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