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 "subvolume.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 /* Inserting into a given leaf node (last stage of insert): */
  82
  83 /* Handle overwrites and do insert, for non extents: */
  84 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  85                                 struct btree_path *path,
  86                                 struct btree *b,
  87                                 struct btree_node_iter *node_iter,
  88                                 struct bkey_i *insert)
  89 {
  90         struct bkey_packed *k;
  91         unsigned clobber_u64s = 0, new_u64s = 0;
  92
  93         EBUG_ON(btree_node_just_written(b));
  94         EBUG_ON(bset_written(b, btree_bset_last(b)));
  95         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  96         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
  97         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
  98         EBUG_ON(insert->k.u64s >
  99                 bch_btree_keys_u64s_remaining(trans->c, b));
 100
 101         k = bch2_btree_node_iter_peek_all(node_iter, b);
 102         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 103                 k = NULL;
 104
 105         /* @k is the key being overwritten/deleted, if any: */
 106         EBUG_ON(k && bkey_deleted(k));
 107
 108         /* Deleting, but not found? nothing to do: */
 109         if (bkey_deleted(&insert->k) && !k)
 110                 return false;
 111
 112         if (bkey_deleted(&insert->k)) {
 113                 /* Deleting: */
 114                 btree_account_key_drop(b, k);
 115                 k->type = KEY_TYPE_deleted;
 116
 117                 if (k->needs_whiteout)
 118                         push_whiteout(trans->c, b, insert->k.p);
 119                 k->needs_whiteout = false;
 120
 121                 if (k >= btree_bset_last(b)->start) {
 122                         clobber_u64s = k->u64s;
 123                         bch2_bset_delete(b, k, clobber_u64s);
 124                         goto fix_iter;
 125                 } else {
 126                         bch2_btree_path_fix_key_modified(trans, b, k);
 127                 }
 128
 129                 return true;
 130         }
 131
 132         if (k) {
 133                 /* Overwriting: */
 134                 btree_account_key_drop(b, k);
 135                 k->type = KEY_TYPE_deleted;
 136
 137                 insert->k.needs_whiteout = k->needs_whiteout;
 138                 k->needs_whiteout = false;
 139
 140                 if (k >= btree_bset_last(b)->start) {
 141                         clobber_u64s = k->u64s;
 142                         goto overwrite;
 143                 } else {
 144                         bch2_btree_path_fix_key_modified(trans, b, k);
 145                 }
 146         }
 147
 148         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 149 overwrite:
 150         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 151         new_u64s = k->u64s;
 152 fix_iter:
 153         if (clobber_u64s != new_u64s)
 154                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 155                                          clobber_u64s, new_u64s);
 156         return true;
 157 }
 158
 159 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 160                                unsigned i, u64 seq)
 161 {
 162         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 163         struct btree_write *w = container_of(pin, struct btree_write, journal);
 164         struct btree *b = container_of(w, struct btree, writes[i]);
 165         struct btree_trans trans;
 166         unsigned long old, new, v;
 167         unsigned idx = w - b->writes;
 168
 169         bch2_trans_init(&trans, c, 0, 0);
 170
 171         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 172         v = READ_ONCE(b->flags);
 173
 174         do {
 175                 old = new = v;
 176
 177                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 178                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 179                     w->journal.seq != seq)
 180                         break;
 181
 182                 new &= ~BTREE_WRITE_TYPE_MASK;
 183                 new |= BTREE_WRITE_journal_reclaim;
 184                 new |= 1 << BTREE_NODE_need_write;
 185         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 186
 187         btree_node_write_if_need(c, b, SIX_LOCK_read);
 188         six_unlock_read(&b->c.lock);
 189
 190         bch2_trans_exit(&trans);
 191         return 0;
 192 }
 193
 194 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 195 {
 196         return __btree_node_flush(j, pin, 0, seq);
 197 }
 198
 199 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 200 {
 201         return __btree_node_flush(j, pin, 1, seq);
 202 }
 203
 204 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 205                                        struct btree *b, u64 seq)
 206 {
 207         struct btree_write *w = btree_current_write(b);
 208
 209         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 210                              btree_node_write_idx(b) == 0
 211                              ? bch2_btree_node_flush0
 212                              : bch2_btree_node_flush1);
 213 }
 214
 215 /**
 216  * btree_insert_key - insert a key one key into a leaf node
 217  */
 218 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 219                                        struct btree_path *path,
 220                                        struct bkey_i *insert,
 221                                        u64 journal_seq)
 222 {
 223         struct bch_fs *c = trans->c;
 224         struct btree *b = path_l(path)->b;
 225         struct bset_tree *t = bset_tree_last(b);
 226         struct bset *i = bset(b, t);
 227         int old_u64s = bset_u64s(t);
 228         int old_live_u64s = b->nr.live_u64s;
 229         int live_u64s_added, u64s_added;
 230
 231         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 232                                         &path_l(path)->iter, insert)))
 233                 return;
 234
 235         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 236
 237         bch2_btree_add_journal_pin(c, b, journal_seq);
 238
 239         if (unlikely(!btree_node_dirty(b))) {
 240                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 241                 set_btree_node_dirty_acct(c, b);
 242         }
 243
 244         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 245         u64s_added = (int) bset_u64s(t) - old_u64s;
 246
 247         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 248                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 249         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 250                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 251
 252         if (u64s_added > live_u64s_added &&
 253             bch2_maybe_compact_whiteouts(c, b))
 254                 bch2_trans_node_reinit_iter(trans, b);
 255 }
 256
 257 /* Cached btree updates: */
 258
 259 /* Normal update interface: */
 260
 261 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 262                                              struct btree_insert_entry *i)
 263 {
 264         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 265         BUG_ON(i->cached        != i->path->cached);
 266         BUG_ON(i->level         != i->path->level);
 267         BUG_ON(i->btree_id      != i->path->btree_id);
 268         EBUG_ON(!i->level &&
 269                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 270                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 271                 i->k->k.p.snapshot &&
 272                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 273 }
 274
 275 static noinline int
 276 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
 277                                    unsigned long trace_ip)
 278 {
 279         return drop_locks_do(trans,
 280                 bch2_journal_preres_get(&trans->c->journal,
 281                         &trans->journal_preres,
 282                         trans->journal_preres_u64s,
 283                         (flags & BCH_WATERMARK_MASK)));
 284 }
 285
 286 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 287                                                       unsigned flags)
 288 {
 289         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 290                                     trans->journal_u64s, flags);
 291 }
 292
 293 #define JSET_ENTRY_LOG_U64s             4
 294
 295 static noinline void journal_transaction_name(struct btree_trans *trans)
 296 {
 297         struct bch_fs *c = trans->c;
 298         struct journal *j = &c->journal;
 299         struct jset_entry *entry =
 300                 bch2_journal_add_entry(j, &trans->journal_res,
 301                                        BCH_JSET_ENTRY_log, 0, 0,
 302                                        JSET_ENTRY_LOG_U64s);
 303         struct jset_entry_log *l =
 304                 container_of(entry, struct jset_entry_log, entry);
 305
 306         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 307 }
 308
 309 static inline int btree_key_can_insert(struct btree_trans *trans,
 310                                        struct btree *b, unsigned u64s)
 311 {
 312         struct bch_fs *c = trans->c;
 313
 314         if (!bch2_btree_node_insert_fits(c, b, u64s))
 315                 return -BCH_ERR_btree_insert_btree_node_full;
 316
 317         return 0;
 318 }
 319
 320 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 321                                        struct btree_path *path, unsigned u64s)
 322 {
 323         struct bch_fs *c = trans->c;
 324         struct bkey_cached *ck = (void *) path->l[0].b;
 325         struct btree_insert_entry *i;
 326         unsigned new_u64s;
 327         struct bkey_i *new_k;
 328
 329         EBUG_ON(path->level);
 330
 331         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 332             bch2_btree_key_cache_must_wait(c) &&
 333             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 334                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 335
 336         /*
 337          * bch2_varint_decode can read past the end of the buffer by at most 7
 338          * bytes (it won't be used):
 339          */
 340         u64s += 1;
 341
 342         if (u64s <= ck->u64s)
 343                 return 0;
 344
 345         new_u64s        = roundup_pow_of_two(u64s);
 346         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 347         if (!new_k) {
 348                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 349                         bch2_btree_ids[path->btree_id], new_u64s);
 350                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 351         }
 352
 353         trans_for_each_update(trans, i)
 354                 if (i->old_v == &ck->k->v)
 355                         i->old_v = &new_k->v;
 356
 357         ck->u64s        = new_u64s;
 358         ck->k           = new_k;
 359         return 0;
 360 }
 361
 362 /* Triggers: */
 363
 364 static int run_one_mem_trigger(struct btree_trans *trans,
 365                                struct btree_insert_entry *i,
 366                                unsigned flags)
 367 {
 368         struct bkey_s_c old = { &i->old_k, i->old_v };
 369         struct bkey_i *new = i->k;
 370         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 371         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 372         int ret;
 373
 374         verify_update_old_key(trans, i);
 375
 376         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 377                 return 0;
 378
 379         if (!btree_node_type_needs_gc((enum btree_node_type) i->btree_id))
 380                 return 0;
 381
 382         if (old_ops->atomic_trigger == new_ops->atomic_trigger &&
 383             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 384                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 385                                 old, bkey_i_to_s_c(new),
 386                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 387         } else {
 388                 struct bkey             _deleted = KEY(0, 0, 0);
 389                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 390
 391                 _deleted.p = i->path->pos;
 392
 393                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 394                                 deleted, bkey_i_to_s_c(new),
 395                                 BTREE_TRIGGER_INSERT|flags) ?:
 396                         bch2_mark_key(trans, i->btree_id, i->level,
 397                                 old, deleted,
 398                                 BTREE_TRIGGER_OVERWRITE|flags);
 399         }
 400
 401         return ret;
 402 }
 403
 404 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 405                                  bool overwrite)
 406 {
 407         /*
 408          * Transactional triggers create new btree_insert_entries, so we can't
 409          * pass them a pointer to a btree_insert_entry, that memory is going to
 410          * move:
 411          */
 412         struct bkey old_k = i->old_k;
 413         struct bkey_s_c old = { &old_k, i->old_v };
 414         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 415         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 416
 417         verify_update_old_key(trans, i);
 418
 419         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 420             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 421                 return 0;
 422
 423         if (!i->insert_trigger_run &&
 424             !i->overwrite_trigger_run &&
 425             old_ops->trans_trigger == new_ops->trans_trigger &&
 426             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 427                 i->overwrite_trigger_run = true;
 428                 i->insert_trigger_run = true;
 429                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 430                                            BTREE_TRIGGER_INSERT|
 431                                            BTREE_TRIGGER_OVERWRITE|
 432                                            i->flags) ?: 1;
 433         } else if (overwrite && !i->overwrite_trigger_run) {
 434                 i->overwrite_trigger_run = true;
 435                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 436         } else if (!overwrite && !i->insert_trigger_run) {
 437                 i->insert_trigger_run = true;
 438                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 439         } else {
 440                 return 0;
 441         }
 442 }
 443
 444 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 445                               struct btree_insert_entry *btree_id_start)
 446 {
 447         struct btree_insert_entry *i;
 448         bool trans_trigger_run;
 449         int ret, overwrite;
 450
 451         for (overwrite = 1; overwrite >= 0; --overwrite) {
 452
 453                 /*
 454                  * Running triggers will append more updates to the list of updates as
 455                  * we're walking it:
 456                  */
 457                 do {
 458                         trans_trigger_run = false;
 459
 460                         for (i = btree_id_start;
 461                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 462                              i++) {
 463                                 if (i->btree_id != btree_id)
 464                                         continue;
 465
 466                                 ret = run_one_trans_trigger(trans, i, overwrite);
 467                                 if (ret < 0)
 468                                         return ret;
 469                                 if (ret)
 470                                         trans_trigger_run = true;
 471                         }
 472                 } while (trans_trigger_run);
 473         }
 474
 475         return 0;
 476 }
 477
 478 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 479 {
 480         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 481         unsigned btree_id = 0;
 482         int ret = 0;
 483
 484         /*
 485          *
 486          * For a given btree, this algorithm runs insert triggers before
 487          * overwrite triggers: this is so that when extents are being moved
 488          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 489          * they are re-added.
 490          */
 491         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 492                 if (btree_id == BTREE_ID_alloc)
 493                         continue;
 494
 495                 while (btree_id_start < trans->updates + trans->nr_updates &&
 496                        btree_id_start->btree_id < btree_id)
 497                         btree_id_start++;
 498
 499                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 500                 if (ret)
 501                         return ret;
 502         }
 503
 504         trans_for_each_update(trans, i) {
 505                 if (i->btree_id > BTREE_ID_alloc)
 506                         break;
 507                 if (i->btree_id == BTREE_ID_alloc) {
 508                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 509                         if (ret)
 510                                 return ret;
 511                         break;
 512                 }
 513         }
 514
 515 #ifdef CONFIG_BCACHEFS_DEBUG
 516         trans_for_each_update(trans, i)
 517                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 518                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 519                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 520 #endif
 521         return 0;
 522 }
 523
 524 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 525 {
 526         struct bch_fs *c = trans->c;
 527         struct btree_insert_entry *i;
 528         int ret = 0;
 529
 530         trans_for_each_update(trans, i) {
 531                 /*
 532                  * XXX: synchronization of cached update triggers with gc
 533                  * XXX: synchronization of interior node updates with gc
 534                  */
 535                 BUG_ON(i->cached || i->level);
 536
 537                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 538                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 539                         if (ret)
 540                                 break;
 541                 }
 542         }
 543
 544         return ret;
 545 }
 546
 547 static inline int
 548 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 549                                struct btree_insert_entry **stopped_at,
 550                                unsigned long trace_ip)
 551 {
 552         struct bch_fs *c = trans->c;
 553         struct btree_insert_entry *i;
 554         struct btree_write_buffered_key *wb;
 555         struct btree_trans_commit_hook *h;
 556         unsigned u64s = 0;
 557         bool marking = false;
 558         int ret;
 559
 560         if (race_fault()) {
 561                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 562                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 563         }
 564
 565         /*
 566          * Check if the insert will fit in the leaf node with the write lock
 567          * held, otherwise another thread could write the node changing the
 568          * amount of space available:
 569          */
 570
 571         prefetch(&trans->c->journal.flags);
 572
 573         trans_for_each_update(trans, i) {
 574                 /* Multiple inserts might go to same leaf: */
 575                 if (!same_leaf_as_prev(trans, i))
 576                         u64s = 0;
 577
 578                 u64s += i->k->k.u64s;
 579                 ret = !i->cached
 580                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 581                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 582                 if (ret) {
 583                         *stopped_at = i;
 584                         return ret;
 585                 }
 586
 587                 if (btree_node_type_needs_gc(i->bkey_type))
 588                         marking = true;
 589         }
 590
 591         if (trans->nr_wb_updates &&
 592             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 593                 return -BCH_ERR_btree_insert_need_flush_buffer;
 594
 595         /*
 596          * Don't get journal reservation until after we know insert will
 597          * succeed:
 598          */
 599         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 600                 ret = bch2_trans_journal_res_get(trans,
 601                                 (flags & BCH_WATERMARK_MASK)|
 602                                 JOURNAL_RES_GET_NONBLOCK);
 603                 if (ret)
 604                         return ret;
 605
 606                 if (unlikely(trans->journal_transaction_names))
 607                         journal_transaction_name(trans);
 608         } else {
 609                 trans->journal_res.seq = c->journal.replay_journal_seq;
 610         }
 611
 612         /*
 613          * Not allowed to fail after we've gotten our journal reservation - we
 614          * have to use it:
 615          */
 616
 617         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 618             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 619                 if (bch2_journal_seq_verify)
 620                         trans_for_each_update(trans, i)
 621                                 i->k->k.version.lo = trans->journal_res.seq;
 622                 else if (bch2_inject_invalid_keys)
 623                         trans_for_each_update(trans, i)
 624                                 i->k->k.version = MAX_VERSION;
 625         }
 626
 627         if (trans->fs_usage_deltas &&
 628             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 629                 return -BCH_ERR_btree_insert_need_mark_replicas;
 630
 631         if (trans->nr_wb_updates) {
 632                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 633
 634                 ret = bch2_btree_insert_keys_write_buffer(trans);
 635                 if (ret)
 636                         goto revert_fs_usage;
 637         }
 638
 639         h = trans->hooks;
 640         while (h) {
 641                 ret = h->fn(trans, h);
 642                 if (ret)
 643                         goto revert_fs_usage;
 644                 h = h->next;
 645         }
 646
 647         trans_for_each_update(trans, i)
 648                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 649                         ret = run_one_mem_trigger(trans, i, i->flags);
 650                         if (ret)
 651                                 goto fatal_err;
 652                 }
 653
 654         if (unlikely(c->gc_pos.phase)) {
 655                 ret = bch2_trans_commit_run_gc_triggers(trans);
 656                 if  (ret)
 657                         goto fatal_err;
 658         }
 659
 660         if (unlikely(trans->extra_journal_entries.nr)) {
 661                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 662                                   trans->extra_journal_entries.data,
 663                                   trans->extra_journal_entries.nr);
 664
 665                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 666                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 667         }
 668
 669         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 670                 struct journal *j = &c->journal;
 671                 struct jset_entry *entry;
 672
 673                 trans_for_each_update(trans, i) {
 674                         if (i->key_cache_already_flushed)
 675                                 continue;
 676
 677                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 678                                 continue;
 679
 680                         verify_update_old_key(trans, i);
 681
 682                         if (trans->journal_transaction_names) {
 683                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 684                                                        BCH_JSET_ENTRY_overwrite,
 685                                                        i->btree_id, i->level,
 686                                                        i->old_k.u64s);
 687                                 bkey_reassemble(&entry->start[0],
 688                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 689                         }
 690
 691                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 692                                                BCH_JSET_ENTRY_btree_keys,
 693                                                i->btree_id, i->level,
 694                                                i->k->k.u64s);
 695                         bkey_copy(&entry->start[0], i->k);
 696                 }
 697
 698                 trans_for_each_wb_update(trans, wb) {
 699                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 700                                                BCH_JSET_ENTRY_btree_keys,
 701                                                wb->btree, 0,
 702                                                wb->k.k.u64s);
 703                         bkey_copy(&entry->start[0], &wb->k);
 704                 }
 705
 706                 if (trans->journal_seq)
 707                         *trans->journal_seq = trans->journal_res.seq;
 708         }
 709
 710         trans_for_each_update(trans, i) {
 711                 i->k->k.needs_whiteout = false;
 712
 713                 if (!i->cached) {
 714                         u64 seq = trans->journal_res.seq;
 715
 716                         if (i->flags & BTREE_UPDATE_PREJOURNAL)
 717                                 seq = i->seq;
 718
 719                         bch2_btree_insert_key_leaf(trans, i->path, i->k, seq);
 720                 } else if (!i->key_cache_already_flushed)
 721                         bch2_btree_insert_key_cached(trans, flags, i);
 722                 else {
 723                         bch2_btree_key_cache_drop(trans, i->path);
 724                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 725                 }
 726         }
 727
 728         return 0;
 729 fatal_err:
 730         bch2_fatal_error(c);
 731 revert_fs_usage:
 732         if (trans->fs_usage_deltas)
 733                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 734         return ret;
 735 }
 736
 737 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 738 {
 739         while (--i >= trans->updates) {
 740                 if (same_leaf_as_prev(trans, i))
 741                         continue;
 742
 743                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 744         }
 745
 746         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 747         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 748 }
 749
 750 static inline int trans_lock_write(struct btree_trans *trans)
 751 {
 752         struct btree_insert_entry *i;
 753
 754         trans_for_each_update(trans, i) {
 755                 if (same_leaf_as_prev(trans, i))
 756                         continue;
 757
 758                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 759                         return trans_lock_write_fail(trans, i);
 760
 761                 if (!i->cached)
 762                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 763         }
 764
 765         return 0;
 766 }
 767
 768 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 769 {
 770         struct btree_insert_entry *i;
 771         struct btree_write_buffered_key *wb;
 772
 773         trans_for_each_update(trans, i)
 774                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 775
 776         trans_for_each_wb_update(trans, wb)
 777                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 778 }
 779
 780 #ifdef CONFIG_BCACHEFS_DEBUG
 781 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 782                                                    struct btree_insert_entry *i,
 783                                                    struct printbuf *err)
 784 {
 785         struct bch_fs *c = trans->c;
 786         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 787
 788         printbuf_reset(err);
 789         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 790                    trans->fn, (void *) i->ip_allocated);
 791         prt_newline(err);
 792         printbuf_indent_add(err, 2);
 793
 794         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 795         prt_newline(err);
 796
 797         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 798                           i->bkey_type, rw, err);
 799         bch2_print_string_as_lines(KERN_ERR, err->buf);
 800
 801         bch2_inconsistent_error(c);
 802         bch2_dump_trans_updates(trans);
 803         printbuf_exit(err);
 804
 805         return -EINVAL;
 806 }
 807 #endif
 808
 809 /*
 810  * Get journal reservation, take write locks, and attempt to do btree update(s):
 811  */
 812 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 813                                        struct btree_insert_entry **stopped_at,
 814                                        unsigned long trace_ip)
 815 {
 816         struct bch_fs *c = trans->c;
 817         struct btree_insert_entry *i;
 818         int ret = 0, u64s_delta = 0;
 819
 820 #ifdef CONFIG_BCACHEFS_DEBUG
 821         trans_for_each_update(trans, i) {
 822                 struct printbuf buf = PRINTBUF;
 823                 enum bkey_invalid_flags invalid_flags = 0;
 824
 825                 if (!(flags & BTREE_INSERT_JOURNAL_REPLAY))
 826                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
 827
 828                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 829                                                i->bkey_type, invalid_flags, &buf)))
 830                         ret = bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 831                 btree_insert_entry_checks(trans, i);
 832                 printbuf_exit(&buf);
 833
 834                 if (ret)
 835                         return ret;
 836         }
 837 #endif
 838
 839         trans_for_each_update(trans, i) {
 840                 if (i->cached)
 841                         continue;
 842
 843                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 844                 u64s_delta -= i->old_btree_u64s;
 845
 846                 if (!same_leaf_as_next(trans, i)) {
 847                         if (u64s_delta <= 0) {
 848                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 849                                                         i->level, flags);
 850                                 if (unlikely(ret))
 851                                         return ret;
 852                         }
 853
 854                         u64s_delta = 0;
 855                 }
 856         }
 857
 858         ret = bch2_journal_preres_get(&c->journal,
 859                         &trans->journal_preres, trans->journal_preres_u64s,
 860                         (flags & BCH_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 861         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 862                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 863         if (unlikely(ret))
 864                 return ret;
 865
 866         ret = trans_lock_write(trans);
 867         if (unlikely(ret))
 868                 return ret;
 869
 870         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 871
 872         if (!ret && unlikely(trans->journal_replay_not_finished))
 873                 bch2_drop_overwrites_from_journal(trans);
 874
 875         trans_for_each_update(trans, i)
 876                 if (!same_leaf_as_prev(trans, i))
 877                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 878                                                         insert_l(i)->b);
 879
 880         if (!ret && trans->journal_pin)
 881                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 882                                      trans->journal_pin, NULL);
 883
 884         /*
 885          * Drop journal reservation after dropping write locks, since dropping
 886          * the journal reservation may kick off a journal write:
 887          */
 888         bch2_journal_res_put(&c->journal, &trans->journal_res);
 889
 890         if (unlikely(ret))
 891                 return ret;
 892
 893         bch2_trans_downgrade(trans);
 894
 895         return 0;
 896 }
 897
 898 static int journal_reclaim_wait_done(struct bch_fs *c)
 899 {
 900         int ret = bch2_journal_error(&c->journal) ?:
 901                 !bch2_btree_key_cache_must_wait(c);
 902
 903         if (!ret)
 904                 journal_reclaim_kick(&c->journal);
 905         return ret;
 906 }
 907
 908 static noinline
 909 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 910                             struct btree_insert_entry *i,
 911                             int ret, unsigned long trace_ip)
 912 {
 913         struct bch_fs *c = trans->c;
 914
 915         switch (ret) {
 916         case -BCH_ERR_btree_insert_btree_node_full:
 917                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 918                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 919                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 920                 break;
 921         case -BCH_ERR_btree_insert_need_mark_replicas:
 922                 ret = drop_locks_do(trans,
 923                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 924                 break;
 925         case -BCH_ERR_journal_res_get_blocked:
 926                 /*
 927                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 928                  * flag
 929                  */
 930                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 931                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 932                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 933                         break;
 934                 }
 935
 936                 ret = drop_locks_do(trans,
 937                         bch2_trans_journal_res_get(trans,
 938                                         (flags & BCH_WATERMARK_MASK)|
 939                                         JOURNAL_RES_GET_CHECK));
 940                 break;
 941         case -BCH_ERR_btree_insert_need_journal_reclaim:
 942                 bch2_trans_unlock(trans);
 943
 944                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 945
 946                 wait_event_freezable(c->journal.reclaim_wait,
 947                                      (ret = journal_reclaim_wait_done(c)));
 948                 if (ret < 0)
 949                         break;
 950
 951                 ret = bch2_trans_relock(trans);
 952                 break;
 953         case -BCH_ERR_btree_insert_need_flush_buffer: {
 954                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 955
 956                 ret = 0;
 957
 958                 if (wb->state.nr > wb->size * 3 / 4) {
 959                         bch2_trans_unlock(trans);
 960                         mutex_lock(&wb->flush_lock);
 961
 962                         if (wb->state.nr > wb->size * 3 / 4) {
 963                                 bch2_trans_begin(trans);
 964                                 ret = __bch2_btree_write_buffer_flush(trans,
 965                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
 966                                 if (!ret) {
 967                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
 968                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
 969                                 }
 970                         } else {
 971                                 mutex_unlock(&wb->flush_lock);
 972                                 ret = bch2_trans_relock(trans);
 973                         }
 974                 }
 975                 break;
 976         }
 977         default:
 978                 BUG_ON(ret >= 0);
 979                 break;
 980         }
 981
 982         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 983
 984         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 985                                 !(flags & BTREE_INSERT_NOWAIT) &&
 986                                 (flags & BTREE_INSERT_NOFAIL), c,
 987                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 988
 989         return ret;
 990 }
 991
 992 static noinline int
 993 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 994 {
 995         struct bch_fs *c = trans->c;
 996         int ret;
 997
 998         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
 999             test_bit(BCH_FS_STARTED, &c->flags))
1000                 return -BCH_ERR_erofs_trans_commit;
1001
1002         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1003         if (ret)
1004                 return ret;
1005
1006         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1007         return 0;
1008 }
1009
1010 /*
1011  * This is for updates done in the early part of fsck - btree_gc - before we've
1012  * gone RW. we only add the new key to the list of keys for journal replay to
1013  * do.
1014  */
1015 static noinline int
1016 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1017 {
1018         struct bch_fs *c = trans->c;
1019         struct btree_insert_entry *i;
1020         int ret = 0;
1021
1022         trans_for_each_update(trans, i) {
1023                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1024                 if (ret)
1025                         break;
1026         }
1027
1028         return ret;
1029 }
1030
1031 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1032 {
1033         struct bch_fs *c = trans->c;
1034         struct btree_insert_entry *i = NULL;
1035         struct btree_write_buffered_key *wb;
1036         unsigned u64s;
1037         int ret = 0;
1038
1039         if (!trans->nr_updates &&
1040             !trans->nr_wb_updates &&
1041             !trans->extra_journal_entries.nr)
1042                 goto out_reset;
1043
1044         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1045                 lockdep_assert_held(&c->gc_lock);
1046
1047         ret = bch2_trans_commit_run_triggers(trans);
1048         if (ret)
1049                 goto out_reset;
1050
1051         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1052                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1053                 goto out_reset;
1054         }
1055
1056         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1057             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1058                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1059                 if (ret)
1060                         goto out_reset;
1061         }
1062
1063         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1064             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1065                 bch2_trans_begin(trans);
1066                 bch2_trans_unlock(trans);
1067
1068                 ret = __bch2_btree_write_buffer_flush(trans,
1069                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1070                 if (!ret) {
1071                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1072                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1073                 }
1074                 goto out;
1075         }
1076
1077         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1078
1079         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1080
1081         trans->journal_u64s             = trans->extra_journal_entries.nr;
1082         trans->journal_preres_u64s      = 0;
1083
1084         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1085
1086         if (trans->journal_transaction_names)
1087                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1088
1089         trans_for_each_update(trans, i) {
1090                 EBUG_ON(!i->path->should_be_locked);
1091
1092                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1093                 if (unlikely(ret))
1094                         goto out;
1095
1096                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1097
1098                 if (i->key_cache_already_flushed)
1099                         continue;
1100
1101                 /* we're going to journal the key being updated: */
1102                 u64s = jset_u64s(i->k->k.u64s);
1103                 if (i->cached &&
1104                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1105                         trans->journal_preres_u64s += u64s;
1106
1107                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1108                         continue;
1109
1110                 trans->journal_u64s += u64s;
1111
1112                 /* and we're also going to log the overwrite: */
1113                 if (trans->journal_transaction_names)
1114                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1115         }
1116
1117         trans_for_each_wb_update(trans, wb)
1118                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1119
1120         if (trans->extra_journal_res) {
1121                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1122                                 trans->extra_journal_res,
1123                                 (flags & BTREE_INSERT_NOFAIL)
1124                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1125                 if (ret)
1126                         goto err;
1127         }
1128 retry:
1129         bch2_trans_verify_not_in_restart(trans);
1130         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1131
1132         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1133
1134         /* make sure we didn't drop or screw up locks: */
1135         bch2_trans_verify_locks(trans);
1136
1137         if (ret)
1138                 goto err;
1139
1140         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1141 out:
1142         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1143
1144         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1145                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1146 out_reset:
1147         bch2_trans_reset_updates(trans);
1148
1149         return ret;
1150 err:
1151         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1152         if (ret)
1153                 goto out;
1154
1155         goto retry;
1156 }