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 "disk_accounting.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "journal.h"
  17 #include "journal_io.h"
  18 #include "journal_reclaim.h"
  19 #include "replicas.h"
  20 #include "snapshot.h"
  21
  22 #include <linux/prefetch.h>
  23
  24 static const char * const trans_commit_flags_strs[] = {
  25 #define x(n, ...) #n,
  26         BCH_TRANS_COMMIT_FLAGS()
  27 #undef x
  28         NULL
  29 };
  30
  31 void bch2_trans_commit_flags_to_text(struct printbuf *out, enum bch_trans_commit_flags flags)
  32 {
  33         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
  34
  35         prt_printf(out, "watermark=%s", bch2_watermarks[watermark]);
  36
  37         flags >>= BCH_WATERMARK_BITS;
  38         if (flags) {
  39                 prt_char(out, ' ');
  40                 bch2_prt_bitflags(out, trans_commit_flags_strs, flags);
  41         }
  42 }
  43
  44 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  45 {
  46 #ifdef CONFIG_BCACHEFS_DEBUG
  47         struct bch_fs *c = trans->c;
  48         struct bkey u;
  49         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(trans->paths + i->path, &u);
  50
  51         if (unlikely(trans->journal_replay_not_finished)) {
  52                 struct bkey_i *j_k =
  53                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  54
  55                 if (j_k)
  56                         k = bkey_i_to_s_c(j_k);
  57         }
  58
  59         u = *k.k;
  60         u.needs_whiteout = i->old_k.needs_whiteout;
  61
  62         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  63         BUG_ON(i->old_v != k.v);
  64 #endif
  65 }
  66
  67 static inline struct btree_path_level *insert_l(struct btree_trans *trans, struct btree_insert_entry *i)
  68 {
  69         return (trans->paths + i->path)->l + i->level;
  70 }
  71
  72 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  73                                      struct btree_insert_entry *i)
  74 {
  75         return i != trans->updates &&
  76                 insert_l(trans, &i[0])->b == insert_l(trans, &i[-1])->b;
  77 }
  78
  79 static inline bool same_leaf_as_next(struct btree_trans *trans,
  80                                      struct btree_insert_entry *i)
  81 {
  82         return i + 1 < trans->updates + trans->nr_updates &&
  83                 insert_l(trans, &i[0])->b == insert_l(trans, &i[1])->b;
  84 }
  85
  86 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  87                                            struct btree_path *path,
  88                                            struct btree *b)
  89 {
  90         struct bch_fs *c = trans->c;
  91
  92         if (unlikely(btree_node_just_written(b)) &&
  93             bch2_btree_post_write_cleanup(c, b))
  94                 bch2_trans_node_reinit_iter(trans, b);
  95
  96         /*
  97          * If the last bset has been written, or if it's gotten too big - start
  98          * a new bset to insert into:
  99          */
 100         if (want_new_bset(c, b))
 101                 bch2_btree_init_next(trans, b);
 102 }
 103
 104 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 105 {
 106         while (--i >= trans->updates) {
 107                 if (same_leaf_as_prev(trans, i))
 108                         continue;
 109
 110                 bch2_btree_node_unlock_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 111         }
 112
 113         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 114         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 115 }
 116
 117 static inline int bch2_trans_lock_write(struct btree_trans *trans)
 118 {
 119         EBUG_ON(trans->write_locked);
 120
 121         trans_for_each_update(trans, i) {
 122                 if (same_leaf_as_prev(trans, i))
 123                         continue;
 124
 125                 if (bch2_btree_node_lock_write(trans, trans->paths + i->path, &insert_l(trans, i)->b->c))
 126                         return trans_lock_write_fail(trans, i);
 127
 128                 if (!i->cached)
 129                         bch2_btree_node_prep_for_write(trans, trans->paths + i->path, insert_l(trans, i)->b);
 130         }
 131
 132         trans->write_locked = true;
 133         return 0;
 134 }
 135
 136 static inline void bch2_trans_unlock_updates_write(struct btree_trans *trans)
 137 {
 138         if (likely(trans->write_locked)) {
 139                 trans_for_each_update(trans, i)
 140                         if (btree_node_locked_type(trans->paths + i->path, i->level) ==
 141                             BTREE_NODE_WRITE_LOCKED)
 142                                 bch2_btree_node_unlock_write_inlined(trans,
 143                                                 trans->paths + i->path, insert_l(trans, i)->b);
 144                 trans->write_locked = false;
 145         }
 146 }
 147
 148 /* Inserting into a given leaf node (last stage of insert): */
 149
 150 /* Handle overwrites and do insert, for non extents: */
 151 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 152                                 struct btree_path *path,
 153                                 struct btree *b,
 154                                 struct btree_node_iter *node_iter,
 155                                 struct bkey_i *insert)
 156 {
 157         struct bkey_packed *k;
 158         unsigned clobber_u64s = 0, new_u64s = 0;
 159
 160         EBUG_ON(btree_node_just_written(b));
 161         EBUG_ON(bset_written(b, btree_bset_last(b)));
 162         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 163         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 164         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 165         EBUG_ON(insert->k.u64s > bch2_btree_keys_u64s_remaining(b));
 166         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 167
 168         k = bch2_btree_node_iter_peek_all(node_iter, b);
 169         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 170                 k = NULL;
 171
 172         /* @k is the key being overwritten/deleted, if any: */
 173         EBUG_ON(k && bkey_deleted(k));
 174
 175         /* Deleting, but not found? nothing to do: */
 176         if (bkey_deleted(&insert->k) && !k)
 177                 return false;
 178
 179         if (bkey_deleted(&insert->k)) {
 180                 /* Deleting: */
 181                 btree_account_key_drop(b, k);
 182                 k->type = KEY_TYPE_deleted;
 183
 184                 if (k->needs_whiteout)
 185                         push_whiteout(b, insert->k.p);
 186                 k->needs_whiteout = false;
 187
 188                 if (k >= btree_bset_last(b)->start) {
 189                         clobber_u64s = k->u64s;
 190                         bch2_bset_delete(b, k, clobber_u64s);
 191                         goto fix_iter;
 192                 } else {
 193                         bch2_btree_path_fix_key_modified(trans, b, k);
 194                 }
 195
 196                 return true;
 197         }
 198
 199         if (k) {
 200                 /* Overwriting: */
 201                 btree_account_key_drop(b, k);
 202                 k->type = KEY_TYPE_deleted;
 203
 204                 insert->k.needs_whiteout = k->needs_whiteout;
 205                 k->needs_whiteout = false;
 206
 207                 if (k >= btree_bset_last(b)->start) {
 208                         clobber_u64s = k->u64s;
 209                         goto overwrite;
 210                 } else {
 211                         bch2_btree_path_fix_key_modified(trans, b, k);
 212                 }
 213         }
 214
 215         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 216 overwrite:
 217         bch2_bset_insert(b, k, insert, clobber_u64s);
 218         new_u64s = k->u64s;
 219 fix_iter:
 220         if (clobber_u64s != new_u64s)
 221                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 222                                          clobber_u64s, new_u64s);
 223         return true;
 224 }
 225
 226 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 227                                unsigned i, u64 seq)
 228 {
 229         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 230         struct btree_write *w = container_of(pin, struct btree_write, journal);
 231         struct btree *b = container_of(w, struct btree, writes[i]);
 232         struct btree_trans *trans = bch2_trans_get(c);
 233         unsigned long old, new;
 234         unsigned idx = w - b->writes;
 235
 236         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 237
 238         old = READ_ONCE(b->flags);
 239         do {
 240                 new = old;
 241
 242                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 243                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 244                     w->journal.seq != seq)
 245                         break;
 246
 247                 new &= ~BTREE_WRITE_TYPE_MASK;
 248                 new |= BTREE_WRITE_journal_reclaim;
 249                 new |= 1 << BTREE_NODE_need_write;
 250         } while (!try_cmpxchg(&b->flags, &old, new));
 251
 252         btree_node_write_if_need(trans, b, SIX_LOCK_read);
 253         six_unlock_read(&b->c.lock);
 254
 255         bch2_trans_put(trans);
 256         return 0;
 257 }
 258
 259 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 260 {
 261         return __btree_node_flush(j, pin, 0, seq);
 262 }
 263
 264 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 265 {
 266         return __btree_node_flush(j, pin, 1, seq);
 267 }
 268
 269 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 270                                        struct btree *b, u64 seq)
 271 {
 272         struct btree_write *w = btree_current_write(b);
 273
 274         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 275                              btree_node_write_idx(b) == 0
 276                              ? bch2_btree_node_flush0
 277                              : bch2_btree_node_flush1);
 278 }
 279
 280 /**
 281  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 282  * @trans:              btree transaction object
 283  * @path:               path pointing to @insert's pos
 284  * @insert:             key to insert
 285  * @journal_seq:        sequence number of journal reservation
 286  */
 287 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 288                                        struct btree_path *path,
 289                                        struct bkey_i *insert,
 290                                        u64 journal_seq)
 291 {
 292         struct bch_fs *c = trans->c;
 293         struct btree *b = path_l(path)->b;
 294         struct bset_tree *t = bset_tree_last(b);
 295         struct bset *i = bset(b, t);
 296         int old_u64s = bset_u64s(t);
 297         int old_live_u64s = b->nr.live_u64s;
 298         int live_u64s_added, u64s_added;
 299
 300         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 301                                         &path_l(path)->iter, insert)))
 302                 return;
 303
 304         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 305
 306         bch2_btree_add_journal_pin(c, b, journal_seq);
 307
 308         if (unlikely(!btree_node_dirty(b))) {
 309                 EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
 310                 set_btree_node_dirty_acct(c, b);
 311         }
 312
 313         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 314         u64s_added = (int) bset_u64s(t) - old_u64s;
 315
 316         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 317                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 318         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 319                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 320
 321         if (u64s_added > live_u64s_added &&
 322             bch2_maybe_compact_whiteouts(c, b))
 323                 bch2_trans_node_reinit_iter(trans, b);
 324 }
 325
 326 /* Cached btree updates: */
 327
 328 /* Normal update interface: */
 329
 330 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 331                                              struct btree_insert_entry *i)
 332 {
 333         struct btree_path *path = trans->paths + i->path;
 334
 335         BUG_ON(!bpos_eq(i->k->k.p, path->pos));
 336         BUG_ON(i->cached        != path->cached);
 337         BUG_ON(i->level         != path->level);
 338         BUG_ON(i->btree_id      != path->btree_id);
 339         EBUG_ON(!i->level &&
 340                 btree_type_has_snapshots(i->btree_id) &&
 341                 !(i->flags & BTREE_UPDATE_internal_snapshot_node) &&
 342                 test_bit(JOURNAL_replay_done, &trans->c->journal.flags) &&
 343                 i->k->k.p.snapshot &&
 344                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot) > 0);
 345 }
 346
 347 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 348                                                       unsigned flags)
 349 {
 350         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 351                                     trans->journal_u64s, flags, trans);
 352 }
 353
 354 #define JSET_ENTRY_LOG_U64s             4
 355
 356 static noinline void journal_transaction_name(struct btree_trans *trans)
 357 {
 358         struct bch_fs *c = trans->c;
 359         struct journal *j = &c->journal;
 360         struct jset_entry *entry =
 361                 bch2_journal_add_entry(j, &trans->journal_res,
 362                                        BCH_JSET_ENTRY_log, 0, 0,
 363                                        JSET_ENTRY_LOG_U64s);
 364         struct jset_entry_log *l =
 365                 container_of(entry, struct jset_entry_log, entry);
 366
 367         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 368 }
 369
 370 static inline int btree_key_can_insert(struct btree_trans *trans,
 371                                        struct btree *b, unsigned u64s)
 372 {
 373         if (!bch2_btree_node_insert_fits(b, u64s))
 374                 return -BCH_ERR_btree_insert_btree_node_full;
 375
 376         return 0;
 377 }
 378
 379 noinline static int
 380 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 381                                      struct btree_path *path, unsigned new_u64s)
 382 {
 383         struct bkey_cached *ck = (void *) path->l[0].b;
 384         struct bkey_i *new_k;
 385         int ret;
 386
 387         bch2_trans_unlock_updates_write(trans);
 388         bch2_trans_unlock(trans);
 389
 390         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 391         if (!new_k) {
 392                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 393                         bch2_btree_id_str(path->btree_id), new_u64s);
 394                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 395         }
 396
 397         ret =   bch2_trans_relock(trans) ?:
 398                 bch2_trans_lock_write(trans);
 399         if (unlikely(ret)) {
 400                 kfree(new_k);
 401                 return ret;
 402         }
 403
 404         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 405
 406         trans_for_each_update(trans, i)
 407                 if (i->old_v == &ck->k->v)
 408                         i->old_v = &new_k->v;
 409
 410         kfree(ck->k);
 411         ck->u64s        = new_u64s;
 412         ck->k           = new_k;
 413         return 0;
 414 }
 415
 416 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 417                                        struct btree_path *path, unsigned u64s)
 418 {
 419         struct bch_fs *c = trans->c;
 420         struct bkey_cached *ck = (void *) path->l[0].b;
 421         unsigned new_u64s;
 422         struct bkey_i *new_k;
 423         unsigned watermark = flags & BCH_WATERMARK_MASK;
 424
 425         EBUG_ON(path->level);
 426
 427         if (watermark < BCH_WATERMARK_reclaim &&
 428             !test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 429             bch2_btree_key_cache_must_wait(c))
 430                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 431
 432         /*
 433          * bch2_varint_decode can read past the end of the buffer by at most 7
 434          * bytes (it won't be used):
 435          */
 436         u64s += 1;
 437
 438         if (u64s <= ck->u64s)
 439                 return 0;
 440
 441         new_u64s        = roundup_pow_of_two(u64s);
 442         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT|__GFP_NOWARN);
 443         if (unlikely(!new_k))
 444                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 445
 446         trans_for_each_update(trans, i)
 447                 if (i->old_v == &ck->k->v)
 448                         i->old_v = &new_k->v;
 449
 450         ck->u64s        = new_u64s;
 451         ck->k           = new_k;
 452         return 0;
 453 }
 454
 455 /* Triggers: */
 456
 457 static int run_one_mem_trigger(struct btree_trans *trans,
 458                                struct btree_insert_entry *i,
 459                                unsigned flags)
 460 {
 461         verify_update_old_key(trans, i);
 462
 463         if (unlikely(flags & BTREE_TRIGGER_norun))
 464                 return 0;
 465
 466         struct bkey_s_c old = { &i->old_k, i->old_v };
 467         struct bkey_i *new = i->k;
 468         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 469         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 470
 471         if (old_ops->trigger == new_ops->trigger)
 472                 return bch2_key_trigger(trans, i->btree_id, i->level,
 473                                 old, bkey_i_to_s(new),
 474                                 BTREE_TRIGGER_insert|BTREE_TRIGGER_overwrite|flags);
 475         else
 476                 return bch2_key_trigger_new(trans, i->btree_id, i->level,
 477                                 bkey_i_to_s(new), flags) ?:
 478                        bch2_key_trigger_old(trans, i->btree_id, i->level,
 479                                 old, flags);
 480 }
 481
 482 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i)
 483 {
 484         verify_update_old_key(trans, i);
 485
 486         if ((i->flags & BTREE_TRIGGER_norun) ||
 487             !btree_node_type_has_trans_triggers(i->bkey_type))
 488                 return 0;
 489
 490         /*
 491          * Transactional triggers create new btree_insert_entries, so we can't
 492          * pass them a pointer to a btree_insert_entry, that memory is going to
 493          * move:
 494          */
 495         struct bkey old_k = i->old_k;
 496         struct bkey_s_c old = { &old_k, i->old_v };
 497         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 498         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 499         unsigned flags = i->flags|BTREE_TRIGGER_transactional;
 500
 501         if (!i->insert_trigger_run &&
 502             !i->overwrite_trigger_run &&
 503             old_ops->trigger == new_ops->trigger) {
 504                 i->overwrite_trigger_run = true;
 505                 i->insert_trigger_run = true;
 506                 return bch2_key_trigger(trans, i->btree_id, i->level, old, bkey_i_to_s(i->k),
 507                                         BTREE_TRIGGER_insert|
 508                                         BTREE_TRIGGER_overwrite|flags) ?: 1;
 509         } else if (!i->overwrite_trigger_run) {
 510                 i->overwrite_trigger_run = true;
 511                 return bch2_key_trigger_old(trans, i->btree_id, i->level, old, flags) ?: 1;
 512         } else if (!i->insert_trigger_run) {
 513                 i->insert_trigger_run = true;
 514                 return bch2_key_trigger_new(trans, i->btree_id, i->level, bkey_i_to_s(i->k), flags) ?: 1;
 515         } else {
 516                 return 0;
 517         }
 518 }
 519
 520 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 521                               unsigned *btree_id_updates_start)
 522 {
 523         bool trans_trigger_run;
 524
 525         /*
 526          * Running triggers will append more updates to the list of updates as
 527          * we're walking it:
 528          */
 529         do {
 530                 trans_trigger_run = false;
 531
 532                 for (unsigned i = *btree_id_updates_start;
 533                      i < trans->nr_updates && trans->updates[i].btree_id <= btree_id;
 534                      i++) {
 535                         if (trans->updates[i].btree_id < btree_id) {
 536                                 *btree_id_updates_start = i;
 537                                 continue;
 538                         }
 539
 540                         int ret = run_one_trans_trigger(trans, trans->updates + i);
 541                         if (ret < 0)
 542                                 return ret;
 543                         if (ret)
 544                                 trans_trigger_run = true;
 545                 }
 546         } while (trans_trigger_run);
 547
 548         trans_for_each_update(trans, i)
 549                 BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
 550                        i->btree_id == btree_id &&
 551                        btree_node_type_has_trans_triggers(i->bkey_type) &&
 552                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 553
 554         return 0;
 555 }
 556
 557 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 558 {
 559         unsigned btree_id = 0, btree_id_updates_start = 0;
 560         int ret = 0;
 561
 562         /*
 563          *
 564          * For a given btree, this algorithm runs insert triggers before
 565          * overwrite triggers: this is so that when extents are being moved
 566          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 567          * they are re-added.
 568          */
 569         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 570                 if (btree_id == BTREE_ID_alloc)
 571                         continue;
 572
 573                 ret = run_btree_triggers(trans, btree_id, &btree_id_updates_start);
 574                 if (ret)
 575                         return ret;
 576         }
 577
 578         btree_id_updates_start = 0;
 579         ret = run_btree_triggers(trans, BTREE_ID_alloc, &btree_id_updates_start);
 580         if (ret)
 581                 return ret;
 582
 583 #ifdef CONFIG_BCACHEFS_DEBUG
 584         trans_for_each_update(trans, i)
 585                 BUG_ON(!(i->flags & BTREE_TRIGGER_norun) &&
 586                        btree_node_type_has_trans_triggers(i->bkey_type) &&
 587                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 588 #endif
 589         return 0;
 590 }
 591
 592 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 593 {
 594         trans_for_each_update(trans, i)
 595                 if (btree_node_type_has_triggers(i->bkey_type) &&
 596                     gc_visited(trans->c, gc_pos_btree(i->btree_id, i->level, i->k->k.p))) {
 597                         int ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_gc);
 598                         if (ret)
 599                                 return ret;
 600                 }
 601
 602         return 0;
 603 }
 604
 605 static inline int
 606 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 607                                struct btree_insert_entry **stopped_at,
 608                                unsigned long trace_ip)
 609 {
 610         struct bch_fs *c = trans->c;
 611         struct btree_trans_commit_hook *h;
 612         unsigned u64s = 0;
 613         int ret = 0;
 614
 615         bch2_trans_verify_not_unlocked_or_in_restart(trans);
 616
 617         if (race_fault()) {
 618                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 619                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_fault_inject);
 620         }
 621
 622         /*
 623          * Check if the insert will fit in the leaf node with the write lock
 624          * held, otherwise another thread could write the node changing the
 625          * amount of space available:
 626          */
 627
 628         prefetch(&trans->c->journal.flags);
 629
 630         trans_for_each_update(trans, i) {
 631                 /* Multiple inserts might go to same leaf: */
 632                 if (!same_leaf_as_prev(trans, i))
 633                         u64s = 0;
 634
 635                 u64s += i->k->k.u64s;
 636                 ret = !i->cached
 637                         ? btree_key_can_insert(trans, insert_l(trans, i)->b, u64s)
 638                         : btree_key_can_insert_cached(trans, flags, trans->paths + i->path, u64s);
 639                 if (ret) {
 640                         *stopped_at = i;
 641                         return ret;
 642                 }
 643
 644                 i->k->k.needs_whiteout = false;
 645         }
 646
 647         /*
 648          * Don't get journal reservation until after we know insert will
 649          * succeed:
 650          */
 651         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 652                 ret = bch2_trans_journal_res_get(trans,
 653                                 (flags & BCH_WATERMARK_MASK)|
 654                                 JOURNAL_RES_GET_NONBLOCK);
 655                 if (ret)
 656                         return ret;
 657
 658                 if (unlikely(trans->journal_transaction_names))
 659                         journal_transaction_name(trans);
 660         }
 661
 662         /*
 663          * Not allowed to fail after we've gotten our journal reservation - we
 664          * have to use it:
 665          */
 666
 667         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 668             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 669                 if (bch2_journal_seq_verify)
 670                         trans_for_each_update(trans, i)
 671                                 i->k->k.bversion.lo = trans->journal_res.seq;
 672                 else if (bch2_inject_invalid_keys)
 673                         trans_for_each_update(trans, i)
 674                                 i->k->k.bversion = MAX_VERSION;
 675         }
 676
 677         h = trans->hooks;
 678         while (h) {
 679                 ret = h->fn(trans, h);
 680                 if (ret)
 681                         return ret;
 682                 h = h->next;
 683         }
 684
 685         struct jset_entry *entry = trans->journal_entries;
 686
 687         percpu_down_read(&c->mark_lock);
 688         for (entry = trans->journal_entries;
 689              entry != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
 690              entry = vstruct_next(entry))
 691                 if (entry->type == BCH_JSET_ENTRY_write_buffer_keys &&
 692                     entry->start->k.type == KEY_TYPE_accounting) {
 693                         ret = bch2_accounting_trans_commit_hook(trans, bkey_i_to_accounting(entry->start), flags);
 694                         if (ret)
 695                                 goto revert_fs_usage;
 696                 }
 697         percpu_up_read(&c->mark_lock);
 698
 699         /* XXX: we only want to run this if deltas are nonzero */
 700         bch2_trans_account_disk_usage_change(trans);
 701
 702         trans_for_each_update(trans, i)
 703                 if (btree_node_type_has_atomic_triggers(i->bkey_type)) {
 704                         ret = run_one_mem_trigger(trans, i, BTREE_TRIGGER_atomic|i->flags);
 705                         if (ret)
 706                                 goto fatal_err;
 707                 }
 708
 709         if (unlikely(c->gc_pos.phase)) {
 710                 ret = bch2_trans_commit_run_gc_triggers(trans);
 711                 if  (ret)
 712                         goto fatal_err;
 713         }
 714
 715         struct bkey_validate_context validate_context = { .from = BKEY_VALIDATE_commit };
 716
 717         if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
 718                 validate_context.flags = BCH_VALIDATE_write|BCH_VALIDATE_commit;
 719
 720         for (struct jset_entry *i = trans->journal_entries;
 721              i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
 722              i = vstruct_next(i)) {
 723                 ret = bch2_journal_entry_validate(c, NULL, i,
 724                                                   bcachefs_metadata_version_current,
 725                                                   CPU_BIG_ENDIAN, validate_context);
 726                 if (unlikely(ret)) {
 727                         bch2_trans_inconsistent(trans, "invalid journal entry on insert from %s\n",
 728                                                 trans->fn);
 729                         goto fatal_err;
 730                 }
 731         }
 732
 733         trans_for_each_update(trans, i) {
 734                 validate_context.level  = i->level;
 735                 validate_context.btree  = i->btree_id;
 736
 737                 ret = bch2_bkey_validate(c, bkey_i_to_s_c(i->k), validate_context);
 738                 if (unlikely(ret)){
 739                         bch2_trans_inconsistent(trans, "invalid bkey on insert from %s -> %ps\n",
 740                                                 trans->fn, (void *) i->ip_allocated);
 741                         goto fatal_err;
 742                 }
 743                 btree_insert_entry_checks(trans, i);
 744         }
 745
 746         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 747                 struct journal *j = &c->journal;
 748                 struct jset_entry *entry;
 749
 750                 trans_for_each_update(trans, i) {
 751                         if (i->key_cache_already_flushed)
 752                                 continue;
 753
 754                         if (i->flags & BTREE_UPDATE_nojournal)
 755                                 continue;
 756
 757                         verify_update_old_key(trans, i);
 758
 759                         if (trans->journal_transaction_names) {
 760                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 761                                                        BCH_JSET_ENTRY_overwrite,
 762                                                        i->btree_id, i->level,
 763                                                        i->old_k.u64s);
 764                                 bkey_reassemble((struct bkey_i *) entry->start,
 765                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 766                         }
 767
 768                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 769                                                BCH_JSET_ENTRY_btree_keys,
 770                                                i->btree_id, i->level,
 771                                                i->k->k.u64s);
 772                         bkey_copy((struct bkey_i *) entry->start, i->k);
 773                 }
 774
 775                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 776                                   trans->journal_entries,
 777                                   trans->journal_entries_u64s);
 778
 779                 trans->journal_res.offset       += trans->journal_entries_u64s;
 780                 trans->journal_res.u64s         -= trans->journal_entries_u64s;
 781
 782                 if (trans->journal_seq)
 783                         *trans->journal_seq = trans->journal_res.seq;
 784         }
 785
 786         trans_for_each_update(trans, i) {
 787                 struct btree_path *path = trans->paths + i->path;
 788
 789                 if (!i->cached)
 790                         bch2_btree_insert_key_leaf(trans, path, i->k, trans->journal_res.seq);
 791                 else if (!i->key_cache_already_flushed)
 792                         bch2_btree_insert_key_cached(trans, flags, i);
 793                 else
 794                         bch2_btree_key_cache_drop(trans, path);
 795         }
 796
 797         return 0;
 798 fatal_err:
 799         bch2_fs_fatal_error(c, "fatal error in transaction commit: %s", bch2_err_str(ret));
 800         percpu_down_read(&c->mark_lock);
 801 revert_fs_usage:
 802         for (struct jset_entry *entry2 = trans->journal_entries;
 803              entry2 != entry;
 804              entry2 = vstruct_next(entry2))
 805                 if (entry2->type == BCH_JSET_ENTRY_write_buffer_keys &&
 806                     entry2->start->k.type == KEY_TYPE_accounting)
 807                         bch2_accounting_trans_commit_revert(trans,
 808                                         bkey_i_to_accounting(entry2->start), flags);
 809         percpu_up_read(&c->mark_lock);
 810         return ret;
 811 }
 812
 813 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 814 {
 815         /*
 816          * Accounting keys aren't deduped in the journal: we have to compare
 817          * each individual update against what's in the btree to see if it has
 818          * been applied yet, and accounting updates also don't overwrite,
 819          * they're deltas that accumulate.
 820          */
 821         trans_for_each_update(trans, i)
 822                 if (i->k->k.type != KEY_TYPE_accounting)
 823                         bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 824 }
 825
 826 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 827                                 struct journal_entry_pin *_pin, u64 seq)
 828 {
 829         return 0;
 830 }
 831
 832 /*
 833  * Get journal reservation, take write locks, and attempt to do btree update(s):
 834  */
 835 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 836                                        struct btree_insert_entry **stopped_at,
 837                                        unsigned long trace_ip)
 838 {
 839         struct bch_fs *c = trans->c;
 840         int ret = 0, u64s_delta = 0;
 841
 842         for (unsigned idx = 0; idx < trans->nr_updates; idx++) {
 843                 struct btree_insert_entry *i = trans->updates + idx;
 844                 if (i->cached)
 845                         continue;
 846
 847                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 848                 u64s_delta -= i->old_btree_u64s;
 849
 850                 if (!same_leaf_as_next(trans, i)) {
 851                         if (u64s_delta <= 0) {
 852                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 853                                                         i->level, flags);
 854                                 if (unlikely(ret))
 855                                         return ret;
 856                         }
 857
 858                         u64s_delta = 0;
 859                 }
 860         }
 861
 862         ret = bch2_trans_lock_write(trans);
 863         if (unlikely(ret))
 864                 return ret;
 865
 866         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 867
 868         if (!ret && unlikely(trans->journal_replay_not_finished))
 869                 bch2_drop_overwrites_from_journal(trans);
 870
 871         bch2_trans_unlock_updates_write(trans);
 872
 873         if (!ret && trans->journal_pin)
 874                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 875                                      trans->journal_pin,
 876                                      bch2_trans_commit_journal_pin_flush);
 877
 878         /*
 879          * Drop journal reservation after dropping write locks, since dropping
 880          * the journal reservation may kick off a journal write:
 881          */
 882         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 883                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 884
 885         return ret;
 886 }
 887
 888 static int journal_reclaim_wait_done(struct bch_fs *c)
 889 {
 890         int ret = bch2_journal_error(&c->journal) ?:
 891                 bch2_btree_key_cache_wait_done(c);
 892
 893         if (!ret)
 894                 journal_reclaim_kick(&c->journal);
 895         return ret;
 896 }
 897
 898 static noinline
 899 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 900                             struct btree_insert_entry *i,
 901                             int ret, unsigned long trace_ip)
 902 {
 903         struct bch_fs *c = trans->c;
 904         enum bch_watermark watermark = flags & BCH_WATERMARK_MASK;
 905
 906         switch (ret) {
 907         case -BCH_ERR_btree_insert_btree_node_full:
 908                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 909                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 910                         trace_and_count(c, trans_restart_btree_node_split, trans,
 911                                         trace_ip, trans->paths + i->path);
 912                 break;
 913         case -BCH_ERR_btree_insert_need_mark_replicas:
 914                 ret = drop_locks_do(trans,
 915                         bch2_accounting_update_sb(trans));
 916                 break;
 917         case -BCH_ERR_journal_res_get_blocked:
 918                 /*
 919                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 920                  * flag
 921                  */
 922                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 923                     watermark < BCH_WATERMARK_reclaim) {
 924                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 925                         break;
 926                 }
 927
 928                 ret = drop_locks_do(trans,
 929                         bch2_trans_journal_res_get(trans,
 930                                         (flags & BCH_WATERMARK_MASK)|
 931                                         JOURNAL_RES_GET_CHECK));
 932                 break;
 933         case -BCH_ERR_btree_insert_need_journal_reclaim:
 934                 bch2_trans_unlock(trans);
 935
 936                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 937                 track_event_change(&c->times[BCH_TIME_blocked_key_cache_flush], true);
 938
 939                 wait_event_freezable(c->journal.reclaim_wait,
 940                                      (ret = journal_reclaim_wait_done(c)));
 941
 942                 track_event_change(&c->times[BCH_TIME_blocked_key_cache_flush], false);
 943
 944                 if (ret < 0)
 945                         break;
 946
 947                 ret = bch2_trans_relock(trans);
 948                 break;
 949         default:
 950                 BUG_ON(ret >= 0);
 951                 break;
 952         }
 953
 954         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 955
 956         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 957                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 958                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 959
 960         return ret;
 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
 973         BUG_ON(current != c->recovery_task);
 974
 975         trans_for_each_update(trans, i) {
 976                 int ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 977                 if (ret)
 978                         return ret;
 979         }
 980
 981         for (struct jset_entry *i = trans->journal_entries;
 982              i != (void *) ((u64 *) trans->journal_entries + trans->journal_entries_u64s);
 983              i = vstruct_next(i))
 984                 if (i->type == BCH_JSET_ENTRY_btree_keys ||
 985                     i->type == BCH_JSET_ENTRY_write_buffer_keys) {
 986                         int ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->start);
 987                         if (ret)
 988                                 return ret;
 989                 }
 990
 991         return 0;
 992 }
 993
 994 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 995 {
 996         struct btree_insert_entry *errored_at = NULL;
 997         struct bch_fs *c = trans->c;
 998         int ret = 0;
 999
1000         bch2_trans_verify_not_unlocked_or_in_restart(trans);
1001
1002         ret = trans_maybe_inject_restart(trans, _RET_IP_);
1003         if (unlikely(ret))
1004                 goto out_reset;
1005
1006         if (!trans->nr_updates &&
1007             !trans->journal_entries_u64s)
1008                 goto out_reset;
1009
1010         ret = bch2_trans_commit_run_triggers(trans);
1011         if (ret)
1012                 goto out_reset;
1013
1014         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1015             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1016                 if (unlikely(!test_bit(BCH_FS_may_go_rw, &c->flags)))
1017                         ret = do_bch2_trans_commit_to_journal_replay(trans);
1018                 else
1019                         ret = -BCH_ERR_erofs_trans_commit;
1020                 goto out_reset;
1021         }
1022
1023         EBUG_ON(test_bit(BCH_FS_clean_shutdown, &c->flags));
1024
1025         trans->journal_u64s             = trans->journal_entries_u64s;
1026         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1027         if (trans->journal_transaction_names)
1028                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1029
1030         trans_for_each_update(trans, i) {
1031                 struct btree_path *path = trans->paths + i->path;
1032
1033                 EBUG_ON(!path->should_be_locked);
1034
1035                 ret = bch2_btree_path_upgrade(trans, path, i->level + 1);
1036                 if (unlikely(ret))
1037                         goto out;
1038
1039                 EBUG_ON(!btree_node_intent_locked(path, i->level));
1040
1041                 if (i->key_cache_already_flushed)
1042                         continue;
1043
1044                 if (i->flags & BTREE_UPDATE_nojournal)
1045                         continue;
1046
1047                 /* we're going to journal the key being updated: */
1048                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1049
1050                 /* and we're also going to log the overwrite: */
1051                 if (trans->journal_transaction_names)
1052                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1053         }
1054
1055         if (trans->extra_disk_res) {
1056                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1057                                 trans->extra_disk_res,
1058                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1059                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1060                 if (ret)
1061                         goto err;
1062         }
1063 retry:
1064         errored_at = NULL;
1065         bch2_trans_verify_not_unlocked_or_in_restart(trans);
1066         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1067                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1068         memset(&trans->fs_usage_delta, 0, sizeof(trans->fs_usage_delta));
1069
1070         ret = do_bch2_trans_commit(trans, flags, &errored_at, _RET_IP_);
1071
1072         /* make sure we didn't drop or screw up locks: */
1073         bch2_trans_verify_locks(trans);
1074
1075         if (ret)
1076                 goto err;
1077
1078         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1079 out:
1080         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1081                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1082 out_reset:
1083         if (!ret)
1084                 bch2_trans_downgrade(trans);
1085         bch2_trans_reset_updates(trans);
1086
1087         return ret;
1088 err:
1089         ret = bch2_trans_commit_error(trans, flags, errored_at, ret, _RET_IP_);
1090         if (ret)
1091                 goto out;
1092
1093         /*
1094          * We might have done another transaction commit in the error path -
1095          * i.e. btree write buffer flush - which will have made use of
1096          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1097          * how the journal sequence number to pin is passed in - so we must
1098          * restart:
1099          */
1100         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1101                 ret = -BCH_ERR_transaction_restart_nested;
1102                 goto out;
1103         }
1104
1105         goto retry;
1106 }