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 void __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 }
 149
 150 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 151 {
 152         return __btree_node_flush(j, pin, 0, seq);
 153 }
 154
 155 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 156 {
 157         return __btree_node_flush(j, pin, 1, seq);
 158 }
 159
 160 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 161                                        struct btree *b, u64 seq)
 162 {
 163         struct btree_write *w = btree_current_write(b);
 164
 165         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 166                              btree_node_write_idx(b) == 0
 167                              ? btree_node_flush0
 168                              : btree_node_flush1);
 169 }
 170
 171 /**
 172  * btree_insert_key - insert a key one key into a leaf node
 173  */
 174 static bool btree_insert_key_leaf(struct btree_trans *trans,
 175                                   struct btree_iter *iter,
 176                                   struct bkey_i *insert)
 177 {
 178         struct bch_fs *c = trans->c;
 179         struct btree *b = iter_l(iter)->b;
 180         struct bset_tree *t = bset_tree_last(b);
 181         struct bset *i = bset(b, t);
 182         int old_u64s = bset_u64s(t);
 183         int old_live_u64s = b->nr.live_u64s;
 184         int live_u64s_added, u64s_added;
 185
 186         EBUG_ON(!iter->level &&
 187                 !test_bit(BCH_FS_BTREE_INTERIOR_REPLAY_DONE, &c->flags));
 188
 189         if (unlikely(!bch2_btree_bset_insert_key(iter, b,
 190                                         &iter_l(iter)->iter, insert)))
 191                 return false;
 192
 193         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 194                                          le64_to_cpu(i->journal_seq)));
 195
 196         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 197
 198         if (unlikely(!btree_node_dirty(b)))
 199                 set_btree_node_dirty(c, b);
 200
 201         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 202         u64s_added = (int) bset_u64s(t) - old_u64s;
 203
 204         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 205                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 206         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 207                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 208
 209         if (u64s_added > live_u64s_added &&
 210             bch2_maybe_compact_whiteouts(c, b))
 211                 bch2_btree_iter_reinit_node(iter, b);
 212
 213         trace_btree_insert_key(c, b, insert);
 214         return true;
 215 }
 216
 217 /* Cached btree updates: */
 218
 219 /* Normal update interface: */
 220
 221 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 222                                              struct btree_insert_entry *i)
 223 {
 224         struct bch_fs *c = trans->c;
 225
 226         if (bch2_debug_check_bkeys) {
 227                 const char *invalid = bch2_bkey_invalid(c,
 228                                 bkey_i_to_s_c(i->k), i->bkey_type);
 229                 if (invalid) {
 230                         char buf[200];
 231
 232                         bch2_bkey_val_to_text(&PBUF(buf), c, bkey_i_to_s_c(i->k));
 233                         panic("invalid bkey %s on insert: %s\n", buf, invalid);
 234                 }
 235         }
 236         BUG_ON(!i->is_extent && bpos_cmp(i->k->k.p, i->iter->real_pos));
 237         BUG_ON(i->level         != i->iter->level);
 238         BUG_ON(i->btree_id      != i->iter->btree_id);
 239 }
 240
 241 static noinline int
 242 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 243 {
 244         struct bch_fs *c = trans->c;
 245         int ret;
 246
 247         bch2_trans_unlock(trans);
 248
 249         ret = bch2_journal_preres_get(&c->journal,
 250                         &trans->journal_preres, u64s, 0);
 251         if (ret)
 252                 return ret;
 253
 254         if (!bch2_trans_relock(trans)) {
 255                 trace_trans_restart_journal_preres_get(trans->ip);
 256                 return -EINTR;
 257         }
 258
 259         return 0;
 260 }
 261
 262 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 263                                              unsigned flags)
 264 {
 265         struct bch_fs *c = trans->c;
 266         int ret;
 267
 268         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 269                 flags |= JOURNAL_RES_GET_RESERVED;
 270
 271         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 272                                    trans->journal_u64s, flags);
 273
 274         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 275 }
 276
 277 static enum btree_insert_ret
 278 btree_key_can_insert(struct btree_trans *trans,
 279                      struct btree_iter *iter,
 280                      unsigned u64s)
 281 {
 282         struct bch_fs *c = trans->c;
 283         struct btree *b = iter_l(iter)->b;
 284
 285         if (!bch2_btree_node_insert_fits(c, b, u64s))
 286                 return BTREE_INSERT_BTREE_NODE_FULL;
 287
 288         return BTREE_INSERT_OK;
 289 }
 290
 291 static enum btree_insert_ret
 292 btree_key_can_insert_cached(struct btree_trans *trans,
 293                             struct btree_iter *iter,
 294                             unsigned u64s)
 295 {
 296         struct bkey_cached *ck = (void *) iter->l[0].b;
 297         unsigned new_u64s;
 298         struct bkey_i *new_k;
 299
 300         BUG_ON(iter->level);
 301
 302         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 303             bch2_btree_key_cache_must_wait(trans->c) &&
 304             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 305                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 306
 307         if (u64s <= ck->u64s)
 308                 return BTREE_INSERT_OK;
 309
 310         new_u64s        = roundup_pow_of_two(u64s);
 311         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 312         if (!new_k)
 313                 return -ENOMEM;
 314
 315         ck->u64s        = new_u64s;
 316         ck->k           = new_k;
 317         return BTREE_INSERT_OK;
 318 }
 319
 320 static inline void do_btree_insert_one(struct btree_trans *trans,
 321                                        struct btree_iter *iter,
 322                                        struct bkey_i *insert)
 323 {
 324         struct bch_fs *c = trans->c;
 325         struct journal *j = &c->journal;
 326         bool did_work;
 327
 328         EBUG_ON(trans->journal_res.ref !=
 329                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 330
 331         insert->k.needs_whiteout = false;
 332
 333         did_work = (btree_iter_type(iter) != BTREE_ITER_CACHED)
 334                 ? btree_insert_key_leaf(trans, iter, insert)
 335                 : bch2_btree_insert_key_cached(trans, iter, insert);
 336         if (!did_work)
 337                 return;
 338
 339         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 340                 bch2_journal_add_keys(j, &trans->journal_res,
 341                                       iter->btree_id, insert);
 342
 343                 bch2_journal_set_has_inode(j, &trans->journal_res,
 344                                            insert->k.p.inode);
 345
 346                 if (trans->journal_seq)
 347                         *trans->journal_seq = trans->journal_res.seq;
 348         }
 349 }
 350
 351 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 352 {
 353         __bch2_btree_iter_unlock(iter);
 354 }
 355
 356 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct btree_insert_entry *i;
 360
 361         trans_for_each_update(trans, i) {
 362                 /*
 363                  * XXX: synchronization of cached update triggers with gc
 364                  */
 365                 BUG_ON(btree_iter_type(i->iter) == BTREE_ITER_CACHED);
 366
 367                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 368                         bch2_mark_update(trans, i->iter, i->k, NULL,
 369                                          i->trigger_flags|BTREE_TRIGGER_GC);
 370         }
 371 }
 372
 373 static inline int
 374 bch2_trans_commit_write_locked(struct btree_trans *trans,
 375                                struct btree_insert_entry **stopped_at)
 376 {
 377         struct bch_fs *c = trans->c;
 378         struct bch_fs_usage *fs_usage = NULL;
 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                 fs_usage = bch2_fs_usage_scratch_get(c);
 427         }
 428
 429         /*
 430          * Don't get journal reservation until after we know insert will
 431          * succeed:
 432          */
 433         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 434                 ret = bch2_trans_journal_res_get(trans,
 435                                 JOURNAL_RES_GET_NONBLOCK);
 436                 if (ret)
 437                         goto err;
 438         } else {
 439                 trans->journal_res.seq = c->journal.replay_journal_seq;
 440         }
 441
 442         if (unlikely(trans->extra_journal_entry_u64s)) {
 443                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 444                                   trans->extra_journal_entries,
 445                                   trans->extra_journal_entry_u64s);
 446
 447                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 448                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 449         }
 450
 451         /*
 452          * Not allowed to fail after we've gotten our journal reservation - we
 453          * have to use it:
 454          */
 455
 456         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 457                 if (bch2_journal_seq_verify)
 458                         trans_for_each_update2(trans, i)
 459                                 i->k->k.version.lo = trans->journal_res.seq;
 460                 else if (bch2_inject_invalid_keys)
 461                         trans_for_each_update2(trans, i)
 462                                 i->k->k.version = MAX_VERSION;
 463         }
 464
 465         /* Must be called under mark_lock: */
 466         if (marking && trans->fs_usage_deltas &&
 467             bch2_replicas_delta_list_apply(c, fs_usage,
 468                                            trans->fs_usage_deltas)) {
 469                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 470                 goto err;
 471         }
 472
 473         trans_for_each_update(trans, i)
 474                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type))
 475                         bch2_mark_update(trans, i->iter, i->k,
 476                                          fs_usage, i->trigger_flags);
 477
 478         if (marking)
 479                 bch2_trans_fs_usage_apply(trans, fs_usage);
 480
 481         if (unlikely(c->gc_pos.phase))
 482                 bch2_trans_mark_gc(trans);
 483
 484         trans_for_each_update2(trans, i)
 485                 do_btree_insert_one(trans, i->iter, i->k);
 486 err:
 487         if (marking) {
 488                 bch2_fs_usage_scratch_put(c, fs_usage);
 489                 percpu_up_read(&c->mark_lock);
 490         }
 491
 492         return ret;
 493 }
 494
 495 /*
 496  * Get journal reservation, take write locks, and attempt to do btree update(s):
 497  */
 498 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 499                                        struct btree_insert_entry **stopped_at)
 500 {
 501         struct btree_insert_entry *i;
 502         struct btree_iter *iter;
 503         int ret;
 504
 505         trans_for_each_update2(trans, i)
 506                 BUG_ON(!btree_node_intent_locked(i->iter, i->iter->level));
 507
 508         ret = bch2_journal_preres_get(&trans->c->journal,
 509                         &trans->journal_preres, trans->journal_preres_u64s,
 510                         JOURNAL_RES_GET_NONBLOCK|
 511                         ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM)
 512                          ? JOURNAL_RES_GET_RECLAIM : 0));
 513         if (unlikely(ret == -EAGAIN))
 514                 ret = bch2_trans_journal_preres_get_cold(trans,
 515                                                 trans->journal_preres_u64s);
 516         if (unlikely(ret))
 517                 return ret;
 518
 519         /*
 520          * Can't be holding any read locks when we go to take write locks:
 521          * another thread could be holding an intent lock on the same node we
 522          * have a read lock on, and it'll block trying to take a write lock
 523          * (because we hold a read lock) and it could be blocking us by holding
 524          * its own read lock (while we're trying to to take write locks).
 525          *
 526          * note - this must be done after bch2_trans_journal_preres_get_cold()
 527          * or anything else that might call bch2_trans_relock(), since that
 528          * would just retake the read locks:
 529          */
 530         trans_for_each_iter(trans, iter) {
 531                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 532                         if (btree_iter_keep(trans, iter)) {
 533                                 if (!bch2_btree_iter_upgrade(iter, 1)) {
 534                                         trace_trans_restart_upgrade(trans->ip);
 535                                         return -EINTR;
 536                                 }
 537                         } else {
 538                                 bch2_btree_iter_unlock_noinline(iter);
 539                         }
 540                 }
 541         }
 542
 543         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 544                 trans_for_each_update2(trans, i)
 545                         btree_insert_entry_checks(trans, i);
 546         bch2_btree_trans_verify_locks(trans);
 547
 548         trans_for_each_update2(trans, i)
 549                 if (!same_leaf_as_prev(trans, i))
 550                         bch2_btree_node_lock_for_insert(trans->c,
 551                                         iter_l(i->iter)->b, i->iter);
 552
 553         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 554
 555         trans_for_each_update2(trans, i)
 556                 if (!same_leaf_as_prev(trans, i))
 557                         bch2_btree_node_unlock_write_inlined(iter_l(i->iter)->b,
 558                                                              i->iter);
 559
 560         if (!ret && trans->journal_pin)
 561                 bch2_journal_pin_add(&trans->c->journal, trans->journal_res.seq,
 562                                      trans->journal_pin, NULL);
 563
 564         /*
 565          * Drop journal reservation after dropping write locks, since dropping
 566          * the journal reservation may kick off a journal write:
 567          */
 568         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 569
 570         if (unlikely(ret))
 571                 return ret;
 572
 573         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 574                 trans->nounlock = true;
 575
 576         if (!(trans->flags & BTREE_INSERT_NOUNLOCK))
 577         trans_for_each_update2(trans, i)
 578                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 579                     !same_leaf_as_prev(trans, i))
 580                         bch2_foreground_maybe_merge(trans->c, i->iter,
 581                                                     0, trans->flags);
 582
 583         trans->nounlock = false;
 584
 585         bch2_trans_downgrade(trans);
 586
 587         return 0;
 588 }
 589
 590 static noinline
 591 int bch2_trans_commit_error(struct btree_trans *trans,
 592                             struct btree_insert_entry *i,
 593                             int ret)
 594 {
 595         struct bch_fs *c = trans->c;
 596         unsigned flags = trans->flags;
 597
 598         /*
 599          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 600          * update; if we haven't done anything yet it doesn't apply
 601          */
 602         flags &= ~BTREE_INSERT_NOUNLOCK;
 603
 604         switch (ret) {
 605         case BTREE_INSERT_BTREE_NODE_FULL:
 606                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 607
 608                 /*
 609                  * if the split succeeded without dropping locks the insert will
 610                  * still be atomic (what the caller peeked() and is overwriting
 611                  * won't have changed)
 612                  */
 613 #if 0
 614                 /*
 615                  * XXX:
 616                  * split -> btree node merging (of parent node) might still drop
 617                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 618                  *
 619                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 620                  * will inhibit merging - but we don't have a reliable way yet
 621                  * (do we?) of checking if we dropped locks in this path
 622                  */
 623                 if (!ret)
 624                         goto retry;
 625 #endif
 626
 627                 /*
 628                  * don't care if we got ENOSPC because we told split it
 629                  * couldn't block:
 630                  */
 631                 if (!ret ||
 632                     ret == -EINTR ||
 633                     (flags & BTREE_INSERT_NOUNLOCK)) {
 634                         trace_trans_restart_btree_node_split(trans->ip);
 635                         ret = -EINTR;
 636                 }
 637                 break;
 638         case BTREE_INSERT_ENOSPC:
 639                 ret = -ENOSPC;
 640                 break;
 641         case BTREE_INSERT_NEED_MARK_REPLICAS:
 642                 bch2_trans_unlock(trans);
 643
 644                 trans_for_each_update(trans, i) {
 645                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 646                         if (ret)
 647                                 return ret;
 648                 }
 649
 650                 if (bch2_trans_relock(trans))
 651                         return 0;
 652
 653                 trace_trans_restart_mark_replicas(trans->ip);
 654                 ret = -EINTR;
 655                 break;
 656         case BTREE_INSERT_NEED_JOURNAL_RES:
 657                 bch2_trans_unlock(trans);
 658
 659                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 660                 if (ret)
 661                         return ret;
 662
 663                 if (bch2_trans_relock(trans))
 664                         return 0;
 665
 666                 trace_trans_restart_journal_res_get(trans->ip);
 667                 ret = -EINTR;
 668                 break;
 669         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 670                 bch2_trans_unlock(trans);
 671
 672                 do {
 673                         mutex_lock(&c->journal.reclaim_lock);
 674                         ret = bch2_journal_reclaim(&c->journal);
 675                         mutex_unlock(&c->journal.reclaim_lock);
 676                 } while (!ret && bch2_btree_key_cache_must_wait(c));
 677
 678                 if (!ret && bch2_trans_relock(trans))
 679                         return 0;
 680
 681                 trace_trans_restart_journal_reclaim(trans->ip);
 682                 ret = -EINTR;
 683                 break;
 684         default:
 685                 BUG_ON(ret >= 0);
 686                 break;
 687         }
 688
 689         return ret;
 690 }
 691
 692 static noinline int
 693 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 694 {
 695         struct bch_fs *c = trans->c;
 696         int ret;
 697
 698         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 699                 return -EROFS;
 700
 701         bch2_trans_unlock(trans);
 702
 703         ret = bch2_fs_read_write_early(c);
 704         if (ret)
 705                 return ret;
 706
 707         percpu_ref_get(&c->writes);
 708         return 0;
 709 }
 710
 711 static void __bch2_trans_update2(struct btree_trans *trans,
 712                                  struct btree_insert_entry n)
 713 {
 714         struct btree_insert_entry *i;
 715
 716         btree_insert_entry_checks(trans, &n);
 717
 718         EBUG_ON(trans->nr_updates2 >= BTREE_ITER_MAX);
 719
 720         n.iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
 721
 722         trans_for_each_update2(trans, i)
 723                 if (btree_insert_entry_cmp(&n, i) <= 0)
 724                         break;
 725
 726         if (i < trans->updates2 + trans->nr_updates2 &&
 727             !btree_insert_entry_cmp(&n, i))
 728                 *i = n;
 729         else
 730                 array_insert_item(trans->updates2, trans->nr_updates2,
 731                                   i - trans->updates2, n);
 732 }
 733
 734 static void bch2_trans_update2(struct btree_trans *trans,
 735                                struct btree_iter *iter,
 736                                struct bkey_i *insert)
 737 {
 738         __bch2_trans_update2(trans, (struct btree_insert_entry) {
 739                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 740                 .btree_id       = iter->btree_id,
 741                 .level          = iter->level,
 742                 .iter           = iter,
 743                 .k              = insert,
 744         });
 745 }
 746
 747 static int extent_update_to_keys(struct btree_trans *trans,
 748                                  struct btree_insert_entry n)
 749 {
 750         int ret;
 751
 752         if (bkey_deleted(&n.k->k))
 753                 return 0;
 754
 755         ret = bch2_extent_can_insert(trans, n.iter, n.k);
 756         if (ret)
 757                 return ret;
 758
 759         n.iter = bch2_trans_get_iter(trans, n.iter->btree_id, n.k->k.p,
 760                                      BTREE_ITER_INTENT|
 761                                      BTREE_ITER_NOT_EXTENTS);
 762         n.is_extent = false;
 763
 764         __bch2_trans_update2(trans, n);
 765         bch2_trans_iter_put(trans, n.iter);
 766         return 0;
 767 }
 768
 769 static int extent_handle_overwrites(struct btree_trans *trans,
 770                                     enum btree_id btree_id,
 771                                     struct bkey_i *insert)
 772 {
 773         struct btree_iter *iter, *update_iter;
 774         struct bpos start = bkey_start_pos(&insert->k);
 775         struct bkey_i *update;
 776         struct bkey_s_c k;
 777         int ret = 0;
 778
 779         iter = bch2_trans_get_iter(trans, btree_id, start,
 780                                    BTREE_ITER_INTENT);
 781         k = bch2_btree_iter_peek_with_updates(iter);
 782
 783         while (k.k && !(ret = bkey_err(k))) {
 784                 if (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) <= 0)
 785                         break;
 786
 787                 if (bkey_cmp(bkey_start_pos(k.k), start) < 0) {
 788                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 789                         if ((ret = PTR_ERR_OR_ZERO(update)))
 790                                 break;
 791
 792                         bkey_reassemble(update, k);
 793
 794                         bch2_cut_back(start, update);
 795
 796                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 797                                                           BTREE_ITER_NOT_EXTENTS|
 798                                                           BTREE_ITER_INTENT);
 799                         bch2_trans_update2(trans, update_iter, update);
 800                         bch2_trans_iter_put(trans, update_iter);
 801                 }
 802
 803                 if (bkey_cmp(k.k->p, insert->k.p) < 0 ||
 804                     (!bkey_cmp(k.k->p, insert->k.p) && bkey_deleted(&insert->k))) {
 805                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey));
 806                         if ((ret = PTR_ERR_OR_ZERO(update)))
 807                                 break;
 808
 809                         bkey_init(&update->k);
 810                         update->k.p = k.k->p;
 811
 812                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 813                                                           BTREE_ITER_NOT_EXTENTS|
 814                                                           BTREE_ITER_INTENT);
 815                         bch2_trans_update2(trans, update_iter, update);
 816                         bch2_trans_iter_put(trans, update_iter);
 817                 }
 818
 819                 if (bkey_cmp(k.k->p, insert->k.p) > 0) {
 820                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
 821                         if ((ret = PTR_ERR_OR_ZERO(update)))
 822                                 break;
 823
 824                         bkey_reassemble(update, k);
 825                         bch2_cut_front(insert->k.p, update);
 826
 827                         update_iter = bch2_trans_get_iter(trans, btree_id, update->k.p,
 828                                                           BTREE_ITER_NOT_EXTENTS|
 829                                                           BTREE_ITER_INTENT);
 830                         bch2_trans_update2(trans, update_iter, update);
 831                         bch2_trans_iter_put(trans, update_iter);
 832                         break;
 833                 }
 834
 835                 k = bch2_btree_iter_next_with_updates(iter);
 836         }
 837         bch2_trans_iter_put(trans, iter);
 838
 839         return ret;
 840 }
 841
 842 int __bch2_trans_commit(struct btree_trans *trans)
 843 {
 844         struct btree_insert_entry *i = NULL;
 845         struct btree_iter *iter;
 846         bool trans_trigger_run;
 847         unsigned u64s, reset_flags = 0;
 848         int ret = 0;
 849
 850         if (!trans->nr_updates)
 851                 goto out_reset;
 852
 853         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 854                 lockdep_assert_held(&trans->c->gc_lock);
 855
 856         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 857
 858         trans->journal_u64s             = trans->extra_journal_entry_u64s;
 859         trans->journal_preres_u64s      = 0;
 860
 861         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 862             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 863                 ret = bch2_trans_commit_get_rw_cold(trans);
 864                 if (ret)
 865                         goto out_reset;
 866         }
 867
 868 #ifdef CONFIG_BCACHEFS_DEBUG
 869         trans_for_each_update(trans, i)
 870                 if (btree_iter_type(i->iter) != BTREE_ITER_CACHED &&
 871                     !(i->trigger_flags & BTREE_TRIGGER_NORUN))
 872                         bch2_btree_key_cache_verify_clean(trans,
 873                                         i->btree_id, i->k->k.p);
 874 #endif
 875
 876         /*
 877          * Running triggers will append more updates to the list of updates as
 878          * we're walking it:
 879          */
 880         do {
 881                 trans_trigger_run = false;
 882
 883                 trans_for_each_update(trans, i) {
 884                         if ((BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 885                             !i->trans_triggers_run) {
 886                                 i->trans_triggers_run = true;
 887                                 trans_trigger_run = true;
 888
 889                                 ret = bch2_trans_mark_update(trans, i->iter, i->k,
 890                                                              i->trigger_flags);
 891                                 if (unlikely(ret)) {
 892                                         if (ret == -EINTR)
 893                                                 trace_trans_restart_mark(trans->ip);
 894                                         goto out;
 895                                 }
 896                         }
 897                 }
 898         } while (trans_trigger_run);
 899
 900         /* Turn extents updates into keys: */
 901         trans_for_each_update(trans, i)
 902                 if (i->is_extent) {
 903                         ret = extent_handle_overwrites(trans, i->btree_id, i->k);
 904                         if (unlikely(ret))
 905                                 goto out;
 906                 }
 907
 908         trans_for_each_update(trans, i) {
 909                 ret = i->is_extent
 910                         ? extent_update_to_keys(trans, *i)
 911                         : (__bch2_trans_update2(trans, *i), 0);
 912                 if (unlikely(ret))
 913                         goto out;
 914         }
 915
 916         trans_for_each_update2(trans, i) {
 917                 ret = bch2_btree_iter_traverse(i->iter);
 918                 if (unlikely(ret)) {
 919                         trace_trans_restart_traverse(trans->ip);
 920                         goto out;
 921                 }
 922
 923                 /*
 924                  * We're not using bch2_btree_iter_upgrade here because
 925                  * we know trans->nounlock can't be set:
 926                  */
 927                 if (unlikely(!btree_node_intent_locked(i->iter, i->iter->level) &&
 928                              !__bch2_btree_iter_upgrade(i->iter, i->iter->level + 1))) {
 929                         trace_trans_restart_upgrade(trans->ip);
 930                         ret = -EINTR;
 931                         goto out;
 932                 }
 933
 934                 u64s = jset_u64s(i->k->k.u64s);
 935                 if (btree_iter_type(i->iter) == BTREE_ITER_CACHED &&
 936                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
 937                         trans->journal_preres_u64s += u64s;
 938                 trans->journal_u64s += u64s;
 939         }
 940 retry:
 941         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 942
 943         ret = do_bch2_trans_commit(trans, &i);
 944
 945         /* make sure we didn't drop or screw up locks: */
 946         bch2_btree_trans_verify_locks(trans);
 947
 948         if (ret)
 949                 goto err;
 950
 951         trans_for_each_iter(trans, iter)
 952                 if (btree_iter_live(trans, iter) &&
 953                     (iter->flags & BTREE_ITER_SET_POS_AFTER_COMMIT))
 954                         bch2_btree_iter_set_pos(iter, iter->pos_after_commit);
 955 out:
 956         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 957
 958         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 959                 percpu_ref_put(&trans->c->writes);
 960 out_reset:
 961         if (!ret)
 962                 reset_flags |= TRANS_RESET_NOTRAVERSE;
 963         if (!ret && (trans->flags & BTREE_INSERT_NOUNLOCK))
 964                 reset_flags |= TRANS_RESET_NOUNLOCK;
 965         bch2_trans_reset(trans, reset_flags);
 966
 967         return ret;
 968 err:
 969         ret = bch2_trans_commit_error(trans, i, ret);
 970         if (ret)
 971                 goto out;
 972
 973         goto retry;
 974 }
 975
 976 int bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 977                       struct bkey_i *k, enum btree_trigger_flags flags)
 978 {
 979         struct btree_insert_entry *i, n = (struct btree_insert_entry) {
 980                 .trigger_flags  = flags,
 981                 .bkey_type      = __btree_node_type(iter->level, iter->btree_id),
 982                 .btree_id       = iter->btree_id,
 983                 .level          = iter->level,
 984                 .is_extent      = (iter->flags & BTREE_ITER_IS_EXTENTS) != 0,
 985                 .iter           = iter,
 986                 .k              = k
 987         };
 988
 989         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
 990
 991 #ifdef CONFIG_BCACHEFS_DEBUG
 992         BUG_ON(bkey_cmp(iter->pos,
 993                         n.is_extent ? bkey_start_pos(&k->k) : k->k.p));
 994
 995         trans_for_each_update(trans, i) {
 996                 BUG_ON(bkey_cmp(i->iter->pos,
 997                                 i->is_extent ? bkey_start_pos(&i->k->k) : i->k->k.p));
 998
 999                 BUG_ON(i != trans->updates &&
1000                        btree_insert_entry_cmp(i - 1, i) >= 0);
1001         }
1002 #endif
1003
1004         iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1005
1006         if (n.is_extent) {
1007                 iter->pos_after_commit = k->k.p;
1008                 iter->flags |= BTREE_ITER_SET_POS_AFTER_COMMIT;
1009         }
1010
1011         /*
1012          * Pending updates are kept sorted: first, find position of new update,
1013          * then delete/trim any updates the new update overwrites:
1014          */
1015         if (!n.is_extent) {
1016                 trans_for_each_update(trans, i)
1017                         if (btree_insert_entry_cmp(&n, i) <= 0)
1018                                 break;
1019
1020                 if (i < trans->updates + trans->nr_updates &&
1021                     !btree_insert_entry_cmp(&n, i))
1022                         *i = n;
1023                 else
1024                         array_insert_item(trans->updates, trans->nr_updates,
1025                                           i - trans->updates, n);
1026         } else {
1027                 trans_for_each_update(trans, i)
1028                         if (btree_insert_entry_cmp(&n, i) < 0)
1029                                 break;
1030
1031                 while (i > trans->updates &&
1032                        i[-1].btree_id == n.btree_id &&
1033                        bkey_cmp(bkey_start_pos(&n.k->k),
1034                                 bkey_start_pos(&i[-1].k->k)) <= 0) {
1035                         --i;
1036                         array_remove_item(trans->updates, trans->nr_updates,
1037                                           i - trans->updates);
1038                 }
1039
1040                 if (i > trans->updates &&
1041                     i[-1].btree_id == n.btree_id &&
1042                     bkey_cmp(bkey_start_pos(&n.k->k), i[-1].k->k.p) < 0)
1043                         bch2_cut_back(bkey_start_pos(&n.k->k), i[-1].k);
1044
1045                 if (i < trans->updates + trans->nr_updates &&
1046                     i->btree_id == n.btree_id &&
1047                     bkey_cmp(n.k->k.p, bkey_start_pos(&i->k->k)) > 0) {
1048                         /* We don't handle splitting extents here: */
1049                         BUG_ON(bkey_cmp(bkey_start_pos(&n.k->k),
1050                                         bkey_start_pos(&i->k->k)) > 0);
1051
1052                         /*
1053                          * When we have an extent that overwrites the start of another
1054                          * update, trimming that extent will mean the iterator's
1055                          * position has to change since the iterator position has to
1056                          * match the extent's start pos - but we don't want to change
1057                          * the iterator pos if some other code is using it, so we may
1058                          * need to clone it:
1059                          */
1060                         if (btree_iter_live(trans, i->iter)) {
1061                                 i->iter = bch2_trans_copy_iter(trans, i->iter);
1062
1063                                 i->iter->flags |= BTREE_ITER_KEEP_UNTIL_COMMIT;
1064                                 bch2_trans_iter_put(trans, i->iter);
1065                         }
1066
1067                         bch2_cut_front(n.k->k.p, i->k);
1068                         bch2_btree_iter_set_pos(i->iter, n.k->k.p);
1069                 }
1070
1071                 array_insert_item(trans->updates, trans->nr_updates,
1072                                   i - trans->updates, n);
1073         }
1074
1075         return 0;
1076 }
1077
1078 void bch2_trans_commit_hook(struct btree_trans *trans,
1079                             struct btree_trans_commit_hook *h)
1080 {
1081         h->next = trans->hooks;
1082         trans->hooks = h;
1083 }
1084
1085 int __bch2_btree_insert(struct btree_trans *trans,
1086                         enum btree_id id, struct bkey_i *k)
1087 {
1088         struct btree_iter *iter;
1089         int ret;
1090
1091         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
1092                                    BTREE_ITER_INTENT);
1093
1094         ret   = bch2_btree_iter_traverse(iter) ?:
1095                 bch2_trans_update(trans, iter, k, 0);
1096         bch2_trans_iter_put(trans, iter);
1097         return ret;
1098 }
1099
1100 /**
1101  * bch2_btree_insert - insert keys into the extent btree
1102  * @c:                  pointer to struct bch_fs
1103  * @id:                 btree to insert into
1104  * @insert_keys:        list of keys to insert
1105  * @hook:               insert callback
1106  */
1107 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1108                       struct bkey_i *k,
1109                       struct disk_reservation *disk_res,
1110                       u64 *journal_seq, int flags)
1111 {
1112         return bch2_trans_do(c, disk_res, journal_seq, flags,
1113                              __bch2_btree_insert(&trans, id, k));
1114 }
1115
1116 int bch2_btree_delete_at(struct btree_trans *trans,
1117                          struct btree_iter *iter, unsigned flags)
1118 {
1119         struct bkey_i k;
1120
1121         bkey_init(&k.k);
1122         k.k.p = iter->pos;
1123
1124         bch2_trans_update(trans, iter, &k, 0);
1125         return bch2_trans_commit(trans, NULL, NULL,
1126                                  BTREE_INSERT_NOFAIL|flags);
1127 }
1128
1129 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1130                                   struct bpos start, struct bpos end,
1131                                   u64 *journal_seq)
1132 {
1133         struct btree_iter *iter;
1134         struct bkey_s_c k;
1135         int ret = 0;
1136
1137         iter = bch2_trans_get_iter(trans, id, start, BTREE_ITER_INTENT);
1138 retry:
1139         while ((k = bch2_btree_iter_peek(iter)).k &&
1140                !(ret = bkey_err(k)) &&
1141                bkey_cmp(iter->pos, end) < 0) {
1142                 struct bkey_i delete;
1143
1144                 bch2_trans_begin(trans);
1145
1146                 bkey_init(&delete.k);
1147
1148                 /*
1149                  * This could probably be more efficient for extents:
1150                  */
1151
1152                 /*
1153                  * For extents, iter.pos won't necessarily be the same as
1154                  * bkey_start_pos(k.k) (for non extents they always will be the
1155                  * same). It's important that we delete starting from iter.pos
1156                  * because the range we want to delete could start in the middle
1157                  * of k.
1158                  *
1159                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1160                  * bkey_start_pos(k.k)).
1161                  */
1162                 delete.k.p = iter->pos;
1163
1164                 if (btree_node_type_is_extents(iter->btree_id)) {
1165                         unsigned max_sectors =
1166                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1167
1168                         /* create the biggest key we can */
1169                         bch2_key_resize(&delete.k, max_sectors);
1170                         bch2_cut_back(end, &delete);
1171
1172                         ret = bch2_extent_trim_atomic(&delete, iter);
1173                         if (ret)
1174                                 break;
1175                 }
1176
1177                 bch2_trans_update(trans, iter, &delete, 0);
1178                 ret = bch2_trans_commit(trans, NULL, journal_seq,
1179                                         BTREE_INSERT_NOFAIL);
1180                 if (ret)
1181                         break;
1182
1183                 bch2_trans_cond_resched(trans);
1184         }
1185
1186         if (ret == -EINTR) {
1187                 ret = 0;
1188                 goto retry;
1189         }
1190
1191         bch2_trans_iter_put(trans, iter);
1192         return ret;
1193 }
1194
1195 /*
1196  * bch_btree_delete_range - delete everything within a given range
1197  *
1198  * Range is a half open interval - [start, end)
1199  */
1200 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1201                             struct bpos start, struct bpos end,
1202                             u64 *journal_seq)
1203 {
1204         return bch2_trans_do(c, NULL, journal_seq, 0,
1205                              bch2_btree_delete_range_trans(&trans, id, start, end, journal_seq));
1206 }