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 "buckets.h"
  12 #include "debug.h"
  13 #include "errcode.h"
  14 #include "error.h"
  15 #include "extent_update.h"
  16 #include "journal.h"
  17 #include "journal_reclaim.h"
  18 #include "keylist.h"
  19 #include "recovery.h"
  20 #include "subvolume.h"
  21 #include "replicas.h"
  22
  23 #include <linux/prefetch.h>
  24 #include <linux/sort.h>
  25 #include <trace/events/bcachefs.h>
  26
  27 static int __must_check
  28 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  29                           struct bkey_i *, enum btree_update_flags);
  30
  31 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  32                                          const struct btree_insert_entry *r)
  33 {
  34         return   cmp_int(l->btree_id,   r->btree_id) ?:
  35                  cmp_int(l->cached,     r->cached) ?:
  36                  -cmp_int(l->level,     r->level) ?:
  37                  bpos_cmp(l->k->k.p,    r->k->k.p);
  38 }
  39
  40 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  41 {
  42         return i->path->l + i->level;
  43 }
  44
  45 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  46                                      struct btree_insert_entry *i)
  47 {
  48         return i != trans->updates &&
  49                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  50 }
  51
  52 static inline bool same_leaf_as_next(struct btree_trans *trans,
  53                                      struct btree_insert_entry *i)
  54 {
  55         return i + 1 < trans->updates + trans->nr_updates &&
  56                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  57 }
  58
  59 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  60                                                   struct btree_path *path,
  61                                                   struct btree *b)
  62 {
  63         struct bch_fs *c = trans->c;
  64
  65         if (path->cached)
  66                 return;
  67
  68         if (unlikely(btree_node_just_written(b)) &&
  69             bch2_btree_post_write_cleanup(c, b))
  70                 bch2_trans_node_reinit_iter(trans, b);
  71
  72         /*
  73          * If the last bset has been written, or if it's gotten too big - start
  74          * a new bset to insert into:
  75          */
  76         if (want_new_bset(c, b))
  77                 bch2_btree_init_next(trans, b);
  78 }
  79
  80 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
  81                                      struct btree_path *path,
  82                                      struct btree *b)
  83 {
  84         bch2_btree_node_lock_write(trans, path, b);
  85         bch2_btree_node_prep_for_write(trans, path, b);
  86 }
  87
  88 /* Inserting into a given leaf node (last stage of insert): */
  89
  90 /* Handle overwrites and do insert, for non extents: */
  91 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  92                                 struct btree_path *path,
  93                                 struct btree *b,
  94                                 struct btree_node_iter *node_iter,
  95                                 struct bkey_i *insert)
  96 {
  97         struct bkey_packed *k;
  98         unsigned clobber_u64s = 0, new_u64s = 0;
  99
 100         EBUG_ON(btree_node_just_written(b));
 101         EBUG_ON(bset_written(b, btree_bset_last(b)));
 102         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 103         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
 104         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
 105         EBUG_ON(insert->k.u64s >
 106                 bch_btree_keys_u64s_remaining(trans->c, b));
 107
 108         k = bch2_btree_node_iter_peek_all(node_iter, b);
 109         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 110                 k = NULL;
 111
 112         /* @k is the key being overwritten/deleted, if any: */
 113         EBUG_ON(k && bkey_deleted(k));
 114
 115         /* Deleting, but not found? nothing to do: */
 116         if (bkey_deleted(&insert->k) && !k)
 117                 return false;
 118
 119         if (bkey_deleted(&insert->k)) {
 120                 /* Deleting: */
 121                 btree_account_key_drop(b, k);
 122                 k->type = KEY_TYPE_deleted;
 123
 124                 if (k->needs_whiteout)
 125                         push_whiteout(trans->c, b, insert->k.p);
 126                 k->needs_whiteout = false;
 127
 128                 if (k >= btree_bset_last(b)->start) {
 129                         clobber_u64s = k->u64s;
 130                         bch2_bset_delete(b, k, clobber_u64s);
 131                         goto fix_iter;
 132                 } else {
 133                         bch2_btree_path_fix_key_modified(trans, b, k);
 134                 }
 135
 136                 return true;
 137         }
 138
 139         if (k) {
 140                 /* Overwriting: */
 141                 btree_account_key_drop(b, k);
 142                 k->type = KEY_TYPE_deleted;
 143
 144                 insert->k.needs_whiteout = k->needs_whiteout;
 145                 k->needs_whiteout = false;
 146
 147                 if (k >= btree_bset_last(b)->start) {
 148                         clobber_u64s = k->u64s;
 149                         goto overwrite;
 150                 } else {
 151                         bch2_btree_path_fix_key_modified(trans, b, k);
 152                 }
 153         }
 154
 155         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 156 overwrite:
 157         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 158         new_u64s = k->u64s;
 159 fix_iter:
 160         if (clobber_u64s != new_u64s)
 161                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 162                                          clobber_u64s, new_u64s);
 163         return true;
 164 }
 165
 166 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 167                                unsigned i, u64 seq)
 168 {
 169         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 170         struct btree_write *w = container_of(pin, struct btree_write, journal);
 171         struct btree *b = container_of(w, struct btree, writes[i]);
 172         unsigned long old, new, v;
 173         unsigned idx = w - b->writes;
 174
 175         six_lock_read(&b->c.lock, NULL, NULL);
 176         v = READ_ONCE(b->flags);
 177
 178         do {
 179                 old = new = v;
 180
 181                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 182                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 183                     w->journal.seq != seq)
 184                         break;
 185
 186                 new |= 1 << BTREE_NODE_need_write;
 187         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 188
 189         btree_node_write_if_need(c, b, SIX_LOCK_read);
 190         six_unlock_read(&b->c.lock);
 191         return 0;
 192 }
 193
 194 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 195 {
 196         return __btree_node_flush(j, pin, 0, seq);
 197 }
 198
 199 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 200 {
 201         return __btree_node_flush(j, pin, 1, seq);
 202 }
 203
 204 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 205                                        struct btree *b, u64 seq)
 206 {
 207         struct btree_write *w = btree_current_write(b);
 208
 209         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 210                              btree_node_write_idx(b) == 0
 211                              ? btree_node_flush0
 212                              : btree_node_flush1);
 213 }
 214
 215 /**
 216  * btree_insert_key - insert a key one key into a leaf node
 217  */
 218 static void btree_insert_key_leaf(struct btree_trans *trans,
 219                                   struct btree_insert_entry *insert)
 220 {
 221         struct bch_fs *c = trans->c;
 222         struct btree *b = insert_l(insert)->b;
 223         struct bset_tree *t = bset_tree_last(b);
 224         struct bset *i = bset(b, t);
 225         int old_u64s = bset_u64s(t);
 226         int old_live_u64s = b->nr.live_u64s;
 227         int live_u64s_added, u64s_added;
 228
 229         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 230                                         &insert_l(insert)->iter, insert->k)))
 231                 return;
 232
 233         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 234                                          le64_to_cpu(i->journal_seq)));
 235
 236         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 237
 238         if (unlikely(!btree_node_dirty(b)))
 239                 set_btree_node_dirty_acct(c, b);
 240
 241         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 242         u64s_added = (int) bset_u64s(t) - old_u64s;
 243
 244         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 245                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 246         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 247                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 248
 249         if (u64s_added > live_u64s_added &&
 250             bch2_maybe_compact_whiteouts(c, b))
 251                 bch2_trans_node_reinit_iter(trans, b);
 252 }
 253
 254 /* Cached btree updates: */
 255
 256 /* Normal update interface: */
 257
 258 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 259                                              struct btree_insert_entry *i)
 260 {
 261         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 262         BUG_ON(i->cached        != i->path->cached);
 263         BUG_ON(i->level         != i->path->level);
 264         BUG_ON(i->btree_id      != i->path->btree_id);
 265         EBUG_ON(!i->level &&
 266                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 267                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 268                 i->k->k.p.snapshot &&
 269                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 270 }
 271
 272 static noinline int
 273 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 274                                    unsigned long trace_ip)
 275 {
 276         struct bch_fs *c = trans->c;
 277         int ret;
 278
 279         bch2_trans_unlock(trans);
 280
 281         ret = bch2_journal_preres_get(&c->journal,
 282                         &trans->journal_preres, u64s, 0);
 283         if (ret)
 284                 return ret;
 285
 286         ret = bch2_trans_relock(trans);
 287         if (ret) {
 288                 trace_trans_restart_journal_preres_get(trans, trace_ip);
 289                 return ret;
 290         }
 291
 292         return 0;
 293 }
 294
 295 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 296                                              unsigned flags)
 297 {
 298         struct bch_fs *c = trans->c;
 299         int ret;
 300
 301         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 302                                    trans->journal_u64s,
 303                                    flags|
 304                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 305
 306         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 307 }
 308
 309 #define JSET_ENTRY_LOG_U64s             4
 310
 311 static void journal_transaction_name(struct btree_trans *trans)
 312 {
 313         struct bch_fs *c = trans->c;
 314         struct journal *j = &c->journal;
 315         struct jset_entry *entry =
 316                 bch2_journal_add_entry(j, &trans->journal_res,
 317                                        BCH_JSET_ENTRY_log, 0, 0,
 318                                        JSET_ENTRY_LOG_U64s);
 319         struct jset_entry_log *l =
 320                 container_of(entry, struct jset_entry_log, entry);
 321
 322         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 323 }
 324
 325 static inline enum btree_insert_ret
 326 btree_key_can_insert(struct btree_trans *trans,
 327                      struct btree *b,
 328                      unsigned u64s)
 329 {
 330         struct bch_fs *c = trans->c;
 331
 332         if (!bch2_btree_node_insert_fits(c, b, u64s))
 333                 return BTREE_INSERT_BTREE_NODE_FULL;
 334
 335         return BTREE_INSERT_OK;
 336 }
 337
 338 static enum btree_insert_ret
 339 btree_key_can_insert_cached(struct btree_trans *trans,
 340                             struct btree_path *path,
 341                             unsigned u64s)
 342 {
 343         struct bch_fs *c = trans->c;
 344         struct bkey_cached *ck = (void *) path->l[0].b;
 345         unsigned old_u64s = ck->u64s, new_u64s;
 346         struct bkey_i *new_k;
 347
 348         EBUG_ON(path->level);
 349
 350         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 351             bch2_btree_key_cache_must_wait(c) &&
 352             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 353                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 354
 355         /*
 356          * bch2_varint_decode can read past the end of the buffer by at most 7
 357          * bytes (it won't be used):
 358          */
 359         u64s += 1;
 360
 361         if (u64s <= ck->u64s)
 362                 return BTREE_INSERT_OK;
 363
 364         new_u64s        = roundup_pow_of_two(u64s);
 365         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 366         if (!new_k) {
 367                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 368                         bch2_btree_ids[path->btree_id], new_u64s);
 369                 return -ENOMEM;
 370         }
 371
 372         ck->u64s        = new_u64s;
 373         ck->k           = new_k;
 374         /*
 375          * Keys returned by peek() are no longer valid pointers, so we need a
 376          * transaction restart:
 377          */
 378         trace_trans_restart_key_cache_key_realloced(trans, _RET_IP_, path, old_u64s, new_u64s);
 379         return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_key_cache_realloced);
 380 }
 381
 382 /* Triggers: */
 383
 384 static int run_one_mem_trigger(struct btree_trans *trans,
 385                                struct btree_insert_entry *i,
 386                                unsigned flags)
 387 {
 388         struct bkey_s_c old = { &i->old_k, i->old_v };
 389         struct bkey_i *new = i->k;
 390         int ret;
 391
 392         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 393                 return 0;
 394
 395         if (!btree_node_type_needs_gc(i->btree_id))
 396                 return 0;
 397
 398         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 399             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 400             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 401                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 402                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 403         } else {
 404                 struct bkey             _deleted = KEY(0, 0, 0);
 405                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 406
 407                 _deleted.p = i->path->pos;
 408
 409                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 410                                 BTREE_TRIGGER_INSERT|flags) ?:
 411                         bch2_mark_key(trans, old, deleted,
 412                                 BTREE_TRIGGER_OVERWRITE|flags);
 413         }
 414
 415         return ret;
 416 }
 417
 418 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 419                                  bool overwrite)
 420 {
 421         /*
 422          * Transactional triggers create new btree_insert_entries, so we can't
 423          * pass them a pointer to a btree_insert_entry, that memory is going to
 424          * move:
 425          */
 426         struct bkey old_k = i->old_k;
 427         struct bkey_s_c old = { &old_k, i->old_v };
 428
 429         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 430             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 431                 return 0;
 432
 433         if (!i->insert_trigger_run &&
 434             !i->overwrite_trigger_run &&
 435             bch2_bkey_ops[old.k->type].trans_trigger ==
 436             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 437             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 438                 i->overwrite_trigger_run = true;
 439                 i->insert_trigger_run = true;
 440                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 441                                            BTREE_TRIGGER_INSERT|
 442                                            BTREE_TRIGGER_OVERWRITE|
 443                                            i->flags) ?: 1;
 444         } else if (overwrite && !i->overwrite_trigger_run) {
 445                 i->overwrite_trigger_run = true;
 446                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 447         } else if (!overwrite && !i->insert_trigger_run) {
 448                 i->insert_trigger_run = true;
 449                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 450         } else {
 451                 return 0;
 452         }
 453 }
 454
 455 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 456                               struct btree_insert_entry *btree_id_start)
 457 {
 458         struct btree_insert_entry *i;
 459         bool trans_trigger_run;
 460         int ret, overwrite;
 461
 462         for (overwrite = 1; overwrite >= 0; --overwrite) {
 463
 464                 /*
 465                  * Running triggers will append more updates to the list of updates as
 466                  * we're walking it:
 467                  */
 468                 do {
 469                         trans_trigger_run = false;
 470
 471                         for (i = btree_id_start;
 472                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 473                              i++) {
 474                                 if (i->btree_id != btree_id)
 475                                         continue;
 476
 477                                 ret = run_one_trans_trigger(trans, i, overwrite);
 478                                 if (ret < 0)
 479                                         return ret;
 480                                 if (ret)
 481                                         trans_trigger_run = true;
 482                         }
 483                 } while (trans_trigger_run);
 484         }
 485
 486         return 0;
 487 }
 488
 489 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 490 {
 491         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 492         unsigned btree_id = 0;
 493         int ret = 0;
 494
 495         /*
 496          *
 497          * For a given btree, this algorithm runs insert triggers before
 498          * overwrite triggers: this is so that when extents are being moved
 499          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 500          * they are re-added.
 501          */
 502         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 503                 if (btree_id == BTREE_ID_alloc)
 504                         continue;
 505
 506                 while (btree_id_start < trans->updates + trans->nr_updates &&
 507                        btree_id_start->btree_id < btree_id)
 508                         btree_id_start++;
 509
 510                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 511                 if (ret)
 512                         return ret;
 513         }
 514
 515         trans_for_each_update(trans, i) {
 516                 if (i->btree_id > BTREE_ID_alloc)
 517                         break;
 518                 if (i->btree_id == BTREE_ID_alloc) {
 519                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 520                         if (ret)
 521                                 return ret;
 522                         break;
 523                 }
 524         }
 525
 526         trans_for_each_update(trans, i)
 527                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 528                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 529                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 530
 531         return 0;
 532 }
 533
 534 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 535 {
 536         struct bch_fs *c = trans->c;
 537         struct btree_insert_entry *i;
 538         int ret = 0;
 539
 540         trans_for_each_update(trans, i) {
 541                 /*
 542                  * XXX: synchronization of cached update triggers with gc
 543                  * XXX: synchronization of interior node updates with gc
 544                  */
 545                 BUG_ON(i->cached || i->level);
 546
 547                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 548                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 549                         if (ret)
 550                                 break;
 551                 }
 552         }
 553
 554         return ret;
 555 }
 556
 557 static inline int
 558 bch2_trans_commit_write_locked(struct btree_trans *trans,
 559                                struct btree_insert_entry **stopped_at,
 560                                unsigned long trace_ip)
 561 {
 562         struct bch_fs *c = trans->c;
 563         struct btree_insert_entry *i;
 564         struct btree_trans_commit_hook *h;
 565         unsigned u64s = 0;
 566         bool marking = false;
 567         int ret;
 568
 569         if (race_fault()) {
 570                 trace_trans_restart_fault_inject(trans, trace_ip);
 571                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 572         }
 573
 574         /*
 575          * Check if the insert will fit in the leaf node with the write lock
 576          * held, otherwise another thread could write the node changing the
 577          * amount of space available:
 578          */
 579
 580         prefetch(&trans->c->journal.flags);
 581
 582         h = trans->hooks;
 583         while (h) {
 584                 ret = h->fn(trans, h);
 585                 if (ret)
 586                         return ret;
 587                 h = h->next;
 588         }
 589
 590         trans_for_each_update(trans, i) {
 591                 /* Multiple inserts might go to same leaf: */
 592                 if (!same_leaf_as_prev(trans, i))
 593                         u64s = 0;
 594
 595                 u64s += i->k->k.u64s;
 596                 ret = !i->cached
 597                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 598                         : btree_key_can_insert_cached(trans, i->path, u64s);
 599                 if (ret) {
 600                         *stopped_at = i;
 601                         return ret;
 602                 }
 603
 604                 if (btree_node_type_needs_gc(i->bkey_type))
 605                         marking = true;
 606
 607                 /*
 608                  * Revalidate before calling mem triggers - XXX, ugly:
 609                  *
 610                  * - successful btree node splits don't cause transaction
 611                  *   restarts and will have invalidated the pointer to the bkey
 612                  *   value
 613                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 614                  *   when it has to resort
 615                  * - btree_key_can_insert_cached() when it has to reallocate
 616                  *
 617                  *   Ugly because we currently have no way to tell if the
 618                  *   pointer's been invalidated, which means it's debatabale
 619                  *   whether we should be stashing the old key at all.
 620                  */
 621                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 622
 623                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 624                         struct bkey_i *j_k =
 625                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 626                                                             i->k->k.p);
 627
 628                         if (j_k) {
 629                                 i->old_k = j_k->k;
 630                                 i->old_v = &j_k->v;
 631                         }
 632                 }
 633         }
 634
 635         /*
 636          * Don't get journal reservation until after we know insert will
 637          * succeed:
 638          */
 639         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 640                 ret = bch2_trans_journal_res_get(trans,
 641                                 JOURNAL_RES_GET_NONBLOCK);
 642                 if (ret)
 643                         return ret;
 644
 645                 journal_transaction_name(trans);
 646         } else {
 647                 trans->journal_res.seq = c->journal.replay_journal_seq;
 648         }
 649
 650         if (unlikely(trans->extra_journal_entries.nr)) {
 651                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 652                                   trans->extra_journal_entries.data,
 653                                   trans->extra_journal_entries.nr);
 654
 655                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 656                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 657         }
 658
 659         /*
 660          * Not allowed to fail after we've gotten our journal reservation - we
 661          * have to use it:
 662          */
 663
 664         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 665                 if (bch2_journal_seq_verify)
 666                         trans_for_each_update(trans, i)
 667                                 i->k->k.version.lo = trans->journal_res.seq;
 668                 else if (bch2_inject_invalid_keys)
 669                         trans_for_each_update(trans, i)
 670                                 i->k->k.version = MAX_VERSION;
 671         }
 672
 673         if (trans->fs_usage_deltas &&
 674             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 675                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 676
 677         trans_for_each_update(trans, i)
 678                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 679                         ret = run_one_mem_trigger(trans, i, i->flags);
 680                         if (ret)
 681                                 return ret;
 682                 }
 683
 684         if (unlikely(c->gc_pos.phase)) {
 685                 ret = bch2_trans_commit_run_gc_triggers(trans);
 686                 if  (ret)
 687                         return ret;
 688         }
 689
 690         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 691                 trans_for_each_update(trans, i) {
 692                         struct journal *j = &c->journal;
 693                         struct jset_entry *entry;
 694
 695                         if (i->key_cache_already_flushed)
 696                                 continue;
 697
 698                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 699                                                BCH_JSET_ENTRY_overwrite,
 700                                                i->btree_id, i->level,
 701                                                i->old_k.u64s);
 702                         bkey_reassemble(&entry->start[0],
 703                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 704
 705                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 706                                                BCH_JSET_ENTRY_btree_keys,
 707                                                i->btree_id, i->level,
 708                                                i->k->k.u64s);
 709                         bkey_copy(&entry->start[0], i->k);
 710                 }
 711
 712                 if (trans->journal_seq)
 713                         *trans->journal_seq = trans->journal_res.seq;
 714         }
 715
 716         trans_for_each_update(trans, i) {
 717                 i->k->k.needs_whiteout = false;
 718
 719                 if (!i->cached)
 720                         btree_insert_key_leaf(trans, i);
 721                 else if (!i->key_cache_already_flushed)
 722                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 723                 else {
 724                         bch2_btree_key_cache_drop(trans, i->path);
 725                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 726                 }
 727         }
 728
 729         return ret;
 730 }
 731
 732 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 733 {
 734         unsigned l;
 735
 736         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 737                 if (btree_node_read_locked(path, l))
 738                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 739 }
 740
 741 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 742 {
 743         struct btree *b = path_l(path)->b;
 744
 745         do {
 746                 if (path->nodes_locked &&
 747                     path->nodes_locked != path->nodes_intent_locked)
 748                         path_upgrade_readers(trans, path);
 749         } while ((path = prev_btree_path(trans, path)) &&
 750                  path_l(path)->b == b);
 751 }
 752
 753 /*
 754  * Check for nodes that we have both read and intent locks on, and upgrade the
 755  * readers to intent:
 756  */
 757 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 758 {
 759         struct btree_path *path;
 760         unsigned i, nr_read = 0, nr_intent = 0;
 761
 762         trans_for_each_path_inorder(trans, path, i) {
 763                 struct btree_path *next = i + 1 < trans->nr_sorted
 764                         ? trans->paths + trans->sorted[i + 1]
 765                         : NULL;
 766
 767                 if (path->nodes_locked) {
 768                         if (path->nodes_intent_locked)
 769                                 nr_intent++;
 770                         else
 771                                 nr_read++;
 772                 }
 773
 774                 if (!next || path_l(path)->b != path_l(next)->b) {
 775                         if (nr_read && nr_intent)
 776                                 upgrade_readers(trans, path);
 777
 778                         nr_read = nr_intent = 0;
 779                 }
 780         }
 781
 782         bch2_trans_verify_locks(trans);
 783 }
 784
 785 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 786 {
 787         struct btree_path *path;
 788         unsigned i;
 789
 790         trans_for_each_path_inorder(trans, path, i) {
 791                 //if (path == pos)
 792                 //      break;
 793
 794                 if (path->nodes_locked != path->nodes_intent_locked &&
 795                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 796                         return true;
 797         }
 798
 799         return false;
 800 }
 801
 802 static inline int trans_lock_write(struct btree_trans *trans)
 803 {
 804         struct btree_insert_entry *i;
 805         int ret;
 806
 807         trans_for_each_update(trans, i) {
 808                 if (same_leaf_as_prev(trans, i))
 809                         continue;
 810
 811                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 812                         if (have_conflicting_read_lock(trans, i->path))
 813                                 goto fail;
 814
 815                         ret = btree_node_lock_type(trans, i->path,
 816                                              insert_l(i)->b,
 817                                              i->path->pos, i->level,
 818                                              SIX_LOCK_write, NULL, NULL);
 819                         BUG_ON(ret);
 820                 }
 821
 822                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 823         }
 824
 825         return 0;
 826 fail:
 827         while (--i >= trans->updates) {
 828                 if (same_leaf_as_prev(trans, i))
 829                         continue;
 830
 831                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 832         }
 833
 834         trace_trans_restart_would_deadlock_write(trans);
 835         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 836 }
 837
 838 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 839 {
 840         struct btree_insert_entry *i;
 841
 842         trans_for_each_update(trans, i)
 843                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 844 }
 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,
 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         struct printbuf buf = PRINTBUF;
 856         int ret, u64s_delta = 0;
 857         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 858
 859         trans_for_each_update(trans, i) {
 860                 if (bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 861                                       i->bkey_type, rw, &buf)) {
 862                         printbuf_reset(&buf);
 863                         prt_printf(&buf, "invalid bkey on insert from %s -> %ps",
 864                                trans->fn, (void *) i->ip_allocated);
 865                         prt_newline(&buf);
 866                         printbuf_indent_add(&buf, 2);
 867
 868                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
 869                         prt_newline(&buf);
 870
 871                         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 872                                           i->bkey_type, rw, &buf);
 873
 874                         bch2_trans_inconsistent(trans, "%s", buf.buf);
 875                         printbuf_exit(&buf);
 876                         return -EINVAL;
 877                 }
 878                 btree_insert_entry_checks(trans, i);
 879         }
 880
 881         printbuf_exit(&buf);
 882
 883         trans_for_each_update(trans, i) {
 884                 if (i->cached)
 885                         continue;
 886
 887                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 888                 u64s_delta -= i->old_btree_u64s;
 889
 890                 if (!same_leaf_as_next(trans, i)) {
 891                         if (u64s_delta <= 0) {
 892                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 893                                                         i->level, trans->flags);
 894                                 if (unlikely(ret))
 895                                         return ret;
 896                         }
 897
 898                         u64s_delta = 0;
 899                 }
 900         }
 901
 902         ret = bch2_journal_preres_get(&c->journal,
 903                         &trans->journal_preres, trans->journal_preres_u64s,
 904                         JOURNAL_RES_GET_NONBLOCK|
 905                         (trans->flags & JOURNAL_WATERMARK_MASK));
 906         if (unlikely(ret == -EAGAIN))
 907                 ret = bch2_trans_journal_preres_get_cold(trans,
 908                                                 trans->journal_preres_u64s, trace_ip);
 909         if (unlikely(ret))
 910                 return ret;
 911
 912         normalize_read_intent_locks(trans);
 913
 914         ret = trans_lock_write(trans);
 915         if (unlikely(ret))
 916                 return ret;
 917
 918         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 919
 920         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 921                 bch2_drop_overwrites_from_journal(trans);
 922
 923         trans_for_each_update(trans, i)
 924                 if (!same_leaf_as_prev(trans, i))
 925                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 926                                                         insert_l(i)->b);
 927
 928         if (!ret && trans->journal_pin)
 929                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 930                                      trans->journal_pin, NULL);
 931
 932         /*
 933          * Drop journal reservation after dropping write locks, since dropping
 934          * the journal reservation may kick off a journal write:
 935          */
 936         bch2_journal_res_put(&c->journal, &trans->journal_res);
 937
 938         if (unlikely(ret))
 939                 return ret;
 940
 941         bch2_trans_downgrade(trans);
 942
 943         return 0;
 944 }
 945
 946 static int journal_reclaim_wait_done(struct bch_fs *c)
 947 {
 948         int ret = bch2_journal_error(&c->journal) ?:
 949                 !bch2_btree_key_cache_must_wait(c);
 950
 951         if (!ret)
 952                 journal_reclaim_kick(&c->journal);
 953         return ret;
 954 }
 955
 956 static noinline
 957 int bch2_trans_commit_error(struct btree_trans *trans,
 958                             struct btree_insert_entry *i,
 959                             int ret, unsigned long trace_ip)
 960 {
 961         struct bch_fs *c = trans->c;
 962
 963         switch (ret) {
 964         case BTREE_INSERT_BTREE_NODE_FULL:
 965                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 966                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 967                         trace_trans_restart_btree_node_split(trans, trace_ip, i->path);
 968                 break;
 969         case BTREE_INSERT_NEED_MARK_REPLICAS:
 970                 bch2_trans_unlock(trans);
 971
 972                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 973                 if (ret)
 974                         break;
 975
 976                 ret = bch2_trans_relock(trans);
 977                 if (ret)
 978                         trace_trans_restart_mark_replicas(trans, trace_ip);
 979                 break;
 980         case BTREE_INSERT_NEED_JOURNAL_RES:
 981                 bch2_trans_unlock(trans);
 982
 983                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 984                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 985                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 986                         break;
 987                 }
 988
 989                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 990                 if (ret)
 991                         break;
 992
 993                 ret = bch2_trans_relock(trans);
 994                 if (ret)
 995                         trace_trans_restart_journal_res_get(trans, trace_ip);
 996                 break;
 997         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 998                 bch2_trans_unlock(trans);
 999
1000                 trace_trans_blocked_journal_reclaim(trans, trace_ip);
1001
1002                 wait_event_freezable(c->journal.reclaim_wait,
1003                                      (ret = journal_reclaim_wait_done(c)));
1004                 if (ret < 0)
1005                         break;
1006
1007                 ret = bch2_trans_relock(trans);
1008                 if (ret)
1009                         trace_trans_restart_journal_reclaim(trans, trace_ip);
1010                 break;
1011         default:
1012                 BUG_ON(ret >= 0);
1013                 break;
1014         }
1015
1016         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1017         BUG_ON(ret == -ENOSPC &&
1018                !(trans->flags & BTREE_INSERT_NOWAIT) &&
1019                (trans->flags & BTREE_INSERT_NOFAIL));
1020
1021         return ret;
1022 }
1023
1024 static noinline int
1025 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
1026 {
1027         struct bch_fs *c = trans->c;
1028         int ret;
1029
1030         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
1031             test_bit(BCH_FS_STARTED, &c->flags))
1032                 return -EROFS;
1033
1034         bch2_trans_unlock(trans);
1035
1036         ret =   bch2_fs_read_write_early(c) ?:
1037                 bch2_trans_relock(trans);
1038         if (ret)
1039                 return ret;
1040
1041         percpu_ref_get(&c->writes);
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)
1067 {
1068         struct bch_fs *c = trans->c;
1069         struct btree_insert_entry *i = NULL;
1070         unsigned u64s;
1071         int ret = 0;
1072
1073         if (!trans->nr_updates &&
1074             !trans->extra_journal_entries.nr)
1075                 goto out_reset;
1076
1077         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1078                 lockdep_assert_held(&c->gc_lock);
1079
1080         ret = bch2_trans_commit_run_triggers(trans);
1081         if (ret)
1082                 goto out_reset;
1083
1084         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1085                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1086                 goto out_reset;
1087         }
1088
1089         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1090             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1091                 ret = bch2_trans_commit_get_rw_cold(trans);
1092                 if (ret)
1093                         goto out_reset;
1094         }
1095
1096         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1097
1098         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1099
1100         trans->journal_u64s             = trans->extra_journal_entries.nr;
1101         trans->journal_preres_u64s      = 0;
1102
1103         /* For journalling transaction name: */
1104         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1105
1106         trans_for_each_update(trans, i) {
1107                 BUG_ON(!i->path->should_be_locked);
1108
1109                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1110                         trace_trans_restart_upgrade(trans, _RET_IP_, i->path);
1111                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_upgrade);
1112                         goto out;
1113                 }
1114
1115                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1116
1117                 if (i->key_cache_already_flushed)
1118                         continue;
1119
1120                 /* we're going to journal the key being updated: */
1121                 u64s = jset_u64s(i->k->k.u64s);
1122                 if (i->cached &&
1123                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1124                         trans->journal_preres_u64s += u64s;
1125                 trans->journal_u64s += u64s;
1126
1127                 /* and we're also going to log the overwrite: */
1128                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1129         }
1130
1131         if (trans->extra_journal_res) {
1132                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1133                                 trans->extra_journal_res,
1134                                 (trans->flags & BTREE_INSERT_NOFAIL)
1135                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1136                 if (ret)
1137                         goto err;
1138         }
1139 retry:
1140         BUG_ON(trans->restarted);
1141         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1142
1143         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1144
1145         /* make sure we didn't drop or screw up locks: */
1146         bch2_trans_verify_locks(trans);
1147
1148         if (ret)
1149                 goto err;
1150
1151         trace_transaction_commit(trans, _RET_IP_);
1152 out:
1153         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1154
1155         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1156                 percpu_ref_put(&c->writes);
1157 out_reset:
1158         bch2_trans_reset_updates(trans);
1159
1160         if (trans->fs_usage_deltas) {
1161                 trans->fs_usage_deltas->used = 0;
1162                 memset((void *) trans->fs_usage_deltas +
1163                        offsetof(struct replicas_delta_list, memset_start), 0,
1164                        (void *) &trans->fs_usage_deltas->memset_end -
1165                        (void *) &trans->fs_usage_deltas->memset_start);
1166         }
1167
1168         return ret;
1169 err:
1170         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1171         if (ret)
1172                 goto out;
1173
1174         goto retry;
1175 }
1176
1177 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1178                                           enum btree_id id,
1179                                           struct bpos pos)
1180 {
1181         struct bch_fs *c = trans->c;
1182         struct btree_iter iter;
1183         struct bkey_s_c k;
1184         int ret;
1185
1186         if (!btree_type_has_snapshots(id))
1187                 return 0;
1188
1189         if (!snapshot_t(c, pos.snapshot)->children[0])
1190                 return 0;
1191
1192         bch2_trans_iter_init(trans, &iter, id, pos,
1193                              BTREE_ITER_NOT_EXTENTS|
1194                              BTREE_ITER_ALL_SNAPSHOTS);
1195         while (1) {
1196                 k = bch2_btree_iter_prev(&iter);
1197                 ret = bkey_err(k);
1198                 if (ret)
1199                         break;
1200
1201                 if (!k.k)
1202                         break;
1203
1204                 if (bkey_cmp(pos, k.k->p))
1205                         break;
1206
1207                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1208                         ret = 1;
1209                         break;
1210                 }
1211         }
1212         bch2_trans_iter_exit(trans, &iter);
1213
1214         return ret;
1215 }
1216
1217 int bch2_trans_update_extent(struct btree_trans *trans,
1218                              struct btree_iter *orig_iter,
1219                              struct bkey_i *insert,
1220                              enum btree_update_flags flags)
1221 {
1222         struct bch_fs *c = trans->c;
1223         struct btree_iter iter, update_iter;
1224         struct bpos start = bkey_start_pos(&insert->k);
1225         struct bkey_i *update;
1226         struct bkey_s_c k;
1227         enum btree_id btree_id = orig_iter->btree_id;
1228         int ret = 0, compressed_sectors;
1229
1230         bch2_trans_iter_init(trans, &iter, btree_id, start,
1231                              BTREE_ITER_INTENT|
1232                              BTREE_ITER_WITH_UPDATES|
1233                              BTREE_ITER_NOT_EXTENTS);
1234         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1235         if ((ret = bkey_err(k)))
1236                 goto err;
1237         if (!k.k)
1238                 goto out;
1239
1240         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1241                 /*
1242                  * We can't merge extents if they belong to interior snapshot
1243                  * tree nodes, and there's a snapshot in which one extent is
1244                  * visible and the other is not - i.e. if visibility is
1245                  * different.
1246                  *
1247                  * Instead of checking if visibilitiy of the two extents is
1248                  * different, for now we just check if either has been
1249                  * overwritten:
1250                  */
1251                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1252                 if (ret < 0)
1253                         goto err;
1254                 if (ret)
1255                         goto nomerge1;
1256
1257                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1258                 if (ret < 0)
1259                         goto err;
1260                 if (ret)
1261                         goto nomerge1;
1262
1263                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1264                 if ((ret = PTR_ERR_OR_ZERO(update)))
1265                         goto err;
1266
1267                 bkey_reassemble(update, k);
1268
1269                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1270                         ret = bch2_btree_delete_at(trans, &iter, flags);
1271                         if (ret)
1272                                 goto err;
1273
1274                         insert = update;
1275                         goto next;
1276                 }
1277         }
1278 nomerge1:
1279         ret = 0;
1280         if (!bkey_cmp(k.k->p, start))
1281                 goto next;
1282
1283         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1284                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1285                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1286
1287                 /*
1288                  * If we're going to be splitting a compressed extent, note it
1289                  * so that __bch2_trans_commit() can increase our disk
1290                  * reservation:
1291                  */
1292                 if (((front_split && back_split) ||
1293                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1294                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1295                         trans->extra_journal_res += compressed_sectors;
1296
1297                 if (front_split) {
1298                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1299                         if ((ret = PTR_ERR_OR_ZERO(update)))
1300                                 goto err;
1301
1302                         bkey_reassemble(update, k);
1303
1304                         bch2_cut_back(start, update);
1305
1306                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1307                                              BTREE_ITER_NOT_EXTENTS|
1308                                              BTREE_ITER_ALL_SNAPSHOTS|
1309                                              BTREE_ITER_INTENT);
1310                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1311                                 bch2_trans_update(trans, &update_iter, update,
1312                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1313                                                   flags);
1314                         bch2_trans_iter_exit(trans, &update_iter);
1315
1316                         if (ret)
1317                                 goto err;
1318                 }
1319
1320                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1321                     (front_split || back_split)) {
1322                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1323                         if ((ret = PTR_ERR_OR_ZERO(update)))
1324                                 goto err;
1325
1326                         bkey_reassemble(update, k);
1327
1328                         bch2_cut_front(start, update);
1329                         bch2_cut_back(insert->k.p, update);
1330
1331                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1332                                              BTREE_ITER_NOT_EXTENTS|
1333                                              BTREE_ITER_ALL_SNAPSHOTS|
1334                                              BTREE_ITER_INTENT);
1335                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1336                                 bch2_trans_update(trans, &update_iter, update,
1337                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1338                                                   flags);
1339                         bch2_trans_iter_exit(trans, &update_iter);
1340                         if (ret)
1341                                 goto err;
1342                 }
1343
1344                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1345                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1346                         if ((ret = PTR_ERR_OR_ZERO(update)))
1347                                 goto err;
1348
1349                         bkey_init(&update->k);
1350                         update->k.p = k.k->p;
1351
1352                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1353                                 update->k.p.snapshot = insert->k.p.snapshot;
1354                                 update->k.type = KEY_TYPE_whiteout;
1355                         }
1356
1357                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1358                                              BTREE_ITER_NOT_EXTENTS|
1359                                              BTREE_ITER_INTENT);
1360                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1361                                 bch2_trans_update(trans, &update_iter, update,
1362                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1363                                                   flags);
1364                         bch2_trans_iter_exit(trans, &update_iter);
1365
1366                         if (ret)
1367                                 goto err;
1368                 }
1369
1370                 if (back_split) {
1371                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1372                         if ((ret = PTR_ERR_OR_ZERO(update)))
1373                                 goto err;
1374
1375                         bkey_reassemble(update, k);
1376                         bch2_cut_front(insert->k.p, update);
1377
1378                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1379                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1380                                                   flags);
1381                         if (ret)
1382                                 goto err;
1383                         goto out;
1384                 }
1385 next:
1386                 bch2_btree_iter_advance(&iter);
1387                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1388                 if ((ret = bkey_err(k)))
1389                         goto err;
1390                 if (!k.k)
1391                         goto out;
1392         }
1393
1394         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1395                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1396                 if (ret < 0)
1397                         goto err;
1398                 if (ret)
1399                         goto nomerge2;
1400
1401                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1402                 if (ret < 0)
1403                         goto err;
1404                 if (ret)
1405                         goto nomerge2;
1406
1407                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1408         }
1409 nomerge2:
1410         ret = 0;
1411 out:
1412         if (!bkey_deleted(&insert->k)) {
1413                 /*
1414                  * Rewinding iterators is expensive: get a new one and the one
1415                  * that points to the start of insert will be cloned from:
1416                  */
1417                 bch2_trans_iter_exit(trans, &iter);
1418                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1419                                      BTREE_ITER_NOT_EXTENTS|
1420                                      BTREE_ITER_INTENT);
1421                 ret   = bch2_btree_iter_traverse(&iter) ?:
1422                         bch2_trans_update(trans, &iter, insert, flags);
1423         }
1424 err:
1425         bch2_trans_iter_exit(trans, &iter);
1426
1427         return ret;
1428 }
1429
1430 /*
1431  * When deleting, check if we need to emit a whiteout (because we're overwriting
1432  * something in an ancestor snapshot)
1433  */
1434 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1435                                       enum btree_id btree_id, struct bpos pos)
1436 {
1437         struct btree_iter iter;
1438         struct bkey_s_c k;
1439         u32 snapshot = pos.snapshot;
1440         int ret;
1441
1442         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1443                 return 0;
1444
1445         pos.snapshot++;
1446
1447         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1448                            BTREE_ITER_ALL_SNAPSHOTS|
1449                            BTREE_ITER_NOPRESERVE, k, ret) {
1450                 if (bkey_cmp(k.k->p, pos))
1451                         break;
1452
1453                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1454                                               k.k->p.snapshot)) {
1455                         ret = !bkey_whiteout(k.k);
1456                         break;
1457                 }
1458         }
1459         bch2_trans_iter_exit(trans, &iter);
1460
1461         return ret;
1462 }
1463
1464 static int __must_check
1465 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1466                                 struct bkey_i *k, enum btree_update_flags flags,
1467                                 unsigned long ip)
1468 {
1469         struct bch_fs *c = trans->c;
1470         struct btree_insert_entry *i, n;
1471         int ret = 0;
1472
1473         BUG_ON(!path->should_be_locked);
1474
1475         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1476         BUG_ON(bpos_cmp(k->k.p, path->pos));
1477
1478         n = (struct btree_insert_entry) {
1479                 .flags          = flags,
1480                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1481                 .btree_id       = path->btree_id,
1482                 .level          = path->level,
1483                 .cached         = path->cached,
1484                 .path           = path,
1485                 .k              = k,
1486                 .ip_allocated   = ip,
1487         };
1488
1489 #ifdef CONFIG_BCACHEFS_DEBUG
1490         trans_for_each_update(trans, i)
1491                 BUG_ON(i != trans->updates &&
1492                        btree_insert_entry_cmp(i - 1, i) >= 0);
1493 #endif
1494
1495         /*
1496          * Pending updates are kept sorted: first, find position of new update,
1497          * then delete/trim any updates the new update overwrites:
1498          */
1499         trans_for_each_update(trans, i)
1500                 if (btree_insert_entry_cmp(&n, i) <= 0)
1501                         break;
1502
1503         if (i < trans->updates + trans->nr_updates &&
1504             !btree_insert_entry_cmp(&n, i)) {
1505                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1506
1507                 bch2_path_put(trans, i->path, true);
1508                 i->flags        = n.flags;
1509                 i->cached       = n.cached;
1510                 i->k            = n.k;
1511                 i->path         = n.path;
1512                 i->ip_allocated = n.ip_allocated;
1513         } else {
1514                 array_insert_item(trans->updates, trans->nr_updates,
1515                                   i - trans->updates, n);
1516
1517                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1518                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1519
1520                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1521                         struct bkey_i *j_k =
1522                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1523
1524                         if (j_k) {
1525                                 i->old_k = j_k->k;
1526                                 i->old_v = &j_k->v;
1527                         }
1528                 }
1529         }
1530
1531         __btree_path_get(i->path, true);
1532
1533         /*
1534          * If a key is present in the key cache, it must also exist in the
1535          * btree - this is necessary for cache coherency. When iterating over
1536          * a btree that's cached in the key cache, the btree iter code checks
1537          * the key cache - but the key has to exist in the btree for that to
1538          * work:
1539          */
1540         if (path->cached &&
1541             bkey_deleted(&i->old_k)) {
1542                 struct btree_path *btree_path;
1543
1544                 i->key_cache_already_flushed = true;
1545                 i->flags |= BTREE_TRIGGER_NORUN;
1546
1547                 btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1548                                            BTREE_ITER_INTENT, _THIS_IP_);
1549
1550                 ret = bch2_btree_path_traverse(trans, btree_path, 0);
1551                 if (ret)
1552                         goto err;
1553
1554                 btree_path_set_should_be_locked(btree_path);
1555                 ret = bch2_trans_update_by_path_trace(trans, btree_path, k, flags, ip);
1556 err:
1557                 bch2_path_put(trans, btree_path, true);
1558         }
1559
1560         return ret;
1561 }
1562
1563 static int __must_check
1564 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1565                           struct bkey_i *k, enum btree_update_flags flags)
1566 {
1567         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1568 }
1569
1570 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1571                                    struct bkey_i *k, enum btree_update_flags flags)
1572 {
1573         struct btree_path *path = iter->update_path ?: iter->path;
1574         struct bkey_cached *ck;
1575         int ret;
1576
1577         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1578                 return bch2_trans_update_extent(trans, iter, k, flags);
1579
1580         if (bkey_deleted(&k->k) &&
1581             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1582             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1583                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1584                 if (unlikely(ret < 0))
1585                         return ret;
1586
1587                 if (ret)
1588                         k->k.type = KEY_TYPE_whiteout;
1589         }
1590
1591         /*
1592          * Ensure that updates to cached btrees go to the key cache:
1593          */
1594         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1595             !path->cached &&
1596             !path->level &&
1597             btree_id_cached(trans->c, path->btree_id)) {
1598                 if (!iter->key_cache_path ||
1599                     !iter->key_cache_path->should_be_locked ||
1600                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1601                         if (!iter->key_cache_path)
1602                                 iter->key_cache_path =
1603                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1604                                                       BTREE_ITER_INTENT|
1605                                                       BTREE_ITER_CACHED, _THIS_IP_);
1606
1607                         iter->key_cache_path =
1608                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1609                                                         iter->flags & BTREE_ITER_INTENT,
1610                                                         _THIS_IP_);
1611
1612                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1613                                                        BTREE_ITER_CACHED);
1614                         if (unlikely(ret))
1615                                 return ret;
1616
1617                         ck = (void *) iter->key_cache_path->l[0].b;
1618
1619                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1620                                 trace_trans_restart_key_cache_raced(trans, _RET_IP_);
1621                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1622                         }
1623
1624                         btree_path_set_should_be_locked(iter->key_cache_path);
1625                 }
1626
1627                 path = iter->key_cache_path;
1628         }
1629
1630         return bch2_trans_update_by_path(trans, path, k, flags);
1631 }
1632
1633 void bch2_trans_commit_hook(struct btree_trans *trans,
1634                             struct btree_trans_commit_hook *h)
1635 {
1636         h->next = trans->hooks;
1637         trans->hooks = h;
1638 }
1639
1640 int __bch2_btree_insert(struct btree_trans *trans,
1641                         enum btree_id id, struct bkey_i *k)
1642 {
1643         struct btree_iter iter;
1644         int ret;
1645
1646         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1647                              BTREE_ITER_INTENT);
1648         ret   = bch2_btree_iter_traverse(&iter) ?:
1649                 bch2_trans_update(trans, &iter, k, 0);
1650         bch2_trans_iter_exit(trans, &iter);
1651         return ret;
1652 }
1653
1654 /**
1655  * bch2_btree_insert - insert keys into the extent btree
1656  * @c:                  pointer to struct bch_fs
1657  * @id:                 btree to insert into
1658  * @insert_keys:        list of keys to insert
1659  * @hook:               insert callback
1660  */
1661 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1662                       struct bkey_i *k,
1663                       struct disk_reservation *disk_res,
1664                       u64 *journal_seq, int flags)
1665 {
1666         return bch2_trans_do(c, disk_res, journal_seq, flags,
1667                              __bch2_btree_insert(&trans, id, k));
1668 }
1669
1670 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1671                                 unsigned len, unsigned update_flags)
1672 {
1673         struct bkey_i *k;
1674
1675         k = bch2_trans_kmalloc(trans, sizeof(*k));
1676         if (IS_ERR(k))
1677                 return PTR_ERR(k);
1678
1679         bkey_init(&k->k);
1680         k->k.p = iter->pos;
1681         bch2_key_resize(&k->k, len);
1682         return bch2_trans_update(trans, iter, k, update_flags);
1683 }
1684
1685 int bch2_btree_delete_at(struct btree_trans *trans,
1686                          struct btree_iter *iter, unsigned update_flags)
1687 {
1688         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1689 }
1690
1691 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1692                                   struct bpos start, struct bpos end,
1693                                   unsigned update_flags,
1694                                   u64 *journal_seq)
1695 {
1696         struct btree_iter iter;
1697         struct bkey_s_c k;
1698         int ret = 0;
1699
1700         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1701 retry:
1702         while ((bch2_trans_begin(trans),
1703                (k = bch2_btree_iter_peek(&iter)).k) &&
1704                !(ret = bkey_err(k)) &&
1705                bkey_cmp(iter.pos, end) < 0) {
1706                 struct disk_reservation disk_res =
1707                         bch2_disk_reservation_init(trans->c, 0);
1708                 struct bkey_i delete;
1709
1710                 bkey_init(&delete.k);
1711
1712                 /*
1713                  * This could probably be more efficient for extents:
1714                  */
1715
1716                 /*
1717                  * For extents, iter.pos won't necessarily be the same as
1718                  * bkey_start_pos(k.k) (for non extents they always will be the
1719                  * same). It's important that we delete starting from iter.pos
1720                  * because the range we want to delete could start in the middle
1721                  * of k.
1722                  *
1723                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1724                  * bkey_start_pos(k.k)).
1725                  */
1726                 delete.k.p = iter.pos;
1727
1728                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1729                         unsigned max_sectors =
1730                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1731
1732                         /* create the biggest key we can */
1733                         bch2_key_resize(&delete.k, max_sectors);
1734                         bch2_cut_back(end, &delete);
1735
1736                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1737                         if (ret)
1738                                 break;
1739                 }
1740
1741                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1742                         bch2_trans_commit(trans, &disk_res, journal_seq,
1743                                           BTREE_INSERT_NOFAIL);
1744                 bch2_disk_reservation_put(trans->c, &disk_res);
1745                 if (ret)
1746                         break;
1747         }
1748
1749         if (bch2_err_matches(ret, BCH_ERR_transaction_restart)) {
1750                 ret = 0;
1751                 goto retry;
1752         }
1753
1754         bch2_trans_iter_exit(trans, &iter);
1755         return ret;
1756 }
1757
1758 /*
1759  * bch_btree_delete_range - delete everything within a given range
1760  *
1761  * Range is a half open interval - [start, end)
1762  */
1763 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1764                             struct bpos start, struct bpos end,
1765                             unsigned update_flags,
1766                             u64 *journal_seq)
1767 {
1768         return bch2_trans_run(c,
1769                         bch2_btree_delete_range_trans(&trans, id, start, end, update_flags, journal_seq));
1770 }
1771
1772 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1773 {
1774         unsigned len = strlen(msg);
1775         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1776         struct jset_entry_log *l;
1777         int ret;
1778
1779         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1780         if (ret)
1781                 return ret;
1782
1783         l = (void *) &darray_top(trans->extra_journal_entries);
1784         l->entry.u64s           = cpu_to_le16(u64s);
1785         l->entry.btree_id       = 0;
1786         l->entry.level          = 1;
1787         l->entry.type           = BCH_JSET_ENTRY_log;
1788         l->entry.pad[0]         = 0;
1789         l->entry.pad[1]         = 0;
1790         l->entry.pad[2]         = 0;
1791         memcpy(l->d, msg, len);
1792         while (len & 7)
1793                 l->d[len++] = '\0';
1794
1795         trans->extra_journal_entries.nr += jset_u64s(u64s);
1796         return 0;
1797 }