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->trigger == new_ops->trigger) {
 455                 ret   = bch2_key_trigger(trans, i->btree_id, i->level,
 456                                 old, bkey_i_to_s(new),
 457                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 458         } else {
 459                 ret   = bch2_key_trigger_new(trans, i->btree_id, i->level,
 460                                 bkey_i_to_s(new), flags) ?:
 461                         bch2_key_trigger_old(trans, i->btree_id, i->level,
 462                                 old, flags);
 463         }
 464
 465         return ret;
 466 }
 467
 468 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 469                                  bool overwrite)
 470 {
 471         /*
 472          * Transactional triggers create new btree_insert_entries, so we can't
 473          * pass them a pointer to a btree_insert_entry, that memory is going to
 474          * move:
 475          */
 476         struct bkey old_k = i->old_k;
 477         struct bkey_s_c old = { &old_k, i->old_v };
 478         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 479         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 480         unsigned flags = i->flags|BTREE_TRIGGER_TRANSACTIONAL;
 481
 482         verify_update_old_key(trans, i);
 483
 484         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 485             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 486                 return 0;
 487
 488         if (!i->insert_trigger_run &&
 489             !i->overwrite_trigger_run &&
 490             old_ops->trigger == new_ops->trigger) {
 491                 i->overwrite_trigger_run = true;
 492                 i->insert_trigger_run = true;
 493                 return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
 494                                         BTREE_TRIGGER_INSERT|
 495                                         BTREE_TRIGGER_OVERWRITE|flags) ?: 1;
 496         } else if (overwrite && !i->overwrite_trigger_run) {
 497                 i->overwrite_trigger_run = true;
 498                 return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
 499         } else if (!overwrite && !i->insert_trigger_run) {
 500                 i->insert_trigger_run = true;
 501                 return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 502         } else {
 503                 return 0;
 504         }
 505 }
 506
 507 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 508                               struct btree_insert_entry *btree_id_start)
 509 {
 510         struct btree_insert_entry *i;
 511         bool trans_trigger_run;
 512         int ret, overwrite;
 513
 514         for (overwrite = 1; overwrite >= 0; --overwrite) {
 515
 516                 /*
 517                  * Running triggers will append more updates to the list of updates as
 518                  * we're walking it:
 519                  */
 520                 do {
 521                         trans_trigger_run = false;
 522
 523                         for (i = btree_id_start;
 524                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 525                              i++) {
 526                                 if (i->btree_id != btree_id)
 527                                         continue;
 528
 529                                 ret = run_one_trans_trigger(trans, i, overwrite);
 530                                 if (ret < 0)
 531                                         return ret;
 532                                 if (ret)
 533                                         trans_trigger_run = true;
 534                         }
 535                 } while (trans_trigger_run);
 536         }
 537
 538         return 0;
 539 }
 540
 541 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 542 {
 543         struct btree_insert_entry *btree_id_start = trans->updates;
 544         unsigned btree_id = 0;
 545         int ret = 0;
 546
 547         /*
 548          *
 549          * For a given btree, this algorithm runs insert triggers before
 550          * overwrite triggers: this is so that when extents are being moved
 551          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 552          * they are re-added.
 553          */
 554         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 555                 if (btree_id == BTREE_ID_alloc)
 556                         continue;
 557
 558                 while (btree_id_start < trans->updates + trans->nr_updates &&
 559                        btree_id_start->btree_id < btree_id)
 560                         btree_id_start++;
 561
 562                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 563                 if (ret)
 564                         return ret;
 565         }
 566
 567         trans_for_each_update(trans, i) {
 568                 if (i->btree_id > BTREE_ID_alloc)
 569                         break;
 570                 if (i->btree_id == BTREE_ID_alloc) {
 571                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 572                         if (ret)
 573                                 return ret;
 574                         break;
 575                 }
 576         }
 577
 578 #ifdef CONFIG_BCACHEFS_DEBUG
 579         trans_for_each_update(trans, i)
 580                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 581                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 582                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 583 #endif
 584         return 0;
 585 }
 586
 587 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 588 {
 589         struct bch_fs *c = trans->c;
 590         int ret = 0;
 591
 592         trans_for_each_update(trans, i) {
 593                 /*
 594                  * XXX: synchronization of cached update triggers with gc
 595                  * XXX: synchronization of interior node updates with gc
 596                  */
 597                 BUG_ON(i->cached || i->level);
 598
 599                 if (gc_visited(c, gc_pos_btree_node(insert_l(trans, i)->b))) {
 600                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 601                         if (ret)
 602                                 break;
 603                 }
 604         }
 605
 606         return ret;
 607 }
 608
 609 static inline int
 610 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 611                                struct btree_insert_entry **stopped_at,
 612                                unsigned long trace_ip)
 613 {
 614         struct bch_fs *c = trans->c;
 615         struct btree_trans_commit_hook *h;
 616         unsigned u64s = 0;
 617         int ret;
 618
 619         if (race_fault()) {
 620                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 621                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 622         }
 623
 624         /*
 625          * Check if the insert will fit in the leaf node with the write lock
 626          * held, otherwise another thread could write the node changing the
 627          * amount of space available:
 628          */
 629
 630         prefetch(&trans->c->journal.flags);
 631
 632         trans_for_each_update(trans, i) {
 633                 /* Multiple inserts might go to same leaf: */
 634                 if (!same_leaf_as_prev(trans, i))
 635                         u64s = 0;
 636
 637                 u64s += i->k->k.u64s;
 638                 ret = !i->cached
 639                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 640                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 641                 if (ret) {
 642                         *stopped_at = i;
 643                         return ret;
 644                 }
 645
 646                 i->k->k.needs_whiteout = false;
 647         }
 648
 649         /*
 650          * Don't get journal reservation until after we know insert will
 651          * succeed:
 652          */
 653         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 654                 ret = bch2_trans_journal_res_get(trans,
 655                                 (flags & BCH_WATERMARK_MASK)|
 656                                 JOURNAL_RES_GET_NONBLOCK);
 657                 if (ret)
 658                         return ret;
 659
 660                 if (unlikely(trans->journal_transaction_names))
 661                         journal_transaction_name(trans);
 662         }
 663
 664         /*
 665          * Not allowed to fail after we've gotten our journal reservation - we
 666          * have to use it:
 667          */
 668
 669         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 670             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 671                 if (bch2_journal_seq_verify)
 672                         trans_for_each_update(trans, i)
 673                                 i->k->k.version.lo = trans->journal_res.seq;
 674                 else if (bch2_inject_invalid_keys)
 675                         trans_for_each_update(trans, i)
 676                                 i->k->k.version = MAX_VERSION;
 677         }
 678
 679         if (trans->fs_usage_deltas &&
 680             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 681                 return -BCH_ERR_btree_insert_need_mark_replicas;
 682
 683         h = trans->hooks;
 684         while (h) {
 685                 ret = h->fn(trans, h);
 686                 if (ret)
 687                         goto revert_fs_usage;
 688                 h = h->next;
 689         }
 690
 691         trans_for_each_update(trans, i)
 692                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 693                         ret = run_one_mem_trigger(trans, i, i->flags);
 694                         if (ret)
 695                                 goto fatal_err;
 696                 }
 697
 698         if (unlikely(c->gc_pos.phase)) {
 699                 ret = bch2_trans_commit_run_gc_triggers(trans);
 700                 if  (ret)
 701                         goto fatal_err;
 702         }
 703
 704         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 705                 struct journal *j = &c->journal;
 706                 struct jset_entry *entry;
 707
 708                 trans_for_each_update(trans, i) {
 709                         if (i->key_cache_already_flushed)
 710                                 continue;
 711
 712                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 713                                 continue;
 714
 715                         verify_update_old_key(trans, i);
 716
 717                         if (trans->journal_transaction_names) {
 718                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 719                                                        BCH_JSET_ENTRY_overwrite,
 720                                                        i->btree_id, i->level,
 721                                                        i->old_k.u64s);
 722                                 bkey_reassemble((struct bkey_i *) entry->start,
 723                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 724                         }
 725
 726                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 727                                                BCH_JSET_ENTRY_btree_keys,
 728                                                i->btree_id, i->level,
 729                                                i->k->k.u64s);
 730                         bkey_copy((struct bkey_i *) entry->start, i->k);
 731                 }
 732
 733                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 734                                   trans->journal_entries,
 735                                   trans->journal_entries_u64s);
 736
 737                 trans->journal_res.offset       += trans->journal_entries_u64s;
 738                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 739
 740                 if (trans->journal_seq)
 741                         *trans->journal_seq = trans->journal_res.seq;
 742         }
 743
 744         trans_for_each_update(trans, i) {
 745                 struct btree_path *path = trans->paths + i->path;
 746
 747                 if (!i->cached) {
 748                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 749                 } else if (!i->key_cache_already_flushed)
 750                         bch2_btree_insert_key_cached(trans, flags, i);
 751                 else {
 752                         bch2_btree_key_cache_drop(trans, path);
 753                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 754                 }
 755         }
 756
 757         return 0;
 758 fatal_err:
 759         bch2_fatal_error(c);
 760 revert_fs_usage:
 761         if (trans->fs_usage_deltas)
 762                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 763         return ret;
 764 }
 765
 766 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 767 {
 768         trans_for_each_update(trans, i)
 769                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 770 }
 771
 772 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 773                                                    enum bkey_invalid_flags flags,
 774                                                    struct btree_insert_entry *i,
 775                                                    struct printbuf *err)
 776 {
 777         struct bch_fs *c = trans->c;
 778
 779         printbuf_reset(err);
 780         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 781                    trans->fn, (void *) i->ip_allocated);
 782         prt_newline(err);
 783         printbuf_indent_add(err, 2);
 784
 785         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 786         prt_newline(err);
 787
 788         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 789         bch2_print_string_as_lines(KERN_ERR, err->buf);
 790
 791         bch2_inconsistent_error(c);
 792         bch2_dump_trans_updates(trans);
 793
 794         return -EINVAL;
 795 }
 796
 797 static noinline int bch2_trans_commit_journal_entry_invalid(struct btree_trans *trans,
 798                                                    struct jset_entry *i)
 799 {
 800         struct bch_fs *c = trans->c;
 801         struct printbuf buf = PRINTBUF;
 802
 803         prt_printf(&buf, "invalid bkey on insert from %s", trans->fn);
 804         prt_newline(&buf);
 805         printbuf_indent_add(&buf, 2);
 806
 807         bch2_journal_entry_to_text(&buf, c, i);
 808         prt_newline(&buf);
 809
 810         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 811
 812         bch2_inconsistent_error(c);
 813         bch2_dump_trans_updates(trans);
 814
 815         return -EINVAL;
 816 }
 817
 818 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 819                                 struct journal_entry_pin *_pin, u64 seq)
 820 {
 821         return 0;
 822 }
 823
 824 /*
 825  * Get journal reservation, take write locks, and attempt to do btree update(s):
 826  */
 827 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 828                                        struct btree_insert_entry **stopped_at,
 829                                        unsigned long trace_ip)
 830 {
 831         struct bch_fs *c = trans->c;
 832         int ret = 0, u64s_delta = 0;
 833
 834         trans_for_each_update(trans, i) {
 835                 if (i->cached)
 836                         continue;
 837
 838                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 839                 u64s_delta -= i->old_btree_u64s;
 840
 841                 if (!same_leaf_as_next(trans, i)) {
 842                         if (u64s_delta <= 0) {
 843                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 844                                                         i->level, flags);
 845                                 if (unlikely(ret))
 846                                         return ret;
 847                         }
 848
 849                         u64s_delta = 0;
 850                 }
 851         }
 852
 853         ret = bch2_trans_lock_write(trans);
 854         if (unlikely(ret))
 855                 return ret;
 856
 857         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 858
 859         if (!ret && unlikely(trans->journal_replay_not_finished))
 860                 bch2_drop_overwrites_from_journal(trans);
 861
 862         bch2_trans_unlock_write(trans);
 863
 864         if (!ret && trans->journal_pin)
 865                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 866                                      trans->journal_pin,
 867                                      bch2_trans_commit_journal_pin_flush);
 868
 869         /*
 870          * Drop journal reservation after dropping write locks, since dropping
 871          * the journal reservation may kick off a journal write:
 872          */
 873         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 874                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 875
 876         return ret;
 877 }
 878
 879 static int journal_reclaim_wait_done(struct bch_fs *c)
 880 {
 881         int ret = bch2_journal_error(&c->journal) ?:
 882                 !bch2_btree_key_cache_must_wait(c);
 883
 884         if (!ret)
 885                 journal_reclaim_kick(&c->journal);
 886         return ret;
 887 }
 888
 889 static noinline
 890 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 891                             struct btree_insert_entry *i,
 892                             int ret, unsigned long trace_ip)
 893 {
 894         struct bch_fs *c = trans->c;
 895
 896         switch (ret) {
 897         case -BCH_ERR_btree_insert_btree_node_full:
 898                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 899                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 900                         trace_and_count(c, trans_restart_btree_node_split, trans,
 901                                         trace_ip, trans->paths + i->path);
 902                 break;
 903         case -BCH_ERR_btree_insert_need_mark_replicas:
 904                 ret = drop_locks_do(trans,
 905                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 906                 break;
 907         case -BCH_ERR_journal_res_get_blocked:
 908                 /*
 909                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 910                  * flag
 911                  */
 912                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 913                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 914                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 915                         break;
 916                 }
 917
 918                 ret = drop_locks_do(trans,
 919                         bch2_trans_journal_res_get(trans,
 920                                         (flags & BCH_WATERMARK_MASK)|
 921                                         JOURNAL_RES_GET_CHECK));
 922                 break;
 923         case -BCH_ERR_btree_insert_need_journal_reclaim:
 924                 bch2_trans_unlock(trans);
 925
 926                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 927
 928                 wait_event_freezable(c->journal.reclaim_wait,
 929                                      (ret = journal_reclaim_wait_done(c)));
 930                 if (ret < 0)
 931                         break;
 932
 933                 ret = bch2_trans_relock(trans);
 934                 break;
 935         default:
 936                 BUG_ON(ret >= 0);
 937                 break;
 938         }
 939
 940         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 941
 942         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 943                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 944                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 945
 946         return ret;
 947 }
 948
 949 static noinline int
 950 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 951 {
 952         struct bch_fs *c = trans->c;
 953         int ret;
 954
 955         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 956             test_bit(BCH_FS_started, &c->flags))
 957                 return -BCH_ERR_erofs_trans_commit;
 958
 959         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 960         if (ret)
 961                 return ret;
 962
 963         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 964         return 0;
 965 }
 966
 967 /*
 968  * This is for updates done in the early part of fsck - btree_gc - before we've
 969  * gone RW. we only add the new key to the list of keys for journal replay to
 970  * do.
 971  */
 972 static noinline int
 973 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 974 {
 975         struct bch_fs *c = trans->c;
 976         int ret = 0;
 977
 978         trans_for_each_update(trans, i) {
 979                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 980                 if (ret)
 981                         break;
 982         }
 983
 984         return ret;
 985 }
 986
 987 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 988 {
 989         struct btree_insert_entry *errored_at = NULL;
 990         struct bch_fs *c = trans->c;
 991         int ret = 0;
 992
 993         if (!trans->nr_updates &&
 994             !trans->journal_entries_u64s)
 995                 goto out_reset;
 996
 997         ret = bch2_trans_commit_run_triggers(trans);
 998         if (ret)
 999                 goto out_reset;
1000
1001         trans_for_each_update(trans, i) {
1002                 struct printbuf buf = PRINTBUF;
1003                 enum bkey_invalid_flags invalid_flags = 0;
1004
1005                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1006                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1007
1008                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1009                                                i->bkey_type, invalid_flags, &buf)))
1010                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1011                 btree_insert_entry_checks(trans, i);
1012                 printbuf_exit(&buf);
1013
1014                 if (ret)
1015                         return ret;
1016         }
1017
1018         for (struct jset_entry *i = trans->journal_entries;
1019              i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
1020              i = vstruct_next(i)) {
1021                 if (!jset_entry_is_key(i))
1022                         continue;
1023
1024                 enum bkey_invalid_flags invalid_flags = 0;
1025
1026                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1027                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1028
1029                 if (unlikely(bch2_journal_entry_validate(c, NULL, i,
1030                                         bcachefs_metadata_version_current,
1031                                         CPU_BIG_ENDIAN, invalid_flags)))
1032                         ret = bch2_trans_commit_journal_entry_invalid(trans, i);
1033
1034                 if (ret)
1035                         return ret;
1036         }
1037
1038         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1039                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1040                 goto out_reset;
1041         }
1042
1043         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1044             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1045                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1046                 if (ret)
1047                         goto out_reset;
1048         }
1049
1050         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1051
1052         trans->journal_u64s             = trans->journal_entries_u64s;
1053         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1054         if (trans->journal_transaction_names)
1055                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1056
1057         trans_for_each_update(trans, i) {
1058                 struct btree_path *path = trans->paths + i->path;
1059
1060                 EBUG_ON(!path->should_be_locked);
1061
1062                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1063                 if (unlikely(ret))
1064                         goto out;
1065
1066                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1067
1068                 if (i->key_cache_already_flushed)
1069                         continue;
1070
1071                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1072                         continue;
1073
1074                 /* we're going to journal the key being updated: */
1075                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1076
1077                 /* and we're also going to log the overwrite: */
1078                 if (trans->journal_transaction_names)
1079                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1080         }
1081
1082         if (trans->extra_disk_res) {
1083                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1084                                 trans->extra_disk_res,
1085                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1086                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1087                 if (ret)
1088                         goto err;
1089         }
1090 retry:
1091         errored_at = NULL;
1092         bch2_trans_verify_not_in_restart(trans);
1093         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1094                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1095
1096         ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1097
1098         /* make sure we didn't drop or screw up locks: */
1099         bch2_trans_verify_locks(trans);
1100
1101         if (ret)
1102                 goto err;
1103
1104         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1105 out:
1106         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1107                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1108 out_reset:
1109         if (!ret)
1110                 bch2_trans_downgrade(trans);
1111         bch2_trans_reset_updates(trans);
1112
1113         return ret;
1114 err:
1115         ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1116         if (ret)
1117                 goto out;
1118
1119         /*
1120          * We might have done another transaction commit in the error path -
1121          * i.e. btree write buffer flush - which will have made use of
1122          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1123          * how the journal sequence number to pin is passed in - so we must
1124          * restart:
1125          */
1126         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1127                 ret = -BCH_ERR_transaction_restart_nested;
1128                 goto out;
1129         }
1130
1131         goto retry;
1132 }