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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_gc.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_journal_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_write_buffer.h"
  11 #include "buckets.h"
  12 #include "errcode.h"
  13 #include "error.h"
  14 #include "journal.h"
  15 #include "journal_reclaim.h"
  16 #include "replicas.h"
  17 #include "snapshot.h"
  18
  19 #include <linux/prefetch.h>
  20
  21 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  22 {
  23 #ifdef CONFIG_BCACHEFS_DEBUG
  24         struct bch_fs *c = trans->c;
  25         struct bkey u;
  26         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
  27
  28         if (unlikely(trans->journal_replay_not_finished)) {
  29                 struct bkey_i *j_k =
  30                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  31
  32                 if (j_k)
  33                         k = bkey_i_to_s_c(j_k);
  34         }
  35
  36         u = *k.k;
  37         u.needs_whiteout = i->old_k.needs_whiteout;
  38
  39         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  40         BUG_ON(i->old_v != k.v);
  41 #endif
  42 }
  43
  44 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  45 {
  46         return i->path->l + i->level;
  47 }
  48
  49 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  50                                      struct btree_insert_entry *i)
  51 {
  52         return i != trans->updates &&
  53                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  54 }
  55
  56 static inline bool same_leaf_as_next(struct btree_trans *trans,
  57                                      struct btree_insert_entry *i)
  58 {
  59         return i + 1 < trans->updates + trans->nr_updates &&
  60                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  61 }
  62
  63 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  64                                            struct btree_path *path,
  65                                            struct btree *b)
  66 {
  67         struct bch_fs *c = trans->c;
  68
  69         if (unlikely(btree_node_just_written(b)) &&
  70             bch2_btree_post_write_cleanup(c, b))
  71                 bch2_trans_node_reinit_iter(trans, b);
  72
  73         /*
  74          * If the last bset has been written, or if it's gotten too big - start
  75          * a new bset to insert into:
  76          */
  77         if (want_new_bset(c, b))
  78                 bch2_btree_init_next(trans, b);
  79 }
  80
  81 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
  82 {
  83         while (--i >= trans->updates) {
  84                 if (same_leaf_as_prev(trans, i))
  85                         continue;
  86
  87                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
  88         }
  89
  90         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
  91         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
  92 }
  93
  94 static inline int bch2_trans_lock_write(struct btree_trans *trans)
  95 {
  96         struct btree_insert_entry *i;
  97
  98         EBUG_ON(trans->write_locked);
  99
 100         trans_for_each_update(trans, i) {
 101                 if (same_leaf_as_prev(trans, i))
 102                         continue;
 103
 104                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 105                         return trans_lock_write_fail(trans, i);
 106
 107                 if (!i->cached)
 108                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 109         }
 110
 111         trans->write_locked = true;
 112         return 0;
 113 }
 114
 115 static inline void bch2_trans_unlock_write(struct btree_trans *trans)
 116 {
 117         if (likely(trans->write_locked)) {
 118                 struct btree_insert_entry *i;
 119
 120                 trans_for_each_update(trans, i)
 121                         if (!same_leaf_as_prev(trans, i))
 122                                 bch2_btree_node_unlock_write_inlined(trans, i->path,
 123                                                                      insert_l(i)->b);
 124                 trans->write_locked = false;
 125         }
 126 }
 127
 128 /* Inserting into a given leaf node (last stage of insert): */
 129
 130 /* Handle overwrites and do insert, for non extents: */
 131 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 132                                 struct btree_path *path,
 133                                 struct btree *b,
 134                                 struct btree_node_iter *node_iter,
 135                                 struct bkey_i *insert)
 136 {
 137         struct bkey_packed *k;
 138         unsigned clobber_u64s = 0, new_u64s = 0;
 139
 140         EBUG_ON(btree_node_just_written(b));
 141         EBUG_ON(bset_written(b, btree_bset_last(b)));
 142         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 143         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 144         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 145         EBUG_ON(insert->k.u64s >
 146                 bch_btree_keys_u64s_remaining(trans->c, b));
 147         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 148
 149         k = bch2_btree_node_iter_peek_all(node_iter, b);
 150         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 151                 k = NULL;
 152
 153         /* @k is the key being overwritten/deleted, if any: */
 154         EBUG_ON(k && bkey_deleted(k));
 155
 156         /* Deleting, but not found? nothing to do: */
 157         if (bkey_deleted(&insert->k) && !k)
 158                 return false;
 159
 160         if (bkey_deleted(&insert->k)) {
 161                 /* Deleting: */
 162                 btree_account_key_drop(b, k);
 163                 k->type = KEY_TYPE_deleted;
 164
 165                 if (k->needs_whiteout)
 166                         push_whiteout(trans->c, b, insert->k.p);
 167                 k->needs_whiteout = false;
 168
 169                 if (k >= btree_bset_last(b)->start) {
 170                         clobber_u64s = k->u64s;
 171                         bch2_bset_delete(b, k, clobber_u64s);
 172                         goto fix_iter;
 173                 } else {
 174                         bch2_btree_path_fix_key_modified(trans, b, k);
 175                 }
 176
 177                 return true;
 178         }
 179
 180         if (k) {
 181                 /* Overwriting: */
 182                 btree_account_key_drop(b, k);
 183                 k->type = KEY_TYPE_deleted;
 184
 185                 insert->k.needs_whiteout = k->needs_whiteout;
 186                 k->needs_whiteout = false;
 187
 188                 if (k >= btree_bset_last(b)->start) {
 189                         clobber_u64s = k->u64s;
 190                         goto overwrite;
 191                 } else {
 192                         bch2_btree_path_fix_key_modified(trans, b, k);
 193                 }
 194         }
 195
 196         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 197 overwrite:
 198         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 199         new_u64s = k->u64s;
 200 fix_iter:
 201         if (clobber_u64s != new_u64s)
 202                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 203                                          clobber_u64s, new_u64s);
 204         return true;
 205 }
 206
 207 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 208                                unsigned i, u64 seq)
 209 {
 210         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 211         struct btree_write *w = container_of(pin, struct btree_write, journal);
 212         struct btree *b = container_of(w, struct btree, writes[i]);
 213         struct btree_trans *trans = bch2_trans_get(c);
 214         unsigned long old, new, v;
 215         unsigned idx = w - b->writes;
 216
 217         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 218         v = READ_ONCE(b->flags);
 219
 220         do {
 221                 old = new = v;
 222
 223                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 224                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 225                     w->journal.seq != seq)
 226                         break;
 227
 228                 new &= ~BTREE_WRITE_TYPE_MASK;
 229                 new |= BTREE_WRITE_journal_reclaim;
 230                 new |= 1 << BTREE_NODE_need_write;
 231         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 232
 233         btree_node_write_if_need(c, b, SIX_LOCK_read);
 234         six_unlock_read(&b->c.lock);
 235
 236         bch2_trans_put(trans);
 237         return 0;
 238 }
 239
 240 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 241 {
 242         return __btree_node_flush(j, pin, 0, seq);
 243 }
 244
 245 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 246 {
 247         return __btree_node_flush(j, pin, 1, seq);
 248 }
 249
 250 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 251                                        struct btree *b, u64 seq)
 252 {
 253         struct btree_write *w = btree_current_write(b);
 254
 255         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 256                              btree_node_write_idx(b) == 0
 257                              ? bch2_btree_node_flush0
 258                              : bch2_btree_node_flush1);
 259 }
 260
 261 /**
 262  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 263  * @trans:              btree transaction object
 264  * @path:               path pointing to @insert's pos
 265  * @insert:             key to insert
 266  * @journal_seq:        sequence number of journal reservation
 267  */
 268 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 269                                        struct btree_path *path,
 270                                        struct bkey_i *insert,
 271                                        u64 journal_seq)
 272 {
 273         struct bch_fs *c = trans->c;
 274         struct btree *b = path_l(path)->b;
 275         struct bset_tree *t = bset_tree_last(b);
 276         struct bset *i = bset(b, t);
 277         int old_u64s = bset_u64s(t);
 278         int old_live_u64s = b->nr.live_u64s;
 279         int live_u64s_added, u64s_added;
 280
 281         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 282                                         &path_l(path)->iter, insert)))
 283                 return;
 284
 285         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 286
 287         bch2_btree_add_journal_pin(c, b, journal_seq);
 288
 289         if (unlikely(!btree_node_dirty(b))) {
 290                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 291                 set_btree_node_dirty_acct(c, b);
 292         }
 293
 294         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 295         u64s_added = (int) bset_u64s(t) - old_u64s;
 296
 297         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 298                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 299         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 300                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 301
 302         if (u64s_added > live_u64s_added &&
 303             bch2_maybe_compact_whiteouts(c, b))
 304                 bch2_trans_node_reinit_iter(trans, b);
 305 }
 306
 307 /* Cached btree updates: */
 308
 309 /* Normal update interface: */
 310
 311 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 312                                              struct btree_insert_entry *i)
 313 {
 314         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 315         BUG_ON(i->cached        != i->path->cached);
 316         BUG_ON(i->level         != i->path->level);
 317         BUG_ON(i->btree_id      != i->path->btree_id);
 318         EBUG_ON(!i->level &&
 319                 btree_type_has_snapshots(i->btree_id) &&
 320                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 321                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 322                 i->k->k.p.snapshot &&
 323                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 324 }
 325
 326 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 327                                                       unsigned flags)
 328 {
 329         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 330                                     trans->journal_u64s, flags);
 331 }
 332
 333 #define JSET_ENTRY_LOG_U64s             4
 334
 335 static noinline void journal_transaction_name(struct btree_trans *trans)
 336 {
 337         struct bch_fs *c = trans->c;
 338         struct journal *j = &c->journal;
 339         struct jset_entry *entry =
 340                 bch2_journal_add_entry(j, &trans->journal_res,
 341                                        BCH_JSET_ENTRY_log, 0, 0,
 342                                        JSET_ENTRY_LOG_U64s);
 343         struct jset_entry_log *l =
 344                 container_of(entry, struct jset_entry_log, entry);
 345
 346         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 347 }
 348
 349 static inline int btree_key_can_insert(struct btree_trans *trans,
 350                                        struct btree *b, unsigned u64s)
 351 {
 352         struct bch_fs *c = trans->c;
 353
 354         if (!bch2_btree_node_insert_fits(c, b, u64s))
 355                 return -BCH_ERR_btree_insert_btree_node_full;
 356
 357         return 0;
 358 }
 359
 360 noinline static int
 361 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 362                                      struct btree_path *path, unsigned new_u64s)
 363 {
 364         struct bch_fs *c = trans->c;
 365         struct btree_insert_entry *i;
 366         struct bkey_cached *ck = (void *) path->l[0].b;
 367         struct bkey_i *new_k;
 368         int ret;
 369
 370         bch2_trans_unlock_write(trans);
 371         bch2_trans_unlock(trans);
 372
 373         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 374         if (!new_k) {
 375                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 376                         bch2_btree_id_str(path->btree_id), new_u64s);
 377                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 378         }
 379
 380         ret =   bch2_trans_relock(trans) ?:
 381                 bch2_trans_lock_write(trans);
 382         if (unlikely(ret)) {
 383                 kfree(new_k);
 384                 return ret;
 385         }
 386
 387         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 388
 389         trans_for_each_update(trans, i)
 390                 if (i->old_v == &ck->k->v)
 391                         i->old_v = &new_k->v;
 392
 393         kfree(ck->k);
 394         ck->u64s        = new_u64s;
 395         ck->k           = new_k;
 396         return 0;
 397 }
 398
 399 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 400                                        struct btree_path *path, unsigned u64s)
 401 {
 402         struct bch_fs *c = trans->c;
 403         struct bkey_cached *ck = (void *) path->l[0].b;
 404         struct btree_insert_entry *i;
 405         unsigned new_u64s;
 406         struct bkey_i *new_k;
 407
 408         EBUG_ON(path->level);
 409
 410         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 411             bch2_btree_key_cache_must_wait(c) &&
 412             !(flags & BCH_TRANS_COMMIT_journal_reclaim))
 413                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 414
 415         /*
 416          * bch2_varint_decode can read past the end of the buffer by at most 7
 417          * bytes (it won't be used):
 418          */
 419         u64s += 1;
 420
 421         if (u64s <= ck->u64s)
 422                 return 0;
 423
 424         new_u64s        = roundup_pow_of_two(u64s);
 425         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT);
 426         if (unlikely(!new_k))
 427                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 428
 429         trans_for_each_update(trans, i)
 430                 if (i->old_v == &ck->k->v)
 431                         i->old_v = &new_k->v;
 432
 433         ck->u64s        = new_u64s;
 434         ck->k           = new_k;
 435         return 0;
 436 }
 437
 438 /* Triggers: */
 439
 440 static int run_one_mem_trigger(struct btree_trans *trans,
 441                                struct btree_insert_entry *i,
 442                                unsigned flags)
 443 {
 444         struct bkey_s_c old = { &i->old_k, i->old_v };
 445         struct bkey_i *new = i->k;
 446         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 447         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 448         int ret;
 449
 450         verify_update_old_key(trans, i);
 451
 452         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 453                 return 0;
 454
 455         if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
 456                 return 0;
 457
 458         if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
 459                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 460                                 old, bkey_i_to_s_c(new),
 461                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 462         } else {
 463                 struct bkey             _deleted = KEY(0, 0, 0);
 464                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 465
 466                 _deleted.p = i->path->pos;
 467
 468                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 469                                 deleted, bkey_i_to_s_c(new),
 470                                 BTREE_TRIGGER_INSERT|flags) ?:
 471                         bch2_mark_key(trans, i->btree_id, i->level,
 472                                 old, deleted,
 473                                 BTREE_TRIGGER_OVERWRITE|flags);
 474         }
 475
 476         return ret;
 477 }
 478
 479 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 480                                  bool overwrite)
 481 {
 482         /*
 483          * Transactional triggers create new btree_insert_entries, so we can't
 484          * pass them a pointer to a btree_insert_entry, that memory is going to
 485          * move:
 486          */
 487         struct bkey old_k = i->old_k;
 488         struct bkey_s_c old = { &old_k, i->old_v };
 489         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 490         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 491
 492         verify_update_old_key(trans, i);
 493
 494         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 495             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 496                 return 0;
 497
 498         if (!i->insert_trigger_run &&
 499             !i->overwrite_trigger_run &&
 500             old_ops->trans_trigger == new_ops->trans_trigger) {
 501                 i->overwrite_trigger_run = true;
 502                 i->insert_trigger_run = true;
 503                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 504                                            BTREE_TRIGGER_INSERT|
 505                                            BTREE_TRIGGER_OVERWRITE|
 506                                            i->flags) ?: 1;
 507         } else if (overwrite && !i->overwrite_trigger_run) {
 508                 i->overwrite_trigger_run = true;
 509                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 510         } else if (!overwrite && !i->insert_trigger_run) {
 511                 i->insert_trigger_run = true;
 512                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 513         } else {
 514                 return 0;
 515         }
 516 }
 517
 518 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 519                               struct btree_insert_entry *btree_id_start)
 520 {
 521         struct btree_insert_entry *i;
 522         bool trans_trigger_run;
 523         int ret, overwrite;
 524
 525         for (overwrite = 1; overwrite >= 0; --overwrite) {
 526
 527                 /*
 528                  * Running triggers will append more updates to the list of updates as
 529                  * we're walking it:
 530                  */
 531                 do {
 532                         trans_trigger_run = false;
 533
 534                         for (i = btree_id_start;
 535                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 536                              i++) {
 537                                 if (i->btree_id != btree_id)
 538                                         continue;
 539
 540                                 ret = run_one_trans_trigger(trans, i, overwrite);
 541                                 if (ret < 0)
 542                                         return ret;
 543                                 if (ret)
 544                                         trans_trigger_run = true;
 545                         }
 546                 } while (trans_trigger_run);
 547         }
 548
 549         return 0;
 550 }
 551
 552 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 553 {
 554         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 555         unsigned btree_id = 0;
 556         int ret = 0;
 557
 558         /*
 559          *
 560          * For a given btree, this algorithm runs insert triggers before
 561          * overwrite triggers: this is so that when extents are being moved
 562          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 563          * they are re-added.
 564          */
 565         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 566                 if (btree_id == BTREE_ID_alloc)
 567                         continue;
 568
 569                 while (btree_id_start < trans->updates + trans->nr_updates &&
 570                        btree_id_start->btree_id < btree_id)
 571                         btree_id_start++;
 572
 573                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 574                 if (ret)
 575                         return ret;
 576         }
 577
 578         trans_for_each_update(trans, i) {
 579                 if (i->btree_id > BTREE_ID_alloc)
 580                         break;
 581                 if (i->btree_id == BTREE_ID_alloc) {
 582                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 583                         if (ret)
 584                                 return ret;
 585                         break;
 586                 }
 587         }
 588
 589 #ifdef CONFIG_BCACHEFS_DEBUG
 590         trans_for_each_update(trans, i)
 591                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 592                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 593                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 594 #endif
 595         return 0;
 596 }
 597
 598 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 599 {
 600         struct bch_fs *c = trans->c;
 601         struct btree_insert_entry *i;
 602         int ret = 0;
 603
 604         trans_for_each_update(trans, i) {
 605                 /*
 606                  * XXX: synchronization of cached update triggers with gc
 607                  * XXX: synchronization of interior node updates with gc
 608                  */
 609                 BUG_ON(i->cached || i->level);
 610
 611                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 612                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 613                         if (ret)
 614                                 break;
 615                 }
 616         }
 617
 618         return ret;
 619 }
 620
 621 static inline int
 622 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 623                                struct btree_insert_entry **stopped_at,
 624                                unsigned long trace_ip)
 625 {
 626         struct bch_fs *c = trans->c;
 627         struct btree_insert_entry *i;
 628         struct btree_write_buffered_key *wb;
 629         struct btree_trans_commit_hook *h;
 630         unsigned u64s = 0;
 631         int ret;
 632
 633         if (race_fault()) {
 634                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 635                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 636         }
 637
 638         /*
 639          * Check if the insert will fit in the leaf node with the write lock
 640          * held, otherwise another thread could write the node changing the
 641          * amount of space available:
 642          */
 643
 644         prefetch(&trans->c->journal.flags);
 645
 646         trans_for_each_update(trans, i) {
 647                 /* Multiple inserts might go to same leaf: */
 648                 if (!same_leaf_as_prev(trans, i))
 649                         u64s = 0;
 650
 651                 u64s += i->k->k.u64s;
 652                 ret = !i->cached
 653                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 654                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 655                 if (ret) {
 656                         *stopped_at = i;
 657                         return ret;
 658                 }
 659
 660                 i->k->k.needs_whiteout = false;
 661         }
 662
 663         if (trans->nr_wb_updates &&
 664             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 665                 return -BCH_ERR_btree_insert_need_flush_buffer;
 666
 667         /*
 668          * Don't get journal reservation until after we know insert will
 669          * succeed:
 670          */
 671         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 672                 ret = bch2_trans_journal_res_get(trans,
 673                                 (flags & BCH_WATERMARK_MASK)|
 674                                 JOURNAL_RES_GET_NONBLOCK);
 675                 if (ret)
 676                         return ret;
 677
 678                 if (unlikely(trans->journal_transaction_names))
 679                         journal_transaction_name(trans);
 680         }
 681
 682         /*
 683          * Not allowed to fail after we've gotten our journal reservation - we
 684          * have to use it:
 685          */
 686
 687         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 688             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 689                 if (bch2_journal_seq_verify)
 690                         trans_for_each_update(trans, i)
 691                                 i->k->k.version.lo = trans->journal_res.seq;
 692                 else if (bch2_inject_invalid_keys)
 693                         trans_for_each_update(trans, i)
 694                                 i->k->k.version = MAX_VERSION;
 695         }
 696
 697         if (trans->fs_usage_deltas &&
 698             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 699                 return -BCH_ERR_btree_insert_need_mark_replicas;
 700
 701         if (trans->nr_wb_updates) {
 702                 EBUG_ON(flags & BCH_TRANS_COMMIT_no_journal_res);
 703
 704                 ret = bch2_btree_insert_keys_write_buffer(trans);
 705                 if (ret)
 706                         goto revert_fs_usage;
 707         }
 708
 709         h = trans->hooks;
 710         while (h) {
 711                 ret = h->fn(trans, h);
 712                 if (ret)
 713                         goto revert_fs_usage;
 714                 h = h->next;
 715         }
 716
 717         trans_for_each_update(trans, i)
 718                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 719                         ret = run_one_mem_trigger(trans, i, i->flags);
 720                         if (ret)
 721                                 goto fatal_err;
 722                 }
 723
 724         if (unlikely(c->gc_pos.phase)) {
 725                 ret = bch2_trans_commit_run_gc_triggers(trans);
 726                 if  (ret)
 727                         goto fatal_err;
 728         }
 729
 730         if (unlikely(trans->extra_journal_entries.nr)) {
 731                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 732                                   trans->extra_journal_entries.data,
 733                                   trans->extra_journal_entries.nr);
 734
 735                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 736                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 737         }
 738
 739         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 740                 struct journal *j = &c->journal;
 741                 struct jset_entry *entry;
 742
 743                 trans_for_each_update(trans, i) {
 744                         if (i->key_cache_already_flushed)
 745                                 continue;
 746
 747                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 748                                 continue;
 749
 750                         verify_update_old_key(trans, i);
 751
 752                         if (trans->journal_transaction_names) {
 753                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 754                                                        BCH_JSET_ENTRY_overwrite,
 755                                                        i->btree_id, i->level,
 756                                                        i->old_k.u64s);
 757                                 bkey_reassemble((struct bkey_i *) entry->start,
 758                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 759                         }
 760
 761                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 762                                                BCH_JSET_ENTRY_btree_keys,
 763                                                i->btree_id, i->level,
 764                                                i->k->k.u64s);
 765                         bkey_copy((struct bkey_i *) entry->start, i->k);
 766                 }
 767
 768                 trans_for_each_wb_update(trans, wb) {
 769                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 770                                                BCH_JSET_ENTRY_btree_keys,
 771                                                wb->btree, 0,
 772                                                wb->k.k.u64s);
 773                         bkey_copy((struct bkey_i *) entry->start, &wb->k);
 774                 }
 775
 776                 if (trans->journal_seq)
 777                         *trans->journal_seq = trans->journal_res.seq;
 778         }
 779
 780         trans_for_each_update(trans, i) {
 781                 if (!i->cached) {
 782                         bch2_btree_insert_key_leaf(trans, i->path, i->k, trans->journal_res.seq);
 783                 } else if (!i->key_cache_already_flushed)
 784                         bch2_btree_insert_key_cached(trans, flags, i);
 785                 else {
 786                         bch2_btree_key_cache_drop(trans, i->path);
 787                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 788                 }
 789         }
 790
 791         return 0;
 792 fatal_err:
 793         bch2_fatal_error(c);
 794 revert_fs_usage:
 795         if (trans->fs_usage_deltas)
 796                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 797         return ret;
 798 }
 799
 800 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 801 {
 802         struct btree_insert_entry *i;
 803         struct btree_write_buffered_key *wb;
 804
 805         trans_for_each_update(trans, i)
 806                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 807
 808         trans_for_each_wb_update(trans, wb)
 809                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 810 }
 811
 812 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 813                                                    enum bkey_invalid_flags flags,
 814                                                    struct btree_insert_entry *i,
 815                                                    struct printbuf *err)
 816 {
 817         struct bch_fs *c = trans->c;
 818
 819         printbuf_reset(err);
 820         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 821                    trans->fn, (void *) i->ip_allocated);
 822         prt_newline(err);
 823         printbuf_indent_add(err, 2);
 824
 825         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 826         prt_newline(err);
 827
 828         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 829         bch2_print_string_as_lines(KERN_ERR, err->buf);
 830
 831         bch2_inconsistent_error(c);
 832         bch2_dump_trans_updates(trans);
 833
 834         return -EINVAL;
 835 }
 836
 837 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 838                                 struct journal_entry_pin *_pin, u64 seq)
 839 {
 840         return 0;
 841 }
 842
 843 /*
 844  * Get journal reservation, take write locks, and attempt to do btree update(s):
 845  */
 846 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 847                                        struct btree_insert_entry **stopped_at,
 848                                        unsigned long trace_ip)
 849 {
 850         struct bch_fs *c = trans->c;
 851         struct btree_insert_entry *i;
 852         int ret = 0, u64s_delta = 0;
 853
 854         trans_for_each_update(trans, i) {
 855                 if (i->cached)
 856                         continue;
 857
 858                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 859                 u64s_delta -= i->old_btree_u64s;
 860
 861                 if (!same_leaf_as_next(trans, i)) {
 862                         if (u64s_delta <= 0) {
 863                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 864                                                         i->level, flags);
 865                                 if (unlikely(ret))
 866                                         return ret;
 867                         }
 868
 869                         u64s_delta = 0;
 870                 }
 871         }
 872
 873         ret = bch2_trans_lock_write(trans);
 874         if (unlikely(ret))
 875                 return ret;
 876
 877         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 878
 879         if (!ret && unlikely(trans->journal_replay_not_finished))
 880                 bch2_drop_overwrites_from_journal(trans);
 881
 882         bch2_trans_unlock_write(trans);
 883
 884         if (!ret && trans->journal_pin)
 885                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 886                                      trans->journal_pin,
 887                                      bch2_trans_commit_journal_pin_flush);
 888
 889         /*
 890          * Drop journal reservation after dropping write locks, since dropping
 891          * the journal reservation may kick off a journal write:
 892          */
 893         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 894                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 895
 896         return ret;
 897 }
 898
 899 static int journal_reclaim_wait_done(struct bch_fs *c)
 900 {
 901         int ret = bch2_journal_error(&c->journal) ?:
 902                 !bch2_btree_key_cache_must_wait(c);
 903
 904         if (!ret)
 905                 journal_reclaim_kick(&c->journal);
 906         return ret;
 907 }
 908
 909 static noinline
 910 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 911                             struct btree_insert_entry *i,
 912                             int ret, unsigned long trace_ip)
 913 {
 914         struct bch_fs *c = trans->c;
 915
 916         switch (ret) {
 917         case -BCH_ERR_btree_insert_btree_node_full:
 918                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 919                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 920                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 921                 break;
 922         case -BCH_ERR_btree_insert_need_mark_replicas:
 923                 ret = drop_locks_do(trans,
 924                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 925                 break;
 926         case -BCH_ERR_journal_res_get_blocked:
 927                 /*
 928                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 929                  * flag
 930                  */
 931                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 932                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 933                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 934                         break;
 935                 }
 936
 937                 ret = drop_locks_do(trans,
 938                         bch2_trans_journal_res_get(trans,
 939                                         (flags & BCH_WATERMARK_MASK)|
 940                                         JOURNAL_RES_GET_CHECK));
 941                 break;
 942         case -BCH_ERR_btree_insert_need_journal_reclaim:
 943                 bch2_trans_unlock(trans);
 944
 945                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 946
 947                 wait_event_freezable(c->journal.reclaim_wait,
 948                                      (ret = journal_reclaim_wait_done(c)));
 949                 if (ret < 0)
 950                         break;
 951
 952                 ret = bch2_trans_relock(trans);
 953                 break;
 954         case -BCH_ERR_btree_insert_need_flush_buffer: {
 955                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 956
 957                 ret = 0;
 958
 959                 if (wb->state.nr > wb->size * 3 / 4) {
 960                         bch2_trans_unlock(trans);
 961                         mutex_lock(&wb->flush_lock);
 962
 963                         if (wb->state.nr > wb->size * 3 / 4) {
 964                                 bch2_trans_begin(trans);
 965                                 ret = __bch2_btree_write_buffer_flush(trans,
 966                                                 flags|BCH_TRANS_COMMIT_no_check_rw, true);
 967                                 if (!ret) {
 968                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
 969                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
 970                                 }
 971                         } else {
 972                                 mutex_unlock(&wb->flush_lock);
 973                                 ret = bch2_trans_relock(trans);
 974                         }
 975                 }
 976                 break;
 977         }
 978         default:
 979                 BUG_ON(ret >= 0);
 980                 break;
 981         }
 982
 983         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 984
 985         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 986                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 987                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 988
 989         return ret;
 990 }
 991
 992 static noinline int
 993 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 994 {
 995         struct bch_fs *c = trans->c;
 996         int ret;
 997
 998         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 999             test_bit(BCH_FS_STARTED, &c->flags))
1000                 return -BCH_ERR_erofs_trans_commit;
1001
1002         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1003         if (ret)
1004                 return ret;
1005
1006         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1007         return 0;
1008 }
1009
1010 /*
1011  * This is for updates done in the early part of fsck - btree_gc - before we've
1012  * gone RW. we only add the new key to the list of keys for journal replay to
1013  * do.
1014  */
1015 static noinline int
1016 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1017 {
1018         struct bch_fs *c = trans->c;
1019         struct btree_insert_entry *i;
1020         int ret = 0;
1021
1022         trans_for_each_update(trans, i) {
1023                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1024                 if (ret)
1025                         break;
1026         }
1027
1028         return ret;
1029 }
1030
1031 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1032 {
1033         struct bch_fs *c = trans->c;
1034         struct btree_insert_entry *i = NULL;
1035         struct btree_write_buffered_key *wb;
1036         int ret = 0;
1037
1038         if (!trans->nr_updates &&
1039             !trans->nr_wb_updates &&
1040             !trans->extra_journal_entries.nr)
1041                 goto out_reset;
1042
1043         ret = bch2_trans_commit_run_triggers(trans);
1044         if (ret)
1045                 goto out_reset;
1046
1047         trans_for_each_update(trans, i) {
1048                 struct printbuf buf = PRINTBUF;
1049                 enum bkey_invalid_flags invalid_flags = 0;
1050
1051                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1052                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1053
1054                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1055                                                i->bkey_type, invalid_flags, &buf)))
1056                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1057                 btree_insert_entry_checks(trans, i);
1058                 printbuf_exit(&buf);
1059
1060                 if (ret)
1061                         return ret;
1062         }
1063
1064         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1065                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1066                 goto out_reset;
1067         }
1068
1069         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1070             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1071                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1072                 if (ret)
1073                         goto out_reset;
1074         }
1075
1076         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1077             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1078                 bch2_trans_begin(trans);
1079                 bch2_trans_unlock(trans);
1080
1081                 ret = __bch2_btree_write_buffer_flush(trans,
1082                                         flags|BCH_TRANS_COMMIT_no_check_rw, true);
1083                 if (!ret) {
1084                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1085                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1086                 }
1087                 goto out;
1088         }
1089
1090         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1091
1092         trans->journal_u64s             = trans->extra_journal_entries.nr;
1093         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1094         if (trans->journal_transaction_names)
1095                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1096
1097         trans_for_each_update(trans, i) {
1098                 EBUG_ON(!i->path->should_be_locked);
1099
1100                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1101                 if (unlikely(ret))
1102                         goto out;
1103
1104                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1105
1106                 if (i->key_cache_already_flushed)
1107                         continue;
1108
1109                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1110                         continue;
1111
1112                 /* we're going to journal the key being updated: */
1113                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1114
1115                 /* and we're also going to log the overwrite: */
1116                 if (trans->journal_transaction_names)
1117                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1118         }
1119
1120         trans_for_each_wb_update(trans, wb)
1121                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1122
1123         if (trans->extra_journal_res) {
1124                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1125                                 trans->extra_journal_res,
1126                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1127                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1128                 if (ret)
1129                         goto err;
1130         }
1131 retry:
1132         bch2_trans_verify_not_in_restart(trans);
1133         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1134                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1135
1136         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1137
1138         /* make sure we didn't drop or screw up locks: */
1139         bch2_trans_verify_locks(trans);
1140
1141         if (ret)
1142                 goto err;
1143
1144         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1145 out:
1146         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1147                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1148 out_reset:
1149         if (!ret)
1150                 bch2_trans_downgrade(trans);
1151         bch2_trans_reset_updates(trans);
1152
1153         return ret;
1154 err:
1155         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1156         if (ret)
1157                 goto out;
1158
1159         /*
1160          * We might have done another transaction commit in the error path -
1161          * i.e. btree write buffer flush - which will have made use of
1162          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1163          * how the journal sequence number to pin is passed in - so we must
1164          * restart:
1165          */
1166         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1167                 ret = -BCH_ERR_transaction_restart_nested;
1168                 goto out;
1169         }
1170
1171         goto retry;
1172 }