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 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 /* Inserting into a given leaf node (last stage of insert): */
  81
  82 /* Handle overwrites and do insert, for non extents: */
  83 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  84                                 struct btree_path *path,
  85                                 struct btree *b,
  86                                 struct btree_node_iter *node_iter,
  87                                 struct bkey_i *insert)
  88 {
  89         struct bkey_packed *k;
  90         unsigned clobber_u64s = 0, new_u64s = 0;
  91
  92         EBUG_ON(btree_node_just_written(b));
  93         EBUG_ON(bset_written(b, btree_bset_last(b)));
  94         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
  95         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
  96         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
  97         EBUG_ON(insert->k.u64s >
  98                 bch_btree_keys_u64s_remaining(trans->c, b));
  99
 100         k = bch2_btree_node_iter_peek_all(node_iter, b);
 101         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 102                 k = NULL;
 103
 104         /* @k is the key being overwritten/deleted, if any: */
 105         EBUG_ON(k && bkey_deleted(k));
 106
 107         /* Deleting, but not found? nothing to do: */
 108         if (bkey_deleted(&insert->k) && !k)
 109                 return false;
 110
 111         if (bkey_deleted(&insert->k)) {
 112                 /* Deleting: */
 113                 btree_account_key_drop(b, k);
 114                 k->type = KEY_TYPE_deleted;
 115
 116                 if (k->needs_whiteout)
 117                         push_whiteout(trans->c, b, insert->k.p);
 118                 k->needs_whiteout = false;
 119
 120                 if (k >= btree_bset_last(b)->start) {
 121                         clobber_u64s = k->u64s;
 122                         bch2_bset_delete(b, k, clobber_u64s);
 123                         goto fix_iter;
 124                 } else {
 125                         bch2_btree_path_fix_key_modified(trans, b, k);
 126                 }
 127
 128                 return true;
 129         }
 130
 131         if (k) {
 132                 /* Overwriting: */
 133                 btree_account_key_drop(b, k);
 134                 k->type = KEY_TYPE_deleted;
 135
 136                 insert->k.needs_whiteout = k->needs_whiteout;
 137                 k->needs_whiteout = false;
 138
 139                 if (k >= btree_bset_last(b)->start) {
 140                         clobber_u64s = k->u64s;
 141                         goto overwrite;
 142                 } else {
 143                         bch2_btree_path_fix_key_modified(trans, b, k);
 144                 }
 145         }
 146
 147         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 148 overwrite:
 149         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 150         new_u64s = k->u64s;
 151 fix_iter:
 152         if (clobber_u64s != new_u64s)
 153                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 154                                          clobber_u64s, new_u64s);
 155         return true;
 156 }
 157
 158 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 159                                unsigned i, u64 seq)
 160 {
 161         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 162         struct btree_write *w = container_of(pin, struct btree_write, journal);
 163         struct btree *b = container_of(w, struct btree, writes[i]);
 164         struct btree_trans trans;
 165         unsigned long old, new, v;
 166         unsigned idx = w - b->writes;
 167
 168         bch2_trans_init(&trans, c, 0, 0);
 169
 170         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 171         v = READ_ONCE(b->flags);
 172
 173         do {
 174                 old = new = v;
 175
 176                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 177                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 178                     w->journal.seq != seq)
 179                         break;
 180
 181                 new &= ~BTREE_WRITE_TYPE_MASK;
 182                 new |= BTREE_WRITE_journal_reclaim;
 183                 new |= 1 << BTREE_NODE_need_write;
 184         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 185
 186         btree_node_write_if_need(c, b, SIX_LOCK_read);
 187         six_unlock_read(&b->c.lock);
 188
 189         bch2_trans_exit(&trans);
 190         return 0;
 191 }
 192
 193 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 194 {
 195         return __btree_node_flush(j, pin, 0, seq);
 196 }
 197
 198 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 199 {
 200         return __btree_node_flush(j, pin, 1, seq);
 201 }
 202
 203 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 204                                        struct btree *b, u64 seq)
 205 {
 206         struct btree_write *w = btree_current_write(b);
 207
 208         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 209                              btree_node_write_idx(b) == 0
 210                              ? btree_node_flush0
 211                              : btree_node_flush1);
 212 }
 213
 214 /**
 215  * btree_insert_key - insert a key one key into a leaf node
 216  */
 217 static void btree_insert_key_leaf(struct btree_trans *trans,
 218                                   struct btree_insert_entry *insert)
 219 {
 220         struct bch_fs *c = trans->c;
 221         struct btree *b = insert_l(insert)->b;
 222         struct bset_tree *t = bset_tree_last(b);
 223         struct bset *i = bset(b, t);
 224         int old_u64s = bset_u64s(t);
 225         int old_live_u64s = b->nr.live_u64s;
 226         int live_u64s_added, u64s_added;
 227
 228         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 229                                         &insert_l(insert)->iter, insert->k)))
 230                 return;
 231
 232         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 233                                          le64_to_cpu(i->journal_seq)));
 234
 235         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 236
 237         if (unlikely(!btree_node_dirty(b)))
 238                 set_btree_node_dirty_acct(c, b);
 239
 240         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 241         u64s_added = (int) bset_u64s(t) - old_u64s;
 242
 243         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 244                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 245         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 246                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 247
 248         if (u64s_added > live_u64s_added &&
 249             bch2_maybe_compact_whiteouts(c, b))
 250                 bch2_trans_node_reinit_iter(trans, b);
 251 }
 252
 253 /* Cached btree updates: */
 254
 255 /* Normal update interface: */
 256
 257 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 258                                              struct btree_insert_entry *i)
 259 {
 260         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 261         BUG_ON(i->cached        != i->path->cached);
 262         BUG_ON(i->level         != i->path->level);
 263         BUG_ON(i->btree_id      != i->path->btree_id);
 264         EBUG_ON(!i->level &&
 265                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 266                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 267                 i->k->k.p.snapshot &&
 268                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 269 }
 270
 271 static noinline int
 272 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 273                                    unsigned long trace_ip)
 274 {
 275         struct bch_fs *c = trans->c;
 276         int ret;
 277
 278         bch2_trans_unlock(trans);
 279
 280         ret = bch2_journal_preres_get(&c->journal,
 281                         &trans->journal_preres, u64s, 0);
 282         if (ret)
 283                 return ret;
 284
 285         ret = bch2_trans_relock(trans);
 286         if (ret) {
 287                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 288                 return ret;
 289         }
 290
 291         return 0;
 292 }
 293
 294 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 295                                              unsigned flags)
 296 {
 297         struct bch_fs *c = trans->c;
 298         int ret;
 299
 300         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 301                                    trans->journal_u64s,
 302                                    flags|
 303                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 304
 305         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 306 }
 307
 308 #define JSET_ENTRY_LOG_U64s             4
 309
 310 static void journal_transaction_name(struct btree_trans *trans)
 311 {
 312         struct bch_fs *c = trans->c;
 313         struct journal *j = &c->journal;
 314         struct jset_entry *entry =
 315                 bch2_journal_add_entry(j, &trans->journal_res,
 316                                        BCH_JSET_ENTRY_log, 0, 0,
 317                                        JSET_ENTRY_LOG_U64s);
 318         struct jset_entry_log *l =
 319                 container_of(entry, struct jset_entry_log, entry);
 320
 321         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 322 }
 323
 324 static inline enum btree_insert_ret
 325 btree_key_can_insert(struct btree_trans *trans,
 326                      struct btree *b,
 327                      unsigned u64s)
 328 {
 329         struct bch_fs *c = trans->c;
 330
 331         if (!bch2_btree_node_insert_fits(c, b, u64s))
 332                 return BTREE_INSERT_BTREE_NODE_FULL;
 333
 334         return BTREE_INSERT_OK;
 335 }
 336
 337 static enum btree_insert_ret
 338 btree_key_can_insert_cached(struct btree_trans *trans,
 339                             struct btree_path *path,
 340                             unsigned u64s)
 341 {
 342         struct bch_fs *c = trans->c;
 343         struct bkey_cached *ck = (void *) path->l[0].b;
 344         unsigned new_u64s;
 345         struct bkey_i *new_k;
 346
 347         EBUG_ON(path->level);
 348
 349         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 350             bch2_btree_key_cache_must_wait(c) &&
 351             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 352                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 353
 354         /*
 355          * bch2_varint_decode can read past the end of the buffer by at most 7
 356          * bytes (it won't be used):
 357          */
 358         u64s += 1;
 359
 360         if (u64s <= ck->u64s)
 361                 return BTREE_INSERT_OK;
 362
 363         new_u64s        = roundup_pow_of_two(u64s);
 364         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 365         if (!new_k) {
 366                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 367                         bch2_btree_ids[path->btree_id], new_u64s);
 368                 return -ENOMEM;
 369         }
 370
 371         ck->u64s        = new_u64s;
 372         ck->k           = new_k;
 373         return 0;
 374 }
 375
 376 /* Triggers: */
 377
 378 static int run_one_mem_trigger(struct btree_trans *trans,
 379                                struct btree_insert_entry *i,
 380                                unsigned flags)
 381 {
 382         struct bkey_s_c old = { &i->old_k, i->old_v };
 383         struct bkey_i *new = i->k;
 384         int ret;
 385
 386         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 387                 return 0;
 388
 389         if (!btree_node_type_needs_gc(i->btree_id))
 390                 return 0;
 391
 392         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 393             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 394             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 395                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 396                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 397         } else {
 398                 struct bkey             _deleted = KEY(0, 0, 0);
 399                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 400
 401                 _deleted.p = i->path->pos;
 402
 403                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 404                                 BTREE_TRIGGER_INSERT|flags) ?:
 405                         bch2_mark_key(trans, old, deleted,
 406                                 BTREE_TRIGGER_OVERWRITE|flags);
 407         }
 408
 409         return ret;
 410 }
 411
 412 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 413                                  bool overwrite)
 414 {
 415         /*
 416          * Transactional triggers create new btree_insert_entries, so we can't
 417          * pass them a pointer to a btree_insert_entry, that memory is going to
 418          * move:
 419          */
 420         struct bkey old_k = i->old_k;
 421         struct bkey_s_c old = { &old_k, i->old_v };
 422
 423         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 424             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 425                 return 0;
 426
 427         if (!i->insert_trigger_run &&
 428             !i->overwrite_trigger_run &&
 429             bch2_bkey_ops[old.k->type].trans_trigger ==
 430             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 431             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 432                 i->overwrite_trigger_run = true;
 433                 i->insert_trigger_run = true;
 434                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 435                                            BTREE_TRIGGER_INSERT|
 436                                            BTREE_TRIGGER_OVERWRITE|
 437                                            i->flags) ?: 1;
 438         } else if (overwrite && !i->overwrite_trigger_run) {
 439                 i->overwrite_trigger_run = true;
 440                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 441         } else if (!overwrite && !i->insert_trigger_run) {
 442                 i->insert_trigger_run = true;
 443                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 444         } else {
 445                 return 0;
 446         }
 447 }
 448
 449 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 450                               struct btree_insert_entry *btree_id_start)
 451 {
 452         struct btree_insert_entry *i;
 453         bool trans_trigger_run;
 454         int ret, overwrite;
 455
 456         for (overwrite = 1; overwrite >= 0; --overwrite) {
 457
 458                 /*
 459                  * Running triggers will append more updates to the list of updates as
 460                  * we're walking it:
 461                  */
 462                 do {
 463                         trans_trigger_run = false;
 464
 465                         for (i = btree_id_start;
 466                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 467                              i++) {
 468                                 if (i->btree_id != btree_id)
 469                                         continue;
 470
 471                                 ret = run_one_trans_trigger(trans, i, overwrite);
 472                                 if (ret < 0)
 473                                         return ret;
 474                                 if (ret)
 475                                         trans_trigger_run = true;
 476                         }
 477                 } while (trans_trigger_run);
 478         }
 479
 480         return 0;
 481 }
 482
 483 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 484 {
 485         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 486         unsigned btree_id = 0;
 487         int ret = 0;
 488
 489         /*
 490          *
 491          * For a given btree, this algorithm runs insert triggers before
 492          * overwrite triggers: this is so that when extents are being moved
 493          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 494          * they are re-added.
 495          */
 496         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 497                 if (btree_id == BTREE_ID_alloc)
 498                         continue;
 499
 500                 while (btree_id_start < trans->updates + trans->nr_updates &&
 501                        btree_id_start->btree_id < btree_id)
 502                         btree_id_start++;
 503
 504                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 505                 if (ret)
 506                         return ret;
 507         }
 508
 509         trans_for_each_update(trans, i) {
 510                 if (i->btree_id > BTREE_ID_alloc)
 511                         break;
 512                 if (i->btree_id == BTREE_ID_alloc) {
 513                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 514                         if (ret)
 515                                 return ret;
 516                         break;
 517                 }
 518         }
 519
 520 #ifdef CONFIG_BCACHEFS_DEBUG
 521         trans_for_each_update(trans, i)
 522                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 523                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 524                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 525 #endif
 526         return 0;
 527 }
 528
 529 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 530 {
 531         struct bch_fs *c = trans->c;
 532         struct btree_insert_entry *i;
 533         int ret = 0;
 534
 535         trans_for_each_update(trans, i) {
 536                 /*
 537                  * XXX: synchronization of cached update triggers with gc
 538                  * XXX: synchronization of interior node updates with gc
 539                  */
 540                 BUG_ON(i->cached || i->level);
 541
 542                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 543                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 544                         if (ret)
 545                                 break;
 546                 }
 547         }
 548
 549         return ret;
 550 }
 551
 552 static inline int
 553 bch2_trans_commit_write_locked(struct btree_trans *trans,
 554                                struct btree_insert_entry **stopped_at,
 555                                unsigned long trace_ip)
 556 {
 557         struct bch_fs *c = trans->c;
 558         struct btree_insert_entry *i;
 559         struct btree_trans_commit_hook *h;
 560         unsigned u64s = 0;
 561         bool marking = false;
 562         int ret;
 563
 564         if (race_fault()) {
 565                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 566                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 567         }
 568
 569         /*
 570          * Check if the insert will fit in the leaf node with the write lock
 571          * held, otherwise another thread could write the node changing the
 572          * amount of space available:
 573          */
 574
 575         prefetch(&trans->c->journal.flags);
 576
 577         h = trans->hooks;
 578         while (h) {
 579                 ret = h->fn(trans, h);
 580                 if (ret)
 581                         return ret;
 582                 h = h->next;
 583         }
 584
 585         trans_for_each_update(trans, i) {
 586                 /* Multiple inserts might go to same leaf: */
 587                 if (!same_leaf_as_prev(trans, i))
 588                         u64s = 0;
 589
 590                 u64s += i->k->k.u64s;
 591                 ret = !i->cached
 592                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 593                         : btree_key_can_insert_cached(trans, i->path, u64s);
 594                 if (ret) {
 595                         *stopped_at = i;
 596                         return ret;
 597                 }
 598
 599                 if (btree_node_type_needs_gc(i->bkey_type))
 600                         marking = true;
 601
 602                 /*
 603                  * Revalidate before calling mem triggers - XXX, ugly:
 604                  *
 605                  * - successful btree node splits don't cause transaction
 606                  *   restarts and will have invalidated the pointer to the bkey
 607                  *   value
 608                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 609                  *   when it has to resort
 610                  * - btree_key_can_insert_cached() when it has to reallocate
 611                  *
 612                  *   Ugly because we currently have no way to tell if the
 613                  *   pointer's been invalidated, which means it's debatabale
 614                  *   whether we should be stashing the old key at all.
 615                  */
 616                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 617
 618                 if (unlikely(trans->journal_replay_not_finished)) {
 619                         struct bkey_i *j_k =
 620                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 621                                                             i->k->k.p);
 622
 623                         if (j_k) {
 624                                 i->old_k = j_k->k;
 625                                 i->old_v = &j_k->v;
 626                         }
 627                 }
 628         }
 629
 630         /*
 631          * Don't get journal reservation until after we know insert will
 632          * succeed:
 633          */
 634         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 635                 ret = bch2_trans_journal_res_get(trans,
 636                                 JOURNAL_RES_GET_NONBLOCK);
 637                 if (ret)
 638                         return ret;
 639
 640                 journal_transaction_name(trans);
 641         } else {
 642                 trans->journal_res.seq = c->journal.replay_journal_seq;
 643         }
 644
 645         if (unlikely(trans->extra_journal_entries.nr)) {
 646                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 647                                   trans->extra_journal_entries.data,
 648                                   trans->extra_journal_entries.nr);
 649
 650                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 651                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 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 (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 660                 if (bch2_journal_seq_verify)
 661                         trans_for_each_update(trans, i)
 662                                 i->k->k.version.lo = trans->journal_res.seq;
 663                 else if (bch2_inject_invalid_keys)
 664                         trans_for_each_update(trans, i)
 665                                 i->k->k.version = MAX_VERSION;
 666         }
 667
 668         if (trans->fs_usage_deltas &&
 669             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 670                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 671
 672         trans_for_each_update(trans, i)
 673                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 674                         ret = run_one_mem_trigger(trans, i, i->flags);
 675                         if (ret)
 676                                 return ret;
 677                 }
 678
 679         if (unlikely(c->gc_pos.phase)) {
 680                 ret = bch2_trans_commit_run_gc_triggers(trans);
 681                 if  (ret)
 682                         return ret;
 683         }
 684
 685         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 686                 trans_for_each_update(trans, i) {
 687                         struct journal *j = &c->journal;
 688                         struct jset_entry *entry;
 689
 690                         if (i->key_cache_already_flushed)
 691                                 continue;
 692
 693                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 694                                                BCH_JSET_ENTRY_overwrite,
 695                                                i->btree_id, i->level,
 696                                                i->old_k.u64s);
 697                         bkey_reassemble(&entry->start[0],
 698                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 699
 700                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 701                                                BCH_JSET_ENTRY_btree_keys,
 702                                                i->btree_id, i->level,
 703                                                i->k->k.u64s);
 704                         bkey_copy(&entry->start[0], i->k);
 705                 }
 706
 707                 if (trans->journal_seq)
 708                         *trans->journal_seq = trans->journal_res.seq;
 709         }
 710
 711         trans_for_each_update(trans, i) {
 712                 i->k->k.needs_whiteout = false;
 713
 714                 if (!i->cached)
 715                         btree_insert_key_leaf(trans, i);
 716                 else if (!i->key_cache_already_flushed)
 717                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 718                 else {
 719                         bch2_btree_key_cache_drop(trans, i->path);
 720                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 721                 }
 722         }
 723
 724         return ret;
 725 }
 726
 727 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 728 {
 729         while (--i >= trans->updates) {
 730                 if (same_leaf_as_prev(trans, i))
 731                         continue;
 732
 733                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 734         }
 735
 736         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 737         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 738 }
 739
 740 static inline int trans_lock_write(struct btree_trans *trans)
 741 {
 742         struct btree_insert_entry *i;
 743
 744         trans_for_each_update(trans, i) {
 745                 if (same_leaf_as_prev(trans, i))
 746                         continue;
 747
 748                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 749                         return trans_lock_write_fail(trans, i);
 750
 751                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 752         }
 753
 754         return 0;
 755 }
 756
 757 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 758 {
 759         struct btree_insert_entry *i;
 760
 761         trans_for_each_update(trans, i)
 762                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 763 }
 764
 765 #ifdef CONFIG_BCACHEFS_DEBUG
 766 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 767                                                    struct btree_insert_entry *i,
 768                                                    struct printbuf *err)
 769 {
 770         struct bch_fs *c = trans->c;
 771         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 772
 773         printbuf_reset(err);
 774         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 775                    trans->fn, (void *) i->ip_allocated);
 776         prt_newline(err);
 777         printbuf_indent_add(err, 2);
 778
 779         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 780         prt_newline(err);
 781
 782         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 783                           i->bkey_type, rw, err);
 784         bch2_print_string_as_lines(KERN_ERR, err->buf);
 785
 786         bch2_inconsistent_error(c);
 787         bch2_dump_trans_updates(trans);
 788         printbuf_exit(err);
 789
 790         return -EINVAL;
 791 }
 792 #endif
 793
 794 /*
 795  * Get journal reservation, take write locks, and attempt to do btree update(s):
 796  */
 797 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 798                                        struct btree_insert_entry **stopped_at,
 799                                        unsigned long trace_ip)
 800 {
 801         struct bch_fs *c = trans->c;
 802         struct btree_insert_entry *i;
 803         struct printbuf buf = PRINTBUF;
 804         int ret, u64s_delta = 0;
 805
 806 #ifdef CONFIG_BCACHEFS_DEBUG
 807         trans_for_each_update(trans, i) {
 808                 int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 809
 810                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 811                                                i->bkey_type, rw, &buf)))
 812                         return bch2_trans_commit_bkey_invalid(trans, i, &buf);
 813                 btree_insert_entry_checks(trans, i);
 814         }
 815 #endif
 816         printbuf_exit(&buf);
 817
 818         trans_for_each_update(trans, i) {
 819                 if (i->cached)
 820                         continue;
 821
 822                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 823                 u64s_delta -= i->old_btree_u64s;
 824
 825                 if (!same_leaf_as_next(trans, i)) {
 826                         if (u64s_delta <= 0) {
 827                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 828                                                         i->level, trans->flags);
 829                                 if (unlikely(ret))
 830                                         return ret;
 831                         }
 832
 833                         u64s_delta = 0;
 834                 }
 835         }
 836
 837         ret = bch2_journal_preres_get(&c->journal,
 838                         &trans->journal_preres, trans->journal_preres_u64s,
 839                         JOURNAL_RES_GET_NONBLOCK|
 840                         (trans->flags & JOURNAL_WATERMARK_MASK));
 841         if (unlikely(ret == -EAGAIN))
 842                 ret = bch2_trans_journal_preres_get_cold(trans,
 843                                                 trans->journal_preres_u64s, trace_ip);
 844         if (unlikely(ret))
 845                 return ret;
 846
 847         ret = trans_lock_write(trans);
 848         if (unlikely(ret))
 849                 return ret;
 850
 851         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 852
 853         if (!ret && unlikely(trans->journal_replay_not_finished))
 854                 bch2_drop_overwrites_from_journal(trans);
 855
 856         trans_for_each_update(trans, i)
 857                 if (!same_leaf_as_prev(trans, i))
 858                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 859                                                         insert_l(i)->b);
 860
 861         if (!ret && trans->journal_pin)
 862                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 863                                      trans->journal_pin, NULL);
 864
 865         /*
 866          * Drop journal reservation after dropping write locks, since dropping
 867          * the journal reservation may kick off a journal write:
 868          */
 869         bch2_journal_res_put(&c->journal, &trans->journal_res);
 870
 871         if (unlikely(ret))
 872                 return ret;
 873
 874         bch2_trans_downgrade(trans);
 875
 876         return 0;
 877 }
 878
 879 static int journal_reclaim_wait_done(struct bch_fs *c)
 880 {
 881         int ret = bch2_journal_error(&c->journal) ?:
 882                 !bch2_btree_key_cache_must_wait(c);
 883
 884         if (!ret)
 885                 journal_reclaim_kick(&c->journal);
 886         return ret;
 887 }
 888
 889 static noinline
 890 int bch2_trans_commit_error(struct btree_trans *trans,
 891                             struct btree_insert_entry *i,
 892                             int ret, unsigned long trace_ip)
 893 {
 894         struct bch_fs *c = trans->c;
 895
 896         switch (ret) {
 897         case BTREE_INSERT_BTREE_NODE_FULL:
 898                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 899                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 900                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 901                 break;
 902         case BTREE_INSERT_NEED_MARK_REPLICAS:
 903                 bch2_trans_unlock(trans);
 904
 905                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 906                 if (ret)
 907                         break;
 908
 909                 ret = bch2_trans_relock(trans);
 910                 if (ret)
 911                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 912                 break;
 913         case BTREE_INSERT_NEED_JOURNAL_RES:
 914                 bch2_trans_unlock(trans);
 915
 916                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 917                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 918                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 919                         break;
 920                 }
 921
 922                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 923                 if (ret)
 924                         break;
 925
 926                 ret = bch2_trans_relock(trans);
 927                 if (ret)
 928                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 929                 break;
 930         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 931                 bch2_trans_unlock(trans);
 932
 933                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 934
 935                 wait_event_freezable(c->journal.reclaim_wait,
 936                                      (ret = journal_reclaim_wait_done(c)));
 937                 if (ret < 0)
 938                         break;
 939
 940                 ret = bch2_trans_relock(trans);
 941                 if (ret)
 942                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 943                 break;
 944         default:
 945                 BUG_ON(ret >= 0);
 946                 break;
 947         }
 948
 949         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 950
 951         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 952                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 953                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 954                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 955
 956         return ret;
 957 }
 958
 959 static noinline int
 960 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 961 {
 962         struct bch_fs *c = trans->c;
 963         int ret;
 964
 965         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 966             test_bit(BCH_FS_STARTED, &c->flags))
 967                 return -EROFS;
 968
 969         bch2_trans_unlock(trans);
 970
 971         ret =   bch2_fs_read_write_early(c) ?:
 972                 bch2_trans_relock(trans);
 973         if (ret)
 974                 return ret;
 975
 976         percpu_ref_get(&c->writes);
 977         return 0;
 978 }
 979
 980 /*
 981  * This is for updates done in the early part of fsck - btree_gc - before we've
 982  * gone RW. we only add the new key to the list of keys for journal replay to
 983  * do.
 984  */
 985 static noinline int
 986 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 987 {
 988         struct bch_fs *c = trans->c;
 989         struct btree_insert_entry *i;
 990         int ret = 0;
 991
 992         trans_for_each_update(trans, i) {
 993                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 994                 if (ret)
 995                         break;
 996         }
 997
 998         return ret;
 999 }
1000
1001 int __bch2_trans_commit(struct btree_trans *trans)
1002 {
1003         struct bch_fs *c = trans->c;
1004         struct btree_insert_entry *i = NULL;
1005         unsigned u64s;
1006         int ret = 0;
1007
1008         if (!trans->nr_updates &&
1009             !trans->extra_journal_entries.nr)
1010                 goto out_reset;
1011
1012         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1013                 lockdep_assert_held(&c->gc_lock);
1014
1015         ret = bch2_trans_commit_run_triggers(trans);
1016         if (ret)
1017                 goto out_reset;
1018
1019         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1020                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1021                 goto out_reset;
1022         }
1023
1024         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1025             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1026                 ret = bch2_trans_commit_get_rw_cold(trans);
1027                 if (ret)
1028                         goto out_reset;
1029         }
1030
1031         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1032
1033         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1034
1035         trans->journal_u64s             = trans->extra_journal_entries.nr;
1036         trans->journal_preres_u64s      = 0;
1037
1038         /* For journalling transaction name: */
1039         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1040
1041         trans_for_each_update(trans, i) {
1042                 EBUG_ON(!i->path->should_be_locked);
1043
1044                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1045                 if (unlikely(ret))
1046                         goto out;
1047
1048                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1049
1050                 if (i->key_cache_already_flushed)
1051                         continue;
1052
1053                 /* we're going to journal the key being updated: */
1054                 u64s = jset_u64s(i->k->k.u64s);
1055                 if (i->cached &&
1056                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1057                         trans->journal_preres_u64s += u64s;
1058                 trans->journal_u64s += u64s;
1059
1060                 /* and we're also going to log the overwrite: */
1061                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1062         }
1063
1064         if (trans->extra_journal_res) {
1065                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1066                                 trans->extra_journal_res,
1067                                 (trans->flags & BTREE_INSERT_NOFAIL)
1068                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1069                 if (ret)
1070                         goto err;
1071         }
1072 retry:
1073         EBUG_ON(trans->restarted);
1074         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1075
1076         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1077
1078         /* make sure we didn't drop or screw up locks: */
1079         bch2_trans_verify_locks(trans);
1080
1081         if (ret)
1082                 goto err;
1083
1084         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1085 out:
1086         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1087
1088         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1089                 percpu_ref_put(&c->writes);
1090 out_reset:
1091         bch2_trans_reset_updates(trans);
1092
1093         if (trans->fs_usage_deltas) {
1094                 trans->fs_usage_deltas->used = 0;
1095                 memset((void *) trans->fs_usage_deltas +
1096                        offsetof(struct replicas_delta_list, memset_start), 0,
1097                        (void *) &trans->fs_usage_deltas->memset_end -
1098                        (void *) &trans->fs_usage_deltas->memset_start);
1099         }
1100
1101         return ret;
1102 err:
1103         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1104         if (ret)
1105                 goto out;
1106
1107         goto retry;
1108 }
1109
1110 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1111                                           enum btree_id id,
1112                                           struct bpos pos)
1113 {
1114         struct bch_fs *c = trans->c;
1115         struct btree_iter iter;
1116         struct bkey_s_c k;
1117         int ret;
1118
1119         bch2_trans_iter_init(trans, &iter, id, pos,
1120                              BTREE_ITER_NOT_EXTENTS|
1121                              BTREE_ITER_ALL_SNAPSHOTS);
1122         while (1) {
1123                 k = bch2_btree_iter_prev(&iter);
1124                 ret = bkey_err(k);
1125                 if (ret)
1126                         break;
1127
1128                 if (!k.k)
1129                         break;
1130
1131                 if (!bkey_eq(pos, k.k->p))
1132                         break;
1133
1134                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1135                         ret = 1;
1136                         break;
1137                 }
1138         }
1139         bch2_trans_iter_exit(trans, &iter);
1140
1141         return ret;
1142 }
1143
1144 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1145                                           enum btree_id id,
1146                                           struct bpos pos)
1147 {
1148         if (!btree_type_has_snapshots(id) ||
1149             pos.snapshot == U32_MAX ||
1150             !snapshot_t(trans->c, pos.snapshot)->children[0])
1151                 return 0;
1152
1153         return __check_pos_snapshot_overwritten(trans, id, pos);
1154 }
1155
1156 int bch2_trans_update_extent(struct btree_trans *trans,
1157                              struct btree_iter *orig_iter,
1158                              struct bkey_i *insert,
1159                              enum btree_update_flags flags)
1160 {
1161         struct bch_fs *c = trans->c;
1162         struct btree_iter iter, update_iter;
1163         struct bpos start = bkey_start_pos(&insert->k);
1164         struct bkey_i *update;
1165         struct bkey_s_c k;
1166         enum btree_id btree_id = orig_iter->btree_id;
1167         int ret = 0, compressed_sectors;
1168
1169         bch2_trans_iter_init(trans, &iter, btree_id, start,
1170                              BTREE_ITER_INTENT|
1171                              BTREE_ITER_WITH_UPDATES|
1172                              BTREE_ITER_NOT_EXTENTS);
1173         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1174         if ((ret = bkey_err(k)))
1175                 goto err;
1176         if (!k.k)
1177                 goto out;
1178
1179         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1180                 /*
1181                  * We can't merge extents if they belong to interior snapshot
1182                  * tree nodes, and there's a snapshot in which one extent is
1183                  * visible and the other is not - i.e. if visibility is
1184                  * different.
1185                  *
1186                  * Instead of checking if visibilitiy of the two extents is
1187                  * different, for now we just check if either has been
1188                  * overwritten:
1189                  */
1190                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1191                 if (ret < 0)
1192                         goto err;
1193                 if (ret)
1194                         goto nomerge1;
1195
1196                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1197                 if (ret < 0)
1198                         goto err;
1199                 if (ret)
1200                         goto nomerge1;
1201
1202                 update = bch2_bkey_make_mut(trans, k);
1203                 if ((ret = PTR_ERR_OR_ZERO(update)))
1204                         goto err;
1205
1206                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1207                         ret = bch2_btree_delete_at(trans, &iter, flags);
1208                         if (ret)
1209                                 goto err;
1210
1211                         insert = update;
1212                         goto next;
1213                 }
1214         }
1215 nomerge1:
1216         ret = 0;
1217         if (bkey_eq(k.k->p, start))
1218                 goto next;
1219
1220         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1221                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1222                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1223
1224                 /*
1225                  * If we're going to be splitting a compressed extent, note it
1226                  * so that __bch2_trans_commit() can increase our disk
1227                  * reservation:
1228                  */
1229                 if (((front_split && back_split) ||
1230                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1231                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1232                         trans->extra_journal_res += compressed_sectors;
1233
1234                 if (front_split) {
1235                         update = bch2_bkey_make_mut(trans, k);
1236                         if ((ret = PTR_ERR_OR_ZERO(update)))
1237                                 goto err;
1238
1239                         bch2_cut_back(start, update);
1240
1241                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1242                                              BTREE_ITER_NOT_EXTENTS|
1243                                              BTREE_ITER_ALL_SNAPSHOTS|
1244                                              BTREE_ITER_INTENT);
1245                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1246                                 bch2_trans_update(trans, &update_iter, update,
1247                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1248                                                   flags);
1249                         bch2_trans_iter_exit(trans, &update_iter);
1250
1251                         if (ret)
1252                                 goto err;
1253                 }
1254
1255                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1256                     (front_split || back_split)) {
1257                         update = bch2_bkey_make_mut(trans, k);
1258                         if ((ret = PTR_ERR_OR_ZERO(update)))
1259                                 goto err;
1260
1261                         bch2_cut_front(start, update);
1262                         bch2_cut_back(insert->k.p, update);
1263
1264                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1265                                              BTREE_ITER_NOT_EXTENTS|
1266                                              BTREE_ITER_ALL_SNAPSHOTS|
1267                                              BTREE_ITER_INTENT);
1268                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1269                                 bch2_trans_update(trans, &update_iter, update,
1270                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1271                                                   flags);
1272                         bch2_trans_iter_exit(trans, &update_iter);
1273                         if (ret)
1274                                 goto err;
1275                 }
1276
1277                 if (bkey_le(k.k->p, insert->k.p)) {
1278                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1279                         if ((ret = PTR_ERR_OR_ZERO(update)))
1280                                 goto err;
1281
1282                         bkey_init(&update->k);
1283                         update->k.p = k.k->p;
1284
1285                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1286                                 update->k.p.snapshot = insert->k.p.snapshot;
1287                                 update->k.type = KEY_TYPE_whiteout;
1288                         }
1289
1290                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1291                                              BTREE_ITER_NOT_EXTENTS|
1292                                              BTREE_ITER_INTENT);
1293                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1294                                 bch2_trans_update(trans, &update_iter, update,
1295                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1296                                                   flags);
1297                         bch2_trans_iter_exit(trans, &update_iter);
1298
1299                         if (ret)
1300                                 goto err;
1301                 }
1302
1303                 if (back_split) {
1304                         update = bch2_bkey_make_mut(trans, k);
1305                         if ((ret = PTR_ERR_OR_ZERO(update)))
1306                                 goto err;
1307
1308                         bch2_cut_front(insert->k.p, update);
1309
1310                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1311                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1312                                                   flags);
1313                         if (ret)
1314                                 goto err;
1315                         goto out;
1316                 }
1317 next:
1318                 bch2_btree_iter_advance(&iter);
1319                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1320                 if ((ret = bkey_err(k)))
1321                         goto err;
1322                 if (!k.k)
1323                         goto out;
1324         }
1325
1326         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1327                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1328                 if (ret < 0)
1329                         goto err;
1330                 if (ret)
1331                         goto nomerge2;
1332
1333                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1334                 if (ret < 0)
1335                         goto err;
1336                 if (ret)
1337                         goto nomerge2;
1338
1339                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1340         }
1341 nomerge2:
1342         ret = 0;
1343 out:
1344         if (!bkey_deleted(&insert->k)) {
1345                 /*
1346                  * Rewinding iterators is expensive: get a new one and the one
1347                  * that points to the start of insert will be cloned from:
1348                  */
1349                 bch2_trans_iter_exit(trans, &iter);
1350                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1351                                      BTREE_ITER_NOT_EXTENTS|
1352                                      BTREE_ITER_INTENT);
1353                 ret   = bch2_btree_iter_traverse(&iter) ?:
1354                         bch2_trans_update(trans, &iter, insert, flags);
1355         }
1356 err:
1357         bch2_trans_iter_exit(trans, &iter);
1358
1359         return ret;
1360 }
1361
1362 /*
1363  * When deleting, check if we need to emit a whiteout (because we're overwriting
1364  * something in an ancestor snapshot)
1365  */
1366 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1367                                       enum btree_id btree_id, struct bpos pos)
1368 {
1369         struct btree_iter iter;
1370         struct bkey_s_c k;
1371         u32 snapshot = pos.snapshot;
1372         int ret;
1373
1374         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1375                 return 0;
1376
1377         pos.snapshot++;
1378
1379         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1380                            BTREE_ITER_ALL_SNAPSHOTS|
1381                            BTREE_ITER_NOPRESERVE, k, ret) {
1382                 if (!bkey_eq(k.k->p, pos))
1383                         break;
1384
1385                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1386                                               k.k->p.snapshot)) {
1387                         ret = !bkey_whiteout(k.k);
1388                         break;
1389                 }
1390         }
1391         bch2_trans_iter_exit(trans, &iter);
1392
1393         return ret;
1394 }
1395
1396 static int __must_check
1397 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1398                                 struct bkey_i *k, enum btree_update_flags flags,
1399                                 unsigned long ip);
1400
1401 static noinline int flush_new_cached_update(struct btree_trans *trans,
1402                                             struct btree_path *path,
1403                                             struct btree_insert_entry *i,
1404                                             enum btree_update_flags flags,
1405                                             unsigned long ip)
1406 {
1407         struct btree_path *btree_path;
1408         int ret;
1409
1410         i->key_cache_already_flushed = true;
1411         i->flags |= BTREE_TRIGGER_NORUN;
1412
1413         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1414                                    BTREE_ITER_INTENT, _THIS_IP_);
1415
1416         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1417         if (ret)
1418                 goto err;
1419
1420         btree_path_set_should_be_locked(btree_path);
1421         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1422 err:
1423         bch2_path_put(trans, btree_path, true);
1424         return ret;
1425 }
1426
1427 static int __must_check
1428 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1429                                 struct bkey_i *k, enum btree_update_flags flags,
1430                                 unsigned long ip)
1431 {
1432         struct bch_fs *c = trans->c;
1433         struct btree_insert_entry *i, n;
1434         int cmp;
1435
1436         EBUG_ON(!path->should_be_locked);
1437         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1438         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1439
1440         n = (struct btree_insert_entry) {
1441                 .flags          = flags,
1442                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1443                 .btree_id       = path->btree_id,
1444                 .level          = path->level,
1445                 .cached         = path->cached,
1446                 .path           = path,
1447                 .k              = k,
1448                 .ip_allocated   = ip,
1449         };
1450
1451 #ifdef CONFIG_BCACHEFS_DEBUG
1452         trans_for_each_update(trans, i)
1453                 BUG_ON(i != trans->updates &&
1454                        btree_insert_entry_cmp(i - 1, i) >= 0);
1455 #endif
1456
1457         /*
1458          * Pending updates are kept sorted: first, find position of new update,
1459          * then delete/trim any updates the new update overwrites:
1460          */
1461         trans_for_each_update(trans, i) {
1462                 cmp = btree_insert_entry_cmp(&n, i);
1463                 if (cmp <= 0)
1464                         break;
1465         }
1466
1467         if (!cmp && i < trans->updates + trans->nr_updates) {
1468                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1469
1470                 bch2_path_put(trans, i->path, true);
1471                 i->flags        = n.flags;
1472                 i->cached       = n.cached;
1473                 i->k            = n.k;
1474                 i->path         = n.path;
1475                 i->ip_allocated = n.ip_allocated;
1476         } else {
1477                 array_insert_item(trans->updates, trans->nr_updates,
1478                                   i - trans->updates, n);
1479
1480                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1481                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1482
1483                 if (unlikely(trans->journal_replay_not_finished)) {
1484                         struct bkey_i *j_k =
1485                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1486
1487                         if (j_k) {
1488                                 i->old_k = j_k->k;
1489                                 i->old_v = &j_k->v;
1490                         }
1491                 }
1492         }
1493
1494         __btree_path_get(i->path, true);
1495
1496         /*
1497          * If a key is present in the key cache, it must also exist in the
1498          * btree - this is necessary for cache coherency. When iterating over
1499          * a btree that's cached in the key cache, the btree iter code checks
1500          * the key cache - but the key has to exist in the btree for that to
1501          * work:
1502          */
1503         if (unlikely(path->cached && bkey_deleted(&i->old_k)))
1504                 return flush_new_cached_update(trans, path, i, flags, ip);
1505
1506         return 0;
1507 }
1508
1509 static inline int __must_check
1510 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1511                           struct bkey_i *k, enum btree_update_flags flags)
1512 {
1513         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1514 }
1515
1516 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1517                                    struct bkey_i *k, enum btree_update_flags flags)
1518 {
1519         struct btree_path *path = iter->update_path ?: iter->path;
1520         struct bkey_cached *ck;
1521         int ret;
1522
1523         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1524                 return bch2_trans_update_extent(trans, iter, k, flags);
1525
1526         if (bkey_deleted(&k->k) &&
1527             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1528             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1529                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1530                 if (unlikely(ret < 0))
1531                         return ret;
1532
1533                 if (ret)
1534                         k->k.type = KEY_TYPE_whiteout;
1535         }
1536
1537         /*
1538          * Ensure that updates to cached btrees go to the key cache:
1539          */
1540         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1541             !path->cached &&
1542             !path->level &&
1543             btree_id_cached(trans->c, path->btree_id)) {
1544                 if (!iter->key_cache_path ||
1545                     !iter->key_cache_path->should_be_locked ||
1546                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1547                         if (!iter->key_cache_path)
1548                                 iter->key_cache_path =
1549                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1550                                                       BTREE_ITER_INTENT|
1551                                                       BTREE_ITER_CACHED, _THIS_IP_);
1552
1553                         iter->key_cache_path =
1554                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1555                                                         iter->flags & BTREE_ITER_INTENT,
1556                                                         _THIS_IP_);
1557
1558                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1559                                                        BTREE_ITER_CACHED);
1560                         if (unlikely(ret))
1561                                 return ret;
1562
1563                         ck = (void *) iter->key_cache_path->l[0].b;
1564
1565                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1566                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1567                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1568                         }
1569
1570                         btree_path_set_should_be_locked(iter->key_cache_path);
1571                 }
1572
1573                 path = iter->key_cache_path;
1574         }
1575
1576         return bch2_trans_update_by_path(trans, path, k, flags);
1577 }
1578
1579 void bch2_trans_commit_hook(struct btree_trans *trans,
1580                             struct btree_trans_commit_hook *h)
1581 {
1582         h->next = trans->hooks;
1583         trans->hooks = h;
1584 }
1585
1586 int __bch2_btree_insert(struct btree_trans *trans,
1587                         enum btree_id id, struct bkey_i *k)
1588 {
1589         struct btree_iter iter;
1590         int ret;
1591
1592         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1593                              BTREE_ITER_INTENT);
1594         ret   = bch2_btree_iter_traverse(&iter) ?:
1595                 bch2_trans_update(trans, &iter, k, 0);
1596         bch2_trans_iter_exit(trans, &iter);
1597         return ret;
1598 }
1599
1600 /**
1601  * bch2_btree_insert - insert keys into the extent btree
1602  * @c:                  pointer to struct bch_fs
1603  * @id:                 btree to insert into
1604  * @insert_keys:        list of keys to insert
1605  * @hook:               insert callback
1606  */
1607 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1608                       struct bkey_i *k,
1609                       struct disk_reservation *disk_res,
1610                       u64 *journal_seq, int flags)
1611 {
1612         return bch2_trans_do(c, disk_res, journal_seq, flags,
1613                              __bch2_btree_insert(&trans, id, k));
1614 }
1615
1616 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1617                                 unsigned len, unsigned update_flags)
1618 {
1619         struct bkey_i *k;
1620
1621         k = bch2_trans_kmalloc(trans, sizeof(*k));
1622         if (IS_ERR(k))
1623                 return PTR_ERR(k);
1624
1625         bkey_init(&k->k);
1626         k->k.p = iter->pos;
1627         bch2_key_resize(&k->k, len);
1628         return bch2_trans_update(trans, iter, k, update_flags);
1629 }
1630
1631 int bch2_btree_delete_at(struct btree_trans *trans,
1632                          struct btree_iter *iter, unsigned update_flags)
1633 {
1634         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1635 }
1636
1637 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1638                                   struct bpos start, struct bpos end,
1639                                   unsigned update_flags,
1640                                   u64 *journal_seq)
1641 {
1642         u32 restart_count = trans->restart_count;
1643         struct btree_iter iter;
1644         struct bkey_s_c k;
1645         int ret = 0;
1646
1647         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1648         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1649                 struct disk_reservation disk_res =
1650                         bch2_disk_reservation_init(trans->c, 0);
1651                 struct bkey_i delete;
1652
1653                 ret = bkey_err(k);
1654                 if (ret)
1655                         goto err;
1656
1657                 bkey_init(&delete.k);
1658
1659                 /*
1660                  * This could probably be more efficient for extents:
1661                  */
1662
1663                 /*
1664                  * For extents, iter.pos won't necessarily be the same as
1665                  * bkey_start_pos(k.k) (for non extents they always will be the
1666                  * same). It's important that we delete starting from iter.pos
1667                  * because the range we want to delete could start in the middle
1668                  * of k.
1669                  *
1670                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1671                  * bkey_start_pos(k.k)).
1672                  */
1673                 delete.k.p = iter.pos;
1674
1675                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1676                         unsigned max_sectors =
1677                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1678
1679                         /* create the biggest key we can */
1680                         bch2_key_resize(&delete.k, max_sectors);
1681                         bch2_cut_back(end, &delete);
1682
1683                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1684                         if (ret)
1685                                 goto err;
1686                 }
1687
1688                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1689                         bch2_trans_commit(trans, &disk_res, journal_seq,
1690                                           BTREE_INSERT_NOFAIL);
1691                 bch2_disk_reservation_put(trans->c, &disk_res);
1692 err:
1693                 /*
1694                  * the bch2_trans_begin() call is in a weird place because we
1695                  * need to call it after every transaction commit, to avoid path
1696                  * overflow, but don't want to call it if the delete operation
1697                  * is a no-op and we have no work to do:
1698                  */
1699                 bch2_trans_begin(trans);
1700
1701                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1702                         ret = 0;
1703                 if (ret)
1704                         break;
1705         }
1706         bch2_trans_iter_exit(trans, &iter);
1707
1708         if (!ret && trans_was_restarted(trans, restart_count))
1709                 ret = -BCH_ERR_transaction_restart_nested;
1710         return ret;
1711 }
1712
1713 /*
1714  * bch_btree_delete_range - delete everything within a given range
1715  *
1716  * Range is a half open interval - [start, end)
1717  */
1718 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1719                             struct bpos start, struct bpos end,
1720                             unsigned update_flags,
1721                             u64 *journal_seq)
1722 {
1723         int ret = bch2_trans_run(c,
1724                         bch2_btree_delete_range_trans(&trans, id, start, end,
1725                                                       update_flags, journal_seq));
1726         if (ret == -BCH_ERR_transaction_restart_nested)
1727                 ret = 0;
1728         return ret;
1729 }
1730
1731 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1732 {
1733         unsigned len = strlen(msg);
1734         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1735         struct jset_entry_log *l;
1736         int ret;
1737
1738         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1739         if (ret)
1740                 return ret;
1741
1742         l = (void *) &darray_top(trans->extra_journal_entries);
1743         l->entry.u64s           = cpu_to_le16(u64s);
1744         l->entry.btree_id       = 0;
1745         l->entry.level          = 1;
1746         l->entry.type           = BCH_JSET_ENTRY_log;
1747         l->entry.pad[0]         = 0;
1748         l->entry.pad[1]         = 0;
1749         l->entry.pad[2]         = 0;
1750         memcpy(l->d, msg, len);
1751         while (len & 7)
1752                 l->d[len++] = '\0';
1753
1754         trans->extra_journal_entries.nr += jset_u64s(u64s);
1755         return 0;
1756 }