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(trans->paths + 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_trans *trans, struct btree_insert_entry *i)
  45 {
  46         return (trans->paths + 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(trans, &i[0])->b == insert_l(trans, &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(trans, &i[0])->b == insert_l(trans, &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, trans->paths + i->path, insert_l(trans, 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, trans->paths + i->path, &insert_l(trans, i)->b->c))
 105                         return trans_lock_write_fail(trans, i);
 106
 107                 if (!i->cached)
 108                         bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, 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,
 123                                                 trans->paths + i->path, insert_l(trans, 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         struct btree_path *path = trans->paths + i->path;
 315
 316         BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 317         BUG_ON(i->cached        != path->cached);
 318         BUG_ON(i->level         != path->level);
 319         BUG_ON(i->btree_id      != path->btree_id);
 320         EBUG_ON(!i->level &&
 321                 btree_type_has_snapshots(i->btree_id) &&
 322                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 323                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 324                 i->k->k.p.snapshot &&
 325                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 326 }
 327
 328 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 329                                                       unsigned flags)
 330 {
 331         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 332                                     trans->journal_u64s, flags);
 333 }
 334
 335 #define JSET_ENTRY_LOG_U64s             4
 336
 337 static noinline void journal_transaction_name(struct btree_trans *trans)
 338 {
 339         struct bch_fs *c = trans->c;
 340         struct journal *j = &c->journal;
 341         struct jset_entry *entry =
 342                 bch2_journal_add_entry(j, &trans->journal_res,
 343                                        BCH_JSET_ENTRY_log, 0, 0,
 344                                        JSET_ENTRY_LOG_U64s);
 345         struct jset_entry_log *l =
 346                 container_of(entry, struct jset_entry_log, entry);
 347
 348         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 349 }
 350
 351 static inline int btree_key_can_insert(struct btree_trans *trans,
 352                                        struct btree *b, unsigned u64s)
 353 {
 354         struct bch_fs *c = trans->c;
 355
 356         if (!bch2_btree_node_insert_fits(c, b, u64s))
 357                 return -BCH_ERR_btree_insert_btree_node_full;
 358
 359         return 0;
 360 }
 361
 362 noinline static int
 363 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 364                                      struct btree_path *path, unsigned new_u64s)
 365 {
 366         struct btree_insert_entry *i;
 367         struct bkey_cached *ck = (void *) path->l[0].b;
 368         struct bkey_i *new_k;
 369         int ret;
 370
 371         bch2_trans_unlock_write(trans);
 372         bch2_trans_unlock(trans);
 373
 374         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 375         if (!new_k) {
 376                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 377                         bch2_btree_id_str(path->btree_id), new_u64s);
 378                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 379         }
 380
 381         ret =   bch2_trans_relock(trans) ?:
 382                 bch2_trans_lock_write(trans);
 383         if (unlikely(ret)) {
 384                 kfree(new_k);
 385                 return ret;
 386         }
 387
 388         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 389
 390         trans_for_each_update(trans, i)
 391                 if (i->old_v == &ck->k->v)
 392                         i->old_v = &new_k->v;
 393
 394         kfree(ck->k);
 395         ck->u64s        = new_u64s;
 396         ck->k           = new_k;
 397         return 0;
 398 }
 399
 400 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 401                                        struct btree_path *path, unsigned u64s)
 402 {
 403         struct bch_fs *c = trans->c;
 404         struct bkey_cached *ck = (void *) path->l[0].b;
 405         struct btree_insert_entry *i;
 406         unsigned new_u64s;
 407         struct bkey_i *new_k;
 408
 409         EBUG_ON(path->level);
 410
 411         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 412             bch2_btree_key_cache_must_wait(c) &&
 413             !(flags & BCH_TRANS_COMMIT_journal_reclaim))
 414                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 415
 416         /*
 417          * bch2_varint_decode can read past the end of the buffer by at most 7
 418          * bytes (it won't be used):
 419          */
 420         u64s += 1;
 421
 422         if (u64s <= ck->u64s)
 423                 return 0;
 424
 425         new_u64s        = roundup_pow_of_two(u64s);
 426         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT);
 427         if (unlikely(!new_k))
 428                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 429
 430         trans_for_each_update(trans, i)
 431                 if (i->old_v == &ck->k->v)
 432                         i->old_v = &new_k->v;
 433
 434         ck->u64s        = new_u64s;
 435         ck->k           = new_k;
 436         return 0;
 437 }
 438
 439 /* Triggers: */
 440
 441 static int run_one_mem_trigger(struct btree_trans *trans,
 442                                struct btree_insert_entry *i,
 443                                unsigned flags)
 444 {
 445         struct bkey_s_c old = { &i->old_k, i->old_v };
 446         struct bkey_i *new = i->k;
 447         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 448         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 449         int ret;
 450
 451         verify_update_old_key(trans, i);
 452
 453         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 454                 return 0;
 455
 456         if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
 457                 return 0;
 458
 459         if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
 460                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 461                                 old, bkey_i_to_s_c(new),
 462                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 463         } else {
 464                 struct bkey             _deleted = POS_KEY((trans->paths + i->path)->pos);
 465                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 466
 467                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 468                                 deleted, bkey_i_to_s_c(new),
 469                                 BTREE_TRIGGER_INSERT|flags) ?:
 470                         bch2_mark_key(trans, i->btree_id, i->level,
 471                                 old, deleted,
 472                                 BTREE_TRIGGER_OVERWRITE|flags);
 473         }
 474
 475         return ret;
 476 }
 477
 478 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 479                                  bool overwrite)
 480 {
 481         /*
 482          * Transactional triggers create new btree_insert_entries, so we can't
 483          * pass them a pointer to a btree_insert_entry, that memory is going to
 484          * move:
 485          */
 486         struct bkey old_k = i->old_k;
 487         struct bkey_s_c old = { &old_k, i->old_v };
 488         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 489         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 490
 491         verify_update_old_key(trans, i);
 492
 493         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 494             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 495                 return 0;
 496
 497         if (!i->insert_trigger_run &&
 498             !i->overwrite_trigger_run &&
 499             old_ops->trans_trigger == new_ops->trans_trigger) {
 500                 i->overwrite_trigger_run = true;
 501                 i->insert_trigger_run = true;
 502                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 503                                            BTREE_TRIGGER_INSERT|
 504                                            BTREE_TRIGGER_OVERWRITE|
 505                                            i->flags) ?: 1;
 506         } else if (overwrite && !i->overwrite_trigger_run) {
 507                 i->overwrite_trigger_run = true;
 508                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 509         } else if (!overwrite && !i->insert_trigger_run) {
 510                 i->insert_trigger_run = true;
 511                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 512         } else {
 513                 return 0;
 514         }
 515 }
 516
 517 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 518                               struct btree_insert_entry *btree_id_start)
 519 {
 520         struct btree_insert_entry *i;
 521         bool trans_trigger_run;
 522         int ret, overwrite;
 523
 524         for (overwrite = 1; overwrite >= 0; --overwrite) {
 525
 526                 /*
 527                  * Running triggers will append more updates to the list of updates as
 528                  * we're walking it:
 529                  */
 530                 do {
 531                         trans_trigger_run = false;
 532
 533                         for (i = btree_id_start;
 534                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 535                              i++) {
 536                                 if (i->btree_id != btree_id)
 537                                         continue;
 538
 539                                 ret = run_one_trans_trigger(trans, i, overwrite);
 540                                 if (ret < 0)
 541                                         return ret;
 542                                 if (ret)
 543                                         trans_trigger_run = true;
 544                         }
 545                 } while (trans_trigger_run);
 546         }
 547
 548         return 0;
 549 }
 550
 551 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 552 {
 553         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 554         unsigned btree_id = 0;
 555         int ret = 0;
 556
 557         /*
 558          *
 559          * For a given btree, this algorithm runs insert triggers before
 560          * overwrite triggers: this is so that when extents are being moved
 561          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 562          * they are re-added.
 563          */
 564         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 565                 if (btree_id == BTREE_ID_alloc)
 566                         continue;
 567
 568                 while (btree_id_start < trans->updates + trans->nr_updates &&
 569                        btree_id_start->btree_id < btree_id)
 570                         btree_id_start++;
 571
 572                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 573                 if (ret)
 574                         return ret;
 575         }
 576
 577         trans_for_each_update(trans, i) {
 578                 if (i->btree_id > BTREE_ID_alloc)
 579                         break;
 580                 if (i->btree_id == BTREE_ID_alloc) {
 581                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 582                         if (ret)
 583                                 return ret;
 584                         break;
 585                 }
 586         }
 587
 588 #ifdef CONFIG_BCACHEFS_DEBUG
 589         trans_for_each_update(trans, i)
 590                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 591                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 592                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 593 #endif
 594         return 0;
 595 }
 596
 597 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 598 {
 599         struct bch_fs *c = trans->c;
 600         struct btree_insert_entry *i;
 601         int ret = 0;
 602
 603         trans_for_each_update(trans, i) {
 604                 /*
 605                  * XXX: synchronization of cached update triggers with gc
 606                  * XXX: synchronization of interior node updates with gc
 607                  */
 608                 BUG_ON(i->cached || i->level);
 609
 610                 if (gc_visited(c, gc_pos_btree_node(insert_l(trans, i)->b))) {
 611                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 612                         if (ret)
 613                                 break;
 614                 }
 615         }
 616
 617         return ret;
 618 }
 619
 620 static inline int
 621 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 622                                struct btree_insert_entry **stopped_at,
 623                                unsigned long trace_ip)
 624 {
 625         struct bch_fs *c = trans->c;
 626         struct btree_insert_entry *i;
 627         struct btree_trans_commit_hook *h;
 628         unsigned u64s = 0;
 629         int ret;
 630
 631         if (race_fault()) {
 632                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 633                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 634         }
 635
 636         /*
 637          * Check if the insert will fit in the leaf node with the write lock
 638          * held, otherwise another thread could write the node changing the
 639          * amount of space available:
 640          */
 641
 642         prefetch(&trans->c->journal.flags);
 643
 644         trans_for_each_update(trans, i) {
 645                 /* Multiple inserts might go to same leaf: */
 646                 if (!same_leaf_as_prev(trans, i))
 647                         u64s = 0;
 648
 649                 u64s += i->k->k.u64s;
 650                 ret = !i->cached
 651                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 652                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 653                 if (ret) {
 654                         *stopped_at = i;
 655                         return ret;
 656                 }
 657
 658                 i->k->k.needs_whiteout = false;
 659         }
 660
 661         /*
 662          * Don't get journal reservation until after we know insert will
 663          * succeed:
 664          */
 665         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 666                 ret = bch2_trans_journal_res_get(trans,
 667                                 (flags & BCH_WATERMARK_MASK)|
 668                                 JOURNAL_RES_GET_NONBLOCK);
 669                 if (ret)
 670                         return ret;
 671
 672                 if (unlikely(trans->journal_transaction_names))
 673                         journal_transaction_name(trans);
 674         }
 675
 676         /*
 677          * Not allowed to fail after we've gotten our journal reservation - we
 678          * have to use it:
 679          */
 680
 681         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 682             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 683                 if (bch2_journal_seq_verify)
 684                         trans_for_each_update(trans, i)
 685                                 i->k->k.version.lo = trans->journal_res.seq;
 686                 else if (bch2_inject_invalid_keys)
 687                         trans_for_each_update(trans, i)
 688                                 i->k->k.version = MAX_VERSION;
 689         }
 690
 691         if (trans->fs_usage_deltas &&
 692             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 693                 return -BCH_ERR_btree_insert_need_mark_replicas;
 694
 695         h = trans->hooks;
 696         while (h) {
 697                 ret = h->fn(trans, h);
 698                 if (ret)
 699                         goto revert_fs_usage;
 700                 h = h->next;
 701         }
 702
 703         trans_for_each_update(trans, i)
 704                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 705                         ret = run_one_mem_trigger(trans, i, i->flags);
 706                         if (ret)
 707                                 goto fatal_err;
 708                 }
 709
 710         if (unlikely(c->gc_pos.phase)) {
 711                 ret = bch2_trans_commit_run_gc_triggers(trans);
 712                 if  (ret)
 713                         goto fatal_err;
 714         }
 715
 716         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 717                 struct journal *j = &c->journal;
 718                 struct jset_entry *entry;
 719
 720                 trans_for_each_update(trans, i) {
 721                         if (i->key_cache_already_flushed)
 722                                 continue;
 723
 724                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 725                                 continue;
 726
 727                         verify_update_old_key(trans, i);
 728
 729                         if (trans->journal_transaction_names) {
 730                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 731                                                        BCH_JSET_ENTRY_overwrite,
 732                                                        i->btree_id, i->level,
 733                                                        i->old_k.u64s);
 734                                 bkey_reassemble((struct bkey_i *) entry->start,
 735                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 736                         }
 737
 738                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 739                                                BCH_JSET_ENTRY_btree_keys,
 740                                                i->btree_id, i->level,
 741                                                i->k->k.u64s);
 742                         bkey_copy((struct bkey_i *) entry->start, i->k);
 743                 }
 744
 745                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 746                                   trans->journal_entries,
 747                                   trans->journal_entries_u64s);
 748
 749                 trans->journal_res.offset       += trans->journal_entries_u64s;
 750                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 751
 752                 if (trans->journal_seq)
 753                         *trans->journal_seq = trans->journal_res.seq;
 754         }
 755
 756         trans_for_each_update(trans, i) {
 757                 struct btree_path *path = trans->paths + i->path;
 758
 759                 if (!i->cached) {
 760                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 761                 } else if (!i->key_cache_already_flushed)
 762                         bch2_btree_insert_key_cached(trans, flags, i);
 763                 else {
 764                         bch2_btree_key_cache_drop(trans, path);
 765                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 766                 }
 767         }
 768
 769         return 0;
 770 fatal_err:
 771         bch2_fatal_error(c);
 772 revert_fs_usage:
 773         if (trans->fs_usage_deltas)
 774                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 775         return ret;
 776 }
 777
 778 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 779 {
 780         struct btree_insert_entry *i;
 781
 782         trans_for_each_update(trans, i)
 783                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 784 }
 785
 786 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 787                                                    enum bkey_invalid_flags flags,
 788                                                    struct btree_insert_entry *i,
 789                                                    struct printbuf *err)
 790 {
 791         struct bch_fs *c = trans->c;
 792
 793         printbuf_reset(err);
 794         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 795                    trans->fn, (void *) i->ip_allocated);
 796         prt_newline(err);
 797         printbuf_indent_add(err, 2);
 798
 799         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 800         prt_newline(err);
 801
 802         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 803         bch2_print_string_as_lines(KERN_ERR, err->buf);
 804
 805         bch2_inconsistent_error(c);
 806         bch2_dump_trans_updates(trans);
 807
 808         return -EINVAL;
 809 }
 810
 811 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 812                                 struct journal_entry_pin *_pin, u64 seq)
 813 {
 814         return 0;
 815 }
 816
 817 /*
 818  * Get journal reservation, take write locks, and attempt to do btree update(s):
 819  */
 820 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 821                                        struct btree_insert_entry **stopped_at,
 822                                        unsigned long trace_ip)
 823 {
 824         struct bch_fs *c = trans->c;
 825         struct btree_insert_entry *i;
 826         int ret = 0, u64s_delta = 0;
 827
 828         trans_for_each_update(trans, i) {
 829                 if (i->cached)
 830                         continue;
 831
 832                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 833                 u64s_delta -= i->old_btree_u64s;
 834
 835                 if (!same_leaf_as_next(trans, i)) {
 836                         if (u64s_delta <= 0) {
 837                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 838                                                         i->level, flags);
 839                                 if (unlikely(ret))
 840                                         return ret;
 841                         }
 842
 843                         u64s_delta = 0;
 844                 }
 845         }
 846
 847         ret = bch2_trans_lock_write(trans);
 848         if (unlikely(ret))
 849                 return ret;
 850
 851         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 852
 853         if (!ret && unlikely(trans->journal_replay_not_finished))
 854                 bch2_drop_overwrites_from_journal(trans);
 855
 856         bch2_trans_unlock_write(trans);
 857
 858         if (!ret && trans->journal_pin)
 859                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 860                                      trans->journal_pin,
 861                                      bch2_trans_commit_journal_pin_flush);
 862
 863         /*
 864          * Drop journal reservation after dropping write locks, since dropping
 865          * the journal reservation may kick off a journal write:
 866          */
 867         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 868                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 869
 870         return ret;
 871 }
 872
 873 static int journal_reclaim_wait_done(struct bch_fs *c)
 874 {
 875         int ret = bch2_journal_error(&c->journal) ?:
 876                 !bch2_btree_key_cache_must_wait(c);
 877
 878         if (!ret)
 879                 journal_reclaim_kick(&c->journal);
 880         return ret;
 881 }
 882
 883 static noinline
 884 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 885                             struct btree_insert_entry *i,
 886                             int ret, unsigned long trace_ip)
 887 {
 888         struct bch_fs *c = trans->c;
 889
 890         switch (ret) {
 891         case -BCH_ERR_btree_insert_btree_node_full:
 892                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 893                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 894                         trace_and_count(c, trans_restart_btree_node_split, trans,
 895                                         trace_ip, trans->paths + i->path);
 896                 break;
 897         case -BCH_ERR_btree_insert_need_mark_replicas:
 898                 ret = drop_locks_do(trans,
 899                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 900                 break;
 901         case -BCH_ERR_journal_res_get_blocked:
 902                 /*
 903                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 904                  * flag
 905                  */
 906                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 907                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 908                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 909                         break;
 910                 }
 911
 912                 ret = drop_locks_do(trans,
 913                         bch2_trans_journal_res_get(trans,
 914                                         (flags & BCH_WATERMARK_MASK)|
 915                                         JOURNAL_RES_GET_CHECK));
 916                 break;
 917         case -BCH_ERR_btree_insert_need_journal_reclaim:
 918                 bch2_trans_unlock(trans);
 919
 920                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 921
 922                 wait_event_freezable(c->journal.reclaim_wait,
 923                                      (ret = journal_reclaim_wait_done(c)));
 924                 if (ret < 0)
 925                         break;
 926
 927                 ret = bch2_trans_relock(trans);
 928                 break;
 929         default:
 930                 BUG_ON(ret >= 0);
 931                 break;
 932         }
 933
 934         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 935
 936         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 937                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 938                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 939
 940         return ret;
 941 }
 942
 943 static noinline int
 944 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 945 {
 946         struct bch_fs *c = trans->c;
 947         int ret;
 948
 949         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 950             test_bit(BCH_FS_started, &c->flags))
 951                 return -BCH_ERR_erofs_trans_commit;
 952
 953         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 954         if (ret)
 955                 return ret;
 956
 957         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 958         return 0;
 959 }
 960
 961 /*
 962  * This is for updates done in the early part of fsck - btree_gc - before we've
 963  * gone RW. we only add the new key to the list of keys for journal replay to
 964  * do.
 965  */
 966 static noinline int
 967 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 968 {
 969         struct bch_fs *c = trans->c;
 970         struct btree_insert_entry *i;
 971         int ret = 0;
 972
 973         trans_for_each_update(trans, i) {
 974                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 975                 if (ret)
 976                         break;
 977         }
 978
 979         return ret;
 980 }
 981
 982 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 983 {
 984         struct bch_fs *c = trans->c;
 985         struct btree_insert_entry *i = NULL;
 986         int ret = 0;
 987
 988         if (!trans->nr_updates &&
 989             !trans->journal_entries_u64s)
 990                 goto out_reset;
 991
 992         ret = bch2_trans_commit_run_triggers(trans);
 993         if (ret)
 994                 goto out_reset;
 995
 996         trans_for_each_update(trans, i) {
 997                 struct printbuf buf = PRINTBUF;
 998                 enum bkey_invalid_flags invalid_flags = 0;
 999
1000                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1001                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1002
1003                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1004                                                i->bkey_type, invalid_flags, &buf)))
1005                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1006                 btree_insert_entry_checks(trans, i);
1007                 printbuf_exit(&buf);
1008
1009                 if (ret)
1010                         return ret;
1011         }
1012
1013         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1014                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1015                 goto out_reset;
1016         }
1017
1018         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1019             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1020                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1021                 if (ret)
1022                         goto out_reset;
1023         }
1024
1025         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1026
1027         trans->journal_u64s             = trans->journal_entries_u64s;
1028         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1029         if (trans->journal_transaction_names)
1030                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1031
1032         trans_for_each_update(trans, i) {
1033                 struct btree_path *path = trans->paths + i->path;
1034
1035                 EBUG_ON(!path->should_be_locked);
1036
1037                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1038                 if (unlikely(ret))
1039                         goto out;
1040
1041                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1042
1043                 if (i->key_cache_already_flushed)
1044                         continue;
1045
1046                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1047                         continue;
1048
1049                 /* we're going to journal the key being updated: */
1050                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1051
1052                 /* and we're also going to log the overwrite: */
1053                 if (trans->journal_transaction_names)
1054                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1055         }
1056
1057         if (trans->extra_disk_res) {
1058                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1059                                 trans->extra_disk_res,
1060                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1061                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1062                 if (ret)
1063                         goto err;
1064         }
1065 retry:
1066         bch2_trans_verify_not_in_restart(trans);
1067         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1068                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1069
1070         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1071
1072         /* make sure we didn't drop or screw up locks: */
1073         bch2_trans_verify_locks(trans);
1074
1075         if (ret)
1076                 goto err;
1077
1078         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1079 out:
1080         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1081                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1082 out_reset:
1083         if (!ret)
1084                 bch2_trans_downgrade(trans);
1085         bch2_trans_reset_updates(trans);
1086
1087         return ret;
1088 err:
1089         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1090         if (ret)
1091                 goto out;
1092
1093         /*
1094          * We might have done another transaction commit in the error path -
1095          * i.e. btree write buffer flush - which will have made use of
1096          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1097          * how the journal sequence number to pin is passed in - so we must
1098          * restart:
1099          */
1100         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1101                 ret = -BCH_ERR_transaction_restart_nested;
1102                 goto out;
1103         }
1104
1105         goto retry;
1106 }