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                 /*
 962                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 963                  * flag
 964                  */
 965                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 966                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 967                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 968                         break;
 969                 }
 970
 971                 ret = drop_locks_do(trans,
 972                         bch2_trans_journal_res_get(trans,
 973                                         (flags & BCH_WATERMARK_MASK)|
 974                                         JOURNAL_RES_GET_CHECK));
 975                 break;
 976         case -BCH_ERR_btree_insert_need_journal_reclaim:
 977                 bch2_trans_unlock(trans);
 978
 979                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 980
 981                 wait_event_freezable(c->journal.reclaim_wait,
 982                                      (ret = journal_reclaim_wait_done(c)));
 983                 if (ret < 0)
 984                         break;
 985
 986                 ret = bch2_trans_relock(trans);
 987                 break;
 988         case -BCH_ERR_btree_insert_need_flush_buffer: {
 989                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 990
 991                 ret = 0;
 992
 993                 if (wb->state.nr > wb->size * 3 / 4) {
 994                         bch2_trans_unlock(trans);
 995                         mutex_lock(&wb->flush_lock);
 996
 997                         if (wb->state.nr > wb->size * 3 / 4) {
 998                                 bch2_trans_begin(trans);
 999                                 ret = __bch2_btree_write_buffer_flush(trans,
1000                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
1001                                 if (!ret) {
1002                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1003                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1004                                 }
1005                         } else {
1006                                 mutex_unlock(&wb->flush_lock);
1007                                 ret = bch2_trans_relock(trans);
1008                         }
1009                 }
1010                 break;
1011         }
1012         default:
1013                 BUG_ON(ret >= 0);
1014                 break;
1015         }
1016
1017         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1018
1019         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1020                                 !(flags & BTREE_INSERT_NOWAIT) &&
1021                                 (flags & BTREE_INSERT_NOFAIL), c,
1022                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1023
1024         return ret;
1025 }
1026
1027 static noinline int
1028 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1029 {
1030         struct bch_fs *c = trans->c;
1031         int ret;
1032
1033         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1034             test_bit(BCH_FS_STARTED, &c->flags))
1035                 return -BCH_ERR_erofs_trans_commit;
1036
1037         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1038         if (ret)
1039                 return ret;
1040
1041         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1042         return 0;
1043 }
1044
1045 /*
1046  * This is for updates done in the early part of fsck - btree_gc - before we've
1047  * gone RW. we only add the new key to the list of keys for journal replay to
1048  * do.
1049  */
1050 static noinline int
1051 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1052 {
1053         struct bch_fs *c = trans->c;
1054         struct btree_insert_entry *i;
1055         int ret = 0;
1056
1057         trans_for_each_update(trans, i) {
1058                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1059                 if (ret)
1060                         break;
1061         }
1062
1063         return ret;
1064 }
1065
1066 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1067 {
1068         struct bch_fs *c = trans->c;
1069         struct btree_insert_entry *i = NULL;
1070         struct btree_write_buffered_key *wb;
1071         unsigned u64s;
1072         int ret = 0;
1073
1074         if (!trans->nr_updates &&
1075             !trans->nr_wb_updates &&
1076             !trans->extra_journal_entries.nr)
1077                 goto out_reset;
1078
1079         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1080                 lockdep_assert_held(&c->gc_lock);
1081
1082         ret = bch2_trans_commit_run_triggers(trans);
1083         if (ret)
1084                 goto out_reset;
1085
1086         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1087                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1088                 goto out_reset;
1089         }
1090
1091         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1092             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1093                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1094                 if (ret)
1095                         goto out_reset;
1096         }
1097
1098         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1099             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1100                 bch2_trans_begin(trans);
1101                 bch2_trans_unlock(trans);
1102
1103                 ret = __bch2_btree_write_buffer_flush(trans,
1104                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1105                 if (!ret) {
1106                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1107                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1108                 }
1109                 goto out;
1110         }
1111
1112         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1113
1114         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1115
1116         trans->journal_u64s             = trans->extra_journal_entries.nr;
1117         trans->journal_preres_u64s      = 0;
1118
1119         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1120
1121         if (trans->journal_transaction_names)
1122                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1123
1124         trans_for_each_update(trans, i) {
1125                 EBUG_ON(!i->path->should_be_locked);
1126
1127                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1128                 if (unlikely(ret))
1129                         goto out;
1130
1131                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1132
1133                 if (i->key_cache_already_flushed)
1134                         continue;
1135
1136                 /* we're going to journal the key being updated: */
1137                 u64s = jset_u64s(i->k->k.u64s);
1138                 if (i->cached &&
1139                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1140                         trans->journal_preres_u64s += u64s;
1141
1142                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1143                         continue;
1144
1145                 trans->journal_u64s += u64s;
1146
1147                 /* and we're also going to log the overwrite: */
1148                 if (trans->journal_transaction_names)
1149                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1150         }
1151
1152         trans_for_each_wb_update(trans, wb)
1153                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1154
1155         if (trans->extra_journal_res) {
1156                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1157                                 trans->extra_journal_res,
1158                                 (flags & BTREE_INSERT_NOFAIL)
1159                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1160                 if (ret)
1161                         goto err;
1162         }
1163 retry:
1164         bch2_trans_verify_not_in_restart(trans);
1165         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1166
1167         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1168
1169         /* make sure we didn't drop or screw up locks: */
1170         bch2_trans_verify_locks(trans);
1171
1172         if (ret)
1173                 goto err;
1174
1175         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1176 out:
1177         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1178
1179         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1180                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1181 out_reset:
1182         bch2_trans_reset_updates(trans);
1183
1184         return ret;
1185 err:
1186         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1187         if (ret)
1188                 goto out;
1189
1190         goto retry;
1191 }
1192
1193 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1194                                           enum btree_id id,
1195                                           struct bpos pos)
1196 {
1197         struct bch_fs *c = trans->c;
1198         struct btree_iter iter;
1199         struct bkey_s_c k;
1200         int ret;
1201
1202         bch2_trans_iter_init(trans, &iter, id, pos,
1203                              BTREE_ITER_NOT_EXTENTS|
1204                              BTREE_ITER_ALL_SNAPSHOTS);
1205         while (1) {
1206                 k = bch2_btree_iter_prev(&iter);
1207                 ret = bkey_err(k);
1208                 if (ret)
1209                         break;
1210
1211                 if (!k.k)
1212                         break;
1213
1214                 if (!bkey_eq(pos, k.k->p))
1215                         break;
1216
1217                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1218                         ret = 1;
1219                         break;
1220                 }
1221         }
1222         bch2_trans_iter_exit(trans, &iter);
1223
1224         return ret;
1225 }
1226
1227 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1228                                           enum btree_id id,
1229                                           struct bpos pos)
1230 {
1231         if (!btree_type_has_snapshots(id) ||
1232             !snapshot_t(trans->c, pos.snapshot)->children[0])
1233                 return 0;
1234
1235         return __check_pos_snapshot_overwritten(trans, id, pos);
1236 }
1237
1238 static noinline int extent_front_merge(struct btree_trans *trans,
1239                                        struct btree_iter *iter,
1240                                        struct bkey_s_c k,
1241                                        struct bkey_i **insert,
1242                                        enum btree_update_flags flags)
1243 {
1244         struct bch_fs *c = trans->c;
1245         struct bkey_i *update;
1246         int ret;
1247
1248         update = bch2_bkey_make_mut_noupdate(trans, k);
1249         ret = PTR_ERR_OR_ZERO(update);
1250         if (ret)
1251                 return ret;
1252
1253         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1254                 return 0;
1255
1256         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1257                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1258         if (ret < 0)
1259                 return ret;
1260         if (ret)
1261                 return 0;
1262
1263         ret = bch2_btree_delete_at(trans, iter, flags);
1264         if (ret)
1265                 return ret;
1266
1267         *insert = update;
1268         return 0;
1269 }
1270
1271 static noinline int extent_back_merge(struct btree_trans *trans,
1272                                       struct btree_iter *iter,
1273                                       struct bkey_i *insert,
1274                                       struct bkey_s_c k)
1275 {
1276         struct bch_fs *c = trans->c;
1277         int ret;
1278
1279         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1280                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1281         if (ret < 0)
1282                 return ret;
1283         if (ret)
1284                 return 0;
1285
1286         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1287         return 0;
1288 }
1289
1290 /*
1291  * When deleting, check if we need to emit a whiteout (because we're overwriting
1292  * something in an ancestor snapshot)
1293  */
1294 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1295                                       enum btree_id btree_id, struct bpos pos)
1296 {
1297         struct btree_iter iter;
1298         struct bkey_s_c k;
1299         u32 snapshot = pos.snapshot;
1300         int ret;
1301
1302         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1303                 return 0;
1304
1305         pos.snapshot++;
1306
1307         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1308                            BTREE_ITER_ALL_SNAPSHOTS|
1309                            BTREE_ITER_NOPRESERVE, k, ret) {
1310                 if (!bkey_eq(k.k->p, pos))
1311                         break;
1312
1313                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1314                                               k.k->p.snapshot)) {
1315                         ret = !bkey_whiteout(k.k);
1316                         break;
1317                 }
1318         }
1319         bch2_trans_iter_exit(trans, &iter);
1320
1321         return ret;
1322 }
1323
1324 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
1325                                    enum btree_id id,
1326                                    struct bpos old_pos,
1327                                    struct bpos new_pos)
1328 {
1329         struct bch_fs *c = trans->c;
1330         struct btree_iter old_iter, new_iter;
1331         struct bkey_s_c old_k, new_k;
1332         snapshot_id_list s;
1333         struct bkey_i *update;
1334         int ret;
1335
1336         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
1337                 return 0;
1338
1339         darray_init(&s);
1340
1341         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
1342                              BTREE_ITER_NOT_EXTENTS|
1343                              BTREE_ITER_ALL_SNAPSHOTS);
1344         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
1345                !(ret = bkey_err(old_k)) &&
1346                bkey_eq(old_pos, old_k.k->p)) {
1347                 struct bpos whiteout_pos =
1348                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
1349
1350                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
1351                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
1352                         continue;
1353
1354                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
1355                                            BTREE_ITER_NOT_EXTENTS|
1356                                            BTREE_ITER_INTENT);
1357                 ret = bkey_err(new_k);
1358                 if (ret)
1359                         break;
1360
1361                 if (new_k.k->type == KEY_TYPE_deleted) {
1362                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
1363                         ret = PTR_ERR_OR_ZERO(update);
1364                         if (ret)
1365                                 break;
1366
1367                         bkey_init(&update->k);
1368                         update->k.p             = whiteout_pos;
1369                         update->k.type          = KEY_TYPE_whiteout;
1370
1371                         ret = bch2_trans_update(trans, &new_iter, update,
1372                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1373                 }
1374                 bch2_trans_iter_exit(trans, &new_iter);
1375
1376                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
1377                 if (ret)
1378                         break;
1379         }
1380         bch2_trans_iter_exit(trans, &old_iter);
1381         darray_exit(&s);
1382
1383         return ret;
1384 }
1385
1386 int bch2_trans_update_extent(struct btree_trans *trans,
1387                              struct btree_iter *orig_iter,
1388                              struct bkey_i *insert,
1389                              enum btree_update_flags flags)
1390 {
1391         struct btree_iter iter;
1392         struct bpos start = bkey_start_pos(&insert->k);
1393         struct bkey_i *update;
1394         struct bkey_s_c k;
1395         enum btree_id btree_id = orig_iter->btree_id;
1396         int ret = 0, compressed_sectors;
1397
1398         bch2_trans_iter_init(trans, &iter, btree_id, start,
1399                              BTREE_ITER_INTENT|
1400                              BTREE_ITER_WITH_UPDATES|
1401                              BTREE_ITER_NOT_EXTENTS);
1402         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1403         if ((ret = bkey_err(k)))
1404                 goto err;
1405         if (!k.k)
1406                 goto out;
1407
1408         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1409                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1410                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1411                         if (ret)
1412                                 goto err;
1413                 }
1414
1415                 goto next;
1416         }
1417
1418         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1419                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1420                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1421
1422                 /*
1423                  * If we're going to be splitting a compressed extent, note it
1424                  * so that __bch2_trans_commit() can increase our disk
1425                  * reservation:
1426                  */
1427                 if (((front_split && back_split) ||
1428                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1429                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1430                         trans->extra_journal_res += compressed_sectors;
1431
1432                 if (front_split) {
1433                         update = bch2_bkey_make_mut_noupdate(trans, k);
1434                         if ((ret = PTR_ERR_OR_ZERO(update)))
1435                                 goto err;
1436
1437                         bch2_cut_back(start, update);
1438
1439                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1440                                                 k.k->p, update->k.p) ?:
1441                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1442                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1443                         if (ret)
1444                                 goto err;
1445                 }
1446
1447                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1448                     (front_split || back_split)) {
1449                         update = bch2_bkey_make_mut_noupdate(trans, k);
1450                         if ((ret = PTR_ERR_OR_ZERO(update)))
1451                                 goto err;
1452
1453                         bch2_cut_front(start, update);
1454                         bch2_cut_back(insert->k.p, update);
1455
1456                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1457                                                 k.k->p, update->k.p) ?:
1458                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1459                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1460                         if (ret)
1461                                 goto err;
1462                 }
1463
1464                 if (bkey_le(k.k->p, insert->k.p)) {
1465                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1466                         if ((ret = PTR_ERR_OR_ZERO(update)))
1467                                 goto err;
1468
1469                         bkey_init(&update->k);
1470                         update->k.p = k.k->p;
1471                         update->k.p.snapshot = insert->k.p.snapshot;
1472
1473                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1474                                 update->k.type = KEY_TYPE_whiteout;
1475                         } else if (btree_type_has_snapshots(btree_id)) {
1476                                 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1477                                 if (ret < 0)
1478                                         goto err;
1479                                 if (ret)
1480                                         update->k.type = KEY_TYPE_whiteout;
1481                         }
1482
1483                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1484                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1485                         if (ret)
1486                                 goto err;
1487                 }
1488
1489                 if (back_split) {
1490                         update = bch2_bkey_make_mut_noupdate(trans, k);
1491                         if ((ret = PTR_ERR_OR_ZERO(update)))
1492                                 goto err;
1493
1494                         bch2_cut_front(insert->k.p, update);
1495
1496                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1497                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1498                                                   flags);
1499                         if (ret)
1500                                 goto err;
1501                         goto out;
1502                 }
1503 next:
1504                 bch2_btree_iter_advance(&iter);
1505                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1506                 if ((ret = bkey_err(k)))
1507                         goto err;
1508                 if (!k.k)
1509                         goto out;
1510         }
1511
1512         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1513                 ret = extent_back_merge(trans, &iter, insert, k);
1514                 if (ret)
1515                         goto err;
1516         }
1517 out:
1518         if (!bkey_deleted(&insert->k)) {
1519                 /*
1520                  * Rewinding iterators is expensive: get a new one and the one
1521                  * that points to the start of insert will be cloned from:
1522                  */
1523                 bch2_trans_iter_exit(trans, &iter);
1524                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1525                                      BTREE_ITER_NOT_EXTENTS|
1526                                      BTREE_ITER_INTENT);
1527                 ret   = bch2_btree_iter_traverse(&iter) ?:
1528                         bch2_trans_update(trans, &iter, insert, flags);
1529         }
1530 err:
1531         bch2_trans_iter_exit(trans, &iter);
1532
1533         return ret;
1534 }
1535
1536 static int __must_check
1537 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1538                                 struct bkey_i *k, enum btree_update_flags flags,
1539                                 unsigned long ip);
1540
1541 static noinline int flush_new_cached_update(struct btree_trans *trans,
1542                                             struct btree_path *path,
1543                                             struct btree_insert_entry *i,
1544                                             enum btree_update_flags flags,
1545                                             unsigned long ip)
1546 {
1547         struct btree_path *btree_path;
1548         struct bkey k;
1549         int ret;
1550
1551         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1552                                    BTREE_ITER_INTENT, _THIS_IP_);
1553         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1554         if (ret)
1555                 goto out;
1556
1557         /*
1558          * The old key in the insert entry might actually refer to an existing
1559          * key in the btree that has been deleted from cache and not yet
1560          * flushed. Check for this and skip the flush so we don't run triggers
1561          * against a stale key.
1562          */
1563         bch2_btree_path_peek_slot_exact(btree_path, &k);
1564         if (!bkey_deleted(&k))
1565                 goto out;
1566
1567         i->key_cache_already_flushed = true;
1568         i->flags |= BTREE_TRIGGER_NORUN;
1569
1570         btree_path_set_should_be_locked(btree_path);
1571         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1572 out:
1573         bch2_path_put(trans, btree_path, true);
1574         return ret;
1575 }
1576
1577 static int __must_check
1578 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1579                                 struct bkey_i *k, enum btree_update_flags flags,
1580                                 unsigned long ip)
1581 {
1582         struct bch_fs *c = trans->c;
1583         struct btree_insert_entry *i, n;
1584         int cmp;
1585
1586         EBUG_ON(!path->should_be_locked);
1587         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1588         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1589
1590         n = (struct btree_insert_entry) {
1591                 .flags          = flags,
1592                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1593                 .btree_id       = path->btree_id,
1594                 .level          = path->level,
1595                 .cached         = path->cached,
1596                 .path           = path,
1597                 .k              = k,
1598                 .ip_allocated   = ip,
1599         };
1600
1601 #ifdef CONFIG_BCACHEFS_DEBUG
1602         trans_for_each_update(trans, i)
1603                 BUG_ON(i != trans->updates &&
1604                        btree_insert_entry_cmp(i - 1, i) >= 0);
1605 #endif
1606
1607         /*
1608          * Pending updates are kept sorted: first, find position of new update,
1609          * then delete/trim any updates the new update overwrites:
1610          */
1611         trans_for_each_update(trans, i) {
1612                 cmp = btree_insert_entry_cmp(&n, i);
1613                 if (cmp <= 0)
1614                         break;
1615         }
1616
1617         if (!cmp && i < trans->updates + trans->nr_updates) {
1618                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1619
1620                 bch2_path_put(trans, i->path, true);
1621                 i->flags        = n.flags;
1622                 i->cached       = n.cached;
1623                 i->k            = n.k;
1624                 i->path         = n.path;
1625                 i->ip_allocated = n.ip_allocated;
1626         } else {
1627                 array_insert_item(trans->updates, trans->nr_updates,
1628                                   i - trans->updates, n);
1629
1630                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1631                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1632
1633                 if (unlikely(trans->journal_replay_not_finished)) {
1634                         struct bkey_i *j_k =
1635                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1636
1637                         if (j_k) {
1638                                 i->old_k = j_k->k;
1639                                 i->old_v = &j_k->v;
1640                         }
1641                 }
1642         }
1643
1644         __btree_path_get(i->path, true);
1645
1646         /*
1647          * If a key is present in the key cache, it must also exist in the
1648          * btree - this is necessary for cache coherency. When iterating over
1649          * a btree that's cached in the key cache, the btree iter code checks
1650          * the key cache - but the key has to exist in the btree for that to
1651          * work:
1652          */
1653         if (path->cached && bkey_deleted(&i->old_k))
1654                 return flush_new_cached_update(trans, path, i, flags, ip);
1655
1656         return 0;
1657 }
1658
1659 static inline int __must_check
1660 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1661                           struct bkey_i *k, enum btree_update_flags flags)
1662 {
1663         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1664 }
1665
1666 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1667                                    struct bkey_i *k, enum btree_update_flags flags)
1668 {
1669         struct btree_path *path = iter->update_path ?: iter->path;
1670         struct bkey_cached *ck;
1671         int ret;
1672
1673         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1674                 return bch2_trans_update_extent(trans, iter, k, flags);
1675
1676         if (bkey_deleted(&k->k) &&
1677             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1678             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1679                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1680                 if (unlikely(ret < 0))
1681                         return ret;
1682
1683                 if (ret)
1684                         k->k.type = KEY_TYPE_whiteout;
1685         }
1686
1687         /*
1688          * Ensure that updates to cached btrees go to the key cache:
1689          */
1690         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1691             !path->cached &&
1692             !path->level &&
1693             btree_id_cached(trans->c, path->btree_id)) {
1694                 if (!iter->key_cache_path ||
1695                     !iter->key_cache_path->should_be_locked ||
1696                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1697                         if (!iter->key_cache_path)
1698                                 iter->key_cache_path =
1699                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1700                                                       BTREE_ITER_INTENT|
1701                                                       BTREE_ITER_CACHED, _THIS_IP_);
1702
1703                         iter->key_cache_path =
1704                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1705                                                         iter->flags & BTREE_ITER_INTENT,
1706                                                         _THIS_IP_);
1707
1708                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1709                                                        BTREE_ITER_CACHED);
1710                         if (unlikely(ret))
1711                                 return ret;
1712
1713                         ck = (void *) iter->key_cache_path->l[0].b;
1714
1715                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1716                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1717                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1718                         }
1719
1720                         btree_path_set_should_be_locked(iter->key_cache_path);
1721                 }
1722
1723                 path = iter->key_cache_path;
1724         }
1725
1726         return bch2_trans_update_by_path(trans, path, k, flags);
1727 }
1728
1729 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1730                                             enum btree_id btree,
1731                                             struct bkey_i *k)
1732 {
1733         struct btree_write_buffered_key *i;
1734         int ret;
1735
1736         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1737         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1738
1739         trans_for_each_wb_update(trans, i) {
1740                 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
1741                         bkey_copy(&i->k, k);
1742                         return 0;
1743                 }
1744         }
1745
1746         if (!trans->wb_updates ||
1747             trans->nr_wb_updates == trans->wb_updates_size) {
1748                 struct btree_write_buffered_key *u;
1749
1750                 if (trans->nr_wb_updates == trans->wb_updates_size) {
1751                         struct btree_transaction_stats *s = btree_trans_stats(trans);
1752
1753                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1754                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1755                         if (s)
1756                                 s->wb_updates_size = trans->wb_updates_size;
1757                 }
1758
1759                 u = bch2_trans_kmalloc_nomemzero(trans,
1760                                         trans->wb_updates_size *
1761                                         sizeof(struct btree_write_buffered_key));
1762                 ret = PTR_ERR_OR_ZERO(u);
1763                 if (ret)
1764                         return ret;
1765
1766                 if (trans->nr_wb_updates)
1767                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1768                                sizeof(struct btree_write_buffered_key));
1769                 trans->wb_updates = u;
1770         }
1771
1772         trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
1773                 .btree  = btree,
1774         };
1775
1776         bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
1777         trans->nr_wb_updates++;
1778
1779         return 0;
1780 }
1781
1782 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1783                              enum btree_id btree, struct bpos end)
1784 {
1785         struct bkey_s_c k;
1786         int ret = 0;
1787
1788         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1789         k = bch2_btree_iter_prev(iter);
1790         ret = bkey_err(k);
1791         if (ret)
1792                 goto err;
1793
1794         bch2_btree_iter_advance(iter);
1795         k = bch2_btree_iter_peek_slot(iter);
1796         ret = bkey_err(k);
1797         if (ret)
1798                 goto err;
1799
1800         BUG_ON(k.k->type != KEY_TYPE_deleted);
1801
1802         if (bkey_gt(k.k->p, end)) {
1803                 ret = -BCH_ERR_ENOSPC_btree_slot;
1804                 goto err;
1805         }
1806
1807         return 0;
1808 err:
1809         bch2_trans_iter_exit(trans, iter);
1810         return ret;
1811 }
1812
1813 void bch2_trans_commit_hook(struct btree_trans *trans,
1814                             struct btree_trans_commit_hook *h)
1815 {
1816         h->next = trans->hooks;
1817         trans->hooks = h;
1818 }
1819
1820 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1821                                 enum btree_id btree, struct bkey_i *k,
1822                                 enum btree_update_flags flags)
1823 {
1824         struct btree_iter iter;
1825         int ret;
1826
1827         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1828                              BTREE_ITER_NOT_EXTENTS|
1829                              BTREE_ITER_INTENT);
1830         ret   = bch2_btree_iter_traverse(&iter) ?:
1831                 bch2_trans_update(trans, &iter, k, flags);
1832         bch2_trans_iter_exit(trans, &iter);
1833         return ret;
1834 }
1835
1836 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1837                         struct bkey_i *k, enum btree_update_flags flags)
1838 {
1839         struct btree_iter iter;
1840         int ret;
1841
1842         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1843                              BTREE_ITER_CACHED|
1844                              BTREE_ITER_INTENT);
1845         ret   = bch2_btree_iter_traverse(&iter) ?:
1846                 bch2_trans_update(trans, &iter, k, flags);
1847         bch2_trans_iter_exit(trans, &iter);
1848         return ret;
1849 }
1850
1851 /**
1852  * bch2_btree_insert - insert keys into the extent btree
1853  * @c:                  pointer to struct bch_fs
1854  * @id:                 btree to insert into
1855  * @insert_keys:        list of keys to insert
1856  * @hook:               insert callback
1857  */
1858 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1859                       struct bkey_i *k,
1860                       struct disk_reservation *disk_res,
1861                       u64 *journal_seq, int flags)
1862 {
1863         return bch2_trans_do(c, disk_res, journal_seq, flags,
1864                              __bch2_btree_insert(&trans, id, k, 0));
1865 }
1866
1867 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1868                                 unsigned len, unsigned update_flags)
1869 {
1870         struct bkey_i *k;
1871
1872         k = bch2_trans_kmalloc(trans, sizeof(*k));
1873         if (IS_ERR(k))
1874                 return PTR_ERR(k);
1875
1876         bkey_init(&k->k);
1877         k->k.p = iter->pos;
1878         bch2_key_resize(&k->k, len);
1879         return bch2_trans_update(trans, iter, k, update_flags);
1880 }
1881
1882 int bch2_btree_delete_at(struct btree_trans *trans,
1883                          struct btree_iter *iter, unsigned update_flags)
1884 {
1885         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1886 }
1887
1888 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1889                                   enum btree_id btree, struct bpos pos)
1890 {
1891         struct bkey_i *k;
1892
1893         k = bch2_trans_kmalloc(trans, sizeof(*k));
1894         if (IS_ERR(k))
1895                 return PTR_ERR(k);
1896
1897         bkey_init(&k->k);
1898         k->k.p = pos;
1899         return bch2_trans_update_buffered(trans, btree, k);
1900 }
1901
1902 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1903                                   struct bpos start, struct bpos end,
1904                                   unsigned update_flags,
1905                                   u64 *journal_seq)
1906 {
1907         u32 restart_count = trans->restart_count;
1908         struct btree_iter iter;
1909         struct bkey_s_c k;
1910         int ret = 0;
1911
1912         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1913         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1914                 struct disk_reservation disk_res =
1915                         bch2_disk_reservation_init(trans->c, 0);
1916                 struct bkey_i delete;
1917
1918                 ret = bkey_err(k);
1919                 if (ret)
1920                         goto err;
1921
1922                 bkey_init(&delete.k);
1923
1924                 /*
1925                  * This could probably be more efficient for extents:
1926                  */
1927
1928                 /*
1929                  * For extents, iter.pos won't necessarily be the same as
1930                  * bkey_start_pos(k.k) (for non extents they always will be the
1931                  * same). It's important that we delete starting from iter.pos
1932                  * because the range we want to delete could start in the middle
1933                  * of k.
1934                  *
1935                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1936                  * bkey_start_pos(k.k)).
1937                  */
1938                 delete.k.p = iter.pos;
1939
1940                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1941                         bch2_key_resize(&delete.k,
1942                                         bpos_min(end, k.k->p).offset -
1943                                         iter.pos.offset);
1944
1945                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1946                         bch2_trans_commit(trans, &disk_res, journal_seq,
1947                                           BTREE_INSERT_NOFAIL);
1948                 bch2_disk_reservation_put(trans->c, &disk_res);
1949 err:
1950                 /*
1951                  * the bch2_trans_begin() call is in a weird place because we
1952                  * need to call it after every transaction commit, to avoid path
1953                  * overflow, but don't want to call it if the delete operation
1954                  * is a no-op and we have no work to do:
1955                  */
1956                 bch2_trans_begin(trans);
1957
1958                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1959                         ret = 0;
1960                 if (ret)
1961                         break;
1962         }
1963         bch2_trans_iter_exit(trans, &iter);
1964
1965         if (!ret && trans_was_restarted(trans, restart_count))
1966                 ret = -BCH_ERR_transaction_restart_nested;
1967         return ret;
1968 }
1969
1970 /*
1971  * bch_btree_delete_range - delete everything within a given range
1972  *
1973  * Range is a half open interval - [start, end)
1974  */
1975 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1976                             struct bpos start, struct bpos end,
1977                             unsigned update_flags,
1978                             u64 *journal_seq)
1979 {
1980         int ret = bch2_trans_run(c,
1981                         bch2_btree_delete_range_trans(&trans, id, start, end,
1982                                                       update_flags, journal_seq));
1983         if (ret == -BCH_ERR_transaction_restart_nested)
1984                 ret = 0;
1985         return ret;
1986 }
1987
1988 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
1989 {
1990         struct printbuf buf = PRINTBUF;
1991         struct jset_entry_log *l;
1992         unsigned u64s;
1993         int ret;
1994
1995         prt_vprintf(&buf, fmt, args);
1996         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
1997         if (ret)
1998                 goto err;
1999
2000         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
2001
2002         ret = darray_make_room(entries, jset_u64s(u64s));
2003         if (ret)
2004                 goto err;
2005
2006         l = (void *) &darray_top(*entries);
2007         l->entry.u64s           = cpu_to_le16(u64s);
2008         l->entry.btree_id       = 0;
2009         l->entry.level          = 1;
2010         l->entry.type           = BCH_JSET_ENTRY_log;
2011         l->entry.pad[0]         = 0;
2012         l->entry.pad[1]         = 0;
2013         l->entry.pad[2]         = 0;
2014         memcpy(l->d, buf.buf, buf.pos);
2015         while (buf.pos & 7)
2016                 l->d[buf.pos++] = '\0';
2017
2018         entries->nr += jset_u64s(u64s);
2019 err:
2020         printbuf_exit(&buf);
2021         return ret;
2022 }
2023
2024 static int
2025 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
2026                   va_list args)
2027 {
2028         int ret;
2029
2030         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
2031                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
2032         } else {
2033                 ret = bch2_trans_do(c, NULL, NULL,
2034                         BTREE_INSERT_LAZY_RW|commit_flags,
2035                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
2036         }
2037
2038         return ret;
2039 }
2040
2041 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
2042 {
2043         va_list args;
2044         int ret;
2045
2046         va_start(args, fmt);
2047         ret = __bch2_fs_log_msg(c, 0, fmt, args);
2048         va_end(args);
2049         return ret;
2050 }
2051
2052 /*
2053  * Use for logging messages during recovery to enable reserved space and avoid
2054  * blocking.
2055  */
2056 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2057 {
2058         va_list args;
2059         int ret;
2060
2061         va_start(args, fmt);
2062         ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
2063         va_end(args);
2064         return ret;
2065 }