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_io.h"
  16 #include "journal_reclaim.h"
  17 #include "replicas.h"
  18 #include "snapshot.h"
  19
  20 #include <linux/prefetch.h>
  21
  22 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  23 {
  24 #ifdef CONFIG_BCACHEFS_DEBUG
  25         struct bch_fs *c = trans->c;
  26         struct bkey u;
  27         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(trans->paths + i->path, &u);
  28
  29         if (unlikely(trans->journal_replay_not_finished)) {
  30                 struct bkey_i *j_k =
  31                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  32
  33                 if (j_k)
  34                         k = bkey_i_to_s_c(j_k);
  35         }
  36
  37         u = *k.k;
  38         u.needs_whiteout = i->old_k.needs_whiteout;
  39
  40         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  41         BUG_ON(i->old_v != k.v);
  42 #endif
  43 }
  44
  45 static inline struct btree_path_level *insert_l(struct btree_trans *trans, struct btree_insert_entry *i)
  46 {
  47         return (trans->paths + i->path)->l + i->level;
  48 }
  49
  50 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  51                                      struct btree_insert_entry *i)
  52 {
  53         return i != trans->updates &&
  54                 insert_l(trans, &i[0])->b == insert_l(trans, &i[-1])->b;
  55 }
  56
  57 static inline bool same_leaf_as_next(struct btree_trans *trans,
  58                                      struct btree_insert_entry *i)
  59 {
  60         return i + 1 < trans->updates + trans->nr_updates &&
  61                 insert_l(trans, &i[0])->b == insert_l(trans, &i[1])->b;
  62 }
  63
  64 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  65                                            struct btree_path *path,
  66                                            struct btree *b)
  67 {
  68         struct bch_fs *c = trans->c;
  69
  70         if (unlikely(btree_node_just_written(b)) &&
  71             bch2_btree_post_write_cleanup(c, b))
  72                 bch2_trans_node_reinit_iter(trans, b);
  73
  74         /*
  75          * If the last bset has been written, or if it's gotten too big - start
  76          * a new bset to insert into:
  77          */
  78         if (want_new_bset(c, b))
  79                 bch2_btree_init_next(trans, b);
  80 }
  81
  82 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
  83 {
  84         while (--i >= trans->updates) {
  85                 if (same_leaf_as_prev(trans, i))
  86                         continue;
  87
  88                 bch2_btree_node_unlock_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
  89         }
  90
  91         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
  92         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
  93 }
  94
  95 static inline int bch2_trans_lock_write(struct btree_trans *trans)
  96 {
  97         EBUG_ON(trans->write_locked);
  98
  99         trans_for_each_update(trans, i) {
 100                 if (same_leaf_as_prev(trans, i))
 101                         continue;
 102
 103                 if (bch2_btree_node_lock_write(trans, trans->paths + i->path, &insert_l(trans, i)->b->c))
 104                         return trans_lock_write_fail(trans, i);
 105
 106                 if (!i->cached)
 107                         bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 108         }
 109
 110         trans->write_locked = true;
 111         return 0;
 112 }
 113
 114 static inline void bch2_trans_unlock_write(struct btree_trans *trans)
 115 {
 116         if (likely(trans->write_locked)) {
 117                 trans_for_each_update(trans, i)
 118                         if (!same_leaf_as_prev(trans, i))
 119                                 bch2_btree_node_unlock_write_inlined(trans,
 120                                                 trans->paths + i->path, insert_l(trans, i)->b);
 121                 trans->write_locked = false;
 122         }
 123 }
 124
 125 /* Inserting into a given leaf node (last stage of insert): */
 126
 127 /* Handle overwrites and do insert, for non extents: */
 128 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 129                                 struct btree_path *path,
 130                                 struct btree *b,
 131                                 struct btree_node_iter *node_iter,
 132                                 struct bkey_i *insert)
 133 {
 134         struct bkey_packed *k;
 135         unsigned clobber_u64s = 0, new_u64s = 0;
 136
 137         EBUG_ON(btree_node_just_written(b));
 138         EBUG_ON(bset_written(b, btree_bset_last(b)));
 139         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 140         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 141         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 142         EBUG_ON(insert->k.u64s >
 143                 bch_btree_keys_u64s_remaining(trans->c, b));
 144         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 145
 146         k = bch2_btree_node_iter_peek_all(node_iter, b);
 147         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 148                 k = NULL;
 149
 150         /* @k is the key being overwritten/deleted, if any: */
 151         EBUG_ON(k && bkey_deleted(k));
 152
 153         /* Deleting, but not found? nothing to do: */
 154         if (bkey_deleted(&insert->k) && !k)
 155                 return false;
 156
 157         if (bkey_deleted(&insert->k)) {
 158                 /* Deleting: */
 159                 btree_account_key_drop(b, k);
 160                 k->type = KEY_TYPE_deleted;
 161
 162                 if (k->needs_whiteout)
 163                         push_whiteout(trans->c, b, insert->k.p);
 164                 k->needs_whiteout = false;
 165
 166                 if (k >= btree_bset_last(b)->start) {
 167                         clobber_u64s = k->u64s;
 168                         bch2_bset_delete(b, k, clobber_u64s);
 169                         goto fix_iter;
 170                 } else {
 171                         bch2_btree_path_fix_key_modified(trans, b, k);
 172                 }
 173
 174                 return true;
 175         }
 176
 177         if (k) {
 178                 /* Overwriting: */
 179                 btree_account_key_drop(b, k);
 180                 k->type = KEY_TYPE_deleted;
 181
 182                 insert->k.needs_whiteout = k->needs_whiteout;
 183                 k->needs_whiteout = false;
 184
 185                 if (k >= btree_bset_last(b)->start) {
 186                         clobber_u64s = k->u64s;
 187                         goto overwrite;
 188                 } else {
 189                         bch2_btree_path_fix_key_modified(trans, b, k);
 190                 }
 191         }
 192
 193         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 194 overwrite:
 195         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 196         new_u64s = k->u64s;
 197 fix_iter:
 198         if (clobber_u64s != new_u64s)
 199                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 200                                          clobber_u64s, new_u64s);
 201         return true;
 202 }
 203
 204 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 205                                unsigned i, u64 seq)
 206 {
 207         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 208         struct btree_write *w = container_of(pin, struct btree_write, journal);
 209         struct btree *b = container_of(w, struct btree, writes[i]);
 210         struct btree_trans *trans = bch2_trans_get(c);
 211         unsigned long old, new, v;
 212         unsigned idx = w - b->writes;
 213
 214         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 215         v = READ_ONCE(b->flags);
 216
 217         do {
 218                 old = new = v;
 219
 220                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 221                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 222                     w->journal.seq != seq)
 223                         break;
 224
 225                 new &= ~BTREE_WRITE_TYPE_MASK;
 226                 new |= BTREE_WRITE_journal_reclaim;
 227                 new |= 1 << BTREE_NODE_need_write;
 228         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 229
 230         btree_node_write_if_need(c, b, SIX_LOCK_read);
 231         six_unlock_read(&b->c.lock);
 232
 233         bch2_trans_put(trans);
 234         return 0;
 235 }
 236
 237 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 238 {
 239         return __btree_node_flush(j, pin, 0, seq);
 240 }
 241
 242 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 243 {
 244         return __btree_node_flush(j, pin, 1, seq);
 245 }
 246
 247 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 248                                        struct btree *b, u64 seq)
 249 {
 250         struct btree_write *w = btree_current_write(b);
 251
 252         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 253                              btree_node_write_idx(b) == 0
 254                              ? bch2_btree_node_flush0
 255                              : bch2_btree_node_flush1);
 256 }
 257
 258 /**
 259  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 260  * @trans:              btree transaction object
 261  * @path:               path pointing to @insert's pos
 262  * @insert:             key to insert
 263  * @journal_seq:        sequence number of journal reservation
 264  */
 265 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 266                                        struct btree_path *path,
 267                                        struct bkey_i *insert,
 268                                        u64 journal_seq)
 269 {
 270         struct bch_fs *c = trans->c;
 271         struct btree *b = path_l(path)->b;
 272         struct bset_tree *t = bset_tree_last(b);
 273         struct bset *i = bset(b, t);
 274         int old_u64s = bset_u64s(t);
 275         int old_live_u64s = b->nr.live_u64s;
 276         int live_u64s_added, u64s_added;
 277
 278         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 279                                         &path_l(path)->iter, insert)))
 280                 return;
 281
 282         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 283
 284         bch2_btree_add_journal_pin(c, b, journal_seq);
 285
 286         if (unlikely(!btree_node_dirty(b))) {
 287                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 288                 set_btree_node_dirty_acct(c, b);
 289         }
 290
 291         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 292         u64s_added = (int) bset_u64s(t) - old_u64s;
 293
 294         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 295                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 296         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 297                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 298
 299         if (u64s_added > live_u64s_added &&
 300             bch2_maybe_compact_whiteouts(c, b))
 301                 bch2_trans_node_reinit_iter(trans, b);
 302 }
 303
 304 /* Cached btree updates: */
 305
 306 /* Normal update interface: */
 307
 308 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 309                                              struct btree_insert_entry *i)
 310 {
 311         struct btree_path *path = trans->paths + i->path;
 312
 313         BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 314         BUG_ON(i->cached        != path->cached);
 315         BUG_ON(i->level         != path->level);
 316         BUG_ON(i->btree_id      != path->btree_id);
 317         EBUG_ON(!i->level &&
 318                 btree_type_has_snapshots(i->btree_id) &&
 319                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 320                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 321                 i->k->k.p.snapshot &&
 322                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 323 }
 324
 325 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 326                                                       unsigned flags)
 327 {
 328         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 329                                     trans->journal_u64s, flags);
 330 }
 331
 332 #define JSET_ENTRY_LOG_U64s             4
 333
 334 static noinline void journal_transaction_name(struct btree_trans *trans)
 335 {
 336         struct bch_fs *c = trans->c;
 337         struct journal *j = &c->journal;
 338         struct jset_entry *entry =
 339                 bch2_journal_add_entry(j, &trans->journal_res,
 340                                        BCH_JSET_ENTRY_log, 0, 0,
 341                                        JSET_ENTRY_LOG_U64s);
 342         struct jset_entry_log *l =
 343                 container_of(entry, struct jset_entry_log, entry);
 344
 345         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 346 }
 347
 348 static inline int btree_key_can_insert(struct btree_trans *trans,
 349                                        struct btree *b, unsigned u64s)
 350 {
 351         struct bch_fs *c = trans->c;
 352
 353         if (!bch2_btree_node_insert_fits(c, b, u64s))
 354                 return -BCH_ERR_btree_insert_btree_node_full;
 355
 356         return 0;
 357 }
 358
 359 noinline static int
 360 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 361                                      struct btree_path *path, unsigned new_u64s)
 362 {
 363         struct bkey_cached *ck = (void *) path->l[0].b;
 364         struct bkey_i *new_k;
 365         int ret;
 366
 367         bch2_trans_unlock_write(trans);
 368         bch2_trans_unlock(trans);
 369
 370         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 371         if (!new_k) {
 372                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 373                         bch2_btree_id_str(path->btree_id), new_u64s);
 374                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 375         }
 376
 377         ret =   bch2_trans_relock(trans) ?:
 378                 bch2_trans_lock_write(trans);
 379         if (unlikely(ret)) {
 380                 kfree(new_k);
 381                 return ret;
 382         }
 383
 384         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 385
 386         trans_for_each_update(trans, i)
 387                 if (i->old_v == &ck->k->v)
 388                         i->old_v = &new_k->v;
 389
 390         kfree(ck->k);
 391         ck->u64s        = new_u64s;
 392         ck->k           = new_k;
 393         return 0;
 394 }
 395
 396 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 397                                        struct btree_path *path, unsigned u64s)
 398 {
 399         struct bch_fs *c = trans->c;
 400         struct bkey_cached *ck = (void *) path->l[0].b;
 401         unsigned new_u64s;
 402         struct bkey_i *new_k;
 403
 404         EBUG_ON(path->level);
 405
 406         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 407             bch2_btree_key_cache_must_wait(c) &&
 408             !(flags & BCH_TRANS_COMMIT_journal_reclaim))
 409                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 410
 411         /*
 412          * bch2_varint_decode can read past the end of the buffer by at most 7
 413          * bytes (it won't be used):
 414          */
 415         u64s += 1;
 416
 417         if (u64s <= ck->u64s)
 418                 return 0;
 419
 420         new_u64s        = roundup_pow_of_two(u64s);
 421         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT);
 422         if (unlikely(!new_k))
 423                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 424
 425         trans_for_each_update(trans, i)
 426                 if (i->old_v == &ck->k->v)
 427                         i->old_v = &new_k->v;
 428
 429         ck->u64s        = new_u64s;
 430         ck->k           = new_k;
 431         return 0;
 432 }
 433
 434 /* Triggers: */
 435
 436 static int run_one_mem_trigger(struct btree_trans *trans,
 437                                struct btree_insert_entry *i,
 438                                unsigned flags)
 439 {
 440         struct bkey_s_c old = { &i->old_k, i->old_v };
 441         struct bkey_i *new = i->k;
 442         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 443         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 444         int ret;
 445
 446         verify_update_old_key(trans, i);
 447
 448         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 449                 return 0;
 450
 451         if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
 452                 return 0;
 453
 454         if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
 455                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 456                                 old, bkey_i_to_s_c(new),
 457                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 458         } else {
 459                 struct bkey             _deleted = POS_KEY((trans->paths + i->path)->pos);
 460                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 461
 462                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 463                                 deleted, bkey_i_to_s_c(new),
 464                                 BTREE_TRIGGER_INSERT|flags) ?:
 465                         bch2_mark_key(trans, i->btree_id, i->level,
 466                                 old, deleted,
 467                                 BTREE_TRIGGER_OVERWRITE|flags);
 468         }
 469
 470         return ret;
 471 }
 472
 473 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 474                                  bool overwrite)
 475 {
 476         /*
 477          * Transactional triggers create new btree_insert_entries, so we can't
 478          * pass them a pointer to a btree_insert_entry, that memory is going to
 479          * move:
 480          */
 481         struct bkey old_k = i->old_k;
 482         struct bkey_s_c old = { &old_k, i->old_v };
 483         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 484         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 485
 486         verify_update_old_key(trans, i);
 487
 488         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 489             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 490                 return 0;
 491
 492         if (!i->insert_trigger_run &&
 493             !i->overwrite_trigger_run &&
 494             old_ops->trans_trigger == new_ops->trans_trigger) {
 495                 i->overwrite_trigger_run = true;
 496                 i->insert_trigger_run = true;
 497                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 498                                            BTREE_TRIGGER_INSERT|
 499                                            BTREE_TRIGGER_OVERWRITE|
 500                                            i->flags) ?: 1;
 501         } else if (overwrite && !i->overwrite_trigger_run) {
 502                 i->overwrite_trigger_run = true;
 503                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 504         } else if (!overwrite && !i->insert_trigger_run) {
 505                 i->insert_trigger_run = true;
 506                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 507         } else {
 508                 return 0;
 509         }
 510 }
 511
 512 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 513                               struct btree_insert_entry *btree_id_start)
 514 {
 515         struct btree_insert_entry *i;
 516         bool trans_trigger_run;
 517         int ret, overwrite;
 518
 519         for (overwrite = 1; overwrite >= 0; --overwrite) {
 520
 521                 /*
 522                  * Running triggers will append more updates to the list of updates as
 523                  * we're walking it:
 524                  */
 525                 do {
 526                         trans_trigger_run = false;
 527
 528                         for (i = btree_id_start;
 529                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 530                              i++) {
 531                                 if (i->btree_id != btree_id)
 532                                         continue;
 533
 534                                 ret = run_one_trans_trigger(trans, i, overwrite);
 535                                 if (ret < 0)
 536                                         return ret;
 537                                 if (ret)
 538                                         trans_trigger_run = true;
 539                         }
 540                 } while (trans_trigger_run);
 541         }
 542
 543         return 0;
 544 }
 545
 546 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 547 {
 548         struct btree_insert_entry *btree_id_start = trans->updates;
 549         unsigned btree_id = 0;
 550         int ret = 0;
 551
 552         /*
 553          *
 554          * For a given btree, this algorithm runs insert triggers before
 555          * overwrite triggers: this is so that when extents are being moved
 556          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 557          * they are re-added.
 558          */
 559         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 560                 if (btree_id == BTREE_ID_alloc)
 561                         continue;
 562
 563                 while (btree_id_start < trans->updates + trans->nr_updates &&
 564                        btree_id_start->btree_id < btree_id)
 565                         btree_id_start++;
 566
 567                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 568                 if (ret)
 569                         return ret;
 570         }
 571
 572         trans_for_each_update(trans, i) {
 573                 if (i->btree_id > BTREE_ID_alloc)
 574                         break;
 575                 if (i->btree_id == BTREE_ID_alloc) {
 576                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 577                         if (ret)
 578                                 return ret;
 579                         break;
 580                 }
 581         }
 582
 583 #ifdef CONFIG_BCACHEFS_DEBUG
 584         trans_for_each_update(trans, i)
 585                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 586                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 587                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 588 #endif
 589         return 0;
 590 }
 591
 592 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 593 {
 594         struct bch_fs *c = trans->c;
 595         int ret = 0;
 596
 597         trans_for_each_update(trans, i) {
 598                 /*
 599                  * XXX: synchronization of cached update triggers with gc
 600                  * XXX: synchronization of interior node updates with gc
 601                  */
 602                 BUG_ON(i->cached || i->level);
 603
 604                 if (gc_visited(c, gc_pos_btree_node(insert_l(trans, i)->b))) {
 605                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 606                         if (ret)
 607                                 break;
 608                 }
 609         }
 610
 611         return ret;
 612 }
 613
 614 static inline int
 615 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 616                                struct btree_insert_entry **stopped_at,
 617                                unsigned long trace_ip)
 618 {
 619         struct bch_fs *c = trans->c;
 620         struct btree_trans_commit_hook *h;
 621         unsigned u64s = 0;
 622         int ret;
 623
 624         if (race_fault()) {
 625                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 626                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 627         }
 628
 629         /*
 630          * Check if the insert will fit in the leaf node with the write lock
 631          * held, otherwise another thread could write the node changing the
 632          * amount of space available:
 633          */
 634
 635         prefetch(&trans->c->journal.flags);
 636
 637         trans_for_each_update(trans, i) {
 638                 /* Multiple inserts might go to same leaf: */
 639                 if (!same_leaf_as_prev(trans, i))
 640                         u64s = 0;
 641
 642                 u64s += i->k->k.u64s;
 643                 ret = !i->cached
 644                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 645                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 646                 if (ret) {
 647                         *stopped_at = i;
 648                         return ret;
 649                 }
 650
 651                 i->k->k.needs_whiteout = false;
 652         }
 653
 654         /*
 655          * Don't get journal reservation until after we know insert will
 656          * succeed:
 657          */
 658         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 659                 ret = bch2_trans_journal_res_get(trans,
 660                                 (flags & BCH_WATERMARK_MASK)|
 661                                 JOURNAL_RES_GET_NONBLOCK);
 662                 if (ret)
 663                         return ret;
 664
 665                 if (unlikely(trans->journal_transaction_names))
 666                         journal_transaction_name(trans);
 667         }
 668
 669         /*
 670          * Not allowed to fail after we've gotten our journal reservation - we
 671          * have to use it:
 672          */
 673
 674         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 675             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 676                 if (bch2_journal_seq_verify)
 677                         trans_for_each_update(trans, i)
 678                                 i->k->k.version.lo = trans->journal_res.seq;
 679                 else if (bch2_inject_invalid_keys)
 680                         trans_for_each_update(trans, i)
 681                                 i->k->k.version = MAX_VERSION;
 682         }
 683
 684         if (trans->fs_usage_deltas &&
 685             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 686                 return -BCH_ERR_btree_insert_need_mark_replicas;
 687
 688         h = trans->hooks;
 689         while (h) {
 690                 ret = h->fn(trans, h);
 691                 if (ret)
 692                         goto revert_fs_usage;
 693                 h = h->next;
 694         }
 695
 696         trans_for_each_update(trans, i)
 697                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 698                         ret = run_one_mem_trigger(trans, i, i->flags);
 699                         if (ret)
 700                                 goto fatal_err;
 701                 }
 702
 703         if (unlikely(c->gc_pos.phase)) {
 704                 ret = bch2_trans_commit_run_gc_triggers(trans);
 705                 if  (ret)
 706                         goto fatal_err;
 707         }
 708
 709         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 710                 struct journal *j = &c->journal;
 711                 struct jset_entry *entry;
 712
 713                 trans_for_each_update(trans, i) {
 714                         if (i->key_cache_already_flushed)
 715                                 continue;
 716
 717                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 718                                 continue;
 719
 720                         verify_update_old_key(trans, i);
 721
 722                         if (trans->journal_transaction_names) {
 723                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 724                                                        BCH_JSET_ENTRY_overwrite,
 725                                                        i->btree_id, i->level,
 726                                                        i->old_k.u64s);
 727                                 bkey_reassemble((struct bkey_i *) entry->start,
 728                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 729                         }
 730
 731                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 732                                                BCH_JSET_ENTRY_btree_keys,
 733                                                i->btree_id, i->level,
 734                                                i->k->k.u64s);
 735                         bkey_copy((struct bkey_i *) entry->start, i->k);
 736                 }
 737
 738                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 739                                   trans->journal_entries,
 740                                   trans->journal_entries_u64s);
 741
 742                 trans->journal_res.offset       += trans->journal_entries_u64s;
 743                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 744
 745                 if (trans->journal_seq)
 746                         *trans->journal_seq = trans->journal_res.seq;
 747         }
 748
 749         trans_for_each_update(trans, i) {
 750                 struct btree_path *path = trans->paths + i->path;
 751
 752                 if (!i->cached) {
 753                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 754                 } else if (!i->key_cache_already_flushed)
 755                         bch2_btree_insert_key_cached(trans, flags, i);
 756                 else {
 757                         bch2_btree_key_cache_drop(trans, path);
 758                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 759                 }
 760         }
 761
 762         return 0;
 763 fatal_err:
 764         bch2_fatal_error(c);
 765 revert_fs_usage:
 766         if (trans->fs_usage_deltas)
 767                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 768         return ret;
 769 }
 770
 771 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 772 {
 773         trans_for_each_update(trans, i)
 774                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 775 }
 776
 777 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 778                                                    enum bkey_invalid_flags flags,
 779                                                    struct btree_insert_entry *i,
 780                                                    struct printbuf *err)
 781 {
 782         struct bch_fs *c = trans->c;
 783
 784         printbuf_reset(err);
 785         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 786                    trans->fn, (void *) i->ip_allocated);
 787         prt_newline(err);
 788         printbuf_indent_add(err, 2);
 789
 790         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 791         prt_newline(err);
 792
 793         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 794         bch2_print_string_as_lines(KERN_ERR, err->buf);
 795
 796         bch2_inconsistent_error(c);
 797         bch2_dump_trans_updates(trans);
 798
 799         return -EINVAL;
 800 }
 801
 802 static noinline int bch2_trans_commit_journal_entry_invalid(struct btree_trans *trans,
 803                                                    struct jset_entry *i)
 804 {
 805         struct bch_fs *c = trans->c;
 806         struct printbuf buf = PRINTBUF;
 807
 808         prt_printf(&buf, "invalid bkey on insert from %s", trans->fn);
 809         prt_newline(&buf);
 810         printbuf_indent_add(&buf, 2);
 811
 812         bch2_journal_entry_to_text(&buf, c, i);
 813         prt_newline(&buf);
 814
 815         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 816
 817         bch2_inconsistent_error(c);
 818         bch2_dump_trans_updates(trans);
 819
 820         return -EINVAL;
 821 }
 822
 823 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 824                                 struct journal_entry_pin *_pin, u64 seq)
 825 {
 826         return 0;
 827 }
 828
 829 /*
 830  * Get journal reservation, take write locks, and attempt to do btree update(s):
 831  */
 832 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 833                                        struct btree_insert_entry **stopped_at,
 834                                        unsigned long trace_ip)
 835 {
 836         struct bch_fs *c = trans->c;
 837         int ret = 0, u64s_delta = 0;
 838
 839         trans_for_each_update(trans, i) {
 840                 if (i->cached)
 841                         continue;
 842
 843                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 844                 u64s_delta -= i->old_btree_u64s;
 845
 846                 if (!same_leaf_as_next(trans, i)) {
 847                         if (u64s_delta <= 0) {
 848                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 849                                                         i->level, flags);
 850                                 if (unlikely(ret))
 851                                         return ret;
 852                         }
 853
 854                         u64s_delta = 0;
 855                 }
 856         }
 857
 858         ret = bch2_trans_lock_write(trans);
 859         if (unlikely(ret))
 860                 return ret;
 861
 862         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 863
 864         if (!ret && unlikely(trans->journal_replay_not_finished))
 865                 bch2_drop_overwrites_from_journal(trans);
 866
 867         bch2_trans_unlock_write(trans);
 868
 869         if (!ret && trans->journal_pin)
 870                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 871                                      trans->journal_pin,
 872                                      bch2_trans_commit_journal_pin_flush);
 873
 874         /*
 875          * Drop journal reservation after dropping write locks, since dropping
 876          * the journal reservation may kick off a journal write:
 877          */
 878         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 879                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 880
 881         return ret;
 882 }
 883
 884 static int journal_reclaim_wait_done(struct bch_fs *c)
 885 {
 886         int ret = bch2_journal_error(&c->journal) ?:
 887                 !bch2_btree_key_cache_must_wait(c);
 888
 889         if (!ret)
 890                 journal_reclaim_kick(&c->journal);
 891         return ret;
 892 }
 893
 894 static noinline
 895 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 896                             struct btree_insert_entry *i,
 897                             int ret, unsigned long trace_ip)
 898 {
 899         struct bch_fs *c = trans->c;
 900
 901         switch (ret) {
 902         case -BCH_ERR_btree_insert_btree_node_full:
 903                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 904                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 905                         trace_and_count(c, trans_restart_btree_node_split, trans,
 906                                         trace_ip, trans->paths + i->path);
 907                 break;
 908         case -BCH_ERR_btree_insert_need_mark_replicas:
 909                 ret = drop_locks_do(trans,
 910                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 911                 break;
 912         case -BCH_ERR_journal_res_get_blocked:
 913                 /*
 914                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 915                  * flag
 916                  */
 917                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 918                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 919                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 920                         break;
 921                 }
 922
 923                 ret = drop_locks_do(trans,
 924                         bch2_trans_journal_res_get(trans,
 925                                         (flags & BCH_WATERMARK_MASK)|
 926                                         JOURNAL_RES_GET_CHECK));
 927                 break;
 928         case -BCH_ERR_btree_insert_need_journal_reclaim:
 929                 bch2_trans_unlock(trans);
 930
 931                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 932
 933                 wait_event_freezable(c->journal.reclaim_wait,
 934                                      (ret = journal_reclaim_wait_done(c)));
 935                 if (ret < 0)
 936                         break;
 937
 938                 ret = bch2_trans_relock(trans);
 939                 break;
 940         default:
 941                 BUG_ON(ret >= 0);
 942                 break;
 943         }
 944
 945         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 946
 947         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 948                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 949                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 950
 951         return ret;
 952 }
 953
 954 static noinline int
 955 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 956 {
 957         struct bch_fs *c = trans->c;
 958         int ret;
 959
 960         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 961             test_bit(BCH_FS_started, &c->flags))
 962                 return -BCH_ERR_erofs_trans_commit;
 963
 964         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 965         if (ret)
 966                 return ret;
 967
 968         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 969         return 0;
 970 }
 971
 972 /*
 973  * This is for updates done in the early part of fsck - btree_gc - before we've
 974  * gone RW. we only add the new key to the list of keys for journal replay to
 975  * do.
 976  */
 977 static noinline int
 978 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 979 {
 980         struct bch_fs *c = trans->c;
 981         int ret = 0;
 982
 983         trans_for_each_update(trans, i) {
 984                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 985                 if (ret)
 986                         break;
 987         }
 988
 989         return ret;
 990 }
 991
 992 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 993 {
 994         struct btree_insert_entry *errored_at = NULL;
 995         struct bch_fs *c = trans->c;
 996         int ret = 0;
 997
 998         if (!trans->nr_updates &&
 999             !trans->journal_entries_u64s)
1000                 goto out_reset;
1001
1002         ret = bch2_trans_commit_run_triggers(trans);
1003         if (ret)
1004                 goto out_reset;
1005
1006         trans_for_each_update(trans, i) {
1007                 struct printbuf buf = PRINTBUF;
1008                 enum bkey_invalid_flags invalid_flags = 0;
1009
1010                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1011                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1012
1013                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1014                                                i->bkey_type, invalid_flags, &buf)))
1015                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1016                 btree_insert_entry_checks(trans, i);
1017                 printbuf_exit(&buf);
1018
1019                 if (ret)
1020                         return ret;
1021         }
1022
1023         for (struct jset_entry *i = trans->journal_entries;
1024              i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
1025              i = vstruct_next(i)) {
1026                 if (!jset_entry_is_key(i))
1027                         continue;
1028
1029                 enum bkey_invalid_flags invalid_flags = 0;
1030
1031                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1032                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1033
1034                 if (unlikely(bch2_journal_entry_validate(c, NULL, i,
1035                                         bcachefs_metadata_version_current,
1036                                         CPU_BIG_ENDIAN, invalid_flags)))
1037                         ret = bch2_trans_commit_journal_entry_invalid(trans, i);
1038
1039                 if (ret)
1040                         return ret;
1041         }
1042
1043         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1044                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1045                 goto out_reset;
1046         }
1047
1048         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1049             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1050                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1051                 if (ret)
1052                         goto out_reset;
1053         }
1054
1055         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1056
1057         trans->journal_u64s             = trans->journal_entries_u64s;
1058         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1059         if (trans->journal_transaction_names)
1060                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1061
1062         trans_for_each_update(trans, i) {
1063                 struct btree_path *path = trans->paths + i->path;
1064
1065                 EBUG_ON(!path->should_be_locked);
1066
1067                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1068                 if (unlikely(ret))
1069                         goto out;
1070
1071                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1072
1073                 if (i->key_cache_already_flushed)
1074                         continue;
1075
1076                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1077                         continue;
1078
1079                 /* we're going to journal the key being updated: */
1080                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1081
1082                 /* and we're also going to log the overwrite: */
1083                 if (trans->journal_transaction_names)
1084                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1085         }
1086
1087         if (trans->extra_disk_res) {
1088                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1089                                 trans->extra_disk_res,
1090                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1091                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1092                 if (ret)
1093                         goto err;
1094         }
1095 retry:
1096         errored_at = NULL;
1097         bch2_trans_verify_not_in_restart(trans);
1098         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1099                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1100
1101         ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1102
1103         /* make sure we didn't drop or screw up locks: */
1104         bch2_trans_verify_locks(trans);
1105
1106         if (ret)
1107                 goto err;
1108
1109         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1110 out:
1111         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1112                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1113 out_reset:
1114         if (!ret)
1115                 bch2_trans_downgrade(trans);
1116         bch2_trans_reset_updates(trans);
1117
1118         return ret;
1119 err:
1120         ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1121         if (ret)
1122                 goto out;
1123
1124         /*
1125          * We might have done another transaction commit in the error path -
1126          * i.e. btree write buffer flush - which will have made use of
1127          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1128          * how the journal sequence number to pin is passed in - so we must
1129          * restart:
1130          */
1131         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1132                 ret = -BCH_ERR_transaction_restart_nested;
1133                 goto out;
1134         }
1135
1136         goto retry;
1137 }