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_cmp(insert->k.p, b->data->min_key) < 0);
  96         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
  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 |= 1 << BTREE_NODE_need_write;
 182         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 183
 184         b->write_type = BTREE_WRITE_journal_reclaim;
 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_cmp(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         trans_for_each_update(trans, i)
 521                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 522                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 523                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 524
 525         return 0;
 526 }
 527
 528 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 529 {
 530         struct bch_fs *c = trans->c;
 531         struct btree_insert_entry *i;
 532         int ret = 0;
 533
 534         trans_for_each_update(trans, i) {
 535                 /*
 536                  * XXX: synchronization of cached update triggers with gc
 537                  * XXX: synchronization of interior node updates with gc
 538                  */
 539                 BUG_ON(i->cached || i->level);
 540
 541                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 542                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 543                         if (ret)
 544                                 break;
 545                 }
 546         }
 547
 548         return ret;
 549 }
 550
 551 static inline int
 552 bch2_trans_commit_write_locked(struct btree_trans *trans,
 553                                struct btree_insert_entry **stopped_at,
 554                                unsigned long trace_ip)
 555 {
 556         struct bch_fs *c = trans->c;
 557         struct btree_insert_entry *i;
 558         struct btree_trans_commit_hook *h;
 559         unsigned u64s = 0;
 560         bool marking = false;
 561         int ret;
 562
 563         if (race_fault()) {
 564                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 565                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 566         }
 567
 568         /*
 569          * Check if the insert will fit in the leaf node with the write lock
 570          * held, otherwise another thread could write the node changing the
 571          * amount of space available:
 572          */
 573
 574         prefetch(&trans->c->journal.flags);
 575
 576         h = trans->hooks;
 577         while (h) {
 578                 ret = h->fn(trans, h);
 579                 if (ret)
 580                         return ret;
 581                 h = h->next;
 582         }
 583
 584         trans_for_each_update(trans, i) {
 585                 /* Multiple inserts might go to same leaf: */
 586                 if (!same_leaf_as_prev(trans, i))
 587                         u64s = 0;
 588
 589                 u64s += i->k->k.u64s;
 590                 ret = !i->cached
 591                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 592                         : btree_key_can_insert_cached(trans, i->path, u64s);
 593                 if (ret) {
 594                         *stopped_at = i;
 595                         return ret;
 596                 }
 597
 598                 if (btree_node_type_needs_gc(i->bkey_type))
 599                         marking = true;
 600
 601                 /*
 602                  * Revalidate before calling mem triggers - XXX, ugly:
 603                  *
 604                  * - successful btree node splits don't cause transaction
 605                  *   restarts and will have invalidated the pointer to the bkey
 606                  *   value
 607                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 608                  *   when it has to resort
 609                  * - btree_key_can_insert_cached() when it has to reallocate
 610                  *
 611                  *   Ugly because we currently have no way to tell if the
 612                  *   pointer's been invalidated, which means it's debatabale
 613                  *   whether we should be stashing the old key at all.
 614                  */
 615                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 616
 617                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 618                         struct bkey_i *j_k =
 619                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 620                                                             i->k->k.p);
 621
 622                         if (j_k) {
 623                                 i->old_k = j_k->k;
 624                                 i->old_v = &j_k->v;
 625                         }
 626                 }
 627         }
 628
 629         /*
 630          * Don't get journal reservation until after we know insert will
 631          * succeed:
 632          */
 633         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 634                 ret = bch2_trans_journal_res_get(trans,
 635                                 JOURNAL_RES_GET_NONBLOCK);
 636                 if (ret)
 637                         return ret;
 638
 639                 journal_transaction_name(trans);
 640         } else {
 641                 trans->journal_res.seq = c->journal.replay_journal_seq;
 642         }
 643
 644         if (unlikely(trans->extra_journal_entries.nr)) {
 645                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 646                                   trans->extra_journal_entries.data,
 647                                   trans->extra_journal_entries.nr);
 648
 649                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 650                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 651         }
 652
 653         /*
 654          * Not allowed to fail after we've gotten our journal reservation - we
 655          * have to use it:
 656          */
 657
 658         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 659                 if (bch2_journal_seq_verify)
 660                         trans_for_each_update(trans, i)
 661                                 i->k->k.version.lo = trans->journal_res.seq;
 662                 else if (bch2_inject_invalid_keys)
 663                         trans_for_each_update(trans, i)
 664                                 i->k->k.version = MAX_VERSION;
 665         }
 666
 667         if (trans->fs_usage_deltas &&
 668             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 669                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 670
 671         trans_for_each_update(trans, i)
 672                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 673                         ret = run_one_mem_trigger(trans, i, i->flags);
 674                         if (ret)
 675                                 return ret;
 676                 }
 677
 678         if (unlikely(c->gc_pos.phase)) {
 679                 ret = bch2_trans_commit_run_gc_triggers(trans);
 680                 if  (ret)
 681                         return ret;
 682         }
 683
 684         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 685                 trans_for_each_update(trans, i) {
 686                         struct journal *j = &c->journal;
 687                         struct jset_entry *entry;
 688
 689                         if (i->key_cache_already_flushed)
 690                                 continue;
 691
 692                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 693                                                BCH_JSET_ENTRY_overwrite,
 694                                                i->btree_id, i->level,
 695                                                i->old_k.u64s);
 696                         bkey_reassemble(&entry->start[0],
 697                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 698
 699                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 700                                                BCH_JSET_ENTRY_btree_keys,
 701                                                i->btree_id, i->level,
 702                                                i->k->k.u64s);
 703                         bkey_copy(&entry->start[0], i->k);
 704                 }
 705
 706                 if (trans->journal_seq)
 707                         *trans->journal_seq = trans->journal_res.seq;
 708         }
 709
 710         trans_for_each_update(trans, i) {
 711                 i->k->k.needs_whiteout = false;
 712
 713                 if (!i->cached)
 714                         btree_insert_key_leaf(trans, i);
 715                 else if (!i->key_cache_already_flushed)
 716                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 717                 else {
 718                         bch2_btree_key_cache_drop(trans, i->path);
 719                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 720                 }
 721         }
 722
 723         return ret;
 724 }
 725
 726 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 727 {
 728         while (--i >= trans->updates) {
 729                 if (same_leaf_as_prev(trans, i))
 730                         continue;
 731
 732                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 733         }
 734
 735         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 736         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 737 }
 738
 739 static inline int trans_lock_write(struct btree_trans *trans)
 740 {
 741         struct btree_insert_entry *i;
 742
 743         trans_for_each_update(trans, i) {
 744                 if (same_leaf_as_prev(trans, i))
 745                         continue;
 746
 747                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 748                         return trans_lock_write_fail(trans, i);
 749
 750                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 751         }
 752
 753         return 0;
 754 }
 755
 756 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 757 {
 758         struct btree_insert_entry *i;
 759
 760         trans_for_each_update(trans, i)
 761                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 762 }
 763
 764 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 765                                                    struct btree_insert_entry *i,
 766                                                    struct printbuf *err)
 767 {
 768         struct bch_fs *c = trans->c;
 769         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 770
 771         printbuf_reset(err);
 772         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 773                    trans->fn, (void *) i->ip_allocated);
 774         prt_newline(err);
 775         printbuf_indent_add(err, 2);
 776
 777         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 778         prt_newline(err);
 779
 780         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 781                           i->bkey_type, rw, err);
 782         bch2_print_string_as_lines(KERN_ERR, err->buf);
 783
 784         bch2_inconsistent_error(c);
 785         bch2_dump_trans_updates(trans);
 786         printbuf_exit(err);
 787
 788         return -EINVAL;
 789 }
 790
 791 /*
 792  * Get journal reservation, take write locks, and attempt to do btree update(s):
 793  */
 794 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 795                                        struct btree_insert_entry **stopped_at,
 796                                        unsigned long trace_ip)
 797 {
 798         struct bch_fs *c = trans->c;
 799         struct btree_insert_entry *i;
 800         struct printbuf buf = PRINTBUF;
 801         int ret, u64s_delta = 0;
 802         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 803
 804         trans_for_each_update(trans, i) {
 805                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 806                                                i->bkey_type, rw, &buf)))
 807                         return bch2_trans_commit_bkey_invalid(trans, i, &buf);
 808                 btree_insert_entry_checks(trans, i);
 809         }
 810
 811         printbuf_exit(&buf);
 812
 813         trans_for_each_update(trans, i) {
 814                 if (i->cached)
 815                         continue;
 816
 817                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 818                 u64s_delta -= i->old_btree_u64s;
 819
 820                 if (!same_leaf_as_next(trans, i)) {
 821                         if (u64s_delta <= 0) {
 822                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 823                                                         i->level, trans->flags);
 824                                 if (unlikely(ret))
 825                                         return ret;
 826                         }
 827
 828                         u64s_delta = 0;
 829                 }
 830         }
 831
 832         ret = bch2_journal_preres_get(&c->journal,
 833                         &trans->journal_preres, trans->journal_preres_u64s,
 834                         JOURNAL_RES_GET_NONBLOCK|
 835                         (trans->flags & JOURNAL_WATERMARK_MASK));
 836         if (unlikely(ret == -EAGAIN))
 837                 ret = bch2_trans_journal_preres_get_cold(trans,
 838                                                 trans->journal_preres_u64s, trace_ip);
 839         if (unlikely(ret))
 840                 return ret;
 841
 842         ret = trans_lock_write(trans);
 843         if (unlikely(ret))
 844                 return ret;
 845
 846         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 847
 848         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 849                 bch2_drop_overwrites_from_journal(trans);
 850
 851         trans_for_each_update(trans, i)
 852                 if (!same_leaf_as_prev(trans, i))
 853                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 854                                                         insert_l(i)->b);
 855
 856         if (!ret && trans->journal_pin)
 857                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 858                                      trans->journal_pin, NULL);
 859
 860         /*
 861          * Drop journal reservation after dropping write locks, since dropping
 862          * the journal reservation may kick off a journal write:
 863          */
 864         bch2_journal_res_put(&c->journal, &trans->journal_res);
 865
 866         if (unlikely(ret))
 867                 return ret;
 868
 869         bch2_trans_downgrade(trans);
 870
 871         return 0;
 872 }
 873
 874 static int journal_reclaim_wait_done(struct bch_fs *c)
 875 {
 876         int ret = bch2_journal_error(&c->journal) ?:
 877                 !bch2_btree_key_cache_must_wait(c);
 878
 879         if (!ret)
 880                 journal_reclaim_kick(&c->journal);
 881         return ret;
 882 }
 883
 884 static noinline
 885 int bch2_trans_commit_error(struct btree_trans *trans,
 886                             struct btree_insert_entry *i,
 887                             int ret, unsigned long trace_ip)
 888 {
 889         struct bch_fs *c = trans->c;
 890
 891         switch (ret) {
 892         case BTREE_INSERT_BTREE_NODE_FULL:
 893                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 894                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 895                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 896                 break;
 897         case BTREE_INSERT_NEED_MARK_REPLICAS:
 898                 bch2_trans_unlock(trans);
 899
 900                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 901                 if (ret)
 902                         break;
 903
 904                 ret = bch2_trans_relock(trans);
 905                 if (ret)
 906                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 907                 break;
 908         case BTREE_INSERT_NEED_JOURNAL_RES:
 909                 bch2_trans_unlock(trans);
 910
 911                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 912                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 913                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 914                         break;
 915                 }
 916
 917                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 918                 if (ret)
 919                         break;
 920
 921                 ret = bch2_trans_relock(trans);
 922                 if (ret)
 923                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 924                 break;
 925         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 926                 bch2_trans_unlock(trans);
 927
 928                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 929
 930                 wait_event_freezable(c->journal.reclaim_wait,
 931                                      (ret = journal_reclaim_wait_done(c)));
 932                 if (ret < 0)
 933                         break;
 934
 935                 ret = bch2_trans_relock(trans);
 936                 if (ret)
 937                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 938                 break;
 939         default:
 940                 BUG_ON(ret >= 0);
 941                 break;
 942         }
 943
 944         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 945
 946         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 947                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 948                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 949                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 950
 951         return ret;
 952 }
 953
 954 static noinline int
 955 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 956 {
 957         struct bch_fs *c = trans->c;
 958         int ret;
 959
 960         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 961             test_bit(BCH_FS_STARTED, &c->flags))
 962                 return -EROFS;
 963
 964         bch2_trans_unlock(trans);
 965
 966         ret =   bch2_fs_read_write_early(c) ?:
 967                 bch2_trans_relock(trans);
 968         if (ret)
 969                 return ret;
 970
 971         percpu_ref_get(&c->writes);
 972         return 0;
 973 }
 974
 975 /*
 976  * This is for updates done in the early part of fsck - btree_gc - before we've
 977  * gone RW. we only add the new key to the list of keys for journal replay to
 978  * do.
 979  */
 980 static noinline int
 981 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 982 {
 983         struct bch_fs *c = trans->c;
 984         struct btree_insert_entry *i;
 985         int ret = 0;
 986
 987         trans_for_each_update(trans, i) {
 988                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 989                 if (ret)
 990                         break;
 991         }
 992
 993         return ret;
 994 }
 995
 996 int __bch2_trans_commit(struct btree_trans *trans)
 997 {
 998         struct bch_fs *c = trans->c;
 999         struct btree_insert_entry *i = NULL;
1000         unsigned u64s;
1001         int ret = 0;
1002
1003         if (!trans->nr_updates &&
1004             !trans->extra_journal_entries.nr)
1005                 goto out_reset;
1006
1007         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1008                 lockdep_assert_held(&c->gc_lock);
1009
1010         ret = bch2_trans_commit_run_triggers(trans);
1011         if (ret)
1012                 goto out_reset;
1013
1014         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1015                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1016                 goto out_reset;
1017         }
1018
1019         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1020             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1021                 ret = bch2_trans_commit_get_rw_cold(trans);
1022                 if (ret)
1023                         goto out_reset;
1024         }
1025
1026         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1027
1028         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1029
1030         trans->journal_u64s             = trans->extra_journal_entries.nr;
1031         trans->journal_preres_u64s      = 0;
1032
1033         /* For journalling transaction name: */
1034         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1035
1036         trans_for_each_update(trans, i) {
1037                 BUG_ON(!i->path->should_be_locked);
1038
1039                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1040                 if (unlikely(ret))
1041                         goto out;
1042
1043                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1044
1045                 if (i->key_cache_already_flushed)
1046                         continue;
1047
1048                 /* we're going to journal the key being updated: */
1049                 u64s = jset_u64s(i->k->k.u64s);
1050                 if (i->cached &&
1051                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1052                         trans->journal_preres_u64s += u64s;
1053                 trans->journal_u64s += u64s;
1054
1055                 /* and we're also going to log the overwrite: */
1056                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1057         }
1058
1059         if (trans->extra_journal_res) {
1060                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1061                                 trans->extra_journal_res,
1062                                 (trans->flags & BTREE_INSERT_NOFAIL)
1063                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1064                 if (ret)
1065                         goto err;
1066         }
1067 retry:
1068         BUG_ON(trans->restarted);
1069         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1070
1071         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1072
1073         /* make sure we didn't drop or screw up locks: */
1074         bch2_trans_verify_locks(trans);
1075
1076         if (ret)
1077                 goto err;
1078
1079         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1080 out:
1081         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1082
1083         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1084                 percpu_ref_put(&c->writes);
1085 out_reset:
1086         bch2_trans_reset_updates(trans);
1087
1088         if (trans->fs_usage_deltas) {
1089                 trans->fs_usage_deltas->used = 0;
1090                 memset((void *) trans->fs_usage_deltas +
1091                        offsetof(struct replicas_delta_list, memset_start), 0,
1092                        (void *) &trans->fs_usage_deltas->memset_end -
1093                        (void *) &trans->fs_usage_deltas->memset_start);
1094         }
1095
1096         return ret;
1097 err:
1098         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1099         if (ret)
1100                 goto out;
1101
1102         goto retry;
1103 }
1104
1105 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1106                                           enum btree_id id,
1107                                           struct bpos pos)
1108 {
1109         struct bch_fs *c = trans->c;
1110         struct btree_iter iter;
1111         struct bkey_s_c k;
1112         int ret;
1113
1114         bch2_trans_iter_init(trans, &iter, id, pos,
1115                              BTREE_ITER_NOT_EXTENTS|
1116                              BTREE_ITER_ALL_SNAPSHOTS);
1117         while (1) {
1118                 k = bch2_btree_iter_prev(&iter);
1119                 ret = bkey_err(k);
1120                 if (ret)
1121                         break;
1122
1123                 if (!k.k)
1124                         break;
1125
1126                 if (bkey_cmp(pos, k.k->p))
1127                         break;
1128
1129                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1130                         ret = 1;
1131                         break;
1132                 }
1133         }
1134         bch2_trans_iter_exit(trans, &iter);
1135
1136         return ret;
1137 }
1138
1139 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1140                                           enum btree_id id,
1141                                           struct bpos pos)
1142 {
1143         if (!btree_type_has_snapshots(id) ||
1144             pos.snapshot == U32_MAX ||
1145             !snapshot_t(trans->c, pos.snapshot)->children[0])
1146                 return 0;
1147
1148         return __check_pos_snapshot_overwritten(trans, id, pos);
1149 }
1150
1151 int bch2_trans_update_extent(struct btree_trans *trans,
1152                              struct btree_iter *orig_iter,
1153                              struct bkey_i *insert,
1154                              enum btree_update_flags flags)
1155 {
1156         struct bch_fs *c = trans->c;
1157         struct btree_iter iter, update_iter;
1158         struct bpos start = bkey_start_pos(&insert->k);
1159         struct bkey_i *update;
1160         struct bkey_s_c k;
1161         enum btree_id btree_id = orig_iter->btree_id;
1162         int ret = 0, compressed_sectors;
1163
1164         bch2_trans_iter_init(trans, &iter, btree_id, start,
1165                              BTREE_ITER_INTENT|
1166                              BTREE_ITER_WITH_UPDATES|
1167                              BTREE_ITER_NOT_EXTENTS);
1168         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1169         if ((ret = bkey_err(k)))
1170                 goto err;
1171         if (!k.k)
1172                 goto out;
1173
1174         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1175                 /*
1176                  * We can't merge extents if they belong to interior snapshot
1177                  * tree nodes, and there's a snapshot in which one extent is
1178                  * visible and the other is not - i.e. if visibility is
1179                  * different.
1180                  *
1181                  * Instead of checking if visibilitiy of the two extents is
1182                  * different, for now we just check if either has been
1183                  * overwritten:
1184                  */
1185                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1186                 if (ret < 0)
1187                         goto err;
1188                 if (ret)
1189                         goto nomerge1;
1190
1191                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1192                 if (ret < 0)
1193                         goto err;
1194                 if (ret)
1195                         goto nomerge1;
1196
1197                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1198                 if ((ret = PTR_ERR_OR_ZERO(update)))
1199                         goto err;
1200
1201                 bkey_reassemble(update, k);
1202
1203                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1204                         ret = bch2_btree_delete_at(trans, &iter, flags);
1205                         if (ret)
1206                                 goto err;
1207
1208                         insert = update;
1209                         goto next;
1210                 }
1211         }
1212 nomerge1:
1213         ret = 0;
1214         if (!bkey_cmp(k.k->p, start))
1215                 goto next;
1216
1217         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1218                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1219                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1220
1221                 /*
1222                  * If we're going to be splitting a compressed extent, note it
1223                  * so that __bch2_trans_commit() can increase our disk
1224                  * reservation:
1225                  */
1226                 if (((front_split && back_split) ||
1227                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1228                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1229                         trans->extra_journal_res += compressed_sectors;
1230
1231                 if (front_split) {
1232                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1233                         if ((ret = PTR_ERR_OR_ZERO(update)))
1234                                 goto err;
1235
1236                         bkey_reassemble(update, k);
1237
1238                         bch2_cut_back(start, update);
1239
1240                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1241                                              BTREE_ITER_NOT_EXTENTS|
1242                                              BTREE_ITER_ALL_SNAPSHOTS|
1243                                              BTREE_ITER_INTENT);
1244                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1245                                 bch2_trans_update(trans, &update_iter, update,
1246                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1247                                                   flags);
1248                         bch2_trans_iter_exit(trans, &update_iter);
1249
1250                         if (ret)
1251                                 goto err;
1252                 }
1253
1254                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1255                     (front_split || back_split)) {
1256                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1257                         if ((ret = PTR_ERR_OR_ZERO(update)))
1258                                 goto err;
1259
1260                         bkey_reassemble(update, k);
1261
1262                         bch2_cut_front(start, update);
1263                         bch2_cut_back(insert->k.p, update);
1264
1265                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1266                                              BTREE_ITER_NOT_EXTENTS|
1267                                              BTREE_ITER_ALL_SNAPSHOTS|
1268                                              BTREE_ITER_INTENT);
1269                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1270                                 bch2_trans_update(trans, &update_iter, update,
1271                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1272                                                   flags);
1273                         bch2_trans_iter_exit(trans, &update_iter);
1274                         if (ret)
1275                                 goto err;
1276                 }
1277
1278                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1279                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1280                         if ((ret = PTR_ERR_OR_ZERO(update)))
1281                                 goto err;
1282
1283                         bkey_init(&update->k);
1284                         update->k.p = k.k->p;
1285
1286                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1287                                 update->k.p.snapshot = insert->k.p.snapshot;
1288                                 update->k.type = KEY_TYPE_whiteout;
1289                         }
1290
1291                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1292                                              BTREE_ITER_NOT_EXTENTS|
1293                                              BTREE_ITER_INTENT);
1294                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1295                                 bch2_trans_update(trans, &update_iter, update,
1296                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1297                                                   flags);
1298                         bch2_trans_iter_exit(trans, &update_iter);
1299
1300                         if (ret)
1301                                 goto err;
1302                 }
1303
1304                 if (back_split) {
1305                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1306                         if ((ret = PTR_ERR_OR_ZERO(update)))
1307                                 goto err;
1308
1309                         bkey_reassemble(update, k);
1310                         bch2_cut_front(insert->k.p, update);
1311
1312                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1313                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1314                                                   flags);
1315                         if (ret)
1316                                 goto err;
1317                         goto out;
1318                 }
1319 next:
1320                 bch2_btree_iter_advance(&iter);
1321                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1322                 if ((ret = bkey_err(k)))
1323                         goto err;
1324                 if (!k.k)
1325                         goto out;
1326         }
1327
1328         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1329                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1330                 if (ret < 0)
1331                         goto err;
1332                 if (ret)
1333                         goto nomerge2;
1334
1335                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1336                 if (ret < 0)
1337                         goto err;
1338                 if (ret)
1339                         goto nomerge2;
1340
1341                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1342         }
1343 nomerge2:
1344         ret = 0;
1345 out:
1346         if (!bkey_deleted(&insert->k)) {
1347                 /*
1348                  * Rewinding iterators is expensive: get a new one and the one
1349                  * that points to the start of insert will be cloned from:
1350                  */
1351                 bch2_trans_iter_exit(trans, &iter);
1352                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1353                                      BTREE_ITER_NOT_EXTENTS|
1354                                      BTREE_ITER_INTENT);
1355                 ret   = bch2_btree_iter_traverse(&iter) ?:
1356                         bch2_trans_update(trans, &iter, insert, flags);
1357         }
1358 err:
1359         bch2_trans_iter_exit(trans, &iter);
1360
1361         return ret;
1362 }
1363
1364 /*
1365  * When deleting, check if we need to emit a whiteout (because we're overwriting
1366  * something in an ancestor snapshot)
1367  */
1368 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1369                                       enum btree_id btree_id, struct bpos pos)
1370 {
1371         struct btree_iter iter;
1372         struct bkey_s_c k;
1373         u32 snapshot = pos.snapshot;
1374         int ret;
1375
1376         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1377                 return 0;
1378
1379         pos.snapshot++;
1380
1381         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1382                            BTREE_ITER_ALL_SNAPSHOTS|
1383                            BTREE_ITER_NOPRESERVE, k, ret) {
1384                 if (bkey_cmp(k.k->p, pos))
1385                         break;
1386
1387                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1388                                               k.k->p.snapshot)) {
1389                         ret = !bkey_whiteout(k.k);
1390                         break;
1391                 }
1392         }
1393         bch2_trans_iter_exit(trans, &iter);
1394
1395         return ret;
1396 }
1397
1398 static int __must_check
1399 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1400                                 struct bkey_i *k, enum btree_update_flags flags,
1401                                 unsigned long ip);
1402
1403 static noinline int flush_new_cached_update(struct btree_trans *trans,
1404                                             struct btree_path *path,
1405                                             struct btree_insert_entry *i,
1406                                             enum btree_update_flags flags,
1407                                             unsigned long ip)
1408 {
1409         struct btree_path *btree_path;
1410         int ret;
1411
1412         i->key_cache_already_flushed = true;
1413         i->flags |= BTREE_TRIGGER_NORUN;
1414
1415         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1416                                    BTREE_ITER_INTENT, _THIS_IP_);
1417
1418         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1419         if (ret)
1420                 goto err;
1421
1422         btree_path_set_should_be_locked(btree_path);
1423         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1424 err:
1425         bch2_path_put(trans, btree_path, true);
1426         return ret;
1427 }
1428
1429 static int __must_check
1430 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1431                                 struct bkey_i *k, enum btree_update_flags flags,
1432                                 unsigned long ip)
1433 {
1434         struct bch_fs *c = trans->c;
1435         struct btree_insert_entry *i, n;
1436
1437         BUG_ON(!path->should_be_locked);
1438
1439         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1440         BUG_ON(bpos_cmp(k->k.p, path->pos));
1441
1442         n = (struct btree_insert_entry) {
1443                 .flags          = flags,
1444                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1445                 .btree_id       = path->btree_id,
1446                 .level          = path->level,
1447                 .cached         = path->cached,
1448                 .path           = path,
1449                 .k              = k,
1450                 .ip_allocated   = ip,
1451         };
1452
1453 #ifdef CONFIG_BCACHEFS_DEBUG
1454         trans_for_each_update(trans, i)
1455                 BUG_ON(i != trans->updates &&
1456                        btree_insert_entry_cmp(i - 1, i) >= 0);
1457 #endif
1458
1459         /*
1460          * Pending updates are kept sorted: first, find position of new update,
1461          * then delete/trim any updates the new update overwrites:
1462          */
1463         trans_for_each_update(trans, i)
1464                 if (btree_insert_entry_cmp(&n, i) <= 0)
1465                         break;
1466
1467         if (i < trans->updates + trans->nr_updates &&
1468             !btree_insert_entry_cmp(&n, i)) {
1469                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1470
1471                 bch2_path_put(trans, i->path, true);
1472                 i->flags        = n.flags;
1473                 i->cached       = n.cached;
1474                 i->k            = n.k;
1475                 i->path         = n.path;
1476                 i->ip_allocated = n.ip_allocated;
1477         } else {
1478                 array_insert_item(trans->updates, trans->nr_updates,
1479                                   i - trans->updates, n);
1480
1481                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1482                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1483
1484                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1485                         struct bkey_i *j_k =
1486                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1487
1488                         if (j_k) {
1489                                 i->old_k = j_k->k;
1490                                 i->old_v = &j_k->v;
1491                         }
1492                 }
1493         }
1494
1495         __btree_path_get(i->path, true);
1496
1497         /*
1498          * If a key is present in the key cache, it must also exist in the
1499          * btree - this is necessary for cache coherency. When iterating over
1500          * a btree that's cached in the key cache, the btree iter code checks
1501          * the key cache - but the key has to exist in the btree for that to
1502          * work:
1503          */
1504         if (unlikely(path->cached && bkey_deleted(&i->old_k)))
1505                 return flush_new_cached_update(trans, path, i, flags, ip);
1506
1507         return 0;
1508 }
1509
1510 static int __must_check
1511 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1512                           struct bkey_i *k, enum btree_update_flags flags)
1513 {
1514         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1515 }
1516
1517 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1518                                    struct bkey_i *k, enum btree_update_flags flags)
1519 {
1520         struct btree_path *path = iter->update_path ?: iter->path;
1521         struct bkey_cached *ck;
1522         int ret;
1523
1524         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1525                 return bch2_trans_update_extent(trans, iter, k, flags);
1526
1527         if (bkey_deleted(&k->k) &&
1528             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1529             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1530                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1531                 if (unlikely(ret < 0))
1532                         return ret;
1533
1534                 if (ret)
1535                         k->k.type = KEY_TYPE_whiteout;
1536         }
1537
1538         /*
1539          * Ensure that updates to cached btrees go to the key cache:
1540          */
1541         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1542             !path->cached &&
1543             !path->level &&
1544             btree_id_cached(trans->c, path->btree_id)) {
1545                 if (!iter->key_cache_path ||
1546                     !iter->key_cache_path->should_be_locked ||
1547                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1548                         if (!iter->key_cache_path)
1549                                 iter->key_cache_path =
1550                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1551                                                       BTREE_ITER_INTENT|
1552                                                       BTREE_ITER_CACHED, _THIS_IP_);
1553
1554                         iter->key_cache_path =
1555                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1556                                                         iter->flags & BTREE_ITER_INTENT,
1557                                                         _THIS_IP_);
1558
1559                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1560                                                        BTREE_ITER_CACHED);
1561                         if (unlikely(ret))
1562                                 return ret;
1563
1564                         ck = (void *) iter->key_cache_path->l[0].b;
1565
1566                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1567                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1568                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1569                         }
1570
1571                         btree_path_set_should_be_locked(iter->key_cache_path);
1572                 }
1573
1574                 path = iter->key_cache_path;
1575         }
1576
1577         return bch2_trans_update_by_path(trans, path, k, flags);
1578 }
1579
1580 void bch2_trans_commit_hook(struct btree_trans *trans,
1581                             struct btree_trans_commit_hook *h)
1582 {
1583         h->next = trans->hooks;
1584         trans->hooks = h;
1585 }
1586
1587 int __bch2_btree_insert(struct btree_trans *trans,
1588                         enum btree_id id, struct bkey_i *k)
1589 {
1590         struct btree_iter iter;
1591         int ret;
1592
1593         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1594                              BTREE_ITER_INTENT);
1595         ret   = bch2_btree_iter_traverse(&iter) ?:
1596                 bch2_trans_update(trans, &iter, k, 0);
1597         bch2_trans_iter_exit(trans, &iter);
1598         return ret;
1599 }
1600
1601 /**
1602  * bch2_btree_insert - insert keys into the extent btree
1603  * @c:                  pointer to struct bch_fs
1604  * @id:                 btree to insert into
1605  * @insert_keys:        list of keys to insert
1606  * @hook:               insert callback
1607  */
1608 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1609                       struct bkey_i *k,
1610                       struct disk_reservation *disk_res,
1611                       u64 *journal_seq, int flags)
1612 {
1613         return bch2_trans_do(c, disk_res, journal_seq, flags,
1614                              __bch2_btree_insert(&trans, id, k));
1615 }
1616
1617 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1618                                 unsigned len, unsigned update_flags)
1619 {
1620         struct bkey_i *k;
1621
1622         k = bch2_trans_kmalloc(trans, sizeof(*k));
1623         if (IS_ERR(k))
1624                 return PTR_ERR(k);
1625
1626         bkey_init(&k->k);
1627         k->k.p = iter->pos;
1628         bch2_key_resize(&k->k, len);
1629         return bch2_trans_update(trans, iter, k, update_flags);
1630 }
1631
1632 int bch2_btree_delete_at(struct btree_trans *trans,
1633                          struct btree_iter *iter, unsigned update_flags)
1634 {
1635         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1636 }
1637
1638 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1639                                   struct bpos start, struct bpos end,
1640                                   unsigned update_flags,
1641                                   u64 *journal_seq)
1642 {
1643         u32 restart_count = trans->restart_count;
1644         struct btree_iter iter;
1645         struct bkey_s_c k;
1646         int ret = 0;
1647
1648         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1649         while ((k = bch2_btree_iter_peek(&iter)).k) {
1650                 struct disk_reservation disk_res =
1651                         bch2_disk_reservation_init(trans->c, 0);
1652                 struct bkey_i delete;
1653
1654                 ret = bkey_err(k);
1655                 if (ret)
1656                         goto err;
1657
1658                 if (bkey_cmp(iter.pos, end) >= 0)
1659                         break;
1660
1661                 bkey_init(&delete.k);
1662
1663                 /*
1664                  * This could probably be more efficient for extents:
1665                  */
1666
1667                 /*
1668                  * For extents, iter.pos won't necessarily be the same as
1669                  * bkey_start_pos(k.k) (for non extents they always will be the
1670                  * same). It's important that we delete starting from iter.pos
1671                  * because the range we want to delete could start in the middle
1672                  * of k.
1673                  *
1674                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1675                  * bkey_start_pos(k.k)).
1676                  */
1677                 delete.k.p = iter.pos;
1678
1679                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1680                         unsigned max_sectors =
1681                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1682
1683                         /* create the biggest key we can */
1684                         bch2_key_resize(&delete.k, max_sectors);
1685                         bch2_cut_back(end, &delete);
1686
1687                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1688                         if (ret)
1689                                 goto err;
1690                 }
1691
1692                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1693                         bch2_trans_commit(trans, &disk_res, journal_seq,
1694                                           BTREE_INSERT_NOFAIL);
1695                 bch2_disk_reservation_put(trans->c, &disk_res);
1696 err:
1697                 /*
1698                  * the bch2_trans_begin() call is in a weird place because we
1699                  * need to call it after every transaction commit, to avoid path
1700                  * overflow, but don't want to call it if the delete operation
1701                  * is a no-op and we have no work to do:
1702                  */
1703                 bch2_trans_begin(trans);
1704
1705                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1706                         ret = 0;
1707                 if (ret)
1708                         break;
1709         }
1710         bch2_trans_iter_exit(trans, &iter);
1711
1712         if (!ret && trans_was_restarted(trans, restart_count))
1713                 ret = -BCH_ERR_transaction_restart_nested;
1714         return ret;
1715 }
1716
1717 /*
1718  * bch_btree_delete_range - delete everything within a given range
1719  *
1720  * Range is a half open interval - [start, end)
1721  */
1722 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1723                             struct bpos start, struct bpos end,
1724                             unsigned update_flags,
1725                             u64 *journal_seq)
1726 {
1727         int ret = bch2_trans_run(c,
1728                         bch2_btree_delete_range_trans(&trans, id, start, end,
1729                                                       update_flags, journal_seq));
1730         if (ret == -BCH_ERR_transaction_restart_nested)
1731                 ret = 0;
1732         return ret;
1733 }
1734
1735 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1736 {
1737         unsigned len = strlen(msg);
1738         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1739         struct jset_entry_log *l;
1740         int ret;
1741
1742         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1743         if (ret)
1744                 return ret;
1745
1746         l = (void *) &darray_top(trans->extra_journal_entries);
1747         l->entry.u64s           = cpu_to_le16(u64s);
1748         l->entry.btree_id       = 0;
1749         l->entry.level          = 1;
1750         l->entry.type           = BCH_JSET_ENTRY_log;
1751         l->entry.pad[0]         = 0;
1752         l->entry.pad[1]         = 0;
1753         l->entry.pad[2]         = 0;
1754         memcpy(l->d, msg, len);
1755         while (len & 7)
1756                 l->d[len++] = '\0';
1757
1758         trans->extra_journal_entries.nr += jset_u64s(u64s);
1759         return 0;
1760 }