git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_leaf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "buckets.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extent_update.h"
  14 #include "journal.h"
  15 #include "journal_reclaim.h"
  16 #include "keylist.h"
  17 #include "replicas.h"
  18
  19 #include <linux/prefetch.h>
  20 #include <linux/sort.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  24                                      struct btree_insert_entry *i)
  25 {
  26         return i != trans->updates2 &&
  27                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  28 }
  29
  30 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  31                                             struct btree_iter *iter)
  32 {
  33         bch2_btree_node_lock_write(b, iter);
  34
  35         if (unlikely(btree_node_just_written(b)) &&
  36             bch2_btree_post_write_cleanup(c, b))
  37                 bch2_btree_iter_reinit_node(iter, b);
  38
  39         /*
  40          * If the last bset has been written, or if it's gotten too big - start
  41          * a new bset to insert into:
  42          */
  43         if (want_new_bset(c, b))
  44                 bch2_btree_init_next(c, b, iter);
  45 }
  46
  47 /* Inserting into a given leaf node (last stage of insert): */
  48
  49 /* Handle overwrites and do insert, for non extents: */
  50 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  51                                 struct btree *b,
  52                                 struct btree_node_iter *node_iter,
  53                                 struct bkey_i *insert)
  54 {
  55         struct bkey_packed *k;
  56         unsigned clobber_u64s = 0, new_u64s = 0;
  57
  58         EBUG_ON(btree_node_just_written(b));
  59         EBUG_ON(bset_written(b, btree_bset_last(b)));
  60         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  61         EBUG_ON(bkey_cmp(b->data->min_key, POS_MIN) &&
  62                 bkey_cmp(bkey_start_pos(&insert->k),
  63                          bkey_predecessor(b->data->min_key)) < 0);
  64         EBUG_ON(bkey_cmp(insert->k.p, b->data->min_key) < 0);
  65         EBUG_ON(bkey_cmp(insert->k.p, b->data->max_key) > 0);
  66         EBUG_ON(insert->k.u64s >
  67                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  68         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  69
  70         k = bch2_btree_node_iter_peek_all(node_iter, b);
  71         if (k && bkey_cmp_packed(b, k, &insert->k))
  72                 k = NULL;
  73
  74         /* @k is the key being overwritten/deleted, if any: */
  75         EBUG_ON(k && bkey_whiteout(k));
  76
  77         /* Deleting, but not found? nothing to do: */
  78         if (bkey_whiteout(&insert->k) && !k)
  79                 return false;
  80
  81         if (bkey_whiteout(&insert->k)) {
  82                 /* Deleting: */
  83                 btree_account_key_drop(b, k);
  84                 k->type = KEY_TYPE_deleted;
  85
  86                 if (k->needs_whiteout)
  87                         push_whiteout(iter->trans->c, b, insert->k.p);
  88                 k->needs_whiteout = false;
  89
  90                 if (k >= btree_bset_last(b)->start) {
  91                         clobber_u64s = k->u64s;
  92                         bch2_bset_delete(b, k, clobber_u64s);
  93                         goto fix_iter;
  94                 } else {
  95                         bch2_btree_iter_fix_key_modified(iter, b, k);
  96                 }
  97
  98                 return true;
  99         }
 100
 101         if (k) {
 102                 /* Overwriting: */
 103                 btree_account_key_drop(b, k);
 104                 k->type = KEY_TYPE_deleted;
 105
 106                 insert->k.needs_whiteout = k->needs_whiteout;
 107                 k->needs_whiteout = false;
 108
 109                 if (k >= btree_bset_last(b)->start) {
 110                         clobber_u64s = k->u64s;
 111                         goto overwrite;
 112                 } else {
 113                         bch2_btree_iter_fix_key_modified(iter, b, k);
 114                 }
 115         }
 116
 117         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 118 overwrite:
 119         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 120         new_u64s = k->u64s;
 121 fix_iter:
 122         if (clobber_u64s != new_u64s)
 123                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 124                                          clobber_u64s, new_u64s);
 125         return true;
 126 }
 127
 128 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 129                                unsigned i, u64 seq)
 130 {
 131         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 132         struct btree_write *w = container_of(pin, struct btree_write, journal);
 133         struct btree *b = container_of(w, struct btree, writes[i]);
 134
 135         btree_node_lock_type(c, b, SIX_LOCK_read);
 136         bch2_btree_node_write_cond(c, b,
 137                 (btree_current_write(b) == w && w->journal.seq == seq));
 138         six_unlock_read(&b->lock);
 139 }
 140
 141 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 142 {
 143         return __btree_node_flush(j, pin, 0, seq);
 144 }
 145
 146 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 147 {
 148         return __btree_node_flush(j, pin, 1, seq);
 149 }
 150
 151 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 152                                        struct btree *b, u64 seq)
 153 {
 154         struct btree_write *w = btree_current_write(b);
 155
 156         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 157                              btree_node_write_idx(b) == 0
 158                              ? btree_node_flush0
 159                              : btree_node_flush1);
 160 }
 161
 162 static inline void __btree_journal_key(struct btree_trans *trans,
 163                                        enum btree_id btree_id,
 164                                        struct bkey_i *insert)
 165 {
 166         struct journal *j = &trans->c->journal;
 167         u64 seq = trans->journal_res.seq;
 168         bool needs_whiteout = insert->k.needs_whiteout;
 169
 170         /* ick */
 171         insert->k.needs_whiteout = false;
 172         bch2_journal_add_keys(j, &trans->journal_res,
 173                               btree_id, insert);
 174         insert->k.needs_whiteout = needs_whiteout;
 175
 176         bch2_journal_set_has_inode(j, &trans->journal_res,
 177                                    insert->k.p.inode);
 178
 179         if (trans->journal_seq)
 180                 *trans->journal_seq = seq;
 181 }
 182
 183 static void bch2_btree_journal_key(struct btree_trans *trans,
 184                                    struct btree_iter *iter,
 185                                    struct bkey_i *insert)
 186 {
 187         struct bch_fs *c = trans->c;
 188         struct journal *j = &c->journal;
 189         struct btree *b = iter_l(iter)->b;
 190
 191         EBUG_ON(trans->journal_res.ref !=
 192                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 193
 194         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 195                 __btree_journal_key(trans, iter->btree_id, insert);
 196                 btree_bset_last(b)->journal_seq =
 197                         cpu_to_le64(trans->journal_res.seq);
 198         }
 199
 200         bch2_btree_add_journal_pin(c, b,
 201                 likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 202                         ? trans->journal_res.seq
 203                         : j->replay_journal_seq);
 204
 205         if (unlikely(!btree_node_dirty(b)))
 206                 set_btree_node_dirty(b);
 207 }
 208
 209 /**
 210  * btree_insert_key - insert a key one key into a leaf node
 211  */
 212 static void btree_insert_key_leaf(struct btree_trans *trans,
 213                                   struct btree_iter *iter,
 214                                   struct bkey_i *insert)
 215 {
 216         struct bch_fs *c = trans->c;
 217         struct btree *b = iter_l(iter)->b;
 218         struct bset_tree *t = bset_tree_last(b);
 219         int old_u64s = bset_u64s(t);
 220         int old_live_u64s = b->nr.live_u64s;
 221         int live_u64s_added, u64s_added;
 222
 223         insert->k.needs_whiteout = false;
 224
 225         if (likely(bch2_btree_bset_insert_key(iter, b, &iter_l(iter)->iter, insert)))
 226                 bch2_btree_journal_key(trans, iter, insert);
 227
 228         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 229         u64s_added = (int) bset_u64s(t) - old_u64s;
 230
 231         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 232                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 233         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 234                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 235
 236         if (u64s_added > live_u64s_added &&
 237             bch2_maybe_compact_whiteouts(c, b))
 238                 bch2_btree_iter_reinit_node(iter, b);
 239
 240         trace_btree_insert_key(c, b, insert);
 241 }
 242
 243 /* Normal update interface: */
 244
 245 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 246                                              struct btree_iter *iter,
 247                                              struct bkey_i *insert)
 248 {
 249         struct bch_fs *c = trans->c;
 250
 251         BUG_ON(bkey_cmp(insert->k.p, iter->pos));
 252         BUG_ON(debug_check_bkeys(c) &&
 253                bch2_bkey_invalid(c, bkey_i_to_s_c(insert), iter->btree_id));
 254 }
 255
 256 static noinline int
 257 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 258 {
 259         struct bch_fs *c = trans->c;
 260         int ret;
 261
 262         bch2_trans_unlock(trans);
 263
 264         ret = bch2_journal_preres_get(&c->journal,
 265                         &trans->journal_preres, u64s, 0);
 266         if (ret)
 267                 return ret;
 268
 269         if (!bch2_trans_relock(trans)) {
 270                 trace_trans_restart_journal_preres_get(trans->ip);
 271                 return -EINTR;
 272         }
 273
 274         return 0;
 275 }
 276
 277 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 278                                              unsigned flags)
 279 {
 280         struct bch_fs *c = trans->c;
 281         int ret;
 282
 283         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 284                 flags |= JOURNAL_RES_GET_RESERVED;
 285
 286         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 287                                    trans->journal_u64s, flags);
 288
 289         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 290 }
 291
 292 static enum btree_insert_ret
 293 btree_key_can_insert(struct btree_trans *trans,
 294                      struct btree_iter *iter,
 295                      struct bkey_i *insert,
 296                      unsigned *u64s)
 297 {
 298         struct bch_fs *c = trans->c;
 299         struct btree *b = iter_l(iter)->b;
 300         static enum btree_insert_ret ret;
 301
 302         if (unlikely(btree_node_fake(b)))
 303                 return BTREE_INSERT_BTREE_NODE_FULL;
 304
 305         /*
 306          * old bch2_extent_sort_fix_overlapping() algorithm won't work with new
 307          * style extent updates:
 308          */
 309         if (unlikely(btree_node_old_extent_overwrite(b)))
 310                 return BTREE_INSERT_BTREE_NODE_FULL;
 311
 312         ret = !btree_node_is_extents(b)
 313                 ? BTREE_INSERT_OK
 314                 : bch2_extent_can_insert(trans, iter, insert);
 315         if (ret)
 316                 return ret;
 317
 318         if (*u64s > bch_btree_keys_u64s_remaining(c, b))
 319                 return BTREE_INSERT_BTREE_NODE_FULL;
 320
 321         return BTREE_INSERT_OK;
 322 }
 323
 324 static inline void do_btree_insert_one(struct btree_trans *trans,
 325                                        struct btree_iter *iter,
 326                                        struct bkey_i *insert)
 327 {
 328         btree_insert_key_leaf(trans, iter, insert);
 329 }
 330
 331 static inline bool iter_has_trans_triggers(struct btree_iter *iter)
 332 {
 333         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << iter->btree_id);
 334 }
 335
 336 static inline bool iter_has_nontrans_triggers(struct btree_iter *iter)
 337 {
 338         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 339                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 340                 (1U << iter->btree_id);
 341 }
 342
 343 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 344 {
 345         __bch2_btree_iter_unlock(iter);
 346 }
 347
 348 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 349 {
 350         struct bch_fs *c = trans->c;
 351         struct btree_insert_entry *i;
 352
 353         trans_for_each_update(trans, i)
 354                 if (gc_visited(c, gc_pos_btree_node(iter_l(i->iter)->b)))
 355                         bch2_mark_update(trans, i->iter, i->k, NULL,
 356                                          i->trigger_flags|BTREE_TRIGGER_GC);
 357 }
 358
 359 static inline int
 360 bch2_trans_commit_write_locked(struct btree_trans *trans,
 361                                struct btree_insert_entry **stopped_at)
 362 {
 363         struct bch_fs *c = trans->c;
 364         struct bch_fs_usage *fs_usage = NULL;
 365         struct btree_insert_entry *i;
 366         unsigned u64s = 0;
 367         bool marking = false;
 368         int ret;
 369
 370         if (race_fault()) {
 371                 trace_trans_restart_fault_inject(trans->ip);
 372                 return -EINTR;
 373         }
 374
 375         /*
 376          * Check if the insert will fit in the leaf node with the write lock
 377          * held, otherwise another thread could write the node changing the
 378          * amount of space available:
 379          */
 380
 381         prefetch(&trans->c->journal.flags);
 382
 383         trans_for_each_update2(trans, i) {
 384                 /* Multiple inserts might go to same leaf: */
 385                 if (!same_leaf_as_prev(trans, i))
 386                         u64s = 0;
 387
 388                 u64s += i->k->k.u64s;
 389                 ret = btree_key_can_insert(trans, i->iter, i->k, &u64s);
 390                 if (ret) {
 391                         *stopped_at = i;
 392                         return ret;
 393                 }
 394
 395                 if (btree_node_type_needs_gc(i->iter->btree_id))
 396                         marking = true;
 397         }
 398
 399         if (marking) {
 400                 percpu_down_read(&c->mark_lock);
 401                 fs_usage = bch2_fs_usage_scratch_get(c);
 402         }
 403
 404         /*
 405          * Don't get journal reservation until after we know insert will
 406          * succeed:
 407          */
 408         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 409                 ret = bch2_trans_journal_res_get(trans,
 410                                 JOURNAL_RES_GET_NONBLOCK);
 411                 if (ret)
 412                         goto err;
 413         }
 414
 415         /*
 416          * Not allowed to fail after we've gotten our journal reservation - we
 417          * have to use it:
 418          */
 419
 420         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 421                 if (journal_seq_verify(c))
 422                         trans_for_each_update2(trans, i)
 423                                 i->k->k.version.lo = trans->journal_res.seq;
 424                 else if (inject_invalid_keys(c))
 425                         trans_for_each_update2(trans, i)
 426                                 i->k->k.version = MAX_VERSION;
 427         }
 428
 429         /* Must be called under mark_lock: */
 430         if (marking && trans->fs_usage_deltas &&
 431             bch2_replicas_delta_list_apply(c, fs_usage,
 432                                            trans->fs_usage_deltas)) {
 433                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 434                 goto err;
 435         }
 436
 437         trans_for_each_update(trans, i)
 438                 if (iter_has_nontrans_triggers(i->iter))
 439                         bch2_mark_update(trans, i->iter, i->k,
 440                                          fs_usage, i->trigger_flags);
 441
 442         if (marking)
 443                 bch2_trans_fs_usage_apply(trans, fs_usage);
 444
 445         if (unlikely(c->gc_pos.phase))
 446                 bch2_trans_mark_gc(trans);
 447
 448         trans_for_each_update2(trans, i)
 449                 do_btree_insert_one(trans, i->iter, i->k);
 450 err:
 451         if (marking) {
 452                 bch2_fs_usage_scratch_put(c, fs_usage);
 453                 percpu_up_read(&c->mark_lock);
 454         }
 455
 456         return ret;
 457 }
 458
 459 /*
 460  * Get journal reservation, take write locks, and attempt to do btree update(s):
 461  */
 462 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 463                                        struct btree_insert_entry **stopped_at)
 464 {
 465         struct btree_insert_entry *i;
 466         struct btree_iter *iter;
 467         int ret;
 468
 469         trans_for_each_update2(trans, i)
 470                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 471
 472         ret = bch2_journal_preres_get(&trans->c->journal,
 473                         &trans->journal_preres, trans->journal_preres_u64s,
 474                         JOURNAL_RES_GET_NONBLOCK);
 475         if (unlikely(ret == -EAGAIN))
 476                 ret = bch2_trans_journal_preres_get_cold(trans,
 477                                                 trans->journal_preres_u64s);
 478         if (unlikely(ret))
 479                 return ret;
 480
 481         /*
 482          * Can't be holding any read locks when we go to take write locks:
 483          *
 484          * note - this must be done after bch2_trans_journal_preres_get_cold()
 485          * or anything else that might call bch2_trans_relock(), since that
 486          * would just retake the read locks:
 487          */
 488         trans_for_each_iter_all(trans, iter) {
 489                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 490                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 491                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 492                         bch2_btree_iter_unlock_noinline(iter);
 493                 }
 494         }
 495
 496         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 497                 trans_for_each_update2(trans, i)
 498                         btree_insert_entry_checks(trans, i->iter, i->k);
 499         bch2_btree_trans_verify_locks(trans);
 500
 501         trans_for_each_update2(trans, i)
 502                 if (!same_leaf_as_prev(trans, i))
 503                         bch2_btree_node_lock_for_insert(trans->c,
 504                                         iter_l(i->iter)->b, i->iter);
 505
 506         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 507
 508         trans_for_each_update2(trans, i)
 509                 if (!same_leaf_as_prev(trans, i))
 510                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 511                                                              i->iter);
 512
 513         /*
 514          * Drop journal reservation after dropping write locks, since dropping
 515          * the journal reservation may kick off a journal write:
 516          */
 517         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 518
 519         if (unlikely(ret))
 520                 return ret;
 521
 522         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 523                 trans->nounlock = true;
 524
 525         trans_for_each_update2(trans, i)
 526                 if (!same_leaf_as_prev(trans, i))
 527                         bch2_foreground_maybe_merge(trans->c, i->iter,
 528                                                     0, trans->flags);
 529
 530         trans->nounlock = false;
 531
 532         trans_for_each_update2(trans, i)
 533                 bch2_btree_iter_downgrade(i->iter);
 534
 535         return 0;
 536 }
 537
 538 static noinline
 539 int bch2_trans_commit_error(struct btree_trans *trans,
 540                             struct btree_insert_entry *i,
 541                             int ret)
 542 {
 543         struct bch_fs *c = trans->c;
 544         unsigned flags = trans->flags;
 545
 546         /*
 547          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 548          * update; if we haven't done anything yet it doesn't apply
 549          */
 550         flags &= ~BTREE_INSERT_NOUNLOCK;
 551
 552         switch (ret) {
 553         case BTREE_INSERT_BTREE_NODE_FULL:
 554                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 555
 556                 /*
 557                  * if the split succeeded without dropping locks the insert will
 558                  * still be atomic (what the caller peeked() and is overwriting
 559                  * won't have changed)
 560                  */
 561 #if 0
 562                 /*
 563                  * XXX:
 564                  * split -> btree node merging (of parent node) might still drop
 565                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 566                  *
 567                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 568                  * will inhibit merging - but we don't have a reliable way yet
 569                  * (do we?) of checking if we dropped locks in this path
 570                  */
 571                 if (!ret)
 572                         goto retry;
 573 #endif
 574
 575                 /*
 576                  * don't care if we got ENOSPC because we told split it
 577                  * couldn't block:
 578                  */
 579                 if (!ret ||
 580                     ret == -EINTR ||
 581                     (flags & BTREE_INSERT_NOUNLOCK)) {
 582                         trace_trans_restart_btree_node_split(trans->ip);
 583                         ret = -EINTR;
 584                 }
 585                 break;
 586         case BTREE_INSERT_ENOSPC:
 587                 ret = -ENOSPC;
 588                 break;
 589         case BTREE_INSERT_NEED_MARK_REPLICAS:
 590                 bch2_trans_unlock(trans);
 591
 592                 trans_for_each_update(trans, i) {
 593                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 594                         if (ret)
 595                                 return ret;
 596                 }
 597
 598                 if (bch2_trans_relock(trans))
 599                         return 0;
 600
 601                 trace_trans_restart_mark_replicas(trans->ip);
 602                 ret = -EINTR;
 603                 break;
 604         case BTREE_INSERT_NEED_JOURNAL_RES:
 605                 bch2_trans_unlock(trans);
 606
 607                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 608                 if (ret)
 609                         return ret;
 610
 611                 if (bch2_trans_relock(trans))
 612                         return 0;
 613
 614                 trace_trans_restart_journal_res_get(trans->ip);
 615                 ret = -EINTR;
 616                 break;
 617         default:
 618                 BUG_ON(ret >= 0);
 619                 break;
 620         }
 621
 622         if (ret == -EINTR) {
 623                 int ret2 = bch2_btree_iter_traverse_all(trans);
 624
 625                 if (ret2) {
 626                         trace_trans_restart_traverse(trans->ip);
 627                         return ret2;
 628                 }
 629
 630                 trace_trans_restart_atomic(trans->ip);
 631         }
 632
 633         return ret;
 634 }
 635
 636 static noinline int
 637 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 638 {
 639         struct bch_fs *c = trans->c;
 640         int ret;
 641
 642         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 643                 return -EROFS;
 644
 645         bch2_trans_unlock(trans);
 646
 647         ret = bch2_fs_read_write_early(c);
 648         if (ret)
 649                 return ret;
 650
 651         percpu_ref_get(&c->writes);
 652         return 0;
 653 }
 654
 655 static void bch2_trans_update2(struct btree_trans *trans,
 656                                struct btree_iter *iter,
 657                                struct bkey_i *insert)
 658 {
 659         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 660                 .iter = iter, .k = insert
 661         };
 662
 663         btree_insert_entry_checks(trans, n.iter, n.k);
 664
 665         BUG_ON(iter->uptodate > BTREE_ITER_NEED_PEEK);
 666
 667         EBUG_ON(trans->nr_updates2 >= trans->nr_iters);
 668
 669         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 670
 671         trans_for_each_update2(trans, i) {
 672                 if (btree_iter_cmp(n.iter, i->iter) == 0) {
 673                         *i = n;
 674                         return;
 675                 }
 676
 677                 if (btree_iter_cmp(n.iter, i->iter) <= 0)
 678                         break;
 679         }
 680
 681         array_insert_item(trans->updates2, trans->nr_updates2,
 682                           i - trans->updates2, n);
 683 }
 684
 685 static int extent_update_to_keys(struct btree_trans *trans,
 686                                  struct btree_iter *orig_iter,
 687                                  struct bkey_i *insert)
 688 {
 689         struct btree_iter *iter;
 690
 691         if (bkey_deleted(&insert->k))
 692                 return 0;
 693
 694         iter = bch2_trans_copy_iter(trans, orig_iter);
 695         if (IS_ERR(iter))
 696                 return PTR_ERR(iter);
 697
 698         iter->flags |= BTREE_ITER_INTENT;
 699         __bch2_btree_iter_set_pos(iter, insert->k.p, false);
 700         bch2_trans_update2(trans, iter, insert);
 701         bch2_trans_iter_put(trans, iter);
 702         return 0;
 703 }
 704
 705 static int extent_handle_overwrites(struct btree_trans *trans,
 706                                     enum btree_id btree_id,
 707                                     struct bpos start, struct bpos end)
 708 {
 709         struct btree_iter *iter = NULL, *update_iter;
 710         struct bkey_i *update;
 711         struct bkey_s_c k;
 712         int ret = 0;
 713
 714         iter = bch2_trans_get_iter(trans, btree_id, start, BTREE_ITER_INTENT);
 715         ret = PTR_ERR_OR_ZERO(iter);
 716         if (ret)
 717                 return ret;
 718
 719         k = bch2_btree_iter_peek_with_updates(iter);
 720
 721         while (k.k && !(ret = bkey_err(k))) {
 722                 if (bkey_cmp(end, bkey_start_pos(k.k)) <= 0)
 723                         break;
 724
 725                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 726                         update_iter = bch2_trans_copy_iter(trans, iter);
 727                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 728                                 goto err;
 729
 730                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 731                         if ((ret = PTR_ERR_OR_ZERO(update)))
 732                                 goto err;
 733
 734                         bkey_reassemble(update, k);
 735                         bch2_cut_back(start, update);
 736
 737                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 738                         bch2_trans_update2(trans, update_iter, update);
 739                         bch2_trans_iter_put(trans, update_iter);
 740                 }
 741
 742                 if (bkey_cmp(k.k->p, end) > 0) {
 743                         update_iter = bch2_trans_copy_iter(trans, iter);
 744                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 745                                 goto err;
 746
 747                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 748                         if ((ret = PTR_ERR_OR_ZERO(update)))
 749                                 goto err;
 750
 751                         bkey_reassemble(update, k);
 752                         bch2_cut_front(end, update);
 753
 754                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 755                         bch2_trans_update2(trans, update_iter, update);
 756                         bch2_trans_iter_put(trans, update_iter);
 757                 } else {
 758                         update_iter = bch2_trans_copy_iter(trans, iter);
 759                         if ((ret = PTR_ERR_OR_ZERO(update_iter)))
 760                                 goto err;
 761
 762                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 763                         if ((ret = PTR_ERR_OR_ZERO(update)))
 764                                 goto err;
 765
 766                         update->k = *k.k;
 767                         set_bkey_val_u64s(&update->k, 0);
 768                         update->k.type = KEY_TYPE_deleted;
 769                         update->k.size = 0;
 770
 771                         __bch2_btree_iter_set_pos(update_iter, update->k.p, false);
 772                         bch2_trans_update2(trans, update_iter, update);
 773                         bch2_trans_iter_put(trans, update_iter);
 774                 }
 775
 776                 k = bch2_btree_iter_next_with_updates(iter);
 777         }
 778 err:
 779         if (!IS_ERR_OR_NULL(iter))
 780                 bch2_trans_iter_put(trans, iter);
 781         return ret;
 782 }
 783
 784 int __bch2_trans_commit(struct btree_trans *trans)
 785 {
 786         struct btree_insert_entry *i = NULL;
 787         struct btree_iter *iter;
 788         bool trans_trigger_run;
 789         unsigned u64s;
 790         int ret = 0;
 791
 792         BUG_ON(trans->need_reset);
 793
 794         if (!trans->nr_updates)
 795                 goto out_noupdates;
 796
 797         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 798                 lockdep_assert_held(&trans->c->gc_lock);
 799
 800         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 801
 802         trans->journal_u64s             = 0;
 803         trans->journal_preres_u64s      = 0;
 804
 805         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 806             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 807                 ret = bch2_trans_commit_get_rw_cold(trans);
 808                 if (ret)
 809                         return ret;
 810         }
 811
 812         /*
 813          * Running triggers will append more updates to the list of updates as
 814          * we're walking it:
 815          */
 816         do {
 817                 trans_trigger_run = false;
 818
 819                 trans_for_each_update(trans, i) {
 820                         if (unlikely(i->iter->uptodate > BTREE_ITER_NEED_PEEK)) {
 821                                 trace_trans_restart_traverse(trans->ip);
 822                                 ret = -EINTR;
 823                                 goto out;
 824                         }
 825
 826                         /*
 827                          * We're not using bch2_btree_iter_upgrade here because
 828                          * we know trans->nounlock can't be set:
 829                          */
 830                         if (unlikely(i->iter->locks_want < 1 &&
 831                                      !__bch2_btree_iter_upgrade(i->iter, 1))) {
 832                                 trace_trans_restart_upgrade(trans->ip);
 833                                 ret = -EINTR;
 834                                 goto out;
 835                         }
 836
 837                         if (iter_has_trans_triggers(i->iter) &&
 838                             !i->trans_triggers_run) {
 839                                 i->trans_triggers_run = true;
 840                                 trans_trigger_run = true;
 841
 842                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 843                                                              i->trigger_flags);
 844                                 if (unlikely(ret)) {
 845                                         if (ret == -EINTR)
 846                                                 trace_trans_restart_mark(trans->ip);
 847                                         goto out;
 848                                 }
 849                         }
 850                 }
 851         } while (trans_trigger_run);
 852
 853         /* Turn extents updates into keys: */
 854         trans_for_each_update(trans, i)
 855                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 856                         struct bpos start = bkey_start_pos(&i->k->k);
 857
 858                         while (i + 1 < trans->updates + trans->nr_updates &&
 859                                i[0].iter->btree_id == i[1].iter->btree_id &&
 860                                !bkey_cmp(i[0].k->k.p, bkey_start_pos(&i[1].k->k)))
 861                                 i++;
 862
 863                         ret = extent_handle_overwrites(trans, i->iter->btree_id,
 864                                                        start, i->k->k.p);
 865                         if (ret)
 866                                 goto out;
 867                 }
 868
 869         trans_for_each_update(trans, i) {
 870                 if (i->iter->flags & BTREE_ITER_IS_EXTENTS) {
 871                         ret = extent_update_to_keys(trans, i->iter, i->k);
 872                         if (ret)
 873                                 goto out;
 874                 } else {
 875                         bch2_trans_update2(trans, i->iter, i->k);
 876                 }
 877         }
 878
 879         trans_for_each_update2(trans, i) {
 880                 BUG_ON(i->iter->uptodate > BTREE_ITER_NEED_PEEK);
 881                 BUG_ON(i->iter->locks_want < 1);
 882
 883                 u64s = jset_u64s(i->k->k.u64s);
 884                 if (0)
 885                         trans->journal_preres_u64s += u64s;
 886                 trans->journal_u64s += u64s;
 887         }
 888 retry:
 889         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 890
 891         ret = do_bch2_trans_commit(trans, &i);
 892
 893         /* make sure we didn't drop or screw up locks: */
 894         bch2_btree_trans_verify_locks(trans);
 895
 896         if (ret)
 897                 goto err;
 898
 899         trans_for_each_iter(trans, iter)
 900                 if ((trans->iters_live & (1ULL << iter->idx)) &&
 901                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT)) {
 902                         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 903                                 bch2_btree_iter_set_pos_same_leaf(iter, iter->pos_after_commit);
 904                         else
 905                                 bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 906                 }
 907 out:
 908         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 909
 910         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 911                 percpu_ref_put(&trans->c->writes);
 912 out_noupdates:
 913         bch2_trans_reset(trans, !ret ? TRANS_RESET_NOTRAVERSE : 0);
 914
 915         return ret;
 916 err:
 917         ret = bch2_trans_commit_error(trans, i, ret);
 918         if (ret)
 919                 goto out;
 920
 921         goto retry;
 922 }
 923
 924 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 925                       struct bkey_i *k, enum btree_trigger_flags flags)
 926 {
 927         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 928                 .trigger_flags = flags, .iter = iter, .k = k
 929         };
 930
 931         EBUG_ON(bkey_cmp(iter->pos,
 932                          (iter->flags & BTREE_ITER_IS_EXTENTS)
 933                          ? bkey_start_pos(&k->k)
 934                          : k->k.p));
 935
 936         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 937
 938         if (btree_node_type_is_extents(iter->btree_id)) {
 939                 iter->pos_after_commit = k->k.p;
 940                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
 941         }
 942
 943         /*
 944          * Pending updates are kept sorted: first, find position of new update:
 945          */
 946         trans_for_each_update(trans, i)
 947                 if (btree_iter_cmp(iter, i->iter) <= 0)
 948                         break;
 949
 950         /*
 951          * Now delete/trim any updates the new update overwrites:
 952          */
 953         if (i > trans->updates &&
 954             i[-1].iter->btree_id == iter->btree_id &&
 955             bkey_cmp(iter->pos, i[-1].k->k.p) < 0)
 956                 bch2_cut_back(n.iter->pos, i[-1].k);
 957
 958         while (i < trans->updates + trans->nr_updates &&
 959                iter->btree_id == i->iter->btree_id &&
 960                bkey_cmp(n.k->k.p, i->k->k.p) >= 0)
 961                 array_remove_item(trans->updates, trans->nr_updates,
 962                                   i - trans->updates);
 963
 964         if (i < trans->updates + trans->nr_updates &&
 965             iter->btree_id == i->iter->btree_id &&
 966             bkey_cmp(n.k->k.p, i->iter->pos) > 0) {
 967                 /*
 968                  * When we have an extent that overwrites the start of another
 969                  * update, trimming that extent will mean the iterator's
 970                  * position has to change since the iterator position has to
 971                  * match the extent's start pos - but we don't want to change
 972                  * the iterator pos if some other code is using it, so we may
 973                  * need to clone it:
 974                  */
 975                 if (trans->iters_live & (1ULL << i->iter->idx)) {
 976                         i->iter = bch2_trans_copy_iter(trans, i->iter);
 977                         if (IS_ERR(i->iter)) {
 978                                 trans->need_reset = true;
 979                                 return PTR_ERR(i->iter);
 980                         }
 981
 982                         i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 983                         bch2_trans_iter_put(trans, i->iter);
 984                 }
 985
 986                 bch2_cut_front(n.k->k.p, i->k);
 987                 bch2_btree_iter_set_pos(i->iter, n.k->k.p);
 988         }
 989
 990         EBUG_ON(trans->nr_updates >= trans->nr_iters);
 991
 992         array_insert_item(trans->updates, trans->nr_updates,
 993                           i - trans->updates, n);
 994         return 0;
 995 }
 996
 997 int __bch2_btree_insert(struct btree_trans *trans,
 998                         enum btree_id id, struct bkey_i *k)
 999 {
1000         struct btree_iter *iter;
1001         int ret;
1002
1003         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1004                                    BTREE_ITER_INTENT);
1005         if (IS_ERR(iter))
1006                 return PTR_ERR(iter);
1007
1008         ret   = bch2_btree_iter_traverse(iter) ?:
1009                 bch2_trans_update(trans, iter, k, 0);
1010         bch2_trans_iter_put(trans, iter);
1011         return ret;
1012 }
1013
1014 /**
1015  * bch2_btree_insert - insert keys into the extent btree
1016  * @c:                  pointer to struct bch_fs
1017  * @id:                 btree to insert into
1018  * @insert_keys:        list of keys to insert
1019  * @hook:               insert callback
1020  */
1021 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1022                       struct bkey_i *k,
1023                       struct disk_reservation *disk_res,
1024                       u64 *journal_seq, int flags)
1025 {
1026         return bch2_trans_do(c, disk_res, journal_seq, flags,
1027                              __bch2_btree_insert(&trans, id, k));
1028 }
1029
1030 int bch2_btree_delete_at_range(struct btree_trans *trans,
1031                                struct btree_iter *iter,
1032                                struct bpos end,
1033                                u64 *journal_seq)
1034 {
1035         struct bkey_s_c k;
1036         int ret = 0;
1037 retry:
1038         while ((k = bch2_btree_iter_peek(iter)).k &&
1039                !(ret = bkey_err(k)) &&
1040                bkey_cmp(iter->pos, end) < 0) {
1041                 struct bkey_i delete;
1042
1043                 bch2_trans_begin(trans);
1044
1045                 bkey_init(&delete.k);
1046
1047                 /*
1048                  * For extents, iter.pos won't necessarily be the same as
1049                  * bkey_start_pos(k.k) (for non extents they always will be the
1050                  * same). It's important that we delete starting from iter.pos
1051                  * because the range we want to delete could start in the middle
1052                  * of k.
1053                  *
1054                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1055                  * bkey_start_pos(k.k)).
1056                  */
1057                 delete.k.p = iter->pos;
1058
1059                 if (btree_node_type_is_extents(iter->btree_id)) {
1060                         unsigned max_sectors =
1061                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1062
1063                         /* create the biggest key we can */
1064                         bch2_key_resize(&delete.k, max_sectors);
1065                         bch2_cut_back(end, &delete);
1066
1067                         ret = bch2_extent_trim_atomic(&delete, iter);
1068                         if (ret)
1069                                 break;
1070                 }
1071
1072                 bch2_trans_update(trans, iter, &delete, 0);
1073                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1074                                         BTREE_INSERT_NOFAIL);
1075                 if (ret)
1076                         break;
1077
1078                 bch2_trans_cond_resched(trans);
1079         }
1080
1081         if (ret == -EINTR) {
1082                 ret = 0;
1083                 goto retry;
1084         }
1085
1086         return ret;
1087
1088 }
1089
1090 int bch2_btree_delete_at(struct btree_trans *trans,
1091                          struct btree_iter *iter, unsigned flags)
1092 {
1093         struct bkey_i k;
1094
1095         bkey_init(&k.k);
1096         k.k.p = iter->pos;
1097
1098         bch2_trans_update(trans, iter, &k, 0);
1099         return bch2_trans_commit(trans, NULL, NULL,
1100                                  BTREE_INSERT_NOFAIL|
1101                                  BTREE_INSERT_USE_RESERVE|flags);
1102 }
1103
1104 /*
1105  * bch_btree_delete_range - delete everything within a given range
1106  *
1107  * Range is a half open interval - [start, end)
1108  */
1109 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1110                             struct bpos start, struct bpos end,
1111                             u64 *journal_seq)
1112 {
1113         struct btree_trans trans;
1114         struct btree_iter *iter;
1115         int ret = 0;
1116
1117         /*
1118          * XXX: whether we need mem/more iters depends on whether this btree id
1119          * has triggers
1120          */
1121         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
1122
1123         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
1124
1125         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
1126         ret = bch2_trans_exit(&trans) ?: ret;
1127
1128         BUG_ON(ret == -EINTR);
1129         return ret;
1130 }