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 "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "recovery.h"
  19 #include "subvolume.h"
  20 #include "replicas.h"
  21
  22 #include <linux/prefetch.h>
  23 #include <linux/sort.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static int __must_check
  27 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  28                           struct bkey_i *, enum btree_update_flags);
  29
  30 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  31                                          const struct btree_insert_entry *r)
  32 {
  33         return   cmp_int(l->btree_id,   r->btree_id) ?:
  34                  cmp_int(l->cached,     r->cached) ?:
  35                  -cmp_int(l->level,     r->level) ?:
  36                  bpos_cmp(l->k->k.p,    r->k->k.p);
  37 }
  38
  39 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  40 {
  41         return i->path->l + i->level;
  42 }
  43
  44 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  45                                      struct btree_insert_entry *i)
  46 {
  47         return i != trans->updates &&
  48                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  49 }
  50
  51 static inline bool same_leaf_as_next(struct btree_trans *trans,
  52                                      struct btree_insert_entry *i)
  53 {
  54         return i + 1 < trans->updates + trans->nr_updates &&
  55                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  56 }
  57
  58 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  59                                                   struct btree_path *path,
  60                                                   struct btree *b)
  61 {
  62         struct bch_fs *c = trans->c;
  63
  64         if (path->cached)
  65                 return;
  66
  67         if (unlikely(btree_node_just_written(b)) &&
  68             bch2_btree_post_write_cleanup(c, b))
  69                 bch2_trans_node_reinit_iter(trans, b);
  70
  71         /*
  72          * If the last bset has been written, or if it's gotten too big - start
  73          * a new bset to insert into:
  74          */
  75         if (want_new_bset(c, b))
  76                 bch2_btree_init_next(trans, b);
  77 }
  78
  79 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
  80                                      struct btree_path *path,
  81                                      struct btree *b)
  82 {
  83         bch2_btree_node_lock_write(trans, path, b);
  84         bch2_btree_node_prep_for_write(trans, path, b);
  85 }
  86
  87 /* Inserting into a given leaf node (last stage of insert): */
  88
  89 /* Handle overwrites and do insert, for non extents: */
  90 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  91                                 struct btree_path *path,
  92                                 struct btree *b,
  93                                 struct btree_node_iter *node_iter,
  94                                 struct bkey_i *insert)
  95 {
  96         struct bkey_packed *k;
  97         unsigned clobber_u64s = 0, new_u64s = 0;
  98
  99         EBUG_ON(btree_node_just_written(b));
 100         EBUG_ON(bset_written(b, btree_bset_last(b)));
 101         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 102         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
 103         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
 104         EBUG_ON(insert->k.u64s >
 105                 bch_btree_keys_u64s_remaining(trans->c, b));
 106
 107         k = bch2_btree_node_iter_peek_all(node_iter, b);
 108         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 109                 k = NULL;
 110
 111         /* @k is the key being overwritten/deleted, if any: */
 112         EBUG_ON(k && bkey_deleted(k));
 113
 114         /* Deleting, but not found? nothing to do: */
 115         if (bkey_deleted(&insert->k) && !k)
 116                 return false;
 117
 118         if (bkey_deleted(&insert->k)) {
 119                 /* Deleting: */
 120                 btree_account_key_drop(b, k);
 121                 k->type = KEY_TYPE_deleted;
 122
 123                 if (k->needs_whiteout)
 124                         push_whiteout(trans->c, b, insert->k.p);
 125                 k->needs_whiteout = false;
 126
 127                 if (k >= btree_bset_last(b)->start) {
 128                         clobber_u64s = k->u64s;
 129                         bch2_bset_delete(b, k, clobber_u64s);
 130                         goto fix_iter;
 131                 } else {
 132                         bch2_btree_path_fix_key_modified(trans, b, k);
 133                 }
 134
 135                 return true;
 136         }
 137
 138         if (k) {
 139                 /* Overwriting: */
 140                 btree_account_key_drop(b, k);
 141                 k->type = KEY_TYPE_deleted;
 142
 143                 insert->k.needs_whiteout = k->needs_whiteout;
 144                 k->needs_whiteout = false;
 145
 146                 if (k >= btree_bset_last(b)->start) {
 147                         clobber_u64s = k->u64s;
 148                         goto overwrite;
 149                 } else {
 150                         bch2_btree_path_fix_key_modified(trans, b, k);
 151                 }
 152         }
 153
 154         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 155 overwrite:
 156         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 157         new_u64s = k->u64s;
 158 fix_iter:
 159         if (clobber_u64s != new_u64s)
 160                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 161                                          clobber_u64s, new_u64s);
 162         return true;
 163 }
 164
 165 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 166                                unsigned i, u64 seq)
 167 {
 168         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 169         struct btree_write *w = container_of(pin, struct btree_write, journal);
 170         struct btree *b = container_of(w, struct btree, writes[i]);
 171         unsigned long old, new, v;
 172         unsigned idx = w - b->writes;
 173
 174         six_lock_read(&b->c.lock, NULL, NULL);
 175         v = READ_ONCE(b->flags);
 176
 177         do {
 178                 old = new = v;
 179
 180                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 181                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 182                     w->journal.seq != seq)
 183                         break;
 184
 185                 new |= 1 << BTREE_NODE_need_write;
 186         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 187
 188         btree_node_write_if_need(c, b, SIX_LOCK_read);
 189         six_unlock_read(&b->c.lock);
 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         if (!bch2_trans_relock(trans)) {
 286                 trace_trans_restart_journal_preres_get(trans->fn, trace_ip);
 287                 return -EINTR;
 288         }
 289
 290         return 0;
 291 }
 292
 293 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 294                                              unsigned flags)
 295 {
 296         struct bch_fs *c = trans->c;
 297         int ret;
 298
 299         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 300                                    trans->journal_u64s,
 301                                    flags|
 302                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 303
 304         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 305 }
 306
 307 #define JSET_ENTRY_LOG_U64s             4
 308
 309 static void journal_transaction_name(struct btree_trans *trans)
 310 {
 311         struct bch_fs *c = trans->c;
 312         struct journal *j = &c->journal;
 313         struct jset_entry *entry =
 314                 bch2_journal_add_entry(j, &trans->journal_res,
 315                                        BCH_JSET_ENTRY_log, 0, 0,
 316                                        JSET_ENTRY_LOG_U64s);
 317         struct jset_entry_log *l =
 318                 container_of(entry, struct jset_entry_log, entry);
 319
 320         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 321 }
 322
 323 static inline enum btree_insert_ret
 324 btree_key_can_insert(struct btree_trans *trans,
 325                      struct btree *b,
 326                      unsigned u64s)
 327 {
 328         struct bch_fs *c = trans->c;
 329
 330         if (!bch2_btree_node_insert_fits(c, b, u64s))
 331                 return BTREE_INSERT_BTREE_NODE_FULL;
 332
 333         return BTREE_INSERT_OK;
 334 }
 335
 336 static enum btree_insert_ret
 337 btree_key_can_insert_cached(struct btree_trans *trans,
 338                             struct btree_path *path,
 339                             unsigned u64s)
 340 {
 341         struct bch_fs *c = trans->c;
 342         struct bkey_cached *ck = (void *) path->l[0].b;
 343         unsigned old_u64s = ck->u64s, new_u64s;
 344         struct bkey_i *new_k;
 345
 346         EBUG_ON(path->level);
 347
 348         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 349             bch2_btree_key_cache_must_wait(c) &&
 350             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 351                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 352
 353         /*
 354          * bch2_varint_decode can read past the end of the buffer by at most 7
 355          * bytes (it won't be used):
 356          */
 357         u64s += 1;
 358
 359         if (u64s <= ck->u64s)
 360                 return BTREE_INSERT_OK;
 361
 362         new_u64s        = roundup_pow_of_two(u64s);
 363         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 364         if (!new_k) {
 365                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 366                         bch2_btree_ids[path->btree_id], new_u64s);
 367                 return -ENOMEM;
 368         }
 369
 370         ck->u64s        = new_u64s;
 371         ck->k           = new_k;
 372         /*
 373          * Keys returned by peek() are no longer valid pointers, so we need a
 374          * transaction restart:
 375          */
 376         trace_trans_restart_key_cache_key_realloced(trans->fn, _RET_IP_,
 377                                              path->btree_id, &path->pos,
 378                                              old_u64s, new_u64s);
 379         /*
 380          * Not using btree_trans_restart() because we can't unlock here, we have
 381          * write locks held:
 382          */
 383         trans->restarted = true;
 384         return -EINTR;
 385 }
 386
 387 /* Triggers: */
 388
 389 static int run_one_mem_trigger(struct btree_trans *trans,
 390                                struct btree_insert_entry *i,
 391                                unsigned flags)
 392 {
 393         struct bkey_s_c old = { &i->old_k, i->old_v };
 394         struct bkey_i *new = i->k;
 395         int ret;
 396
 397         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 398                 return 0;
 399
 400         if (!btree_node_type_needs_gc(i->btree_id))
 401                 return 0;
 402
 403         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 404             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 405             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 406                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 407                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 408         } else {
 409                 struct bkey             _deleted = KEY(0, 0, 0);
 410                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 411
 412                 _deleted.p = i->path->pos;
 413
 414                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 415                                 BTREE_TRIGGER_INSERT|flags) ?:
 416                         bch2_mark_key(trans, old, deleted,
 417                                 BTREE_TRIGGER_OVERWRITE|flags);
 418         }
 419
 420         return ret;
 421 }
 422
 423 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 424                                  bool overwrite)
 425 {
 426         /*
 427          * Transactional triggers create new btree_insert_entries, so we can't
 428          * pass them a pointer to a btree_insert_entry, that memory is going to
 429          * move:
 430          */
 431         struct bkey old_k = i->old_k;
 432         struct bkey_s_c old = { &old_k, i->old_v };
 433
 434         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 435             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 436                 return 0;
 437
 438         if (!i->insert_trigger_run &&
 439             !i->overwrite_trigger_run &&
 440             bch2_bkey_ops[old.k->type].trans_trigger ==
 441             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 442             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 443                 i->overwrite_trigger_run = true;
 444                 i->insert_trigger_run = true;
 445                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 446                                            BTREE_TRIGGER_INSERT|
 447                                            BTREE_TRIGGER_OVERWRITE|
 448                                            i->flags) ?: 1;
 449         } else if (overwrite && !i->overwrite_trigger_run) {
 450                 i->overwrite_trigger_run = true;
 451                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 452         } else if (!overwrite && !i->insert_trigger_run) {
 453                 i->insert_trigger_run = true;
 454                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 455         } else {
 456                 return 0;
 457         }
 458 }
 459
 460 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 461                               struct btree_insert_entry *btree_id_start)
 462 {
 463         struct btree_insert_entry *i;
 464         bool trans_trigger_run;
 465         int ret, overwrite;
 466
 467         for (overwrite = 1; overwrite >= 0; --overwrite) {
 468
 469                 /*
 470                  * Running triggers will append more updates to the list of updates as
 471                  * we're walking it:
 472                  */
 473                 do {
 474                         trans_trigger_run = false;
 475
 476                         for (i = btree_id_start;
 477                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 478                              i++) {
 479                                 if (i->btree_id != btree_id)
 480                                         continue;
 481
 482                                 ret = run_one_trans_trigger(trans, i, overwrite);
 483                                 if (ret < 0)
 484                                         return ret;
 485                                 if (ret)
 486                                         trans_trigger_run = true;
 487                         }
 488                 } while (trans_trigger_run);
 489         }
 490
 491         return 0;
 492 }
 493
 494 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 495 {
 496         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 497         unsigned btree_id = 0;
 498         int ret = 0;
 499
 500         /*
 501          *
 502          * For a given btree, this algorithm runs insert triggers before
 503          * overwrite triggers: this is so that when extents are being moved
 504          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 505          * they are re-added.
 506          */
 507         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 508                 if (btree_id == BTREE_ID_alloc)
 509                         continue;
 510
 511                 while (btree_id_start < trans->updates + trans->nr_updates &&
 512                        btree_id_start->btree_id < btree_id)
 513                         btree_id_start++;
 514
 515                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 516                 if (ret)
 517                         return ret;
 518         }
 519
 520         trans_for_each_update(trans, i) {
 521                 if (i->btree_id > BTREE_ID_alloc)
 522                         break;
 523                 if (i->btree_id == BTREE_ID_alloc) {
 524                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 525                         if (ret)
 526                                 return ret;
 527                         break;
 528                 }
 529         }
 530
 531         trans_for_each_update(trans, i)
 532                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 533                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 534                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 535
 536         return 0;
 537 }
 538
 539 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 540 {
 541         struct bch_fs *c = trans->c;
 542         struct btree_insert_entry *i;
 543         int ret = 0;
 544
 545         trans_for_each_update(trans, i) {
 546                 /*
 547                  * XXX: synchronization of cached update triggers with gc
 548                  * XXX: synchronization of interior node updates with gc
 549                  */
 550                 BUG_ON(i->cached || i->level);
 551
 552                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 553                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 554                         if (ret)
 555                                 break;
 556                 }
 557         }
 558
 559         return ret;
 560 }
 561
 562 static inline int
 563 bch2_trans_commit_write_locked(struct btree_trans *trans,
 564                                struct btree_insert_entry **stopped_at,
 565                                unsigned long trace_ip)
 566 {
 567         struct bch_fs *c = trans->c;
 568         struct btree_insert_entry *i;
 569         struct btree_trans_commit_hook *h;
 570         unsigned u64s = 0;
 571         bool marking = false;
 572         int ret;
 573
 574         if (race_fault()) {
 575                 trace_trans_restart_fault_inject(trans->fn, trace_ip);
 576                 trans->restarted = true;
 577                 return -EINTR;
 578         }
 579
 580         /*
 581          * Check if the insert will fit in the leaf node with the write lock
 582          * held, otherwise another thread could write the node changing the
 583          * amount of space available:
 584          */
 585
 586         prefetch(&trans->c->journal.flags);
 587
 588         h = trans->hooks;
 589         while (h) {
 590                 ret = h->fn(trans, h);
 591                 if (ret)
 592                         return ret;
 593                 h = h->next;
 594         }
 595
 596         trans_for_each_update(trans, i) {
 597                 /* Multiple inserts might go to same leaf: */
 598                 if (!same_leaf_as_prev(trans, i))
 599                         u64s = 0;
 600
 601                 u64s += i->k->k.u64s;
 602                 ret = !i->cached
 603                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 604                         : btree_key_can_insert_cached(trans, i->path, u64s);
 605                 if (ret) {
 606                         *stopped_at = i;
 607                         return ret;
 608                 }
 609
 610                 if (btree_node_type_needs_gc(i->bkey_type))
 611                         marking = true;
 612
 613                 /*
 614                  * Revalidate before calling mem triggers - XXX, ugly:
 615                  *
 616                  * - successful btree node splits don't cause transaction
 617                  *   restarts and will have invalidated the pointer to the bkey
 618                  *   value
 619                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 620                  *   when it has to resort
 621                  * - btree_key_can_insert_cached() when it has to reallocate
 622                  *
 623                  *   Ugly because we currently have no way to tell if the
 624                  *   pointer's been invalidated, which means it's debatabale
 625                  *   whether we should be stashing the old key at all.
 626                  */
 627                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 628
 629                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 630                         struct bkey_i *j_k =
 631                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 632                                                             i->k->k.p);
 633
 634                         if (j_k) {
 635                                 i->old_k = j_k->k;
 636                                 i->old_v = &j_k->v;
 637                         }
 638                 }
 639         }
 640
 641         /*
 642          * Don't get journal reservation until after we know insert will
 643          * succeed:
 644          */
 645         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 646                 ret = bch2_trans_journal_res_get(trans,
 647                                 JOURNAL_RES_GET_NONBLOCK);
 648                 if (ret)
 649                         return ret;
 650
 651                 journal_transaction_name(trans);
 652         } else {
 653                 trans->journal_res.seq = c->journal.replay_journal_seq;
 654         }
 655
 656         if (unlikely(trans->extra_journal_entries.nr)) {
 657                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 658                                   trans->extra_journal_entries.data,
 659                                   trans->extra_journal_entries.nr);
 660
 661                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 662                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 663         }
 664
 665         /*
 666          * Not allowed to fail after we've gotten our journal reservation - we
 667          * have to use it:
 668          */
 669
 670         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 671                 if (bch2_journal_seq_verify)
 672                         trans_for_each_update(trans, i)
 673                                 i->k->k.version.lo = trans->journal_res.seq;
 674                 else if (bch2_inject_invalid_keys)
 675                         trans_for_each_update(trans, i)
 676                                 i->k->k.version = MAX_VERSION;
 677         }
 678
 679         if (trans->fs_usage_deltas &&
 680             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 681                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 682
 683         trans_for_each_update(trans, i)
 684                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 685                         ret = run_one_mem_trigger(trans, i, i->flags);
 686                         if (ret)
 687                                 return ret;
 688                 }
 689
 690         if (unlikely(c->gc_pos.phase)) {
 691                 ret = bch2_trans_commit_run_gc_triggers(trans);
 692                 if  (ret)
 693                         return ret;
 694         }
 695
 696         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 697                 trans_for_each_update(trans, i) {
 698                         struct journal *j = &c->journal;
 699                         struct jset_entry *entry;
 700
 701                         if (i->key_cache_already_flushed)
 702                                 continue;
 703
 704                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 705                                                BCH_JSET_ENTRY_overwrite,
 706                                                i->btree_id, i->level,
 707                                                i->old_k.u64s);
 708                         bkey_reassemble(&entry->start[0],
 709                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 710
 711                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 712                                                BCH_JSET_ENTRY_btree_keys,
 713                                                i->btree_id, i->level,
 714                                                i->k->k.u64s);
 715                         bkey_copy(&entry->start[0], i->k);
 716                 }
 717
 718                 if (trans->journal_seq)
 719                         *trans->journal_seq = trans->journal_res.seq;
 720         }
 721
 722         trans_for_each_update(trans, i) {
 723                 i->k->k.needs_whiteout = false;
 724
 725                 if (!i->cached)
 726                         btree_insert_key_leaf(trans, i);
 727                 else if (!i->key_cache_already_flushed)
 728                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 729                 else
 730                         bch2_btree_key_cache_drop(trans, i->path);
 731         }
 732
 733         return ret;
 734 }
 735
 736 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 737 {
 738         unsigned l;
 739
 740         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 741                 if (btree_node_read_locked(path, l))
 742                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 743 }
 744
 745 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 746 {
 747         struct btree *b = path_l(path)->b;
 748
 749         do {
 750                 if (path->nodes_locked &&
 751                     path->nodes_locked != path->nodes_intent_locked)
 752                         path_upgrade_readers(trans, path);
 753         } while ((path = prev_btree_path(trans, path)) &&
 754                  path_l(path)->b == b);
 755 }
 756
 757 /*
 758  * Check for nodes that we have both read and intent locks on, and upgrade the
 759  * readers to intent:
 760  */
 761 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 762 {
 763         struct btree_path *path;
 764         unsigned i, nr_read = 0, nr_intent = 0;
 765
 766         trans_for_each_path_inorder(trans, path, i) {
 767                 struct btree_path *next = i + 1 < trans->nr_sorted
 768                         ? trans->paths + trans->sorted[i + 1]
 769                         : NULL;
 770
 771                 if (path->nodes_locked) {
 772                         if (path->nodes_intent_locked)
 773                                 nr_intent++;
 774                         else
 775                                 nr_read++;
 776                 }
 777
 778                 if (!next || path_l(path)->b != path_l(next)->b) {
 779                         if (nr_read && nr_intent)
 780                                 upgrade_readers(trans, path);
 781
 782                         nr_read = nr_intent = 0;
 783                 }
 784         }
 785
 786         bch2_trans_verify_locks(trans);
 787 }
 788
 789 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 790 {
 791         struct btree_path *path;
 792         unsigned i;
 793
 794         trans_for_each_path_inorder(trans, path, i) {
 795                 //if (path == pos)
 796                 //      break;
 797
 798                 if (path->nodes_locked != path->nodes_intent_locked &&
 799                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 800                         return true;
 801         }
 802
 803         return false;
 804 }
 805
 806 static inline int trans_lock_write(struct btree_trans *trans)
 807 {
 808         struct btree_insert_entry *i;
 809
 810         trans_for_each_update(trans, i) {
 811                 if (same_leaf_as_prev(trans, i))
 812                         continue;
 813
 814                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 815                         if (have_conflicting_read_lock(trans, i->path))
 816                                 goto fail;
 817
 818                         btree_node_lock_type(trans, i->path,
 819                                              insert_l(i)->b,
 820                                              i->path->pos, i->level,
 821                                              SIX_LOCK_write, NULL, NULL);
 822                 }
 823
 824                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 825         }
 826
 827         return 0;
 828 fail:
 829         while (--i >= trans->updates) {
 830                 if (same_leaf_as_prev(trans, i))
 831                         continue;
 832
 833                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 834         }
 835
 836         trace_trans_restart_would_deadlock_write(trans->fn);
 837         return btree_trans_restart(trans);
 838 }
 839
 840 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 841 {
 842         struct btree_insert_entry *i;
 843
 844         trans_for_each_update(trans, i)
 845                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 846 }
 847
 848 /*
 849  * Get journal reservation, take write locks, and attempt to do btree update(s):
 850  */
 851 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 852                                        struct btree_insert_entry **stopped_at,
 853                                        unsigned long trace_ip)
 854 {
 855         struct bch_fs *c = trans->c;
 856         struct btree_insert_entry *i;
 857         struct printbuf buf = PRINTBUF;
 858         int ret, u64s_delta = 0;
 859         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 860
 861         trans_for_each_update(trans, i) {
 862                 if (bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 863                                       i->bkey_type, rw, &buf)) {
 864                         printbuf_reset(&buf);
 865                         prt_printf(&buf, "invalid bkey on insert from %s -> %ps",
 866                                trans->fn, (void *) i->ip_allocated);
 867                         prt_newline(&buf);
 868                         printbuf_indent_add(&buf, 2);
 869
 870                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
 871                         prt_newline(&buf);
 872
 873                         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 874                                           i->bkey_type, rw, &buf);
 875
 876                         bch2_trans_inconsistent(trans, "%s", buf.buf);
 877                         printbuf_exit(&buf);
 878                         return -EINVAL;
 879                 }
 880                 btree_insert_entry_checks(trans, i);
 881         }
 882
 883         printbuf_exit(&buf);
 884
 885         trans_for_each_update(trans, i) {
 886                 if (i->cached)
 887                         continue;
 888
 889                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 890                 u64s_delta -= i->old_btree_u64s;
 891
 892                 if (!same_leaf_as_next(trans, i)) {
 893                         if (u64s_delta <= 0) {
 894                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 895                                                         i->level, trans->flags);
 896                                 if (unlikely(ret))
 897                                         return ret;
 898                         }
 899
 900                         u64s_delta = 0;
 901                 }
 902         }
 903
 904         ret = bch2_journal_preres_get(&c->journal,
 905                         &trans->journal_preres, trans->journal_preres_u64s,
 906                         JOURNAL_RES_GET_NONBLOCK|
 907                         (trans->flags & JOURNAL_WATERMARK_MASK));
 908         if (unlikely(ret == -EAGAIN))
 909                 ret = bch2_trans_journal_preres_get_cold(trans,
 910                                                 trans->journal_preres_u64s, trace_ip);
 911         if (unlikely(ret))
 912                 return ret;
 913
 914         normalize_read_intent_locks(trans);
 915
 916         ret = trans_lock_write(trans);
 917         if (unlikely(ret))
 918                 return ret;
 919
 920         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 921
 922         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 923                 bch2_drop_overwrites_from_journal(trans);
 924
 925         trans_for_each_update(trans, i)
 926                 if (!same_leaf_as_prev(trans, i))
 927                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 928                                                         insert_l(i)->b);
 929
 930         if (!ret && trans->journal_pin)
 931                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 932                                      trans->journal_pin, NULL);
 933
 934         /*
 935          * Drop journal reservation after dropping write locks, since dropping
 936          * the journal reservation may kick off a journal write:
 937          */
 938         bch2_journal_res_put(&c->journal, &trans->journal_res);
 939
 940         if (unlikely(ret))
 941                 return ret;
 942
 943         bch2_trans_downgrade(trans);
 944
 945         return 0;
 946 }
 947
 948 static int journal_reclaim_wait_done(struct bch_fs *c)
 949 {
 950         int ret = bch2_journal_error(&c->journal) ?:
 951                 !bch2_btree_key_cache_must_wait(c);
 952
 953         if (!ret)
 954                 journal_reclaim_kick(&c->journal);
 955         return ret;
 956 }
 957
 958 static noinline
 959 int bch2_trans_commit_error(struct btree_trans *trans,
 960                             struct btree_insert_entry *i,
 961                             int ret, unsigned long trace_ip)
 962 {
 963         struct bch_fs *c = trans->c;
 964
 965         switch (ret) {
 966         case BTREE_INSERT_BTREE_NODE_FULL:
 967                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 968                 if (!ret)
 969                         return 0;
 970
 971                 if (ret == -EINTR)
 972                         trace_trans_restart_btree_node_split(trans->fn, trace_ip,
 973                                                 i->btree_id, &i->path->pos);
 974                 break;
 975         case BTREE_INSERT_NEED_MARK_REPLICAS:
 976                 bch2_trans_unlock(trans);
 977
 978                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 979                 if (ret)
 980                         break;
 981
 982                 if (bch2_trans_relock(trans))
 983                         return 0;
 984
 985                 trace_trans_restart_mark_replicas(trans->fn, trace_ip);
 986                 ret = -EINTR;
 987                 break;
 988         case BTREE_INSERT_NEED_JOURNAL_RES:
 989                 bch2_trans_unlock(trans);
 990
 991                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 992                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 993                         trans->restarted = true;
 994                         ret = -EAGAIN;
 995                         break;
 996                 }
 997
 998                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 999                 if (ret)
1000                         break;
1001
1002                 if (bch2_trans_relock(trans))
1003                         return 0;
1004
1005                 trace_trans_restart_journal_res_get(trans->fn, trace_ip);
1006                 ret = -EINTR;
1007                 break;
1008         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
1009                 bch2_trans_unlock(trans);
1010
1011                 trace_trans_blocked_journal_reclaim(trans->fn, trace_ip);
1012
1013                 wait_event_freezable(c->journal.reclaim_wait,
1014                                      (ret = journal_reclaim_wait_done(c)));
1015                 if (ret < 0)
1016                         break;
1017
1018                 if (bch2_trans_relock(trans))
1019                         return 0;
1020
1021                 trace_trans_restart_journal_reclaim(trans->fn, trace_ip);
1022                 ret = -EINTR;
1023                 break;
1024         default:
1025                 BUG_ON(ret >= 0);
1026                 break;
1027         }
1028
1029         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
1030         BUG_ON(ret == -ENOSPC &&
1031                !(trans->flags & BTREE_INSERT_NOWAIT) &&
1032                (trans->flags & BTREE_INSERT_NOFAIL));
1033
1034         return ret;
1035 }
1036
1037 static noinline int
1038 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
1039 {
1040         struct bch_fs *c = trans->c;
1041         int ret;
1042
1043         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
1044             test_bit(BCH_FS_STARTED, &c->flags))
1045                 return -EROFS;
1046
1047         bch2_trans_unlock(trans);
1048
1049         ret = bch2_fs_read_write_early(c);
1050         if (ret)
1051                 return ret;
1052
1053         if (!bch2_trans_relock(trans))
1054                 return -EINTR;
1055
1056         percpu_ref_get(&c->writes);
1057         return 0;
1058 }
1059
1060 /*
1061  * This is for updates done in the early part of fsck - btree_gc - before we've
1062  * gone RW. we only add the new key to the list of keys for journal replay to
1063  * do.
1064  */
1065 static noinline int
1066 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1067 {
1068         struct bch_fs *c = trans->c;
1069         struct btree_insert_entry *i;
1070         int ret = 0;
1071
1072         trans_for_each_update(trans, i) {
1073                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1074                 if (ret)
1075                         break;
1076         }
1077
1078         return ret;
1079 }
1080
1081 int __bch2_trans_commit(struct btree_trans *trans)
1082 {
1083         struct bch_fs *c = trans->c;
1084         struct btree_insert_entry *i = NULL;
1085         unsigned u64s;
1086         int ret = 0;
1087
1088         if (!trans->nr_updates &&
1089             !trans->extra_journal_entries.nr)
1090                 goto out_reset;
1091
1092         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1093                 lockdep_assert_held(&c->gc_lock);
1094
1095         ret = bch2_trans_commit_run_triggers(trans);
1096         if (ret)
1097                 goto out_reset;
1098
1099         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1100                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1101                 goto out_reset;
1102         }
1103
1104         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1105             unlikely(!percpu_ref_tryget(&c->writes))) {
1106                 ret = bch2_trans_commit_get_rw_cold(trans);
1107                 if (ret)
1108                         goto out_reset;
1109         }
1110
1111         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1112
1113         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1114
1115         trans->journal_u64s             = trans->extra_journal_entries.nr;
1116         trans->journal_preres_u64s      = 0;
1117
1118         /* For journalling transaction name: */
1119         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1120
1121         trans_for_each_update(trans, i) {
1122                 BUG_ON(!i->path->should_be_locked);
1123
1124                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1125                         trace_trans_restart_upgrade(trans->fn, _RET_IP_,
1126                                                     i->btree_id, &i->path->pos);
1127                         ret = btree_trans_restart(trans);
1128                         goto out;
1129                 }
1130
1131                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1132
1133                 if (i->key_cache_already_flushed)
1134                         continue;
1135
1136                 /* we're going to journal the key being updated: */
1137                 u64s = jset_u64s(i->k->k.u64s);
1138                 if (i->cached &&
1139                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1140                         trans->journal_preres_u64s += u64s;
1141                 trans->journal_u64s += u64s;
1142
1143                 /* and we're also going to log the overwrite: */
1144                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1145         }
1146
1147         if (trans->extra_journal_res) {
1148                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1149                                 trans->extra_journal_res,
1150                                 (trans->flags & BTREE_INSERT_NOFAIL)
1151                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1152                 if (ret)
1153                         goto err;
1154         }
1155 retry:
1156         BUG_ON(trans->restarted);
1157         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1158
1159         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1160
1161         /* make sure we didn't drop or screw up locks: */
1162         bch2_trans_verify_locks(trans);
1163
1164         if (ret)
1165                 goto err;
1166
1167         trace_transaction_commit(trans->fn, _RET_IP_);
1168 out:
1169         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1170
1171         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1172                 percpu_ref_put(&c->writes);
1173 out_reset:
1174         bch2_trans_reset_updates(trans);
1175
1176         if (trans->fs_usage_deltas) {
1177                 trans->fs_usage_deltas->used = 0;
1178                 memset((void *) trans->fs_usage_deltas +
1179                        offsetof(struct replicas_delta_list, memset_start), 0,
1180                        (void *) &trans->fs_usage_deltas->memset_end -
1181                        (void *) &trans->fs_usage_deltas->memset_start);
1182         }
1183
1184         return ret;
1185 err:
1186         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1187         if (ret)
1188                 goto out;
1189
1190         goto retry;
1191 }
1192
1193 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1194                                           enum btree_id id,
1195                                           struct bpos pos)
1196 {
1197         struct bch_fs *c = trans->c;
1198         struct btree_iter iter;
1199         struct bkey_s_c k;
1200         int ret;
1201
1202         if (!btree_type_has_snapshots(id))
1203                 return 0;
1204
1205         if (!snapshot_t(c, pos.snapshot)->children[0])
1206                 return 0;
1207
1208         bch2_trans_iter_init(trans, &iter, id, pos,
1209                              BTREE_ITER_NOT_EXTENTS|
1210                              BTREE_ITER_ALL_SNAPSHOTS);
1211         while (1) {
1212                 k = bch2_btree_iter_prev(&iter);
1213                 ret = bkey_err(k);
1214                 if (ret)
1215                         break;
1216
1217                 if (!k.k)
1218                         break;
1219
1220                 if (bkey_cmp(pos, k.k->p))
1221                         break;
1222
1223                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1224                         ret = 1;
1225                         break;
1226                 }
1227         }
1228         bch2_trans_iter_exit(trans, &iter);
1229
1230         return ret;
1231 }
1232
1233 int bch2_trans_update_extent(struct btree_trans *trans,
1234                              struct btree_iter *orig_iter,
1235                              struct bkey_i *insert,
1236                              enum btree_update_flags flags)
1237 {
1238         struct bch_fs *c = trans->c;
1239         struct btree_iter iter, update_iter;
1240         struct bpos start = bkey_start_pos(&insert->k);
1241         struct bkey_i *update;
1242         struct bkey_s_c k;
1243         enum btree_id btree_id = orig_iter->btree_id;
1244         int ret = 0, compressed_sectors;
1245
1246         bch2_trans_iter_init(trans, &iter, btree_id, start,
1247                              BTREE_ITER_INTENT|
1248                              BTREE_ITER_WITH_UPDATES|
1249                              BTREE_ITER_NOT_EXTENTS);
1250         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1251         if ((ret = bkey_err(k)))
1252                 goto err;
1253         if (!k.k)
1254                 goto out;
1255
1256         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1257                 /*
1258                  * We can't merge extents if they belong to interior snapshot
1259                  * tree nodes, and there's a snapshot in which one extent is
1260                  * visible and the other is not - i.e. if visibility is
1261                  * different.
1262                  *
1263                  * Instead of checking if visibilitiy of the two extents is
1264                  * different, for now we just check if either has been
1265                  * overwritten:
1266                  */
1267                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1268                 if (ret < 0)
1269                         goto err;
1270                 if (ret)
1271                         goto nomerge1;
1272
1273                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1274                 if (ret < 0)
1275                         goto err;
1276                 if (ret)
1277                         goto nomerge1;
1278
1279                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1280                 if ((ret = PTR_ERR_OR_ZERO(update)))
1281                         goto err;
1282
1283                 bkey_reassemble(update, k);
1284
1285                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1286                         ret = bch2_btree_delete_at(trans, &iter, flags);
1287                         if (ret)
1288                                 goto err;
1289
1290                         insert = update;
1291                         goto next;
1292                 }
1293         }
1294 nomerge1:
1295         ret = 0;
1296         if (!bkey_cmp(k.k->p, start))
1297                 goto next;
1298
1299         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1300                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1301                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1302
1303                 /*
1304                  * If we're going to be splitting a compressed extent, note it
1305                  * so that __bch2_trans_commit() can increase our disk
1306                  * reservation:
1307                  */
1308                 if (((front_split && back_split) ||
1309                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1310                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1311                         trans->extra_journal_res += compressed_sectors;
1312
1313                 if (front_split) {
1314                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1315                         if ((ret = PTR_ERR_OR_ZERO(update)))
1316                                 goto err;
1317
1318                         bkey_reassemble(update, k);
1319
1320                         bch2_cut_back(start, update);
1321
1322                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1323                                              BTREE_ITER_NOT_EXTENTS|
1324                                              BTREE_ITER_ALL_SNAPSHOTS|
1325                                              BTREE_ITER_INTENT);
1326                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1327                                 bch2_trans_update(trans, &update_iter, update,
1328                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1329                                                   flags);
1330                         bch2_trans_iter_exit(trans, &update_iter);
1331
1332                         if (ret)
1333                                 goto err;
1334                 }
1335
1336                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1337                     (front_split || back_split)) {
1338                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1339                         if ((ret = PTR_ERR_OR_ZERO(update)))
1340                                 goto err;
1341
1342                         bkey_reassemble(update, k);
1343
1344                         bch2_cut_front(start, update);
1345                         bch2_cut_back(insert->k.p, update);
1346
1347                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1348                                              BTREE_ITER_NOT_EXTENTS|
1349                                              BTREE_ITER_ALL_SNAPSHOTS|
1350                                              BTREE_ITER_INTENT);
1351                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1352                                 bch2_trans_update(trans, &update_iter, update,
1353                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1354                                                   flags);
1355                         bch2_trans_iter_exit(trans, &update_iter);
1356                         if (ret)
1357                                 goto err;
1358                 }
1359
1360                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1361                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1362                         if ((ret = PTR_ERR_OR_ZERO(update)))
1363                                 goto err;
1364
1365                         bkey_init(&update->k);
1366                         update->k.p = k.k->p;
1367
1368                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1369                                 update->k.p.snapshot = insert->k.p.snapshot;
1370                                 update->k.type = KEY_TYPE_whiteout;
1371                         }
1372
1373                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1374                                              BTREE_ITER_NOT_EXTENTS|
1375                                              BTREE_ITER_INTENT);
1376                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1377                                 bch2_trans_update(trans, &update_iter, update,
1378                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1379                                                   flags);
1380                         bch2_trans_iter_exit(trans, &update_iter);
1381
1382                         if (ret)
1383                                 goto err;
1384                 }
1385
1386                 if (back_split) {
1387                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1388                         if ((ret = PTR_ERR_OR_ZERO(update)))
1389                                 goto err;
1390
1391                         bkey_reassemble(update, k);
1392                         bch2_cut_front(insert->k.p, update);
1393
1394                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1395                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1396                                                   flags);
1397                         if (ret)
1398                                 goto err;
1399                         goto out;
1400                 }
1401 next:
1402                 bch2_btree_iter_advance(&iter);
1403                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1404                 if ((ret = bkey_err(k)))
1405                         goto err;
1406                 if (!k.k)
1407                         goto out;
1408         }
1409
1410         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1411                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1412                 if (ret < 0)
1413                         goto err;
1414                 if (ret)
1415                         goto nomerge2;
1416
1417                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1418                 if (ret < 0)
1419                         goto err;
1420                 if (ret)
1421                         goto nomerge2;
1422
1423                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1424         }
1425 nomerge2:
1426         ret = 0;
1427 out:
1428         if (!bkey_deleted(&insert->k)) {
1429                 /*
1430                  * Rewinding iterators is expensive: get a new one and the one
1431                  * that points to the start of insert will be cloned from:
1432                  */
1433                 bch2_trans_iter_exit(trans, &iter);
1434                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1435                                      BTREE_ITER_NOT_EXTENTS|
1436                                      BTREE_ITER_INTENT);
1437                 ret   = bch2_btree_iter_traverse(&iter) ?:
1438                         bch2_trans_update(trans, &iter, insert, flags);
1439         }
1440 err:
1441         bch2_trans_iter_exit(trans, &iter);
1442
1443         return ret;
1444 }
1445
1446 /*
1447  * When deleting, check if we need to emit a whiteout (because we're overwriting
1448  * something in an ancestor snapshot)
1449  */
1450 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1451                                       enum btree_id btree_id, struct bpos pos)
1452 {
1453         struct btree_iter iter;
1454         struct bkey_s_c k;
1455         u32 snapshot = pos.snapshot;
1456         int ret;
1457
1458         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1459                 return 0;
1460
1461         pos.snapshot++;
1462
1463         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1464                            BTREE_ITER_ALL_SNAPSHOTS|
1465                            BTREE_ITER_NOPRESERVE, k, ret) {
1466                 if (bkey_cmp(k.k->p, pos))
1467                         break;
1468
1469                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1470                                               k.k->p.snapshot)) {
1471                         ret = !bkey_whiteout(k.k);
1472                         break;
1473                 }
1474         }
1475         bch2_trans_iter_exit(trans, &iter);
1476
1477         return ret;
1478 }
1479
1480 static int __must_check
1481 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1482                                 struct bkey_i *k, enum btree_update_flags flags,
1483                                 unsigned long ip)
1484 {
1485         struct bch_fs *c = trans->c;
1486         struct btree_insert_entry *i, n;
1487         int ret = 0;
1488
1489         BUG_ON(!path->should_be_locked);
1490
1491         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1492         BUG_ON(bpos_cmp(k->k.p, path->pos));
1493
1494         n = (struct btree_insert_entry) {
1495                 .flags          = flags,
1496                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1497                 .btree_id       = path->btree_id,
1498                 .level          = path->level,
1499                 .cached         = path->cached,
1500                 .path           = path,
1501                 .k              = k,
1502                 .ip_allocated   = ip,
1503         };
1504
1505 #ifdef CONFIG_BCACHEFS_DEBUG
1506         trans_for_each_update(trans, i)
1507                 BUG_ON(i != trans->updates &&
1508                        btree_insert_entry_cmp(i - 1, i) >= 0);
1509 #endif
1510
1511         /*
1512          * Pending updates are kept sorted: first, find position of new update,
1513          * then delete/trim any updates the new update overwrites:
1514          */
1515         trans_for_each_update(trans, i)
1516                 if (btree_insert_entry_cmp(&n, i) <= 0)
1517                         break;
1518
1519         if (i < trans->updates + trans->nr_updates &&
1520             !btree_insert_entry_cmp(&n, i)) {
1521                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1522
1523                 bch2_path_put(trans, i->path, true);
1524                 i->flags        = n.flags;
1525                 i->cached       = n.cached;
1526                 i->k            = n.k;
1527                 i->path         = n.path;
1528                 i->ip_allocated = n.ip_allocated;
1529         } else {
1530                 array_insert_item(trans->updates, trans->nr_updates,
1531                                   i - trans->updates, n);
1532
1533                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1534                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1535
1536                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1537                         struct bkey_i *j_k =
1538                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1539
1540                         if (j_k) {
1541                                 i->old_k = j_k->k;
1542                                 i->old_v = &j_k->v;
1543                         }
1544                 }
1545         }
1546
1547         __btree_path_get(i->path, true);
1548
1549         /*
1550          * If a key is present in the key cache, it must also exist in the
1551          * btree - this is necessary for cache coherency. When iterating over
1552          * a btree that's cached in the key cache, the btree iter code checks
1553          * the key cache - but the key has to exist in the btree for that to
1554          * work:
1555          */
1556         if (path->cached &&
1557             bkey_deleted(&i->old_k)) {
1558                 struct btree_path *btree_path;
1559
1560                 i->key_cache_already_flushed = true;
1561                 i->flags |= BTREE_TRIGGER_NORUN;
1562
1563                 btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1564                                            BTREE_ITER_INTENT, _THIS_IP_);
1565
1566                 ret = bch2_btree_path_traverse(trans, btree_path, 0);
1567                 if (ret)
1568                         goto err;
1569
1570                 btree_path->should_be_locked = true;
1571                 ret = bch2_trans_update_by_path_trace(trans, btree_path, k, flags, ip);
1572 err:
1573                 bch2_path_put(trans, btree_path, true);
1574         }
1575
1576         return ret;
1577 }
1578
1579 static int __must_check
1580 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1581                           struct bkey_i *k, enum btree_update_flags flags)
1582 {
1583         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1584 }
1585
1586 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1587                                    struct bkey_i *k, enum btree_update_flags flags)
1588 {
1589         struct btree_path *path = iter->update_path ?: iter->path;
1590         struct bkey_cached *ck;
1591         int ret;
1592
1593         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1594                 return bch2_trans_update_extent(trans, iter, k, flags);
1595
1596         if (bkey_deleted(&k->k) &&
1597             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1598             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1599                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1600                 if (unlikely(ret < 0))
1601                         return ret;
1602
1603                 if (ret)
1604                         k->k.type = KEY_TYPE_whiteout;
1605         }
1606
1607         /*
1608          * Ensure that updates to cached btrees go to the key cache:
1609          */
1610         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1611             !path->cached &&
1612             !path->level &&
1613             btree_id_cached(trans->c, path->btree_id)) {
1614                 if (!iter->key_cache_path ||
1615                     !iter->key_cache_path->should_be_locked ||
1616                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1617                         if (!iter->key_cache_path)
1618                                 iter->key_cache_path =
1619                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1620                                                       BTREE_ITER_INTENT|
1621                                                       BTREE_ITER_CACHED, _THIS_IP_);
1622
1623                         iter->key_cache_path =
1624                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1625                                                         iter->flags & BTREE_ITER_INTENT,
1626                                                         _THIS_IP_);
1627
1628                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1629                                                        BTREE_ITER_CACHED);
1630                         if (unlikely(ret))
1631                                 return ret;
1632
1633                         ck = (void *) iter->key_cache_path->l[0].b;
1634
1635                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1636                                 trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
1637                                 btree_trans_restart(trans);
1638                                 return -EINTR;
1639                         }
1640
1641                         iter->key_cache_path->should_be_locked = true;
1642                 }
1643
1644                 path = iter->key_cache_path;
1645         }
1646
1647         return bch2_trans_update_by_path(trans, path, k, flags);
1648 }
1649
1650 void bch2_trans_commit_hook(struct btree_trans *trans,
1651                             struct btree_trans_commit_hook *h)
1652 {
1653         h->next = trans->hooks;
1654         trans->hooks = h;
1655 }
1656
1657 int __bch2_btree_insert(struct btree_trans *trans,
1658                         enum btree_id id, struct bkey_i *k)
1659 {
1660         struct btree_iter iter;
1661         int ret;
1662
1663         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1664                              BTREE_ITER_INTENT);
1665         ret   = bch2_btree_iter_traverse(&iter) ?:
1666                 bch2_trans_update(trans, &iter, k, 0);
1667         bch2_trans_iter_exit(trans, &iter);
1668         return ret;
1669 }
1670
1671 /**
1672  * bch2_btree_insert - insert keys into the extent btree
1673  * @c:                  pointer to struct bch_fs
1674  * @id:                 btree to insert into
1675  * @insert_keys:        list of keys to insert
1676  * @hook:               insert callback
1677  */
1678 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1679                       struct bkey_i *k,
1680                       struct disk_reservation *disk_res,
1681                       u64 *journal_seq, int flags)
1682 {
1683         return bch2_trans_do(c, disk_res, journal_seq, flags,
1684                              __bch2_btree_insert(&trans, id, k));
1685 }
1686
1687 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1688                                 unsigned len, unsigned update_flags)
1689 {
1690         struct bkey_i *k;
1691
1692         k = bch2_trans_kmalloc(trans, sizeof(*k));
1693         if (IS_ERR(k))
1694                 return PTR_ERR(k);
1695
1696         bkey_init(&k->k);
1697         k->k.p = iter->pos;
1698         bch2_key_resize(&k->k, len);
1699         return bch2_trans_update(trans, iter, k, update_flags);
1700 }
1701
1702 int bch2_btree_delete_at(struct btree_trans *trans,
1703                          struct btree_iter *iter, unsigned update_flags)
1704 {
1705         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1706 }
1707
1708 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1709                                   struct bpos start, struct bpos end,
1710                                   unsigned update_flags,
1711                                   u64 *journal_seq)
1712 {
1713         struct btree_iter iter;
1714         struct bkey_s_c k;
1715         int ret = 0;
1716
1717         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1718 retry:
1719         while ((bch2_trans_begin(trans),
1720                (k = bch2_btree_iter_peek(&iter)).k) &&
1721                !(ret = bkey_err(k)) &&
1722                bkey_cmp(iter.pos, end) < 0) {
1723                 struct disk_reservation disk_res =
1724                         bch2_disk_reservation_init(trans->c, 0);
1725                 struct bkey_i delete;
1726
1727                 bkey_init(&delete.k);
1728
1729                 /*
1730                  * This could probably be more efficient for extents:
1731                  */
1732
1733                 /*
1734                  * For extents, iter.pos won't necessarily be the same as
1735                  * bkey_start_pos(k.k) (for non extents they always will be the
1736                  * same). It's important that we delete starting from iter.pos
1737                  * because the range we want to delete could start in the middle
1738                  * of k.
1739                  *
1740                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1741                  * bkey_start_pos(k.k)).
1742                  */
1743                 delete.k.p = iter.pos;
1744
1745                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1746                         unsigned max_sectors =
1747                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1748
1749                         /* create the biggest key we can */
1750                         bch2_key_resize(&delete.k, max_sectors);
1751                         bch2_cut_back(end, &delete);
1752
1753                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1754                         if (ret)
1755                                 break;
1756                 }
1757
1758                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1759                         bch2_trans_commit(trans, &disk_res, journal_seq,
1760                                           BTREE_INSERT_NOFAIL);
1761                 bch2_disk_reservation_put(trans->c, &disk_res);
1762                 if (ret)
1763                         break;
1764         }
1765
1766         if (ret == -EINTR) {
1767                 ret = 0;
1768                 goto retry;
1769         }
1770
1771         bch2_trans_iter_exit(trans, &iter);
1772         return ret;
1773 }
1774
1775 /*
1776  * bch_btree_delete_range - delete everything within a given range
1777  *
1778  * Range is a half open interval - [start, end)
1779  */
1780 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1781                             struct bpos start, struct bpos end,
1782                             unsigned update_flags,
1783                             u64 *journal_seq)
1784 {
1785         return bch2_trans_do(c, NULL, journal_seq, 0,
1786                              bch2_btree_delete_range_trans(&trans, id, start, end,
1787                                                            update_flags, journal_seq));
1788 }
1789
1790 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1791 {
1792         unsigned len = strlen(msg);
1793         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1794         struct jset_entry_log *l;
1795         int ret;
1796
1797         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1798         if (ret)
1799                 return ret;
1800
1801         l = (void *) &darray_top(trans->extra_journal_entries);
1802         l->entry.u64s           = cpu_to_le16(u64s);
1803         l->entry.btree_id       = 0;
1804         l->entry.level          = 1;
1805         l->entry.type           = BCH_JSET_ENTRY_log;
1806         l->entry.pad[0]         = 0;
1807         l->entry.pad[1]         = 0;
1808         l->entry.pad[2]         = 0;
1809         memcpy(l->d, msg, len);
1810         while (len & 7)
1811                 l->d[len++] = '\0';
1812
1813         trans->extra_journal_entries.nr += jset_u64s(u64s);
1814         return 0;
1815 }