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