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