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 void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  28 {
  29 #ifdef CONFIG_BCACHEFS_DEBUG
  30         struct bch_fs *c = trans->c;
  31         struct bkey u;
  32         struct bkey_s_c k = bch2_btree_path_peek_slot(i->path, &u);
  33
  34         if (unlikely(trans->journal_replay_not_finished)) {
  35                 struct bkey_i *j_k =
  36                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  37
  38                 if (j_k)
  39                         k = bkey_i_to_s_c(j_k);
  40         }
  41
  42         i->old_k.needs_whiteout = k.k->needs_whiteout;
  43
  44         BUG_ON(memcmp(&i->old_k, k.k, sizeof(struct bkey)));
  45         BUG_ON(i->old_v != k.v);
  46 #endif
  47 }
  48
  49 static int __must_check
  50 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  51                           struct bkey_i *, enum btree_update_flags);
  52
  53 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  54                                          const struct btree_insert_entry *r)
  55 {
  56         return   cmp_int(l->btree_id,   r->btree_id) ?:
  57                  cmp_int(l->cached,     r->cached) ?:
  58                  -cmp_int(l->level,     r->level) ?:
  59                  bpos_cmp(l->k->k.p,    r->k->k.p);
  60 }
  61
  62 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  63 {
  64         return i->path->l + i->level;
  65 }
  66
  67 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  68                                      struct btree_insert_entry *i)
  69 {
  70         return i != trans->updates &&
  71                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  72 }
  73
  74 static inline bool same_leaf_as_next(struct btree_trans *trans,
  75                                      struct btree_insert_entry *i)
  76 {
  77         return i + 1 < trans->updates + trans->nr_updates &&
  78                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  79 }
  80
  81 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  82                                            struct btree_path *path,
  83                                            struct btree *b)
  84 {
  85         struct bch_fs *c = trans->c;
  86
  87         if (path->cached)
  88                 return;
  89
  90         if (unlikely(btree_node_just_written(b)) &&
  91             bch2_btree_post_write_cleanup(c, b))
  92                 bch2_trans_node_reinit_iter(trans, b);
  93
  94         /*
  95          * If the last bset has been written, or if it's gotten too big - start
  96          * a new bset to insert into:
  97          */
  98         if (want_new_bset(c, b))
  99                 bch2_btree_init_next(trans, b);
 100 }
 101
 102 /* Inserting into a given leaf node (last stage of insert): */
 103
 104 /* Handle overwrites and do insert, for non extents: */
 105 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 106                                 struct btree_path *path,
 107                                 struct btree *b,
 108                                 struct btree_node_iter *node_iter,
 109                                 struct bkey_i *insert)
 110 {
 111         struct bkey_packed *k;
 112         unsigned clobber_u64s = 0, new_u64s = 0;
 113
 114         EBUG_ON(btree_node_just_written(b));
 115         EBUG_ON(bset_written(b, btree_bset_last(b)));
 116         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 117         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 118         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 119         EBUG_ON(insert->k.u64s >
 120                 bch_btree_keys_u64s_remaining(trans->c, b));
 121
 122         k = bch2_btree_node_iter_peek_all(node_iter, b);
 123         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 124                 k = NULL;
 125
 126         /* @k is the key being overwritten/deleted, if any: */
 127         EBUG_ON(k && bkey_deleted(k));
 128
 129         /* Deleting, but not found? nothing to do: */
 130         if (bkey_deleted(&insert->k) && !k)
 131                 return false;
 132
 133         if (bkey_deleted(&insert->k)) {
 134                 /* Deleting: */
 135                 btree_account_key_drop(b, k);
 136                 k->type = KEY_TYPE_deleted;
 137
 138                 if (k->needs_whiteout)
 139                         push_whiteout(trans->c, b, insert->k.p);
 140                 k->needs_whiteout = false;
 141
 142                 if (k >= btree_bset_last(b)->start) {
 143                         clobber_u64s = k->u64s;
 144                         bch2_bset_delete(b, k, clobber_u64s);
 145                         goto fix_iter;
 146                 } else {
 147                         bch2_btree_path_fix_key_modified(trans, b, k);
 148                 }
 149
 150                 return true;
 151         }
 152
 153         if (k) {
 154                 /* Overwriting: */
 155                 btree_account_key_drop(b, k);
 156                 k->type = KEY_TYPE_deleted;
 157
 158                 insert->k.needs_whiteout = k->needs_whiteout;
 159                 k->needs_whiteout = false;
 160
 161                 if (k >= btree_bset_last(b)->start) {
 162                         clobber_u64s = k->u64s;
 163                         goto overwrite;
 164                 } else {
 165                         bch2_btree_path_fix_key_modified(trans, b, k);
 166                 }
 167         }
 168
 169         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 170 overwrite:
 171         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 172         new_u64s = k->u64s;
 173 fix_iter:
 174         if (clobber_u64s != new_u64s)
 175                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 176                                          clobber_u64s, new_u64s);
 177         return true;
 178 }
 179
 180 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 181                                unsigned i, u64 seq)
 182 {
 183         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 184         struct btree_write *w = container_of(pin, struct btree_write, journal);
 185         struct btree *b = container_of(w, struct btree, writes[i]);
 186         struct btree_trans trans;
 187         unsigned long old, new, v;
 188         unsigned idx = w - b->writes;
 189
 190         bch2_trans_init(&trans, c, 0, 0);
 191
 192         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 193         v = READ_ONCE(b->flags);
 194
 195         do {
 196                 old = new = v;
 197
 198                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 199                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 200                     w->journal.seq != seq)
 201                         break;
 202
 203                 new &= ~BTREE_WRITE_TYPE_MASK;
 204                 new |= BTREE_WRITE_journal_reclaim;
 205                 new |= 1 << BTREE_NODE_need_write;
 206         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 207
 208         btree_node_write_if_need(c, b, SIX_LOCK_read);
 209         six_unlock_read(&b->c.lock);
 210
 211         bch2_trans_exit(&trans);
 212         return 0;
 213 }
 214
 215 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 216 {
 217         return __btree_node_flush(j, pin, 0, seq);
 218 }
 219
 220 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 221 {
 222         return __btree_node_flush(j, pin, 1, seq);
 223 }
 224
 225 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 226                                        struct btree *b, u64 seq)
 227 {
 228         struct btree_write *w = btree_current_write(b);
 229
 230         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 231                              btree_node_write_idx(b) == 0
 232                              ? btree_node_flush0
 233                              : btree_node_flush1);
 234 }
 235
 236 /**
 237  * btree_insert_key - insert a key one key into a leaf node
 238  */
 239 static void btree_insert_key_leaf(struct btree_trans *trans,
 240                                   struct btree_insert_entry *insert)
 241 {
 242         struct bch_fs *c = trans->c;
 243         struct btree *b = insert_l(insert)->b;
 244         struct bset_tree *t = bset_tree_last(b);
 245         struct bset *i = bset(b, t);
 246         int old_u64s = bset_u64s(t);
 247         int old_live_u64s = b->nr.live_u64s;
 248         int live_u64s_added, u64s_added;
 249
 250         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 251                                         &insert_l(insert)->iter, insert->k)))
 252                 return;
 253
 254         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 255                                          le64_to_cpu(i->journal_seq)));
 256
 257         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 258
 259         if (unlikely(!btree_node_dirty(b)))
 260                 set_btree_node_dirty_acct(c, b);
 261
 262         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 263         u64s_added = (int) bset_u64s(t) - old_u64s;
 264
 265         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 266                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 267         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 268                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 269
 270         if (u64s_added > live_u64s_added &&
 271             bch2_maybe_compact_whiteouts(c, b))
 272                 bch2_trans_node_reinit_iter(trans, b);
 273 }
 274
 275 /* Cached btree updates: */
 276
 277 /* Normal update interface: */
 278
 279 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 280                                              struct btree_insert_entry *i)
 281 {
 282         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 283         BUG_ON(i->cached        != i->path->cached);
 284         BUG_ON(i->level         != i->path->level);
 285         BUG_ON(i->btree_id      != i->path->btree_id);
 286         EBUG_ON(!i->level &&
 287                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 288                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 289                 i->k->k.p.snapshot &&
 290                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 291 }
 292
 293 static noinline int
 294 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 295                                    unsigned long trace_ip)
 296 {
 297         struct bch_fs *c = trans->c;
 298         int ret;
 299
 300         bch2_trans_unlock(trans);
 301
 302         ret = bch2_journal_preres_get(&c->journal,
 303                         &trans->journal_preres, u64s, 0);
 304         if (ret)
 305                 return ret;
 306
 307         ret = bch2_trans_relock(trans);
 308         if (ret) {
 309                 trace_and_count(c, trans_restart_journal_preres_get, trans, trace_ip, 0);
 310                 return ret;
 311         }
 312
 313         return 0;
 314 }
 315
 316 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 317                                              unsigned flags)
 318 {
 319         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 320                                     trans->journal_u64s,
 321                                     flags|
 322                                     (trans->flags & JOURNAL_WATERMARK_MASK));
 323 }
 324
 325 #define JSET_ENTRY_LOG_U64s             4
 326
 327 static void journal_transaction_name(struct btree_trans *trans)
 328 {
 329         struct bch_fs *c = trans->c;
 330         struct journal *j = &c->journal;
 331         struct jset_entry *entry =
 332                 bch2_journal_add_entry(j, &trans->journal_res,
 333                                        BCH_JSET_ENTRY_log, 0, 0,
 334                                        JSET_ENTRY_LOG_U64s);
 335         struct jset_entry_log *l =
 336                 container_of(entry, struct jset_entry_log, entry);
 337
 338         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 339 }
 340
 341 static inline int btree_key_can_insert(struct btree_trans *trans,
 342                                        struct btree *b, unsigned u64s)
 343 {
 344         struct bch_fs *c = trans->c;
 345
 346         if (!bch2_btree_node_insert_fits(c, b, u64s))
 347                 return -BCH_ERR_btree_insert_btree_node_full;
 348
 349         return 0;
 350 }
 351
 352 static int btree_key_can_insert_cached(struct btree_trans *trans,
 353                                        struct btree_path *path,
 354                                        unsigned u64s)
 355 {
 356         struct bch_fs *c = trans->c;
 357         struct bkey_cached *ck = (void *) path->l[0].b;
 358         struct btree_insert_entry *i;
 359         unsigned new_u64s;
 360         struct bkey_i *new_k;
 361
 362         EBUG_ON(path->level);
 363
 364         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 365             bch2_btree_key_cache_must_wait(c) &&
 366             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 367                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 368
 369         /*
 370          * bch2_varint_decode can read past the end of the buffer by at most 7
 371          * bytes (it won't be used):
 372          */
 373         u64s += 1;
 374
 375         if (u64s <= ck->u64s)
 376                 return 0;
 377
 378         new_u64s        = roundup_pow_of_two(u64s);
 379         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 380         if (!new_k) {
 381                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 382                         bch2_btree_ids[path->btree_id], new_u64s);
 383                 return -ENOMEM;
 384         }
 385
 386         trans_for_each_update(trans, i)
 387                 if (i->old_v == &ck->k->v)
 388                         i->old_v = &new_k->v;
 389
 390         ck->u64s        = new_u64s;
 391         ck->k           = new_k;
 392         return 0;
 393 }
 394
 395 /* Triggers: */
 396
 397 static int run_one_mem_trigger(struct btree_trans *trans,
 398                                struct btree_insert_entry *i,
 399                                unsigned flags)
 400 {
 401         struct bkey_s_c old = { &i->old_k, i->old_v };
 402         struct bkey_i *new = i->k;
 403         int ret;
 404
 405         verify_update_old_key(trans, i);
 406
 407         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 408                 return 0;
 409
 410         if (!btree_node_type_needs_gc(i->btree_id))
 411                 return 0;
 412
 413         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 414             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 415             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 416                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 417                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 418         } else {
 419                 struct bkey             _deleted = KEY(0, 0, 0);
 420                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 421
 422                 _deleted.p = i->path->pos;
 423
 424                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 425                                 BTREE_TRIGGER_INSERT|flags) ?:
 426                         bch2_mark_key(trans, old, deleted,
 427                                 BTREE_TRIGGER_OVERWRITE|flags);
 428         }
 429
 430         return ret;
 431 }
 432
 433 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 434                                  bool overwrite)
 435 {
 436         /*
 437          * Transactional triggers create new btree_insert_entries, so we can't
 438          * pass them a pointer to a btree_insert_entry, that memory is going to
 439          * move:
 440          */
 441         struct bkey old_k = i->old_k;
 442         struct bkey_s_c old = { &old_k, i->old_v };
 443
 444         verify_update_old_key(trans, i);
 445
 446         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 447             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 448                 return 0;
 449
 450         if (!i->insert_trigger_run &&
 451             !i->overwrite_trigger_run &&
 452             bch2_bkey_ops[old.k->type].trans_trigger ==
 453             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 454             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 455                 i->overwrite_trigger_run = true;
 456                 i->insert_trigger_run = true;
 457                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 458                                            BTREE_TRIGGER_INSERT|
 459                                            BTREE_TRIGGER_OVERWRITE|
 460                                            i->flags) ?: 1;
 461         } else if (overwrite && !i->overwrite_trigger_run) {
 462                 i->overwrite_trigger_run = true;
 463                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 464         } else if (!overwrite && !i->insert_trigger_run) {
 465                 i->insert_trigger_run = true;
 466                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 467         } else {
 468                 return 0;
 469         }
 470 }
 471
 472 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 473                               struct btree_insert_entry *btree_id_start)
 474 {
 475         struct btree_insert_entry *i;
 476         bool trans_trigger_run;
 477         int ret, overwrite;
 478
 479         for (overwrite = 1; overwrite >= 0; --overwrite) {
 480
 481                 /*
 482                  * Running triggers will append more updates to the list of updates as
 483                  * we're walking it:
 484                  */
 485                 do {
 486                         trans_trigger_run = false;
 487
 488                         for (i = btree_id_start;
 489                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 490                              i++) {
 491                                 if (i->btree_id != btree_id)
 492                                         continue;
 493
 494                                 ret = run_one_trans_trigger(trans, i, overwrite);
 495                                 if (ret < 0)
 496                                         return ret;
 497                                 if (ret)
 498                                         trans_trigger_run = true;
 499                         }
 500                 } while (trans_trigger_run);
 501         }
 502
 503         return 0;
 504 }
 505
 506 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 507 {
 508         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 509         unsigned btree_id = 0;
 510         int ret = 0;
 511
 512         /*
 513          *
 514          * For a given btree, this algorithm runs insert triggers before
 515          * overwrite triggers: this is so that when extents are being moved
 516          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 517          * they are re-added.
 518          */
 519         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 520                 if (btree_id == BTREE_ID_alloc)
 521                         continue;
 522
 523                 while (btree_id_start < trans->updates + trans->nr_updates &&
 524                        btree_id_start->btree_id < btree_id)
 525                         btree_id_start++;
 526
 527                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 528                 if (ret)
 529                         return ret;
 530         }
 531
 532         trans_for_each_update(trans, i) {
 533                 if (i->btree_id > BTREE_ID_alloc)
 534                         break;
 535                 if (i->btree_id == BTREE_ID_alloc) {
 536                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 537                         if (ret)
 538                                 return ret;
 539                         break;
 540                 }
 541         }
 542
 543 #ifdef CONFIG_BCACHEFS_DEBUG
 544         trans_for_each_update(trans, i)
 545                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 546                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 547                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 548 #endif
 549         return 0;
 550 }
 551
 552 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 553 {
 554         struct bch_fs *c = trans->c;
 555         struct btree_insert_entry *i;
 556         int ret = 0;
 557
 558         trans_for_each_update(trans, i) {
 559                 /*
 560                  * XXX: synchronization of cached update triggers with gc
 561                  * XXX: synchronization of interior node updates with gc
 562                  */
 563                 BUG_ON(i->cached || i->level);
 564
 565                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 566                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 567                         if (ret)
 568                                 break;
 569                 }
 570         }
 571
 572         return ret;
 573 }
 574
 575 static inline int
 576 bch2_trans_commit_write_locked(struct btree_trans *trans,
 577                                struct btree_insert_entry **stopped_at,
 578                                unsigned long trace_ip)
 579 {
 580         struct bch_fs *c = trans->c;
 581         struct btree_insert_entry *i;
 582         struct btree_trans_commit_hook *h;
 583         unsigned u64s = 0;
 584         bool marking = false;
 585         int ret;
 586
 587         if (race_fault()) {
 588                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 589                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 590         }
 591
 592         /*
 593          * Check if the insert will fit in the leaf node with the write lock
 594          * held, otherwise another thread could write the node changing the
 595          * amount of space available:
 596          */
 597
 598         prefetch(&trans->c->journal.flags);
 599
 600         h = trans->hooks;
 601         while (h) {
 602                 ret = h->fn(trans, h);
 603                 if (ret)
 604                         return ret;
 605                 h = h->next;
 606         }
 607
 608         trans_for_each_update(trans, i) {
 609                 /* Multiple inserts might go to same leaf: */
 610                 if (!same_leaf_as_prev(trans, i))
 611                         u64s = 0;
 612
 613                 u64s += i->k->k.u64s;
 614                 ret = !i->cached
 615                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 616                         : btree_key_can_insert_cached(trans, i->path, u64s);
 617                 if (ret) {
 618                         *stopped_at = i;
 619                         return ret;
 620                 }
 621
 622                 if (btree_node_type_needs_gc(i->bkey_type))
 623                         marking = true;
 624         }
 625
 626         /*
 627          * Don't get journal reservation until after we know insert will
 628          * succeed:
 629          */
 630         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 631                 ret = bch2_trans_journal_res_get(trans,
 632                                 JOURNAL_RES_GET_NONBLOCK);
 633                 if (ret)
 634                         return ret;
 635
 636                 journal_transaction_name(trans);
 637         } else {
 638                 trans->journal_res.seq = c->journal.replay_journal_seq;
 639         }
 640
 641         /*
 642          * Not allowed to fail after we've gotten our journal reservation - we
 643          * have to use it:
 644          */
 645
 646         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 647             !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 648                 if (bch2_journal_seq_verify)
 649                         trans_for_each_update(trans, i)
 650                                 i->k->k.version.lo = trans->journal_res.seq;
 651                 else if (bch2_inject_invalid_keys)
 652                         trans_for_each_update(trans, i)
 653                                 i->k->k.version = MAX_VERSION;
 654         }
 655
 656         if (trans->fs_usage_deltas &&
 657             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 658                 return -BCH_ERR_btree_insert_need_mark_replicas;
 659
 660         trans_for_each_update(trans, i)
 661                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 662                         ret = run_one_mem_trigger(trans, i, i->flags);
 663                         if (ret)
 664                                 return ret;
 665                 }
 666
 667         if (unlikely(c->gc_pos.phase)) {
 668                 ret = bch2_trans_commit_run_gc_triggers(trans);
 669                 if  (ret)
 670                         return ret;
 671         }
 672
 673         if (unlikely(trans->extra_journal_entries.nr)) {
 674                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 675                                   trans->extra_journal_entries.data,
 676                                   trans->extra_journal_entries.nr);
 677
 678                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 679                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 680         }
 681
 682         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 683                 trans_for_each_update(trans, i) {
 684                         struct journal *j = &c->journal;
 685                         struct jset_entry *entry;
 686
 687                         if (i->key_cache_already_flushed)
 688                                 continue;
 689
 690                         verify_update_old_key(trans, i);
 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 #ifdef CONFIG_BCACHEFS_DEBUG
 765 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 766                                                    struct btree_insert_entry *i,
 767                                                    struct printbuf *err)
 768 {
 769         struct bch_fs *c = trans->c;
 770         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 771
 772         printbuf_reset(err);
 773         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 774                    trans->fn, (void *) i->ip_allocated);
 775         prt_newline(err);
 776         printbuf_indent_add(err, 2);
 777
 778         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 779         prt_newline(err);
 780
 781         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 782                           i->bkey_type, rw, err);
 783         bch2_print_string_as_lines(KERN_ERR, err->buf);
 784
 785         bch2_inconsistent_error(c);
 786         bch2_dump_trans_updates(trans);
 787         printbuf_exit(err);
 788
 789         return -EINVAL;
 790 }
 791 #endif
 792
 793 /*
 794  * Get journal reservation, take write locks, and attempt to do btree update(s):
 795  */
 796 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 797                                        struct btree_insert_entry **stopped_at,
 798                                        unsigned long trace_ip)
 799 {
 800         struct bch_fs *c = trans->c;
 801         struct btree_insert_entry *i;
 802         struct printbuf buf = PRINTBUF;
 803         int ret, u64s_delta = 0;
 804
 805 #ifdef CONFIG_BCACHEFS_DEBUG
 806         trans_for_each_update(trans, i) {
 807                 int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 808
 809                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 810                                                i->bkey_type, rw, &buf)))
 811                         return bch2_trans_commit_bkey_invalid(trans, i, &buf);
 812                 btree_insert_entry_checks(trans, i);
 813         }
 814 #endif
 815         printbuf_exit(&buf);
 816
 817         trans_for_each_update(trans, i) {
 818                 if (i->cached)
 819                         continue;
 820
 821                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 822                 u64s_delta -= i->old_btree_u64s;
 823
 824                 if (!same_leaf_as_next(trans, i)) {
 825                         if (u64s_delta <= 0) {
 826                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 827                                                         i->level, trans->flags);
 828                                 if (unlikely(ret))
 829                                         return ret;
 830                         }
 831
 832                         u64s_delta = 0;
 833                 }
 834         }
 835
 836         ret = bch2_journal_preres_get(&c->journal,
 837                         &trans->journal_preres, trans->journal_preres_u64s,
 838                         JOURNAL_RES_GET_NONBLOCK|
 839                         (trans->flags & JOURNAL_WATERMARK_MASK));
 840         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 841                 ret = bch2_trans_journal_preres_get_cold(trans,
 842                                                 trans->journal_preres_u64s, trace_ip);
 843         if (unlikely(ret))
 844                 return ret;
 845
 846         ret = trans_lock_write(trans);
 847         if (unlikely(ret))
 848                 return ret;
 849
 850         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 851
 852         if (!ret && unlikely(trans->journal_replay_not_finished))
 853                 bch2_drop_overwrites_from_journal(trans);
 854
 855         trans_for_each_update(trans, i)
 856                 if (!same_leaf_as_prev(trans, i))
 857                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 858                                                         insert_l(i)->b);
 859
 860         if (!ret && trans->journal_pin)
 861                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 862                                      trans->journal_pin, NULL);
 863
 864         /*
 865          * Drop journal reservation after dropping write locks, since dropping
 866          * the journal reservation may kick off a journal write:
 867          */
 868         bch2_journal_res_put(&c->journal, &trans->journal_res);
 869
 870         if (unlikely(ret))
 871                 return ret;
 872
 873         bch2_trans_downgrade(trans);
 874
 875         return 0;
 876 }
 877
 878 static int journal_reclaim_wait_done(struct bch_fs *c)
 879 {
 880         int ret = bch2_journal_error(&c->journal) ?:
 881                 !bch2_btree_key_cache_must_wait(c);
 882
 883         if (!ret)
 884                 journal_reclaim_kick(&c->journal);
 885         return ret;
 886 }
 887
 888 static noinline
 889 int bch2_trans_commit_error(struct btree_trans *trans,
 890                             struct btree_insert_entry *i,
 891                             int ret, unsigned long trace_ip)
 892 {
 893         struct bch_fs *c = trans->c;
 894
 895         switch (ret) {
 896         case -BCH_ERR_btree_insert_btree_node_full:
 897                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 898                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 899                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 900                 break;
 901         case -BCH_ERR_btree_insert_need_mark_replicas:
 902                 bch2_trans_unlock(trans);
 903
 904                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 905                 if (ret)
 906                         break;
 907
 908                 ret = bch2_trans_relock(trans);
 909                 if (ret)
 910                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 911                 break;
 912         case -BCH_ERR_journal_res_get_blocked:
 913                 bch2_trans_unlock(trans);
 914
 915                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 916                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 917                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 918                         break;
 919                 }
 920
 921                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 922                 if (ret)
 923                         break;
 924
 925                 ret = bch2_trans_relock(trans);
 926                 if (ret)
 927                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 928                 break;
 929         case -BCH_ERR_btree_insert_need_journal_reclaim:
 930                 bch2_trans_unlock(trans);
 931
 932                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 933
 934                 wait_event_freezable(c->journal.reclaim_wait,
 935                                      (ret = journal_reclaim_wait_done(c)));
 936                 if (ret < 0)
 937                         break;
 938
 939                 ret = bch2_trans_relock(trans);
 940                 if (ret)
 941                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 942                 break;
 943         default:
 944                 BUG_ON(ret >= 0);
 945                 break;
 946         }
 947
 948         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 949
 950         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 951                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 952                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 953                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 954
 955         return ret;
 956 }
 957
 958 static noinline int
 959 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 960 {
 961         struct bch_fs *c = trans->c;
 962         int ret;
 963
 964         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 965             test_bit(BCH_FS_STARTED, &c->flags))
 966                 return -BCH_ERR_erofs_trans_commit;
 967
 968         bch2_trans_unlock(trans);
 969
 970         ret =   bch2_fs_read_write_early(c) ?:
 971                 bch2_trans_relock(trans);
 972         if (ret)
 973                 return ret;
 974
 975         percpu_ref_get(&c->writes);
 976         return 0;
 977 }
 978
 979 /*
 980  * This is for updates done in the early part of fsck - btree_gc - before we've
 981  * gone RW. we only add the new key to the list of keys for journal replay to
 982  * do.
 983  */
 984 static noinline int
 985 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 986 {
 987         struct bch_fs *c = trans->c;
 988         struct btree_insert_entry *i;
 989         int ret = 0;
 990
 991         trans_for_each_update(trans, i) {
 992                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 993                 if (ret)
 994                         break;
 995         }
 996
 997         return ret;
 998 }
 999
1000 int __bch2_trans_commit(struct btree_trans *trans)
1001 {
1002         struct bch_fs *c = trans->c;
1003         struct btree_insert_entry *i = NULL;
1004         unsigned u64s;
1005         int ret = 0;
1006
1007         if (!trans->nr_updates &&
1008             !trans->extra_journal_entries.nr)
1009                 goto out_reset;
1010
1011         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1012                 lockdep_assert_held(&c->gc_lock);
1013
1014         ret = bch2_trans_commit_run_triggers(trans);
1015         if (ret)
1016                 goto out_reset;
1017
1018         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1019                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1020                 goto out_reset;
1021         }
1022
1023         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1024             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1025                 ret = bch2_trans_commit_get_rw_cold(trans);
1026                 if (ret)
1027                         goto out_reset;
1028         }
1029
1030         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1031
1032         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1033
1034         trans->journal_u64s             = trans->extra_journal_entries.nr;
1035         trans->journal_preres_u64s      = 0;
1036
1037         /* For journalling transaction name: */
1038         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1039
1040         trans_for_each_update(trans, i) {
1041                 EBUG_ON(!i->path->should_be_locked);
1042
1043                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1044                 if (unlikely(ret))
1045                         goto out;
1046
1047                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1048
1049                 if (i->key_cache_already_flushed)
1050                         continue;
1051
1052                 /* we're going to journal the key being updated: */
1053                 u64s = jset_u64s(i->k->k.u64s);
1054                 if (i->cached &&
1055                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1056                         trans->journal_preres_u64s += u64s;
1057                 trans->journal_u64s += u64s;
1058
1059                 /* and we're also going to log the overwrite: */
1060                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1061         }
1062
1063         if (trans->extra_journal_res) {
1064                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1065                                 trans->extra_journal_res,
1066                                 (trans->flags & BTREE_INSERT_NOFAIL)
1067                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1068                 if (ret)
1069                         goto err;
1070         }
1071 retry:
1072         EBUG_ON(trans->restarted);
1073         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1074
1075         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1076
1077         /* make sure we didn't drop or screw up locks: */
1078         bch2_trans_verify_locks(trans);
1079
1080         if (ret)
1081                 goto err;
1082
1083         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1084 out:
1085         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1086
1087         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1088                 percpu_ref_put(&c->writes);
1089 out_reset:
1090         bch2_trans_reset_updates(trans);
1091
1092         if (trans->fs_usage_deltas) {
1093                 trans->fs_usage_deltas->used = 0;
1094                 memset((void *) trans->fs_usage_deltas +
1095                        offsetof(struct replicas_delta_list, memset_start), 0,
1096                        (void *) &trans->fs_usage_deltas->memset_end -
1097                        (void *) &trans->fs_usage_deltas->memset_start);
1098         }
1099
1100         return ret;
1101 err:
1102         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1103         if (ret)
1104                 goto out;
1105
1106         goto retry;
1107 }
1108
1109 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1110                                           enum btree_id id,
1111                                           struct bpos pos)
1112 {
1113         struct bch_fs *c = trans->c;
1114         struct btree_iter iter;
1115         struct bkey_s_c k;
1116         int ret;
1117
1118         bch2_trans_iter_init(trans, &iter, id, pos,
1119                              BTREE_ITER_NOT_EXTENTS|
1120                              BTREE_ITER_ALL_SNAPSHOTS);
1121         while (1) {
1122                 k = bch2_btree_iter_prev(&iter);
1123                 ret = bkey_err(k);
1124                 if (ret)
1125                         break;
1126
1127                 if (!k.k)
1128                         break;
1129
1130                 if (!bkey_eq(pos, k.k->p))
1131                         break;
1132
1133                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1134                         ret = 1;
1135                         break;
1136                 }
1137         }
1138         bch2_trans_iter_exit(trans, &iter);
1139
1140         return ret;
1141 }
1142
1143 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1144                                           enum btree_id id,
1145                                           struct bpos pos)
1146 {
1147         if (!btree_type_has_snapshots(id) ||
1148             pos.snapshot == U32_MAX ||
1149             !snapshot_t(trans->c, pos.snapshot)->children[0])
1150                 return 0;
1151
1152         return __check_pos_snapshot_overwritten(trans, id, pos);
1153 }
1154
1155 int bch2_trans_update_extent(struct btree_trans *trans,
1156                              struct btree_iter *orig_iter,
1157                              struct bkey_i *insert,
1158                              enum btree_update_flags flags)
1159 {
1160         struct bch_fs *c = trans->c;
1161         struct btree_iter iter, update_iter;
1162         struct bpos start = bkey_start_pos(&insert->k);
1163         struct bkey_i *update;
1164         struct bkey_s_c k;
1165         enum btree_id btree_id = orig_iter->btree_id;
1166         int ret = 0, compressed_sectors;
1167
1168         bch2_trans_iter_init(trans, &iter, btree_id, start,
1169                              BTREE_ITER_INTENT|
1170                              BTREE_ITER_WITH_UPDATES|
1171                              BTREE_ITER_NOT_EXTENTS);
1172         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1173         if ((ret = bkey_err(k)))
1174                 goto err;
1175         if (!k.k)
1176                 goto out;
1177
1178         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1179                 /*
1180                  * We can't merge extents if they belong to interior snapshot
1181                  * tree nodes, and there's a snapshot in which one extent is
1182                  * visible and the other is not - i.e. if visibility is
1183                  * different.
1184                  *
1185                  * Instead of checking if visibilitiy of the two extents is
1186                  * different, for now we just check if either has been
1187                  * overwritten:
1188                  */
1189                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1190                 if (ret < 0)
1191                         goto err;
1192                 if (ret)
1193                         goto nomerge1;
1194
1195                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1196                 if (ret < 0)
1197                         goto err;
1198                 if (ret)
1199                         goto nomerge1;
1200
1201                 update = bch2_bkey_make_mut(trans, k);
1202                 if ((ret = PTR_ERR_OR_ZERO(update)))
1203                         goto err;
1204
1205                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1206                         ret = bch2_btree_delete_at(trans, &iter, flags);
1207                         if (ret)
1208                                 goto err;
1209
1210                         insert = update;
1211                         goto next;
1212                 }
1213         }
1214 nomerge1:
1215         ret = 0;
1216         if (bkey_eq(k.k->p, start))
1217                 goto next;
1218
1219         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1220                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1221                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1222
1223                 /*
1224                  * If we're going to be splitting a compressed extent, note it
1225                  * so that __bch2_trans_commit() can increase our disk
1226                  * reservation:
1227                  */
1228                 if (((front_split && back_split) ||
1229                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1230                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1231                         trans->extra_journal_res += compressed_sectors;
1232
1233                 if (front_split) {
1234                         update = bch2_bkey_make_mut(trans, k);
1235                         if ((ret = PTR_ERR_OR_ZERO(update)))
1236                                 goto err;
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_bkey_make_mut(trans, k);
1257                         if ((ret = PTR_ERR_OR_ZERO(update)))
1258                                 goto err;
1259
1260                         bch2_cut_front(start, update);
1261                         bch2_cut_back(insert->k.p, update);
1262
1263                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1264                                              BTREE_ITER_NOT_EXTENTS|
1265                                              BTREE_ITER_ALL_SNAPSHOTS|
1266                                              BTREE_ITER_INTENT);
1267                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1268                                 bch2_trans_update(trans, &update_iter, update,
1269                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1270                                                   flags);
1271                         bch2_trans_iter_exit(trans, &update_iter);
1272                         if (ret)
1273                                 goto err;
1274                 }
1275
1276                 if (bkey_le(k.k->p, insert->k.p)) {
1277                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1278                         if ((ret = PTR_ERR_OR_ZERO(update)))
1279                                 goto err;
1280
1281                         bkey_init(&update->k);
1282                         update->k.p = k.k->p;
1283
1284                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1285                                 update->k.p.snapshot = insert->k.p.snapshot;
1286                                 update->k.type = KEY_TYPE_whiteout;
1287                         }
1288
1289                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1290                                              BTREE_ITER_NOT_EXTENTS|
1291                                              BTREE_ITER_INTENT);
1292                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1293                                 bch2_trans_update(trans, &update_iter, update,
1294                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1295                                                   flags);
1296                         bch2_trans_iter_exit(trans, &update_iter);
1297
1298                         if (ret)
1299                                 goto err;
1300                 }
1301
1302                 if (back_split) {
1303                         update = bch2_bkey_make_mut(trans, k);
1304                         if ((ret = PTR_ERR_OR_ZERO(update)))
1305                                 goto err;
1306
1307                         bch2_cut_front(insert->k.p, update);
1308
1309                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1310                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1311                                                   flags);
1312                         if (ret)
1313                                 goto err;
1314                         goto out;
1315                 }
1316 next:
1317                 bch2_btree_iter_advance(&iter);
1318                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1319                 if ((ret = bkey_err(k)))
1320                         goto err;
1321                 if (!k.k)
1322                         goto out;
1323         }
1324
1325         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1326                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1327                 if (ret < 0)
1328                         goto err;
1329                 if (ret)
1330                         goto nomerge2;
1331
1332                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1333                 if (ret < 0)
1334                         goto err;
1335                 if (ret)
1336                         goto nomerge2;
1337
1338                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1339         }
1340 nomerge2:
1341         ret = 0;
1342 out:
1343         if (!bkey_deleted(&insert->k)) {
1344                 /*
1345                  * Rewinding iterators is expensive: get a new one and the one
1346                  * that points to the start of insert will be cloned from:
1347                  */
1348                 bch2_trans_iter_exit(trans, &iter);
1349                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1350                                      BTREE_ITER_NOT_EXTENTS|
1351                                      BTREE_ITER_INTENT);
1352                 ret   = bch2_btree_iter_traverse(&iter) ?:
1353                         bch2_trans_update(trans, &iter, insert, flags);
1354         }
1355 err:
1356         bch2_trans_iter_exit(trans, &iter);
1357
1358         return ret;
1359 }
1360
1361 /*
1362  * When deleting, check if we need to emit a whiteout (because we're overwriting
1363  * something in an ancestor snapshot)
1364  */
1365 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1366                                       enum btree_id btree_id, struct bpos pos)
1367 {
1368         struct btree_iter iter;
1369         struct bkey_s_c k;
1370         u32 snapshot = pos.snapshot;
1371         int ret;
1372
1373         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1374                 return 0;
1375
1376         pos.snapshot++;
1377
1378         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1379                            BTREE_ITER_ALL_SNAPSHOTS|
1380                            BTREE_ITER_NOPRESERVE, k, ret) {
1381                 if (!bkey_eq(k.k->p, pos))
1382                         break;
1383
1384                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1385                                               k.k->p.snapshot)) {
1386                         ret = !bkey_whiteout(k.k);
1387                         break;
1388                 }
1389         }
1390         bch2_trans_iter_exit(trans, &iter);
1391
1392         return ret;
1393 }
1394
1395 static int __must_check
1396 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1397                                 struct bkey_i *k, enum btree_update_flags flags,
1398                                 unsigned long ip);
1399
1400 static noinline int flush_new_cached_update(struct btree_trans *trans,
1401                                             struct btree_path *path,
1402                                             struct btree_insert_entry *i,
1403                                             enum btree_update_flags flags,
1404                                             unsigned long ip)
1405 {
1406         struct btree_path *btree_path;
1407         int ret;
1408
1409         i->key_cache_already_flushed = true;
1410         i->flags |= BTREE_TRIGGER_NORUN;
1411
1412         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1413                                    BTREE_ITER_INTENT, _THIS_IP_);
1414
1415         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1416         if (ret)
1417                 goto err;
1418
1419         btree_path_set_should_be_locked(btree_path);
1420         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1421 err:
1422         bch2_path_put(trans, btree_path, true);
1423         return ret;
1424 }
1425
1426 static int __must_check
1427 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1428                                 struct bkey_i *k, enum btree_update_flags flags,
1429                                 unsigned long ip)
1430 {
1431         struct bch_fs *c = trans->c;
1432         struct btree_insert_entry *i, n;
1433         int cmp;
1434
1435         EBUG_ON(!path->should_be_locked);
1436         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1437         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1438
1439         n = (struct btree_insert_entry) {
1440                 .flags          = flags,
1441                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1442                 .btree_id       = path->btree_id,
1443                 .level          = path->level,
1444                 .cached         = path->cached,
1445                 .path           = path,
1446                 .k              = k,
1447                 .ip_allocated   = ip,
1448         };
1449
1450 #ifdef CONFIG_BCACHEFS_DEBUG
1451         trans_for_each_update(trans, i)
1452                 BUG_ON(i != trans->updates &&
1453                        btree_insert_entry_cmp(i - 1, i) >= 0);
1454 #endif
1455
1456         /*
1457          * Pending updates are kept sorted: first, find position of new update,
1458          * then delete/trim any updates the new update overwrites:
1459          */
1460         trans_for_each_update(trans, i) {
1461                 cmp = btree_insert_entry_cmp(&n, i);
1462                 if (cmp <= 0)
1463                         break;
1464         }
1465
1466         if (!cmp && i < trans->updates + trans->nr_updates) {
1467                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1468
1469                 bch2_path_put(trans, i->path, true);
1470                 i->flags        = n.flags;
1471                 i->cached       = n.cached;
1472                 i->k            = n.k;
1473                 i->path         = n.path;
1474                 i->ip_allocated = n.ip_allocated;
1475         } else {
1476                 array_insert_item(trans->updates, trans->nr_updates,
1477                                   i - trans->updates, n);
1478
1479                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1480                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1481
1482                 if (unlikely(trans->journal_replay_not_finished)) {
1483                         struct bkey_i *j_k =
1484                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1485
1486                         if (j_k) {
1487                                 i->old_k = j_k->k;
1488                                 i->old_v = &j_k->v;
1489                         }
1490                 }
1491         }
1492
1493         __btree_path_get(i->path, true);
1494
1495         /*
1496          * If a key is present in the key cache, it must also exist in the
1497          * btree - this is necessary for cache coherency. When iterating over
1498          * a btree that's cached in the key cache, the btree iter code checks
1499          * the key cache - but the key has to exist in the btree for that to
1500          * work:
1501          */
1502         if (unlikely(path->cached && bkey_deleted(&i->old_k)))
1503                 return flush_new_cached_update(trans, path, i, flags, ip);
1504
1505         return 0;
1506 }
1507
1508 static inline int __must_check
1509 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1510                           struct bkey_i *k, enum btree_update_flags flags)
1511 {
1512         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1513 }
1514
1515 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1516                                    struct bkey_i *k, enum btree_update_flags flags)
1517 {
1518         struct btree_path *path = iter->update_path ?: iter->path;
1519         struct bkey_cached *ck;
1520         int ret;
1521
1522         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1523                 return bch2_trans_update_extent(trans, iter, k, flags);
1524
1525         if (bkey_deleted(&k->k) &&
1526             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1527             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1528                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1529                 if (unlikely(ret < 0))
1530                         return ret;
1531
1532                 if (ret)
1533                         k->k.type = KEY_TYPE_whiteout;
1534         }
1535
1536         /*
1537          * Ensure that updates to cached btrees go to the key cache:
1538          */
1539         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1540             !path->cached &&
1541             !path->level &&
1542             btree_id_cached(trans->c, path->btree_id)) {
1543                 if (!iter->key_cache_path ||
1544                     !iter->key_cache_path->should_be_locked ||
1545                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1546                         if (!iter->key_cache_path)
1547                                 iter->key_cache_path =
1548                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1549                                                       BTREE_ITER_INTENT|
1550                                                       BTREE_ITER_CACHED, _THIS_IP_);
1551
1552                         iter->key_cache_path =
1553                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1554                                                         iter->flags & BTREE_ITER_INTENT,
1555                                                         _THIS_IP_);
1556
1557                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1558                                                        BTREE_ITER_CACHED);
1559                         if (unlikely(ret))
1560                                 return ret;
1561
1562                         ck = (void *) iter->key_cache_path->l[0].b;
1563
1564                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1565                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1566                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1567                         }
1568
1569                         btree_path_set_should_be_locked(iter->key_cache_path);
1570                 }
1571
1572                 path = iter->key_cache_path;
1573         }
1574
1575         return bch2_trans_update_by_path(trans, path, k, flags);
1576 }
1577
1578 void bch2_trans_commit_hook(struct btree_trans *trans,
1579                             struct btree_trans_commit_hook *h)
1580 {
1581         h->next = trans->hooks;
1582         trans->hooks = h;
1583 }
1584
1585 int __bch2_btree_insert(struct btree_trans *trans,
1586                         enum btree_id id, struct bkey_i *k)
1587 {
1588         struct btree_iter iter;
1589         int ret;
1590
1591         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1592                              BTREE_ITER_INTENT);
1593         ret   = bch2_btree_iter_traverse(&iter) ?:
1594                 bch2_trans_update(trans, &iter, k, 0);
1595         bch2_trans_iter_exit(trans, &iter);
1596         return ret;
1597 }
1598
1599 /**
1600  * bch2_btree_insert - insert keys into the extent btree
1601  * @c:                  pointer to struct bch_fs
1602  * @id:                 btree to insert into
1603  * @insert_keys:        list of keys to insert
1604  * @hook:               insert callback
1605  */
1606 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1607                       struct bkey_i *k,
1608                       struct disk_reservation *disk_res,
1609                       u64 *journal_seq, int flags)
1610 {
1611         return bch2_trans_do(c, disk_res, journal_seq, flags,
1612                              __bch2_btree_insert(&trans, id, k));
1613 }
1614
1615 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1616                                 unsigned len, unsigned update_flags)
1617 {
1618         struct bkey_i *k;
1619
1620         k = bch2_trans_kmalloc(trans, sizeof(*k));
1621         if (IS_ERR(k))
1622                 return PTR_ERR(k);
1623
1624         bkey_init(&k->k);
1625         k->k.p = iter->pos;
1626         bch2_key_resize(&k->k, len);
1627         return bch2_trans_update(trans, iter, k, update_flags);
1628 }
1629
1630 int bch2_btree_delete_at(struct btree_trans *trans,
1631                          struct btree_iter *iter, unsigned update_flags)
1632 {
1633         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1634 }
1635
1636 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1637                                   struct bpos start, struct bpos end,
1638                                   unsigned update_flags,
1639                                   u64 *journal_seq)
1640 {
1641         u32 restart_count = trans->restart_count;
1642         struct btree_iter iter;
1643         struct bkey_s_c k;
1644         int ret = 0;
1645
1646         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1647         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1648                 struct disk_reservation disk_res =
1649                         bch2_disk_reservation_init(trans->c, 0);
1650                 struct bkey_i delete;
1651
1652                 ret = bkey_err(k);
1653                 if (ret)
1654                         goto err;
1655
1656                 bkey_init(&delete.k);
1657
1658                 /*
1659                  * This could probably be more efficient for extents:
1660                  */
1661
1662                 /*
1663                  * For extents, iter.pos won't necessarily be the same as
1664                  * bkey_start_pos(k.k) (for non extents they always will be the
1665                  * same). It's important that we delete starting from iter.pos
1666                  * because the range we want to delete could start in the middle
1667                  * of k.
1668                  *
1669                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1670                  * bkey_start_pos(k.k)).
1671                  */
1672                 delete.k.p = iter.pos;
1673
1674                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1675                         unsigned max_sectors =
1676                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1677
1678                         /* create the biggest key we can */
1679                         bch2_key_resize(&delete.k, max_sectors);
1680                         bch2_cut_back(end, &delete);
1681
1682                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1683                         if (ret)
1684                                 goto err;
1685                 }
1686
1687                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1688                         bch2_trans_commit(trans, &disk_res, journal_seq,
1689                                           BTREE_INSERT_NOFAIL);
1690                 bch2_disk_reservation_put(trans->c, &disk_res);
1691 err:
1692                 /*
1693                  * the bch2_trans_begin() call is in a weird place because we
1694                  * need to call it after every transaction commit, to avoid path
1695                  * overflow, but don't want to call it if the delete operation
1696                  * is a no-op and we have no work to do:
1697                  */
1698                 bch2_trans_begin(trans);
1699
1700                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1701                         ret = 0;
1702                 if (ret)
1703                         break;
1704         }
1705         bch2_trans_iter_exit(trans, &iter);
1706
1707         if (!ret && trans_was_restarted(trans, restart_count))
1708                 ret = -BCH_ERR_transaction_restart_nested;
1709         return ret;
1710 }
1711
1712 /*
1713  * bch_btree_delete_range - delete everything within a given range
1714  *
1715  * Range is a half open interval - [start, end)
1716  */
1717 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1718                             struct bpos start, struct bpos end,
1719                             unsigned update_flags,
1720                             u64 *journal_seq)
1721 {
1722         int ret = bch2_trans_run(c,
1723                         bch2_btree_delete_range_trans(&trans, id, start, end,
1724                                                       update_flags, journal_seq));
1725         if (ret == -BCH_ERR_transaction_restart_nested)
1726                 ret = 0;
1727         return ret;
1728 }
1729
1730 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1731 {
1732         unsigned len = strlen(msg);
1733         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1734         struct jset_entry_log *l;
1735         int ret;
1736
1737         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1738         if (ret)
1739                 return ret;
1740
1741         l = (void *) &darray_top(trans->extra_journal_entries);
1742         l->entry.u64s           = cpu_to_le16(u64s);
1743         l->entry.btree_id       = 0;
1744         l->entry.level          = 1;
1745         l->entry.type           = BCH_JSET_ENTRY_log;
1746         l->entry.pad[0]         = 0;
1747         l->entry.pad[1]         = 0;
1748         l->entry.pad[2]         = 0;
1749         memcpy(l->d, msg, len);
1750         while (len & 7)
1751                 l->d[len++] = '\0';
1752
1753         trans->extra_journal_entries.nr += jset_u64s(u64s);
1754         return 0;
1755 }