fs/bcachefs/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 > bch2_btree_keys_u64s_remaining(b));
 143         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 144
 145         k = bch2_btree_node_iter_peek_all(node_iter, b);
 146         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 147                 k = NULL;
 148
 149         /* @k is the key being overwritten/deleted, if any: */
 150         EBUG_ON(k && bkey_deleted(k));
 151
 152         /* Deleting, but not found? nothing to do: */
 153         if (bkey_deleted(&insert->k) && !k)
 154                 return false;
 155
 156         if (bkey_deleted(&insert->k)) {
 157                 /* Deleting: */
 158                 btree_account_key_drop(b, k);
 159                 k->type = KEY_TYPE_deleted;
 160
 161                 if (k->needs_whiteout)
 162                         push_whiteout(b, insert->k.p);
 163                 k->needs_whiteout = false;
 164
 165                 if (k >= btree_bset_last(b)->start) {
 166                         clobber_u64s = k->u64s;
 167                         bch2_bset_delete(b, k, clobber_u64s);
 168                         goto fix_iter;
 169                 } else {
 170                         bch2_btree_path_fix_key_modified(trans, b, k);
 171                 }
 172
 173                 return true;
 174         }
 175
 176         if (k) {
 177                 /* Overwriting: */
 178                 btree_account_key_drop(b, k);
 179                 k->type = KEY_TYPE_deleted;
 180
 181                 insert->k.needs_whiteout = k->needs_whiteout;
 182                 k->needs_whiteout = false;
 183
 184                 if (k >= btree_bset_last(b)->start) {
 185                         clobber_u64s = k->u64s;
 186                         goto overwrite;
 187                 } else {
 188                         bch2_btree_path_fix_key_modified(trans, b, k);
 189                 }
 190         }
 191
 192         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 193 overwrite:
 194         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 195         new_u64s = k->u64s;
 196 fix_iter:
 197         if (clobber_u64s != new_u64s)
 198                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 199                                          clobber_u64s, new_u64s);
 200         return true;
 201 }
 202
 203 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 204                                unsigned i, u64 seq)
 205 {
 206         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 207         struct btree_write *w = container_of(pin, struct btree_write, journal);
 208         struct btree *b = container_of(w, struct btree, writes[i]);
 209         struct btree_trans *trans = bch2_trans_get(c);
 210         unsigned long old, new, v;
 211         unsigned idx = w - b->writes;
 212
 213         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 214         v = READ_ONCE(b->flags);
 215
 216         do {
 217                 old = new = v;
 218
 219                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 220                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 221                     w->journal.seq != seq)
 222                         break;
 223
 224                 new &= ~BTREE_WRITE_TYPE_MASK;
 225                 new |= BTREE_WRITE_journal_reclaim;
 226                 new |= 1 << BTREE_NODE_need_write;
 227         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 228
 229         btree_node_write_if_need(c, b, SIX_LOCK_read);
 230         six_unlock_read(&b->c.lock);
 231
 232         bch2_trans_put(trans);
 233         return 0;
 234 }
 235
 236 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 237 {
 238         return __btree_node_flush(j, pin, 0, seq);
 239 }
 240
 241 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 242 {
 243         return __btree_node_flush(j, pin, 1, seq);
 244 }
 245
 246 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 247                                        struct btree *b, u64 seq)
 248 {
 249         struct btree_write *w = btree_current_write(b);
 250
 251         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 252                              btree_node_write_idx(b) == 0
 253                              ? bch2_btree_node_flush0
 254                              : bch2_btree_node_flush1);
 255 }
 256
 257 /**
 258  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 259  * @trans:              btree transaction object
 260  * @path:               path pointing to @insert's pos
 261  * @insert:             key to insert
 262  * @journal_seq:        sequence number of journal reservation
 263  */
 264 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 265                                        struct btree_path *path,
 266                                        struct bkey_i *insert,
 267                                        u64 journal_seq)
 268 {
 269         struct bch_fs *c = trans->c;
 270         struct btree *b = path_l(path)->b;
 271         struct bset_tree *t = bset_tree_last(b);
 272         struct bset *i = bset(b, t);
 273         int old_u64s = bset_u64s(t);
 274         int old_live_u64s = b->nr.live_u64s;
 275         int live_u64s_added, u64s_added;
 276
 277         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 278                                         &path_l(path)->iter, insert)))
 279                 return;
 280
 281         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 282
 283         bch2_btree_add_journal_pin(c, b, journal_seq);
 284
 285         if (unlikely(!btree_node_dirty(b))) {
 286                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 287                 set_btree_node_dirty_acct(c, b);
 288         }
 289
 290         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 291         u64s_added = (int) bset_u64s(t) - old_u64s;
 292
 293         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 294                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 295         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 296                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 297
 298         if (u64s_added > live_u64s_added &&
 299             bch2_maybe_compact_whiteouts(c, b))
 300                 bch2_trans_node_reinit_iter(trans, b);
 301 }
 302
 303 /* Cached btree updates: */
 304
 305 /* Normal update interface: */
 306
 307 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 308                                              struct btree_insert_entry *i)
 309 {
 310         struct btree_path *path = trans->paths + i->path;
 311
 312         BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 313         BUG_ON(i->cached        != path->cached);
 314         BUG_ON(i->level         != path->level);
 315         BUG_ON(i->btree_id      != path->btree_id);
 316         EBUG_ON(!i->level &&
 317                 btree_type_has_snapshots(i->btree_id) &&
 318                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 319                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 320                 i->k->k.p.snapshot &&
 321                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot) > 0);
 322 }
 323
 324 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 325                                                       unsigned flags)
 326 {
 327         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 328                                     trans->journal_u64s, flags);
 329 }
 330
 331 #define JSET_ENTRY_LOG_U64s             4
 332
 333 static noinline void journal_transaction_name(struct btree_trans *trans)
 334 {
 335         struct bch_fs *c = trans->c;
 336         struct journal *j = &c->journal;
 337         struct jset_entry *entry =
 338                 bch2_journal_add_entry(j, &trans->journal_res,
 339                                        BCH_JSET_ENTRY_log, 0, 0,
 340                                        JSET_ENTRY_LOG_U64s);
 341         struct jset_entry_log *l =
 342                 container_of(entry, struct jset_entry_log, entry);
 343
 344         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 345 }
 346
 347 static inline int btree_key_can_insert(struct btree_trans *trans,
 348                                        struct btree *b, unsigned u64s)
 349 {
 350         if (!bch2_btree_node_insert_fits(b, u64s))
 351                 return -BCH_ERR_btree_insert_btree_node_full;
 352
 353         return 0;
 354 }
 355
 356 noinline static int
 357 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 358                                      struct btree_path *path, unsigned new_u64s)
 359 {
 360         struct bkey_cached *ck = (void *) path->l[0].b;
 361         struct bkey_i *new_k;
 362         int ret;
 363
 364         bch2_trans_unlock_write(trans);
 365         bch2_trans_unlock(trans);
 366
 367         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 368         if (!new_k) {
 369                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 370                         bch2_btree_id_str(path->btree_id), new_u64s);
 371                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 372         }
 373
 374         ret =   bch2_trans_relock(trans) ?:
 375                 bch2_trans_lock_write(trans);
 376         if (unlikely(ret)) {
 377                 kfree(new_k);
 378                 return ret;
 379         }
 380
 381         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 382
 383         trans_for_each_update(trans, i)
 384                 if (i->old_v == &ck->k->v)
 385                         i->old_v = &new_k->v;
 386
 387         kfree(ck->k);
 388         ck->u64s        = new_u64s;
 389         ck->k           = new_k;
 390         return 0;
 391 }
 392
 393 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 394                                        struct btree_path *path, unsigned u64s)
 395 {
 396         struct bch_fs *c = trans->c;
 397         struct bkey_cached *ck = (void *) path->l[0].b;
 398         unsigned new_u64s;
 399         struct bkey_i *new_k;
 400
 401         EBUG_ON(path->level);
 402
 403         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 404             bch2_btree_key_cache_must_wait(c) &&
 405             !(flags & BCH_TRANS_COMMIT_journal_reclaim))
 406                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 407
 408         /*
 409          * bch2_varint_decode can read past the end of the buffer by at most 7
 410          * bytes (it won't be used):
 411          */
 412         u64s += 1;
 413
 414         if (u64s <= ck->u64s)
 415                 return 0;
 416
 417         new_u64s        = roundup_pow_of_two(u64s);
 418         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 419         if (unlikely(!new_k))
 420                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 421
 422         trans_for_each_update(trans, i)
 423                 if (i->old_v == &ck->k->v)
 424                         i->old_v = &new_k->v;
 425
 426         ck->u64s        = new_u64s;
 427         ck->k           = new_k;
 428         return 0;
 429 }
 430
 431 /* Triggers: */
 432
 433 static int run_one_mem_trigger(struct btree_trans *trans,
 434                                struct btree_insert_entry *i,
 435                                unsigned flags)
 436 {
 437         struct bkey_s_c old = { &i->old_k, i->old_v };
 438         struct bkey_i *new = i->k;
 439         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 440         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 441         int ret;
 442
 443         verify_update_old_key(trans, i);
 444
 445         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 446                 return 0;
 447
 448         if (old_ops->trigger == new_ops->trigger) {
 449                 ret   = bch2_key_trigger(trans, i->btree_id, i->level,
 450                                 old, bkey_i_to_s(new),
 451                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 452         } else {
 453                 ret   = bch2_key_trigger_new(trans, i->btree_id, i->level,
 454                                 bkey_i_to_s(new), flags) ?:
 455                         bch2_key_trigger_old(trans, i->btree_id, i->level,
 456                                 old, flags);
 457         }
 458
 459         return ret;
 460 }
 461
 462 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 463                                  bool overwrite)
 464 {
 465         /*
 466          * Transactional triggers create new btree_insert_entries, so we can't
 467          * pass them a pointer to a btree_insert_entry, that memory is going to
 468          * move:
 469          */
 470         struct bkey old_k = i->old_k;
 471         struct bkey_s_c old = { &old_k, i->old_v };
 472         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 473         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 474         unsigned flags = i->flags|BTREE_TRIGGER_TRANSACTIONAL;
 475
 476         verify_update_old_key(trans, i);
 477
 478         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 479             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 480                 return 0;
 481
 482         if (!i->insert_trigger_run &&
 483             !i->overwrite_trigger_run &&
 484             old_ops->trigger == new_ops->trigger) {
 485                 i->overwrite_trigger_run = true;
 486                 i->insert_trigger_run = true;
 487                 return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
 488                                         BTREE_TRIGGER_INSERT|
 489                                         BTREE_TRIGGER_OVERWRITE|flags) ?: 1;
 490         } else if (overwrite && !i->overwrite_trigger_run) {
 491                 i->overwrite_trigger_run = true;
 492                 return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
 493         } else if (!overwrite && !i->insert_trigger_run) {
 494                 i->insert_trigger_run = true;
 495                 return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 496         } else {
 497                 return 0;
 498         }
 499 }
 500
 501 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 502                               struct btree_insert_entry *btree_id_start)
 503 {
 504         struct btree_insert_entry *i;
 505         bool trans_trigger_run;
 506         int ret, overwrite;
 507
 508         for (overwrite = 1; overwrite >= 0; --overwrite) {
 509
 510                 /*
 511                  * Running triggers will append more updates to the list of updates as
 512                  * we're walking it:
 513                  */
 514                 do {
 515                         trans_trigger_run = false;
 516
 517                         for (i = btree_id_start;
 518                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 519                              i++) {
 520                                 if (i->btree_id != btree_id)
 521                                         continue;
 522
 523                                 ret = run_one_trans_trigger(trans, i, overwrite);
 524                                 if (ret < 0)
 525                                         return ret;
 526                                 if (ret)
 527                                         trans_trigger_run = true;
 528                         }
 529                 } while (trans_trigger_run);
 530         }
 531
 532         return 0;
 533 }
 534
 535 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 536 {
 537         struct btree_insert_entry *btree_id_start = trans->updates;
 538         unsigned btree_id = 0;
 539         int ret = 0;
 540
 541         /*
 542          *
 543          * For a given btree, this algorithm runs insert triggers before
 544          * overwrite triggers: this is so that when extents are being moved
 545          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 546          * they are re-added.
 547          */
 548         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 549                 if (btree_id == BTREE_ID_alloc)
 550                         continue;
 551
 552                 while (btree_id_start < trans->updates + trans->nr_updates &&
 553                        btree_id_start->btree_id < btree_id)
 554                         btree_id_start++;
 555
 556                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 557                 if (ret)
 558                         return ret;
 559         }
 560
 561         trans_for_each_update(trans, i) {
 562                 if (i->btree_id > BTREE_ID_alloc)
 563                         break;
 564                 if (i->btree_id == BTREE_ID_alloc) {
 565                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 566                         if (ret)
 567                                 return ret;
 568                         break;
 569                 }
 570         }
 571
 572 #ifdef CONFIG_BCACHEFS_DEBUG
 573         trans_for_each_update(trans, i)
 574                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 575                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 576                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 577 #endif
 578         return 0;
 579 }
 580
 581 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 582 {
 583         trans_for_each_update(trans, i) {
 584                 /*
 585                  * XXX: synchronization of cached update triggers with gc
 586                  * XXX: synchronization of interior node updates with gc
 587                  */
 588                 BUG_ON(i->cached || i->level);
 589
 590                 if (btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)) &&
 591                     gc_visited(trans->c, gc_pos_btree_node(insert_l(trans, i)->b))) {
 592                         int ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 593                         if (ret)
 594                                 return ret;
 595                 }
 596         }
 597
 598         return 0;
 599 }
 600
 601 static inline int
 602 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 603                                struct btree_insert_entry **stopped_at,
 604                                unsigned long trace_ip)
 605 {
 606         struct bch_fs *c = trans->c;
 607         struct btree_trans_commit_hook *h;
 608         unsigned u64s = 0;
 609         int ret;
 610
 611         if (race_fault()) {
 612                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 613                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 614         }
 615
 616         /*
 617          * Check if the insert will fit in the leaf node with the write lock
 618          * held, otherwise another thread could write the node changing the
 619          * amount of space available:
 620          */
 621
 622         prefetch(&trans->c->journal.flags);
 623
 624         trans_for_each_update(trans, i) {
 625                 /* Multiple inserts might go to same leaf: */
 626                 if (!same_leaf_as_prev(trans, i))
 627                         u64s = 0;
 628
 629                 u64s += i->k->k.u64s;
 630                 ret = !i->cached
 631                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 632                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 633                 if (ret) {
 634                         *stopped_at = i;
 635                         return ret;
 636                 }
 637
 638                 i->k->k.needs_whiteout = false;
 639         }
 640
 641         /*
 642          * Don't get journal reservation until after we know insert will
 643          * succeed:
 644          */
 645         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 646                 ret = bch2_trans_journal_res_get(trans,
 647                                 (flags & BCH_WATERMARK_MASK)|
 648                                 JOURNAL_RES_GET_NONBLOCK);
 649                 if (ret)
 650                         return ret;
 651
 652                 if (unlikely(trans->journal_transaction_names))
 653                         journal_transaction_name(trans);
 654         }
 655
 656         /*
 657          * Not allowed to fail after we've gotten our journal reservation - we
 658          * have to use it:
 659          */
 660
 661         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 662             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 663                 if (bch2_journal_seq_verify)
 664                         trans_for_each_update(trans, i)
 665                                 i->k->k.version.lo = trans->journal_res.seq;
 666                 else if (bch2_inject_invalid_keys)
 667                         trans_for_each_update(trans, i)
 668                                 i->k->k.version = MAX_VERSION;
 669         }
 670
 671         if (trans->fs_usage_deltas &&
 672             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 673                 return -BCH_ERR_btree_insert_need_mark_replicas;
 674
 675         /* XXX: we only want to run this if deltas are nonzero */
 676         bch2_trans_account_disk_usage_change(trans);
 677
 678         h = trans->hooks;
 679         while (h) {
 680                 ret = h->fn(trans, h);
 681                 if (ret)
 682                         goto revert_fs_usage;
 683                 h = h->next;
 684         }
 685
 686         trans_for_each_update(trans, i)
 687                 if (BTREE_NODE_TYPE_HAS_ATOMIC_TRIGGERS & (1U << i->bkey_type)) {
 688                         ret = run_one_mem_trigger(trans, i, BTREE_TRIGGER_ATOMIC|i->flags);
 689                         if (ret)
 690                                 goto fatal_err;
 691                 }
 692
 693         if (unlikely(c->gc_pos.phase)) {
 694                 ret = bch2_trans_commit_run_gc_triggers(trans);
 695                 if  (ret)
 696                         goto fatal_err;
 697         }
 698
 699         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 700                 struct journal *j = &c->journal;
 701                 struct jset_entry *entry;
 702
 703                 trans_for_each_update(trans, i) {
 704                         if (i->key_cache_already_flushed)
 705                                 continue;
 706
 707                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 708                                 continue;
 709
 710                         verify_update_old_key(trans, i);
 711
 712                         if (trans->journal_transaction_names) {
 713                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 714                                                        BCH_JSET_ENTRY_overwrite,
 715                                                        i->btree_id, i->level,
 716                                                        i->old_k.u64s);
 717                                 bkey_reassemble((struct bkey_i *) entry->start,
 718                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 719                         }
 720
 721                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 722                                                BCH_JSET_ENTRY_btree_keys,
 723                                                i->btree_id, i->level,
 724                                                i->k->k.u64s);
 725                         bkey_copy((struct bkey_i *) entry->start, i->k);
 726                 }
 727
 728                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 729                                   trans->journal_entries,
 730                                   trans->journal_entries_u64s);
 731
 732                 trans->journal_res.offset       += trans->journal_entries_u64s;
 733                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 734
 735                 if (trans->journal_seq)
 736                         *trans->journal_seq = trans->journal_res.seq;
 737         }
 738
 739         trans_for_each_update(trans, i) {
 740                 struct btree_path *path = trans->paths + i->path;
 741
 742                 if (!i->cached) {
 743                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 744                 } else if (!i->key_cache_already_flushed)
 745                         bch2_btree_insert_key_cached(trans, flags, i);
 746                 else {
 747                         bch2_btree_key_cache_drop(trans, path);
 748                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 749                 }
 750         }
 751
 752         return 0;
 753 fatal_err:
 754         bch2_fatal_error(c);
 755 revert_fs_usage:
 756         if (trans->fs_usage_deltas)
 757                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 758         return ret;
 759 }
 760
 761 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 762 {
 763         trans_for_each_update(trans, i)
 764                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 765 }
 766
 767 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 768                                                    enum bkey_invalid_flags flags,
 769                                                    struct btree_insert_entry *i,
 770                                                    struct printbuf *err)
 771 {
 772         struct bch_fs *c = trans->c;
 773
 774         printbuf_reset(err);
 775         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 776                    trans->fn, (void *) i->ip_allocated);
 777         prt_newline(err);
 778         printbuf_indent_add(err, 2);
 779
 780         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 781         prt_newline(err);
 782
 783         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 784         bch2_print_string_as_lines(KERN_ERR, err->buf);
 785
 786         bch2_inconsistent_error(c);
 787         bch2_dump_trans_updates(trans);
 788
 789         return -EINVAL;
 790 }
 791
 792 static noinline int bch2_trans_commit_journal_entry_invalid(struct btree_trans *trans,
 793                                                    struct jset_entry *i)
 794 {
 795         struct bch_fs *c = trans->c;
 796         struct printbuf buf = PRINTBUF;
 797
 798         prt_printf(&buf, "invalid bkey on insert from %s", trans->fn);
 799         prt_newline(&buf);
 800         printbuf_indent_add(&buf, 2);
 801
 802         bch2_journal_entry_to_text(&buf, c, i);
 803         prt_newline(&buf);
 804
 805         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 806
 807         bch2_inconsistent_error(c);
 808         bch2_dump_trans_updates(trans);
 809
 810         return -EINVAL;
 811 }
 812
 813 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 814                                 struct journal_entry_pin *_pin, u64 seq)
 815 {
 816         return 0;
 817 }
 818
 819 /*
 820  * Get journal reservation, take write locks, and attempt to do btree update(s):
 821  */
 822 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 823                                        struct btree_insert_entry **stopped_at,
 824                                        unsigned long trace_ip)
 825 {
 826         struct bch_fs *c = trans->c;
 827         int ret = 0, u64s_delta = 0;
 828
 829         trans_for_each_update(trans, i) {
 830                 if (i->cached)
 831                         continue;
 832
 833                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 834                 u64s_delta -= i->old_btree_u64s;
 835
 836                 if (!same_leaf_as_next(trans, i)) {
 837                         if (u64s_delta <= 0) {
 838                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 839                                                         i->level, flags);
 840                                 if (unlikely(ret))
 841                                         return ret;
 842                         }
 843
 844                         u64s_delta = 0;
 845                 }
 846         }
 847
 848         ret = bch2_trans_lock_write(trans);
 849         if (unlikely(ret))
 850                 return ret;
 851
 852         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 853
 854         if (!ret && unlikely(trans->journal_replay_not_finished))
 855                 bch2_drop_overwrites_from_journal(trans);
 856
 857         bch2_trans_unlock_write(trans);
 858
 859         if (!ret && trans->journal_pin)
 860                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 861                                      trans->journal_pin,
 862                                      bch2_trans_commit_journal_pin_flush);
 863
 864         /*
 865          * Drop journal reservation after dropping write locks, since dropping
 866          * the journal reservation may kick off a journal write:
 867          */
 868         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 869                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 870
 871         return ret;
 872 }
 873
 874 static int journal_reclaim_wait_done(struct bch_fs *c)
 875 {
 876         int ret = bch2_journal_error(&c->journal) ?:
 877                 !bch2_btree_key_cache_must_wait(c);
 878
 879         if (!ret)
 880                 journal_reclaim_kick(&c->journal);
 881         return ret;
 882 }
 883
 884 static noinline
 885 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 886                             struct btree_insert_entry *i,
 887                             int ret, unsigned long trace_ip)
 888 {
 889         struct bch_fs *c = trans->c;
 890         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 891
 892         switch (ret) {
 893         case -BCH_ERR_btree_insert_btree_node_full:
 894                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 895                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 896                         trace_and_count(c, trans_restart_btree_node_split, trans,
 897                                         trace_ip, trans->paths + i->path);
 898                 break;
 899         case -BCH_ERR_btree_insert_need_mark_replicas:
 900                 ret = drop_locks_do(trans,
 901                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 902                 break;
 903         case -BCH_ERR_journal_res_get_blocked:
 904                 /*
 905                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 906                  * flag
 907                  */
 908                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 909                     watermark < BCH_WATERMARK_reclaim) {
 910                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 911                         break;
 912                 }
 913
 914                 ret = drop_locks_do(trans,
 915                         bch2_trans_journal_res_get(trans,
 916                                         (flags & BCH_WATERMARK_MASK)|
 917                                         JOURNAL_RES_GET_CHECK));
 918                 break;
 919         case -BCH_ERR_btree_insert_need_journal_reclaim:
 920                 bch2_trans_unlock(trans);
 921
 922                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 923
 924                 wait_event_freezable(c->journal.reclaim_wait,
 925                                      (ret = journal_reclaim_wait_done(c)));
 926                 if (ret < 0)
 927                         break;
 928
 929                 ret = bch2_trans_relock(trans);
 930                 break;
 931         default:
 932                 BUG_ON(ret >= 0);
 933                 break;
 934         }
 935
 936         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 937
 938         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 939                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 940                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 941
 942         return ret;
 943 }
 944
 945 static noinline int
 946 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 947 {
 948         struct bch_fs *c = trans->c;
 949         int ret;
 950
 951         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 952             test_bit(BCH_FS_started, &c->flags))
 953                 return -BCH_ERR_erofs_trans_commit;
 954
 955         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 956         if (ret)
 957                 return ret;
 958
 959         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 960         return 0;
 961 }
 962
 963 /*
 964  * This is for updates done in the early part of fsck - btree_gc - before we've
 965  * gone RW. we only add the new key to the list of keys for journal replay to
 966  * do.
 967  */
 968 static noinline int
 969 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 970 {
 971         struct bch_fs *c = trans->c;
 972         int ret = 0;
 973
 974         trans_for_each_update(trans, i) {
 975                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 976                 if (ret)
 977                         break;
 978         }
 979
 980         return ret;
 981 }
 982
 983 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 984 {
 985         struct btree_insert_entry *errored_at = NULL;
 986         struct bch_fs *c = trans->c;
 987         int ret = 0;
 988
 989         if (!trans->nr_updates &&
 990             !trans->journal_entries_u64s)
 991                 goto out_reset;
 992
 993         memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));
 994
 995         ret = bch2_trans_commit_run_triggers(trans);
 996         if (ret)
 997                 goto out_reset;
 998
 999         trans_for_each_update(trans, i) {
1000                 struct printbuf buf = PRINTBUF;
1001                 enum bkey_invalid_flags invalid_flags = 0;
1002
1003                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1004                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1005
1006                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1007                                                i->bkey_type, invalid_flags, &buf)))
1008                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1009                 btree_insert_entry_checks(trans, i);
1010                 printbuf_exit(&buf);
1011
1012                 if (ret)
1013                         return ret;
1014         }
1015
1016         for (struct jset_entry *i = trans->journal_entries;
1017              i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
1018              i = vstruct_next(i)) {
1019                 enum bkey_invalid_flags invalid_flags = 0;
1020
1021                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1022                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1023
1024                 if (unlikely(bch2_journal_entry_validate(c, NULL, i,
1025                                         bcachefs_metadata_version_current,
1026                                         CPU_BIG_ENDIAN, invalid_flags)))
1027                         ret = bch2_trans_commit_journal_entry_invalid(trans, i);
1028
1029                 if (ret)
1030                         return ret;
1031         }
1032
1033         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1034                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1035                 goto out_reset;
1036         }
1037
1038         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1039             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1040                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1041                 if (ret)
1042                         goto out_reset;
1043         }
1044
1045         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1046
1047         trans->journal_u64s             = trans->journal_entries_u64s;
1048         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1049         if (trans->journal_transaction_names)
1050                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1051
1052         trans_for_each_update(trans, i) {
1053                 struct btree_path *path = trans->paths + i->path;
1054
1055                 EBUG_ON(!path->should_be_locked);
1056
1057                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1058                 if (unlikely(ret))
1059                         goto out;
1060
1061                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1062
1063                 if (i->key_cache_already_flushed)
1064                         continue;
1065
1066                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1067                         continue;
1068
1069                 /* we're going to journal the key being updated: */
1070                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1071
1072                 /* and we're also going to log the overwrite: */
1073                 if (trans->journal_transaction_names)
1074                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1075         }
1076
1077         if (trans->extra_disk_res) {
1078                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1079                                 trans->extra_disk_res,
1080                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1081                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1082                 if (ret)
1083                         goto err;
1084         }
1085 retry:
1086         errored_at = NULL;
1087         bch2_trans_verify_not_in_restart(trans);
1088         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1089                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1090
1091         ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1092
1093         /* make sure we didn't drop or screw up locks: */
1094         bch2_trans_verify_locks(trans);
1095
1096         if (ret)
1097                 goto err;
1098
1099         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1100 out:
1101         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1102                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1103 out_reset:
1104         if (!ret)
1105                 bch2_trans_downgrade(trans);
1106         bch2_trans_reset_updates(trans);
1107
1108         return ret;
1109 err:
1110         ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1111         if (ret)
1112                 goto out;
1113
1114         /*
1115          * We might have done another transaction commit in the error path -
1116          * i.e. btree write buffer flush - which will have made use of
1117          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1118          * how the journal sequence number to pin is passed in - so we must
1119          * restart:
1120          */
1121         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1122                 ret = -BCH_ERR_transaction_restart_nested;
1123                 goto out;
1124         }
1125
1126         goto retry;
1127 }