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         struct bch_fs *c = trans->c;
 320         int ret;
 321
 322         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 323                                    trans->journal_u64s,
 324                                    flags|
 325                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 326
 327         return ret == -EAGAIN ? -BCH_ERR_btree_insert_need_journal_res : ret;
 328 }
 329
 330 #define JSET_ENTRY_LOG_U64s             4
 331
 332 static void journal_transaction_name(struct btree_trans *trans)
 333 {
 334         struct bch_fs *c = trans->c;
 335         struct journal *j = &c->journal;
 336         struct jset_entry *entry =
 337                 bch2_journal_add_entry(j, &trans->journal_res,
 338                                        BCH_JSET_ENTRY_log, 0, 0,
 339                                        JSET_ENTRY_LOG_U64s);
 340         struct jset_entry_log *l =
 341                 container_of(entry, struct jset_entry_log, entry);
 342
 343         strncpy(l->d, trans->fn, JSET_ENTRY_LOG_U64s * sizeof(u64));
 344 }
 345
 346 static inline int btree_key_can_insert(struct btree_trans *trans,
 347                                        struct btree *b, unsigned u64s)
 348 {
 349         struct bch_fs *c = trans->c;
 350
 351         if (!bch2_btree_node_insert_fits(c, b, u64s))
 352                 return -BCH_ERR_btree_insert_btree_node_full;
 353
 354         return 0;
 355 }
 356
 357 static int btree_key_can_insert_cached(struct btree_trans *trans,
 358                                        struct btree_path *path,
 359                                        unsigned u64s)
 360 {
 361         struct bch_fs *c = trans->c;
 362         struct bkey_cached *ck = (void *) path->l[0].b;
 363         struct btree_insert_entry *i;
 364         unsigned new_u64s;
 365         struct bkey_i *new_k;
 366
 367         EBUG_ON(path->level);
 368
 369         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 370             bch2_btree_key_cache_must_wait(c) &&
 371             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 372                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 373
 374         /*
 375          * bch2_varint_decode can read past the end of the buffer by at most 7
 376          * bytes (it won't be used):
 377          */
 378         u64s += 1;
 379
 380         if (u64s <= ck->u64s)
 381                 return 0;
 382
 383         new_u64s        = roundup_pow_of_two(u64s);
 384         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 385         if (!new_k) {
 386                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 387                         bch2_btree_ids[path->btree_id], new_u64s);
 388                 return -ENOMEM;
 389         }
 390
 391         trans_for_each_update(trans, i)
 392                 if (i->old_v == &ck->k->v)
 393                         i->old_v = &new_k->v;
 394
 395         ck->u64s        = new_u64s;
 396         ck->k           = new_k;
 397         return 0;
 398 }
 399
 400 /* Triggers: */
 401
 402 static int run_one_mem_trigger(struct btree_trans *trans,
 403                                struct btree_insert_entry *i,
 404                                unsigned flags)
 405 {
 406         struct bkey_s_c old = { &i->old_k, i->old_v };
 407         struct bkey_i *new = i->k;
 408         int ret;
 409
 410         verify_update_old_key(trans, i);
 411
 412         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 413                 return 0;
 414
 415         if (!btree_node_type_needs_gc(i->btree_id))
 416                 return 0;
 417
 418         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 419             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 420             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 421                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 422                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 423         } else {
 424                 struct bkey             _deleted = KEY(0, 0, 0);
 425                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 426
 427                 _deleted.p = i->path->pos;
 428
 429                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 430                                 BTREE_TRIGGER_INSERT|flags) ?:
 431                         bch2_mark_key(trans, old, deleted,
 432                                 BTREE_TRIGGER_OVERWRITE|flags);
 433         }
 434
 435         return ret;
 436 }
 437
 438 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 439                                  bool overwrite)
 440 {
 441         /*
 442          * Transactional triggers create new btree_insert_entries, so we can't
 443          * pass them a pointer to a btree_insert_entry, that memory is going to
 444          * move:
 445          */
 446         struct bkey old_k = i->old_k;
 447         struct bkey_s_c old = { &old_k, i->old_v };
 448
 449         verify_update_old_key(trans, i);
 450
 451         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 452             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 453                 return 0;
 454
 455         if (!i->insert_trigger_run &&
 456             !i->overwrite_trigger_run &&
 457             bch2_bkey_ops[old.k->type].trans_trigger ==
 458             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 459             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 460                 i->overwrite_trigger_run = true;
 461                 i->insert_trigger_run = true;
 462                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 463                                            BTREE_TRIGGER_INSERT|
 464                                            BTREE_TRIGGER_OVERWRITE|
 465                                            i->flags) ?: 1;
 466         } else if (overwrite && !i->overwrite_trigger_run) {
 467                 i->overwrite_trigger_run = true;
 468                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 469         } else if (!overwrite && !i->insert_trigger_run) {
 470                 i->insert_trigger_run = true;
 471                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 472         } else {
 473                 return 0;
 474         }
 475 }
 476
 477 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 478                               struct btree_insert_entry *btree_id_start)
 479 {
 480         struct btree_insert_entry *i;
 481         bool trans_trigger_run;
 482         int ret, overwrite;
 483
 484         for (overwrite = 1; overwrite >= 0; --overwrite) {
 485
 486                 /*
 487                  * Running triggers will append more updates to the list of updates as
 488                  * we're walking it:
 489                  */
 490                 do {
 491                         trans_trigger_run = false;
 492
 493                         for (i = btree_id_start;
 494                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 495                              i++) {
 496                                 if (i->btree_id != btree_id)
 497                                         continue;
 498
 499                                 ret = run_one_trans_trigger(trans, i, overwrite);
 500                                 if (ret < 0)
 501                                         return ret;
 502                                 if (ret)
 503                                         trans_trigger_run = true;
 504                         }
 505                 } while (trans_trigger_run);
 506         }
 507
 508         return 0;
 509 }
 510
 511 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 512 {
 513         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 514         unsigned btree_id = 0;
 515         int ret = 0;
 516
 517         /*
 518          *
 519          * For a given btree, this algorithm runs insert triggers before
 520          * overwrite triggers: this is so that when extents are being moved
 521          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 522          * they are re-added.
 523          */
 524         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 525                 if (btree_id == BTREE_ID_alloc)
 526                         continue;
 527
 528                 while (btree_id_start < trans->updates + trans->nr_updates &&
 529                        btree_id_start->btree_id < btree_id)
 530                         btree_id_start++;
 531
 532                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 533                 if (ret)
 534                         return ret;
 535         }
 536
 537         trans_for_each_update(trans, i) {
 538                 if (i->btree_id > BTREE_ID_alloc)
 539                         break;
 540                 if (i->btree_id == BTREE_ID_alloc) {
 541                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 542                         if (ret)
 543                                 return ret;
 544                         break;
 545                 }
 546         }
 547
 548 #ifdef CONFIG_BCACHEFS_DEBUG
 549         trans_for_each_update(trans, i)
 550                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 551                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 552                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 553 #endif
 554         return 0;
 555 }
 556
 557 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 558 {
 559         struct bch_fs *c = trans->c;
 560         struct btree_insert_entry *i;
 561         int ret = 0;
 562
 563         trans_for_each_update(trans, i) {
 564                 /*
 565                  * XXX: synchronization of cached update triggers with gc
 566                  * XXX: synchronization of interior node updates with gc
 567                  */
 568                 BUG_ON(i->cached || i->level);
 569
 570                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 571                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 572                         if (ret)
 573                                 break;
 574                 }
 575         }
 576
 577         return ret;
 578 }
 579
 580 static inline int
 581 bch2_trans_commit_write_locked(struct btree_trans *trans,
 582                                struct btree_insert_entry **stopped_at,
 583                                unsigned long trace_ip)
 584 {
 585         struct bch_fs *c = trans->c;
 586         struct btree_insert_entry *i;
 587         struct btree_trans_commit_hook *h;
 588         unsigned u64s = 0;
 589         bool marking = false;
 590         int ret;
 591
 592         if (race_fault()) {
 593                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 594                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 595         }
 596
 597         /*
 598          * Check if the insert will fit in the leaf node with the write lock
 599          * held, otherwise another thread could write the node changing the
 600          * amount of space available:
 601          */
 602
 603         prefetch(&trans->c->journal.flags);
 604
 605         h = trans->hooks;
 606         while (h) {
 607                 ret = h->fn(trans, h);
 608                 if (ret)
 609                         return ret;
 610                 h = h->next;
 611         }
 612
 613         trans_for_each_update(trans, i) {
 614                 /* Multiple inserts might go to same leaf: */
 615                 if (!same_leaf_as_prev(trans, i))
 616                         u64s = 0;
 617
 618                 u64s += i->k->k.u64s;
 619                 ret = !i->cached
 620                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 621                         : btree_key_can_insert_cached(trans, i->path, u64s);
 622                 if (ret) {
 623                         *stopped_at = i;
 624                         return ret;
 625                 }
 626
 627                 if (btree_node_type_needs_gc(i->bkey_type))
 628                         marking = true;
 629         }
 630
 631         /*
 632          * Don't get journal reservation until after we know insert will
 633          * succeed:
 634          */
 635         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 636                 ret = bch2_trans_journal_res_get(trans,
 637                                 JOURNAL_RES_GET_NONBLOCK);
 638                 if (ret)
 639                         return ret;
 640
 641                 journal_transaction_name(trans);
 642         } else {
 643                 trans->journal_res.seq = c->journal.replay_journal_seq;
 644         }
 645
 646         if (unlikely(trans->extra_journal_entries.nr)) {
 647                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 648                                   trans->extra_journal_entries.data,
 649                                   trans->extra_journal_entries.nr);
 650
 651                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 652                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 653         }
 654
 655         /*
 656          * Not allowed to fail after we've gotten our journal reservation - we
 657          * have to use it:
 658          */
 659
 660         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 661                 if (bch2_journal_seq_verify)
 662                         trans_for_each_update(trans, i)
 663                                 i->k->k.version.lo = trans->journal_res.seq;
 664                 else if (bch2_inject_invalid_keys)
 665                         trans_for_each_update(trans, i)
 666                                 i->k->k.version = MAX_VERSION;
 667         }
 668
 669         if (trans->fs_usage_deltas &&
 670             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 671                 return -BCH_ERR_btree_insert_need_mark_replicas;
 672
 673         trans_for_each_update(trans, i)
 674                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 675                         ret = run_one_mem_trigger(trans, i, i->flags);
 676                         if (ret)
 677                                 return ret;
 678                 }
 679
 680         if (unlikely(c->gc_pos.phase)) {
 681                 ret = bch2_trans_commit_run_gc_triggers(trans);
 682                 if  (ret)
 683                         return ret;
 684         }
 685
 686         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 687                 trans_for_each_update(trans, i) {
 688                         struct journal *j = &c->journal;
 689                         struct jset_entry *entry;
 690
 691                         if (i->key_cache_already_flushed)
 692                                 continue;
 693
 694                         verify_update_old_key(trans, i);
 695
 696                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 697                                                BCH_JSET_ENTRY_overwrite,
 698                                                i->btree_id, i->level,
 699                                                i->old_k.u64s);
 700                         bkey_reassemble(&entry->start[0],
 701                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 702
 703                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 704                                                BCH_JSET_ENTRY_btree_keys,
 705                                                i->btree_id, i->level,
 706                                                i->k->k.u64s);
 707                         bkey_copy(&entry->start[0], i->k);
 708                 }
 709
 710                 if (trans->journal_seq)
 711                         *trans->journal_seq = trans->journal_res.seq;
 712         }
 713
 714         trans_for_each_update(trans, i) {
 715                 i->k->k.needs_whiteout = false;
 716
 717                 if (!i->cached)
 718                         btree_insert_key_leaf(trans, i);
 719                 else if (!i->key_cache_already_flushed)
 720                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 721                 else {
 722                         bch2_btree_key_cache_drop(trans, i->path);
 723                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 724                 }
 725         }
 726
 727         return ret;
 728 }
 729
 730 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 731 {
 732         while (--i >= trans->updates) {
 733                 if (same_leaf_as_prev(trans, i))
 734                         continue;
 735
 736                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 737         }
 738
 739         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 740         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 741 }
 742
 743 static inline int trans_lock_write(struct btree_trans *trans)
 744 {
 745         struct btree_insert_entry *i;
 746
 747         trans_for_each_update(trans, i) {
 748                 if (same_leaf_as_prev(trans, i))
 749                         continue;
 750
 751                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 752                         return trans_lock_write_fail(trans, i);
 753
 754                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 755         }
 756
 757         return 0;
 758 }
 759
 760 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 761 {
 762         struct btree_insert_entry *i;
 763
 764         trans_for_each_update(trans, i)
 765                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 766 }
 767
 768 #ifdef CONFIG_BCACHEFS_DEBUG
 769 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 770                                                    struct btree_insert_entry *i,
 771                                                    struct printbuf *err)
 772 {
 773         struct bch_fs *c = trans->c;
 774         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 775
 776         printbuf_reset(err);
 777         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 778                    trans->fn, (void *) i->ip_allocated);
 779         prt_newline(err);
 780         printbuf_indent_add(err, 2);
 781
 782         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 783         prt_newline(err);
 784
 785         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 786                           i->bkey_type, rw, err);
 787         bch2_print_string_as_lines(KERN_ERR, err->buf);
 788
 789         bch2_inconsistent_error(c);
 790         bch2_dump_trans_updates(trans);
 791         printbuf_exit(err);
 792
 793         return -EINVAL;
 794 }
 795 #endif
 796
 797 /*
 798  * Get journal reservation, take write locks, and attempt to do btree update(s):
 799  */
 800 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 801                                        struct btree_insert_entry **stopped_at,
 802                                        unsigned long trace_ip)
 803 {
 804         struct bch_fs *c = trans->c;
 805         struct btree_insert_entry *i;
 806         struct printbuf buf = PRINTBUF;
 807         int ret, u64s_delta = 0;
 808
 809 #ifdef CONFIG_BCACHEFS_DEBUG
 810         trans_for_each_update(trans, i) {
 811                 int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 812
 813                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 814                                                i->bkey_type, rw, &buf)))
 815                         return bch2_trans_commit_bkey_invalid(trans, i, &buf);
 816                 btree_insert_entry_checks(trans, i);
 817         }
 818 #endif
 819         printbuf_exit(&buf);
 820
 821         trans_for_each_update(trans, i) {
 822                 if (i->cached)
 823                         continue;
 824
 825                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 826                 u64s_delta -= i->old_btree_u64s;
 827
 828                 if (!same_leaf_as_next(trans, i)) {
 829                         if (u64s_delta <= 0) {
 830                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 831                                                         i->level, trans->flags);
 832                                 if (unlikely(ret))
 833                                         return ret;
 834                         }
 835
 836                         u64s_delta = 0;
 837                 }
 838         }
 839
 840         ret = bch2_journal_preres_get(&c->journal,
 841                         &trans->journal_preres, trans->journal_preres_u64s,
 842                         JOURNAL_RES_GET_NONBLOCK|
 843                         (trans->flags & JOURNAL_WATERMARK_MASK));
 844         if (unlikely(ret == -EAGAIN))
 845                 ret = bch2_trans_journal_preres_get_cold(trans,
 846                                                 trans->journal_preres_u64s, trace_ip);
 847         if (unlikely(ret))
 848                 return ret;
 849
 850         ret = trans_lock_write(trans);
 851         if (unlikely(ret))
 852                 return ret;
 853
 854         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 855
 856         if (!ret && unlikely(trans->journal_replay_not_finished))
 857                 bch2_drop_overwrites_from_journal(trans);
 858
 859         trans_for_each_update(trans, i)
 860                 if (!same_leaf_as_prev(trans, i))
 861                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 862                                                         insert_l(i)->b);
 863
 864         if (!ret && trans->journal_pin)
 865                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 866                                      trans->journal_pin, NULL);
 867
 868         /*
 869          * Drop journal reservation after dropping write locks, since dropping
 870          * the journal reservation may kick off a journal write:
 871          */
 872         bch2_journal_res_put(&c->journal, &trans->journal_res);
 873
 874         if (unlikely(ret))
 875                 return ret;
 876
 877         bch2_trans_downgrade(trans);
 878
 879         return 0;
 880 }
 881
 882 static int journal_reclaim_wait_done(struct bch_fs *c)
 883 {
 884         int ret = bch2_journal_error(&c->journal) ?:
 885                 !bch2_btree_key_cache_must_wait(c);
 886
 887         if (!ret)
 888                 journal_reclaim_kick(&c->journal);
 889         return ret;
 890 }
 891
 892 static noinline
 893 int bch2_trans_commit_error(struct btree_trans *trans,
 894                             struct btree_insert_entry *i,
 895                             int ret, unsigned long trace_ip)
 896 {
 897         struct bch_fs *c = trans->c;
 898
 899         switch (ret) {
 900         case -BCH_ERR_btree_insert_btree_node_full:
 901                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 902                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 903                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 904                 break;
 905         case -BCH_ERR_btree_insert_need_mark_replicas:
 906                 bch2_trans_unlock(trans);
 907
 908                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 909                 if (ret)
 910                         break;
 911
 912                 ret = bch2_trans_relock(trans);
 913                 if (ret)
 914                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 915                 break;
 916         case -BCH_ERR_btree_insert_need_journal_res:
 917                 bch2_trans_unlock(trans);
 918
 919                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 920                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 921                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 922                         break;
 923                 }
 924
 925                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 926                 if (ret)
 927                         break;
 928
 929                 ret = bch2_trans_relock(trans);
 930                 if (ret)
 931                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 932                 break;
 933         case -BCH_ERR_btree_insert_need_journal_reclaim:
 934                 bch2_trans_unlock(trans);
 935
 936                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 937
 938                 wait_event_freezable(c->journal.reclaim_wait,
 939                                      (ret = journal_reclaim_wait_done(c)));
 940                 if (ret < 0)
 941                         break;
 942
 943                 ret = bch2_trans_relock(trans);
 944                 if (ret)
 945                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 946                 break;
 947         default:
 948                 BUG_ON(ret >= 0);
 949                 break;
 950         }
 951
 952         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 953
 954         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 955                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 956                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 957                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 958
 959         return ret;
 960 }
 961
 962 static noinline int
 963 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 964 {
 965         struct bch_fs *c = trans->c;
 966         int ret;
 967
 968         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 969             test_bit(BCH_FS_STARTED, &c->flags))
 970                 return -EROFS;
 971
 972         bch2_trans_unlock(trans);
 973
 974         ret =   bch2_fs_read_write_early(c) ?:
 975                 bch2_trans_relock(trans);
 976         if (ret)
 977                 return ret;
 978
 979         percpu_ref_get(&c->writes);
 980         return 0;
 981 }
 982
 983 /*
 984  * This is for updates done in the early part of fsck - btree_gc - before we've
 985  * gone RW. we only add the new key to the list of keys for journal replay to
 986  * do.
 987  */
 988 static noinline int
 989 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 990 {
 991         struct bch_fs *c = trans->c;
 992         struct btree_insert_entry *i;
 993         int ret = 0;
 994
 995         trans_for_each_update(trans, i) {
 996                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
 997                 if (ret)
 998                         break;
 999         }
1000
1001         return ret;
1002 }
1003
1004 int __bch2_trans_commit(struct btree_trans *trans)
1005 {
1006         struct bch_fs *c = trans->c;
1007         struct btree_insert_entry *i = NULL;
1008         unsigned u64s;
1009         int ret = 0;
1010
1011         if (!trans->nr_updates &&
1012             !trans->extra_journal_entries.nr)
1013                 goto out_reset;
1014
1015         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1016                 lockdep_assert_held(&c->gc_lock);
1017
1018         ret = bch2_trans_commit_run_triggers(trans);
1019         if (ret)
1020                 goto out_reset;
1021
1022         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1023                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1024                 goto out_reset;
1025         }
1026
1027         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1028             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1029                 ret = bch2_trans_commit_get_rw_cold(trans);
1030                 if (ret)
1031                         goto out_reset;
1032         }
1033
1034         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1035
1036         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1037
1038         trans->journal_u64s             = trans->extra_journal_entries.nr;
1039         trans->journal_preres_u64s      = 0;
1040
1041         /* For journalling transaction name: */
1042         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1043
1044         trans_for_each_update(trans, i) {
1045                 EBUG_ON(!i->path->should_be_locked);
1046
1047                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1048                 if (unlikely(ret))
1049                         goto out;
1050
1051                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1052
1053                 if (i->key_cache_already_flushed)
1054                         continue;
1055
1056                 /* we're going to journal the key being updated: */
1057                 u64s = jset_u64s(i->k->k.u64s);
1058                 if (i->cached &&
1059                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1060                         trans->journal_preres_u64s += u64s;
1061                 trans->journal_u64s += u64s;
1062
1063                 /* and we're also going to log the overwrite: */
1064                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1065         }
1066
1067         if (trans->extra_journal_res) {
1068                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1069                                 trans->extra_journal_res,
1070                                 (trans->flags & BTREE_INSERT_NOFAIL)
1071                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1072                 if (ret)
1073                         goto err;
1074         }
1075 retry:
1076         EBUG_ON(trans->restarted);
1077         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1078
1079         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1080
1081         /* make sure we didn't drop or screw up locks: */
1082         bch2_trans_verify_locks(trans);
1083
1084         if (ret)
1085                 goto err;
1086
1087         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1088 out:
1089         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1090
1091         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1092                 percpu_ref_put(&c->writes);
1093 out_reset:
1094         bch2_trans_reset_updates(trans);
1095
1096         if (trans->fs_usage_deltas) {
1097                 trans->fs_usage_deltas->used = 0;
1098                 memset((void *) trans->fs_usage_deltas +
1099                        offsetof(struct replicas_delta_list, memset_start), 0,
1100                        (void *) &trans->fs_usage_deltas->memset_end -
1101                        (void *) &trans->fs_usage_deltas->memset_start);
1102         }
1103
1104         return ret;
1105 err:
1106         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1107         if (ret)
1108                 goto out;
1109
1110         goto retry;
1111 }
1112
1113 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1114                                           enum btree_id id,
1115                                           struct bpos pos)
1116 {
1117         struct bch_fs *c = trans->c;
1118         struct btree_iter iter;
1119         struct bkey_s_c k;
1120         int ret;
1121
1122         bch2_trans_iter_init(trans, &iter, id, pos,
1123                              BTREE_ITER_NOT_EXTENTS|
1124                              BTREE_ITER_ALL_SNAPSHOTS);
1125         while (1) {
1126                 k = bch2_btree_iter_prev(&iter);
1127                 ret = bkey_err(k);
1128                 if (ret)
1129                         break;
1130
1131                 if (!k.k)
1132                         break;
1133
1134                 if (!bkey_eq(pos, k.k->p))
1135                         break;
1136
1137                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1138                         ret = 1;
1139                         break;
1140                 }
1141         }
1142         bch2_trans_iter_exit(trans, &iter);
1143
1144         return ret;
1145 }
1146
1147 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1148                                           enum btree_id id,
1149                                           struct bpos pos)
1150 {
1151         if (!btree_type_has_snapshots(id) ||
1152             pos.snapshot == U32_MAX ||
1153             !snapshot_t(trans->c, pos.snapshot)->children[0])
1154                 return 0;
1155
1156         return __check_pos_snapshot_overwritten(trans, id, pos);
1157 }
1158
1159 int bch2_trans_update_extent(struct btree_trans *trans,
1160                              struct btree_iter *orig_iter,
1161                              struct bkey_i *insert,
1162                              enum btree_update_flags flags)
1163 {
1164         struct bch_fs *c = trans->c;
1165         struct btree_iter iter, update_iter;
1166         struct bpos start = bkey_start_pos(&insert->k);
1167         struct bkey_i *update;
1168         struct bkey_s_c k;
1169         enum btree_id btree_id = orig_iter->btree_id;
1170         int ret = 0, compressed_sectors;
1171
1172         bch2_trans_iter_init(trans, &iter, btree_id, start,
1173                              BTREE_ITER_INTENT|
1174                              BTREE_ITER_WITH_UPDATES|
1175                              BTREE_ITER_NOT_EXTENTS);
1176         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1177         if ((ret = bkey_err(k)))
1178                 goto err;
1179         if (!k.k)
1180                 goto out;
1181
1182         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1183                 /*
1184                  * We can't merge extents if they belong to interior snapshot
1185                  * tree nodes, and there's a snapshot in which one extent is
1186                  * visible and the other is not - i.e. if visibility is
1187                  * different.
1188                  *
1189                  * Instead of checking if visibilitiy of the two extents is
1190                  * different, for now we just check if either has been
1191                  * overwritten:
1192                  */
1193                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1194                 if (ret < 0)
1195                         goto err;
1196                 if (ret)
1197                         goto nomerge1;
1198
1199                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1200                 if (ret < 0)
1201                         goto err;
1202                 if (ret)
1203                         goto nomerge1;
1204
1205                 update = bch2_bkey_make_mut(trans, k);
1206                 if ((ret = PTR_ERR_OR_ZERO(update)))
1207                         goto err;
1208
1209                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1210                         ret = bch2_btree_delete_at(trans, &iter, flags);
1211                         if (ret)
1212                                 goto err;
1213
1214                         insert = update;
1215                         goto next;
1216                 }
1217         }
1218 nomerge1:
1219         ret = 0;
1220         if (bkey_eq(k.k->p, start))
1221                 goto next;
1222
1223         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1224                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1225                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1226
1227                 /*
1228                  * If we're going to be splitting a compressed extent, note it
1229                  * so that __bch2_trans_commit() can increase our disk
1230                  * reservation:
1231                  */
1232                 if (((front_split && back_split) ||
1233                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1234                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1235                         trans->extra_journal_res += compressed_sectors;
1236
1237                 if (front_split) {
1238                         update = bch2_bkey_make_mut(trans, k);
1239                         if ((ret = PTR_ERR_OR_ZERO(update)))
1240                                 goto err;
1241
1242                         bch2_cut_back(start, update);
1243
1244                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1245                                              BTREE_ITER_NOT_EXTENTS|
1246                                              BTREE_ITER_ALL_SNAPSHOTS|
1247                                              BTREE_ITER_INTENT);
1248                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1249                                 bch2_trans_update(trans, &update_iter, update,
1250                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1251                                                   flags);
1252                         bch2_trans_iter_exit(trans, &update_iter);
1253
1254                         if (ret)
1255                                 goto err;
1256                 }
1257
1258                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1259                     (front_split || back_split)) {
1260                         update = bch2_bkey_make_mut(trans, k);
1261                         if ((ret = PTR_ERR_OR_ZERO(update)))
1262                                 goto err;
1263
1264                         bch2_cut_front(start, update);
1265                         bch2_cut_back(insert->k.p, update);
1266
1267                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1268                                              BTREE_ITER_NOT_EXTENTS|
1269                                              BTREE_ITER_ALL_SNAPSHOTS|
1270                                              BTREE_ITER_INTENT);
1271                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1272                                 bch2_trans_update(trans, &update_iter, update,
1273                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1274                                                   flags);
1275                         bch2_trans_iter_exit(trans, &update_iter);
1276                         if (ret)
1277                                 goto err;
1278                 }
1279
1280                 if (bkey_le(k.k->p, insert->k.p)) {
1281                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1282                         if ((ret = PTR_ERR_OR_ZERO(update)))
1283                                 goto err;
1284
1285                         bkey_init(&update->k);
1286                         update->k.p = k.k->p;
1287
1288                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1289                                 update->k.p.snapshot = insert->k.p.snapshot;
1290                                 update->k.type = KEY_TYPE_whiteout;
1291                         }
1292
1293                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1294                                              BTREE_ITER_NOT_EXTENTS|
1295                                              BTREE_ITER_INTENT);
1296                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1297                                 bch2_trans_update(trans, &update_iter, update,
1298                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1299                                                   flags);
1300                         bch2_trans_iter_exit(trans, &update_iter);
1301
1302                         if (ret)
1303                                 goto err;
1304                 }
1305
1306                 if (back_split) {
1307                         update = bch2_bkey_make_mut(trans, k);
1308                         if ((ret = PTR_ERR_OR_ZERO(update)))
1309                                 goto err;
1310
1311                         bch2_cut_front(insert->k.p, update);
1312
1313                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1314                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1315                                                   flags);
1316                         if (ret)
1317                                 goto err;
1318                         goto out;
1319                 }
1320 next:
1321                 bch2_btree_iter_advance(&iter);
1322                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1323                 if ((ret = bkey_err(k)))
1324                         goto err;
1325                 if (!k.k)
1326                         goto out;
1327         }
1328
1329         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1330                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1331                 if (ret < 0)
1332                         goto err;
1333                 if (ret)
1334                         goto nomerge2;
1335
1336                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1337                 if (ret < 0)
1338                         goto err;
1339                 if (ret)
1340                         goto nomerge2;
1341
1342                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1343         }
1344 nomerge2:
1345         ret = 0;
1346 out:
1347         if (!bkey_deleted(&insert->k)) {
1348                 /*
1349                  * Rewinding iterators is expensive: get a new one and the one
1350                  * that points to the start of insert will be cloned from:
1351                  */
1352                 bch2_trans_iter_exit(trans, &iter);
1353                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1354                                      BTREE_ITER_NOT_EXTENTS|
1355                                      BTREE_ITER_INTENT);
1356                 ret   = bch2_btree_iter_traverse(&iter) ?:
1357                         bch2_trans_update(trans, &iter, insert, flags);
1358         }
1359 err:
1360         bch2_trans_iter_exit(trans, &iter);
1361
1362         return ret;
1363 }
1364
1365 /*
1366  * When deleting, check if we need to emit a whiteout (because we're overwriting
1367  * something in an ancestor snapshot)
1368  */
1369 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1370                                       enum btree_id btree_id, struct bpos pos)
1371 {
1372         struct btree_iter iter;
1373         struct bkey_s_c k;
1374         u32 snapshot = pos.snapshot;
1375         int ret;
1376
1377         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1378                 return 0;
1379
1380         pos.snapshot++;
1381
1382         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1383                            BTREE_ITER_ALL_SNAPSHOTS|
1384                            BTREE_ITER_NOPRESERVE, k, ret) {
1385                 if (!bkey_eq(k.k->p, pos))
1386                         break;
1387
1388                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1389                                               k.k->p.snapshot)) {
1390                         ret = !bkey_whiteout(k.k);
1391                         break;
1392                 }
1393         }
1394         bch2_trans_iter_exit(trans, &iter);
1395
1396         return ret;
1397 }
1398
1399 static int __must_check
1400 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1401                                 struct bkey_i *k, enum btree_update_flags flags,
1402                                 unsigned long ip);
1403
1404 static noinline int flush_new_cached_update(struct btree_trans *trans,
1405                                             struct btree_path *path,
1406                                             struct btree_insert_entry *i,
1407                                             enum btree_update_flags flags,
1408                                             unsigned long ip)
1409 {
1410         struct btree_path *btree_path;
1411         int ret;
1412
1413         i->key_cache_already_flushed = true;
1414         i->flags |= BTREE_TRIGGER_NORUN;
1415
1416         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1417                                    BTREE_ITER_INTENT, _THIS_IP_);
1418
1419         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1420         if (ret)
1421                 goto err;
1422
1423         btree_path_set_should_be_locked(btree_path);
1424         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1425 err:
1426         bch2_path_put(trans, btree_path, true);
1427         return ret;
1428 }
1429
1430 static int __must_check
1431 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1432                                 struct bkey_i *k, enum btree_update_flags flags,
1433                                 unsigned long ip)
1434 {
1435         struct bch_fs *c = trans->c;
1436         struct btree_insert_entry *i, n;
1437         int cmp;
1438
1439         EBUG_ON(!path->should_be_locked);
1440         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1441         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1442
1443         n = (struct btree_insert_entry) {
1444                 .flags          = flags,
1445                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1446                 .btree_id       = path->btree_id,
1447                 .level          = path->level,
1448                 .cached         = path->cached,
1449                 .path           = path,
1450                 .k              = k,
1451                 .ip_allocated   = ip,
1452         };
1453
1454 #ifdef CONFIG_BCACHEFS_DEBUG
1455         trans_for_each_update(trans, i)
1456                 BUG_ON(i != trans->updates &&
1457                        btree_insert_entry_cmp(i - 1, i) >= 0);
1458 #endif
1459
1460         /*
1461          * Pending updates are kept sorted: first, find position of new update,
1462          * then delete/trim any updates the new update overwrites:
1463          */
1464         trans_for_each_update(trans, i) {
1465                 cmp = btree_insert_entry_cmp(&n, i);
1466                 if (cmp <= 0)
1467                         break;
1468         }
1469
1470         if (!cmp && i < trans->updates + trans->nr_updates) {
1471                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1472
1473                 bch2_path_put(trans, i->path, true);
1474                 i->flags        = n.flags;
1475                 i->cached       = n.cached;
1476                 i->k            = n.k;
1477                 i->path         = n.path;
1478                 i->ip_allocated = n.ip_allocated;
1479         } else {
1480                 array_insert_item(trans->updates, trans->nr_updates,
1481                                   i - trans->updates, n);
1482
1483                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1484                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1485
1486                 if (unlikely(trans->journal_replay_not_finished)) {
1487                         struct bkey_i *j_k =
1488                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1489
1490                         if (j_k) {
1491                                 i->old_k = j_k->k;
1492                                 i->old_v = &j_k->v;
1493                         }
1494                 }
1495         }
1496
1497         __btree_path_get(i->path, true);
1498
1499         /*
1500          * If a key is present in the key cache, it must also exist in the
1501          * btree - this is necessary for cache coherency. When iterating over
1502          * a btree that's cached in the key cache, the btree iter code checks
1503          * the key cache - but the key has to exist in the btree for that to
1504          * work:
1505          */
1506         if (unlikely(path->cached && bkey_deleted(&i->old_k)))
1507                 return flush_new_cached_update(trans, path, i, flags, ip);
1508
1509         return 0;
1510 }
1511
1512 static inline int __must_check
1513 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1514                           struct bkey_i *k, enum btree_update_flags flags)
1515 {
1516         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1517 }
1518
1519 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1520                                    struct bkey_i *k, enum btree_update_flags flags)
1521 {
1522         struct btree_path *path = iter->update_path ?: iter->path;
1523         struct bkey_cached *ck;
1524         int ret;
1525
1526         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1527                 return bch2_trans_update_extent(trans, iter, k, flags);
1528
1529         if (bkey_deleted(&k->k) &&
1530             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1531             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1532                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1533                 if (unlikely(ret < 0))
1534                         return ret;
1535
1536                 if (ret)
1537                         k->k.type = KEY_TYPE_whiteout;
1538         }
1539
1540         /*
1541          * Ensure that updates to cached btrees go to the key cache:
1542          */
1543         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1544             !path->cached &&
1545             !path->level &&
1546             btree_id_cached(trans->c, path->btree_id)) {
1547                 if (!iter->key_cache_path ||
1548                     !iter->key_cache_path->should_be_locked ||
1549                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1550                         if (!iter->key_cache_path)
1551                                 iter->key_cache_path =
1552                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1553                                                       BTREE_ITER_INTENT|
1554                                                       BTREE_ITER_CACHED, _THIS_IP_);
1555
1556                         iter->key_cache_path =
1557                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1558                                                         iter->flags & BTREE_ITER_INTENT,
1559                                                         _THIS_IP_);
1560
1561                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1562                                                        BTREE_ITER_CACHED);
1563                         if (unlikely(ret))
1564                                 return ret;
1565
1566                         ck = (void *) iter->key_cache_path->l[0].b;
1567
1568                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1569                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1570                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1571                         }
1572
1573                         btree_path_set_should_be_locked(iter->key_cache_path);
1574                 }
1575
1576                 path = iter->key_cache_path;
1577         }
1578
1579         return bch2_trans_update_by_path(trans, path, k, flags);
1580 }
1581
1582 void bch2_trans_commit_hook(struct btree_trans *trans,
1583                             struct btree_trans_commit_hook *h)
1584 {
1585         h->next = trans->hooks;
1586         trans->hooks = h;
1587 }
1588
1589 int __bch2_btree_insert(struct btree_trans *trans,
1590                         enum btree_id id, struct bkey_i *k)
1591 {
1592         struct btree_iter iter;
1593         int ret;
1594
1595         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1596                              BTREE_ITER_INTENT);
1597         ret   = bch2_btree_iter_traverse(&iter) ?:
1598                 bch2_trans_update(trans, &iter, k, 0);
1599         bch2_trans_iter_exit(trans, &iter);
1600         return ret;
1601 }
1602
1603 /**
1604  * bch2_btree_insert - insert keys into the extent btree
1605  * @c:                  pointer to struct bch_fs
1606  * @id:                 btree to insert into
1607  * @insert_keys:        list of keys to insert
1608  * @hook:               insert callback
1609  */
1610 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1611                       struct bkey_i *k,
1612                       struct disk_reservation *disk_res,
1613                       u64 *journal_seq, int flags)
1614 {
1615         return bch2_trans_do(c, disk_res, journal_seq, flags,
1616                              __bch2_btree_insert(&trans, id, k));
1617 }
1618
1619 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1620                                 unsigned len, unsigned update_flags)
1621 {
1622         struct bkey_i *k;
1623
1624         k = bch2_trans_kmalloc(trans, sizeof(*k));
1625         if (IS_ERR(k))
1626                 return PTR_ERR(k);
1627
1628         bkey_init(&k->k);
1629         k->k.p = iter->pos;
1630         bch2_key_resize(&k->k, len);
1631         return bch2_trans_update(trans, iter, k, update_flags);
1632 }
1633
1634 int bch2_btree_delete_at(struct btree_trans *trans,
1635                          struct btree_iter *iter, unsigned update_flags)
1636 {
1637         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1638 }
1639
1640 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1641                                   struct bpos start, struct bpos end,
1642                                   unsigned update_flags,
1643                                   u64 *journal_seq)
1644 {
1645         u32 restart_count = trans->restart_count;
1646         struct btree_iter iter;
1647         struct bkey_s_c k;
1648         int ret = 0;
1649
1650         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1651         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1652                 struct disk_reservation disk_res =
1653                         bch2_disk_reservation_init(trans->c, 0);
1654                 struct bkey_i delete;
1655
1656                 ret = bkey_err(k);
1657                 if (ret)
1658                         goto err;
1659
1660                 bkey_init(&delete.k);
1661
1662                 /*
1663                  * This could probably be more efficient for extents:
1664                  */
1665
1666                 /*
1667                  * For extents, iter.pos won't necessarily be the same as
1668                  * bkey_start_pos(k.k) (for non extents they always will be the
1669                  * same). It's important that we delete starting from iter.pos
1670                  * because the range we want to delete could start in the middle
1671                  * of k.
1672                  *
1673                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1674                  * bkey_start_pos(k.k)).
1675                  */
1676                 delete.k.p = iter.pos;
1677
1678                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1679                         unsigned max_sectors =
1680                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1681
1682                         /* create the biggest key we can */
1683                         bch2_key_resize(&delete.k, max_sectors);
1684                         bch2_cut_back(end, &delete);
1685
1686                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1687                         if (ret)
1688                                 goto err;
1689                 }
1690
1691                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1692                         bch2_trans_commit(trans, &disk_res, journal_seq,
1693                                           BTREE_INSERT_NOFAIL);
1694                 bch2_disk_reservation_put(trans->c, &disk_res);
1695 err:
1696                 /*
1697                  * the bch2_trans_begin() call is in a weird place because we
1698                  * need to call it after every transaction commit, to avoid path
1699                  * overflow, but don't want to call it if the delete operation
1700                  * is a no-op and we have no work to do:
1701                  */
1702                 bch2_trans_begin(trans);
1703
1704                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1705                         ret = 0;
1706                 if (ret)
1707                         break;
1708         }
1709         bch2_trans_iter_exit(trans, &iter);
1710
1711         if (!ret && trans_was_restarted(trans, restart_count))
1712                 ret = -BCH_ERR_transaction_restart_nested;
1713         return ret;
1714 }
1715
1716 /*
1717  * bch_btree_delete_range - delete everything within a given range
1718  *
1719  * Range is a half open interval - [start, end)
1720  */
1721 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1722                             struct bpos start, struct bpos end,
1723                             unsigned update_flags,
1724                             u64 *journal_seq)
1725 {
1726         int ret = bch2_trans_run(c,
1727                         bch2_btree_delete_range_trans(&trans, id, start, end,
1728                                                       update_flags, journal_seq));
1729         if (ret == -BCH_ERR_transaction_restart_nested)
1730                 ret = 0;
1731         return ret;
1732 }
1733
1734 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1735 {
1736         unsigned len = strlen(msg);
1737         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1738         struct jset_entry_log *l;
1739         int ret;
1740
1741         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1742         if (ret)
1743                 return ret;
1744
1745         l = (void *) &darray_top(trans->extra_journal_entries);
1746         l->entry.u64s           = cpu_to_le16(u64s);
1747         l->entry.btree_id       = 0;
1748         l->entry.level          = 1;
1749         l->entry.type           = BCH_JSET_ENTRY_log;
1750         l->entry.pad[0]         = 0;
1751         l->entry.pad[1]         = 0;
1752         l->entry.pad[2]         = 0;
1753         memcpy(l->d, msg, len);
1754         while (len & 7)
1755                 l->d[len++] = '\0';
1756
1757         trans->extra_journal_entries.nr += jset_u64s(u64s);
1758         return 0;
1759 }