation after dropping write locks, since dropping
 529          * the journal reservation may kick off a journal write:
 530          */
 531         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 532
 533         if (unlikely(ret))
 534                 return ret;
 535
 536         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 537                 trans->nounlock = true;
 538
 539         trans_for_each_update2(trans, i)
 540                 if (!same_leaf_as_prev(trans, i))
 541                         bch2_foreground_maybe_merge(trans->c, i->iter,
 542                                                     0, trans->flags);
 543
 544         trans->nounlock = false;
 545
 546         trans_for_each_update2(trans, i)
 547                 bch2_btree_iter_downgrade(i->iter);
 548
 549         return 0;
 550 }
 551
 552 static noinline
 553 int bch2_trans_commit_error(struct btree_trans *trans,
 554                             struct btree_insert_entry *i,
 555                             int ret)
 556 {
 557         struct bch_fs *c = trans->c;
 558         unsigned flags = trans->flags;
 559
 560         /*
 561          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 562          * update; if we haven't done anything yet it doesn't apply
 563          */
 564         flags &= ~BTREE_INSERT_NOUNLOCK;
 565
 566         switch (ret) {
 567         case BTREE_INSERT_BTREE_NODE_FULL:
 568                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 569
 570                 /*
 571                  * if the split succeeded without dropping locks the insert will
 572                  * still be atomic (what the caller peeked() and is overwriting
 573                  * won't have changed)
 574                  */
 575 #if 0
 576                 /*
 577                  * XXX:
 578                  * split -> btree node merging (of parent node) might still drop
 579                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 580                  *
 581                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 582                  * will inhibit merging - but we don't have a reliable way yet
 583                  * (do we?) of checking if we dropped locks in this path
 584                  */
 585                 if (!ret)
 586                         goto retry;
 587 #endif
 588
 589                 /*
 590                  * don't care if we got ENOSPC because we told split it
 591                  * couldn't block:
 592                  */
 593                 if (!ret ||
 594                     ret == -EINTR ||
 595                     (flags & BTREE_INSERT_NOUNLOCK)) {
 596                         trace_trans_restart_btree_node_split(trans->ip);
 597                         ret = -EINTR;
 598                 }
 599                 break;
 600         case BTREE_INSERT_ENOSPC:
 601                 ret = -ENOSPC;
 602                 break;
 603         case BTREE_INSERT_NEED_MARK_REPLICAS:
 604                 bch2_trans_unlock(trans);
 605
 606                 trans_for_each_update(trans, i) {
 607                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 608                         if (ret)
 609                                 return ret;
 610                 }
 611
 612                 if (bch2_trans_relock(trans))
 613                         return 0;
 614
 615                 trace_trans_restart_mark_replicas(trans->ip);
 616                 ret = -EINTR;
 617                 break;
 618         case BTREE_INSERT_NEED_JOURNAL_RES:
 619                 bch2_trans_unlock(trans);
 620
 621                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 622                 if (ret)
 623                         return ret;
 624
 625                 if (bch2_trans_relock(trans))
 626                         return 0;
 627
 628                 trace_trans_restart_journal_res_get(trans->ip);
 629                 ret = -EINTR;
 630                 break;
 631         default:
 632                 BUG_ON(ret >= 0);
 633                 break;
 634         }
 635
 636         if (ret == -EINTR) {
 637                 int ret2 = bch2_btree_iter_traverse_all(trans);
 638
 639                 if (ret2) {
 640                         trace_trans_restart_traverse(trans->ip);
 641                         return ret2;
 642                 }
 643
 644                 trace_trans_restart_atomic(trans->ip);
 645         }
 646
 647         return ret;
 648 }
 649
 650 static noinline int
 651 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 652 {
 653         struct bch_fs *c = trans->c;
 654         int ret;
 655
 656         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 657                 return -EROFS;
 658
 659         bch2_trans_unlock(trans);
 660
 661         ret = bch2_fs_read_write_early(c);
 662         if (ret)
 663                 return ret;
 664
 665         percpu_ref_get(&c->writes);
 666         return 0;
 667 }
 668
 669 static void bch2_trans_update2(struct btree_trans *trans,
 670                                struct btree_iter *iter,
 671                                struct bkey_i *insert)
 672 {
 673         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 674                 .iter = iter, .k = insert
 675         };
 676
 677         btree_insert_entry_checks(trans, n.iter, n.k);
 678
 679         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 680
 681         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 682
 683         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 684
 685         trans_for_each_update2(trans, i) {
 686                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 687                         *i = n;
 688                         return;
 689                 }
 690
 691                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 692                         break;
 693         }
 694
 695         array_insert_item(trans->updates2, trans->nr_updates2,
 696                           i - trans->updates2, n);
 697 }
 698
 699 static int extent_update_to_keys(struct btree_trans *trans,
 700                                  struct btree_iter *orig_iter,
 701                                  struct bkey_i *insert)
 702 {
 703         struct btree_iter *iter;
 704
 705         if (bkey_deleted(&insert->k))
 706                 return 0;
 707
 708         iter = bch2_trans_copy_iter(trans, orig_iter);
 709         if (IS_ERR(iter))
 710                 return PTR_ERR(iter);
 711
 712         iter->flags |= BTREE_ITER_INTENT;
 713         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 714         bch2_trans_update2(trans, iter, insert);
 715         bch2_trans_iter_put(trans, iter);
 716         return 0;
 717 }
 718
 719 static int extent_handle_overwrites(struct btree_trans *trans,
 720                                     enum btree_id btree_id,
 721                                     struct bpos start, struct bpos end)
 722 {
 723         struct btree_iter *iter = NULL, *update_iter;
 724         struct bkey_i *update;
 725         struct bkey_s_c k;
 726         int ret = 0;
 727
 728         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 729         ret = PTR_ERR_OR_ZERO(iter);
 730         if (ret)
 731                 return ret;
 732
 733         k = bch2_btree_iter_peek_with_updates(iter);
 734
 735         while (k.k && !(ret = bkey_err(k))) {
 736                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 737                         break;
 738
 739                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 740                         update_iter = bch2_trans_copy_iter(trans, iter);
 741                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 742                                 goto err;
 743
 744                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 745                         if ((ret = PTR_ERR_OR_ZERO(update)))
 746                                 goto err;
 747
 748                         bkey_reassemble(update, k);
 749                         bch2_cut_back(start, update);
 750
 751                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 752                         bch2_trans_update2(trans, update_iter, update);
 753                         bch2_trans_iter_put(trans, update_iter);
 754                 }
 755
 756                 if (bkey_cmp(k.k->p, end) > 0) {
 757                         update_iter = bch2_trans_copy_iter(trans, iter);
 758                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 759                                 goto err;
 760
 761                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 762                         if ((ret = PTR_ERR_OR_ZERO(update)))
 763                                 goto err;
 764
 765                         bkey_reassemble(update, k);
 766                         bch2_cut_front(end, update);
 767
 768                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 769                         bch2_trans_update2(trans, update_iter, update);
 770                         bch2_trans_iter_put(trans, update_iter);
 771                 } else {
 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, sizeof(struct bkey));
 777                         if ((ret = PTR_ERR_OR_ZERO(update)))
 778                                 goto err;
 779
 780                         update->k = *k.k;
 781                         set_bkey_val_u64s(&update->k, 0);
 782                         update->k.type = KEY_TYPE_deleted;
 783                         update->k.size = 0;
 784
 785                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 786                         bch2_trans_update2(trans, update_iter, update);
 787                         bch2_trans_iter_put(trans, update_iter);
 788                 }
 789
 790                 k = bch2_btree_iter_next_with_updates(iter);
 791         }
 792 err:
 793         if (!IS_ERR_OR_NULL(iter))
 794                 bch2_trans_iter_put(trans, iter);
 795         return ret;
 796 }
 797
 798 int __bch2_trans_commit(struct btree_trans *trans)
 799 {
 800         struct btree_insert_entry *i = NULL;
 801         struct btree_iter *iter;
 802         bool trans_trigger_run;
 803         unsigned u64s;
 804         int ret = 0;
 805
 806         BUG_ON(trans->need_reset);
 807
 808         if (!trans->nr_updates)
 809                 goto out_noupdates;
 810
 811         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 812                 lockdep_assert_held(&trans->c->gc_lock);
 813
 814         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 815
 816         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 817         trans->journal_preres_u64s      = 0;
 818
 819         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 820             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 821                 ret = bch2_trans_commit_get_rw_cold(trans);
 822                 if (ret)
 823                         return ret;
 824         }
 825
 826         /*
 827          * Running triggers will append more updates to the list of updates as
 828          * we're walking it:
 829          */
 830         do {
 831                 trans_trigger_run = false;
 832
 833                 trans_for_each_update(trans, i) {
 834                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK)) {
 835                                 trace_trans_restart_traverse(trans->ip);
 836                                 ret = -EINTR;
 837                                 goto out;
 838                         }
 839
 840                         /*
 841                          * We're not using bch2_btree_iter_upgrade here because
 842                          * we know trans->nounlock can't be set:
 843                          */
 844                         if (unlikely(i->iter->locks_want < 1 &&
 845                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 846                                 trace_trans_restart_upgrade(trans->ip);
 847                                 ret = -EINTR;
 848                                 goto out;
 849                         }
 850
 851                         if (iter_has_trans_triggers(i->iter) &&
 852                             !i->trans_triggers_run) {
 853                                 i->trans_triggers_run = true;
 854                                 trans_trigger_run = true;
 855
 856                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 857                                                              i->trigger_flags);
 858                                 if (unlikely(ret)) {
 859                                         if (ret == -EINTR)
 860                                                 trace_trans_restart_mark(trans->ip);
 861                                         goto out;
 862                                 }
 863                         }
 864                 }
 865         } while (trans_trigger_run);
 866
 867         /* Turn extents updates into keys: */
 868         trans_for_each_update(trans, i)
 869                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 870                         struct bpos start = bkey_start_pos(&i->k->k);
 871
 872                         while (i + 1 < trans->updates + trans->nr_updates &&
 873                                i[0].iter->btree_id == i[1].iter->btree_id &&
 874                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 875                                 i++;
 876
 877                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 878                                                        start, i->k->k.p);
 879                         if (ret)
 880                                 goto out;
 881                 }
 882
 883         trans_for_each_update(trans, i) {
 884                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 885                         ret = extent_update_to_keys(trans, i->iter, i->k);
 886                         if (ret)
 887                                 goto out;
 888                 } else {
 889                         bch2_trans_update2(trans, i->iter, i->k);
 890                 }
 891         }
 892
 893         trans_for_each_update2(trans, i) {
 894                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 895                 BUG_ON(i->iter->locks_want < 1);
 896
 897                 u64s = jset_u64s(i->k->k.u64s);
 898                 if (0)
 899                         trans->journal_preres_u64s += u64s;
 900                 trans->journal_u64s += u64s;
 901         }
 902 retry:
 903         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 904
 905         ret = do_bch2_trans_commit(trans, &i);
 906
 907         /* make sure we didn't drop or screw up locks: */
 908         bch2_btree_trans_verify_locks(trans);
 909
 910         if (ret)
 911                 goto err;
 912
 913         trans_for_each_iter(trans, iter)
 914                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 915                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 916                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 917                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 918                         else
 919                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 920                 }
 921 out:
 922         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 923
 924         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 925                 percpu_ref_put(&trans->c->writes);
 926 out_noupdates:
 927         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 928
 929         return ret;
 930 err:
 931         ret = bch2_trans_commit_error(trans, i, ret);
 932         if (ret)
 933                 goto out;
 934
 935         goto retry;
 936 }
 937
 938 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 939                       struct bkey_i *k, enum btree_trigger_flags flags)
 940 {
 941         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 942                 .trigger_flags = flags, .iter = iter, .k = k
 943         };
 944
 945         EBUG_ON(bkey_cmp(iter->pos,
 946                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 947                          ? bkey_start_pos(&k->k)
 948                          : k->k.p));
 949
 950         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 951
 952         if (btree_node_type_is_extents(iter->btree_id)) {
 953                 iter->pos_after_commit = k->k.p;
 954                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 955         }
 956
 957         /*
 958          * Pending updates are kept sorted: first, find position of new update:
 959          */
 960         trans_for_each_update(trans, i)
 961                 if (btree_iter_cmp(iter, i->iter) <= 0)
 962                         break;
 963
 964         /*
 965          * Now delete/trim any updates the new update overwrites:
 966          */
 967         if (i > trans->updates &&
 968             i[-1].iter->btree_id == iter->btree_id &&
 969             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 970                 bch2_cut_back(n.iter->pos, i[-1].k);
 971
 972         while (i < trans->updates + trans->nr_updates &&
 973                iter->btree_id == i->iter->btree_id &&
 974                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
 975                 array_remove_item(trans->updates, trans->nr_updates,
 976                                   i - trans->updates);
 977
 978         if (i < trans->updates + trans->nr_updates &&
 979             iter->btree_id == i->iter->btree_id &&
 980             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
 981                 /*
 982                  * When we have an extent that overwrites the start of another
 983                  * update, trimming that extent will mean the iterator's
 984                  * position has to change since the iterator position has to
 985                  * match the extent's start pos - but we don't want to change
 986                  * the iterator pos if some other code is using it, so we may
 987                  * need to clone it:
 988                  */
 989                 if (trans->iters_live & (1ULL << i->iter->idx)) {
 990                         i->iter = bch2_trans_copy_iter(trans, i->iter);
 991                         if (IS_ERR(i->iter)) {
 992                                 trans->need_reset = true;
 993                                 return PTR_ERR(i->iter);
 994                         }
 995
 996                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 997                         bch2_trans_iter_put(trans, i->iter);
 998                 }
 999
1000                 bch2_cut_front(n.k->k.p, i->k);
1001                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1002         }
1003
1004         EBUG_ON(trans->nr_updates >= trans->nr_iters);
1005
1006         array_insert_item(trans->updates, trans->nr_updates,
1007                           i - trans->updates, n);
1008         return 0;
1009 }
1010
1011 int __bch2_btree_insert(struct btree_trans *trans,
1012                         enum btree_id id, struct bkey_i *k)
1013 {
1014         struct btree_iter *iter;
1015         int ret;
1016
1017         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1018                                    BTREE_ITER_INTENT);
1019         if (IS_ERR(iter))
1020                 return PTR_ERR(iter);
1021
1022         ret   = bch2_btree_iter_traverse(iter) ?:
1023                 bch2_trans_update(trans, iter, k, 0);
1024         bch2_trans_iter_put(trans, iter);
1025         return ret;
1026 }
1027
1028 /**
1029  * bch2_btree_insert - insert keys into the extent btree
1030  * @c:                  pointer to struct bch_fs
1031  * @id:                 btree to insert into
1032  * @insert_keys:        list of keys to insert
1033  * @hook:               insert callback
1034  */
1035 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1036                       struct bkey_i *k,
1037                       struct disk_reservation *disk_res,
1038                       u64 *journal_seq, int flags)
1039 {
1040         return bch2_trans_do(c, disk_res, journal_seq, flags,
1041                              __bch2_btree_insert(&trans, id, k));
1042 }
1043
1044 int bch2_btree_delete_at_range(struct btree_trans *trans,
1045                                struct btree_iter *iter,
1046                                struct bpos end,
1047                                u64 *journal_seq)
1048 {
1049         struct bkey_s_c k;
1050         int ret = 0;
1051 retry:
1052         while ((k = bch2_btree_iter_peek(iter)).k &&
1053                !(ret = bkey_err(k)) &&
1054                bkey_cmp(iter->pos, end) < 0) {
1055                 struct bkey_i delete;
1056
1057                 bch2_trans_begin(trans);
1058
1059                 bkey_init(&delete.k);
1060
1061                 /*
1062                  * For extents, iter.pos won't necessarily be the same as
1063                  * bkey_start_pos(k.k) (for non extents they always will be the
1064                  * same). It's important that we delete starting from iter.pos
1065                  * because the range we want to delete could start in the middle
1066                  * of k.
1067                  *
1068                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1069                  * bkey_start_pos(k.k)).
1070                  */
1071                 delete.k.p = iter->pos;
1072
1073                 if (btree_node_type_is_extents(iter->btree_id)) {
1074                         unsigned max_sectors =
1075                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1076
1077                         /* create the biggest key we can */
1078                         bch2_key_resize(&delete.k, max_sectors);
1079                         bch2_cut_back(end, &delete);
1080
1081                         ret = bch2_extent_trim_atomic(&delete, iter);
1082                         if (ret)
1083                                 break;
1084                 }
1085
1086                 bch2_trans_update(trans, iter, &delete, 0);
1087                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1088                                         BTREE_INSERT_NOFAIL);
1089                 if (ret)
1090                         break;
1091
1092                 bch2_trans_cond_resched(trans);
1093         }
1094
1095         if (ret == -EINTR) {
1096                 ret = 0;
1097                 goto retry;
1098         }
1099
1100         return ret;
1101
1102 }
1103
1104 int bch2_btree_delete_at(struct btree_trans *trans,
1105                          struct btree_iter *iter, unsigned flags)
1106 {
1107         struct bkey_i k;
1108
1109         bkey_init(&k.k);
1110         k.k.p = iter->pos;
1111
1112         bch2_trans_update(trans, iter, &k, 0);
1113         return bch2_trans_commit(trans, NULL, NULL,
1114                                  BTREE_INSERT_NOFAIL|
1115                                  BTREE_INSERT_USE_RESERVE|flags);
1116 }
1117
1118 /*
1119  * bch_btree_delete_range - delete everything within a given range
1120  *
1121  * Range is a half open interval - [start, end)
1122  */
1123 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1124                             struct bpos start, struct bpos end,
1125                             u64 *journal_seq)
1126 {
1127         struct btree_trans trans;
1128         struct btree_iter *iter;
1129         int ret = 0;
1130
1131         /*
1132          * XXX: whether we need mem/more iters depends on whether this btree id
1133          * has triggers
1134          */
1135         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1136
1137         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1138
1139         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1140         ret = bch2_trans_exit(&trans) ?: ret;
1141
1142         BUG_ON(ret == -EINTR);
1143         return ret;
1144 }