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->updates &&
  27                 i[0].iter->l[0].b == i[-1].iter->l[0].b;
  28 }
  29
  30
  31 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  32                                             struct btree_iter *iter)
  33 {
  34         bch2_btree_node_lock_write(b, iter);
  35
  36         if (unlikely(btree_node_just_written(b)) &&
  37             bch2_btree_post_write_cleanup(c, b))
  38                 bch2_btree_iter_reinit_node(iter, b);
  39
  40         /*
  41          * If the last bset has been written, or if it's gotten too big - start
  42          * a new bset to insert into:
  43          */
  44         if (want_new_bset(c, b))
  45                 bch2_btree_init_next(c, b, iter);
  46 }
  47
  48 /* Inserting into a given leaf node (last stage of insert): */
  49
  50 /* Handle overwrites and do insert, for non extents: */
  51 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  52                                 struct btree *b,
  53                                 struct btree_node_iter *node_iter,
  54                                 struct bkey_i *insert)
  55 {
  56         struct bkey_packed *k;
  57         unsigned clobber_u64s = 0, new_u64s = 0;
  58
  59         EBUG_ON(btree_node_just_written(b));
  60         EBUG_ON(bset_written(b, btree_bset_last(b)));
  61         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  62         EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
  63                 bkey_cmp(insert->k.p, b->data->max_key) > 0);
  64
  65         k = bch2_btree_node_iter_peek_all(node_iter, b);
  66         if (k && bkey_cmp_packed(b, k, &insert->k))
  67                 k = NULL;
  68
  69         /* @k is the key being overwritten/deleted, if any: */
  70         EBUG_ON(k && bkey_whiteout(k));
  71
  72         /* Deleting, but not found? nothing to do: */
  73         if (bkey_whiteout(&insert->k) && !k)
  74                 return false;
  75
  76         if (bkey_whiteout(&insert->k)) {
  77                 /* Deleting: */
  78                 btree_account_key_drop(b, k);
  79                 k->type = KEY_TYPE_deleted;
  80
  81                 if (k->needs_whiteout)
  82                         push_whiteout(iter->trans->c, b, k);
  83                 k->needs_whiteout = false;
  84
  85                 if (k >= btree_bset_last(b)->start) {
  86                         clobber_u64s = k->u64s;
  87                         bch2_bset_delete(b, k, clobber_u64s);
  88                         goto fix_iter;
  89                 } else {
  90                         bch2_btree_iter_fix_key_modified(iter, b, k);
  91                 }
  92
  93                 return true;
  94         }
  95
  96         if (k) {
  97                 /* Overwriting: */
  98                 btree_account_key_drop(b, k);
  99                 k->type = KEY_TYPE_deleted;
 100
 101                 insert->k.needs_whiteout = k->needs_whiteout;
 102                 k->needs_whiteout = false;
 103
 104                 if (k >= btree_bset_last(b)->start) {
 105                         clobber_u64s = k->u64s;
 106                         goto overwrite;
 107                 } else {
 108                         bch2_btree_iter_fix_key_modified(iter, b, k);
 109                 }
 110         }
 111
 112         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 113 overwrite:
 114         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 115         new_u64s = k->u64s;
 116 fix_iter:
 117         if (clobber_u64s != new_u64s)
 118                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 119                                          clobber_u64s, new_u64s);
 120         return true;
 121 }
 122
 123 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 124                                unsigned i, u64 seq)
 125 {
 126         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 127         struct btree_write *w = container_of(pin, struct btree_write, journal);
 128         struct btree *b = container_of(w, struct btree, writes[i]);
 129
 130         btree_node_lock_type(c, b, SIX_LOCK_read);
 131         bch2_btree_node_write_cond(c, b,
 132                 (btree_current_write(b) == w && w->journal.seq == seq));
 133         six_unlock_read(&b->lock);
 134 }
 135
 136 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 137 {
 138         return __btree_node_flush(j, pin, 0, seq);
 139 }
 140
 141 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 142 {
 143         return __btree_node_flush(j, pin, 1, seq);
 144 }
 145
 146 static inline void __btree_journal_key(struct btree_trans *trans,
 147                                        enum btree_id btree_id,
 148                                        struct bkey_i *insert)
 149 {
 150         struct journal *j = &trans->c->journal;
 151         u64 seq = trans->journal_res.seq;
 152         bool needs_whiteout = insert->k.needs_whiteout;
 153
 154         /* ick */
 155         insert->k.needs_whiteout = false;
 156         bch2_journal_add_keys(j, &trans->journal_res,
 157                               btree_id, insert);
 158         insert->k.needs_whiteout = needs_whiteout;
 159
 160         bch2_journal_set_has_inode(j, &trans->journal_res,
 161                                    insert->k.p.inode);
 162
 163         if (trans->journal_seq)
 164                 *trans->journal_seq = seq;
 165 }
 166
 167 void bch2_btree_journal_key(struct btree_trans *trans,
 168                            struct btree_iter *iter,
 169                            struct bkey_i *insert)
 170 {
 171         struct bch_fs *c = trans->c;
 172         struct journal *j = &c->journal;
 173         struct btree *b = iter->l[0].b;
 174         struct btree_write *w = btree_current_write(b);
 175
 176         EBUG_ON(iter->level || b->level);
 177         EBUG_ON(trans->journal_res.ref !=
 178                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 179
 180         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 181                 __btree_journal_key(trans, iter->btree_id, insert);
 182                 btree_bset_last(b)->journal_seq =
 183                         cpu_to_le64(trans->journal_res.seq);
 184         }
 185
 186         if (unlikely(!journal_pin_active(&w->journal))) {
 187                 u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 188                         ? trans->journal_res.seq
 189                         : j->replay_journal_seq;
 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
 197         if (unlikely(!btree_node_dirty(b)))
 198                 set_btree_node_dirty(b);
 199 }
 200
 201 static void bch2_insert_fixup_key(struct btree_trans *trans,
 202                                   struct btree_iter *iter,
 203                                   struct bkey_i *insert)
 204 {
 205         struct btree_iter_level *l = &iter->l[0];
 206
 207         EBUG_ON(iter->level);
 208         EBUG_ON(insert->k.u64s >
 209                 bch_btree_keys_u64s_remaining(trans->c, l->b));
 210
 211         if (likely(bch2_btree_bset_insert_key(iter, l->b, &l->iter, insert)))
 212                 bch2_btree_journal_key(trans, iter, insert);
 213 }
 214
 215 /**
 216  * btree_insert_key - insert a key one key into a leaf node
 217  */
 218 static void btree_insert_key_leaf(struct btree_trans *trans,
 219                                   struct btree_iter *iter,
 220                                   struct bkey_i *insert)
 221 {
 222         struct bch_fs *c = trans->c;
 223         struct btree *b = iter->l[0].b;
 224         struct bset_tree *t = bset_tree_last(b);
 225         int old_u64s = bset_u64s(t);
 226         int old_live_u64s = b->nr.live_u64s;
 227         int live_u64s_added, u64s_added;
 228
 229         insert->k.needs_whiteout = false;
 230
 231         if (!btree_node_is_extents(b))
 232                 bch2_insert_fixup_key(trans, iter, insert);
 233         else
 234                 bch2_insert_fixup_extent(trans, iter, insert);
 235
 236         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 237         u64s_added = (int) bset_u64s(t) - old_u64s;
 238
 239         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 240                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 241         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 242                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 243
 244         if (u64s_added > live_u64s_added &&
 245             bch2_maybe_compact_whiteouts(c, b))
 246                 bch2_btree_iter_reinit_node(iter, b);
 247
 248         trace_btree_insert_key(c, b, insert);
 249 }
 250
 251 /* Normal update interface: */
 252
 253 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 254                                              struct btree_iter *iter,
 255                                              struct bkey_i *insert)
 256 {
 257         struct bch_fs *c = trans->c;
 258
 259         BUG_ON(iter->level);
 260         BUG_ON(bkey_cmp(bkey_start_pos(&insert->k), iter->pos));
 261         EBUG_ON((iter->flags & BTREE_ITER_IS_EXTENTS) &&
 262                 bkey_cmp(insert->k.p, iter->l[0].b->key.k.p) > 0);
 263
 264         BUG_ON(debug_check_bkeys(c) &&
 265                !bkey_deleted(&insert->k) &&
 266                bch2_bkey_invalid(c, bkey_i_to_s_c(insert), iter->btree_id));
 267 }
 268
 269 static noinline int
 270 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 271 {
 272         struct bch_fs *c = trans->c;
 273         int ret;
 274
 275         bch2_trans_unlock(trans);
 276
 277         ret = bch2_journal_preres_get(&c->journal,
 278                         &trans->journal_preres, u64s, 0);
 279         if (ret)
 280                 return ret;
 281
 282         if (!bch2_trans_relock(trans)) {
 283                 trace_trans_restart_journal_preres_get(trans->ip);
 284                 return -EINTR;
 285         }
 286
 287         return 0;
 288 }
 289
 290 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 291                                              unsigned flags)
 292 {
 293         struct bch_fs *c = trans->c;
 294         int ret;
 295
 296         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 297                 flags |= JOURNAL_RES_GET_RESERVED;
 298
 299         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 300                                    trans->journal_u64s, flags);
 301
 302         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 303 }
 304
 305 static enum btree_insert_ret
 306 btree_key_can_insert(struct btree_trans *trans,
 307                      struct btree_iter *iter,
 308                      struct bkey_i *insert,
 309                      unsigned *u64s)
 310 {
 311         struct bch_fs *c = trans->c;
 312         struct btree *b = iter->l[0].b;
 313         static enum btree_insert_ret ret;
 314
 315         if (unlikely(btree_node_fake(b)))
 316                 return BTREE_INSERT_BTREE_NODE_FULL;
 317
 318         ret = !btree_node_is_extents(b)
 319                 ? BTREE_INSERT_OK
 320                 : bch2_extent_can_insert(trans, iter, insert, u64s);
 321         if (ret)
 322                 return ret;
 323
 324         if (*u64s > bch_btree_keys_u64s_remaining(c, b))
 325                 return BTREE_INSERT_BTREE_NODE_FULL;
 326
 327         return BTREE_INSERT_OK;
 328 }
 329
 330 static inline void do_btree_insert_one(struct btree_trans *trans,
 331                                        struct btree_iter *iter,
 332                                        struct bkey_i *insert)
 333 {
 334         btree_insert_key_leaf(trans, iter, insert);
 335 }
 336
 337 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 338 {
 339         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 340 }
 341
 342 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 343 {
 344         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 345                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 346                 (1U << iter->btree_id);
 347 }
 348
 349 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 350 {
 351         __bch2_btree_iter_unlock(iter);
 352 }
 353
 354 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 355 {
 356         struct bch_fs *c = trans->c;
 357         struct btree_insert_entry *i;
 358
 359         trans_for_each_update(trans, i)
 360                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 361                         bch2_mark_update(trans, i->iter, i->k, NULL,
 362                                          i->trigger_flags|BTREE_TRIGGER_GC);
 363 }
 364
 365 static inline int
 366 bch2_trans_commit_write_locked(struct btree_trans *trans,
 367                                struct btree_insert_entry **stopped_at)
 368 {
 369         struct bch_fs *c = trans->c;
 370         struct bch_fs_usage *fs_usage = NULL;
 371         struct btree_insert_entry *i;
 372         unsigned u64s = 0;
 373         bool marking = false;
 374         int ret;
 375
 376         if (race_fault()) {
 377                 trace_trans_restart_fault_inject(trans->ip);
 378                 return -EINTR;
 379         }
 380
 381         /*
 382          * Check if the insert will fit in the leaf node with the write lock
 383          * held, otherwise another thread could write the node changing the
 384          * amount of space available:
 385          */
 386
 387         prefetch(&trans->c->journal.flags);
 388
 389         trans_for_each_update(trans, i) {
 390                 /* Multiple inserts might go to same leaf: */
 391                 if (!same_leaf_as_prev(trans, i))
 392                         u64s = 0;
 393
 394                 u64s += i->k->k.u64s;
 395                 ret = btree_key_can_insert(trans, i->iter, i->k, &u64s);
 396                 if (ret) {
 397                         *stopped_at = i;
 398                         return ret;
 399                 }
 400
 401                 if (btree_node_type_needs_gc(i->iter->btree_id))
 402                         marking = true;
 403         }
 404
 405         if (marking) {
 406                 percpu_down_read(&c->mark_lock);
 407                 fs_usage = bch2_fs_usage_scratch_get(c);
 408         }
 409
 410         /*
 411          * Don't get journal reservation until after we know insert will
 412          * succeed:
 413          */
 414         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 415                 ret = bch2_trans_journal_res_get(trans,
 416                                 JOURNAL_RES_GET_NONBLOCK);
 417                 if (ret)
 418                         goto err;
 419         }
 420
 421         /*
 422          * Not allowed to fail after we've gotten our journal reservation - we
 423          * have to use it:
 424          */
 425
 426         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 427                 if (journal_seq_verify(c))
 428                         trans_for_each_update(trans, i)
 429                                 i->k->k.version.lo = trans->journal_res.seq;
 430                 else if (inject_invalid_keys(c))
 431                         trans_for_each_update(trans, i)
 432                                 i->k->k.version = MAX_VERSION;
 433         }
 434
 435         /* Must be called under mark_lock: */
 436         if (marking && trans->fs_usage_deltas &&
 437             bch2_replicas_delta_list_apply(c, fs_usage,
 438                                            trans->fs_usage_deltas)) {
 439                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 440                 goto err;
 441         }
 442
 443         trans_for_each_update(trans, i)
 444                 if (iter_has_nontrans_triggers(i->iter))
 445                         bch2_mark_update(trans, i->iter, i->k,
 446                                          fs_usage, i->trigger_flags);
 447
 448         if (marking)
 449                 bch2_trans_fs_usage_apply(trans, fs_usage);
 450
 451         if (unlikely(c->gc_pos.phase))
 452                 bch2_trans_mark_gc(trans);
 453
 454         trans_for_each_update(trans, i)
 455                 do_btree_insert_one(trans, i->iter, i->k);
 456 err:
 457         if (marking) {
 458                 bch2_fs_usage_scratch_put(c, fs_usage);
 459                 percpu_up_read(&c->mark_lock);
 460         }
 461
 462         return ret;
 463 }
 464
 465 /*
 466  * Get journal reservation, take write locks, and attempt to do btree update(s):
 467  */
 468 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 469                                        struct btree_insert_entry **stopped_at)
 470 {
 471         struct btree_insert_entry *i;
 472         struct btree_iter *iter;
 473         int ret;
 474
 475         trans_for_each_update(trans, i)
 476                 BUG_ON(!btree_node_intent_locked(i->iter, 0));
 477
 478         ret = bch2_journal_preres_get(&trans->c->journal,
 479                         &trans->journal_preres, trans->journal_preres_u64s,
 480                         JOURNAL_RES_GET_NONBLOCK);
 481         if (unlikely(ret == -EAGAIN))
 482                 ret = bch2_trans_journal_preres_get_cold(trans,
 483                                                 trans->journal_preres_u64s);
 484         if (unlikely(ret))
 485                 return ret;
 486
 487         /*
 488          * Can't be holding any read locks when we go to take write locks:
 489          *
 490          * note - this must be done after bch2_trans_journal_preres_get_cold()
 491          * or anything else that might call bch2_trans_relock(), since that
 492          * would just retake the read locks:
 493          */
 494         trans_for_each_iter_all(trans, iter) {
 495                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 496                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 497                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 498                         bch2_btree_iter_unlock_noinline(iter);
 499                 }
 500         }
 501
 502         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 503                 trans_for_each_update(trans, i)
 504                         btree_insert_entry_checks(trans, i->iter, i->k);
 505         bch2_btree_trans_verify_locks(trans);
 506
 507         trans_for_each_update(trans, i)
 508                 if (!same_leaf_as_prev(trans, i))
 509                         bch2_btree_node_lock_for_insert(trans->c,
 510                                                 i->iter->l[0].b, i->iter);
 511
 512         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 513
 514         trans_for_each_update(trans, i)
 515                 if (!same_leaf_as_prev(trans, i))
 516                         bch2_btree_node_unlock_write_inlined(i->iter->l[0].b,
 517                                                              i->iter);
 518
 519         /*
 520          * Drop journal reservation after dropping write locks, since dropping
 521          * the journal reservation may kick off a journal write:
 522          */
 523         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 524
 525         if (unlikely(ret))
 526                 return ret;
 527
 528         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 529                 trans->nounlock = true;
 530
 531         trans_for_each_update(trans, i)
 532                 if (!same_leaf_as_prev(trans, i))
 533                         bch2_foreground_maybe_merge(trans->c, i->iter,
 534                                                     0, trans->flags);
 535
 536         trans->nounlock = false;
 537
 538         trans_for_each_update(trans, i)
 539                 bch2_btree_iter_downgrade(i->iter);
 540
 541         return 0;
 542 }
 543
 544 static noinline
 545 int bch2_trans_commit_error(struct btree_trans *trans,
 546                             struct btree_insert_entry *i,
 547                             int ret)
 548 {
 549         struct bch_fs *c = trans->c;
 550         unsigned flags = trans->flags;
 551
 552         /*
 553          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 554          * update; if we haven't done anything yet it doesn't apply
 555          */
 556         flags &= ~BTREE_INSERT_NOUNLOCK;
 557
 558         switch (ret) {
 559         case BTREE_INSERT_BTREE_NODE_FULL:
 560                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 561
 562                 /*
 563                  * if the split succeeded without dropping locks the insert will
 564                  * still be atomic (what the caller peeked() and is overwriting
 565                  * won't have changed)
 566                  */
 567 #if 0
 568                 /*
 569                  * XXX:
 570                  * split -> btree node merging (of parent node) might still drop
 571                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 572                  *
 573                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 574                  * will inhibit merging - but we don't have a reliable way yet
 575                  * (do we?) of checking if we dropped locks in this path
 576                  */
 577                 if (!ret)
 578                         goto retry;
 579 #endif
 580
 581                 /*
 582                  * don't care if we got ENOSPC because we told split it
 583                  * couldn't block:
 584                  */
 585                 if (!ret ||
 586                     ret == -EINTR ||
 587                     (flags & BTREE_INSERT_NOUNLOCK)) {
 588                         trace_trans_restart_btree_node_split(trans->ip);
 589                         ret = -EINTR;
 590                 }
 591                 break;
 592         case BTREE_INSERT_ENOSPC:
 593                 ret = -ENOSPC;
 594                 break;
 595         case BTREE_INSERT_NEED_MARK_REPLICAS:
 596                 bch2_trans_unlock(trans);
 597
 598                 trans_for_each_update(trans, i) {
 599                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 600                         if (ret)
 601                                 return ret;
 602                 }
 603
 604                 if (bch2_trans_relock(trans))
 605                         return 0;
 606
 607                 trace_trans_restart_mark_replicas(trans->ip);
 608                 ret = -EINTR;
 609                 break;
 610         case BTREE_INSERT_NEED_JOURNAL_RES:
 611                 bch2_trans_unlock(trans);
 612
 613                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 614                 if (ret)
 615                         return ret;
 616
 617                 if (bch2_trans_relock(trans))
 618                         return 0;
 619
 620                 trace_trans_restart_journal_res_get(trans->ip);
 621                 ret = -EINTR;
 622                 break;
 623         default:
 624                 BUG_ON(ret >= 0);
 625                 break;
 626         }
 627
 628         if (ret == -EINTR) {
 629                 int ret2 = bch2_btree_iter_traverse_all(trans);
 630
 631                 if (ret2) {
 632                         trace_trans_restart_traverse(trans->ip);
 633                         return ret2;
 634                 }
 635
 636                 trace_trans_restart_atomic(trans->ip);
 637         }
 638
 639         return ret;
 640 }
 641
 642 static noinline int
 643 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 644 {
 645         struct bch_fs *c = trans->c;
 646         int ret;
 647
 648         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 649                 return -EROFS;
 650
 651         bch2_trans_unlock(trans);
 652
 653         ret = bch2_fs_read_write_early(c);
 654         if (ret)
 655                 return ret;
 656
 657         percpu_ref_get(&c->writes);
 658         return 0;
 659 }
 660
 661 int __bch2_trans_commit(struct btree_trans *trans)
 662 {
 663         struct btree_insert_entry *i = NULL;
 664         struct btree_iter *iter;
 665         bool trans_trigger_run;
 666         unsigned u64s;
 667         int ret = 0;
 668
 669         BUG_ON(trans->need_reset);
 670
 671         if (!trans->nr_updates)
 672                 goto out_noupdates;
 673
 674         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 675                 lockdep_assert_held(&trans->c->gc_lock);
 676
 677         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 678
 679         trans->journal_u64s             = 0;
 680         trans->journal_preres_u64s      = 0;
 681
 682         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 683             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 684                 ret = bch2_trans_commit_get_rw_cold(trans);
 685                 if (ret)
 686                         return ret;
 687         }
 688
 689         /*
 690          * Running triggers will append more updates to the list of updates as
 691          * we're walking it:
 692          */
 693         do {
 694                 trans_trigger_run = false;
 695
 696                 trans_for_each_update(trans, i) {
 697                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK)) {
 698                                 trace_trans_restart_traverse(trans->ip);
 699                                 ret = -EINTR;
 700                                 goto out;
 701                         }
 702
 703                         /*
 704                          * We're not using bch2_btree_iter_upgrade here because
 705                          * we know trans->nounlock can't be set:
 706                          */
 707                         if (unlikely(i->iter->locks_want < 1 &&
 708                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 709                                 trace_trans_restart_upgrade(trans->ip);
 710                                 ret = -EINTR;
 711                                 goto out;
 712                         }
 713
 714                         if (iter_has_trans_triggers(i->iter) &&
 715                             !i->trans_triggers_run) {
 716                                 i->trans_triggers_run = true;
 717                                 trans_trigger_run = true;
 718
 719                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 720                                                              i->trigger_flags);
 721                                 if (unlikely(ret)) {
 722                                         if (ret == -EINTR)
 723                                                 trace_trans_restart_mark(trans->ip);
 724                                         goto out;
 725                                 }
 726                         }
 727                 }
 728         } while (trans_trigger_run);
 729
 730         trans_for_each_update(trans, i) {
 731                 u64s = jset_u64s(i->k->k.u64s);
 732                 if (0)
 733                         trans->journal_preres_u64s += u64s;
 734                 trans->journal_u64s += u64s;
 735         }
 736 retry:
 737         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 738
 739         ret = do_bch2_trans_commit(trans, &i);
 740
 741         /* make sure we didn't drop or screw up locks: */
 742         bch2_btree_trans_verify_locks(trans);
 743
 744         if (ret)
 745                 goto err;
 746
 747         trans_for_each_iter(trans, iter)
 748                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 749                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 750                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 751                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 752                         else
 753                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 754                 }
 755 out:
 756         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 757
 758         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 759                 percpu_ref_put(&trans->c->writes);
 760 out_noupdates:
 761         bch2_trans_reset(trans, TRANS_RESET_MEM|TRANS_RESET_NOTRAVERSE);
 762
 763         return ret;
 764 err:
 765         ret = bch2_trans_commit_error(trans, i, ret);
 766         if (ret)
 767                 goto out;
 768
 769         goto retry;
 770 }
 771
 772 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 773                       struct bkey_i *k, enum btree_trigger_flags flags)
 774 {
 775         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 776                 .trigger_flags = flags, .iter = iter, .k = k
 777         };
 778
 779         EBUG_ON(bkey_cmp(iter->pos, bkey_start_pos(&k->k)));
 780
 781         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 782
 783         if (btree_node_type_is_extents(iter->btree_id)) {
 784                 iter->pos_after_commit = k->k.p;
 785                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 786         }
 787
 788         /*
 789          * Pending updates are kept sorted: first, find position of new update:
 790          */
 791         trans_for_each_update(trans, i)
 792                 if (btree_iter_cmp(iter, i->iter) <= 0)
 793                         break;
 794
 795         /*
 796          * Now delete/trim any updates the new update overwrites:
 797          */
 798         if (i > trans->updates &&
 799             i[-1].iter->btree_id == iter->btree_id &&
 800             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 801                 bch2_cut_back(n.iter->pos, i[-1].k);
 802
 803         while (i < trans->updates + trans->nr_updates &&
 804                iter->btree_id == i->iter->btree_id &&
 805                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
 806                 array_remove_item(trans->updates, trans->nr_updates,
 807                                   i - trans->updates);
 808
 809         if (i < trans->updates + trans->nr_updates &&
 810             iter->btree_id == i->iter->btree_id &&
 811             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
 812                 /*
 813                  * When we have an extent that overwrites the start of another
 814                  * update, trimming that extent will mean the iterator's
 815                  * position has to change since the iterator position has to
 816                  * match the extent's start pos - but we don't want to change
 817                  * the iterator pos if some other code is using it, so we may
 818                  * need to clone it:
 819                  */
 820                 if (trans->iters_live & (1ULL << i->iter->idx)) {
 821                         i->iter = bch2_trans_copy_iter(trans, i->iter);
 822                         if (IS_ERR(i->iter)) {
 823                                 trans->need_reset = true;
 824                                 return PTR_ERR(i->iter);
 825                         }
 826
 827                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 828                         bch2_trans_iter_put(trans, i->iter);
 829                 }
 830
 831                 bch2_cut_front(n.k->k.p, i->k);
 832                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
 833         }
 834
 835         EBUG_ON(trans->nr_updates >= trans->nr_iters);
 836
 837         array_insert_item(trans->updates, trans->nr_updates,
 838                           i - trans->updates, n);
 839         return 0;
 840 }
 841
 842 static int __bch2_btree_insert(struct btree_trans *trans,
 843                                enum btree_id id, struct bkey_i *k)
 844 {
 845         struct btree_iter *iter;
 846
 847         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
 848                                    BTREE_ITER_INTENT);
 849         if (IS_ERR(iter))
 850                 return PTR_ERR(iter);
 851
 852         bch2_trans_update(trans, iter, k, 0);
 853         return 0;
 854 }
 855
 856 /**
 857  * bch2_btree_insert - insert keys into the extent btree
 858  * @c:                  pointer to struct bch_fs
 859  * @id:                 btree to insert into
 860  * @insert_keys:        list of keys to insert
 861  * @hook:               insert callback
 862  */
 863 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 864                       struct bkey_i *k,
 865                       struct disk_reservation *disk_res,
 866                       u64 *journal_seq, int flags)
 867 {
 868         return bch2_trans_do(c, disk_res, journal_seq, flags,
 869                              __bch2_btree_insert(&trans, id, k));
 870 }
 871
 872 int bch2_btree_delete_at_range(struct btree_trans *trans,
 873                                struct btree_iter *iter,
 874                                struct bpos end,
 875                                u64 *journal_seq)
 876 {
 877         struct bkey_s_c k;
 878         int ret = 0;
 879 retry:
 880         while ((k = bch2_btree_iter_peek(iter)).k &&
 881                !(ret = bkey_err(k)) &&
 882                bkey_cmp(iter->pos, end) < 0) {
 883                 struct bkey_i delete;
 884
 885                 bch2_trans_reset(trans, TRANS_RESET_MEM);
 886
 887                 bkey_init(&delete.k);
 888
 889                 /*
 890                  * For extents, iter.pos won't necessarily be the same as
 891                  * bkey_start_pos(k.k) (for non extents they always will be the
 892                  * same). It's important that we delete starting from iter.pos
 893                  * because the range we want to delete could start in the middle
 894                  * of k.
 895                  *
 896                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 897                  * bkey_start_pos(k.k)).
 898                  */
 899                 delete.k.p = iter->pos;
 900
 901                 if (btree_node_type_is_extents(iter->btree_id)) {
 902                         unsigned max_sectors =
 903                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
 904
 905                         /* create the biggest key we can */
 906                         bch2_key_resize(&delete.k, max_sectors);
 907                         bch2_cut_back(end, &delete);
 908
 909                         ret = bch2_extent_trim_atomic(&delete, iter);
 910                         if (ret)
 911                                 break;
 912                 }
 913
 914                 bch2_trans_update(trans, iter, &delete, 0);
 915                 ret = bch2_trans_commit(trans, NULL, journal_seq,
 916                                         BTREE_INSERT_NOFAIL);
 917                 if (ret)
 918                         break;
 919
 920                 bch2_trans_cond_resched(trans);
 921         }
 922
 923         if (ret == -EINTR) {
 924                 ret = 0;
 925                 goto retry;
 926         }
 927
 928         return ret;
 929
 930 }
 931
 932 int bch2_btree_delete_at(struct btree_trans *trans,
 933                          struct btree_iter *iter, unsigned flags)
 934 {
 935         struct bkey_i k;
 936
 937         bkey_init(&k.k);
 938         k.k.p = iter->pos;
 939
 940         bch2_trans_update(trans, iter, &k, 0);
 941         return bch2_trans_commit(trans, NULL, NULL,
 942                                  BTREE_INSERT_NOFAIL|
 943                                  BTREE_INSERT_USE_RESERVE|flags);
 944 }
 945
 946 /*
 947  * bch_btree_delete_range - delete everything within a given range
 948  *
 949  * Range is a half open interval - [start, end)
 950  */
 951 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 952                             struct bpos start, struct bpos end,
 953                             u64 *journal_seq)
 954 {
 955         struct btree_trans trans;
 956         struct btree_iter *iter;
 957         int ret = 0;
 958
 959         /*
 960          * XXX: whether we need mem/more iters depends on whether this btree id
 961          * has triggers
 962          */
 963         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
 964
 965         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
 966
 967         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
 968         ret = bch2_trans_exit(&trans) ?: ret;
 969
 970         BUG_ON(ret == -EINTR);
 971         return ret;
 972 }