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 = bch2_trans_get(c);
 167         unsigned long old, new, v;
 168         unsigned idx = w - b->writes;
 169
 170         btree_node_lock_nopath_nofail(trans, &b->c, SIX_LOCK_read);
 171         v = READ_ONCE(b->flags);
 172
 173         do {
 174                 old = new = v;
 175
 176                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 177                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 178                     w->journal.seq != seq)
 179                         break;
 180
 181                 new &= ~BTREE_WRITE_TYPE_MASK;
 182                 new |= BTREE_WRITE_journal_reclaim;
 183                 new |= 1 << BTREE_NODE_need_write;
 184         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 185
 186         btree_node_write_if_need(c, b, SIX_LOCK_read);
 187         six_unlock_read(&b->c.lock);
 188
 189         bch2_trans_put(trans);
 190         return 0;
 191 }
 192
 193 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 194 {
 195         return __btree_node_flush(j, pin, 0, seq);
 196 }
 197
 198 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 199 {
 200         return __btree_node_flush(j, pin, 1, seq);
 201 }
 202
 203 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 204                                        struct btree *b, u64 seq)
 205 {
 206         struct btree_write *w = btree_current_write(b);
 207
 208         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 209                              btree_node_write_idx(b) == 0
 210                              ? bch2_btree_node_flush0
 211                              : bch2_btree_node_flush1);
 212 }
 213
 214 /**
 215  * bch2_btree_insert_key_leaf() - insert a key one key into a leaf node
 216  * @trans:              btree transaction object
 217  * @path:               path pointing to @insert's pos
 218  * @insert:             key to insert
 219  * @journal_seq:        sequence number of journal reservation
 220  */
 221 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 222                                        struct btree_path *path,
 223                                        struct bkey_i *insert,
 224                                        u64 journal_seq)
 225 {
 226         struct bch_fs *c = trans->c;
 227         struct btree *b = path_l(path)->b;
 228         struct bset_tree *t = bset_tree_last(b);
 229         struct bset *i = bset(b, t);
 230         int old_u64s = bset_u64s(t);
 231         int old_live_u64s = b->nr.live_u64s;
 232         int live_u64s_added, u64s_added;
 233
 234         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 235                                         &path_l(path)->iter, insert)))
 236                 return;
 237
 238         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 239
 240         bch2_btree_add_journal_pin(c, b, journal_seq);
 241
 242         if (unlikely(!btree_node_dirty(b))) {
 243                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 244                 set_btree_node_dirty_acct(c, b);
 245         }
 246
 247         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 248         u64s_added = (int) bset_u64s(t) - old_u64s;
 249
 250         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 251                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 252         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 253                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 254
 255         if (u64s_added > live_u64s_added &&
 256             bch2_maybe_compact_whiteouts(c, b))
 257                 bch2_trans_node_reinit_iter(trans, b);
 258 }
 259
 260 /* Cached btree updates: */
 261
 262 /* Normal update interface: */
 263
 264 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 265                                              struct btree_insert_entry *i)
 266 {
 267         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 268         BUG_ON(i->cached        != i->path->cached);
 269         BUG_ON(i->level         != i->path->level);
 270         BUG_ON(i->btree_id      != i->path->btree_id);
 271         EBUG_ON(!i->level &&
 272                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 273                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 274                 i->k->k.p.snapshot &&
 275                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 276 }
 277
 278 static noinline int
 279 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
 280                                    unsigned long trace_ip)
 281 {
 282         return drop_locks_do(trans,
 283                 bch2_journal_preres_get(&trans->c->journal,
 284                         &trans->journal_preres,
 285                         trans->journal_preres_u64s,
 286                         (flags & BCH_WATERMARK_MASK)));
 287 }
 288
 289 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 290                                                       unsigned flags)
 291 {
 292         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 293                                     trans->journal_u64s, flags);
 294 }
 295
 296 #define JSET_ENTRY_LOG_U64s             4
 297
 298 static noinline void journal_transaction_name(struct btree_trans *trans)
 299 {
 300         struct bch_fs *c = trans->c;
 301         struct journal *j = &c->journal;
 302         struct jset_entry *entry =
 303                 bch2_journal_add_entry(j, &trans->journal_res,
 304                                        BCH_JSET_ENTRY_log, 0, 0,
 305                                        JSET_ENTRY_LOG_U64s);
 306         struct jset_entry_log *l =
 307                 container_of(entry, struct jset_entry_log, entry);
 308
 309         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 310 }
 311
 312 static inline int btree_key_can_insert(struct btree_trans *trans,
 313                                        struct btree *b, unsigned u64s)
 314 {
 315         struct bch_fs *c = trans->c;
 316
 317         if (!bch2_btree_node_insert_fits(c, b, u64s))
 318                 return -BCH_ERR_btree_insert_btree_node_full;
 319
 320         return 0;
 321 }
 322
 323 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 324                                        struct btree_path *path, unsigned u64s)
 325 {
 326         struct bch_fs *c = trans->c;
 327         struct bkey_cached *ck = (void *) path->l[0].b;
 328         struct btree_insert_entry *i;
 329         unsigned new_u64s;
 330         struct bkey_i *new_k;
 331
 332         EBUG_ON(path->level);
 333
 334         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 335             bch2_btree_key_cache_must_wait(c) &&
 336             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 337                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 338
 339         /*
 340          * bch2_varint_decode can read past the end of the buffer by at most 7
 341          * bytes (it won't be used):
 342          */
 343         u64s += 1;
 344
 345         if (u64s <= ck->u64s)
 346                 return 0;
 347
 348         new_u64s        = roundup_pow_of_two(u64s);
 349         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 350         if (!new_k) {
 351                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 352                         bch2_btree_id_str(path->btree_id), new_u64s);
 353                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 354         }
 355
 356         trans_for_each_update(trans, i)
 357                 if (i->old_v == &ck->k->v)
 358                         i->old_v = &new_k->v;
 359
 360         ck->u64s        = new_u64s;
 361         ck->k           = new_k;
 362         return 0;
 363 }
 364
 365 /* Triggers: */
 366
 367 static int run_one_mem_trigger(struct btree_trans *trans,
 368                                struct btree_insert_entry *i,
 369                                unsigned flags)
 370 {
 371         struct bkey_s_c old = { &i->old_k, i->old_v };
 372         struct bkey_i *new = i->k;
 373         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 374         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 375         int ret;
 376
 377         verify_update_old_key(trans, i);
 378
 379         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 380                 return 0;
 381
 382         if (!btree_node_type_needs_gc(__btree_node_type(i->level, i->btree_id)))
 383                 return 0;
 384
 385         if (old_ops->atomic_trigger == new_ops->atomic_trigger) {
 386                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 387                                 old, bkey_i_to_s_c(new),
 388                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 389         } else {
 390                 struct bkey             _deleted = KEY(0, 0, 0);
 391                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 392
 393                 _deleted.p = i->path->pos;
 394
 395                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 396                                 deleted, bkey_i_to_s_c(new),
 397                                 BTREE_TRIGGER_INSERT|flags) ?:
 398                         bch2_mark_key(trans, i->btree_id, i->level,
 399                                 old, deleted,
 400                                 BTREE_TRIGGER_OVERWRITE|flags);
 401         }
 402
 403         return ret;
 404 }
 405
 406 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 407                                  bool overwrite)
 408 {
 409         /*
 410          * Transactional triggers create new btree_insert_entries, so we can't
 411          * pass them a pointer to a btree_insert_entry, that memory is going to
 412          * move:
 413          */
 414         struct bkey old_k = i->old_k;
 415         struct bkey_s_c old = { &old_k, i->old_v };
 416         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 417         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 418
 419         verify_update_old_key(trans, i);
 420
 421         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 422             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 423                 return 0;
 424
 425         if (!i->insert_trigger_run &&
 426             !i->overwrite_trigger_run &&
 427             old_ops->trans_trigger == new_ops->trans_trigger) {
 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         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
 588         if (trans->nr_wb_updates &&
 589             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 590                 return -BCH_ERR_btree_insert_need_flush_buffer;
 591
 592         /*
 593          * Don't get journal reservation until after we know insert will
 594          * succeed:
 595          */
 596         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 597                 ret = bch2_trans_journal_res_get(trans,
 598                                 (flags & BCH_WATERMARK_MASK)|
 599                                 JOURNAL_RES_GET_NONBLOCK);
 600                 if (ret)
 601                         return ret;
 602
 603                 if (unlikely(trans->journal_transaction_names))
 604                         journal_transaction_name(trans);
 605         } else {
 606                 trans->journal_res.seq = c->journal.replay_journal_seq;
 607         }
 608
 609         /*
 610          * Not allowed to fail after we've gotten our journal reservation - we
 611          * have to use it:
 612          */
 613
 614         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 615             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 616                 if (bch2_journal_seq_verify)
 617                         trans_for_each_update(trans, i)
 618                                 i->k->k.version.lo = trans->journal_res.seq;
 619                 else if (bch2_inject_invalid_keys)
 620                         trans_for_each_update(trans, i)
 621                                 i->k->k.version = MAX_VERSION;
 622         }
 623
 624         if (trans->fs_usage_deltas &&
 625             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 626                 return -BCH_ERR_btree_insert_need_mark_replicas;
 627
 628         if (trans->nr_wb_updates) {
 629                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 630
 631                 ret = bch2_btree_insert_keys_write_buffer(trans);
 632                 if (ret)
 633                         goto revert_fs_usage;
 634         }
 635
 636         h = trans->hooks;
 637         while (h) {
 638                 ret = h->fn(trans, h);
 639                 if (ret)
 640                         goto revert_fs_usage;
 641                 h = h->next;
 642         }
 643
 644         trans_for_each_update(trans, i)
 645                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 646                         ret = run_one_mem_trigger(trans, i, i->flags);
 647                         if (ret)
 648                                 goto fatal_err;
 649                 }
 650
 651         if (unlikely(c->gc_pos.phase)) {
 652                 ret = bch2_trans_commit_run_gc_triggers(trans);
 653                 if  (ret)
 654                         goto fatal_err;
 655         }
 656
 657         if (unlikely(trans->extra_journal_entries.nr)) {
 658                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 659                                   trans->extra_journal_entries.data,
 660                                   trans->extra_journal_entries.nr);
 661
 662                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 663                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 664         }
 665
 666         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 667                 struct journal *j = &c->journal;
 668                 struct jset_entry *entry;
 669
 670                 trans_for_each_update(trans, i) {
 671                         if (i->key_cache_already_flushed)
 672                                 continue;
 673
 674                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 675                                 continue;
 676
 677                         verify_update_old_key(trans, i);
 678
 679                         if (trans->journal_transaction_names) {
 680                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 681                                                        BCH_JSET_ENTRY_overwrite,
 682                                                        i->btree_id, i->level,
 683                                                        i->old_k.u64s);
 684                                 bkey_reassemble((struct bkey_i *) entry->start,
 685                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 686                         }
 687
 688                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 689                                                BCH_JSET_ENTRY_btree_keys,
 690                                                i->btree_id, i->level,
 691                                                i->k->k.u64s);
 692                         bkey_copy((struct bkey_i *) entry->start, i->k);
 693                 }
 694
 695                 trans_for_each_wb_update(trans, wb) {
 696                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 697                                                BCH_JSET_ENTRY_btree_keys,
 698                                                wb->btree, 0,
 699                                                wb->k.k.u64s);
 700                         bkey_copy((struct bkey_i *) entry->start, &wb->k);
 701                 }
 702
 703                 if (trans->journal_seq)
 704                         *trans->journal_seq = trans->journal_res.seq;
 705         }
 706
 707         trans_for_each_update(trans, i) {
 708                 i->k->k.needs_whiteout = false;
 709
 710                 if (!i->cached) {
 711                         u64 seq = trans->journal_res.seq;
 712
 713                         if (i->flags & BTREE_UPDATE_PREJOURNAL)
 714                                 seq = i->seq;
 715
 716                         bch2_btree_insert_key_leaf(trans, i->path, i->k, seq);
 717                 } else if (!i->key_cache_already_flushed)
 718                         bch2_btree_insert_key_cached(trans, flags, i);
 719                 else {
 720                         bch2_btree_key_cache_drop(trans, i->path);
 721                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 722                 }
 723         }
 724
 725         return 0;
 726 fatal_err:
 727         bch2_fatal_error(c);
 728 revert_fs_usage:
 729         if (trans->fs_usage_deltas)
 730                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 731         return ret;
 732 }
 733
 734 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 735 {
 736         while (--i >= trans->updates) {
 737                 if (same_leaf_as_prev(trans, i))
 738                         continue;
 739
 740                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 741         }
 742
 743         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 744         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 745 }
 746
 747 static inline int trans_lock_write(struct btree_trans *trans)
 748 {
 749         struct btree_insert_entry *i;
 750
 751         trans_for_each_update(trans, i) {
 752                 if (same_leaf_as_prev(trans, i))
 753                         continue;
 754
 755                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 756                         return trans_lock_write_fail(trans, i);
 757
 758                 if (!i->cached)
 759                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 760         }
 761
 762         return 0;
 763 }
 764
 765 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 766 {
 767         struct btree_insert_entry *i;
 768         struct btree_write_buffered_key *wb;
 769
 770         trans_for_each_update(trans, i)
 771                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 772
 773         trans_for_each_wb_update(trans, wb)
 774                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 775 }
 776
 777 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 778                                                    enum bkey_invalid_flags flags,
 779                                                    struct btree_insert_entry *i,
 780                                                    struct printbuf *err)
 781 {
 782         struct bch_fs *c = trans->c;
 783
 784         printbuf_reset(err);
 785         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 786                    trans->fn, (void *) i->ip_allocated);
 787         prt_newline(err);
 788         printbuf_indent_add(err, 2);
 789
 790         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 791         prt_newline(err);
 792
 793         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k), i->bkey_type, flags, err);
 794         bch2_print_string_as_lines(KERN_ERR, err->buf);
 795
 796         bch2_inconsistent_error(c);
 797         bch2_dump_trans_updates(trans);
 798
 799         return -EINVAL;
 800 }
 801
 802 /*
 803  * Get journal reservation, take write locks, and attempt to do btree update(s):
 804  */
 805 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 806                                        struct btree_insert_entry **stopped_at,
 807                                        unsigned long trace_ip)
 808 {
 809         struct bch_fs *c = trans->c;
 810         struct btree_insert_entry *i;
 811         int ret = 0, u64s_delta = 0;
 812
 813         trans_for_each_update(trans, i) {
 814                 if (i->cached)
 815                         continue;
 816
 817                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 818                 u64s_delta -= i->old_btree_u64s;
 819
 820                 if (!same_leaf_as_next(trans, i)) {
 821                         if (u64s_delta <= 0) {
 822                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 823                                                         i->level, flags);
 824                                 if (unlikely(ret))
 825                                         return ret;
 826                         }
 827
 828                         u64s_delta = 0;
 829                 }
 830         }
 831
 832         ret = bch2_journal_preres_get(&c->journal,
 833                         &trans->journal_preres, trans->journal_preres_u64s,
 834                         (flags & BCH_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 835         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 836                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 837         if (unlikely(ret))
 838                 return ret;
 839
 840         ret = trans_lock_write(trans);
 841         if (unlikely(ret))
 842                 return ret;
 843
 844         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 845
 846         if (!ret && unlikely(trans->journal_replay_not_finished))
 847                 bch2_drop_overwrites_from_journal(trans);
 848
 849         trans_for_each_update(trans, i)
 850                 if (!same_leaf_as_prev(trans, i))
 851                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 852                                                         insert_l(i)->b);
 853
 854         if (!ret && trans->journal_pin)
 855                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 856                                      trans->journal_pin, NULL);
 857
 858         /*
 859          * Drop journal reservation after dropping write locks, since dropping
 860          * the journal reservation may kick off a journal write:
 861          */
 862         bch2_journal_res_put(&c->journal, &trans->journal_res);
 863
 864         return ret;
 865 }
 866
 867 static int journal_reclaim_wait_done(struct bch_fs *c)
 868 {
 869         int ret = bch2_journal_error(&c->journal) ?:
 870                 !bch2_btree_key_cache_must_wait(c);
 871
 872         if (!ret)
 873                 journal_reclaim_kick(&c->journal);
 874         return ret;
 875 }
 876
 877 static noinline
 878 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 879                             struct btree_insert_entry *i,
 880                             int ret, unsigned long trace_ip)
 881 {
 882         struct bch_fs *c = trans->c;
 883
 884         switch (ret) {
 885         case -BCH_ERR_btree_insert_btree_node_full:
 886                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 887                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 888                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 889                 break;
 890         case -BCH_ERR_btree_insert_need_mark_replicas:
 891                 ret = drop_locks_do(trans,
 892                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 893                 break;
 894         case -BCH_ERR_journal_res_get_blocked:
 895                 /*
 896                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 897                  * flag
 898                  */
 899                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 900                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 901                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 902                         break;
 903                 }
 904
 905                 ret = drop_locks_do(trans,
 906                         bch2_trans_journal_res_get(trans,
 907                                         (flags & BCH_WATERMARK_MASK)|
 908                                         JOURNAL_RES_GET_CHECK));
 909                 break;
 910         case -BCH_ERR_btree_insert_need_journal_reclaim:
 911                 bch2_trans_unlock(trans);
 912
 913                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 914
 915                 wait_event_freezable(c->journal.reclaim_wait,
 916                                      (ret = journal_reclaim_wait_done(c)));
 917                 if (ret < 0)
 918                         break;
 919
 920                 ret = bch2_trans_relock(trans);
 921                 break;
 922         case -BCH_ERR_btree_insert_need_flush_buffer: {
 923                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 924
 925                 ret = 0;
 926
 927                 if (wb->state.nr > wb->size * 3 / 4) {
 928                         bch2_trans_unlock(trans);
 929                         mutex_lock(&wb->flush_lock);
 930
 931                         if (wb->state.nr > wb->size * 3 / 4) {
 932                                 bch2_trans_begin(trans);
 933                                 ret = __bch2_btree_write_buffer_flush(trans,
 934                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
 935                                 if (!ret) {
 936                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
 937                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
 938                                 }
 939                         } else {
 940                                 mutex_unlock(&wb->flush_lock);
 941                                 ret = bch2_trans_relock(trans);
 942                         }
 943                 }
 944                 break;
 945         }
 946         default:
 947                 BUG_ON(ret >= 0);
 948                 break;
 949         }
 950
 951         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 952
 953         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 954                                 !(flags & BTREE_INSERT_NOWAIT) &&
 955                                 (flags & BTREE_INSERT_NOFAIL), c,
 956                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 957
 958         return ret;
 959 }
 960
 961 static noinline int
 962 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
 963 {
 964         struct bch_fs *c = trans->c;
 965         int ret;
 966
 967         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
 968             test_bit(BCH_FS_STARTED, &c->flags))
 969                 return -BCH_ERR_erofs_trans_commit;
 970
 971         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
 972         if (ret)
 973                 return ret;
 974
 975         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
 976         return 0;
 977 }
 978
 979 /*
 980  * This is for updates done in the early part of fsck - btree_gc - before we've
 981  * gone RW. we only add the new key to the list of keys for journal replay to
 982  * do.
 983  */
 984 static noinline int
 985 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 986 {
 987         struct bch_fs *c = trans->c;
 988         struct btree_insert_entry *i;
 989         int ret = 0;
 990
 991         trans_for_each_update(trans, i) {
 992                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 993                 if (ret)
 994                         break;
 995         }
 996
 997         return ret;
 998 }
 999
1000 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1001 {
1002         struct bch_fs *c = trans->c;
1003         struct btree_insert_entry *i = NULL;
1004         struct btree_write_buffered_key *wb;
1005         unsigned u64s;
1006         int ret = 0;
1007
1008         if (!trans->nr_updates &&
1009             !trans->nr_wb_updates &&
1010             !trans->extra_journal_entries.nr)
1011                 goto out_reset;
1012
1013         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1014                 lockdep_assert_held(&c->gc_lock);
1015
1016         ret = bch2_trans_commit_run_triggers(trans);
1017         if (ret)
1018                 goto out_reset;
1019
1020         trans_for_each_update(trans, i) {
1021                 struct printbuf buf = PRINTBUF;
1022                 enum bkey_invalid_flags invalid_flags = 0;
1023
1024                 if (!(flags & BTREE_INSERT_JOURNAL_REPLAY))
1025                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
1026
1027                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
1028                                                i->bkey_type, invalid_flags, &buf)))
1029                         ret = bch2_trans_commit_bkey_invalid(trans, invalid_flags, i, &buf);
1030                 btree_insert_entry_checks(trans, i);
1031                 printbuf_exit(&buf);
1032
1033                 if (ret)
1034                         return ret;
1035         }
1036
1037         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1038                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1039                 goto out_reset;
1040         }
1041
1042         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1043             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1044                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1045                 if (ret)
1046                         goto out_reset;
1047         }
1048
1049         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1050             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1051                 bch2_trans_begin(trans);
1052                 bch2_trans_unlock(trans);
1053
1054                 ret = __bch2_btree_write_buffer_flush(trans,
1055                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1056                 if (!ret) {
1057                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1058                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1059                 }
1060                 goto out;
1061         }
1062
1063         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1064
1065         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1066
1067         trans->journal_u64s             = trans->extra_journal_entries.nr;
1068         trans->journal_preres_u64s      = 0;
1069
1070         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1071
1072         if (trans->journal_transaction_names)
1073                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1074
1075         trans_for_each_update(trans, i) {
1076                 EBUG_ON(!i->path->should_be_locked);
1077
1078                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1079                 if (unlikely(ret))
1080                         goto out;
1081
1082                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1083
1084                 if (i->key_cache_already_flushed)
1085                         continue;
1086
1087                 /* we're going to journal the key being updated: */
1088                 u64s = jset_u64s(i->k->k.u64s);
1089                 if (i->cached &&
1090                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1091                         trans->journal_preres_u64s += u64s;
1092
1093                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1094                         continue;
1095
1096                 trans->journal_u64s += u64s;
1097
1098                 /* and we're also going to log the overwrite: */
1099                 if (trans->journal_transaction_names)
1100                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1101         }
1102
1103         trans_for_each_wb_update(trans, wb)
1104                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1105
1106         if (trans->extra_journal_res) {
1107                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1108                                 trans->extra_journal_res,
1109                                 (flags & BTREE_INSERT_NOFAIL)
1110                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1111                 if (ret)
1112                         goto err;
1113         }
1114 retry:
1115         bch2_trans_verify_not_in_restart(trans);
1116         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1117
1118         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1119
1120         /* make sure we didn't drop or screw up locks: */
1121         bch2_trans_verify_locks(trans);
1122
1123         if (ret)
1124                 goto err;
1125
1126         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1127 out:
1128         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1129
1130         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1131                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1132 out_reset:
1133         if (!ret)
1134                 bch2_trans_downgrade(trans);
1135         bch2_trans_reset_updates(trans);
1136
1137         return ret;
1138 err:
1139         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1140         if (ret)
1141                 goto out;
1142
1143         goto retry;
1144 }