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         EBUG_ON(trans->write_locked);
  97
  98         trans_for_each_update(trans, i) {
  99                 if (same_leaf_as_prev(trans, i))
 100                         continue;
 101
 102                 if (bch2_btree_node_lock_write(trans, trans->paths + i->path, &insert_l(trans, i)->b->c))
 103                         return trans_lock_write_fail(trans, i);
 104
 105                 if (!i->cached)
 106                         bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 107         }
 108
 109         trans->write_locked = true;
 110         return 0;
 111 }
 112
 113 static inline void bch2_trans_unlock_write(struct btree_trans *trans)
 114 {
 115         if (likely(trans->write_locked)) {
 116                 trans_for_each_update(trans, i)
 117                         if (!same_leaf_as_prev(trans, i))
 118                                 bch2_btree_node_unlock_write_inlined(trans,
 119                                                 trans->paths + i->path, insert_l(trans, i)->b);
 120                 trans->write_locked = false;
 121         }
 122 }
 123
 124 /* Inserting into a given leaf node (last stage of insert): */
 125
 126 /* Handle overwrites and do insert, for non extents: */
 127 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 128                                 struct btree_path *path,
 129                                 struct btree *b,
 130                                 struct btree_node_iter *node_iter,
 131                                 struct bkey_i *insert)
 132 {
 133         struct bkey_packed *k;
 134         unsigned clobber_u64s = 0, new_u64s = 0;
 135
 136         EBUG_ON(btree_node_just_written(b));
 137         EBUG_ON(bset_written(b, btree_bset_last(b)));
 138         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 139         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 140         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 141         EBUG_ON(insert->k.u64s >
 142                 bch_btree_keys_u64s_remaining(trans->c, b));
 143         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 144
 145         k = bch2_btree_node_iter_peek_all(node_iter, b);
 146         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 147                 k = NULL;
 148
 149         /* @k is the key being overwritten/deleted, if any: */
 150         EBUG_ON(k && bkey_deleted(k));
 151
 152         /* Deleting, but not found? nothing to do: */
 153         if (bkey_deleted(&insert->k) && !k)
 154                 return false;
 155
 156         if (bkey_deleted(&insert->k)) {
 157                 /* Deleting: */
 158                 btree_account_key_drop(b, k);
 159                 k->type = KEY_TYPE_deleted;
 160
 161                 if (k->needs_whiteout)
 162                         push_whiteout(trans->c, b, insert->k.p);
 163                 k->needs_whiteout = false;
 164
 165                 if (k >= btree_bset_last(b)->start) {
 166                         clobber_u64s = k->u64s;
 167                         bch2_bset_delete(b, k, clobber_u64s);
 168                         goto fix_iter;
 169                 } else {
 170                         bch2_btree_path_fix_key_modified(trans, b, k);
 171                 }
 172
 173                 return true;
 174         }
 175
 176         if (k) {
 177                 /* Overwriting: */
 178                 btree_account_key_drop(b, k);
 179                 k->type = KEY_TYPE_deleted;
 180
 181                 insert->k.needs_whiteout = k->needs_whiteout;
 182                 k->needs_whiteout = false;
 183
 184                 if (k >= btree_bset_last(b)->start) {
 185                         clobber_u64s = k->u64s;
 186                         goto overwrite;
 187                 } else {
 188                         bch2_btree_path_fix_key_modified(trans, b, k);
 189                 }
 190         }
 191
 192         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 193 overwrite:
 194         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 195         new_u64s = k->u64s;
 196 fix_iter:
 197         if (clobber_u64s != new_u64s)
 198                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 199                                          clobber_u64s, new_u64s);
 200         return true;
 201 }
 202
 203 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 204                                unsigned i, u64 seq)
 205 {
 206         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 207         struct btree_write *w = container_of(pin, struct btree_write, journal);
 208         struct btree *b = container_of(w, struct btree, writes[i]);
 209         struct btree_trans *trans = bch2_trans_get(c);
 210         unsigned long old, new, v;
 211         unsigned idx = w - b->writes;
 212
 213         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 214         v = READ_ONCE(b->flags);
 215
 216         do {
 217                 old = new = v;
 218
 219                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 220                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 221                     w->journal.seq != seq)
 222                         break;
 223
 224                 new &= ~BTREE_WRITE_TYPE_MASK;
 225                 new |= BTREE_WRITE_journal_reclaim;
 226                 new |= 1 << BTREE_NODE_need_write;
 227         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 228
 229         btree_node_write_if_need(c, b, SIX_LOCK_read);
 230         six_unlock_read(&b->c.lock);
 231
 232         bch2_trans_put(trans);
 233         return 0;
 234 }
 235
 236 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 237 {
 238         return __btree_node_flush(j, pin, 0, seq);
 239 }
 240
 241 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 242 {
 243         return __btree_node_flush(j, pin, 1, seq);
 244 }
 245
 246 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 247                                        struct btree *b, u64 seq)
 248 {
 249         struct btree_write *w = btree_current_write(b);
 250
 251         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 252                              btree_node_write_idx(b) == 0
 253                              ? bch2_btree_node_flush0
 254                              : bch2_btree_node_flush1);
 255 }
 256
 257 /**
 258  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 259  * @trans:              btree transaction object
 260  * @path:               path pointing to @insert's pos
 261  * @insert:             key to insert
 262  * @journal_seq:        sequence number of journal reservation
 263  */
 264 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 265                                        struct btree_path *path,
 266                                        struct bkey_i *insert,
 267                                        u64 journal_seq)
 268 {
 269         struct bch_fs *c = trans->c;
 270         struct btree *b = path_l(path)->b;
 271         struct bset_tree *t = bset_tree_last(b);
 272         struct bset *i = bset(b, t);
 273         int old_u64s = bset_u64s(t);
 274         int old_live_u64s = b->nr.live_u64s;
 275         int live_u64s_added, u64s_added;
 276
 277         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 278                                         &path_l(path)->iter, insert)))
 279                 return;
 280
 281         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 282
 283         bch2_btree_add_journal_pin(c, b, journal_seq);
 284
 285         if (unlikely(!btree_node_dirty(b))) {
 286                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 287                 set_btree_node_dirty_acct(c, b);
 288         }
 289
 290         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 291         u64s_added = (int) bset_u64s(t) - old_u64s;
 292
 293         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 294                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 295         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 296                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 297
 298         if (u64s_added > live_u64s_added &&
 299             bch2_maybe_compact_whiteouts(c, b))
 300                 bch2_trans_node_reinit_iter(trans, b);
 301 }
 302
 303 /* Cached btree updates: */
 304
 305 /* Normal update interface: */
 306
 307 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 308                                              struct btree_insert_entry *i)
 309 {
 310         struct btree_path *path = trans->paths + i->path;
 311
 312         BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 313         BUG_ON(i->cached        != path->cached);
 314         BUG_ON(i->level         != path->level);
 315         BUG_ON(i->btree_id      != path->btree_id);
 316         EBUG_ON(!i->level &&
 317                 btree_type_has_snapshots(i->btree_id) &&
 318                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 319                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 320                 i->k->k.p.snapshot &&
 321                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 322 }
 323
 324 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 325                                                       unsigned flags)
 326 {
 327         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 328                                     trans->journal_u64s, flags);
 329 }
 330
 331 #define JSET_ENTRY_LOG_U64s             4
 332
 333 static noinline void journal_transaction_name(struct btree_trans *trans)
 334 {
 335         struct bch_fs *c = trans->c;
 336         struct journal *j = &c->journal;
 337         struct jset_entry *entry =
 338                 bch2_journal_add_entry(j, &trans->journal_res,
 339                                        BCH_JSET_ENTRY_log, 0, 0,
 340                                        JSET_ENTRY_LOG_U64s);
 341         struct jset_entry_log *l =
 342                 container_of(entry, struct jset_entry_log, entry);
 343
 344         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 345 }
 346
 347 static inline int btree_key_can_insert(struct btree_trans *trans,
 348                                        struct btree *b, unsigned u64s)
 349 {
 350         struct bch_fs *c = trans->c;
 351
 352         if (!bch2_btree_node_insert_fits(c, b, u64s))
 353                 return -BCH_ERR_btree_insert_btree_node_full;
 354
 355         return 0;
 356 }
 357
 358 noinline static int
 359 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 360                                      struct btree_path *path, unsigned new_u64s)
 361 {
 362         struct bkey_cached *ck = (void *) path->l[0].b;
 363         struct bkey_i *new_k;
 364         int ret;
 365
 366         bch2_trans_unlock_write(trans);
 367         bch2_trans_unlock(trans);
 368
 369         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 370         if (!new_k) {
 371                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 372                         bch2_btree_id_str(path->btree_id), new_u64s);
 373                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 374         }
 375
 376         ret =   bch2_trans_relock(trans) ?:
 377                 bch2_trans_lock_write(trans);
 378         if (unlikely(ret)) {
 379                 kfree(new_k);
 380                 return ret;
 381         }
 382
 383         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 384
 385         trans_for_each_update(trans, i)
 386                 if (i->old_v == &ck->k->v)
 387                         i->old_v = &new_k->v;
 388
 389         kfree(ck->k);
 390         ck->u64s        = new_u64s;
 391         ck->k           = new_k;
 392         return 0;
 393 }
 394
 395 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 396                                        struct btree_path *path, unsigned u64s)
 397 {
 398         struct bch_fs *c = trans->c;
 399         struct bkey_cached *ck = (void *) path->l[0].b;
 400         unsigned new_u64s;
 401         struct bkey_i *new_k;
 402
 403         EBUG_ON(path->level);
 404
 405         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 406             bch2_btree_key_cache_must_wait(c) &&
 407             !(flags & BCH_TRANS_COMMIT_journal_reclaim))
 408                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 409
 410         /*
 411          * bch2_varint_decode can read past the end of the buffer by at most 7
 412          * bytes (it won't be used):
 413          */
 414         u64s += 1;
 415
 416         if (u64s <= ck->u64s)
 417                 return 0;
 418
 419         new_u64s        = roundup_pow_of_two(u64s);
 420         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT);
 421         if (unlikely(!new_k))
 422                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 423
 424         trans_for_each_update(trans, i)
 425                 if (i->old_v == &ck->k->v)
 426                         i->old_v = &new_k->v;
 427
 428         ck->u64s        = new_u64s;
 429         ck->k           = new_k;
 430         return 0;
 431 }
 432
 433 /* Triggers: */
 434
 435 static int run_one_mem_trigger(struct btree_trans *trans,
 436                                struct btree_insert_entry *i,
 437                                unsigned flags)
 438 {
 439         struct bkey_s_c old = { &i->old_k, i->old_v };
 440         struct bkey_i *new = i->k;
 441         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 442         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 443         int ret;
 444
 445         verify_update_old_key(trans, i);
 446
 447         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 448                 return 0;
 449
 450         if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
 451                 return 0;
 452
 453         if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
 454                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 455                                 old, bkey_i_to_s_c(new),
 456                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 457         } else {
 458                 struct bkey             _deleted = POS_KEY((trans->paths + i->path)->pos);
 459                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 460
 461                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 462                                 deleted, bkey_i_to_s_c(new),
 463                                 BTREE_TRIGGER_INSERT|flags) ?:
 464                         bch2_mark_key(trans, i->btree_id, i->level,
 465                                 old, deleted,
 466                                 BTREE_TRIGGER_OVERWRITE|flags);
 467         }
 468
 469         return ret;
 470 }
 471
 472 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 473                                  bool overwrite)
 474 {
 475         /*
 476          * Transactional triggers create new btree_insert_entries, so we can't
 477          * pass them a pointer to a btree_insert_entry, that memory is going to
 478          * move:
 479          */
 480         struct bkey old_k = i->old_k;
 481         struct bkey_s_c old = { &old_k, i->old_v };
 482         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 483         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 484
 485         verify_update_old_key(trans, i);
 486
 487         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 488             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 489                 return 0;
 490
 491         if (!i->insert_trigger_run &&
 492             !i->overwrite_trigger_run &&
 493             old_ops->trans_trigger == new_ops->trans_trigger) {
 494                 i->overwrite_trigger_run = true;
 495                 i->insert_trigger_run = true;
 496                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 497                                            BTREE_TRIGGER_INSERT|
 498                                            BTREE_TRIGGER_OVERWRITE|
 499                                            i->flags) ?: 1;
 500         } else if (overwrite && !i->overwrite_trigger_run) {
 501                 i->overwrite_trigger_run = true;
 502                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 503         } else if (!overwrite && !i->insert_trigger_run) {
 504                 i->insert_trigger_run = true;
 505                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 506         } else {
 507                 return 0;
 508         }
 509 }
 510
 511 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 512                               struct btree_insert_entry *btree_id_start)
 513 {
 514         struct btree_insert_entry *i;
 515         bool trans_trigger_run;
 516         int ret, overwrite;
 517
 518         for (overwrite = 1; overwrite >= 0; --overwrite) {
 519
 520                 /*
 521                  * Running triggers will append more updates to the list of updates as
 522                  * we're walking it:
 523                  */
 524                 do {
 525                         trans_trigger_run = false;
 526
 527                         for (i = btree_id_start;
 528                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 529                              i++) {
 530                                 if (i->btree_id != btree_id)
 531                                         continue;
 532
 533                                 ret = run_one_trans_trigger(trans, i, overwrite);
 534                                 if (ret < 0)
 535                                         return ret;
 536                                 if (ret)
 537                                         trans_trigger_run = true;
 538                         }
 539                 } while (trans_trigger_run);
 540         }
 541
 542         return 0;
 543 }
 544
 545 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 546 {
 547         struct btree_insert_entry *btree_id_start = trans->updates;
 548         unsigned btree_id = 0;
 549         int ret = 0;
 550
 551         /*
 552          *
 553          * For a given btree, this algorithm runs insert triggers before
 554          * overwrite triggers: this is so that when extents are being moved
 555          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 556          * they are re-added.
 557          */
 558         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 559                 if (btree_id == BTREE_ID_alloc)
 560                         continue;
 561
 562                 while (btree_id_start < trans->updates + trans->nr_updates &&
 563                        btree_id_start->btree_id < btree_id)
 564                         btree_id_start++;
 565
 566                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 567                 if (ret)
 568                         return ret;
 569         }
 570
 571         trans_for_each_update(trans, i) {
 572                 if (i->btree_id > BTREE_ID_alloc)
 573                         break;
 574                 if (i->btree_id == BTREE_ID_alloc) {
 575                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 576                         if (ret)
 577                                 return ret;
 578                         break;
 579                 }
 580         }
 581
 582 #ifdef CONFIG_BCACHEFS_DEBUG
 583         trans_for_each_update(trans, i)
 584                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 585                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 586                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 587 #endif
 588         return 0;
 589 }
 590
 591 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 592 {
 593         struct bch_fs *c = trans->c;
 594         int ret = 0;
 595
 596         trans_for_each_update(trans, i) {
 597                 /*
 598                  * XXX: synchronization of cached update triggers with gc
 599                  * XXX: synchronization of interior node updates with gc
 600                  */
 601                 BUG_ON(i->cached || i->level);
 602
 603                 if (gc_visited(c, gc_pos_btree_node(insert_l(trans, i)->b))) {
 604                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 605                         if (ret)
 606                                 break;
 607                 }
 608         }
 609
 610         return ret;
 611 }
 612
 613 static inline int
 614 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 615                                struct btree_insert_entry **stopped_at,
 616                                unsigned long trace_ip)
 617 {
 618         struct bch_fs *c = trans->c;
 619         struct btree_trans_commit_hook *h;
 620         unsigned u64s = 0;
 621         int ret;
 622
 623         if (race_fault()) {
 624                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 625                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 626         }
 627
 628         /*
 629          * Check if the insert will fit in the leaf node with the write lock
 630          * held, otherwise another thread could write the node changing the
 631          * amount of space available:
 632          */
 633
 634         prefetch(&trans->c->journal.flags);
 635
 636         trans_for_each_update(trans, i) {
 637                 /* Multiple inserts might go to same leaf: */
 638                 if (!same_leaf_as_prev(trans, i))
 639                         u64s = 0;
 640
 641                 u64s += i->k->k.u64s;
 642                 ret = !i->cached
 643                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 644                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 645                 if (ret) {
 646                         *stopped_at = i;
 647                         return ret;
 648                 }
 649
 650                 i->k->k.needs_whiteout = false;
 651         }
 652
 653         /*
 654          * Don't get journal reservation until after we know insert will
 655          * succeed:
 656          */
 657         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 658                 ret = bch2_trans_journal_res_get(trans,
 659                                 (flags & BCH_WATERMARK_MASK)|
 660                                 JOURNAL_RES_GET_NONBLOCK);
 661                 if (ret)
 662                         return ret;
 663
 664                 if (unlikely(trans->journal_transaction_names))
 665                         journal_transaction_name(trans);
 666         }
 667
 668         /*
 669          * Not allowed to fail after we've gotten our journal reservation - we
 670          * have to use it:
 671          */
 672
 673         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 674             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 675                 if (bch2_journal_seq_verify)
 676                         trans_for_each_update(trans, i)
 677                                 i->k->k.version.lo = trans->journal_res.seq;
 678                 else if (bch2_inject_invalid_keys)
 679                         trans_for_each_update(trans, i)
 680                                 i->k->k.version = MAX_VERSION;
 681         }
 682
 683         if (trans->fs_usage_deltas &&
 684             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 685                 return -BCH_ERR_btree_insert_need_mark_replicas;
 686
 687         h = trans->hooks;
 688         while (h) {
 689                 ret = h->fn(trans, h);
 690                 if (ret)
 691                         goto revert_fs_usage;
 692                 h = h->next;
 693         }
 694
 695         trans_for_each_update(trans, i)
 696                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 697                         ret = run_one_mem_trigger(trans, i, i->flags);
 698                         if (ret)
 699                                 goto fatal_err;
 700                 }
 701
 702         if (unlikely(c->gc_pos.phase)) {
 703                 ret = bch2_trans_commit_run_gc_triggers(trans);
 704                 if  (ret)
 705                         goto fatal_err;
 706         }
 707
 708         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 709                 struct journal *j = &c->journal;
 710                 struct jset_entry *entry;
 711
 712                 trans_for_each_update(trans, i) {
 713                         if (i->key_cache_already_flushed)
 714                                 continue;
 715
 716                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 717                                 continue;
 718
 719                         verify_update_old_key(trans, i);
 720
 721                         if (trans->journal_transaction_names) {
 722                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 723                                                        BCH_JSET_ENTRY_overwrite,
 724                                                        i->btree_id, i->level,
 725                                                        i->old_k.u64s);
 726                                 bkey_reassemble((struct bkey_i *) entry->start,
 727                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 728                         }
 729
 730                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 731                                                BCH_JSET_ENTRY_btree_keys,
 732                                                i->btree_id, i->level,
 733                                                i->k->k.u64s);
 734                         bkey_copy((struct bkey_i *) entry->start, i->k);
 735                 }
 736
 737                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 738                                   trans->journal_entries,
 739                                   trans->journal_entries_u64s);
 740
 741                 trans->journal_res.offset       += trans->journal_entries_u64s;
 742                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 743
 744                 if (trans->journal_seq)
 745                         *trans->journal_seq = trans->journal_res.seq;
 746         }
 747
 748         trans_for_each_update(trans, i) {
 749                 struct btree_path *path = trans->paths + i->path;
 750
 751                 if (!i->cached) {
 752                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 753                 } else if (!i->key_cache_already_flushed)
 754                         bch2_btree_insert_key_cached(trans, flags, i);
 755                 else {
 756                         bch2_btree_key_cache_drop(trans, path);
 757                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 758                 }
 759         }
 760
 761         return 0;
 762 fatal_err:
 763         bch2_fatal_error(c);
 764 revert_fs_usage:
 765         if (trans->fs_usage_deltas)
 766                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 767         return ret;
 768 }
 769
 770 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 771 {
 772         trans_for_each_update(trans, i)
 773                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 774 }
 775
 776 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 777                                                    enum bkey_invalid_flags flags,
 778                                                    struct btree_insert_entry *i,
 779                                                    struct printbuf *err)
 780 {
 781         struct bch_fs *c = trans->c;
 782
 783         printbuf_reset(err);
 784         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 785                    trans->fn, (void *) i->ip_allocated);
 786         prt_newline(err);
 787         printbuf_indent_add(err, 2);
 788
 789         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 790         prt_newline(err);
 791
 792         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 793         bch2_print_string_as_lines(KERN_ERR, err->buf);
 794
 795         bch2_inconsistent_error(c);
 796         bch2_dump_trans_updates(trans);
 797
 798         return -EINVAL;
 799 }
 800
 801 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 802                                 struct journal_entry_pin *_pin, u64 seq)
 803 {
 804         return 0;
 805 }
 806
 807 /*
 808  * Get journal reservation, take write locks, and attempt to do btree update(s):
 809  */
 810 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 811                                        struct btree_insert_entry **stopped_at,
 812                                        unsigned long trace_ip)
 813 {
 814         struct bch_fs *c = trans->c;
 815         int ret = 0, u64s_delta = 0;
 816
 817         trans_for_each_update(trans, i) {
 818                 if (i->cached)
 819                         continue;
 820
 821                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 822                 u64s_delta -= i->old_btree_u64s;
 823
 824                 if (!same_leaf_as_next(trans, i)) {
 825                         if (u64s_delta <= 0) {
 826                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 827                                                         i->level, flags);
 828                                 if (unlikely(ret))
 829                                         return ret;
 830                         }
 831
 832                         u64s_delta = 0;
 833                 }
 834         }
 835
 836         ret = bch2_trans_lock_write(trans);
 837         if (unlikely(ret))
 838                 return ret;
 839
 840         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 841
 842         if (!ret && unlikely(trans->journal_replay_not_finished))
 843                 bch2_drop_overwrites_from_journal(trans);
 844
 845         bch2_trans_unlock_write(trans);
 846
 847         if (!ret && trans->journal_pin)
 848                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 849                                      trans->journal_pin,
 850                                      bch2_trans_commit_journal_pin_flush);
 851
 852         /*
 853          * Drop journal reservation after dropping write locks, since dropping
 854          * the journal reservation may kick off a journal write:
 855          */
 856         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 857                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 858
 859         return ret;
 860 }
 861
 862 static int journal_reclaim_wait_done(struct bch_fs *c)
 863 {
 864         int ret = bch2_journal_error(&c->journal) ?:
 865                 !bch2_btree_key_cache_must_wait(c);
 866
 867         if (!ret)
 868                 journal_reclaim_kick(&c->journal);
 869         return ret;
 870 }
 871
 872 static noinline
 873 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 874                             struct btree_insert_entry *i,
 875                             int ret, unsigned long trace_ip)
 876 {
 877         struct bch_fs *c = trans->c;
 878
 879         switch (ret) {
 880         case -BCH_ERR_btree_insert_btree_node_full:
 881                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 882                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 883                         trace_and_count(c, trans_restart_btree_node_split, trans,
 884                                         trace_ip, trans->paths + i->path);
 885                 break;
 886         case -BCH_ERR_btree_insert_need_mark_replicas:
 887                 ret = drop_locks_do(trans,
 888                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 889                 break;
 890         case -BCH_ERR_journal_res_get_blocked:
 891                 /*
 892                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 893                  * flag
 894                  */
 895                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 896                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 897                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 898                         break;
 899                 }
 900
 901                 ret = drop_locks_do(trans,
 902                         bch2_trans_journal_res_get(trans,
 903                                         (flags & BCH_WATERMARK_MASK)|
 904                                         JOURNAL_RES_GET_CHECK));
 905                 break;
 906         case -BCH_ERR_btree_insert_need_journal_reclaim:
 907                 bch2_trans_unlock(trans);
 908
 909                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 910
 911                 wait_event_freezable(c->journal.reclaim_wait,
 912                                      (ret = journal_reclaim_wait_done(c)));
 913                 if (ret < 0)
 914                         break;
 915
 916                 ret = bch2_trans_relock(trans);
 917                 break;
 918         default:
 919                 BUG_ON(ret >= 0);
 920                 break;
 921         }
 922
 923         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 924
 925         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 926                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 927                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 928
 929         return ret;
 930 }
 931
 932 static noinline int
 933 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 934 {
 935         struct bch_fs *c = trans->c;
 936         int ret;
 937
 938         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 939             test_bit(BCH_FS_started, &c->flags))
 940                 return -BCH_ERR_erofs_trans_commit;
 941
 942         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 943         if (ret)
 944                 return ret;
 945
 946         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 947         return 0;
 948 }
 949
 950 /*
 951  * This is for updates done in the early part of fsck - btree_gc - before we've
 952  * gone RW. we only add the new key to the list of keys for journal replay to
 953  * do.
 954  */
 955 static noinline int
 956 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 957 {
 958         struct bch_fs *c = trans->c;
 959         int ret = 0;
 960
 961         trans_for_each_update(trans, i) {
 962                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 963                 if (ret)
 964                         break;
 965         }
 966
 967         return ret;
 968 }
 969
 970 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 971 {
 972         struct btree_insert_entry *errored_at = NULL;
 973         struct bch_fs *c = trans->c;
 974         int ret = 0;
 975
 976         if (!trans->nr_updates &&
 977             !trans->journal_entries_u64s)
 978                 goto out_reset;
 979
 980         ret = bch2_trans_commit_run_triggers(trans);
 981         if (ret)
 982                 goto out_reset;
 983
 984         trans_for_each_update(trans, i) {
 985                 struct printbuf buf = PRINTBUF;
 986                 enum bkey_invalid_flags invalid_flags = 0;
 987
 988                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
 989                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
 990
 991                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 992                                                i->bkey_type, invalid_flags, &buf)))
 993                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
 994                 btree_insert_entry_checks(trans, i);
 995                 printbuf_exit(&buf);
 996
 997                 if (ret)
 998                         return ret;
 999         }
1000
1001         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1002                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1003                 goto out_reset;
1004         }
1005
1006         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1007             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1008                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1009                 if (ret)
1010                         goto out_reset;
1011         }
1012
1013         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1014
1015         trans->journal_u64s             = trans->journal_entries_u64s;
1016         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1017         if (trans->journal_transaction_names)
1018                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1019
1020         trans_for_each_update(trans, i) {
1021                 struct btree_path *path = trans->paths + i->path;
1022
1023                 EBUG_ON(!path->should_be_locked);
1024
1025                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1026                 if (unlikely(ret))
1027                         goto out;
1028
1029                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1030
1031                 if (i->key_cache_already_flushed)
1032                         continue;
1033
1034                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1035                         continue;
1036
1037                 /* we're going to journal the key being updated: */
1038                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1039
1040                 /* and we're also going to log the overwrite: */
1041                 if (trans->journal_transaction_names)
1042                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1043         }
1044
1045         if (trans->extra_disk_res) {
1046                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1047                                 trans->extra_disk_res,
1048                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1049                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1050                 if (ret)
1051                         goto err;
1052         }
1053 retry:
1054         errored_at = NULL;
1055         bch2_trans_verify_not_in_restart(trans);
1056         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1057                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1058
1059         ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1060
1061         /* make sure we didn't drop or screw up locks: */
1062         bch2_trans_verify_locks(trans);
1063
1064         if (ret)
1065                 goto err;
1066
1067         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1068 out:
1069         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1070                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1071 out_reset:
1072         if (!ret)
1073                 bch2_trans_downgrade(trans);
1074         bch2_trans_reset_updates(trans);
1075
1076         return ret;
1077 err:
1078         ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1079         if (ret)
1080                 goto out;
1081
1082         /*
1083          * We might have done another transaction commit in the error path -
1084          * i.e. btree write buffer flush - which will have made use of
1085          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1086          * how the journal sequence number to pin is passed in - so we must
1087          * restart:
1088          */
1089         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1090                 ret = -BCH_ERR_transaction_restart_nested;
1091                 goto out;
1092         }
1093
1094         goto retry;
1095 }