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_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "buckets.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "replicas.h"
  19
  20 #include <linux/prefetch.h>
  21 #include <linux/sort.h>
  22 #include <trace/events/bcachefs.h>
  23
  24 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  25                                          const struct btree_insert_entry *r)
  26 {
  27         return   cmp_int(l->btree_id,   r->btree_id) ?:
  28                  -cmp_int(l->level,     r->level) ?:
  29                  bpos_cmp(l->k->k.p,    r->k->k.p);
  30 }
  31
  32 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  33                                      struct btree_insert_entry *i)
  34 {
  35         return i != trans->updates2 &&
  36                 iter_l(i[0].iter)->b == iter_l(i[-1].iter)->b;
  37 }
  38
  39 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  40                                             struct btree_iter *iter)
  41 {
  42         bch2_btree_node_lock_write(b, iter);
  43
  44         if (btree_iter_type(iter) == BTREE_ITER_CACHED)
  45                 return;
  46
  47         if (unlikely(btree_node_just_written(b)) &&
  48             bch2_btree_post_write_cleanup(c, b))
  49                 bch2_btree_iter_reinit_node(iter, b);
  50
  51         /*
  52          * If the last bset has been written, or if it's gotten too big - start
  53          * a new bset to insert into:
  54          */
  55         if (want_new_bset(c, b))
  56                 bch2_btree_init_next(c, b, iter);
  57 }
  58
  59 /* Inserting into a given leaf node (last stage of insert): */
  60
  61 /* Handle overwrites and do insert, for non extents: */
  62 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  63                                 struct btree *b,
  64                                 struct btree_node_iter *node_iter,
  65                                 struct bkey_i *insert)
  66 {
  67         struct bkey_packed *k;
  68         unsigned clobber_u64s = 0, new_u64s = 0;
  69
  70         EBUG_ON(btree_node_just_written(b));
  71         EBUG_ON(bset_written(b, btree_bset_last(b)));
  72         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  73         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
  74         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  75         EBUG_ON(insert->k.u64s >
  76                 bch_btree_keys_u64s_remaining(iter->trans->c, b));
  77         EBUG_ON(iter->flags & BTREE_ITER_IS_EXTENTS);
  78
  79         k = bch2_btree_node_iter_peek_all(node_iter, b);
  80         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
  81                 k = NULL;
  82
  83         /* @k is the key being overwritten/deleted, if any: */
  84         EBUG_ON(k && bkey_deleted(k));
  85
  86         /* Deleting, but not found? nothing to do: */
  87         if (bkey_deleted(&insert->k) && !k)
  88                 return false;
  89
  90         if (bkey_deleted(&insert->k)) {
  91                 /* Deleting: */
  92                 btree_account_key_drop(b, k);
  93                 k->type = KEY_TYPE_deleted;
  94
  95                 if (k->needs_whiteout)
  96                         push_whiteout(iter->trans->c, b, insert->k.p);
  97                 k->needs_whiteout = false;
  98
  99                 if (k >= btree_bset_last(b)->start) {
 100                         clobber_u64s = k->u64s;
 101                         bch2_bset_delete(b, k, clobber_u64s);
 102                         goto fix_iter;
 103                 } else {
 104                         bch2_btree_iter_fix_key_modified(iter, b, k);
 105                 }
 106
 107                 return true;
 108         }
 109
 110         if (k) {
 111                 /* Overwriting: */
 112                 btree_account_key_drop(b, k);
 113                 k->type = KEY_TYPE_deleted;
 114
 115                 insert->k.needs_whiteout = k->needs_whiteout;
 116                 k->needs_whiteout = false;
 117
 118                 if (k >= btree_bset_last(b)->start) {
 119                         clobber_u64s = k->u64s;
 120                         goto overwrite;
 121                 } else {
 122                         bch2_btree_iter_fix_key_modified(iter, b, k);
 123                 }
 124         }
 125
 126         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 127 overwrite:
 128         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 129         new_u64s = k->u64s;
 130 fix_iter:
 131         if (clobber_u64s != new_u64s)
 132                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 133                                          clobber_u64s, new_u64s);
 134         return true;
 135 }
 136
 137 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 138                                unsigned i, u64 seq)
 139 {
 140         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 141         struct btree_write *w = container_of(pin, struct btree_write, journal);
 142         struct btree *b = container_of(w, struct btree, writes[i]);
 143
 144         btree_node_lock_type(c, b, SIX_LOCK_read);
 145         bch2_btree_node_write_cond(c, b,
 146                 (btree_current_write(b) == w && w->journal.seq == seq));
 147         six_unlock_read(&b->c.lock);
 148         return 0;
 149 }
 150
 151 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 152 {
 153         return __btree_node_flush(j, pin, 0, seq);
 154 }
 155
 156 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 157 {
 158         return __btree_node_flush(j, pin, 1, seq);
 159 }
 160
 161 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 162                                        struct btree *b, u64 seq)
 163 {
 164         struct btree_write *w = btree_current_write(b);
 165
 166         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 167                              btree_node_write_idx(b) == 0
 168                              ? btree_node_flush0
 169                              : btree_node_flush1);
 170 }
 171
 172 /**
 173  * btree_insert_key - insert a key one key into a leaf node
 174  */
 175 static bool btree_insert_key_leaf(struct btree_trans *trans,
 176                                   struct btree_iter *iter,
 177                                   struct bkey_i *insert)
 178 {
 179         struct bch_fs *c = trans->c;
 180         struct btree *b = iter_l(iter)->b;
 181         struct bset_tree *t = bset_tree_last(b);
 182         struct bset *i = bset(b, t);
 183         int old_u64s = bset_u64s(t);
 184         int old_live_u64s = b->nr.live_u64s;
 185         int live_u64s_added, u64s_added;
 186
 187         EBUG_ON(!iter->level &&
 188                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 189
 190         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 191                                         &iter_l(iter)->iter, insert)))
 192                 return false;
 193
 194         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 195                                          le64_to_cpu(i->journal_seq)));
 196
 197         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 198
 199         if (unlikely(!btree_node_dirty(b)))
 200                 set_btree_node_dirty(c, b);
 201
 202         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 203         u64s_added = (int) bset_u64s(t) - old_u64s;
 204
 205         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 206                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 207         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 208                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 209
 210         if (u64s_added > live_u64s_added &&
 211             bch2_maybe_compact_whiteouts(c, b))
 212                 bch2_btree_iter_reinit_node(iter, b);
 213
 214         trace_btree_insert_key(c, b, insert);
 215         return true;
 216 }
 217
 218 /* Cached btree updates: */
 219
 220 /* Normal update interface: */
 221
 222 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 223                                              struct btree_insert_entry *i)
 224 {
 225         struct bch_fs *c = trans->c;
 226
 227         if (bch2_debug_check_bkeys) {
 228                 const char *invalid = bch2_bkey_invalid(c,
 229                                 bkey_i_to_s_c(i->k), i->bkey_type);
 230                 if (invalid) {
 231                         char buf[200];
 232
 233                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 234                         panic("invalid bkey %s on insert: %s\n", buf, invalid);
 235                 }
 236         }
 237         BUG_ON(!i->is_extent && bpos_cmp(i->k->k.p, i->iter->real_pos));
 238         BUG_ON(i->level         != i->iter->level);
 239         BUG_ON(i->btree_id      != i->iter->btree_id);
 240 }
 241
 242 static noinline int
 243 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 244 {
 245         struct bch_fs *c = trans->c;
 246         int ret;
 247
 248         bch2_trans_unlock(trans);
 249
 250         ret = bch2_journal_preres_get(&c->journal,
 251                         &trans->journal_preres, u64s, 0);
 252         if (ret)
 253                 return ret;
 254
 255         if (!bch2_trans_relock(trans)) {
 256                 trace_trans_restart_journal_preres_get(trans->ip);
 257                 return -EINTR;
 258         }
 259
 260         return 0;
 261 }
 262
 263 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 264                                              unsigned flags)
 265 {
 266         struct bch_fs *c = trans->c;
 267         int ret;
 268
 269         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 270                 flags |= JOURNAL_RES_GET_RESERVED;
 271
 272         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 273                                    trans->journal_u64s, flags);
 274
 275         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 276 }
 277
 278 static enum btree_insert_ret
 279 btree_key_can_insert(struct btree_trans *trans,
 280                      struct btree_iter *iter,
 281                      unsigned u64s)
 282 {
 283         struct bch_fs *c = trans->c;
 284         struct btree *b = iter_l(iter)->b;
 285
 286         if (!bch2_btree_node_insert_fits(c, b, u64s))
 287                 return BTREE_INSERT_BTREE_NODE_FULL;
 288
 289         return BTREE_INSERT_OK;
 290 }
 291
 292 static enum btree_insert_ret
 293 btree_key_can_insert_cached(struct btree_trans *trans,
 294                             struct btree_iter *iter,
 295                             unsigned u64s)
 296 {
 297         struct bkey_cached *ck = (void *) iter->l[0].b;
 298         unsigned new_u64s;
 299         struct bkey_i *new_k;
 300
 301         BUG_ON(iter->level);
 302
 303         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 304             bch2_btree_key_cache_must_wait(trans->c) &&
 305             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 306                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 307
 308         if (u64s <= ck->u64s)
 309                 return BTREE_INSERT_OK;
 310
 311         new_u64s        = roundup_pow_of_two(u64s);
 312         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 313         if (!new_k)
 314                 return -ENOMEM;
 315
 316         ck->u64s        = new_u64s;
 317         ck->k           = new_k;
 318         return BTREE_INSERT_OK;
 319 }
 320
 321 static inline void do_btree_insert_one(struct btree_trans *trans,
 322                                        struct btree_iter *iter,
 323                                        struct bkey_i *insert)
 324 {
 325         struct bch_fs *c = trans->c;
 326         struct journal *j = &c->journal;
 327         bool did_work;
 328
 329         EBUG_ON(trans->journal_res.ref !=
 330                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 331
 332         insert->k.needs_whiteout = false;
 333
 334         did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
 335                 ? btree_insert_key_leaf(trans, iter, insert)
 336                 : bch2_btree_insert_key_cached(trans, iter, insert);
 337         if (!did_work)
 338                 return;
 339
 340         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 341                 bch2_journal_add_keys(j, &trans->journal_res,
 342                                       iter->btree_id, insert);
 343
 344                 bch2_journal_set_has_inode(j, &trans->journal_res,
 345                                            insert->k.p.inode);
 346
 347                 if (trans->journal_seq)
 348                         *trans->journal_seq = trans->journal_res.seq;
 349         }
 350 }
 351
 352 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 353 {
 354         __bch2_btree_iter_unlock(iter);
 355 }
 356
 357 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 358 {
 359         struct bch_fs *c = trans->c;
 360         struct btree_insert_entry *i;
 361
 362         trans_for_each_update(trans, i) {
 363                 /*
 364                  * XXX: synchronization of cached update triggers with gc
 365                  */
 366                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 367
 368                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 369                         bch2_mark_update(trans, i->iter, i->k, NULL,
 370                                          i->trigger_flags|BTREE_TRIGGER_GC);
 371         }
 372 }
 373
 374 static inline int
 375 bch2_trans_commit_write_locked(struct btree_trans *trans,
 376                                struct btree_insert_entry **stopped_at)
 377 {
 378         struct bch_fs *c = trans->c;
 379         struct btree_insert_entry *i;
 380         struct btree_trans_commit_hook *h;
 381         unsigned u64s = 0;
 382         bool marking = false;
 383         int ret;
 384
 385         if (race_fault()) {
 386                 trace_trans_restart_fault_inject(trans->ip);
 387                 return -EINTR;
 388         }
 389
 390         /*
 391          * Check if the insert will fit in the leaf node with the write lock
 392          * held, otherwise another thread could write the node changing the
 393          * amount of space available:
 394          */
 395
 396         prefetch(&trans->c->journal.flags);
 397
 398         h = trans->hooks;
 399         while (h) {
 400                 ret = h->fn(trans, h);
 401                 if (ret)
 402                         return ret;
 403                 h = h->next;
 404         }
 405
 406         trans_for_each_update2(trans, i) {
 407                 /* Multiple inserts might go to same leaf: */
 408                 if (!same_leaf_as_prev(trans, i))
 409                         u64s = 0;
 410
 411                 u64s += i->k->k.u64s;
 412                 ret = btree_iter_type(i->iter) != BTREE_ITER_CACHED
 413                         ? btree_key_can_insert(trans, i->iter, u64s)
 414                         : btree_key_can_insert_cached(trans, i->iter, u64s);
 415                 if (ret) {
 416                         *stopped_at = i;
 417                         return ret;
 418                 }
 419
 420                 if (btree_node_type_needs_gc(i->bkey_type))
 421                         marking = true;
 422         }
 423
 424         if (marking) {
 425                 percpu_down_read(&c->mark_lock);
 426         }
 427
 428         /* Must be called under mark_lock: */
 429         if (marking && trans->fs_usage_deltas &&
 430             !bch2_replicas_delta_list_marked(c, trans->fs_usage_deltas)) {
 431                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 432                 goto err;
 433         }
 434
 435         /*
 436          * Don't get journal reservation until after we know insert will
 437          * succeed:
 438          */
 439         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 440                 ret = bch2_trans_journal_res_get(trans,
 441                                 JOURNAL_RES_GET_NONBLOCK);
 442                 if (ret)
 443                         goto err;
 444         } else {
 445                 trans->journal_res.seq = c->journal.replay_journal_seq;
 446         }
 447
 448         if (unlikely(trans->extra_journal_entry_u64s)) {
 449                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 450                                   trans->extra_journal_entries,
 451                                   trans->extra_journal_entry_u64s);
 452
 453                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 454                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 455         }
 456
 457         /*
 458          * Not allowed to fail after we've gotten our journal reservation - we
 459          * have to use it:
 460          */
 461
 462         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 463                 if (bch2_journal_seq_verify)
 464                         trans_for_each_update2(trans, i)
 465                                 i->k->k.version.lo = trans->journal_res.seq;
 466                 else if (bch2_inject_invalid_keys)
 467                         trans_for_each_update2(trans, i)
 468                                 i->k->k.version = MAX_VERSION;
 469         }
 470
 471         trans_for_each_update(trans, i)
 472                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 473                         bch2_mark_update(trans, i->iter, i->k,
 474                                          NULL, i->trigger_flags);
 475
 476         if (marking && trans->fs_usage_deltas)
 477                 bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas);
 478
 479         if (unlikely(c->gc_pos.phase))
 480                 bch2_trans_mark_gc(trans);
 481
 482         trans_for_each_update2(trans, i)
 483                 do_btree_insert_one(trans, i->iter, i->k);
 484 err:
 485         if (marking) {
 486                 percpu_up_read(&c->mark_lock);
 487         }
 488
 489         return ret;
 490 }
 491
 492 static noinline int maybe_do_btree_merge(struct btree_trans *trans, struct btree_iter *iter)
 493 {
 494         struct btree_insert_entry *i;
 495         struct btree *b = iter_l(iter)->b;
 496         struct bkey_s_c old;
 497         int u64s_delta = 0;
 498         int ret;
 499
 500         /*
 501          * Inserting directly into interior nodes is an uncommon operation with
 502          * various weird edge cases: also, a lot of things about
 503          * BTREE_ITER_NODES iters need to be audited
 504          */
 505         if (unlikely(btree_iter_type(iter) != BTREE_ITER_KEYS))
 506                 return 0;
 507
 508         BUG_ON(iter->level);
 509
 510         trans_for_each_update2(trans, i) {
 511                 if (iter_l(i->iter)->b != b)
 512                         continue;
 513
 514                 old = bch2_btree_iter_peek_slot(i->iter);
 515                 ret = bkey_err(old);
 516                 if (ret)
 517                         return ret;
 518
 519                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 520                 u64s_delta -= !bkey_deleted(old.k) ? old.k->u64s : 0;
 521         }
 522
 523         return u64s_delta <= 0
 524                 ? (bch2_foreground_maybe_merge(trans->c, iter, iter->level,
 525                                 trans->flags & ~BTREE_INSERT_NOUNLOCK) ?: -EINTR)
 526                 : 0;
 527 }
 528
 529 /*
 530  * Get journal reservation, take write locks, and attempt to do btree update(s):
 531  */
 532 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 533                                        struct btree_insert_entry **stopped_at)
 534 {
 535         struct bch_fs *c = trans->c;
 536         struct btree_insert_entry *i;
 537         struct btree_iter *iter;
 538         int ret;
 539
 540         trans_for_each_update2(trans, i) {
 541                 struct btree *b;
 542
 543                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 544
 545                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED)
 546                         continue;
 547
 548                 b = iter_l(i->iter)->b;
 549                 if (b->sib_u64s[0] < c->btree_foreground_merge_threshold ||
 550                     b->sib_u64s[1] < c->btree_foreground_merge_threshold) {
 551                         ret = maybe_do_btree_merge(trans, i->iter);
 552                         if (unlikely(ret))
 553                                 return ret;
 554                 }
 555         }
 556
 557         trans_for_each_update2(trans, i)
 558                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 559
 560         ret = bch2_journal_preres_get(&c->journal,
 561                         &trans->journal_preres, trans->journal_preres_u64s,
 562                         JOURNAL_RES_GET_NONBLOCK|
 563                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 564                          ? JOURNAL_RES_GET_RESERVED : 0));
 565         if (unlikely(ret == -EAGAIN))
 566                 ret = bch2_trans_journal_preres_get_cold(trans,
 567                                                 trans->journal_preres_u64s);
 568         if (unlikely(ret))
 569                 return ret;
 570
 571         /*
 572          * Can't be holding any read locks when we go to take write locks:
 573          * another thread could be holding an intent lock on the same node we
 574          * have a read lock on, and it'll block trying to take a write lock
 575          * (because we hold a read lock) and it could be blocking us by holding
 576          * its own read lock (while we're trying to to take write locks).
 577          *
 578          * note - this must be done after bch2_trans_journal_preres_get_cold()
 579          * or anything else that might call bch2_trans_relock(), since that
 580          * would just retake the read locks:
 581          */
 582         trans_for_each_iter(trans, iter) {
 583                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 584                         if (btree_iter_keep(trans, iter)) {
 585                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 586                                         trace_trans_restart_upgrade(trans->ip);
 587                                         return -EINTR;
 588                                 }
 589                         } else {
 590                                 bch2_btree_iter_unlock_noinline(iter);
 591                         }
 592                 }
 593         }
 594
 595         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 596                 trans_for_each_update2(trans, i)
 597                         btree_insert_entry_checks(trans, i);
 598         bch2_btree_trans_verify_locks(trans);
 599
 600         trans_for_each_update2(trans, i)
 601                 if (!same_leaf_as_prev(trans, i))
 602                         bch2_btree_node_lock_for_insert(c,
 603                                         iter_l(i->iter)->b, i->iter);
 604
 605         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 606
 607         trans_for_each_update2(trans, i)
 608                 if (!same_leaf_as_prev(trans, i))
 609                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 610                                                              i->iter);
 611
 612         if (!ret && trans->journal_pin)
 613                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 614                                      trans->journal_pin, NULL);
 615
 616         /*
 617          * Drop journal reservation after dropping write locks, since dropping
 618          * the journal reservation may kick off a journal write:
 619          */
 620         bch2_journal_res_put(&c->journal, &trans->journal_res);
 621
 622         if (unlikely(ret))
 623                 return ret;
 624
 625         bch2_trans_downgrade(trans);
 626
 627         return 0;
 628 }
 629
 630 static int journal_reclaim_wait_done(struct bch_fs *c)
 631 {
 632         int ret;
 633
 634         ret = bch2_journal_error(&c->journal);
 635         if (ret)
 636                 return ret;
 637
 638         ret = !bch2_btree_key_cache_must_wait(c);
 639         if (ret)
 640                 return ret;
 641
 642         if (mutex_trylock(&c->journal.reclaim_lock)) {
 643                 ret = bch2_journal_reclaim(&c->journal);
 644                 mutex_unlock(&c->journal.reclaim_lock);
 645         }
 646
 647         if (!ret)
 648                 ret = !bch2_btree_key_cache_must_wait(c);
 649         return ret;
 650 }
 651
 652 static noinline
 653 int bch2_trans_commit_error(struct btree_trans *trans,
 654                             struct btree_insert_entry *i,
 655                             int ret)
 656 {
 657         struct bch_fs *c = trans->c;
 658         unsigned flags = trans->flags;
 659
 660         /*
 661          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 662          * update; if we haven't done anything yet it doesn't apply
 663          */
 664         flags &= ~BTREE_INSERT_NOUNLOCK;
 665
 666         switch (ret) {
 667         case BTREE_INSERT_BTREE_NODE_FULL:
 668                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 669
 670                 /*
 671                  * if the split succeeded without dropping locks the insert will
 672                  * still be atomic (what the caller peeked() and is overwriting
 673                  * won't have changed)
 674                  */
 675 #if 0
 676                 /*
 677                  * XXX:
 678                  * split -> btree node merging (of parent node) might still drop
 679                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 680                  *
 681                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 682                  * will inhibit merging - but we don't have a reliable way yet
 683                  * (do we?) of checking if we dropped locks in this path
 684                  */
 685                 if (!ret)
 686                         goto retry;
 687 #endif
 688
 689                 /*
 690                  * don't care if we got ENOSPC because we told split it
 691                  * couldn't block:
 692                  */
 693                 if (!ret ||
 694                     ret == -EINTR ||
 695                     (flags & BTREE_INSERT_NOUNLOCK)) {
 696                         trace_trans_restart_btree_node_split(trans->ip);
 697                         ret = -EINTR;
 698                 }
 699                 break;
 700         case BTREE_INSERT_ENOSPC:
 701                 ret = -ENOSPC;
 702                 break;
 703         case BTREE_INSERT_NEED_MARK_REPLICAS:
 704                 bch2_trans_unlock(trans);
 705
 706                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 707                 if (ret)
 708                         return ret;
 709
 710                 if (bch2_trans_relock(trans))
 711                         return 0;
 712
 713                 trace_trans_restart_mark_replicas(trans->ip);
 714                 ret = -EINTR;
 715                 break;
 716         case BTREE_INSERT_NEED_JOURNAL_RES:
 717                 bch2_trans_unlock(trans);
 718
 719                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 720                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED))
 721                         return -EAGAIN;
 722
 723                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 724                 if (ret)
 725                         return ret;
 726
 727                 if (bch2_trans_relock(trans))
 728                         return 0;
 729
 730                 trace_trans_restart_journal_res_get(trans->ip);
 731                 ret = -EINTR;
 732                 break;
 733         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 734                 bch2_trans_unlock(trans);
 735
 736                 wait_event(c->journal.reclaim_wait,
 737                            (ret = journal_reclaim_wait_done(c)));
 738
 739                 if (!ret && bch2_trans_relock(trans))
 740                         return 0;
 741
 742                 trace_trans_restart_journal_reclaim(trans->ip);
 743                 ret = -EINTR;
 744                 break;
 745         default:
 746                 BUG_ON(ret >= 0);
 747                 break;
 748         }
 749
 750         return ret;
 751 }
 752
 753 static noinline int
 754 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 755 {
 756         struct bch_fs *c = trans->c;
 757         int ret;
 758
 759         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 760                 return -EROFS;
 761
 762         bch2_trans_unlock(trans);
 763
 764         ret = bch2_fs_read_write_early(c);
 765         if (ret)
 766                 return ret;
 767
 768         percpu_ref_get(&c->writes);
 769         return 0;
 770 }
 771
 772 static void __bch2_trans_update2(struct btree_trans *trans,
 773                                  struct btree_insert_entry n)
 774 {
 775         struct btree_insert_entry *i;
 776
 777         btree_insert_entry_checks(trans, &n);
 778
 779         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 780
 781         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 782
 783         trans_for_each_update2(trans, i)
 784                 if (btree_insert_entry_cmp(&n, i) <= 0)
 785                         break;
 786
 787         if (i < trans->updates2 + trans->nr_updates2 &&
 788             !btree_insert_entry_cmp(&n, i))
 789                 *i = n;
 790         else
 791                 array_insert_item(trans->updates2, trans->nr_updates2,
 792                                   i - trans->updates2, n);
 793 }
 794
 795 static void bch2_trans_update2(struct btree_trans *trans,
 796                                struct btree_iter *iter,
 797                                struct bkey_i *insert)
 798 {
 799         __bch2_trans_update2(trans, (struct btree_insert_entry) {
 800                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 801                 .btree_id       = iter->btree_id,
 802                 .level          = iter->level,
 803                 .iter           = iter,
 804                 .k              = insert,
 805         });
 806 }
 807
 808 static int extent_update_to_keys(struct btree_trans *trans,
 809                                  struct btree_insert_entry n)
 810 {
 811         int ret;
 812
 813         if (bkey_deleted(&n.k->k))
 814                 return 0;
 815
 816         ret = bch2_extent_can_insert(trans, n.iter, n.k);
 817         if (ret)
 818                 return ret;
 819
 820         n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p,
 821                                      BTREE_ITER_INTENT|
 822                                      BTREE_ITER_NOT_EXTENTS);
 823         n.is_extent = false;
 824
 825         __bch2_trans_update2(trans, n);
 826         bch2_trans_iter_put(trans, n.iter);
 827         return 0;
 828 }
 829
 830 static int extent_handle_overwrites(struct btree_trans *trans,
 831                                     enum btree_id btree_id,
 832                                     struct bkey_i *insert)
 833 {
 834         struct btree_iter *iter, *update_iter;
 835         struct bpos start = bkey_start_pos(&insert->k);
 836         struct bkey_i *update;
 837         struct bkey_s_c k;
 838         int ret = 0;
 839
 840         iter = bch2_trans_get_iter(trans, btree_id, start,
 841                                    BTREE_ITER_INTENT);
 842         k = bch2_btree_iter_peek_with_updates(iter);
 843
 844         while (k.k && !(ret = bkey_err(k))) {
 845                 if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
 846                         break;
 847
 848                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 849                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 850                         if ((ret = PTR_ERR_OR_ZERO(update)))
 851                                 break;
 852
 853                         bkey_reassemble(update, k);
 854
 855                         bch2_cut_back(start, update);
 856
 857                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 858                                                           BTREE_ITER_NOT_EXTENTS|
 859                                                           BTREE_ITER_INTENT);
 860                         bch2_trans_update2(trans, update_iter, update);
 861                         bch2_trans_iter_put(trans, update_iter);
 862                 }
 863
 864                 if (bkey_cmp(k.k->p, insert->k.p) < 0 ||
 865                     (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) {
 866                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 867                         if ((ret = PTR_ERR_OR_ZERO(update)))
 868                                 break;
 869
 870                         bkey_init(&update->k);
 871                         update->k.p = k.k->p;
 872
 873                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 874                                                           BTREE_ITER_NOT_EXTENTS|
 875                                                           BTREE_ITER_INTENT);
 876                         bch2_trans_update2(trans, update_iter, update);
 877                         bch2_trans_iter_put(trans, update_iter);
 878                 }
 879
 880                 if (bkey_cmp(k.k->p, insert->k.p) > 0) {
 881                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 882                         if ((ret = PTR_ERR_OR_ZERO(update)))
 883                                 break;
 884
 885                         bkey_reassemble(update, k);
 886                         bch2_cut_front(insert->k.p, update);
 887
 888                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 889                                                           BTREE_ITER_NOT_EXTENTS|
 890                                                           BTREE_ITER_INTENT);
 891                         bch2_trans_update2(trans, update_iter, update);
 892                         bch2_trans_iter_put(trans, update_iter);
 893                         break;
 894                 }
 895
 896                 k = bch2_btree_iter_next_with_updates(iter);
 897         }
 898         bch2_trans_iter_put(trans, iter);
 899
 900         return ret;
 901 }
 902
 903 int __bch2_trans_commit(struct btree_trans *trans)
 904 {
 905         struct btree_insert_entry *i = NULL;
 906         struct btree_iter *iter;
 907         bool trans_trigger_run;
 908         unsigned u64s, reset_flags = 0;
 909         int ret = 0;
 910
 911         if (!trans->nr_updates)
 912                 goto out_reset;
 913
 914         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 915                 lockdep_assert_held(&trans->c->gc_lock);
 916
 917         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 918
 919         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 920         trans->journal_preres_u64s      = 0;
 921
 922         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 923             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 924                 ret = bch2_trans_commit_get_rw_cold(trans);
 925                 if (ret)
 926                         goto out_reset;
 927         }
 928
 929 #ifdef CONFIG_BCACHEFS_DEBUG
 930         trans_for_each_update(trans, i)
 931                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 932                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 933                         bch2_btree_key_cache_verify_clean(trans,
 934                                         i->btree_id, i->k->k.p);
 935 #endif
 936
 937         /*
 938          * Running triggers will append more updates to the list of updates as
 939          * we're walking it:
 940          */
 941         do {
 942                 trans_trigger_run = false;
 943
 944                 trans_for_each_update(trans, i) {
 945                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 946                             !i->trans_triggers_run) {
 947                                 i->trans_triggers_run = true;
 948                                 trans_trigger_run = true;
 949
 950                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 951                                                              i->trigger_flags);
 952                                 if (unlikely(ret)) {
 953                                         if (ret == -EINTR)
 954                                                 trace_trans_restart_mark(trans->ip);
 955                                         goto out;
 956                                 }
 957                         }
 958                 }
 959         } while (trans_trigger_run);
 960
 961         /* Turn extents updates into keys: */
 962         trans_for_each_update(trans, i)
 963                 if (i->is_extent) {
 964                         ret = extent_handle_overwrites(trans, i->btree_id, i->k);
 965                         if (unlikely(ret))
 966                                 goto out;
 967                 }
 968
 969         trans_for_each_update(trans, i) {
 970                 ret = i->is_extent
 971                         ? extent_update_to_keys(trans, *i)
 972                         : (__bch2_trans_update2(trans, *i), 0);
 973                 if (unlikely(ret))
 974                         goto out;
 975         }
 976
 977         trans_for_each_update2(trans, i) {
 978                 ret = bch2_btree_iter_traverse(i->iter);
 979                 if (unlikely(ret)) {
 980                         trace_trans_restart_traverse(trans->ip);
 981                         goto out;
 982                 }
 983
 984                 if (unlikely(!bch2_btree_iter_upgrade(i->iter, i->level + 1))) {
 985                         trace_trans_restart_upgrade(trans->ip);
 986                         ret = -EINTR;
 987                         goto out;
 988                 }
 989
 990                 BUG_ON(!btree_node_intent_locked(i->iter, i->level));
 991
 992                 u64s = jset_u64s(i->k->k.u64s);
 993                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 994                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 995                         trans->journal_preres_u64s += u64s;
 996                 trans->journal_u64s += u64s;
 997         }
 998 retry:
 999         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1000
1001         ret = do_bch2_trans_commit(trans, &i);
1002
1003         /* make sure we didn't drop or screw up locks: */
1004         bch2_btree_trans_verify_locks(trans);
1005
1006         if (ret)
1007                 goto err;
1008
1009         trans_for_each_iter(trans, iter)
1010                 if (btree_iter_live(trans, iter) &&
1011                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
1012                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
1013 out:
1014         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
1015
1016         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1017                 percpu_ref_put(&trans->c->writes);
1018 out_reset:
1019         if (!ret)
1020                 reset_flags |= TRANS_RESET_NOTRAVERSE;
1021         if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
1022                 reset_flags |= TRANS_RESET_NOUNLOCK;
1023         bch2_trans_reset(trans, reset_flags);
1024
1025         return ret;
1026 err:
1027         ret = bch2_trans_commit_error(trans, i, ret);
1028         if (ret)
1029                 goto out;
1030
1031         goto retry;
1032 }
1033
1034 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1035                       struct bkey_i *k, enum btree_trigger_flags flags)
1036 {
1037         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
1038                 .trigger_flags  = flags,
1039                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
1040                 .btree_id       = iter->btree_id,
1041                 .level          = iter->level,
1042                 .is_extent      = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0,
1043                 .iter           = iter,
1044                 .k              = k
1045         };
1046
1047         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1048
1049 #ifdef CONFIG_BCACHEFS_DEBUG
1050         BUG_ON(bkey_cmp(iter->pos,
1051                         n.is_extent ? bkey_start_pos(&k->k) : k->k.p));
1052
1053         trans_for_each_update(trans, i) {
1054                 BUG_ON(bkey_cmp(i->iter->pos,
1055                                 i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p));
1056
1057                 BUG_ON(i != trans->updates &&
1058                        btree_insert_entry_cmp(i - 1, i) >= 0);
1059         }
1060 #endif
1061
1062         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1063
1064         if (n.is_extent) {
1065                 iter->pos_after_commit = k->k.p;
1066                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1067         }
1068
1069         /*
1070          * Pending updates are kept sorted: first, find position of new update,
1071          * then delete/trim any updates the new update overwrites:
1072          */
1073         if (!n.is_extent) {
1074                 trans_for_each_update(trans, i)
1075                         if (btree_insert_entry_cmp(&n, i) <= 0)
1076                                 break;
1077
1078                 if (i < trans->updates + trans->nr_updates &&
1079                     !btree_insert_entry_cmp(&n, i))
1080                         *i = n;
1081                 else
1082                         array_insert_item(trans->updates, trans->nr_updates,
1083                                           i - trans->updates, n);
1084         } else {
1085                 trans_for_each_update(trans, i)
1086                         if (btree_insert_entry_cmp(&n, i) < 0)
1087                                 break;
1088
1089                 while (i > trans->updates &&
1090                        i[-1].btree_id == n.btree_id &&
1091                        bkey_cmp(bkey_start_pos(&n.k->k),
1092                                 bkey_start_pos(&i[-1].k->k)) <= 0) {
1093                         --i;
1094                         array_remove_item(trans->updates, trans->nr_updates,
1095                                           i - trans->updates);
1096                 }
1097
1098                 if (i > trans->updates &&
1099                     i[-1].btree_id == n.btree_id &&
1100                     bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0)
1101                         bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k);
1102
1103                 if (i < trans->updates + trans->nr_updates &&
1104                     i->btree_id == n.btree_id &&
1105                     bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) {
1106                         /* We don't handle splitting extents here: */
1107                         BUG_ON(bkey_cmp(bkey_start_pos(&n.k->k),
1108                                         bkey_start_pos(&i->k->k)) > 0);
1109
1110                         /*
1111                          * When we have an extent that overwrites the start of another
1112                          * update, trimming that extent will mean the iterator's
1113                          * position has to change since the iterator position has to
1114                          * match the extent's start pos - but we don't want to change
1115                          * the iterator pos if some other code is using it, so we may
1116                          * need to clone it:
1117                          */
1118                         if (btree_iter_live(trans, i->iter)) {
1119                                 i->iter = bch2_trans_copy_iter(trans, i->iter);
1120
1121                                 i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1122                                 bch2_trans_iter_put(trans, i->iter);
1123                         }
1124
1125                         bch2_cut_front(n.k->k.p, i->k);
1126                         bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1127                 }
1128
1129                 array_insert_item(trans->updates, trans->nr_updates,
1130                                   i - trans->updates, n);
1131         }
1132
1133         return 0;
1134 }
1135
1136 void bch2_trans_commit_hook(struct btree_trans *trans,
1137                             struct btree_trans_commit_hook *h)
1138 {
1139         h->next = trans->hooks;
1140         trans->hooks = h;
1141 }
1142
1143 int __bch2_btree_insert(struct btree_trans *trans,
1144                         enum btree_id id, struct bkey_i *k)
1145 {
1146         struct btree_iter *iter;
1147         int ret;
1148
1149         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1150                                    BTREE_ITER_INTENT);
1151
1152         ret   = bch2_btree_iter_traverse(iter) ?:
1153                 bch2_trans_update(trans, iter, k, 0);
1154         bch2_trans_iter_put(trans, iter);
1155         return ret;
1156 }
1157
1158 /**
1159  * bch2_btree_insert - insert keys into the extent btree
1160  * @c:                  pointer to struct bch_fs
1161  * @id:                 btree to insert into
1162  * @insert_keys:        list of keys to insert
1163  * @hook:               insert callback
1164  */
1165 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1166                       struct bkey_i *k,
1167                       struct disk_reservation *disk_res,
1168                       u64 *journal_seq, int flags)
1169 {
1170         return bch2_trans_do(c, disk_res, journal_seq, flags,
1171                              __bch2_btree_insert(&trans, id, k));
1172 }
1173
1174 int bch2_btree_delete_at(struct btree_trans *trans,
1175                          struct btree_iter *iter, unsigned flags)
1176 {
1177         struct bkey_i k;
1178
1179         bkey_init(&k.k);
1180         k.k.p = iter->pos;
1181
1182         bch2_trans_update(trans, iter, &k, 0);
1183         return bch2_trans_commit(trans, NULL, NULL,
1184                                  BTREE_INSERT_NOFAIL|flags);
1185 }
1186
1187 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1188                                   struct bpos start, struct bpos end,
1189                                   u64 *journal_seq)
1190 {
1191         struct btree_iter *iter;
1192         struct bkey_s_c k;
1193         int ret = 0;
1194
1195         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1196 retry:
1197         while ((k = bch2_btree_iter_peek(iter)).k &&
1198                !(ret = bkey_err(k)) &&
1199                bkey_cmp(iter->pos, end) < 0) {
1200                 struct bkey_i delete;
1201
1202                 bch2_trans_begin(trans);
1203
1204                 bkey_init(&delete.k);
1205
1206                 /*
1207                  * This could probably be more efficient for extents:
1208                  */
1209
1210                 /*
1211                  * For extents, iter.pos won't necessarily be the same as
1212                  * bkey_start_pos(k.k) (for non extents they always will be the
1213                  * same). It's important that we delete starting from iter.pos
1214                  * because the range we want to delete could start in the middle
1215                  * of k.
1216                  *
1217                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1218                  * bkey_start_pos(k.k)).
1219                  */
1220                 delete.k.p = iter->pos;
1221
1222                 if (btree_node_type_is_extents(iter->btree_id)) {
1223                         unsigned max_sectors =
1224                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1225
1226                         /* create the biggest key we can */
1227                         bch2_key_resize(&delete.k, max_sectors);
1228                         bch2_cut_back(end, &delete);
1229
1230                         ret = bch2_extent_trim_atomic(&delete, iter);
1231                         if (ret)
1232                                 break;
1233                 }
1234
1235                 bch2_trans_update(trans, iter, &delete, 0);
1236                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1237                                         BTREE_INSERT_NOFAIL);
1238                 if (ret)
1239                         break;
1240
1241                 bch2_trans_cond_resched(trans);
1242         }
1243
1244         if (ret == -EINTR) {
1245                 ret = 0;
1246                 goto retry;
1247         }
1248
1249         bch2_trans_iter_free(trans, iter);
1250         return ret;
1251 }
1252
1253 /*
1254  * bch_btree_delete_range - delete everything within a given range
1255  *
1256  * Range is a half open interval - [start, end)
1257  */
1258 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1259                             struct bpos start, struct bpos end,
1260                             u64 *journal_seq)
1261 {
1262         return bch2_trans_do(c, NULL, journal_seq, 0,
1263                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1264 }