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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_locking.h"
  10 #include "buckets.h"
  11 #include "debug.h"
  12 #include "error.h"
  13 #include "extent_update.h"
  14 #include "journal.h"
  15 #include "journal_reclaim.h"
  16 #include "keylist.h"
  17 #include "replicas.h"
  18
  19 #include <linux/prefetch.h>
  20 #include <linux/sort.h>
  21 #include <trace/events/bcachefs.h>
  22
  23 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  24                                      unsigned idx)
  25 {
  26         return idx &&
  27                 trans->updates[trans->updates_sorted[idx]].iter->l[0].b ==
  28                 trans->updates[trans->updates_sorted[idx - 1]].iter->l[0].b;
  29 }
  30
  31 #define trans_for_each_update_sorted(_trans, _i, _iter)                 \
  32         for (_iter = 0;                                                 \
  33              _iter < _trans->nr_updates &&                              \
  34              (_i = _trans->updates + _trans->updates_sorted[_iter], 1); \
  35              _iter++)
  36
  37 inline void bch2_btree_node_lock_for_insert(struct bch_fs *c, struct btree *b,
  38                                             struct btree_iter *iter)
  39 {
  40         bch2_btree_node_lock_write(b, iter);
  41
  42         if (unlikely(btree_node_just_written(b)) &&
  43             bch2_btree_post_write_cleanup(c, b))
  44                 bch2_btree_iter_reinit_node(iter, b);
  45
  46         /*
  47          * If the last bset has been written, or if it's gotten too big - start
  48          * a new bset to insert into:
  49          */
  50         if (want_new_bset(c, b))
  51                 bch2_btree_init_next(c, b, iter);
  52 }
  53
  54 static inline void btree_trans_sort_updates(struct btree_trans *trans)
  55 {
  56         struct btree_insert_entry *l, *r;
  57         unsigned nr = 0, pos;
  58
  59         trans_for_each_update(trans, l) {
  60                 for (pos = 0; pos < nr; pos++) {
  61                         r = trans->updates + trans->updates_sorted[pos];
  62
  63                         if (btree_iter_cmp(l->iter, r->iter) <= 0)
  64                                 break;
  65                 }
  66
  67                 memmove(&trans->updates_sorted[pos + 1],
  68                         &trans->updates_sorted[pos],
  69                         (nr - pos) * sizeof(trans->updates_sorted[0]));
  70
  71                 trans->updates_sorted[pos] = l - trans->updates;
  72                 nr++;
  73         }
  74 }
  75
  76 /* Inserting into a given leaf node (last stage of insert): */
  77
  78 /* Handle overwrites and do insert, for non extents: */
  79 bool bch2_btree_bset_insert_key(struct btree_iter *iter,
  80                                 struct btree *b,
  81                                 struct btree_node_iter *node_iter,
  82                                 struct bkey_i *insert)
  83 {
  84         const struct bkey_format *f = &b->format;
  85         struct bkey_packed *k;
  86         unsigned clobber_u64s;
  87
  88         EBUG_ON(btree_node_just_written(b));
  89         EBUG_ON(bset_written(b, btree_bset_last(b)));
  90         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  91         EBUG_ON(bkey_cmp(bkey_start_pos(&insert->k), b->data->min_key) < 0 ||
  92                 bkey_cmp(insert->k.p, b->data->max_key) > 0);
  93
  94         k = bch2_btree_node_iter_peek_all(node_iter, b);
  95         if (k && !bkey_cmp_packed(b, k, &insert->k)) {
  96                 BUG_ON(bkey_whiteout(k));
  97
  98                 if (!bkey_written(b, k) &&
  99                     bkey_val_u64s(&insert->k) == bkeyp_val_u64s(f, k) &&
 100                     !bkey_whiteout(&insert->k)) {
 101                         k->type = insert->k.type;
 102                         memcpy_u64s(bkeyp_val(f, k), &insert->v,
 103                                     bkey_val_u64s(&insert->k));
 104                         return true;
 105                 }
 106
 107                 btree_account_key_drop(b, k);
 108
 109                 if (bkey_whiteout(&insert->k)) {
 110                         unsigned clobber_u64s = k->u64s, new_u64s = k->u64s;
 111
 112                         k->type = KEY_TYPE_deleted;
 113
 114                         if (k->needs_whiteout) {
 115                                 push_whiteout(iter->trans->c, b, k);
 116                                 k->needs_whiteout = false;
 117                         }
 118
 119                         if (k >= btree_bset_last(b)->start) {
 120                                 bch2_bset_delete(b, k, clobber_u64s);
 121                                 new_u64s = 0;
 122                         }
 123
 124                         bch2_btree_node_iter_fix(iter, b, node_iter, k,
 125                                                  clobber_u64s, new_u64s);
 126                         return true;
 127
 128                 }
 129
 130                 if (k >= btree_bset_last(b)->start) {
 131                         clobber_u64s = k->u64s;
 132                         goto overwrite;
 133                 }
 134
 135                 insert->k.needs_whiteout = k->needs_whiteout;
 136                 k->needs_whiteout = false;
 137                 k->type = KEY_TYPE_deleted;
 138                 /*
 139                  * XXX: we should be able to do this without two calls to
 140                  * bch2_btree_node_iter_fix:
 141                  */
 142                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 143                                          k->u64s, k->u64s);
 144         } else {
 145                 /*
 146                  * Deleting, but the key to delete wasn't found - nothing to do:
 147                  */
 148                 if (bkey_whiteout(&insert->k))
 149                         return false;
 150
 151                 insert->k.needs_whiteout = false;
 152         }
 153
 154         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 155         clobber_u64s = 0;
 156 overwrite:
 157         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 158         bch2_btree_node_iter_fix(iter, b, node_iter, k,
 159                                  clobber_u64s, k->u64s);
 160         return true;
 161 }
 162
 163 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 164                                unsigned i, u64 seq)
 165 {
 166         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 167         struct btree_write *w = container_of(pin, struct btree_write, journal);
 168         struct btree *b = container_of(w, struct btree, writes[i]);
 169
 170         btree_node_lock_type(c, b, SIX_LOCK_read);
 171         bch2_btree_node_write_cond(c, b,
 172                 (btree_current_write(b) == w && w->journal.seq == seq));
 173         six_unlock_read(&b->lock);
 174 }
 175
 176 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 177 {
 178         return __btree_node_flush(j, pin, 0, seq);
 179 }
 180
 181 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 182 {
 183         return __btree_node_flush(j, pin, 1, seq);
 184 }
 185
 186 static inline void __btree_journal_key(struct btree_trans *trans,
 187                                        enum btree_id btree_id,
 188                                        struct bkey_i *insert)
 189 {
 190         struct journal *j = &trans->c->journal;
 191         u64 seq = trans->journal_res.seq;
 192         bool needs_whiteout = insert->k.needs_whiteout;
 193
 194         /* ick */
 195         insert->k.needs_whiteout = false;
 196         bch2_journal_add_keys(j, &trans->journal_res,
 197                               btree_id, insert);
 198         insert->k.needs_whiteout = needs_whiteout;
 199
 200         bch2_journal_set_has_inode(j, &trans->journal_res,
 201                                    insert->k.p.inode);
 202
 203         if (trans->journal_seq)
 204                 *trans->journal_seq = seq;
 205 }
 206
 207 void bch2_btree_journal_key(struct btree_trans *trans,
 208                            struct btree_iter *iter,
 209                            struct bkey_i *insert)
 210 {
 211         struct bch_fs *c = trans->c;
 212         struct journal *j = &c->journal;
 213         struct btree *b = iter->l[0].b;
 214         struct btree_write *w = btree_current_write(b);
 215
 216         EBUG_ON(iter->level || b->level);
 217         EBUG_ON(trans->journal_res.ref !=
 218                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 219
 220         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 221                 __btree_journal_key(trans, iter->btree_id, insert);
 222                 btree_bset_last(b)->journal_seq =
 223                         cpu_to_le64(trans->journal_res.seq);
 224         }
 225
 226         if (unlikely(!journal_pin_active(&w->journal))) {
 227                 u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 228                         ? trans->journal_res.seq
 229                         : j->replay_journal_seq;
 230
 231                 bch2_journal_pin_add(j, seq, &w->journal,
 232                                      btree_node_write_idx(b) == 0
 233                                      ? btree_node_flush0
 234                                      : btree_node_flush1);
 235         }
 236
 237         if (unlikely(!btree_node_dirty(b)))
 238                 set_btree_node_dirty(b);
 239 }
 240
 241 static void bch2_insert_fixup_key(struct btree_trans *trans,
 242                                   struct btree_insert_entry *insert)
 243 {
 244         struct btree_iter *iter = insert->iter;
 245         struct btree_iter_level *l = &iter->l[0];
 246
 247         EBUG_ON(iter->level);
 248         EBUG_ON(insert->k->k.u64s >
 249                 bch_btree_keys_u64s_remaining(trans->c, l->b));
 250
 251         if (likely(bch2_btree_bset_insert_key(iter, l->b, &l->iter,
 252                                               insert->k)))
 253                 bch2_btree_journal_key(trans, iter, insert->k);
 254 }
 255
 256 /**
 257  * btree_insert_key - insert a key one key into a leaf node
 258  */
 259 static void btree_insert_key_leaf(struct btree_trans *trans,
 260                                   struct btree_insert_entry *insert)
 261 {
 262         struct bch_fs *c = trans->c;
 263         struct btree_iter *iter = insert->iter;
 264         struct btree *b = iter->l[0].b;
 265         struct bset_tree *t = bset_tree_last(b);
 266         int old_u64s = bset_u64s(t);
 267         int old_live_u64s = b->nr.live_u64s;
 268         int live_u64s_added, u64s_added;
 269
 270         if (!btree_node_is_extents(b))
 271                 bch2_insert_fixup_key(trans, insert);
 272         else
 273                 bch2_insert_fixup_extent(trans, insert);
 274
 275         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 276         u64s_added = (int) bset_u64s(t) - old_u64s;
 277
 278         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 279                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 280         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 281                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 282
 283         if (u64s_added > live_u64s_added &&
 284             bch2_maybe_compact_whiteouts(c, b))
 285                 bch2_btree_iter_reinit_node(iter, b);
 286
 287         trace_btree_insert_key(c, b, insert->k);
 288 }
 289
 290 /* Normal update interface: */
 291
 292 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 293                                              struct btree_insert_entry *i)
 294 {
 295         struct bch_fs *c = trans->c;
 296
 297         BUG_ON(i->iter->level);
 298         BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
 299         EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
 300                 bkey_cmp(i->k->k.p, i->iter->l[0].b->key.k.p) > 0);
 301
 302         BUG_ON(debug_check_bkeys(c) &&
 303                !bkey_deleted(&i->k->k) &&
 304                bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->iter->btree_id));
 305 }
 306
 307 static noinline int
 308 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 309 {
 310         struct bch_fs *c = trans->c;
 311         int ret;
 312
 313         bch2_trans_unlock(trans);
 314
 315         ret = bch2_journal_preres_get(&c->journal,
 316                         &trans->journal_preres, u64s, 0);
 317         if (ret)
 318                 return ret;
 319
 320         if (!bch2_trans_relock(trans)) {
 321                 trace_trans_restart_journal_preres_get(trans->ip);
 322                 return -EINTR;
 323         }
 324
 325         return 0;
 326 }
 327
 328 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 329                                              unsigned flags)
 330 {
 331         struct bch_fs *c = trans->c;
 332         int ret;
 333
 334         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 335                 flags |= JOURNAL_RES_GET_RESERVED;
 336
 337         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 338                                    trans->journal_u64s, flags);
 339
 340         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 341 }
 342
 343 static enum btree_insert_ret
 344 btree_key_can_insert(struct btree_trans *trans,
 345                      struct btree_insert_entry *insert,
 346                      unsigned *u64s)
 347 {
 348         struct bch_fs *c = trans->c;
 349         struct btree *b = insert->iter->l[0].b;
 350         static enum btree_insert_ret ret;
 351
 352         if (unlikely(btree_node_fake(b)))
 353                 return BTREE_INSERT_BTREE_NODE_FULL;
 354
 355         ret = !btree_node_is_extents(b)
 356                 ? BTREE_INSERT_OK
 357                 : bch2_extent_can_insert(trans, insert, u64s);
 358         if (ret)
 359                 return ret;
 360
 361         if (*u64s > bch_btree_keys_u64s_remaining(c, b))
 362                 return BTREE_INSERT_BTREE_NODE_FULL;
 363
 364         return BTREE_INSERT_OK;
 365 }
 366
 367 static inline void do_btree_insert_one(struct btree_trans *trans,
 368                                        struct btree_insert_entry *insert)
 369 {
 370         btree_insert_key_leaf(trans, insert);
 371 }
 372
 373 static inline bool update_has_trans_triggers(struct btree_insert_entry *i)
 374 {
 375         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->iter->btree_id);
 376 }
 377
 378 static inline bool update_has_nontrans_triggers(struct btree_insert_entry *i)
 379 {
 380         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 381                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 382                 (1U << i->iter->btree_id);
 383 }
 384
 385 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 386 {
 387         __bch2_btree_iter_unlock(iter);
 388 }
 389
 390 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 391 {
 392         struct bch_fs *c = trans->c;
 393         struct btree_insert_entry *i;
 394         unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
 395                 ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
 396                 : 0;
 397
 398         if (unlikely(trans->flags & BTREE_INSERT_NOMARK))
 399                 return;
 400
 401         trans_for_each_update(trans, i)
 402                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 403                         bch2_mark_update(trans, i, NULL,
 404                                          mark_flags|BCH_BUCKET_MARK_GC);
 405 }
 406
 407 static inline int
 408 bch2_trans_commit_write_locked(struct btree_trans *trans,
 409                                struct btree_insert_entry **stopped_at)
 410 {
 411         struct bch_fs *c = trans->c;
 412         struct bch_fs_usage *fs_usage = NULL;
 413         struct btree_insert_entry *i;
 414         unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
 415                 ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
 416                 : 0;
 417         unsigned iter, u64s = 0;
 418         bool marking = false;
 419         int ret;
 420
 421         if (race_fault()) {
 422                 trace_trans_restart_fault_inject(trans->ip);
 423                 return -EINTR;
 424         }
 425
 426         /*
 427          * Check if the insert will fit in the leaf node with the write lock
 428          * held, otherwise another thread could write the node changing the
 429          * amount of space available:
 430          */
 431
 432         prefetch(&trans->c->journal.flags);
 433
 434         trans_for_each_update_sorted(trans, i, iter) {
 435                 /* Multiple inserts might go to same leaf: */
 436                 if (!same_leaf_as_prev(trans, iter))
 437                         u64s = 0;
 438
 439                 u64s += i->k->k.u64s;
 440                 ret = btree_key_can_insert(trans, i, &u64s);
 441                 if (ret) {
 442                         *stopped_at = i;
 443                         return ret;
 444                 }
 445
 446                 if (btree_node_type_needs_gc(i->iter->btree_id))
 447                         marking = true;
 448         }
 449
 450         if (marking) {
 451                 percpu_down_read(&c->mark_lock);
 452                 fs_usage = bch2_fs_usage_scratch_get(c);
 453         }
 454
 455         /*
 456          * Don't get journal reservation until after we know insert will
 457          * succeed:
 458          */
 459         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 460                 ret = bch2_trans_journal_res_get(trans,
 461                                 JOURNAL_RES_GET_NONBLOCK);
 462                 if (ret)
 463                         goto err;
 464         }
 465
 466         /*
 467          * Not allowed to fail after we've gotten our journal reservation - we
 468          * have to use it:
 469          */
 470
 471         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 472                 if (journal_seq_verify(c))
 473                         trans_for_each_update(trans, i)
 474                                 i->k->k.version.lo = trans->journal_res.seq;
 475                 else if (inject_invalid_keys(c))
 476                         trans_for_each_update(trans, i)
 477                                 i->k->k.version = MAX_VERSION;
 478         }
 479
 480         /* Must be called under mark_lock: */
 481         if (marking && trans->fs_usage_deltas &&
 482             bch2_replicas_delta_list_apply(c, fs_usage,
 483                                            trans->fs_usage_deltas)) {
 484                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 485                 goto err;
 486         }
 487
 488         trans_for_each_update(trans, i)
 489                 if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
 490                     update_has_nontrans_triggers(i))
 491                         bch2_mark_update(trans, i, fs_usage, mark_flags);
 492
 493         if (marking)
 494                 bch2_trans_fs_usage_apply(trans, fs_usage);
 495
 496         if (unlikely(c->gc_pos.phase))
 497                 bch2_trans_mark_gc(trans);
 498
 499         trans_for_each_update(trans, i)
 500                 do_btree_insert_one(trans, i);
 501 err:
 502         if (marking) {
 503                 bch2_fs_usage_scratch_put(c, fs_usage);
 504                 percpu_up_read(&c->mark_lock);
 505         }
 506
 507         return ret;
 508 }
 509
 510 /*
 511  * Get journal reservation, take write locks, and attempt to do btree update(s):
 512  */
 513 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 514                                        struct btree_insert_entry **stopped_at)
 515 {
 516         struct btree_insert_entry *i;
 517         struct btree_iter *iter;
 518         unsigned idx;
 519         int ret;
 520
 521         trans_for_each_update(trans, i)
 522                 BUG_ON(!btree_node_intent_locked(i->iter, 0));
 523
 524         ret = bch2_journal_preres_get(&trans->c->journal,
 525                         &trans->journal_preres, trans->journal_preres_u64s,
 526                         JOURNAL_RES_GET_NONBLOCK);
 527         if (unlikely(ret == -EAGAIN))
 528                 ret = bch2_trans_journal_preres_get_cold(trans,
 529                                                 trans->journal_preres_u64s);
 530         if (unlikely(ret))
 531                 return ret;
 532
 533         /*
 534          * Can't be holding any read locks when we go to take write locks:
 535          *
 536          * note - this must be done after bch2_trans_journal_preres_get_cold()
 537          * or anything else that might call bch2_trans_relock(), since that
 538          * would just retake the read locks:
 539          */
 540         trans_for_each_iter_all(trans, iter) {
 541                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 542                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 543                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 544                         bch2_btree_iter_unlock_noinline(iter);
 545                 }
 546         }
 547
 548         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 549                 trans_for_each_update(trans, i)
 550                         btree_insert_entry_checks(trans, i);
 551         bch2_btree_trans_verify_locks(trans);
 552
 553         /*
 554          * No more updates can be added - sort updates so we can take write
 555          * locks in the correct order:
 556          */
 557         btree_trans_sort_updates(trans);
 558
 559         trans_for_each_update_sorted(trans, i, idx)
 560                 if (!same_leaf_as_prev(trans, idx))
 561                         bch2_btree_node_lock_for_insert(trans->c,
 562                                                 i->iter->l[0].b, i->iter);
 563
 564         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 565
 566         trans_for_each_update_sorted(trans, i, idx)
 567                 if (!same_leaf_as_prev(trans, idx))
 568                         bch2_btree_node_unlock_write_inlined(i->iter->l[0].b,
 569                                                              i->iter);
 570
 571         /*
 572          * Drop journal reservation after dropping write locks, since dropping
 573          * the journal reservation may kick off a journal write:
 574          */
 575         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 576
 577         if (unlikely(ret))
 578                 return ret;
 579
 580         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 581                 trans->nounlock = true;
 582
 583         trans_for_each_update_sorted(trans, i, idx)
 584                 if (!same_leaf_as_prev(trans, idx))
 585                         bch2_foreground_maybe_merge(trans->c, i->iter,
 586                                                     0, trans->flags);
 587
 588         trans->nounlock = false;
 589
 590         trans_for_each_update(trans, i)
 591                 bch2_btree_iter_downgrade(i->iter);
 592
 593         return 0;
 594 }
 595
 596 static noinline
 597 int bch2_trans_commit_error(struct btree_trans *trans,
 598                             struct btree_insert_entry *i,
 599                             int ret)
 600 {
 601         struct bch_fs *c = trans->c;
 602         unsigned flags = trans->flags;
 603
 604         /*
 605          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 606          * update; if we haven't done anything yet it doesn't apply
 607          */
 608         flags &= ~BTREE_INSERT_NOUNLOCK;
 609
 610         switch (ret) {
 611         case BTREE_INSERT_BTREE_NODE_FULL:
 612                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 613
 614                 /*
 615                  * if the split succeeded without dropping locks the insert will
 616                  * still be atomic (what the caller peeked() and is overwriting
 617                  * won't have changed)
 618                  */
 619 #if 0
 620                 /*
 621                  * XXX:
 622                  * split -> btree node merging (of parent node) might still drop
 623                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 624                  *
 625                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 626                  * will inhibit merging - but we don't have a reliable way yet
 627                  * (do we?) of checking if we dropped locks in this path
 628                  */
 629                 if (!ret)
 630                         goto retry;
 631 #endif
 632
 633                 /*
 634                  * don't care if we got ENOSPC because we told split it
 635                  * couldn't block:
 636                  */
 637                 if (!ret ||
 638                     ret == -EINTR ||
 639                     (flags & BTREE_INSERT_NOUNLOCK)) {
 640                         trace_trans_restart_btree_node_split(trans->ip);
 641                         ret = -EINTR;
 642                 }
 643                 break;
 644         case BTREE_INSERT_ENOSPC:
 645                 ret = -ENOSPC;
 646                 break;
 647         case BTREE_INSERT_NEED_MARK_REPLICAS:
 648                 bch2_trans_unlock(trans);
 649
 650                 trans_for_each_update(trans, i) {
 651                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 652                         if (ret)
 653                                 return ret;
 654                 }
 655
 656                 if (bch2_trans_relock(trans))
 657                         return 0;
 658
 659                 trace_trans_restart_mark_replicas(trans->ip);
 660                 ret = -EINTR;
 661                 break;
 662         case BTREE_INSERT_NEED_JOURNAL_RES:
 663                 bch2_trans_unlock(trans);
 664
 665                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 666                 if (ret)
 667                         return ret;
 668
 669                 if (bch2_trans_relock(trans))
 670                         return 0;
 671
 672                 trace_trans_restart_journal_res_get(trans->ip);
 673                 ret = -EINTR;
 674                 break;
 675         default:
 676                 BUG_ON(ret >= 0);
 677                 break;
 678         }
 679
 680         if (ret == -EINTR) {
 681                 int ret2 = bch2_btree_iter_traverse_all(trans);
 682
 683                 if (ret2) {
 684                         trace_trans_restart_traverse(trans->ip);
 685                         return ret2;
 686                 }
 687
 688                 trace_trans_restart_atomic(trans->ip);
 689         }
 690
 691         return ret;
 692 }
 693
 694 static noinline int
 695 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 696 {
 697         struct bch_fs *c = trans->c;
 698         int ret;
 699
 700         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 701                 return -EROFS;
 702
 703         bch2_trans_unlock(trans);
 704
 705         ret = bch2_fs_read_write_early(c);
 706         if (ret)
 707                 return ret;
 708
 709         percpu_ref_get(&c->writes);
 710         return 0;
 711 }
 712
 713 int __bch2_trans_commit(struct btree_trans *trans)
 714 {
 715         struct btree_insert_entry *i = NULL;
 716         struct btree_iter *iter;
 717         unsigned u64s;
 718         int ret = 0;
 719
 720         if (!trans->nr_updates)
 721                 goto out_noupdates;
 722
 723         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 724                 lockdep_assert_held(&trans->c->gc_lock);
 725
 726         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 727
 728         trans->journal_u64s             = 0;
 729         trans->journal_preres_u64s      = 0;
 730
 731         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 732             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 733                 ret = bch2_trans_commit_get_rw_cold(trans);
 734                 if (ret)
 735                         return ret;
 736         }
 737
 738         /*
 739          * note: running triggers will append more updates to the list of
 740          * updates as we're walking it:
 741          */
 742         trans_for_each_update(trans, i) {
 743                 /* we know trans->nounlock won't be set here: */
 744                 if (unlikely(!(i->iter->locks_want < 1
 745                                ? __bch2_btree_iter_upgrade(i->iter, 1)
 746                                : i->iter->uptodate <= BTREE_ITER_NEED_PEEK))) {
 747                         trace_trans_restart_upgrade(trans->ip);
 748                         ret = -EINTR;
 749                         goto out;
 750                 }
 751
 752                 if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
 753                     update_has_trans_triggers(i)) {
 754                         ret = bch2_trans_mark_update(trans, i->iter, i->k);
 755                         if (unlikely(ret)) {
 756                                 if (ret == -EINTR)
 757                                         trace_trans_restart_mark(trans->ip);
 758                                 goto out;
 759                         }
 760                 }
 761
 762                 u64s = jset_u64s(i->k->k.u64s);
 763                 if (0)
 764                         trans->journal_preres_u64s += u64s;
 765                 trans->journal_u64s += u64s;
 766         }
 767 retry:
 768         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 769
 770         ret = do_bch2_trans_commit(trans, &i);
 771
 772         /* make sure we didn't drop or screw up locks: */
 773         bch2_btree_trans_verify_locks(trans);
 774
 775         if (ret)
 776                 goto err;
 777 out:
 778         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 779
 780         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 781                 percpu_ref_put(&trans->c->writes);
 782 out_noupdates:
 783         trans_for_each_iter_all(trans, iter)
 784                 iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
 785
 786         if (!ret) {
 787                 bch2_trans_unlink_iters(trans);
 788                 trans->iters_touched = 0;
 789         }
 790         trans->nr_updates       = 0;
 791         trans->mem_top          = 0;
 792
 793         if (trans->fs_usage_deltas) {
 794                 trans->fs_usage_deltas->used = 0;
 795                 memset(&trans->fs_usage_deltas->memset_start, 0,
 796                        (void *) &trans->fs_usage_deltas->memset_end -
 797                        (void *) &trans->fs_usage_deltas->memset_start);
 798         }
 799
 800         return ret;
 801 err:
 802         ret = bch2_trans_commit_error(trans, i, ret);
 803         if (ret)
 804                 goto out;
 805
 806         goto retry;
 807 }
 808
 809 static int __bch2_btree_insert(struct btree_trans *trans,
 810                                enum btree_id id, struct bkey_i *k)
 811 {
 812         struct btree_iter *iter;
 813
 814         iter = bch2_trans_get_iter(trans, id, bkey_start_pos(&k->k),
 815                                    BTREE_ITER_INTENT);
 816         if (IS_ERR(iter))
 817                 return PTR_ERR(iter);
 818
 819         bch2_trans_update(trans, iter, k);
 820         return 0;
 821 }
 822
 823 /**
 824  * bch2_btree_insert - insert keys into the extent btree
 825  * @c:                  pointer to struct bch_fs
 826  * @id:                 btree to insert into
 827  * @insert_keys:        list of keys to insert
 828  * @hook:               insert callback
 829  */
 830 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 831                       struct bkey_i *k,
 832                       struct disk_reservation *disk_res,
 833                       u64 *journal_seq, int flags)
 834 {
 835         return bch2_trans_do(c, disk_res, journal_seq, flags,
 836                              __bch2_btree_insert(&trans, id, k));
 837 }
 838
 839 int bch2_btree_delete_at_range(struct btree_trans *trans,
 840                                struct btree_iter *iter,
 841                                struct bpos end,
 842                                u64 *journal_seq)
 843 {
 844         struct bkey_s_c k;
 845         int ret = 0;
 846 retry:
 847         while ((k = bch2_btree_iter_peek(iter)).k &&
 848                !(ret = bkey_err(k)) &&
 849                bkey_cmp(iter->pos, end) < 0) {
 850                 struct bkey_i delete;
 851
 852                 bch2_trans_reset(trans, TRANS_RESET_MEM);
 853
 854                 bkey_init(&delete.k);
 855
 856                 /*
 857                  * For extents, iter.pos won't necessarily be the same as
 858                  * bkey_start_pos(k.k) (for non extents they always will be the
 859                  * same). It's important that we delete starting from iter.pos
 860                  * because the range we want to delete could start in the middle
 861                  * of k.
 862                  *
 863                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 864                  * bkey_start_pos(k.k)).
 865                  */
 866                 delete.k.p = iter->pos;
 867
 868                 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
 869                         unsigned max_sectors =
 870                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
 871
 872                         /* create the biggest key we can */
 873                         bch2_key_resize(&delete.k, max_sectors);
 874                         bch2_cut_back(end, &delete);
 875
 876                         ret = bch2_extent_trim_atomic(&delete, iter);
 877                         if (ret)
 878                                 break;
 879                 }
 880
 881                 bch2_trans_update(trans, iter, &delete);
 882                 ret = bch2_trans_commit(trans, NULL, journal_seq,
 883                                         BTREE_INSERT_NOFAIL);
 884                 if (ret)
 885                         break;
 886
 887                 bch2_trans_cond_resched(trans);
 888         }
 889
 890         if (ret == -EINTR) {
 891                 ret = 0;
 892                 goto retry;
 893         }
 894
 895         return ret;
 896
 897 }
 898
 899 int bch2_btree_delete_at(struct btree_trans *trans,
 900                          struct btree_iter *iter, unsigned flags)
 901 {
 902         struct bkey_i k;
 903
 904         bkey_init(&k.k);
 905         k.k.p = iter->pos;
 906
 907         bch2_trans_update(trans, iter, &k);
 908         return bch2_trans_commit(trans, NULL, NULL,
 909                                  BTREE_INSERT_NOFAIL|
 910                                  BTREE_INSERT_USE_RESERVE|flags);
 911 }
 912
 913 /*
 914  * bch_btree_delete_range - delete everything within a given range
 915  *
 916  * Range is a half open interval - [start, end)
 917  */
 918 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 919                             struct bpos start, struct bpos end,
 920                             u64 *journal_seq)
 921 {
 922         struct btree_trans trans;
 923         struct btree_iter *iter;
 924         int ret = 0;
 925
 926         /*
 927          * XXX: whether we need mem/more iters depends on whether this btree id
 928          * has triggers
 929          */
 930         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
 931
 932         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
 933
 934         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
 935         ret = bch2_trans_exit(&trans) ?: ret;
 936
 937         BUG_ON(ret == -EINTR);
 938         return ret;
 939 }