git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_trans_commit.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_gc.h"
   5 #include "btree_io.h"
   6 #include "btree_iter.h"
   7 #include "btree_journal_iter.h"
   8 #include "btree_key_cache.h"
   9 #include "btree_update_interior.h"
  10 #include "btree_write_buffer.h"
  11 #include "buckets.h"
  12 #include "errcode.h"
  13 #include "error.h"
  14 #include "journal.h"
  15 #include "journal_reclaim.h"
  16 #include "replicas.h"
  17 #include "snapshot.h"
  18
  19 #include <linux/prefetch.h>
  20
  21 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  22 {
  23 #ifdef CONFIG_BCACHEFS_DEBUG
  24         struct bch_fs *c = trans->c;
  25         struct bkey u;
  26         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
  27
  28         if (unlikely(trans->journal_replay_not_finished)) {
  29                 struct bkey_i *j_k =
  30                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  31
  32                 if (j_k)
  33                         k = bkey_i_to_s_c(j_k);
  34         }
  35
  36         u = *k.k;
  37         u.needs_whiteout = i->old_k.needs_whiteout;
  38
  39         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  40         BUG_ON(i->old_v != k.v);
  41 #endif
  42 }
  43
  44 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  45 {
  46         return i->path->l + i->level;
  47 }
  48
  49 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  50                                      struct btree_insert_entry *i)
  51 {
  52         return i != trans->updates &&
  53                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  54 }
  55
  56 static inline bool same_leaf_as_next(struct btree_trans *trans,
  57                                      struct btree_insert_entry *i)
  58 {
  59         return i + 1 < trans->updates + trans->nr_updates &&
  60                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  61 }
  62
  63 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  64                                            struct btree_path *path,
  65                                            struct btree *b)
  66 {
  67         struct bch_fs *c = trans->c;
  68
  69         if (unlikely(btree_node_just_written(b)) &&
  70             bch2_btree_post_write_cleanup(c, b))
  71                 bch2_trans_node_reinit_iter(trans, b);
  72
  73         /*
  74          * If the last bset has been written, or if it's gotten too big - start
  75          * a new bset to insert into:
  76          */
  77         if (want_new_bset(c, b))
  78                 bch2_btree_init_next(trans, b);
  79 }
  80
  81 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
  82 {
  83         while (--i >= trans->updates) {
  84                 if (same_leaf_as_prev(trans, i))
  85                         continue;
  86
  87                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
  88         }
  89
  90         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
  91         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
  92 }
  93
  94 static inline int bch2_trans_lock_write(struct btree_trans *trans)
  95 {
  96         struct btree_insert_entry *i;
  97
  98         EBUG_ON(trans->write_locked);
  99
 100         trans_for_each_update(trans, i) {
 101                 if (same_leaf_as_prev(trans, i))
 102                         continue;
 103
 104                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 105                         return trans_lock_write_fail(trans, i);
 106
 107                 if (!i->cached)
 108                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 109         }
 110
 111         trans->write_locked = true;
 112         return 0;
 113 }
 114
 115 static inline void bch2_trans_unlock_write(struct btree_trans *trans)
 116 {
 117         if (likely(trans->write_locked)) {
 118                 struct btree_insert_entry *i;
 119
 120                 trans_for_each_update(trans, i)
 121                         if (!same_leaf_as_prev(trans, i))
 122                                 bch2_btree_node_unlock_write_inlined(trans, i->path,
 123                                                                      insert_l(i)->b);
 124                 trans->write_locked = false;
 125         }
 126 }
 127
 128 /* Inserting into a given leaf node (last stage of insert): */
 129
 130 /* Handle overwrites and do insert, for non extents: */
 131 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 132                                 struct btree_path *path,
 133                                 struct btree *b,
 134                                 struct btree_node_iter *node_iter,
 135                                 struct bkey_i *insert)
 136 {
 137         struct bkey_packed *k;
 138         unsigned clobber_u64s = 0, new_u64s = 0;
 139
 140         EBUG_ON(btree_node_just_written(b));
 141         EBUG_ON(bset_written(b, btree_bset_last(b)));
 142         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 143         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 144         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 145         EBUG_ON(insert->k.u64s >
 146                 bch_btree_keys_u64s_remaining(trans->c, b));
 147         EBUG_ON(!b->c.level && !bpos_eq(insert->k.p, path->pos));
 148
 149         k = bch2_btree_node_iter_peek_all(node_iter, b);
 150         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 151                 k = NULL;
 152
 153         /* @k is the key being overwritten/deleted, if any: */
 154         EBUG_ON(k && bkey_deleted(k));
 155
 156         /* Deleting, but not found? nothing to do: */
 157         if (bkey_deleted(&insert->k) && !k)
 158                 return false;
 159
 160         if (bkey_deleted(&insert->k)) {
 161                 /* Deleting: */
 162                 btree_account_key_drop(b, k);
 163                 k->type = KEY_TYPE_deleted;
 164
 165                 if (k->needs_whiteout)
 166                         push_whiteout(trans->c, b, insert->k.p);
 167                 k->needs_whiteout = false;
 168
 169                 if (k >= btree_bset_last(b)->start) {
 170                         clobber_u64s = k->u64s;
 171                         bch2_bset_delete(b, k, clobber_u64s);
 172                         goto fix_iter;
 173                 } else {
 174                         bch2_btree_path_fix_key_modified(trans, b, k);
 175                 }
 176
 177                 return true;
 178         }
 179
 180         if (k) {
 181                 /* Overwriting: */
 182                 btree_account_key_drop(b, k);
 183                 k->type = KEY_TYPE_deleted;
 184
 185                 insert->k.needs_whiteout = k->needs_whiteout;
 186                 k->needs_whiteout = false;
 187
 188                 if (k >= btree_bset_last(b)->start) {
 189                         clobber_u64s = k->u64s;
 190                         goto overwrite;
 191                 } else {
 192                         bch2_btree_path_fix_key_modified(trans, b, k);
 193                 }
 194         }
 195
 196         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 197 overwrite:
 198         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 199         new_u64s = k->u64s;
 200 fix_iter:
 201         if (clobber_u64s != new_u64s)
 202                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 203                                          clobber_u64s, new_u64s);
 204         return true;
 205 }
 206
 207 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 208                                unsigned i, u64 seq)
 209 {
 210         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 211         struct btree_write *w = container_of(pin, struct btree_write, journal);
 212         struct btree *b = container_of(w, struct btree, writes[i]);
 213         struct btree_trans *trans = bch2_trans_get(c);
 214         unsigned long old, new, v;
 215         unsigned idx = w - b->writes;
 216
 217         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 218         v = READ_ONCE(b->flags);
 219
 220         do {
 221                 old = new = v;
 222
 223                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 224                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 225                     w->journal.seq != seq)
 226                         break;
 227
 228                 new &= ~BTREE_WRITE_TYPE_MASK;
 229                 new |= BTREE_WRITE_journal_reclaim;
 230                 new |= 1 << BTREE_NODE_need_write;
 231         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 232
 233         btree_node_write_if_need(c, b, SIX_LOCK_read);
 234         six_unlock_read(&b->c.lock);
 235
 236         bch2_trans_put(trans);
 237         return 0;
 238 }
 239
 240 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 241 {
 242         return __btree_node_flush(j, pin, 0, seq);
 243 }
 244
 245 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 246 {
 247         return __btree_node_flush(j, pin, 1, seq);
 248 }
 249
 250 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 251                                        struct btree *b, u64 seq)
 252 {
 253         struct btree_write *w = btree_current_write(b);
 254
 255         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 256                              btree_node_write_idx(b) == 0
 257                              ? bch2_btree_node_flush0
 258                              : bch2_btree_node_flush1);
 259 }
 260
 261 /**
 262  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 263  * @trans:              btree transaction object
 264  * @path:               path pointing to @insert's pos
 265  * @insert:             key to insert
 266  * @journal_seq:        sequence number of journal reservation
 267  */
 268 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 269                                        struct btree_path *path,
 270                                        struct bkey_i *insert,
 271                                        u64 journal_seq)
 272 {
 273         struct bch_fs *c = trans->c;
 274         struct btree *b = path_l(path)->b;
 275         struct bset_tree *t = bset_tree_last(b);
 276         struct bset *i = bset(b, t);
 277         int old_u64s = bset_u64s(t);
 278         int old_live_u64s = b->nr.live_u64s;
 279         int live_u64s_added, u64s_added;
 280
 281         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 282                                         &path_l(path)->iter, insert)))
 283                 return;
 284
 285         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 286
 287         bch2_btree_add_journal_pin(c, b, journal_seq);
 288
 289         if (unlikely(!btree_node_dirty(b))) {
 290                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 291                 set_btree_node_dirty_acct(c, b);
 292         }
 293
 294         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 295         u64s_added = (int) bset_u64s(t) - old_u64s;
 296
 297         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 298                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 299         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 300                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 301
 302         if (u64s_added > live_u64s_added &&
 303             bch2_maybe_compact_whiteouts(c, b))
 304                 bch2_trans_node_reinit_iter(trans, b);
 305 }
 306
 307 /* Cached btree updates: */
 308
 309 /* Normal update interface: */
 310
 311 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 312                                              struct btree_insert_entry *i)
 313 {
 314         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 315         BUG_ON(i->cached        != i->path->cached);
 316         BUG_ON(i->level         != i->path->level);
 317         BUG_ON(i->btree_id      != i->path->btree_id);
 318         EBUG_ON(!i->level &&
 319                 btree_type_has_snapshots(i->btree_id) &&
 320                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 321                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 322                 i->k->k.p.snapshot &&
 323                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 324 }
 325
 326 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 327                                                       unsigned flags)
 328 {
 329         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 330                                     trans->journal_u64s, flags);
 331 }
 332
 333 #define JSET_ENTRY_LOG_U64s             4
 334
 335 static noinline void journal_transaction_name(struct btree_trans *trans)
 336 {
 337         struct bch_fs *c = trans->c;
 338         struct journal *j = &c->journal;
 339         struct jset_entry *entry =
 340                 bch2_journal_add_entry(j, &trans->journal_res,
 341                                        BCH_JSET_ENTRY_log, 0, 0,
 342                                        JSET_ENTRY_LOG_U64s);
 343         struct jset_entry_log *l =
 344                 container_of(entry, struct jset_entry_log, entry);
 345
 346         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 347 }
 348
 349 static inline int btree_key_can_insert(struct btree_trans *trans,
 350                                        struct btree *b, unsigned u64s)
 351 {
 352         struct bch_fs *c = trans->c;
 353
 354         if (!bch2_btree_node_insert_fits(c, b, u64s))
 355                 return -BCH_ERR_btree_insert_btree_node_full;
 356
 357         return 0;
 358 }
 359
 360 noinline static int
 361 btree_key_can_insert_cached_slowpath(struct btree_trans *trans, unsigned flags,
 362                                      struct btree_path *path, unsigned new_u64s)
 363 {
 364         struct btree_insert_entry *i;
 365         struct bkey_cached *ck = (void *) path->l[0].b;
 366         struct bkey_i *new_k;
 367         int ret;
 368
 369         bch2_trans_unlock_write(trans);
 370         bch2_trans_unlock(trans);
 371
 372         new_k = kmalloc(new_u64s * sizeof(u64), GFP_KERNEL);
 373         if (!new_k) {
 374                 bch_err(trans->c, "error allocating memory for key cache key, btree %s u64s %u",
 375                         bch2_btree_id_str(path->btree_id), new_u64s);
 376                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 377         }
 378
 379         ret =   bch2_trans_relock(trans) ?:
 380                 bch2_trans_lock_write(trans);
 381         if (unlikely(ret)) {
 382                 kfree(new_k);
 383                 return ret;
 384         }
 385
 386         memcpy(new_k, ck->k, ck->u64s * sizeof(u64));
 387
 388         trans_for_each_update(trans, i)
 389                 if (i->old_v == &ck->k->v)
 390                         i->old_v = &new_k->v;
 391
 392         kfree(ck->k);
 393         ck->u64s        = new_u64s;
 394         ck->k           = new_k;
 395         return 0;
 396 }
 397
 398 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 399                                        struct btree_path *path, unsigned u64s)
 400 {
 401         struct bch_fs *c = trans->c;
 402         struct bkey_cached *ck = (void *) path->l[0].b;
 403         struct btree_insert_entry *i;
 404         unsigned new_u64s;
 405         struct bkey_i *new_k;
 406
 407         EBUG_ON(path->level);
 408
 409         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 410             bch2_btree_key_cache_must_wait(c) &&
 411             !(flags & BCH_TRANS_COMMIT_journal_reclaim))
 412                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 413
 414         /*
 415          * bch2_varint_decode can read past the end of the buffer by at most 7
 416          * bytes (it won't be used):
 417          */
 418         u64s += 1;
 419
 420         if (u64s <= ck->u64s)
 421                 return 0;
 422
 423         new_u64s        = roundup_pow_of_two(u64s);
 424         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOWAIT);
 425         if (unlikely(!new_k))
 426                 return btree_key_can_insert_cached_slowpath(trans, flags, path, new_u64s);
 427
 428         trans_for_each_update(trans, i)
 429                 if (i->old_v == &ck->k->v)
 430                         i->old_v = &new_k->v;
 431
 432         ck->u64s        = new_u64s;
 433         ck->k           = new_k;
 434         return 0;
 435 }
 436
 437 /* Triggers: */
 438
 439 static int run_one_mem_trigger(struct btree_trans *trans,
 440                                struct btree_insert_entry *i,
 441                                unsigned flags)
 442 {
 443         struct bkey_s_c old = { &i->old_k, i->old_v };
 444         struct bkey_i *new = i->k;
 445         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 446         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 447         int ret;
 448
 449         verify_update_old_key(trans, i);
 450
 451         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 452                 return 0;
 453
 454         if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
 455                 return 0;
 456
 457         if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
 458                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 459                                 old, bkey_i_to_s_c(new),
 460                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 461         } else {
 462                 struct bkey             _deleted = KEY(0, 0, 0);
 463                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 464
 465                 _deleted.p = i->path->pos;
 466
 467                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 468                                 deleted, bkey_i_to_s_c(new),
 469                                 BTREE_TRIGGER_INSERT|flags) ?:
 470                         bch2_mark_key(trans, i->btree_id, i->level,
 471                                 old, deleted,
 472                                 BTREE_TRIGGER_OVERWRITE|flags);
 473         }
 474
 475         return ret;
 476 }
 477
 478 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 479                                  bool overwrite)
 480 {
 481         /*
 482          * Transactional triggers create new btree_insert_entries, so we can't
 483          * pass them a pointer to a btree_insert_entry, that memory is going to
 484          * move:
 485          */
 486         struct bkey old_k = i->old_k;
 487         struct bkey_s_c old = { &old_k, i->old_v };
 488         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 489         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 490
 491         verify_update_old_key(trans, i);
 492
 493         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 494             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 495                 return 0;
 496
 497         if (!i->insert_trigger_run &&
 498             !i->overwrite_trigger_run &&
 499             old_ops->trans_trigger == new_ops->trans_trigger) {
 500                 i->overwrite_trigger_run = true;
 501                 i->insert_trigger_run = true;
 502                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 503                                            BTREE_TRIGGER_INSERT|
 504                                            BTREE_TRIGGER_OVERWRITE|
 505                                            i->flags) ?: 1;
 506         } else if (overwrite && !i->overwrite_trigger_run) {
 507                 i->overwrite_trigger_run = true;
 508                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 509         } else if (!overwrite && !i->insert_trigger_run) {
 510                 i->insert_trigger_run = true;
 511                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 512         } else {
 513                 return 0;
 514         }
 515 }
 516
 517 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 518                               struct btree_insert_entry *btree_id_start)
 519 {
 520         struct btree_insert_entry *i;
 521         bool trans_trigger_run;
 522         int ret, overwrite;
 523
 524         for (overwrite = 1; overwrite >= 0; --overwrite) {
 525
 526                 /*
 527                  * Running triggers will append more updates to the list of updates as
 528                  * we're walking it:
 529                  */
 530                 do {
 531                         trans_trigger_run = false;
 532
 533                         for (i = btree_id_start;
 534                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 535                              i++) {
 536                                 if (i->btree_id != btree_id)
 537                                         continue;
 538
 539                                 ret = run_one_trans_trigger(trans, i, overwrite);
 540                                 if (ret < 0)
 541                                         return ret;
 542                                 if (ret)
 543                                         trans_trigger_run = true;
 544                         }
 545                 } while (trans_trigger_run);
 546         }
 547
 548         return 0;
 549 }
 550
 551 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 552 {
 553         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 554         unsigned btree_id = 0;
 555         int ret = 0;
 556
 557         /*
 558          *
 559          * For a given btree, this algorithm runs insert triggers before
 560          * overwrite triggers: this is so that when extents are being moved
 561          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 562          * they are re-added.
 563          */
 564         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 565                 if (btree_id == BTREE_ID_alloc)
 566                         continue;
 567
 568                 while (btree_id_start < trans->updates + trans->nr_updates &&
 569                        btree_id_start->btree_id < btree_id)
 570                         btree_id_start++;
 571
 572                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 573                 if (ret)
 574                         return ret;
 575         }
 576
 577         trans_for_each_update(trans, i) {
 578                 if (i->btree_id > BTREE_ID_alloc)
 579                         break;
 580                 if (i->btree_id == BTREE_ID_alloc) {
 581                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 582                         if (ret)
 583                                 return ret;
 584                         break;
 585                 }
 586         }
 587
 588 #ifdef CONFIG_BCACHEFS_DEBUG
 589         trans_for_each_update(trans, i)
 590                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 591                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 592                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 593 #endif
 594         return 0;
 595 }
 596
 597 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 598 {
 599         struct bch_fs *c = trans->c;
 600         struct btree_insert_entry *i;
 601         int ret = 0;
 602
 603         trans_for_each_update(trans, i) {
 604                 /*
 605                  * XXX: synchronization of cached update triggers with gc
 606                  * XXX: synchronization of interior node updates with gc
 607                  */
 608                 BUG_ON(i->cached || i->level);
 609
 610                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 611                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 612                         if (ret)
 613                                 break;
 614                 }
 615         }
 616
 617         return ret;
 618 }
 619
 620 static inline int
 621 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 622                                struct btree_insert_entry **stopped_at,
 623                                unsigned long trace_ip)
 624 {
 625         struct bch_fs *c = trans->c;
 626         struct btree_insert_entry *i;
 627         struct btree_write_buffered_key *wb;
 628         struct btree_trans_commit_hook *h;
 629         unsigned u64s = 0;
 630         int ret;
 631
 632         if (race_fault()) {
 633                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 634                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 635         }
 636
 637         /*
 638          * Check if the insert will fit in the leaf node with the write lock
 639          * held, otherwise another thread could write the node changing the
 640          * amount of space available:
 641          */
 642
 643         prefetch(&trans->c->journal.flags);
 644
 645         trans_for_each_update(trans, i) {
 646                 /* Multiple inserts might go to same leaf: */
 647                 if (!same_leaf_as_prev(trans, i))
 648                         u64s = 0;
 649
 650                 u64s += i->k->k.u64s;
 651                 ret = !i->cached
 652                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 653                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 654                 if (ret) {
 655                         *stopped_at = i;
 656                         return ret;
 657                 }
 658
 659                 i->k->k.needs_whiteout = false;
 660         }
 661
 662         /*
 663          * Don't get journal reservation until after we know insert will
 664          * succeed:
 665          */
 666         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 667                 ret = bch2_trans_journal_res_get(trans,
 668                                 (flags & BCH_WATERMARK_MASK)|
 669                                 JOURNAL_RES_GET_NONBLOCK);
 670                 if (ret)
 671                         return ret;
 672
 673                 if (unlikely(trans->journal_transaction_names))
 674                         journal_transaction_name(trans);
 675         }
 676
 677         /*
 678          * Not allowed to fail after we've gotten our journal reservation - we
 679          * have to use it:
 680          */
 681
 682         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 683             !(flags & BCH_TRANS_COMMIT_no_journal_res)) {
 684                 if (bch2_journal_seq_verify)
 685                         trans_for_each_update(trans, i)
 686                                 i->k->k.version.lo = trans->journal_res.seq;
 687                 else if (bch2_inject_invalid_keys)
 688                         trans_for_each_update(trans, i)
 689                                 i->k->k.version = MAX_VERSION;
 690         }
 691
 692         if (trans->fs_usage_deltas &&
 693             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 694                 return -BCH_ERR_btree_insert_need_mark_replicas;
 695
 696         h = trans->hooks;
 697         while (h) {
 698                 ret = h->fn(trans, h);
 699                 if (ret)
 700                         goto revert_fs_usage;
 701                 h = h->next;
 702         }
 703
 704         trans_for_each_update(trans, i)
 705                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 706                         ret = run_one_mem_trigger(trans, i, i->flags);
 707                         if (ret)
 708                                 goto fatal_err;
 709                 }
 710
 711         if (unlikely(c->gc_pos.phase)) {
 712                 ret = bch2_trans_commit_run_gc_triggers(trans);
 713                 if  (ret)
 714                         goto fatal_err;
 715         }
 716
 717         if (unlikely(trans->extra_journal_entries.nr)) {
 718                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 719                                   trans->extra_journal_entries.data,
 720                                   trans->extra_journal_entries.nr);
 721
 722                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 723                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 724         }
 725
 726         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res))) {
 727                 struct journal *j = &c->journal;
 728                 struct jset_entry *entry;
 729
 730                 trans_for_each_update(trans, i) {
 731                         if (i->key_cache_already_flushed)
 732                                 continue;
 733
 734                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 735                                 continue;
 736
 737                         verify_update_old_key(trans, i);
 738
 739                         if (trans->journal_transaction_names) {
 740                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 741                                                        BCH_JSET_ENTRY_overwrite,
 742                                                        i->btree_id, i->level,
 743                                                        i->old_k.u64s);
 744                                 bkey_reassemble((struct bkey_i *) entry->start,
 745                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 746                         }
 747
 748                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 749                                                BCH_JSET_ENTRY_btree_keys,
 750                                                i->btree_id, i->level,
 751                                                i->k->k.u64s);
 752                         bkey_copy((struct bkey_i *) entry->start, i->k);
 753                 }
 754
 755                 trans_for_each_wb_update(trans, wb) {
 756                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 757                                                BCH_JSET_ENTRY_write_buffer_keys,
 758                                                wb->btree, 0,
 759                                                wb->k.k.u64s);
 760                         bkey_copy((struct bkey_i *) entry->start, &wb->k);
 761                 }
 762
 763                 if (trans->journal_seq)
 764                         *trans->journal_seq = trans->journal_res.seq;
 765         }
 766
 767         trans_for_each_update(trans, i) {
 768                 if (!i->cached) {
 769                         bch2_btree_insert_key_leaf(trans, i->path, i->k, trans->journal_res.seq);
 770                 } else if (!i->key_cache_already_flushed)
 771                         bch2_btree_insert_key_cached(trans, flags, i);
 772                 else {
 773                         bch2_btree_key_cache_drop(trans, i->path);
 774                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 775                 }
 776         }
 777
 778         return 0;
 779 fatal_err:
 780         bch2_fatal_error(c);
 781 revert_fs_usage:
 782         if (trans->fs_usage_deltas)
 783                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 784         return ret;
 785 }
 786
 787 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 788 {
 789         struct btree_insert_entry *i;
 790         struct btree_write_buffered_key *wb;
 791
 792         trans_for_each_update(trans, i)
 793                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 794
 795         trans_for_each_wb_update(trans, wb)
 796                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 797 }
 798
 799 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 800                                                    enum bkey_invalid_flags flags,
 801                                                    struct btree_insert_entry *i,
 802                                                    struct printbuf *err)
 803 {
 804         struct bch_fs *c = trans->c;
 805
 806         printbuf_reset(err);
 807         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 808                    trans->fn, (void *) i->ip_allocated);
 809         prt_newline(err);
 810         printbuf_indent_add(err, 2);
 811
 812         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 813         prt_newline(err);
 814
 815         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 816         bch2_print_string_as_lines(KERN_ERR, err->buf);
 817
 818         bch2_inconsistent_error(c);
 819         bch2_dump_trans_updates(trans);
 820
 821         return -EINVAL;
 822 }
 823
 824 static int bch2_trans_commit_journal_pin_flush(struct journal *j,
 825                                 struct journal_entry_pin *_pin, u64 seq)
 826 {
 827         return 0;
 828 }
 829
 830 /*
 831  * Get journal reservation, take write locks, and attempt to do btree update(s):
 832  */
 833 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 834                                        struct btree_insert_entry **stopped_at,
 835                                        unsigned long trace_ip)
 836 {
 837         struct bch_fs *c = trans->c;
 838         struct btree_insert_entry *i;
 839         int ret = 0, u64s_delta = 0;
 840
 841         trans_for_each_update(trans, i) {
 842                 if (i->cached)
 843                         continue;
 844
 845                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 846                 u64s_delta -= i->old_btree_u64s;
 847
 848                 if (!same_leaf_as_next(trans, i)) {
 849                         if (u64s_delta <= 0) {
 850                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 851                                                         i->level, flags);
 852                                 if (unlikely(ret))
 853                                         return ret;
 854                         }
 855
 856                         u64s_delta = 0;
 857                 }
 858         }
 859
 860         ret = bch2_trans_lock_write(trans);
 861         if (unlikely(ret))
 862                 return ret;
 863
 864         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 865
 866         if (!ret && unlikely(trans->journal_replay_not_finished))
 867                 bch2_drop_overwrites_from_journal(trans);
 868
 869         bch2_trans_unlock_write(trans);
 870
 871         if (!ret && trans->journal_pin)
 872                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 873                                      trans->journal_pin,
 874                                      bch2_trans_commit_journal_pin_flush);
 875
 876         /*
 877          * Drop journal reservation after dropping write locks, since dropping
 878          * the journal reservation may kick off a journal write:
 879          */
 880         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
 881                 bch2_journal_res_put(&c->journal, &trans->journal_res);
 882
 883         return ret;
 884 }
 885
 886 static int journal_reclaim_wait_done(struct bch_fs *c)
 887 {
 888         int ret = bch2_journal_error(&c->journal) ?:
 889                 !bch2_btree_key_cache_must_wait(c);
 890
 891         if (!ret)
 892                 journal_reclaim_kick(&c->journal);
 893         return ret;
 894 }
 895
 896 static noinline
 897 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 898                             struct btree_insert_entry *i,
 899                             int ret, unsigned long trace_ip)
 900 {
 901         struct bch_fs *c = trans->c;
 902
 903         switch (ret) {
 904         case -BCH_ERR_btree_insert_btree_node_full:
 905                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 906                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 907                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 908                 break;
 909         case -BCH_ERR_btree_insert_need_mark_replicas:
 910                 ret = drop_locks_do(trans,
 911                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 912                 break;
 913         case -BCH_ERR_journal_res_get_blocked:
 914                 /*
 915                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 916                  * flag
 917                  */
 918                 if ((flags & BCH_TRANS_COMMIT_journal_reclaim) &&
 919                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 920                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 921                         break;
 922                 }
 923
 924                 ret = drop_locks_do(trans,
 925                         bch2_trans_journal_res_get(trans,
 926                                         (flags & BCH_WATERMARK_MASK)|
 927                                         JOURNAL_RES_GET_CHECK));
 928                 break;
 929         case -BCH_ERR_btree_insert_need_journal_reclaim:
 930                 bch2_trans_unlock(trans);
 931
 932                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 933
 934                 wait_event_freezable(c->journal.reclaim_wait,
 935                                      (ret = journal_reclaim_wait_done(c)));
 936                 if (ret < 0)
 937                         break;
 938
 939                 ret = bch2_trans_relock(trans);
 940                 break;
 941         default:
 942                 BUG_ON(ret >= 0);
 943                 break;
 944         }
 945
 946         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 947
 948         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 949                                 (flags & BCH_TRANS_COMMIT_no_enospc), c,
 950                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 951
 952         return ret;
 953 }
 954
 955 static noinline int
 956 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 957 {
 958         struct bch_fs *c = trans->c;
 959         int ret;
 960
 961         if (likely(!(flags & BCH_TRANS_COMMIT_lazy_rw)) ||
 962             test_bit(BCH_FS_STARTED, &c->flags))
 963                 return -BCH_ERR_erofs_trans_commit;
 964
 965         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 966         if (ret)
 967                 return ret;
 968
 969         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 970         return 0;
 971 }
 972
 973 /*
 974  * This is for updates done in the early part of fsck - btree_gc - before we've
 975  * gone RW. we only add the new key to the list of keys for journal replay to
 976  * do.
 977  */
 978 static noinline int
 979 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 980 {
 981         struct bch_fs *c = trans->c;
 982         struct btree_insert_entry *i;
 983         int ret = 0;
 984
 985         trans_for_each_update(trans, i) {
 986                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 987                 if (ret)
 988                         break;
 989         }
 990
 991         return ret;
 992 }
 993
 994 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
 995 {
 996         struct bch_fs *c = trans->c;
 997         struct btree_insert_entry *i = NULL;
 998         struct btree_write_buffered_key *wb;
 999         int ret = 0;
1000
1001         if (!trans->nr_updates &&
1002             !trans->nr_wb_updates &&
1003             !trans->extra_journal_entries.nr)
1004                 goto out_reset;
1005
1006         ret = bch2_trans_commit_run_triggers(trans);
1007         if (ret)
1008                 goto out_reset;
1009
1010         trans_for_each_update(trans, i) {
1011                 struct printbuf buf = PRINTBUF;
1012                 enum bkey_invalid_flags invalid_flags = 0;
1013
1014                 if (!(flags & BCH_TRANS_COMMIT_no_journal_res))
1015                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1016
1017                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1018                                                i->bkey_type, invalid_flags, &buf)))
1019                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1020                 btree_insert_entry_checks(trans, i);
1021                 printbuf_exit(&buf);
1022
1023                 if (ret)
1024                         return ret;
1025         }
1026
1027         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1028                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1029                 goto out_reset;
1030         }
1031
1032         if (!(flags & BCH_TRANS_COMMIT_no_check_rw) &&
1033             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1034                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1035                 if (ret)
1036                         goto out_reset;
1037         }
1038
1039         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1040
1041         trans->journal_u64s             = trans->extra_journal_entries.nr;
1042         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1043         if (trans->journal_transaction_names)
1044                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1045
1046         trans_for_each_update(trans, i) {
1047                 EBUG_ON(!i->path->should_be_locked);
1048
1049                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1050                 if (unlikely(ret))
1051                         goto out;
1052
1053                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1054
1055                 if (i->key_cache_already_flushed)
1056                         continue;
1057
1058                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1059                         continue;
1060
1061                 /* we're going to journal the key being updated: */
1062                 trans->journal_u64s += jset_u64s(i->k->k.u64s);
1063
1064                 /* and we're also going to log the overwrite: */
1065                 if (trans->journal_transaction_names)
1066                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1067         }
1068
1069         trans_for_each_wb_update(trans, wb)
1070                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1071
1072         if (trans->extra_journal_res) {
1073                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1074                                 trans->extra_journal_res,
1075                                 (flags & BCH_TRANS_COMMIT_no_enospc)
1076                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1077                 if (ret)
1078                         goto err;
1079         }
1080 retry:
1081         bch2_trans_verify_not_in_restart(trans);
1082         if (likely(!(flags & BCH_TRANS_COMMIT_no_journal_res)))
1083                 memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1084
1085         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1086
1087         /* make sure we didn't drop or screw up locks: */
1088         bch2_trans_verify_locks(trans);
1089
1090         if (ret)
1091                 goto err;
1092
1093         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1094 out:
1095         if (likely(!(flags & BCH_TRANS_COMMIT_no_check_rw)))
1096                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1097 out_reset:
1098         if (!ret)
1099                 bch2_trans_downgrade(trans);
1100         bch2_trans_reset_updates(trans);
1101
1102         return ret;
1103 err:
1104         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1105         if (ret)
1106                 goto out;
1107
1108         /*
1109          * We might have done another transaction commit in the error path -
1110          * i.e. btree write buffer flush - which will have made use of
1111          * trans->journal_res, but with BCH_TRANS_COMMIT_no_journal_res that is
1112          * how the journal sequence number to pin is passed in - so we must
1113          * restart:
1114          */
1115         if (flags & BCH_TRANS_COMMIT_no_journal_res) {
1116                 ret = -BCH_ERR_transaction_restart_nested;
1117                 goto out;
1118         }
1119
1120         goto retry;
1121 }