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                 insert->k.needs_whiteout = k->needs_whiteout;
 108
 109                 btree_account_key_drop(b, k);
 110
 111                 if (k >= btree_bset_last(b)->start) {
 112                         clobber_u64s = k->u64s;
 113
 114                         /*
 115                          * If we're deleting, and the key we're deleting doesn't
 116                          * need a whiteout (it wasn't overwriting a key that had
 117                          * been written to disk) - just delete it:
 118                          */
 119                         if (bkey_whiteout(&insert->k) && !k->needs_whiteout) {
 120                                 bch2_bset_delete(b, k, clobber_u64s);
 121                                 bch2_btree_node_iter_fix(iter, b, node_iter,
 122                                                          k, clobber_u64s, 0);
 123                                 return true;
 124                         }
 125
 126                         goto overwrite;
 127                 }
 128
 129                 k->type = KEY_TYPE_deleted;
 130                 bch2_btree_node_iter_fix(iter, b, node_iter, k,
 131                                          k->u64s, k->u64s);
 132
 133                 if (bkey_whiteout(&insert->k)) {
 134                         reserve_whiteout(b, k);
 135                         return true;
 136                 } else {
 137                         k->needs_whiteout = false;
 138                 }
 139         } else {
 140                 /*
 141                  * Deleting, but the key to delete wasn't found - nothing to do:
 142                  */
 143                 if (bkey_whiteout(&insert->k))
 144                         return false;
 145
 146                 insert->k.needs_whiteout = false;
 147         }
 148
 149         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 150         clobber_u64s = 0;
 151 overwrite:
 152         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 153         bch2_btree_node_iter_fix(iter, b, node_iter, k,
 154                                  clobber_u64s, k->u64s);
 155         return true;
 156 }
 157
 158 static void __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 159                                unsigned i, u64 seq)
 160 {
 161         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 162         struct btree_write *w = container_of(pin, struct btree_write, journal);
 163         struct btree *b = container_of(w, struct btree, writes[i]);
 164
 165         btree_node_lock_type(c, b, SIX_LOCK_read);
 166         bch2_btree_node_write_cond(c, b,
 167                 (btree_current_write(b) == w && w->journal.seq == seq));
 168         six_unlock_read(&b->lock);
 169 }
 170
 171 static void btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 172 {
 173         return __btree_node_flush(j, pin, 0, seq);
 174 }
 175
 176 static void btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 177 {
 178         return __btree_node_flush(j, pin, 1, seq);
 179 }
 180
 181 static inline void __btree_journal_key(struct btree_trans *trans,
 182                                        enum btree_id btree_id,
 183                                        struct bkey_i *insert)
 184 {
 185         struct journal *j = &trans->c->journal;
 186         u64 seq = trans->journal_res.seq;
 187         bool needs_whiteout = insert->k.needs_whiteout;
 188
 189         /* ick */
 190         insert->k.needs_whiteout = false;
 191         bch2_journal_add_keys(j, &trans->journal_res,
 192                               btree_id, insert);
 193         insert->k.needs_whiteout = needs_whiteout;
 194
 195         bch2_journal_set_has_inode(j, &trans->journal_res,
 196                                    insert->k.p.inode);
 197
 198         if (trans->journal_seq)
 199                 *trans->journal_seq = seq;
 200 }
 201
 202 void bch2_btree_journal_key(struct btree_trans *trans,
 203                            struct btree_iter *iter,
 204                            struct bkey_i *insert)
 205 {
 206         struct bch_fs *c = trans->c;
 207         struct journal *j = &c->journal;
 208         struct btree *b = iter->l[0].b;
 209         struct btree_write *w = btree_current_write(b);
 210
 211         EBUG_ON(iter->level || b->level);
 212         EBUG_ON(trans->journal_res.ref !=
 213                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 214
 215         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 216                 __btree_journal_key(trans, iter->btree_id, insert);
 217                 btree_bset_last(b)->journal_seq =
 218                         cpu_to_le64(trans->journal_res.seq);
 219         }
 220
 221         if (unlikely(!journal_pin_active(&w->journal))) {
 222                 u64 seq = likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))
 223                         ? trans->journal_res.seq
 224                         : j->replay_journal_seq;
 225
 226                 bch2_journal_pin_add(j, seq, &w->journal,
 227                                      btree_node_write_idx(b) == 0
 228                                      ? btree_node_flush0
 229                                      : btree_node_flush1);
 230         }
 231
 232         if (unlikely(!btree_node_dirty(b)))
 233                 set_btree_node_dirty(b);
 234 }
 235
 236 static void bch2_insert_fixup_key(struct btree_trans *trans,
 237                                   struct btree_insert_entry *insert)
 238 {
 239         struct btree_iter *iter = insert->iter;
 240         struct btree_iter_level *l = &iter->l[0];
 241
 242         EBUG_ON(iter->level);
 243         EBUG_ON(insert->k->k.u64s >
 244                 bch_btree_keys_u64s_remaining(trans->c, l->b));
 245
 246         if (likely(bch2_btree_bset_insert_key(iter, l->b, &l->iter,
 247                                               insert->k)))
 248                 bch2_btree_journal_key(trans, iter, insert->k);
 249 }
 250
 251 /**
 252  * btree_insert_key - insert a key one key into a leaf node
 253  */
 254 static void btree_insert_key_leaf(struct btree_trans *trans,
 255                                   struct btree_insert_entry *insert)
 256 {
 257         struct bch_fs *c = trans->c;
 258         struct btree_iter *iter = insert->iter;
 259         struct btree *b = iter->l[0].b;
 260         struct bset_tree *t = bset_tree_last(b);
 261         int old_u64s = bset_u64s(t);
 262         int old_live_u64s = b->nr.live_u64s;
 263         int live_u64s_added, u64s_added;
 264
 265         if (!btree_node_is_extents(b))
 266                 bch2_insert_fixup_key(trans, insert);
 267         else
 268                 bch2_insert_fixup_extent(trans, insert);
 269
 270         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 271         u64s_added = (int) bset_u64s(t) - old_u64s;
 272
 273         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 274                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 275         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 276                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 277
 278         if (u64s_added > live_u64s_added &&
 279             bch2_maybe_compact_whiteouts(c, b))
 280                 bch2_btree_iter_reinit_node(iter, b);
 281
 282         trace_btree_insert_key(c, b, insert->k);
 283 }
 284
 285 /* Normal update interface: */
 286
 287 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 288                                              struct btree_insert_entry *i)
 289 {
 290         struct bch_fs *c = trans->c;
 291
 292         BUG_ON(i->iter->level);
 293         BUG_ON(bkey_cmp(bkey_start_pos(&i->k->k), i->iter->pos));
 294         EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
 295                 bkey_cmp(i->k->k.p, i->iter->l[0].b->key.k.p) > 0);
 296         EBUG_ON((i->iter->flags & BTREE_ITER_IS_EXTENTS) &&
 297                 !(trans->flags & BTREE_INSERT_ATOMIC));
 298
 299         BUG_ON(debug_check_bkeys(c) &&
 300                !bkey_deleted(&i->k->k) &&
 301                bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->iter->btree_id));
 302 }
 303
 304 static noinline int
 305 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s)
 306 {
 307         struct bch_fs *c = trans->c;
 308         int ret;
 309
 310         bch2_trans_unlock(trans);
 311
 312         ret = bch2_journal_preres_get(&c->journal,
 313                         &trans->journal_preres, u64s, 0);
 314         if (ret)
 315                 return ret;
 316
 317         if (!bch2_trans_relock(trans)) {
 318                 trace_trans_restart_journal_preres_get(trans->ip);
 319                 return -EINTR;
 320         }
 321
 322         return 0;
 323 }
 324
 325 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 326                                              unsigned flags)
 327 {
 328         struct bch_fs *c = trans->c;
 329         int ret;
 330
 331         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 332                 flags |= JOURNAL_RES_GET_RESERVED;
 333
 334         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 335                                    trans->journal_u64s, flags);
 336
 337         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 338 }
 339
 340 static enum btree_insert_ret
 341 btree_key_can_insert(struct btree_trans *trans,
 342                      struct btree_insert_entry *insert,
 343                      unsigned *u64s)
 344 {
 345         struct bch_fs *c = trans->c;
 346         struct btree *b = insert->iter->l[0].b;
 347         static enum btree_insert_ret ret;
 348
 349         if (unlikely(btree_node_fake(b)))
 350                 return BTREE_INSERT_BTREE_NODE_FULL;
 351
 352         ret = !btree_node_is_extents(b)
 353                 ? BTREE_INSERT_OK
 354                 : bch2_extent_can_insert(trans, insert, u64s);
 355         if (ret)
 356                 return ret;
 357
 358         if (*u64s > bch_btree_keys_u64s_remaining(c, b))
 359                 return BTREE_INSERT_BTREE_NODE_FULL;
 360
 361         return BTREE_INSERT_OK;
 362 }
 363
 364 static inline void do_btree_insert_one(struct btree_trans *trans,
 365                                        struct btree_insert_entry *insert)
 366 {
 367         btree_insert_key_leaf(trans, insert);
 368 }
 369
 370 static inline bool update_has_trans_triggers(struct btree_insert_entry *i)
 371 {
 372         return BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->iter->btree_id);
 373 }
 374
 375 static inline bool update_has_nontrans_triggers(struct btree_insert_entry *i)
 376 {
 377         return (BTREE_NODE_TYPE_HAS_TRIGGERS &
 378                 ~BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS) &
 379                 (1U << i->iter->btree_id);
 380 }
 381
 382 static noinline void bch2_btree_iter_unlock_noinline(struct btree_iter *iter)
 383 {
 384         __bch2_btree_iter_unlock(iter);
 385 }
 386
 387 static noinline void bch2_trans_mark_gc(struct btree_trans *trans)
 388 {
 389         struct bch_fs *c = trans->c;
 390         struct btree_insert_entry *i;
 391         unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
 392                 ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
 393                 : 0;
 394
 395         if (unlikely(trans->flags & BTREE_INSERT_NOMARK))
 396                 return;
 397
 398         trans_for_each_update(trans, i)
 399                 if (gc_visited(c, gc_pos_btree_node(i->iter->l[0].b)))
 400                         bch2_mark_update(trans, i, NULL,
 401                                          mark_flags|BCH_BUCKET_MARK_GC);
 402 }
 403
 404 static inline int
 405 bch2_trans_commit_write_locked(struct btree_trans *trans,
 406                                struct btree_insert_entry **stopped_at)
 407 {
 408         struct bch_fs *c = trans->c;
 409         struct bch_fs_usage *fs_usage = NULL;
 410         struct btree_insert_entry *i;
 411         unsigned mark_flags = trans->flags & BTREE_INSERT_BUCKET_INVALIDATE
 412                 ? BCH_BUCKET_MARK_BUCKET_INVALIDATE
 413                 : 0;
 414         unsigned iter, u64s = 0;
 415         bool marking = false;
 416         int ret;
 417
 418         if (race_fault()) {
 419                 trace_trans_restart_fault_inject(trans->ip);
 420                 return -EINTR;
 421         }
 422
 423         /*
 424          * Check if the insert will fit in the leaf node with the write lock
 425          * held, otherwise another thread could write the node changing the
 426          * amount of space available:
 427          */
 428
 429         prefetch(&trans->c->journal.flags);
 430
 431         trans_for_each_update_sorted(trans, i, iter) {
 432                 /* Multiple inserts might go to same leaf: */
 433                 if (!same_leaf_as_prev(trans, iter))
 434                         u64s = 0;
 435
 436                 u64s += i->k->k.u64s;
 437                 ret = btree_key_can_insert(trans, i, &u64s);
 438                 if (ret) {
 439                         *stopped_at = i;
 440                         return ret;
 441                 }
 442
 443                 if (btree_node_type_needs_gc(i->iter->btree_id))
 444                         marking = true;
 445         }
 446
 447         if (marking) {
 448                 percpu_down_read(&c->mark_lock);
 449                 fs_usage = bch2_fs_usage_scratch_get(c);
 450         }
 451
 452         /*
 453          * Don't get journal reservation until after we know insert will
 454          * succeed:
 455          */
 456         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 457                 ret = bch2_trans_journal_res_get(trans,
 458                                 JOURNAL_RES_GET_NONBLOCK);
 459                 if (ret)
 460                         goto err;
 461         }
 462
 463         /*
 464          * Not allowed to fail after we've gotten our journal reservation - we
 465          * have to use it:
 466          */
 467
 468         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 469                 if (journal_seq_verify(c))
 470                         trans_for_each_update(trans, i)
 471                                 i->k->k.version.lo = trans->journal_res.seq;
 472                 else if (inject_invalid_keys(c))
 473                         trans_for_each_update(trans, i)
 474                                 i->k->k.version = MAX_VERSION;
 475         }
 476
 477         /* Must be called under mark_lock: */
 478         if (marking && trans->fs_usage_deltas &&
 479             bch2_replicas_delta_list_apply(c, fs_usage,
 480                                            trans->fs_usage_deltas)) {
 481                 ret = BTREE_INSERT_NEED_MARK_REPLICAS;
 482                 goto err;
 483         }
 484
 485         trans_for_each_update(trans, i)
 486                 if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
 487                     update_has_nontrans_triggers(i))
 488                         bch2_mark_update(trans, i, fs_usage, mark_flags);
 489
 490         if (marking)
 491                 bch2_trans_fs_usage_apply(trans, fs_usage);
 492
 493         if (unlikely(c->gc_pos.phase))
 494                 bch2_trans_mark_gc(trans);
 495
 496         trans_for_each_update(trans, i)
 497                 do_btree_insert_one(trans, i);
 498 err:
 499         if (marking) {
 500                 bch2_fs_usage_scratch_put(c, fs_usage);
 501                 percpu_up_read(&c->mark_lock);
 502         }
 503
 504         return ret;
 505 }
 506
 507 /*
 508  * Get journal reservation, take write locks, and attempt to do btree update(s):
 509  */
 510 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 511                                        struct btree_insert_entry **stopped_at)
 512 {
 513         struct btree_insert_entry *i;
 514         struct btree_iter *iter;
 515         unsigned idx, u64s, journal_preres_u64s = 0;
 516         int ret;
 517
 518         /*
 519          * note: running triggers will append more updates to the list of
 520          * updates as we're walking it:
 521          */
 522         trans_for_each_update(trans, i) {
 523                 /* we know trans->nounlock won't be set here: */
 524                 if (unlikely(!(i->iter->locks_want < 1
 525                                ? __bch2_btree_iter_upgrade(i->iter, 1)
 526                                : i->iter->uptodate <= BTREE_ITER_NEED_PEEK))) {
 527                         trace_trans_restart_upgrade(trans->ip);
 528                         return -EINTR;
 529                 }
 530
 531                 if (likely(!(trans->flags & BTREE_INSERT_NOMARK)) &&
 532                     update_has_trans_triggers(i)) {
 533                         ret = bch2_trans_mark_update(trans, i->iter, i->k);
 534                         if (unlikely(ret)) {
 535                                 if (ret == -EINTR)
 536                                         trace_trans_restart_mark(trans->ip);
 537                                 return ret;
 538                         }
 539                 }
 540
 541                 u64s = jset_u64s(i->k->k.u64s);
 542                 if (0)
 543                         journal_preres_u64s += u64s;
 544                 trans->journal_u64s += u64s;
 545         }
 546
 547         ret = bch2_journal_preres_get(&trans->c->journal,
 548                         &trans->journal_preres, journal_preres_u64s,
 549                         JOURNAL_RES_GET_NONBLOCK);
 550         if (unlikely(ret == -EAGAIN))
 551                 ret = bch2_trans_journal_preres_get_cold(trans,
 552                                                 journal_preres_u64s);
 553         if (unlikely(ret))
 554                 return ret;
 555
 556         /*
 557          * Can't be holding any read locks when we go to take write locks:
 558          *
 559          * note - this must be done after bch2_trans_journal_preres_get_cold()
 560          * or anything else that might call bch2_trans_relock(), since that
 561          * would just retake the read locks:
 562          */
 563         trans_for_each_iter_all(trans, iter) {
 564                 if (iter->nodes_locked != iter->nodes_intent_locked) {
 565                         EBUG_ON(iter->flags & BTREE_ITER_KEEP_UNTIL_COMMIT);
 566                         EBUG_ON(trans->iters_live & (1ULL << iter->idx));
 567                         bch2_btree_iter_unlock_noinline(iter);
 568                 }
 569         }
 570
 571         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG))
 572                 trans_for_each_update(trans, i)
 573                         btree_insert_entry_checks(trans, i);
 574         bch2_btree_trans_verify_locks(trans);
 575
 576         /*
 577          * No more updates can be added - sort updates so we can take write
 578          * locks in the correct order:
 579          */
 580         btree_trans_sort_updates(trans);
 581
 582         trans_for_each_update_sorted(trans, i, idx)
 583                 if (!same_leaf_as_prev(trans, idx))
 584                         bch2_btree_node_lock_for_insert(trans->c,
 585                                                 i->iter->l[0].b, i->iter);
 586
 587         ret = bch2_trans_commit_write_locked(trans, stopped_at);
 588
 589         trans_for_each_update_sorted(trans, i, idx)
 590                 if (!same_leaf_as_prev(trans, idx))
 591                         bch2_btree_node_unlock_write_inlined(i->iter->l[0].b,
 592                                                              i->iter);
 593
 594         /*
 595          * Drop journal reservation after dropping write locks, since dropping
 596          * the journal reservation may kick off a journal write:
 597          */
 598         bch2_journal_res_put(&trans->c->journal, &trans->journal_res);
 599
 600         if (unlikely(ret))
 601                 return ret;
 602
 603         if (trans->flags & BTREE_INSERT_NOUNLOCK)
 604                 trans->nounlock = true;
 605
 606         trans_for_each_update_sorted(trans, i, idx)
 607                 if (!same_leaf_as_prev(trans, idx))
 608                         bch2_foreground_maybe_merge(trans->c, i->iter,
 609                                                     0, trans->flags);
 610
 611         trans->nounlock = false;
 612
 613         trans_for_each_update(trans, i)
 614                 bch2_btree_iter_downgrade(i->iter);
 615
 616         return 0;
 617 }
 618
 619 static noinline
 620 int bch2_trans_commit_error(struct btree_trans *trans,
 621                             struct btree_insert_entry *i,
 622                             int ret)
 623 {
 624         struct bch_fs *c = trans->c;
 625         unsigned flags = trans->flags;
 626
 627         /*
 628          * BTREE_INSERT_NOUNLOCK means don't unlock _after_ successful btree
 629          * update; if we haven't done anything yet it doesn't apply
 630          */
 631         flags &= ~BTREE_INSERT_NOUNLOCK;
 632
 633         switch (ret) {
 634         case BTREE_INSERT_BTREE_NODE_FULL:
 635                 ret = bch2_btree_split_leaf(c, i->iter, flags);
 636
 637                 /*
 638                  * if the split succeeded without dropping locks the insert will
 639                  * still be atomic (in the BTREE_INSERT_ATOMIC sense, what the
 640                  * caller peeked() and is overwriting won't have changed)
 641                  */
 642 #if 0
 643                 /*
 644                  * XXX:
 645                  * split -> btree node merging (of parent node) might still drop
 646                  * locks when we're not passing it BTREE_INSERT_NOUNLOCK
 647                  *
 648                  * we don't want to pass BTREE_INSERT_NOUNLOCK to split as that
 649                  * will inhibit merging - but we don't have a reliable way yet
 650                  * (do we?) of checking if we dropped locks in this path
 651                  */
 652                 if (!ret)
 653                         goto retry;
 654 #endif
 655
 656                 /*
 657                  * don't care if we got ENOSPC because we told split it
 658                  * couldn't block:
 659                  */
 660                 if (!ret ||
 661                     ret == -EINTR ||
 662                     (flags & BTREE_INSERT_NOUNLOCK)) {
 663                         trace_trans_restart_btree_node_split(trans->ip);
 664                         ret = -EINTR;
 665                 }
 666                 break;
 667         case BTREE_INSERT_ENOSPC:
 668                 ret = -ENOSPC;
 669                 break;
 670         case BTREE_INSERT_NEED_MARK_REPLICAS:
 671                 bch2_trans_unlock(trans);
 672
 673                 trans_for_each_update(trans, i) {
 674                         ret = bch2_mark_bkey_replicas(c, bkey_i_to_s_c(i->k));
 675                         if (ret)
 676                                 return ret;
 677                 }
 678
 679                 if (bch2_trans_relock(trans))
 680                         return 0;
 681
 682                 trace_trans_restart_mark_replicas(trans->ip);
 683                 ret = -EINTR;
 684                 break;
 685         case BTREE_INSERT_NEED_JOURNAL_RES:
 686                 bch2_trans_unlock(trans);
 687
 688                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 689                 if (ret)
 690                         return ret;
 691
 692                 if (bch2_trans_relock(trans))
 693                         return 0;
 694
 695                 trace_trans_restart_journal_res_get(trans->ip);
 696                 ret = -EINTR;
 697                 break;
 698         default:
 699                 BUG_ON(ret >= 0);
 700                 break;
 701         }
 702
 703         if (ret == -EINTR) {
 704                 int ret2 = bch2_btree_iter_traverse_all(trans);
 705
 706                 if (ret2) {
 707                         trace_trans_restart_traverse(trans->ip);
 708                         return ret2;
 709                 }
 710
 711                 /*
 712                  * BTREE_ITER_ATOMIC means we have to return -EINTR if we
 713                  * dropped locks:
 714                  */
 715                 if (!(flags & BTREE_INSERT_ATOMIC))
 716                         return 0;
 717
 718                 trace_trans_restart_atomic(trans->ip);
 719         }
 720
 721         return ret;
 722 }
 723
 724 static noinline int
 725 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 726 {
 727         struct bch_fs *c = trans->c;
 728         int ret;
 729
 730         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)))
 731                 return -EROFS;
 732
 733         bch2_trans_unlock(trans);
 734
 735         ret = bch2_fs_read_write_early(c);
 736         if (ret)
 737                 return ret;
 738
 739         percpu_ref_get(&c->writes);
 740         return 0;
 741 }
 742
 743 int __bch2_trans_commit(struct btree_trans *trans)
 744 {
 745         struct btree_insert_entry *i = NULL;
 746         struct btree_iter *iter;
 747         unsigned orig_nr_updates        = trans->nr_updates;
 748         unsigned orig_mem_top           = trans->mem_top;
 749         int ret = 0;
 750
 751         if (!trans->nr_updates)
 752                 goto out_noupdates;
 753
 754         /* for the sake of sanity: */
 755         EBUG_ON(trans->nr_updates > 1 && !(trans->flags & BTREE_INSERT_ATOMIC));
 756
 757         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
 758                 lockdep_assert_held(&trans->c->gc_lock);
 759
 760         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
 761
 762         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
 763             unlikely(!percpu_ref_tryget(&trans->c->writes))) {
 764                 ret = bch2_trans_commit_get_rw_cold(trans);
 765                 if (ret)
 766                         return ret;
 767         }
 768 retry:
 769         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
 770         trans->journal_u64s     = 0;
 771
 772         ret = do_bch2_trans_commit(trans, &i);
 773
 774         if (trans->fs_usage_deltas) {
 775                 trans->fs_usage_deltas->used = 0;
 776                 memset(&trans->fs_usage_deltas->memset_start, 0,
 777                        (void *) &trans->fs_usage_deltas->memset_end -
 778                        (void *) &trans->fs_usage_deltas->memset_start);
 779         }
 780
 781         /* make sure we didn't drop or screw up locks: */
 782         bch2_btree_trans_verify_locks(trans);
 783
 784         if (ret)
 785                 goto err;
 786 out:
 787         bch2_journal_preres_put(&trans->c->journal, &trans->journal_preres);
 788
 789         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
 790                 percpu_ref_put(&trans->c->writes);
 791 out_noupdates:
 792         EBUG_ON(!(trans->flags & BTREE_INSERT_ATOMIC) && ret == -EINTR);
 793
 794         trans_for_each_iter_all(trans, iter)
 795                 iter->flags &= ~BTREE_ITER_KEEP_UNTIL_COMMIT;
 796
 797         if (!ret) {
 798                 bch2_trans_unlink_iters(trans);
 799                 trans->iters_touched = 0;
 800         }
 801         trans->nr_updates       = 0;
 802         trans->mem_top          = 0;
 803
 804         return ret;
 805 err:
 806         ret = bch2_trans_commit_error(trans, i, ret);
 807
 808         /* can't loop if it was passed in and we changed it: */
 809         if (unlikely(trans->flags & BTREE_INSERT_NO_CLEAR_REPLICAS) && !ret)
 810                 ret = -EINTR;
 811         if (ret)
 812                 goto out;
 813
 814         /* free updates and memory used by triggers, they'll be reexecuted: */
 815         trans->nr_updates       = orig_nr_updates;
 816         trans->mem_top          = orig_mem_top;
 817         goto retry;
 818 }
 819
 820 /**
 821  * bch2_btree_insert - insert keys into the extent btree
 822  * @c:                  pointer to struct bch_fs
 823  * @id:                 btree to insert into
 824  * @insert_keys:        list of keys to insert
 825  * @hook:               insert callback
 826  */
 827 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
 828                      struct bkey_i *k,
 829                      struct disk_reservation *disk_res,
 830                      u64 *journal_seq, int flags)
 831 {
 832         struct btree_trans trans;
 833         struct btree_iter *iter;
 834         int ret;
 835
 836         bch2_trans_init(&trans, c, 0, 0);
 837 retry:
 838         bch2_trans_begin(&trans);
 839
 840         iter = bch2_trans_get_iter(&trans, id, bkey_start_pos(&k->k),
 841                                    BTREE_ITER_INTENT);
 842
 843         bch2_trans_update(&trans, iter, k);
 844
 845         ret = bch2_trans_commit(&trans, disk_res, journal_seq, flags);
 846         if (ret == -EINTR)
 847                 goto retry;
 848         bch2_trans_exit(&trans);
 849
 850         return ret;
 851 }
 852
 853 int bch2_btree_delete_at_range(struct btree_trans *trans,
 854                                struct btree_iter *iter,
 855                                struct bpos end,
 856                                u64 *journal_seq)
 857 {
 858         struct bkey_s_c k;
 859         int ret = 0;
 860 retry:
 861         while ((k = bch2_btree_iter_peek(iter)).k &&
 862                !(ret = bkey_err(k)) &&
 863                bkey_cmp(iter->pos, end) < 0) {
 864                 struct bkey_i delete;
 865
 866                 bch2_trans_unlink_iters(trans);
 867                 trans->iters_touched &= trans->iters_live;
 868
 869                 bkey_init(&delete.k);
 870
 871                 /*
 872                  * For extents, iter.pos won't necessarily be the same as
 873                  * bkey_start_pos(k.k) (for non extents they always will be the
 874                  * same). It's important that we delete starting from iter.pos
 875                  * because the range we want to delete could start in the middle
 876                  * of k.
 877                  *
 878                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 879                  * bkey_start_pos(k.k)).
 880                  */
 881                 delete.k.p = iter->pos;
 882
 883                 if (iter->flags & BTREE_ITER_IS_EXTENTS) {
 884                         unsigned max_sectors =
 885                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
 886
 887                         /* create the biggest key we can */
 888                         bch2_key_resize(&delete.k, max_sectors);
 889                         bch2_cut_back(end, &delete);
 890
 891                         ret = bch2_extent_trim_atomic(&delete, iter);
 892                         if (ret)
 893                                 break;
 894                 }
 895
 896                 bch2_trans_update(trans, iter, &delete);
 897                 ret = bch2_trans_commit(trans, NULL, journal_seq,
 898                                         BTREE_INSERT_ATOMIC|
 899                                         BTREE_INSERT_NOFAIL);
 900                 if (ret)
 901                         break;
 902
 903                 bch2_trans_cond_resched(trans);
 904         }
 905
 906         if (ret == -EINTR) {
 907                 ret = 0;
 908                 goto retry;
 909         }
 910
 911         return ret;
 912
 913 }
 914
 915 int bch2_btree_delete_at(struct btree_trans *trans,
 916                          struct btree_iter *iter, unsigned flags)
 917 {
 918         struct bkey_i k;
 919
 920         bkey_init(&k.k);
 921         k.k.p = iter->pos;
 922
 923         bch2_trans_update(trans, iter, &k);
 924         return bch2_trans_commit(trans, NULL, NULL,
 925                                  BTREE_INSERT_NOFAIL|
 926                                  BTREE_INSERT_USE_RESERVE|flags);
 927 }
 928
 929 /*
 930  * bch_btree_delete_range - delete everything within a given range
 931  *
 932  * Range is a half open interval - [start, end)
 933  */
 934 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 935                             struct bpos start, struct bpos end,
 936                             u64 *journal_seq)
 937 {
 938         struct btree_trans trans;
 939         struct btree_iter *iter;
 940         int ret = 0;
 941
 942         /*
 943          * XXX: whether we need mem/more iters depends on whether this btree id
 944          * has triggers
 945          */
 946         bch2_trans_init(&trans, c, BTREE_ITER_MAX, 512);
 947
 948         iter = bch2_trans_get_iter(&trans, id, start, BTREE_ITER_INTENT);
 949
 950         ret = bch2_btree_delete_at_range(&trans, iter, end, journal_seq);
 951         ret = bch2_trans_exit(&trans) ?: ret;
 952
 953         BUG_ON(ret == -EINTR);
 954         return ret;
 955 }