fs/bcachefs/btree_trans_commit.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "alloc_foreground.h"
   5 #include "btree_gc.h"
   6 #include "btree_io.h"
   7 #include "btree_iter.h"
   8 #include "btree_journal_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_update_interior.h"
  11 #include "btree_write_buffer.h"
  12 #include "buckets.h"
  13 #include "errcode.h"
  14 #include "error.h"
  15 #include "journal.h"
  16 #include "journal_io.h"
  17 #include "journal_reclaim.h"
  18 #include "replicas.h"
  19 #include "snapshot.h"
  20
  21 #include <linux/prefetch.h>
  22
  23 static const char * const trans_commit_flags_strs[] = {
  24 #define x(n, ...) #n,
  25         BCH_TRANS_COMMIT_FLAGS()
  26 #undef x
  27         NULL
  28 };
  29
  30 void bch2_trans_commit_flags_to_text(struct printbuf *out, enum bch_trans_commit_flags flags)
  31 {
  32         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
  33
  34         prt_printf(out, "watermark=%s", bch2_watermarks[watermark]);
  35
  36         flags >>= BCH_WATERMARK_BITS;
  37         if (flags) {
  38                 prt_char(out, ' ');
  39                 bch2_prt_bitflags(out, trans_commit_flags_strs, flags);
  40         }
  41 }
  42
  43 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  44 {
  45 #ifdef CONFIG_BCACHEFS_DEBUG
  46         struct bch_fs *c = trans->c;
  47         struct bkey u;
  48         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(trans->paths + i->path, &u);
  49
  50         if (unlikely(trans->journal_replay_not_finished)) {
  51                 struct bkey_i *j_k =
  52                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  53
  54                 if (j_k)
  55                         k = bkey_i_to_s_c(j_k);
  56         }
  57
  58         u = *k.k;
  59         u.needs_whiteout = i->old_k.needs_whiteout;
  60
  61         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  62         BUG_ON(i->old_v != k.v);
  63 #endif
  64 }
  65
  66 static inline struct btree_path_level *insert_l(struct btree_trans *trans, struct btree_insert_entry *i)
  67 {
  68         return (trans->paths + i->path)->l + i->level;
  69 }
  70
  71 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  72                                      struct btree_insert_entry *i)
  73 {
  74         return i != trans->updates &&
  75                 insert_l(trans, &i[0])->b == insert_l(trans, &i[-1])->b;
  76 }
  77
  78 static inline bool same_leaf_as_next(struct btree_trans *trans,
  79                                      struct btree_insert_entry *i)
  80 {
  81         return i + 1 < trans->updates + trans->nr_updates &&
  82                 insert_l(trans, &i[0])->b == insert_l(trans, &i[1])->b;
  83 }
  84
  85 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  86                                            struct btree_path *path,
  87                                            struct btree *b)
  88 {
  89         struct bch_fs *c = trans->c;
  90
  91         if (unlikely(btree_node_just_written(b)) &&
  92             bch2_btree_post_write_cleanup(c, b))
  93                 bch2_trans_node_reinit_iter(trans, b);
  94
  95         /*
  96          * If the last bset has been written, or if it's gotten too big - start
  97          * a new bset to insert into:
  98          */
  99         if (want_new_bset(c, b))
 100                 bch2_btree_init_next(trans, b);
 101 }
 102
 103 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 104 {
 105         while (--i >= trans->updates) {
 106                 if (same_leaf_as_prev(trans, i))
 107                         continue;
 108
 109                 bch2_btree_node_unlock_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 110         }
 111
 112         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 113         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 114 }
 115
 116 static inline int bch2_trans_lock_write(struct btree_trans *trans)
 117 {
 118         EBUG_ON(trans->write_locked);
 119
 120         trans_for_each_update(trans, i) {
 121                 if (same_leaf_as_prev(trans, i))
 122                         continue;
 123
 124                 if (bch2_btree_node_lock_write(trans, trans->paths + i->path, &insert_l(trans, i)->b->c))
 125                         return trans_lock_write_fail(trans, i);
 126
 127                 if (!i->cached)
 128                         bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 129         }
 130
 131         trans->write_locked = true;
 132         return 0;
 133 }
 134
 135 static inline void bch2_trans_unlock_write(struct btree_trans *trans)
 136 {
 137         if (likely(trans->write_locked)) {
 138                 trans_for_each_update(trans, i)
 139                         if (!same_leaf_as_prev(trans, i))
 140                                 bch2_btree_node_unlock_write_inlined(trans,
 141                                                 trans->paths + i->path, insert_l(trans, i)->b);
 142                 trans->write_locked = false;
 143         }
 144 }
 145
 146 /* Inserting into a given leaf node (last stage of insert): */
 147
 148 /* Handle overwrites and do insert, for non extents: */
 149 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 150                                 struct btree_path *path,
 151                                 struct btree *b,
 152                                 struct btree_node_iter *node_iter,
 153                                 struct bkey_i *insert)
 154 {
 155         struct bkey_packed *k;
 156         unsigned clobber_u64s = 0, new_u64s = 0;
 157
 158         EBUG_ON(btree_node_just_written(b));
 159         EBUG_ON(bset_written(b, btree_bset_last(b)));
 160         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 161         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 162         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 163         EBUG_ON(insert->k.u64s > bch2_btree_keys_u64s_remaining(b));
 164         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 165
 166         k = bch2_btree_node_iter_peek_all(node_iter, b);
 167         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 168                 k = NULL;
 169
 170         /* @k is the key being overwritten/deleted, if any: */
 171         EBUG_ON(k && bkey_deleted(k));
 172
 173         /* Deleting, but not found? nothing to do: */
 174         if (bkey_deleted(&insert->k) && !k)
 175                 return false;
 176
 177         if (bkey_deleted(&insert->k)) {
 178                 /* Deleting: */
 179                 btree_account_key_drop(b, k);
 180                 k->type = KEY_TYPE_deleted;
 181
 182                 if (k->needs_whiteout)
 183                         push_whiteout(b, insert->k.p);
 184                 k->needs_whiteout = false;
 185
 186                 if (k >= btree_bset_last(b)->start) {
 187                         clobber_u64s = k->u64s;
 188                         bch2_bset_delete(b, k, clobber_u64s);
 189                         goto fix_iter;
 190                 } else {
 191                         bch2_btree_path_fix_key_modified(trans, b, k);
 192                 }
 193
 194                 return true;
 195         }
 196
 197         if (k) {
 198                 /* Overwriting: */
 199                 btree_account_key_drop(b, k);
 200                 k->type = KEY_TYPE_deleted;
 201
 202                 insert->k.needs_whiteout = k->needs_whiteout;
 203                 k->needs_whiteout = false;
 204
 205                 if (k >= btree_bset_last(b)->start) {
 206                         clobber_u64s = k->u64s;
 207                         goto overwrite;
 208                 } else {
 209                         bch2_btree_path_fix_key_modified(trans, b, k);
 210                 }
 211         }
 212
 213         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 214 overwrite:
 215         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 216         new_u64s = k->u64s;
 217 fix_iter:
 218         if (clobber_u64s != new_u64s)
 219                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 220                                          clobber_u64s, new_u64s);
 221         return true;
 222 }
 223
 224 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 225                                unsigned i, u64 seq)
 226 {
 227         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 228         struct btree_write *w = container_of(pin, struct btree_write, journal);
 229         struct btree *b = container_of(w, struct btree, writes[i]);
 230         struct btree_trans *trans = bch2_trans_get(c);
 231         unsigned long old, new, v;
 232         unsigned idx = w - b->writes;
 233
 234         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 235         v = READ_ONCE(b->flags);
 236
 237         do {
 238                 old = new = v;
 239
 240                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 241                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 242                     w->journal.seq != seq)
 243                         break;
 244
 245                 new &= ~BTREE_WRITE_TYPE_MASK;
 246                 new |= BTREE_WRITE_journal_reclaim;
 247                 new |= 1 << BTREE_NODE_need_write;
 248         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 249
 250         btree_node_write_if_need(c, b, SIX_LOCK_read);
 251         six_unlock_read(&b->c.lock);
 252
 253         bch2_trans_put(trans);
 254         return 0;
 255 }
 256
 257 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 258 {
 259         return __btree_node_flush(j, pin, 0, seq);
 260 }
 261
 262 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 263 {
 264         return __btree_node_flush(j, pin, 1, seq);
 265 }
 266
 267 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 268                                        struct btree *b, u64 seq)
 269 {
 270         struct btree_write *w = btree_current_write(b);
 271
 272         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 273                              btree_node_write_idx(b) == 0
 274                              ? bch2_btree_node_flush0
 275                              : bch2_btree_node_flush1);
 276 }
 277
 278 /**
 279  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 280  * @trans:              btree transaction object
 281  * @path:               path pointing to @insert's pos
 282  * @insert:             key to insert
 283  * @journal_seq:        sequence number of journal reservation
 284  */
 285 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 286                                        struct btree_path *path,
 287                                        struct bkey_i *insert,
 288                                        u64 journal_seq)
 289 {
 290         struct bch_fs *c = trans->c;
 291         struct btree *b = path_l(path)->b;
 292         struct bset_tree *t = bset_tree_last(b);
 293         struct bset *i = bset(b, t);
 294         int old_u64s = bset_u64s(t);
 295         int old_live_u64s = b->nr.live_u64s;
 296         int live_u64s_added, u64s_added;
 297
 298         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 299                                         &path_l(path)->iter, insert)))
 300                 return;
 301
 302         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 303
 304         bch2_btree_add_journal_pin(c, b, journal_seq);
 305
 306         if (unlikely(!btree_node_dirty(b))) {
 307                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 308                 set_btree_node_dirty_acct(c, b);
 309         }
 310
 311         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 312         u64s_added = (int) bset_u64s(t) - old_u64s;
 313
 314         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 315                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 316         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 317                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 318
 319         if (u64s_added > live_u64s_added &&
 320             bch2_maybe_compact_whiteouts(c, b))
 321                 bch2_trans_node_reinit_iter(trans, b);
 322 }
 323
 324 /* Cached btree updates: */
 325
 326 /* Normal update interface: */
 327
 328 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 329                                              struct btree_insert_entry *i)
 330 {
 331         struct btree_path *path = trans->paths + i->path;
 332
 333         BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 334         BUG_ON(i->cached        != path->cached);
 335         BUG_ON(i->level         != path->level);
 336         BUG_ON(i->btree_id      != path->btree_id);
 337         EBUG_ON(!i->level &&
 338                 btree_type_has_snapshots(i->btree_id) &&
 339                 !(i->flags & BTREE_UPDATE_internal_snapshot_node) &&
 340                 test_bit(JOURNAL_replay_done, &trans->c->journal.flags) &&
 341                 i->k->k.p.snapshot &&
 342                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot) > 0);
 343 }
 344
 345 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 346                                                       unsigned flags)
 347 {
 348         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 349                                     trans->journal_u64s, flags);
 350 }
 351
 352 #define JSET_ENTRY_LOG_U64s             4
 353
 354 static noinline void journal_transaction_name(struct btree_trans *trans)
 355 {
 356         struct bch_fs *c = trans->c;
 357         struct journal *j = &c->journal;
 358         struct jset_entry *entry =
 359                 bch2_journal_add_entry(j, &trans->journal_res,
 360                                        BCH_JSET_ENTRY_log, 0, 0,
 361                                        JSET_ENTRY_LOG_U64s);
 362         struct jset_entry_log *l =
 363                 container_of(entry, struct jset_entry_log, entry);
 364
 365         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 366 }
 367
 368 static inline int btree_key_can_insert(struct btree_trans *trans,
 369                                        struct btree *b, unsigned u64s)
 370 {
 371         if (!bch2_btree_node_insert_fits(b, u64s))
 372                 return -BCH_ERR_btree_insert_btree_node_full;
 373
 374         return 0;
 375 }
 376
 377 noinline static int
 378 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 379                                      struct btree_path *path, unsigned new_u64s)
 380 {
 381         struct bkey_cached *ck = (void *) path->l[0].b;
 382         struct bkey_i *new_k;
 383         int ret;
 384
 385         bch2_trans_unlock_write(trans);
 386         bch2_trans_unlock(trans);
 387
 388         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 389         if (!new_k) {
 390                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 391                         bch2_btree_id_str(path->btree_id), new_u64s);
 392                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 393         }
 394
 395         ret =   bch2_trans_relock(trans) ?:
 396                 bch2_trans_lock_write(trans);
 397         if (unlikely(ret)) {
 398                 kfree(new_k);
 399                 return ret;
 400         }
 401
 402         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 403
 404         trans_for_each_update(trans, i)
 405                 if (i->old_v == &ck->k->v)
 406                         i->old_v = &new_k->v;
 407
 408         kfree(ck->k);
 409         ck->u64s        = new_u64s;
 410         ck->k           = new_k;
 411         return 0;
 412 }
 413
 414 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 415                                        struct btree_path *path, unsigned u64s)
 416 {
 417         struct bch_fs *c = trans->c;
 418         struct bkey_cached *ck = (void *) path->l[0].b;
 419         unsigned new_u64s;
 420         struct bkey_i *new_k;
 421         unsigned watermark = flags & BCH_WATERMARK_MASK;
 422
 423         EBUG_ON(path->level);
 424
 425         if (watermark < BCH_WATERMARK_reclaim &&
 426             !test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 427             bch2_btree_key_cache_must_wait(c))
 428                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 429
 430         /*
 431          * bch2_varint_decode can read past the end of the buffer by at most 7
 432          * bytes (it won't be used):
 433          */
 434         u64s += 1;
 435
 436         if (u64s <= ck->u64s)
 437                 return 0;
 438
 439         new_u64s        = roundup_pow_of_two(u64s);
 440         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 441         if (unlikely(!new_k))
 442                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 443
 444         trans_for_each_update(trans, i)
 445                 if (i->old_v == &ck->k->v)
 446                         i->old_v = &new_k->v;
 447
 448         ck->u64s        = new_u64s;
 449         ck->k           = new_k;
 450         return 0;
 451 }
 452
 453 /* Triggers: */
 454
 455 static int run_one_mem_trigger(struct btree_trans *trans,
 456                                struct btree_insert_entry *i,
 457                                unsigned flags)
 458 {
 459         struct bkey_s_c old = { &i->old_k, i->old_v };
 460         struct bkey_i *new = i->k;
 461         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 462         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 463         int ret;
 464
 465         verify_update_old_key(trans, i);
 466
 467         if (unlikely(flags & BTREE_TRIGGER_norun))
 468                 return 0;
 469
 470         if (old_ops->trigger == new_ops->trigger) {
 471                 ret   = bch2_key_trigger(trans, i->btree_id, i->level,
 472                                 old, bkey_i_to_s(new),
 473                                 BTREE_TRIGGER_insert|BTREE_TRIGGER_overwrite|flags);
 474         } else {
 475                 ret   = bch2_key_trigger_new(trans, i->btree_id, i->level,
 476                                 bkey_i_to_s(new), flags) ?:
 477                         bch2_key_trigger_old(trans, i->btree_id, i->level,
 478                                 old, flags);
 479         }
 480
 481         return ret;
 482 }
 483
 484 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 485                                  bool overwrite)
 486 {
 487         /*
 488          * Transactional triggers create new btree_insert_entries, so we can't
 489          * pass them a pointer to a btree_insert_entry, that memory is going to
 490          * move:
 491          */
 492         struct bkey old_k = i->old_k;
 493         struct bkey_s_c old = { &old_k, i->old_v };
 494         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 495         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 496         unsigned flags = i->flags|BTREE_TRIGGER_transactional;
 497
 498         verify_update_old_key(trans, i);
 499
 500         if ((i->flags & BTREE_TRIGGER_norun) ||
 501             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 502                 return 0;
 503
 504         if (!i->insert_trigger_run &&
 505             !i->overwrite_trigger_run &&
 506             old_ops->trigger == new_ops->trigger) {
 507                 i->overwrite_trigger_run = true;
 508                 i->insert_trigger_run = true;
 509                 return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
 510                                         BTREE_TRIGGER_insert|
 511                                         BTREE_TRIGGER_overwrite|flags) ?: 1;
 512         } else if (overwrite && !i->overwrite_trigger_run) {
 513                 i->overwrite_trigger_run = true;
 514                 return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
 515         } else if (!overwrite && !i->insert_trigger_run) {
 516                 i->insert_trigger_run = true;
 517                 return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 518         } else {
 519                 return 0;
 520         }
 521 }
 522
 523 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 524                               unsigned btree_id_start)
 525 {
 526         bool trans_trigger_run;
 527         int ret, overwrite;
 528
 529         for (overwrite = 1; overwrite >= 0; --overwrite) {
 530
 531                 /*
 532                  * Running triggers will append more updates to the list of updates as
 533                  * we're walking it:
 534                  */
 535                 do {
 536                         trans_trigger_run = false;
 537
 538                         for (unsigned i = btree_id_start;
 539                              i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
 540                              i++) {
 541                                 if (trans->updates[i].btree_id != btree_id)
 542                                         continue;
 543
 544                                 ret = run_one_trans_trigger(trans, trans->updates + i, overwrite);
 545                                 if (ret < 0)
 546                                         return ret;
 547                                 if (ret)
 548                                         trans_trigger_run = true;
 549                         }
 550                 } while (trans_trigger_run);
 551         }
 552
 553         return 0;
 554 }
 555
 556 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 557 {
 558         unsigned btree_id = 0, btree_id_start = 0;
 559         int ret = 0;
 560
 561         /*
 562          *
 563          * For a given btree, this algorithm runs insert triggers before
 564          * overwrite triggers: this is so that when extents are being moved
 565          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 566          * they are re-added.
 567          */
 568         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 569                 if (btree_id == BTREE_ID_alloc)
 570                         continue;
 571
 572                 while (btree_id_start < trans->nr_updates &&
 573                        trans->updates[btree_id_start].btree_id < btree_id)
 574                         btree_id_start++;
 575
 576                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 577                 if (ret)
 578                         return ret;
 579         }
 580
 581         for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
 582                 struct btree_insert_entry *i = trans->updates + idx;
 583
 584                 if (i->btree_id > BTREE_ID_alloc)
 585                         break;
 586                 if (i->btree_id == BTREE_ID_alloc) {
 587                         ret = run_btree_triggers(trans, BTREE_ID_alloc, idx);
 588                         if (ret)
 589                                 return ret;
 590                         break;
 591                 }
 592         }
 593
 594 #ifdef CONFIG_BCACHEFS_DEBUG
 595         trans_for_each_update(trans, i)
 596                 BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
 597                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 598                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 599 #endif
 600         return 0;
 601 }
 602
 603 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 604 {
 605         trans_for_each_update(trans, i) {
 606                 /*
 607                  * XXX: synchronization of cached update triggers with gc
 608                  * XXX: synchronization of interior node updates with gc
 609                  */
 610                 BUG_ON(i->cached || i->level);
 611
 612                 if (btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)) &&
 613                     gc_visited(trans->c, gc_pos_btree_node(insert_l(trans, i)->b))) {
 614                         int ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_gc);
 615                         if (ret)
 616                                 return ret;
 617                 }
 618         }
 619
 620         return 0;
 621 }
 622
 623 static inline int
 624 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 625                                struct btree_insert_entry **stopped_at,
 626                                unsigned long trace_ip)
 627 {
 628         struct bch_fs *c = trans->c;
 629         struct btree_trans_commit_hook *h;
 630         unsigned u64s = 0;
 631         int ret;
 632
 633         bch2_trans_verify_not_unlocked(trans);
 634         bch2_trans_verify_not_in_restart(trans);
 635
 636         if (race_fault()) {
 637                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 638                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 639         }
 640
 641         /*
 642          * Check if the insert will fit in the leaf node with the write lock
 643          * held, otherwise another thread could write the node changing the
 644          * amount of space available:
 645          */
 646
 647         prefetch(&trans->c->journal.flags);
 648
 649         trans_for_each_update(trans, i) {
 650                 /* Multiple inserts might go to same leaf: */
 651                 if (!same_leaf_as_prev(trans, i))
 652                         u64s = 0;
 653
 654                 u64s += i->k->k.u64s;
 655                 ret = !i->cached
 656                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 657                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 658                 if (ret) {
 659                         *stopped_at = i;
 660                         return ret;
 661                 }
 662
 663                 i->k->k.needs_whiteout = false;
 664         }
 665
 666         /*
 667          * Don't get journal reservation until after we know insert will
 668          * succeed:
 669          */
 670         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 671                 ret = bch2_trans_journal_res_get(trans,
 672                                 (flags & BCH_WATERMARK_MASK)|
 673                                 JOURNAL_RES_GET_NONBLOCK);
 674                 if (ret)
 675                         return ret;
 676
 677                 if (unlikely(trans->journal_transaction_names))
 678                         journal_transaction_name(trans);
 679         }
 680
 681         /*
 682          * Not allowed to fail after we've gotten our journal reservation - we
 683          * have to use it:
 684          */
 685
 686         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 687             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 688                 if (bch2_journal_seq_verify)
 689                         trans_for_each_update(trans, i)
 690                                 i->k->k.version.lo = trans->journal_res.seq;
 691                 else if (bch2_inject_invalid_keys)
 692                         trans_for_each_update(trans, i)
 693                                 i->k->k.version = MAX_VERSION;
 694         }
 695
 696         if (trans->fs_usage_deltas &&
 697             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 698                 return -BCH_ERR_btree_insert_need_mark_replicas;
 699
 700         /* XXX: we only want to run this if deltas are nonzero */
 701         bch2_trans_account_disk_usage_change(trans);
 702
 703         h = trans->hooks;
 704         while (h) {
 705                 ret = h->fn(trans, h);
 706                 if (ret)
 707                         goto revert_fs_usage;
 708                 h = h->next;
 709         }
 710
 711         trans_for_each_update(trans, i)
 712                 if (BTREE_NODE_TYPE_HAS_ATOMIC_TRIGGERS & (1U << i->bkey_type)) {
 713                         ret = run_one_mem_trigger(trans, i, BTREE_TRIGGER_atomic|i->flags);
 714                         if (ret)
 715                                 goto fatal_err;
 716                 }
 717
 718         if (unlikely(c->gc_pos.phase)) {
 719                 ret = bch2_trans_commit_run_gc_triggers(trans);
 720                 if  (ret)
 721                         goto fatal_err;
 722         }
 723
 724         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 725                 struct journal *j = &c->journal;
 726                 struct jset_entry *entry;
 727
 728                 trans_for_each_update(trans, i) {
 729                         if (i->key_cache_already_flushed)
 730                                 continue;
 731
 732                         if (i->flags & BTREE_UPDATE_nojournal)
 733                                 continue;
 734
 735                         verify_update_old_key(trans, i);
 736
 737                         if (trans->journal_transaction_names) {
 738                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 739                                                        BCH_JSET_ENTRY_overwrite,
 740                                                        i->btree_id, i->level,
 741                                                        i->old_k.u64s);
 742                                 bkey_reassemble((struct bkey_i *) entry->start,
 743                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 744                         }
 745
 746                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 747                                                BCH_JSET_ENTRY_btree_keys,
 748                                                i->btree_id, i->level,
 749                                                i->k->k.u64s);
 750                         bkey_copy((struct bkey_i *) entry->start, i->k);
 751                 }
 752
 753                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 754                                   trans->journal_entries,
 755                                   trans->journal_entries_u64s);
 756
 757                 trans->journal_res.offset       += trans->journal_entries_u64s;
 758                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 759
 760                 if (trans->journal_seq)
 761                         *trans->journal_seq = trans->journal_res.seq;
 762         }
 763
 764         trans_for_each_update(trans, i) {
 765                 struct btree_path *path = trans->paths + i->path;
 766
 767                 if (!i->cached) {
 768                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 769                 } else if (!i->key_cache_already_flushed)
 770                         bch2_btree_insert_key_cached(trans, flags, i);
 771                 else {
 772                         bch2_btree_key_cache_drop(trans, path);
 773                         btree_path_set_dirty(path, BTREE_ITER_NEED_TRAVERSE);
 774                 }
 775         }
 776
 777         return 0;
 778 fatal_err:
 779         bch2_fatal_error(c);
 780 revert_fs_usage:
 781         if (trans->fs_usage_deltas)
 782                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 783         return ret;
 784 }
 785
 786 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 787 {
 788         trans_for_each_update(trans, i)
 789                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 790 }
 791
 792 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 793                                                    enum bch_validate_flags flags,
 794                                                    struct btree_insert_entry *i,
 795                                                    struct printbuf *err)
 796 {
 797         struct bch_fs *c = trans->c;
 798
 799         printbuf_reset(err);
 800         prt_printf(err, "invalid bkey on insert from %s -> %ps\n",
 801                    trans->fn, (void *) i->ip_allocated);
 802         printbuf_indent_add(err, 2);
 803
 804         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 805         prt_newline(err);
 806
 807         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 808         bch2_print_string_as_lines(KERN_ERR, err->buf);
 809
 810         bch2_inconsistent_error(c);
 811         bch2_dump_trans_updates(trans);
 812
 813         return -EINVAL;
 814 }
 815
 816 static noinline int bch2_trans_commit_journal_entry_invalid(struct btree_trans *trans,
 817                                                    struct jset_entry *i)
 818 {
 819         struct bch_fs *c = trans->c;
 820         struct printbuf buf = PRINTBUF;
 821
 822         prt_printf(&buf, "invalid bkey on insert from %s\n", trans->fn);
 823         printbuf_indent_add(&buf, 2);
 824
 825         bch2_journal_entry_to_text(&buf, c, i);
 826         prt_newline(&buf);
 827
 828         bch2_print_string_as_lines(KERN_ERR, buf.buf);
 829
 830         bch2_inconsistent_error(c);
 831         bch2_dump_trans_updates(trans);
 832
 833         return -EINVAL;
 834 }
 835
 836 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 837                                 struct journal_entry_pin *_pin, u64 seq)
 838 {
 839         return 0;
 840 }
 841
 842 /*
 843  * Get journal reservation, take write locks, and attempt to do btree update(s):
 844  */
 845 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 846                                        struct btree_insert_entry **stopped_at,
 847                                        unsigned long trace_ip)
 848 {
 849         struct bch_fs *c = trans->c;
 850         int ret = 0, u64s_delta = 0;
 851
 852         for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
 853                 struct btree_insert_entry *i = trans->updates + idx;
 854                 if (i->cached)
 855                         continue;
 856
 857                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 858                 u64s_delta -= i->old_btree_u64s;
 859
 860                 if (!same_leaf_as_next(trans, i)) {
 861                         if (u64s_delta <= 0) {
 862                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 863                                                         i->level, flags);
 864                                 if (unlikely(ret))
 865                                         return ret;
 866                         }
 867
 868                         u64s_delta = 0;
 869                 }
 870         }
 871
 872         ret = bch2_trans_lock_write(trans);
 873         if (unlikely(ret))
 874                 return ret;
 875
 876         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 877
 878         if (!ret && unlikely(trans->journal_replay_not_finished))
 879                 bch2_drop_overwrites_from_journal(trans);
 880
 881         bch2_trans_unlock_write(trans);
 882
 883         if (!ret && trans->journal_pin)
 884                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 885                                      trans->journal_pin,
 886                                      bch2_trans_commit_journal_pin_flush);
 887
 888         /*
 889          * Drop journal reservation after dropping write locks, since dropping
 890          * the journal reservation may kick off a journal write:
 891          */
 892         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 893                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 894
 895         return ret;
 896 }
 897
 898 static int journal_reclaim_wait_done(struct bch_fs *c)
 899 {
 900         int ret = bch2_journal_error(&c->journal) ?:
 901                 !bch2_btree_key_cache_must_wait(c);
 902
 903         if (!ret)
 904                 journal_reclaim_kick(&c->journal);
 905         return ret;
 906 }
 907
 908 static noinline
 909 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 910                             struct btree_insert_entry *i,
 911                             int ret, unsigned long trace_ip)
 912 {
 913         struct bch_fs *c = trans->c;
 914         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 915
 916         switch (ret) {
 917         case -BCH_ERR_btree_insert_btree_node_full:
 918                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 919                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 920                         trace_and_count(c, trans_restart_btree_node_split, trans,
 921                                         trace_ip, trans->paths + i->path);
 922                 break;
 923         case -BCH_ERR_btree_insert_need_mark_replicas:
 924                 ret = drop_locks_do(trans,
 925                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 926                 break;
 927         case -BCH_ERR_journal_res_get_blocked:
 928                 /*
 929                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 930                  * flag
 931                  */
 932                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 933                     watermark < BCH_WATERMARK_reclaim) {
 934                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 935                         break;
 936                 }
 937
 938                 ret = drop_locks_do(trans,
 939                         bch2_trans_journal_res_get(trans,
 940                                         (flags & BCH_WATERMARK_MASK)|
 941                                         JOURNAL_RES_GET_CHECK));
 942                 break;
 943         case -BCH_ERR_btree_insert_need_journal_reclaim:
 944                 bch2_trans_unlock(trans);
 945
 946                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 947
 948                 wait_event_freezable(c->journal.reclaim_wait,
 949                                      (ret = journal_reclaim_wait_done(c)));
 950                 if (ret < 0)
 951                         break;
 952
 953                 ret = bch2_trans_relock(trans);
 954                 break;
 955         default:
 956                 BUG_ON(ret >= 0);
 957                 break;
 958         }
 959
 960         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 961
 962         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 963                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 964                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 965
 966         return ret;
 967 }
 968
 969 static noinline int
 970 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 971 {
 972         struct bch_fs *c = trans->c;
 973         int ret;
 974
 975         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 976             test_bit(BCH_FS_started, &c->flags))
 977                 return -BCH_ERR_erofs_trans_commit;
 978
 979         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 980         if (ret)
 981                 return ret;
 982
 983         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 984         return 0;
 985 }
 986
 987 /*
 988  * This is for updates done in the early part of fsck - btree_gc - before we've
 989  * gone RW. we only add the new key to the list of keys for journal replay to
 990  * do.
 991  */
 992 static noinline int
 993 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 994 {
 995         struct bch_fs *c = trans->c;
 996         int ret = 0;
 997
 998         trans_for_each_update(trans, i) {
 999                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1000                 if (ret)
1001                         break;
1002         }
1003
1004         return ret;
1005 }
1006
1007 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1008 {
1009         struct btree_insert_entry *errored_at = NULL;
1010         struct bch_fs *c = trans->c;
1011         int ret = 0;
1012
1013         bch2_trans_verify_not_unlocked(trans);
1014         bch2_trans_verify_not_in_restart(trans);
1015
1016         if (!trans->nr_updates &&
1017             !trans->journal_entries_u64s)
1018                 goto out_reset;
1019
1020         memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));
1021
1022         ret = bch2_trans_commit_run_triggers(trans);
1023         if (ret)
1024                 goto out_reset;
1025
1026         trans_for_each_update(trans, i) {
1027                 struct printbuf buf = PRINTBUF;
1028                 enum bch_validate_flags invalid_flags = 0;
1029
1030                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1031                         invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
1032
1033                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1034                                                i->bkey_type, invalid_flags, &buf)))
1035                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1036                 btree_insert_entry_checks(trans, i);
1037                 printbuf_exit(&buf);
1038
1039                 if (ret)
1040                         return ret;
1041         }
1042
1043         for (struct jset_entry *i = trans->journal_entries;
1044              i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
1045              i = vstruct_next(i)) {
1046                 enum bch_validate_flags invalid_flags = 0;
1047
1048                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1049                         invalid_flags |= BCH_VALIDATE_write|BCH_VALIDATE_commit;
1050
1051                 if (unlikely(bch2_journal_entry_validate(c, NULL, i,
1052                                         bcachefs_metadata_version_current,
1053                                         CPU_BIG_ENDIAN, invalid_flags)))
1054                         ret = bch2_trans_commit_journal_entry_invalid(trans, i);
1055
1056                 if (ret)
1057                         return ret;
1058         }
1059
1060         if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags))) {
1061                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1062                 goto out_reset;
1063         }
1064
1065         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1066             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1067                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1068                 if (ret)
1069                         goto out_reset;
1070         }
1071
1072         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1073
1074         trans->journal_u64s             = trans->journal_entries_u64s;
1075         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1076         if (trans->journal_transaction_names)
1077                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1078
1079         trans_for_each_update(trans, i) {
1080                 struct btree_path *path = trans->paths + i->path;
1081
1082                 EBUG_ON(!path->should_be_locked);
1083
1084                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1085                 if (unlikely(ret))
1086                         goto out;
1087
1088                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1089
1090                 if (i->key_cache_already_flushed)
1091                         continue;
1092
1093                 if (i->flags & BTREE_UPDATE_nojournal)
1094                         continue;
1095
1096                 /* we're going to journal the key being updated: */
1097                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1098
1099                 /* and we're also going to log the overwrite: */
1100                 if (trans->journal_transaction_names)
1101                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1102         }
1103
1104         if (trans->extra_disk_res) {
1105                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1106                                 trans->extra_disk_res,
1107                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1108                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1109                 if (ret)
1110                         goto err;
1111         }
1112 retry:
1113         errored_at = NULL;
1114         bch2_trans_verify_not_unlocked(trans);
1115         bch2_trans_verify_not_in_restart(trans);
1116         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1117                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1118
1119         ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1120
1121         /* make sure we didn't drop or screw up locks: */
1122         bch2_trans_verify_locks(trans);
1123
1124         if (ret)
1125                 goto err;
1126
1127         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1128 out:
1129         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1130                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1131 out_reset:
1132         if (!ret)
1133                 bch2_trans_downgrade(trans);
1134         bch2_trans_reset_updates(trans);
1135
1136         return ret;
1137 err:
1138         ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1139         if (ret)
1140                 goto out;
1141
1142         /*
1143          * We might have done another transaction commit in the error path -
1144          * i.e. btree write buffer flush - which will have made use of
1145          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1146          * how the journal sequence number to pin is passed in - so we must
1147          * restart:
1148          */
1149         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1150                 ret = -BCH_ERR_transaction_restart_nested;
1151                 goto out;
1152         }
1153
1154         goto retry;
1155 }