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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "btree_write_buffer.h"
  12 #include "buckets.h"
  13 #include "debug.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "extent_update.h"
  17 #include "journal.h"
  18 #include "journal_reclaim.h"
  19 #include "keylist.h"
  20 #include "recovery.h"
  21 #include "subvolume.h"
  22 #include "replicas.h"
  23 #include "trace.h"
  24
  25 #include <linux/prefetch.h>
  26 #include <linux/sort.h>
  27
  28 /*
  29  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
  30  * different snapshot:
  31  */
  32 static struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
  33 {
  34         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
  35
  36         if (k.k && bpos_eq(path->pos, k.k->p))
  37                 return k;
  38
  39         bkey_init(u);
  40         u->p = path->pos;
  41         return (struct bkey_s_c) { u, NULL };
  42 }
  43
  44 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  45 {
  46 #ifdef CONFIG_BCACHEFS_DEBUG
  47         struct bch_fs *c = trans->c;
  48         struct bkey u;
  49         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
  50
  51         if (unlikely(trans->journal_replay_not_finished)) {
  52                 struct bkey_i *j_k =
  53                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  54
  55                 if (j_k)
  56                         k = bkey_i_to_s_c(j_k);
  57         }
  58
  59         u = *k.k;
  60         u.needs_whiteout = i->old_k.needs_whiteout;
  61
  62         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  63         BUG_ON(i->old_v != k.v);
  64 #endif
  65 }
  66
  67 static int __must_check
  68 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  69                           struct bkey_i *, enum btree_update_flags);
  70
  71 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  72                                          const struct btree_insert_entry *r)
  73 {
  74         return   cmp_int(l->btree_id,   r->btree_id) ?:
  75                  cmp_int(l->cached,     r->cached) ?:
  76                  -cmp_int(l->level,     r->level) ?:
  77                  bpos_cmp(l->k->k.p,    r->k->k.p);
  78 }
  79
  80 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  81 {
  82         return i->path->l + i->level;
  83 }
  84
  85 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  86                                      struct btree_insert_entry *i)
  87 {
  88         return i != trans->updates &&
  89                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  90 }
  91
  92 static inline bool same_leaf_as_next(struct btree_trans *trans,
  93                                      struct btree_insert_entry *i)
  94 {
  95         return i + 1 < trans->updates + trans->nr_updates &&
  96                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  97 }
  98
  99 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
 100                                            struct btree_path *path,
 101                                            struct btree *b)
 102 {
 103         struct bch_fs *c = trans->c;
 104
 105         if (unlikely(btree_node_just_written(b)) &&
 106             bch2_btree_post_write_cleanup(c, b))
 107                 bch2_trans_node_reinit_iter(trans, b);
 108
 109         /*
 110          * If the last bset has been written, or if it's gotten too big - start
 111          * a new bset to insert into:
 112          */
 113         if (want_new_bset(c, b))
 114                 bch2_btree_init_next(trans, b);
 115 }
 116
 117 /* Inserting into a given leaf node (last stage of insert): */
 118
 119 /* Handle overwrites and do insert, for non extents: */
 120 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 121                                 struct btree_path *path,
 122                                 struct btree *b,
 123                                 struct btree_node_iter *node_iter,
 124                                 struct bkey_i *insert)
 125 {
 126         struct bkey_packed *k;
 127         unsigned clobber_u64s = 0, new_u64s = 0;
 128
 129         EBUG_ON(btree_node_just_written(b));
 130         EBUG_ON(bset_written(b, btree_bset_last(b)));
 131         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 132         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 133         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 134         EBUG_ON(insert->k.u64s >
 135                 bch_btree_keys_u64s_remaining(trans->c, b));
 136
 137         k = bch2_btree_node_iter_peek_all(node_iter, b);
 138         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 139                 k = NULL;
 140
 141         /* @k is the key being overwritten/deleted, if any: */
 142         EBUG_ON(k && bkey_deleted(k));
 143
 144         /* Deleting, but not found? nothing to do: */
 145         if (bkey_deleted(&insert->k) && !k)
 146                 return false;
 147
 148         if (bkey_deleted(&insert->k)) {
 149                 /* Deleting: */
 150                 btree_account_key_drop(b, k);
 151                 k->type = KEY_TYPE_deleted;
 152
 153                 if (k->needs_whiteout)
 154                         push_whiteout(trans->c, b, insert->k.p);
 155                 k->needs_whiteout = false;
 156
 157                 if (k >= btree_bset_last(b)->start) {
 158                         clobber_u64s = k->u64s;
 159                         bch2_bset_delete(b, k, clobber_u64s);
 160                         goto fix_iter;
 161                 } else {
 162                         bch2_btree_path_fix_key_modified(trans, b, k);
 163                 }
 164
 165                 return true;
 166         }
 167
 168         if (k) {
 169                 /* Overwriting: */
 170                 btree_account_key_drop(b, k);
 171                 k->type = KEY_TYPE_deleted;
 172
 173                 insert->k.needs_whiteout = k->needs_whiteout;
 174                 k->needs_whiteout = false;
 175
 176                 if (k >= btree_bset_last(b)->start) {
 177                         clobber_u64s = k->u64s;
 178                         goto overwrite;
 179                 } else {
 180                         bch2_btree_path_fix_key_modified(trans, b, k);
 181                 }
 182         }
 183
 184         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 185 overwrite:
 186         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 187         new_u64s = k->u64s;
 188 fix_iter:
 189         if (clobber_u64s != new_u64s)
 190                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 191                                          clobber_u64s, new_u64s);
 192         return true;
 193 }
 194
 195 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 196                                unsigned i, u64 seq)
 197 {
 198         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 199         struct btree_write *w = container_of(pin, struct btree_write, journal);
 200         struct btree *b = container_of(w, struct btree, writes[i]);
 201         struct btree_trans trans;
 202         unsigned long old, new, v;
 203         unsigned idx = w - b->writes;
 204
 205         bch2_trans_init(&trans, c, 0, 0);
 206
 207         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 208         v = READ_ONCE(b->flags);
 209
 210         do {
 211                 old = new = v;
 212
 213                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 214                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 215                     w->journal.seq != seq)
 216                         break;
 217
 218                 new &= ~BTREE_WRITE_TYPE_MASK;
 219                 new |= BTREE_WRITE_journal_reclaim;
 220                 new |= 1 << BTREE_NODE_need_write;
 221         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 222
 223         btree_node_write_if_need(c, b, SIX_LOCK_read);
 224         six_unlock_read(&b->c.lock);
 225
 226         bch2_trans_exit(&trans);
 227         return 0;
 228 }
 229
 230 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 231 {
 232         return __btree_node_flush(j, pin, 0, seq);
 233 }
 234
 235 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 236 {
 237         return __btree_node_flush(j, pin, 1, seq);
 238 }
 239
 240 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 241                                        struct btree *b, u64 seq)
 242 {
 243         struct btree_write *w = btree_current_write(b);
 244
 245         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 246                              btree_node_write_idx(b) == 0
 247                              ? bch2_btree_node_flush0
 248                              : bch2_btree_node_flush1);
 249 }
 250
 251 /**
 252  * btree_insert_key - insert a key one key into a leaf node
 253  */
 254 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 255                                        struct btree_path *path,
 256                                        struct bkey_i *insert,
 257                                        u64 journal_seq)
 258 {
 259         struct bch_fs *c = trans->c;
 260         struct btree *b = path_l(path)->b;
 261         struct bset_tree *t = bset_tree_last(b);
 262         struct bset *i = bset(b, t);
 263         int old_u64s = bset_u64s(t);
 264         int old_live_u64s = b->nr.live_u64s;
 265         int live_u64s_added, u64s_added;
 266
 267         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 268                                         &path_l(path)->iter, insert)))
 269                 return;
 270
 271         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 272
 273         bch2_btree_add_journal_pin(c, b, journal_seq);
 274
 275         if (unlikely(!btree_node_dirty(b))) {
 276                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 277                 set_btree_node_dirty_acct(c, b);
 278         }
 279
 280         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 281         u64s_added = (int) bset_u64s(t) - old_u64s;
 282
 283         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 284                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 285         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 286                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 287
 288         if (u64s_added > live_u64s_added &&
 289             bch2_maybe_compact_whiteouts(c, b))
 290                 bch2_trans_node_reinit_iter(trans, b);
 291 }
 292
 293 static void btree_insert_key_leaf(struct btree_trans *trans,
 294                                   struct btree_insert_entry *insert)
 295 {
 296         bch2_btree_insert_key_leaf(trans, insert->path, insert->k, trans->journal_res.seq);
 297 }
 298
 299 /* Cached btree updates: */
 300
 301 /* Normal update interface: */
 302
 303 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 304                                              struct btree_insert_entry *i)
 305 {
 306         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 307         BUG_ON(i->cached        != i->path->cached);
 308         BUG_ON(i->level         != i->path->level);
 309         BUG_ON(i->btree_id      != i->path->btree_id);
 310         EBUG_ON(!i->level &&
 311                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 312                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 313                 i->k->k.p.snapshot &&
 314                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 315 }
 316
 317 static noinline int
 318 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
 319                                    unsigned long trace_ip)
 320 {
 321         return drop_locks_do(trans,
 322                 bch2_journal_preres_get(&trans->c->journal,
 323                         &trans->journal_preres,
 324                         trans->journal_preres_u64s,
 325                         (flags & BCH_WATERMARK_MASK)));
 326 }
 327
 328 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 329                                                       unsigned flags)
 330 {
 331         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 332                                     trans->journal_u64s, flags);
 333 }
 334
 335 #define JSET_ENTRY_LOG_U64s             4
 336
 337 static noinline void journal_transaction_name(struct btree_trans *trans)
 338 {
 339         struct bch_fs *c = trans->c;
 340         struct journal *j = &c->journal;
 341         struct jset_entry *entry =
 342                 bch2_journal_add_entry(j, &trans->journal_res,
 343                                        BCH_JSET_ENTRY_log, 0, 0,
 344                                        JSET_ENTRY_LOG_U64s);
 345         struct jset_entry_log *l =
 346                 container_of(entry, struct jset_entry_log, entry);
 347
 348         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 349 }
 350
 351 static inline int btree_key_can_insert(struct btree_trans *trans,
 352                                        struct btree *b, unsigned u64s)
 353 {
 354         struct bch_fs *c = trans->c;
 355
 356         if (!bch2_btree_node_insert_fits(c, b, u64s))
 357                 return -BCH_ERR_btree_insert_btree_node_full;
 358
 359         return 0;
 360 }
 361
 362 static int btree_key_can_insert_cached(struct btree_trans *trans, unsigned flags,
 363                                        struct btree_path *path, unsigned u64s)
 364 {
 365         struct bch_fs *c = trans->c;
 366         struct bkey_cached *ck = (void *) path->l[0].b;
 367         struct btree_insert_entry *i;
 368         unsigned new_u64s;
 369         struct bkey_i *new_k;
 370
 371         EBUG_ON(path->level);
 372
 373         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 374             bch2_btree_key_cache_must_wait(c) &&
 375             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 376                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 377
 378         /*
 379          * bch2_varint_decode can read past the end of the buffer by at most 7
 380          * bytes (it won't be used):
 381          */
 382         u64s += 1;
 383
 384         if (u64s <= ck->u64s)
 385                 return 0;
 386
 387         new_u64s        = roundup_pow_of_two(u64s);
 388         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 389         if (!new_k) {
 390                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 391                         bch2_btree_ids[path->btree_id], new_u64s);
 392                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 393         }
 394
 395         trans_for_each_update(trans, i)
 396                 if (i->old_v == &ck->k->v)
 397                         i->old_v = &new_k->v;
 398
 399         ck->u64s        = new_u64s;
 400         ck->k           = new_k;
 401         return 0;
 402 }
 403
 404 /* Triggers: */
 405
 406 static int run_one_mem_trigger(struct btree_trans *trans,
 407                                struct btree_insert_entry *i,
 408                                unsigned flags)
 409 {
 410         struct bkey_s_c old = { &i->old_k, i->old_v };
 411         struct bkey_i *new = i->k;
 412         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 413         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 414         int ret;
 415
 416         verify_update_old_key(trans, i);
 417
 418         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 419                 return 0;
 420
 421         if (!btree_node_type_needs_gc(i->btree_id))
 422                 return 0;
 423
 424         if (old_ops->atomic_trigger == new_ops->atomic_trigger &&
 425             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 426                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 427                                 old, bkey_i_to_s_c(new),
 428                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 429         } else {
 430                 struct bkey             _deleted = KEY(0, 0, 0);
 431                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 432
 433                 _deleted.p = i->path->pos;
 434
 435                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 436                                 deleted, bkey_i_to_s_c(new),
 437                                 BTREE_TRIGGER_INSERT|flags) ?:
 438                         bch2_mark_key(trans, i->btree_id, i->level,
 439                                 old, deleted,
 440                                 BTREE_TRIGGER_OVERWRITE|flags);
 441         }
 442
 443         return ret;
 444 }
 445
 446 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 447                                  bool overwrite)
 448 {
 449         /*
 450          * Transactional triggers create new btree_insert_entries, so we can't
 451          * pass them a pointer to a btree_insert_entry, that memory is going to
 452          * move:
 453          */
 454         struct bkey old_k = i->old_k;
 455         struct bkey_s_c old = { &old_k, i->old_v };
 456         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 457         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 458
 459         verify_update_old_key(trans, i);
 460
 461         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 462             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 463                 return 0;
 464
 465         if (!i->insert_trigger_run &&
 466             !i->overwrite_trigger_run &&
 467             old_ops->trans_trigger == new_ops->trans_trigger &&
 468             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 469                 i->overwrite_trigger_run = true;
 470                 i->insert_trigger_run = true;
 471                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 472                                            BTREE_TRIGGER_INSERT|
 473                                            BTREE_TRIGGER_OVERWRITE|
 474                                            i->flags) ?: 1;
 475         } else if (overwrite && !i->overwrite_trigger_run) {
 476                 i->overwrite_trigger_run = true;
 477                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 478         } else if (!overwrite && !i->insert_trigger_run) {
 479                 i->insert_trigger_run = true;
 480                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 481         } else {
 482                 return 0;
 483         }
 484 }
 485
 486 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 487                               struct btree_insert_entry *btree_id_start)
 488 {
 489         struct btree_insert_entry *i;
 490         bool trans_trigger_run;
 491         int ret, overwrite;
 492
 493         for (overwrite = 1; overwrite >= 0; --overwrite) {
 494
 495                 /*
 496                  * Running triggers will append more updates to the list of updates as
 497                  * we're walking it:
 498                  */
 499                 do {
 500                         trans_trigger_run = false;
 501
 502                         for (i = btree_id_start;
 503                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 504                              i++) {
 505                                 if (i->btree_id != btree_id)
 506                                         continue;
 507
 508                                 ret = run_one_trans_trigger(trans, i, overwrite);
 509                                 if (ret < 0)
 510                                         return ret;
 511                                 if (ret)
 512                                         trans_trigger_run = true;
 513                         }
 514                 } while (trans_trigger_run);
 515         }
 516
 517         return 0;
 518 }
 519
 520 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 521 {
 522         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 523         unsigned btree_id = 0;
 524         int ret = 0;
 525
 526         /*
 527          *
 528          * For a given btree, this algorithm runs insert triggers before
 529          * overwrite triggers: this is so that when extents are being moved
 530          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 531          * they are re-added.
 532          */
 533         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 534                 if (btree_id == BTREE_ID_alloc)
 535                         continue;
 536
 537                 while (btree_id_start < trans->updates + trans->nr_updates &&
 538                        btree_id_start->btree_id < btree_id)
 539                         btree_id_start++;
 540
 541                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 542                 if (ret)
 543                         return ret;
 544         }
 545
 546         trans_for_each_update(trans, i) {
 547                 if (i->btree_id > BTREE_ID_alloc)
 548                         break;
 549                 if (i->btree_id == BTREE_ID_alloc) {
 550                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 551                         if (ret)
 552                                 return ret;
 553                         break;
 554                 }
 555         }
 556
 557 #ifdef CONFIG_BCACHEFS_DEBUG
 558         trans_for_each_update(trans, i)
 559                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 560                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 561                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 562 #endif
 563         return 0;
 564 }
 565
 566 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 567 {
 568         struct bch_fs *c = trans->c;
 569         struct btree_insert_entry *i;
 570         int ret = 0;
 571
 572         trans_for_each_update(trans, i) {
 573                 /*
 574                  * XXX: synchronization of cached update triggers with gc
 575                  * XXX: synchronization of interior node updates with gc
 576                  */
 577                 BUG_ON(i->cached || i->level);
 578
 579                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 580                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 581                         if (ret)
 582                                 break;
 583                 }
 584         }
 585
 586         return ret;
 587 }
 588
 589 static inline int
 590 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 591                                struct btree_insert_entry **stopped_at,
 592                                unsigned long trace_ip)
 593 {
 594         struct bch_fs *c = trans->c;
 595         struct btree_insert_entry *i;
 596         struct btree_write_buffered_key *wb;
 597         struct btree_trans_commit_hook *h;
 598         unsigned u64s = 0;
 599         bool marking = false;
 600         int ret;
 601
 602         if (race_fault()) {
 603                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 604                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 605         }
 606
 607         /*
 608          * Check if the insert will fit in the leaf node with the write lock
 609          * held, otherwise another thread could write the node changing the
 610          * amount of space available:
 611          */
 612
 613         prefetch(&trans->c->journal.flags);
 614
 615         trans_for_each_update(trans, i) {
 616                 /* Multiple inserts might go to same leaf: */
 617                 if (!same_leaf_as_prev(trans, i))
 618                         u64s = 0;
 619
 620                 u64s += i->k->k.u64s;
 621                 ret = !i->cached
 622                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 623                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 624                 if (ret) {
 625                         *stopped_at = i;
 626                         return ret;
 627                 }
 628
 629                 if (btree_node_type_needs_gc(i->bkey_type))
 630                         marking = true;
 631         }
 632
 633         if (trans->nr_wb_updates &&
 634             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 635                 return -BCH_ERR_btree_insert_need_flush_buffer;
 636
 637         /*
 638          * Don't get journal reservation until after we know insert will
 639          * succeed:
 640          */
 641         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 642                 ret = bch2_trans_journal_res_get(trans,
 643                                 (flags & BCH_WATERMARK_MASK)|
 644                                 JOURNAL_RES_GET_NONBLOCK);
 645                 if (ret)
 646                         return ret;
 647
 648                 if (unlikely(trans->journal_transaction_names))
 649                         journal_transaction_name(trans);
 650         } else {
 651                 trans->journal_res.seq = c->journal.replay_journal_seq;
 652         }
 653
 654         /*
 655          * Not allowed to fail after we've gotten our journal reservation - we
 656          * have to use it:
 657          */
 658
 659         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 660             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 661                 if (bch2_journal_seq_verify)
 662                         trans_for_each_update(trans, i)
 663                                 i->k->k.version.lo = trans->journal_res.seq;
 664                 else if (bch2_inject_invalid_keys)
 665                         trans_for_each_update(trans, i)
 666                                 i->k->k.version = MAX_VERSION;
 667         }
 668
 669         if (trans->fs_usage_deltas &&
 670             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 671                 return -BCH_ERR_btree_insert_need_mark_replicas;
 672
 673         if (trans->nr_wb_updates) {
 674                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 675
 676                 ret = bch2_btree_insert_keys_write_buffer(trans);
 677                 if (ret)
 678                         goto revert_fs_usage;
 679         }
 680
 681         h = trans->hooks;
 682         while (h) {
 683                 ret = h->fn(trans, h);
 684                 if (ret)
 685                         goto revert_fs_usage;
 686                 h = h->next;
 687         }
 688
 689         trans_for_each_update(trans, i)
 690                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 691                         ret = run_one_mem_trigger(trans, i, i->flags);
 692                         if (ret)
 693                                 goto fatal_err;
 694                 }
 695
 696         if (unlikely(c->gc_pos.phase)) {
 697                 ret = bch2_trans_commit_run_gc_triggers(trans);
 698                 if  (ret)
 699                         goto fatal_err;
 700         }
 701
 702         if (unlikely(trans->extra_journal_entries.nr)) {
 703                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 704                                   trans->extra_journal_entries.data,
 705                                   trans->extra_journal_entries.nr);
 706
 707                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 708                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 709         }
 710
 711         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 712                 struct journal *j = &c->journal;
 713                 struct jset_entry *entry;
 714
 715                 trans_for_each_update(trans, i) {
 716                         if (i->key_cache_already_flushed)
 717                                 continue;
 718
 719                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 720                                 continue;
 721
 722                         verify_update_old_key(trans, i);
 723
 724                         if (trans->journal_transaction_names) {
 725                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 726                                                        BCH_JSET_ENTRY_overwrite,
 727                                                        i->btree_id, i->level,
 728                                                        i->old_k.u64s);
 729                                 bkey_reassemble(&entry->start[0],
 730                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 731                         }
 732
 733                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 734                                                BCH_JSET_ENTRY_btree_keys,
 735                                                i->btree_id, i->level,
 736                                                i->k->k.u64s);
 737                         bkey_copy(&entry->start[0], i->k);
 738                 }
 739
 740                 trans_for_each_wb_update(trans, wb) {
 741                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 742                                                BCH_JSET_ENTRY_btree_keys,
 743                                                wb->btree, 0,
 744                                                wb->k.k.u64s);
 745                         bkey_copy(&entry->start[0], &wb->k);
 746                 }
 747
 748                 if (trans->journal_seq)
 749                         *trans->journal_seq = trans->journal_res.seq;
 750         }
 751
 752         trans_for_each_update(trans, i) {
 753                 i->k->k.needs_whiteout = false;
 754
 755                 if (!i->cached)
 756                         btree_insert_key_leaf(trans, i);
 757                 else if (!i->key_cache_already_flushed)
 758                         bch2_btree_insert_key_cached(trans, flags, i);
 759                 else {
 760                         bch2_btree_key_cache_drop(trans, i->path);
 761                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 762                 }
 763         }
 764
 765         return 0;
 766 fatal_err:
 767         bch2_fatal_error(c);
 768 revert_fs_usage:
 769         if (trans->fs_usage_deltas)
 770                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 771         return ret;
 772 }
 773
 774 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 775 {
 776         while (--i >= trans->updates) {
 777                 if (same_leaf_as_prev(trans, i))
 778                         continue;
 779
 780                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 781         }
 782
 783         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 784         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 785 }
 786
 787 static inline int trans_lock_write(struct btree_trans *trans)
 788 {
 789         struct btree_insert_entry *i;
 790
 791         trans_for_each_update(trans, i) {
 792                 if (same_leaf_as_prev(trans, i))
 793                         continue;
 794
 795                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 796                         return trans_lock_write_fail(trans, i);
 797
 798                 if (!i->cached)
 799                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 800         }
 801
 802         return 0;
 803 }
 804
 805 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 806 {
 807         struct btree_insert_entry *i;
 808         struct btree_write_buffered_key *wb;
 809
 810         trans_for_each_update(trans, i)
 811                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 812
 813         trans_for_each_wb_update(trans, wb)
 814                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 815 }
 816
 817 #ifdef CONFIG_BCACHEFS_DEBUG
 818 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 819                                                    struct btree_insert_entry *i,
 820                                                    struct printbuf *err)
 821 {
 822         struct bch_fs *c = trans->c;
 823         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 824
 825         printbuf_reset(err);
 826         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 827                    trans->fn, (void *) i->ip_allocated);
 828         prt_newline(err);
 829         printbuf_indent_add(err, 2);
 830
 831         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 832         prt_newline(err);
 833
 834         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 835                           i->bkey_type, rw, err);
 836         bch2_print_string_as_lines(KERN_ERR, err->buf);
 837
 838         bch2_inconsistent_error(c);
 839         bch2_dump_trans_updates(trans);
 840         printbuf_exit(err);
 841
 842         return -EINVAL;
 843 }
 844 #endif
 845
 846 /*
 847  * Get journal reservation, take write locks, and attempt to do btree update(s):
 848  */
 849 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 850                                        struct btree_insert_entry **stopped_at,
 851                                        unsigned long trace_ip)
 852 {
 853         struct bch_fs *c = trans->c;
 854         struct btree_insert_entry *i;
 855         int ret, u64s_delta = 0;
 856
 857 #ifdef CONFIG_BCACHEFS_DEBUG
 858         struct printbuf buf = PRINTBUF;
 859
 860         trans_for_each_update(trans, i) {
 861                 enum bkey_invalid_flags invalid_flags = 0;
 862
 863                 if (!(flags & BTREE_INSERT_JOURNAL_REPLAY))
 864                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
 865
 866                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 867                                                i->bkey_type, invalid_flags, &buf)))
 868                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 869                 btree_insert_entry_checks(trans, i);
 870         }
 871         printbuf_exit(&buf);
 872 #endif
 873
 874         trans_for_each_update(trans, i) {
 875                 if (i->cached)
 876                         continue;
 877
 878                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 879                 u64s_delta -= i->old_btree_u64s;
 880
 881                 if (!same_leaf_as_next(trans, i)) {
 882                         if (u64s_delta <= 0) {
 883                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 884                                                         i->level, flags);
 885                                 if (unlikely(ret))
 886                                         return ret;
 887                         }
 888
 889                         u64s_delta = 0;
 890                 }
 891         }
 892
 893         ret = bch2_journal_preres_get(&c->journal,
 894                         &trans->journal_preres, trans->journal_preres_u64s,
 895                         (flags & BCH_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 896         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 897                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 898         if (unlikely(ret))
 899                 return ret;
 900
 901         ret = trans_lock_write(trans);
 902         if (unlikely(ret))
 903                 return ret;
 904
 905         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 906
 907         if (!ret && unlikely(trans->journal_replay_not_finished))
 908                 bch2_drop_overwrites_from_journal(trans);
 909
 910         trans_for_each_update(trans, i)
 911                 if (!same_leaf_as_prev(trans, i))
 912                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 913                                                         insert_l(i)->b);
 914
 915         if (!ret && trans->journal_pin)
 916                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 917                                      trans->journal_pin, NULL);
 918
 919         /*
 920          * Drop journal reservation after dropping write locks, since dropping
 921          * the journal reservation may kick off a journal write:
 922          */
 923         bch2_journal_res_put(&c->journal, &trans->journal_res);
 924
 925         if (unlikely(ret))
 926                 return ret;
 927
 928         bch2_trans_downgrade(trans);
 929
 930         return 0;
 931 }
 932
 933 static int journal_reclaim_wait_done(struct bch_fs *c)
 934 {
 935         int ret = bch2_journal_error(&c->journal) ?:
 936                 !bch2_btree_key_cache_must_wait(c);
 937
 938         if (!ret)
 939                 journal_reclaim_kick(&c->journal);
 940         return ret;
 941 }
 942
 943 static noinline
 944 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 945                             struct btree_insert_entry *i,
 946                             int ret, unsigned long trace_ip)
 947 {
 948         struct bch_fs *c = trans->c;
 949
 950         switch (ret) {
 951         case -BCH_ERR_btree_insert_btree_node_full:
 952                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 953                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 954                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 955                 break;
 956         case -BCH_ERR_btree_insert_need_mark_replicas:
 957                 ret = drop_locks_do(trans,
 958                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 959                 break;
 960         case -BCH_ERR_journal_res_get_blocked:
 961                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 962                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 963                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 964                         break;
 965                 }
 966
 967                 ret = drop_locks_do(trans,
 968                         bch2_trans_journal_res_get(trans,
 969                                         (flags & BCH_WATERMARK_MASK)|
 970                                         JOURNAL_RES_GET_CHECK));
 971                 break;
 972         case -BCH_ERR_btree_insert_need_journal_reclaim:
 973                 bch2_trans_unlock(trans);
 974
 975                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 976
 977                 wait_event_freezable(c->journal.reclaim_wait,
 978                                      (ret = journal_reclaim_wait_done(c)));
 979                 if (ret < 0)
 980                         break;
 981
 982                 ret = bch2_trans_relock(trans);
 983                 break;
 984         case -BCH_ERR_btree_insert_need_flush_buffer: {
 985                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 986
 987                 ret = 0;
 988
 989                 if (wb->state.nr > wb->size * 3 / 4) {
 990                         bch2_trans_unlock(trans);
 991                         mutex_lock(&wb->flush_lock);
 992
 993                         if (wb->state.nr > wb->size * 3 / 4) {
 994                                 bch2_trans_begin(trans);
 995                                 ret = __bch2_btree_write_buffer_flush(trans,
 996                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
 997                                 if (!ret) {
 998                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
 999                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1000                                 }
1001                         } else {
1002                                 mutex_unlock(&wb->flush_lock);
1003                                 ret = bch2_trans_relock(trans);
1004                         }
1005                 }
1006                 break;
1007         }
1008         default:
1009                 BUG_ON(ret >= 0);
1010                 break;
1011         }
1012
1013         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1014
1015         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1016                                 !(flags & BTREE_INSERT_NOWAIT) &&
1017                                 (flags & BTREE_INSERT_NOFAIL), c,
1018                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1019
1020         return ret;
1021 }
1022
1023 static noinline int
1024 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1025 {
1026         struct bch_fs *c = trans->c;
1027         int ret;
1028
1029         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1030             test_bit(BCH_FS_STARTED, &c->flags))
1031                 return -BCH_ERR_erofs_trans_commit;
1032
1033         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1034         if (ret)
1035                 return ret;
1036
1037         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1038         return 0;
1039 }
1040
1041 /*
1042  * This is for updates done in the early part of fsck - btree_gc - before we've
1043  * gone RW. we only add the new key to the list of keys for journal replay to
1044  * do.
1045  */
1046 static noinline int
1047 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1048 {
1049         struct bch_fs *c = trans->c;
1050         struct btree_insert_entry *i;
1051         int ret = 0;
1052
1053         trans_for_each_update(trans, i) {
1054                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1055                 if (ret)
1056                         break;
1057         }
1058
1059         return ret;
1060 }
1061
1062 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1063 {
1064         struct bch_fs *c = trans->c;
1065         struct btree_insert_entry *i = NULL;
1066         struct btree_write_buffered_key *wb;
1067         unsigned u64s;
1068         int ret = 0;
1069
1070         if (!trans->nr_updates &&
1071             !trans->nr_wb_updates &&
1072             !trans->extra_journal_entries.nr)
1073                 goto out_reset;
1074
1075         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1076                 lockdep_assert_held(&c->gc_lock);
1077
1078         ret = bch2_trans_commit_run_triggers(trans);
1079         if (ret)
1080                 goto out_reset;
1081
1082         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1083                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1084                 goto out_reset;
1085         }
1086
1087         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1088             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1089                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1090                 if (ret)
1091                         goto out_reset;
1092         }
1093
1094         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1095             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1096                 bch2_trans_begin(trans);
1097                 bch2_trans_unlock(trans);
1098
1099                 ret = __bch2_btree_write_buffer_flush(trans,
1100                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1101                 if (!ret) {
1102                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1103                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1104                 }
1105                 goto out;
1106         }
1107
1108         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1109
1110         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1111
1112         trans->journal_u64s             = trans->extra_journal_entries.nr;
1113         trans->journal_preres_u64s      = 0;
1114
1115         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1116
1117         if (trans->journal_transaction_names)
1118                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1119
1120         trans_for_each_update(trans, i) {
1121                 EBUG_ON(!i->path->should_be_locked);
1122
1123                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1124                 if (unlikely(ret))
1125                         goto out;
1126
1127                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1128
1129                 if (i->key_cache_already_flushed)
1130                         continue;
1131
1132                 /* we're going to journal the key being updated: */
1133                 u64s = jset_u64s(i->k->k.u64s);
1134                 if (i->cached &&
1135                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1136                         trans->journal_preres_u64s += u64s;
1137
1138                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1139                         continue;
1140
1141                 trans->journal_u64s += u64s;
1142
1143                 /* and we're also going to log the overwrite: */
1144                 if (trans->journal_transaction_names)
1145                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1146         }
1147
1148         trans_for_each_wb_update(trans, wb)
1149                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1150
1151         if (trans->extra_journal_res) {
1152                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1153                                 trans->extra_journal_res,
1154                                 (flags & BTREE_INSERT_NOFAIL)
1155                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1156                 if (ret)
1157                         goto err;
1158         }
1159 retry:
1160         bch2_trans_verify_not_in_restart(trans);
1161         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1162
1163         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1164
1165         /* make sure we didn't drop or screw up locks: */
1166         bch2_trans_verify_locks(trans);
1167
1168         if (ret)
1169                 goto err;
1170
1171         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1172 out:
1173         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1174
1175         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1176                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1177 out_reset:
1178         bch2_trans_reset_updates(trans);
1179
1180         return ret;
1181 err:
1182         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1183         if (ret)
1184                 goto out;
1185
1186         goto retry;
1187 }
1188
1189 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1190                                           enum btree_id id,
1191                                           struct bpos pos)
1192 {
1193         struct bch_fs *c = trans->c;
1194         struct btree_iter iter;
1195         struct bkey_s_c k;
1196         int ret;
1197
1198         bch2_trans_iter_init(trans, &iter, id, pos,
1199                              BTREE_ITER_NOT_EXTENTS|
1200                              BTREE_ITER_ALL_SNAPSHOTS);
1201         while (1) {
1202                 k = bch2_btree_iter_prev(&iter);
1203                 ret = bkey_err(k);
1204                 if (ret)
1205                         break;
1206
1207                 if (!k.k)
1208                         break;
1209
1210                 if (!bkey_eq(pos, k.k->p))
1211                         break;
1212
1213                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1214                         ret = 1;
1215                         break;
1216                 }
1217         }
1218         bch2_trans_iter_exit(trans, &iter);
1219
1220         return ret;
1221 }
1222
1223 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1224                                           enum btree_id id,
1225                                           struct bpos pos)
1226 {
1227         if (!btree_type_has_snapshots(id) ||
1228             !snapshot_t(trans->c, pos.snapshot)->children[0])
1229                 return 0;
1230
1231         return __check_pos_snapshot_overwritten(trans, id, pos);
1232 }
1233
1234 static noinline int extent_front_merge(struct btree_trans *trans,
1235                                        struct btree_iter *iter,
1236                                        struct bkey_s_c k,
1237                                        struct bkey_i **insert,
1238                                        enum btree_update_flags flags)
1239 {
1240         struct bch_fs *c = trans->c;
1241         struct bkey_i *update;
1242         int ret;
1243
1244         update = bch2_bkey_make_mut_noupdate(trans, k);
1245         ret = PTR_ERR_OR_ZERO(update);
1246         if (ret)
1247                 return ret;
1248
1249         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1250                 return 0;
1251
1252         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1253                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1254         if (ret < 0)
1255                 return ret;
1256         if (ret)
1257                 return 0;
1258
1259         ret = bch2_btree_delete_at(trans, iter, flags);
1260         if (ret)
1261                 return ret;
1262
1263         *insert = update;
1264         return 0;
1265 }
1266
1267 static noinline int extent_back_merge(struct btree_trans *trans,
1268                                       struct btree_iter *iter,
1269                                       struct bkey_i *insert,
1270                                       struct bkey_s_c k)
1271 {
1272         struct bch_fs *c = trans->c;
1273         int ret;
1274
1275         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1276                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1277         if (ret < 0)
1278                 return ret;
1279         if (ret)
1280                 return 0;
1281
1282         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1283         return 0;
1284 }
1285
1286 /*
1287  * When deleting, check if we need to emit a whiteout (because we're overwriting
1288  * something in an ancestor snapshot)
1289  */
1290 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1291                                       enum btree_id btree_id, struct bpos pos)
1292 {
1293         struct btree_iter iter;
1294         struct bkey_s_c k;
1295         u32 snapshot = pos.snapshot;
1296         int ret;
1297
1298         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1299                 return 0;
1300
1301         pos.snapshot++;
1302
1303         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1304                            BTREE_ITER_ALL_SNAPSHOTS|
1305                            BTREE_ITER_NOPRESERVE, k, ret) {
1306                 if (!bkey_eq(k.k->p, pos))
1307                         break;
1308
1309                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1310                                               k.k->p.snapshot)) {
1311                         ret = !bkey_whiteout(k.k);
1312                         break;
1313                 }
1314         }
1315         bch2_trans_iter_exit(trans, &iter);
1316
1317         return ret;
1318 }
1319
1320 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
1321                                    enum btree_id id,
1322                                    struct bpos old_pos,
1323                                    struct bpos new_pos)
1324 {
1325         struct bch_fs *c = trans->c;
1326         struct btree_iter old_iter, new_iter;
1327         struct bkey_s_c old_k, new_k;
1328         snapshot_id_list s;
1329         struct bkey_i *update;
1330         int ret;
1331
1332         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
1333                 return 0;
1334
1335         darray_init(&s);
1336
1337         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
1338                              BTREE_ITER_NOT_EXTENTS|
1339                              BTREE_ITER_ALL_SNAPSHOTS);
1340         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
1341                !(ret = bkey_err(old_k)) &&
1342                bkey_eq(old_pos, old_k.k->p)) {
1343                 struct bpos whiteout_pos =
1344                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
1345
1346                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
1347                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
1348                         continue;
1349
1350                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
1351                                            BTREE_ITER_NOT_EXTENTS|
1352                                            BTREE_ITER_INTENT);
1353                 ret = bkey_err(new_k);
1354                 if (ret)
1355                         break;
1356
1357                 if (new_k.k->type == KEY_TYPE_deleted) {
1358                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
1359                         ret = PTR_ERR_OR_ZERO(update);
1360                         if (ret)
1361                                 break;
1362
1363                         bkey_init(&update->k);
1364                         update->k.p             = whiteout_pos;
1365                         update->k.type          = KEY_TYPE_whiteout;
1366
1367                         ret = bch2_trans_update(trans, &new_iter, update,
1368                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1369                 }
1370                 bch2_trans_iter_exit(trans, &new_iter);
1371
1372                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
1373                 if (ret)
1374                         break;
1375         }
1376         bch2_trans_iter_exit(trans, &old_iter);
1377         darray_exit(&s);
1378
1379         return ret;
1380 }
1381
1382 int bch2_trans_update_extent(struct btree_trans *trans,
1383                              struct btree_iter *orig_iter,
1384                              struct bkey_i *insert,
1385                              enum btree_update_flags flags)
1386 {
1387         struct btree_iter iter;
1388         struct bpos start = bkey_start_pos(&insert->k);
1389         struct bkey_i *update;
1390         struct bkey_s_c k;
1391         enum btree_id btree_id = orig_iter->btree_id;
1392         int ret = 0, compressed_sectors;
1393
1394         bch2_trans_iter_init(trans, &iter, btree_id, start,
1395                              BTREE_ITER_INTENT|
1396                              BTREE_ITER_WITH_UPDATES|
1397                              BTREE_ITER_NOT_EXTENTS);
1398         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1399         if ((ret = bkey_err(k)))
1400                 goto err;
1401         if (!k.k)
1402                 goto out;
1403
1404         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1405                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1406                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1407                         if (ret)
1408                                 goto err;
1409                 }
1410
1411                 goto next;
1412         }
1413
1414         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1415                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1416                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1417
1418                 /*
1419                  * If we're going to be splitting a compressed extent, note it
1420                  * so that __bch2_trans_commit() can increase our disk
1421                  * reservation:
1422                  */
1423                 if (((front_split && back_split) ||
1424                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1425                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1426                         trans->extra_journal_res += compressed_sectors;
1427
1428                 if (front_split) {
1429                         update = bch2_bkey_make_mut_noupdate(trans, k);
1430                         if ((ret = PTR_ERR_OR_ZERO(update)))
1431                                 goto err;
1432
1433                         bch2_cut_back(start, update);
1434
1435                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1436                                                 k.k->p, update->k.p) ?:
1437                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1438                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1439                         if (ret)
1440                                 goto err;
1441                 }
1442
1443                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1444                     (front_split || back_split)) {
1445                         update = bch2_bkey_make_mut_noupdate(trans, k);
1446                         if ((ret = PTR_ERR_OR_ZERO(update)))
1447                                 goto err;
1448
1449                         bch2_cut_front(start, update);
1450                         bch2_cut_back(insert->k.p, update);
1451
1452                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1453                                                 k.k->p, update->k.p) ?:
1454                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1455                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1456                         if (ret)
1457                                 goto err;
1458                 }
1459
1460                 if (bkey_le(k.k->p, insert->k.p)) {
1461                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1462                         if ((ret = PTR_ERR_OR_ZERO(update)))
1463                                 goto err;
1464
1465                         bkey_init(&update->k);
1466                         update->k.p = k.k->p;
1467                         update->k.p.snapshot = insert->k.p.snapshot;
1468
1469                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1470                                 update->k.type = KEY_TYPE_whiteout;
1471                         } else if (btree_type_has_snapshots(btree_id)) {
1472                                 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1473                                 if (ret < 0)
1474                                         goto err;
1475                                 if (ret)
1476                                         update->k.type = KEY_TYPE_whiteout;
1477                         }
1478
1479                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1480                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1481                         if (ret)
1482                                 goto err;
1483                 }
1484
1485                 if (back_split) {
1486                         update = bch2_bkey_make_mut_noupdate(trans, k);
1487                         if ((ret = PTR_ERR_OR_ZERO(update)))
1488                                 goto err;
1489
1490                         bch2_cut_front(insert->k.p, update);
1491
1492                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1493                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1494                                                   flags);
1495                         if (ret)
1496                                 goto err;
1497                         goto out;
1498                 }
1499 next:
1500                 bch2_btree_iter_advance(&iter);
1501                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1502                 if ((ret = bkey_err(k)))
1503                         goto err;
1504                 if (!k.k)
1505                         goto out;
1506         }
1507
1508         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1509                 ret = extent_back_merge(trans, &iter, insert, k);
1510                 if (ret)
1511                         goto err;
1512         }
1513 out:
1514         if (!bkey_deleted(&insert->k)) {
1515                 /*
1516                  * Rewinding iterators is expensive: get a new one and the one
1517                  * that points to the start of insert will be cloned from:
1518                  */
1519                 bch2_trans_iter_exit(trans, &iter);
1520                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1521                                      BTREE_ITER_NOT_EXTENTS|
1522                                      BTREE_ITER_INTENT);
1523                 ret   = bch2_btree_iter_traverse(&iter) ?:
1524                         bch2_trans_update(trans, &iter, insert, flags);
1525         }
1526 err:
1527         bch2_trans_iter_exit(trans, &iter);
1528
1529         return ret;
1530 }
1531
1532 static int __must_check
1533 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1534                                 struct bkey_i *k, enum btree_update_flags flags,
1535                                 unsigned long ip);
1536
1537 static noinline int flush_new_cached_update(struct btree_trans *trans,
1538                                             struct btree_path *path,
1539                                             struct btree_insert_entry *i,
1540                                             enum btree_update_flags flags,
1541                                             unsigned long ip)
1542 {
1543         struct btree_path *btree_path;
1544         struct bkey k;
1545         int ret;
1546
1547         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1548                                    BTREE_ITER_INTENT, _THIS_IP_);
1549         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1550         if (ret)
1551                 goto out;
1552
1553         /*
1554          * The old key in the insert entry might actually refer to an existing
1555          * key in the btree that has been deleted from cache and not yet
1556          * flushed. Check for this and skip the flush so we don't run triggers
1557          * against a stale key.
1558          */
1559         bch2_btree_path_peek_slot_exact(btree_path, &k);
1560         if (!bkey_deleted(&k))
1561                 goto out;
1562
1563         i->key_cache_already_flushed = true;
1564         i->flags |= BTREE_TRIGGER_NORUN;
1565
1566         btree_path_set_should_be_locked(btree_path);
1567         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1568 out:
1569         bch2_path_put(trans, btree_path, true);
1570         return ret;
1571 }
1572
1573 static int __must_check
1574 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1575                                 struct bkey_i *k, enum btree_update_flags flags,
1576                                 unsigned long ip)
1577 {
1578         struct bch_fs *c = trans->c;
1579         struct btree_insert_entry *i, n;
1580         int cmp;
1581
1582         EBUG_ON(!path->should_be_locked);
1583         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1584         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1585
1586         n = (struct btree_insert_entry) {
1587                 .flags          = flags,
1588                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1589                 .btree_id       = path->btree_id,
1590                 .level          = path->level,
1591                 .cached         = path->cached,
1592                 .path           = path,
1593                 .k              = k,
1594                 .ip_allocated   = ip,
1595         };
1596
1597 #ifdef CONFIG_BCACHEFS_DEBUG
1598         trans_for_each_update(trans, i)
1599                 BUG_ON(i != trans->updates &&
1600                        btree_insert_entry_cmp(i - 1, i) >= 0);
1601 #endif
1602
1603         /*
1604          * Pending updates are kept sorted: first, find position of new update,
1605          * then delete/trim any updates the new update overwrites:
1606          */
1607         trans_for_each_update(trans, i) {
1608                 cmp = btree_insert_entry_cmp(&n, i);
1609                 if (cmp <= 0)
1610                         break;
1611         }
1612
1613         if (!cmp && i < trans->updates + trans->nr_updates) {
1614                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1615
1616                 bch2_path_put(trans, i->path, true);
1617                 i->flags        = n.flags;
1618                 i->cached       = n.cached;
1619                 i->k            = n.k;
1620                 i->path         = n.path;
1621                 i->ip_allocated = n.ip_allocated;
1622         } else {
1623                 array_insert_item(trans->updates, trans->nr_updates,
1624                                   i - trans->updates, n);
1625
1626                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1627                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1628
1629                 if (unlikely(trans->journal_replay_not_finished)) {
1630                         struct bkey_i *j_k =
1631                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1632
1633                         if (j_k) {
1634                                 i->old_k = j_k->k;
1635                                 i->old_v = &j_k->v;
1636                         }
1637                 }
1638         }
1639
1640         __btree_path_get(i->path, true);
1641
1642         /*
1643          * If a key is present in the key cache, it must also exist in the
1644          * btree - this is necessary for cache coherency. When iterating over
1645          * a btree that's cached in the key cache, the btree iter code checks
1646          * the key cache - but the key has to exist in the btree for that to
1647          * work:
1648          */
1649         if (path->cached && bkey_deleted(&i->old_k))
1650                 return flush_new_cached_update(trans, path, i, flags, ip);
1651
1652         return 0;
1653 }
1654
1655 static inline int __must_check
1656 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1657                           struct bkey_i *k, enum btree_update_flags flags)
1658 {
1659         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1660 }
1661
1662 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1663                                    struct bkey_i *k, enum btree_update_flags flags)
1664 {
1665         struct btree_path *path = iter->update_path ?: iter->path;
1666         struct bkey_cached *ck;
1667         int ret;
1668
1669         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1670                 return bch2_trans_update_extent(trans, iter, k, flags);
1671
1672         if (bkey_deleted(&k->k) &&
1673             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1674             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1675                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1676                 if (unlikely(ret < 0))
1677                         return ret;
1678
1679                 if (ret)
1680                         k->k.type = KEY_TYPE_whiteout;
1681         }
1682
1683         /*
1684          * Ensure that updates to cached btrees go to the key cache:
1685          */
1686         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1687             !path->cached &&
1688             !path->level &&
1689             btree_id_cached(trans->c, path->btree_id)) {
1690                 if (!iter->key_cache_path ||
1691                     !iter->key_cache_path->should_be_locked ||
1692                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1693                         if (!iter->key_cache_path)
1694                                 iter->key_cache_path =
1695                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1696                                                       BTREE_ITER_INTENT|
1697                                                       BTREE_ITER_CACHED, _THIS_IP_);
1698
1699                         iter->key_cache_path =
1700                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1701                                                         iter->flags & BTREE_ITER_INTENT,
1702                                                         _THIS_IP_);
1703
1704                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1705                                                        BTREE_ITER_CACHED);
1706                         if (unlikely(ret))
1707                                 return ret;
1708
1709                         ck = (void *) iter->key_cache_path->l[0].b;
1710
1711                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1712                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1713                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1714                         }
1715
1716                         btree_path_set_should_be_locked(iter->key_cache_path);
1717                 }
1718
1719                 path = iter->key_cache_path;
1720         }
1721
1722         return bch2_trans_update_by_path(trans, path, k, flags);
1723 }
1724
1725 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1726                                             enum btree_id btree,
1727                                             struct bkey_i *k)
1728 {
1729         struct btree_write_buffered_key *i;
1730         int ret;
1731
1732         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1733         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1734
1735         trans_for_each_wb_update(trans, i) {
1736                 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
1737                         bkey_copy(&i->k, k);
1738                         return 0;
1739                 }
1740         }
1741
1742         if (!trans->wb_updates ||
1743             trans->nr_wb_updates == trans->wb_updates_size) {
1744                 struct btree_write_buffered_key *u;
1745
1746                 if (trans->nr_wb_updates == trans->wb_updates_size) {
1747                         struct btree_transaction_stats *s = btree_trans_stats(trans);
1748
1749                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1750                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1751                         if (s)
1752                                 s->wb_updates_size = trans->wb_updates_size;
1753                 }
1754
1755                 u = bch2_trans_kmalloc_nomemzero(trans,
1756                                         trans->wb_updates_size *
1757                                         sizeof(struct btree_write_buffered_key));
1758                 ret = PTR_ERR_OR_ZERO(u);
1759                 if (ret)
1760                         return ret;
1761
1762                 if (trans->nr_wb_updates)
1763                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1764                                sizeof(struct btree_write_buffered_key));
1765                 trans->wb_updates = u;
1766         }
1767
1768         trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
1769                 .btree  = btree,
1770         };
1771
1772         bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
1773         trans->nr_wb_updates++;
1774
1775         return 0;
1776 }
1777
1778 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1779                              enum btree_id btree, struct bpos end)
1780 {
1781         struct bkey_s_c k;
1782         int ret = 0;
1783
1784         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1785         k = bch2_btree_iter_prev(iter);
1786         ret = bkey_err(k);
1787         if (ret)
1788                 goto err;
1789
1790         bch2_btree_iter_advance(iter);
1791         k = bch2_btree_iter_peek_slot(iter);
1792         ret = bkey_err(k);
1793         if (ret)
1794                 goto err;
1795
1796         BUG_ON(k.k->type != KEY_TYPE_deleted);
1797
1798         if (bkey_gt(k.k->p, end)) {
1799                 ret = -BCH_ERR_ENOSPC_btree_slot;
1800                 goto err;
1801         }
1802
1803         return 0;
1804 err:
1805         bch2_trans_iter_exit(trans, iter);
1806         return ret;
1807 }
1808
1809 void bch2_trans_commit_hook(struct btree_trans *trans,
1810                             struct btree_trans_commit_hook *h)
1811 {
1812         h->next = trans->hooks;
1813         trans->hooks = h;
1814 }
1815
1816 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1817                                 enum btree_id btree, struct bkey_i *k,
1818                                 enum btree_update_flags flags)
1819 {
1820         struct btree_iter iter;
1821         int ret;
1822
1823         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1824                              BTREE_ITER_NOT_EXTENTS|
1825                              BTREE_ITER_INTENT);
1826         ret   = bch2_btree_iter_traverse(&iter) ?:
1827                 bch2_trans_update(trans, &iter, k, flags);
1828         bch2_trans_iter_exit(trans, &iter);
1829         return ret;
1830 }
1831
1832 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1833                         struct bkey_i *k, enum btree_update_flags flags)
1834 {
1835         struct btree_iter iter;
1836         int ret;
1837
1838         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1839                              BTREE_ITER_CACHED|
1840                              BTREE_ITER_INTENT);
1841         ret   = bch2_btree_iter_traverse(&iter) ?:
1842                 bch2_trans_update(trans, &iter, k, flags);
1843         bch2_trans_iter_exit(trans, &iter);
1844         return ret;
1845 }
1846
1847 /**
1848  * bch2_btree_insert - insert keys into the extent btree
1849  * @c:                  pointer to struct bch_fs
1850  * @id:                 btree to insert into
1851  * @insert_keys:        list of keys to insert
1852  * @hook:               insert callback
1853  */
1854 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1855                       struct bkey_i *k,
1856                       struct disk_reservation *disk_res,
1857                       u64 *journal_seq, int flags)
1858 {
1859         return bch2_trans_do(c, disk_res, journal_seq, flags,
1860                              __bch2_btree_insert(&trans, id, k, 0));
1861 }
1862
1863 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1864                                 unsigned len, unsigned update_flags)
1865 {
1866         struct bkey_i *k;
1867
1868         k = bch2_trans_kmalloc(trans, sizeof(*k));
1869         if (IS_ERR(k))
1870                 return PTR_ERR(k);
1871
1872         bkey_init(&k->k);
1873         k->k.p = iter->pos;
1874         bch2_key_resize(&k->k, len);
1875         return bch2_trans_update(trans, iter, k, update_flags);
1876 }
1877
1878 int bch2_btree_delete_at(struct btree_trans *trans,
1879                          struct btree_iter *iter, unsigned update_flags)
1880 {
1881         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1882 }
1883
1884 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1885                                   enum btree_id btree, struct bpos pos)
1886 {
1887         struct bkey_i *k;
1888
1889         k = bch2_trans_kmalloc(trans, sizeof(*k));
1890         if (IS_ERR(k))
1891                 return PTR_ERR(k);
1892
1893         bkey_init(&k->k);
1894         k->k.p = pos;
1895         return bch2_trans_update_buffered(trans, btree, k);
1896 }
1897
1898 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1899                                   struct bpos start, struct bpos end,
1900                                   unsigned update_flags,
1901                                   u64 *journal_seq)
1902 {
1903         u32 restart_count = trans->restart_count;
1904         struct btree_iter iter;
1905         struct bkey_s_c k;
1906         int ret = 0;
1907
1908         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1909         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1910                 struct disk_reservation disk_res =
1911                         bch2_disk_reservation_init(trans->c, 0);
1912                 struct bkey_i delete;
1913
1914                 ret = bkey_err(k);
1915                 if (ret)
1916                         goto err;
1917
1918                 bkey_init(&delete.k);
1919
1920                 /*
1921                  * This could probably be more efficient for extents:
1922                  */
1923
1924                 /*
1925                  * For extents, iter.pos won't necessarily be the same as
1926                  * bkey_start_pos(k.k) (for non extents they always will be the
1927                  * same). It's important that we delete starting from iter.pos
1928                  * because the range we want to delete could start in the middle
1929                  * of k.
1930                  *
1931                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1932                  * bkey_start_pos(k.k)).
1933                  */
1934                 delete.k.p = iter.pos;
1935
1936                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1937                         bch2_key_resize(&delete.k,
1938                                         bpos_min(end, k.k->p).offset -
1939                                         iter.pos.offset);
1940
1941                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1942                         bch2_trans_commit(trans, &disk_res, journal_seq,
1943                                           BTREE_INSERT_NOFAIL);
1944                 bch2_disk_reservation_put(trans->c, &disk_res);
1945 err:
1946                 /*
1947                  * the bch2_trans_begin() call is in a weird place because we
1948                  * need to call it after every transaction commit, to avoid path
1949                  * overflow, but don't want to call it if the delete operation
1950                  * is a no-op and we have no work to do:
1951                  */
1952                 bch2_trans_begin(trans);
1953
1954                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1955                         ret = 0;
1956                 if (ret)
1957                         break;
1958         }
1959         bch2_trans_iter_exit(trans, &iter);
1960
1961         if (!ret && trans_was_restarted(trans, restart_count))
1962                 ret = -BCH_ERR_transaction_restart_nested;
1963         return ret;
1964 }
1965
1966 /*
1967  * bch_btree_delete_range - delete everything within a given range
1968  *
1969  * Range is a half open interval - [start, end)
1970  */
1971 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1972                             struct bpos start, struct bpos end,
1973                             unsigned update_flags,
1974                             u64 *journal_seq)
1975 {
1976         int ret = bch2_trans_run(c,
1977                         bch2_btree_delete_range_trans(&trans, id, start, end,
1978                                                       update_flags, journal_seq));
1979         if (ret == -BCH_ERR_transaction_restart_nested)
1980                 ret = 0;
1981         return ret;
1982 }
1983
1984 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1985 {
1986         struct printbuf buf = PRINTBUF;
1987         struct jset_entry_log *l;
1988         unsigned u64s;
1989         int ret;
1990
1991         prt_vprintf(&buf, fmt, args);
1992         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
1993         if (ret)
1994                 goto err;
1995
1996         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
1997
1998         ret = darray_make_room(entries, jset_u64s(u64s));
1999         if (ret)
2000                 goto err;
2001
2002         l = (void *) &darray_top(*entries);
2003         l->entry.u64s           = cpu_to_le16(u64s);
2004         l->entry.btree_id       = 0;
2005         l->entry.level          = 1;
2006         l->entry.type           = BCH_JSET_ENTRY_log;
2007         l->entry.pad[0]         = 0;
2008         l->entry.pad[1]         = 0;
2009         l->entry.pad[2]         = 0;
2010         memcpy(l->d, buf.buf, buf.pos);
2011         while (buf.pos & 7)
2012                 l->d[buf.pos++] = '\0';
2013
2014         entries->nr += jset_u64s(u64s);
2015 err:
2016         printbuf_exit(&buf);
2017         return ret;
2018 }
2019
2020 static int
2021 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
2022                   va_list args)
2023 {
2024         int ret;
2025
2026         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
2027                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
2028         } else {
2029                 ret = bch2_trans_do(c, NULL, NULL,
2030                         BTREE_INSERT_LAZY_RW|commit_flags,
2031                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
2032         }
2033
2034         return ret;
2035 }
2036
2037 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
2038 {
2039         va_list args;
2040         int ret;
2041
2042         va_start(args, fmt);
2043         ret = __bch2_fs_log_msg(c, 0, fmt, args);
2044         va_end(args);
2045         return ret;
2046 }
2047
2048 /*
2049  * Use for logging messages during recovery to enable reserved space and avoid
2050  * blocking.
2051  */
2052 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2053 {
2054         va_list args;
2055         int ret;
2056
2057         va_start(args, fmt);
2058         ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
2059         va_end(args);
2060         return ret;
2061 }