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 ? 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 enum btree_insert_ret
 347 btree_key_can_insert(struct btree_trans *trans,
 348                      struct btree *b,
 349                      unsigned u64s)
 350 {
 351         struct bch_fs *c = trans->c;
 352
 353         if (!bch2_btree_node_insert_fits(c, b, u64s))
 354                 return BTREE_INSERT_BTREE_NODE_FULL;
 355
 356         return BTREE_INSERT_OK;
 357 }
 358
 359 static enum btree_insert_ret
 360 btree_key_can_insert_cached(struct btree_trans *trans,
 361                             struct btree_path *path,
 362                             unsigned u64s)
 363 {
 364         struct bch_fs *c = trans->c;
 365         struct bkey_cached *ck = (void *) path->l[0].b;
 366         struct btree_insert_entry *i;
 367         unsigned new_u64s;
 368         struct bkey_i *new_k;
 369
 370         EBUG_ON(path->level);
 371
 372         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 373             bch2_btree_key_cache_must_wait(c) &&
 374             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 375                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 376
 377         /*
 378          * bch2_varint_decode can read past the end of the buffer by at most 7
 379          * bytes (it won't be used):
 380          */
 381         u64s += 1;
 382
 383         if (u64s <= ck->u64s)
 384                 return BTREE_INSERT_OK;
 385
 386         new_u64s        = roundup_pow_of_two(u64s);
 387         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 388         if (!new_k) {
 389                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 390                         bch2_btree_ids[path->btree_id], new_u64s);
 391                 return -ENOMEM;
 392         }
 393
 394         trans_for_each_update(trans, i)
 395                 if (i->old_v == &ck->k->v)
 396                         i->old_v = &new_k->v;
 397
 398         ck->u64s        = new_u64s;
 399         ck->k           = new_k;
 400         return 0;
 401 }
 402
 403 /* Triggers: */
 404
 405 static int run_one_mem_trigger(struct btree_trans *trans,
 406                                struct btree_insert_entry *i,
 407                                unsigned flags)
 408 {
 409         struct bkey_s_c old = { &i->old_k, i->old_v };
 410         struct bkey_i *new = i->k;
 411         int ret;
 412
 413         verify_update_old_key(trans, i);
 414
 415         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 416                 return 0;
 417
 418         if (!btree_node_type_needs_gc(i->btree_id))
 419                 return 0;
 420
 421         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 422             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 423             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 424                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 425                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 426         } else {
 427                 struct bkey             _deleted = KEY(0, 0, 0);
 428                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 429
 430                 _deleted.p = i->path->pos;
 431
 432                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 433                                 BTREE_TRIGGER_INSERT|flags) ?:
 434                         bch2_mark_key(trans, old, deleted,
 435                                 BTREE_TRIGGER_OVERWRITE|flags);
 436         }
 437
 438         return ret;
 439 }
 440
 441 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 442                                  bool overwrite)
 443 {
 444         /*
 445          * Transactional triggers create new btree_insert_entries, so we can't
 446          * pass them a pointer to a btree_insert_entry, that memory is going to
 447          * move:
 448          */
 449         struct bkey old_k = i->old_k;
 450         struct bkey_s_c old = { &old_k, i->old_v };
 451
 452         verify_update_old_key(trans, i);
 453
 454         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 455             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 456                 return 0;
 457
 458         if (!i->insert_trigger_run &&
 459             !i->overwrite_trigger_run &&
 460             bch2_bkey_ops[old.k->type].trans_trigger ==
 461             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 462             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 463                 i->overwrite_trigger_run = true;
 464                 i->insert_trigger_run = true;
 465                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 466                                            BTREE_TRIGGER_INSERT|
 467                                            BTREE_TRIGGER_OVERWRITE|
 468                                            i->flags) ?: 1;
 469         } else if (overwrite && !i->overwrite_trigger_run) {
 470                 i->overwrite_trigger_run = true;
 471                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 472         } else if (!overwrite && !i->insert_trigger_run) {
 473                 i->insert_trigger_run = true;
 474                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 475         } else {
 476                 return 0;
 477         }
 478 }
 479
 480 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 481                               struct btree_insert_entry *btree_id_start)
 482 {
 483         struct btree_insert_entry *i;
 484         bool trans_trigger_run;
 485         int ret, overwrite;
 486
 487         for (overwrite = 1; overwrite >= 0; --overwrite) {
 488
 489                 /*
 490                  * Running triggers will append more updates to the list of updates as
 491                  * we're walking it:
 492                  */
 493                 do {
 494                         trans_trigger_run = false;
 495
 496                         for (i = btree_id_start;
 497                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 498                              i++) {
 499                                 if (i->btree_id != btree_id)
 500                                         continue;
 501
 502                                 ret = run_one_trans_trigger(trans, i, overwrite);
 503                                 if (ret < 0)
 504                                         return ret;
 505                                 if (ret)
 506                                         trans_trigger_run = true;
 507                         }
 508                 } while (trans_trigger_run);
 509         }
 510
 511         return 0;
 512 }
 513
 514 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 515 {
 516         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 517         unsigned btree_id = 0;
 518         int ret = 0;
 519
 520         /*
 521          *
 522          * For a given btree, this algorithm runs insert triggers before
 523          * overwrite triggers: this is so that when extents are being moved
 524          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 525          * they are re-added.
 526          */
 527         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 528                 if (btree_id == BTREE_ID_alloc)
 529                         continue;
 530
 531                 while (btree_id_start < trans->updates + trans->nr_updates &&
 532                        btree_id_start->btree_id < btree_id)
 533                         btree_id_start++;
 534
 535                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 536                 if (ret)
 537                         return ret;
 538         }
 539
 540         trans_for_each_update(trans, i) {
 541                 if (i->btree_id > BTREE_ID_alloc)
 542                         break;
 543                 if (i->btree_id == BTREE_ID_alloc) {
 544                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 545                         if (ret)
 546                                 return ret;
 547                         break;
 548                 }
 549         }
 550
 551 #ifdef CONFIG_BCACHEFS_DEBUG
 552         trans_for_each_update(trans, i)
 553                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 554                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 555                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 556 #endif
 557         return 0;
 558 }
 559
 560 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 561 {
 562         struct bch_fs *c = trans->c;
 563         struct btree_insert_entry *i;
 564         int ret = 0;
 565
 566         trans_for_each_update(trans, i) {
 567                 /*
 568                  * XXX: synchronization of cached update triggers with gc
 569                  * XXX: synchronization of interior node updates with gc
 570                  */
 571                 BUG_ON(i->cached || i->level);
 572
 573                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 574                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 575                         if (ret)
 576                                 break;
 577                 }
 578         }
 579
 580         return ret;
 581 }
 582
 583 static inline int
 584 bch2_trans_commit_write_locked(struct btree_trans *trans,
 585                                struct btree_insert_entry **stopped_at,
 586                                unsigned long trace_ip)
 587 {
 588         struct bch_fs *c = trans->c;
 589         struct btree_insert_entry *i;
 590         struct btree_trans_commit_hook *h;
 591         unsigned u64s = 0;
 592         bool marking = false;
 593         int ret;
 594
 595         if (race_fault()) {
 596                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 597                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 598         }
 599
 600         /*
 601          * Check if the insert will fit in the leaf node with the write lock
 602          * held, otherwise another thread could write the node changing the
 603          * amount of space available:
 604          */
 605
 606         prefetch(&trans->c->journal.flags);
 607
 608         h = trans->hooks;
 609         while (h) {
 610                 ret = h->fn(trans, h);
 611                 if (ret)
 612                         return ret;
 613                 h = h->next;
 614         }
 615
 616         trans_for_each_update(trans, i) {
 617                 /* Multiple inserts might go to same leaf: */
 618                 if (!same_leaf_as_prev(trans, i))
 619                         u64s = 0;
 620
 621                 u64s += i->k->k.u64s;
 622                 ret = !i->cached
 623                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 624                         : btree_key_can_insert_cached(trans, i->path, u64s);
 625                 if (ret) {
 626                         *stopped_at = i;
 627                         return ret;
 628                 }
 629
 630                 if (btree_node_type_needs_gc(i->bkey_type))
 631                         marking = true;
 632         }
 633
 634         /*
 635          * Don't get journal reservation until after we know insert will
 636          * succeed:
 637          */
 638         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 639                 ret = bch2_trans_journal_res_get(trans,
 640                                 JOURNAL_RES_GET_NONBLOCK);
 641                 if (ret)
 642                         return ret;
 643
 644                 journal_transaction_name(trans);
 645         } else {
 646                 trans->journal_res.seq = c->journal.replay_journal_seq;
 647         }
 648
 649         if (unlikely(trans->extra_journal_entries.nr)) {
 650                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 651                                   trans->extra_journal_entries.data,
 652                                   trans->extra_journal_entries.nr);
 653
 654                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 655                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 656         }
 657
 658         /*
 659          * Not allowed to fail after we've gotten our journal reservation - we
 660          * have to use it:
 661          */
 662
 663         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 664                 if (bch2_journal_seq_verify)
 665                         trans_for_each_update(trans, i)
 666                                 i->k->k.version.lo = trans->journal_res.seq;
 667                 else if (bch2_inject_invalid_keys)
 668                         trans_for_each_update(trans, i)
 669                                 i->k->k.version = MAX_VERSION;
 670         }
 671
 672         if (trans->fs_usage_deltas &&
 673             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 674                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 675
 676         trans_for_each_update(trans, i)
 677                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 678                         ret = run_one_mem_trigger(trans, i, i->flags);
 679                         if (ret)
 680                                 return ret;
 681                 }
 682
 683         if (unlikely(c->gc_pos.phase)) {
 684                 ret = bch2_trans_commit_run_gc_triggers(trans);
 685                 if  (ret)
 686                         return ret;
 687         }
 688
 689         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 690                 trans_for_each_update(trans, i) {
 691                         struct journal *j = &c->journal;
 692                         struct jset_entry *entry;
 693
 694                         if (i->key_cache_already_flushed)
 695                                 continue;
 696
 697                         verify_update_old_key(trans, i);
 698
 699                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 700                                                BCH_JSET_ENTRY_overwrite,
 701                                                i->btree_id, i->level,
 702                                                i->old_k.u64s);
 703                         bkey_reassemble(&entry->start[0],
 704                                         (struct bkey_s_c) { &i->old_k, i->old_v });
 705
 706                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 707                                                BCH_JSET_ENTRY_btree_keys,
 708                                                i->btree_id, i->level,
 709                                                i->k->k.u64s);
 710                         bkey_copy(&entry->start[0], i->k);
 711                 }
 712
 713                 if (trans->journal_seq)
 714                         *trans->journal_seq = trans->journal_res.seq;
 715         }
 716
 717         trans_for_each_update(trans, i) {
 718                 i->k->k.needs_whiteout = false;
 719
 720                 if (!i->cached)
 721                         btree_insert_key_leaf(trans, i);
 722                 else if (!i->key_cache_already_flushed)
 723                         bch2_btree_insert_key_cached(trans, i->path, i->k);
 724                 else {
 725                         bch2_btree_key_cache_drop(trans, i->path);
 726                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 727                 }
 728         }
 729
 730         return ret;
 731 }
 732
 733 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 734 {
 735         while (--i >= trans->updates) {
 736                 if (same_leaf_as_prev(trans, i))
 737                         continue;
 738
 739                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 740         }
 741
 742         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 743         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 744 }
 745
 746 static inline int trans_lock_write(struct btree_trans *trans)
 747 {
 748         struct btree_insert_entry *i;
 749
 750         trans_for_each_update(trans, i) {
 751                 if (same_leaf_as_prev(trans, i))
 752                         continue;
 753
 754                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 755                         return trans_lock_write_fail(trans, i);
 756
 757                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 758         }
 759
 760         return 0;
 761 }
 762
 763 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 764 {
 765         struct btree_insert_entry *i;
 766
 767         trans_for_each_update(trans, i)
 768                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 769 }
 770
 771 #ifdef CONFIG_BCACHEFS_DEBUG
 772 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans,
 773                                                    struct btree_insert_entry *i,
 774                                                    struct printbuf *err)
 775 {
 776         struct bch_fs *c = trans->c;
 777         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 778
 779         printbuf_reset(err);
 780         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 781                    trans->fn, (void *) i->ip_allocated);
 782         prt_newline(err);
 783         printbuf_indent_add(err, 2);
 784
 785         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 786         prt_newline(err);
 787
 788         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 789                           i->bkey_type, rw, err);
 790         bch2_print_string_as_lines(KERN_ERR, err->buf);
 791
 792         bch2_inconsistent_error(c);
 793         bch2_dump_trans_updates(trans);
 794         printbuf_exit(err);
 795
 796         return -EINVAL;
 797 }
 798 #endif
 799
 800 /*
 801  * Get journal reservation, take write locks, and attempt to do btree update(s):
 802  */
 803 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 804                                        struct btree_insert_entry **stopped_at,
 805                                        unsigned long trace_ip)
 806 {
 807         struct bch_fs *c = trans->c;
 808         struct btree_insert_entry *i;
 809         struct printbuf buf = PRINTBUF;
 810         int ret, u64s_delta = 0;
 811
 812 #ifdef CONFIG_BCACHEFS_DEBUG
 813         trans_for_each_update(trans, i) {
 814                 int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 815
 816                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 817                                                i->bkey_type, rw, &buf)))
 818                         return bch2_trans_commit_bkey_invalid(trans, i, &buf);
 819                 btree_insert_entry_checks(trans, i);
 820         }
 821 #endif
 822         printbuf_exit(&buf);
 823
 824         trans_for_each_update(trans, i) {
 825                 if (i->cached)
 826                         continue;
 827
 828                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 829                 u64s_delta -= i->old_btree_u64s;
 830
 831                 if (!same_leaf_as_next(trans, i)) {
 832                         if (u64s_delta <= 0) {
 833                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 834                                                         i->level, trans->flags);
 835                                 if (unlikely(ret))
 836                                         return ret;
 837                         }
 838
 839                         u64s_delta = 0;
 840                 }
 841         }
 842
 843         ret = bch2_journal_preres_get(&c->journal,
 844                         &trans->journal_preres, trans->journal_preres_u64s,
 845                         JOURNAL_RES_GET_NONBLOCK|
 846                         (trans->flags & JOURNAL_WATERMARK_MASK));
 847         if (unlikely(ret == -EAGAIN))
 848                 ret = bch2_trans_journal_preres_get_cold(trans,
 849                                                 trans->journal_preres_u64s, trace_ip);
 850         if (unlikely(ret))
 851                 return ret;
 852
 853         ret = trans_lock_write(trans);
 854         if (unlikely(ret))
 855                 return ret;
 856
 857         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 858
 859         if (!ret && unlikely(trans->journal_replay_not_finished))
 860                 bch2_drop_overwrites_from_journal(trans);
 861
 862         trans_for_each_update(trans, i)
 863                 if (!same_leaf_as_prev(trans, i))
 864                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 865                                                         insert_l(i)->b);
 866
 867         if (!ret && trans->journal_pin)
 868                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 869                                      trans->journal_pin, NULL);
 870
 871         /*
 872          * Drop journal reservation after dropping write locks, since dropping
 873          * the journal reservation may kick off a journal write:
 874          */
 875         bch2_journal_res_put(&c->journal, &trans->journal_res);
 876
 877         if (unlikely(ret))
 878                 return ret;
 879
 880         bch2_trans_downgrade(trans);
 881
 882         return 0;
 883 }
 884
 885 static int journal_reclaim_wait_done(struct bch_fs *c)
 886 {
 887         int ret = bch2_journal_error(&c->journal) ?:
 888                 !bch2_btree_key_cache_must_wait(c);
 889
 890         if (!ret)
 891                 journal_reclaim_kick(&c->journal);
 892         return ret;
 893 }
 894
 895 static noinline
 896 int bch2_trans_commit_error(struct btree_trans *trans,
 897                             struct btree_insert_entry *i,
 898                             int ret, unsigned long trace_ip)
 899 {
 900         struct bch_fs *c = trans->c;
 901
 902         switch (ret) {
 903         case BTREE_INSERT_BTREE_NODE_FULL:
 904                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 905                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 906                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 907                 break;
 908         case BTREE_INSERT_NEED_MARK_REPLICAS:
 909                 bch2_trans_unlock(trans);
 910
 911                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 912                 if (ret)
 913                         break;
 914
 915                 ret = bch2_trans_relock(trans);
 916                 if (ret)
 917                         trace_and_count(c, trans_restart_mark_replicas, trans, trace_ip);
 918                 break;
 919         case BTREE_INSERT_NEED_JOURNAL_RES:
 920                 bch2_trans_unlock(trans);
 921
 922                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 923                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 924                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 925                         break;
 926                 }
 927
 928                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 929                 if (ret)
 930                         break;
 931
 932                 ret = bch2_trans_relock(trans);
 933                 if (ret)
 934                         trace_and_count(c, trans_restart_journal_res_get, trans, trace_ip);
 935                 break;
 936         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
 937                 bch2_trans_unlock(trans);
 938
 939                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 940
 941                 wait_event_freezable(c->journal.reclaim_wait,
 942                                      (ret = journal_reclaim_wait_done(c)));
 943                 if (ret < 0)
 944                         break;
 945
 946                 ret = bch2_trans_relock(trans);
 947                 if (ret)
 948                         trace_and_count(c, trans_restart_journal_reclaim, trans, trace_ip);
 949                 break;
 950         default:
 951                 BUG_ON(ret >= 0);
 952                 break;
 953         }
 954
 955         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
 956
 957         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
 958                                 !(trans->flags & BTREE_INSERT_NOWAIT) &&
 959                                 (trans->flags & BTREE_INSERT_NOFAIL), c,
 960                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
 961
 962         return ret;
 963 }
 964
 965 static noinline int
 966 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
 967 {
 968         struct bch_fs *c = trans->c;
 969         int ret;
 970
 971         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
 972             test_bit(BCH_FS_STARTED, &c->flags))
 973                 return -EROFS;
 974
 975         bch2_trans_unlock(trans);
 976
 977         ret =   bch2_fs_read_write_early(c) ?:
 978                 bch2_trans_relock(trans);
 979         if (ret)
 980                 return ret;
 981
 982         percpu_ref_get(&c->writes);
 983         return 0;
 984 }
 985
 986 /*
 987  * This is for updates done in the early part of fsck - btree_gc - before we've
 988  * gone RW. we only add the new key to the list of keys for journal replay to
 989  * do.
 990  */
 991 static noinline int
 992 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
 993 {
 994         struct bch_fs *c = trans->c;
 995         struct btree_insert_entry *i;
 996         int ret = 0;
 997
 998         trans_for_each_update(trans, i) {
 999                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1000                 if (ret)
1001                         break;
1002         }
1003
1004         return ret;
1005 }
1006
1007 int __bch2_trans_commit(struct btree_trans *trans)
1008 {
1009         struct bch_fs *c = trans->c;
1010         struct btree_insert_entry *i = NULL;
1011         unsigned u64s;
1012         int ret = 0;
1013
1014         if (!trans->nr_updates &&
1015             !trans->extra_journal_entries.nr)
1016                 goto out_reset;
1017
1018         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1019                 lockdep_assert_held(&c->gc_lock);
1020
1021         ret = bch2_trans_commit_run_triggers(trans);
1022         if (ret)
1023                 goto out_reset;
1024
1025         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1026                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1027                 goto out_reset;
1028         }
1029
1030         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1031             unlikely(!percpu_ref_tryget_live(&c->writes))) {
1032                 ret = bch2_trans_commit_get_rw_cold(trans);
1033                 if (ret)
1034                         goto out_reset;
1035         }
1036
1037         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1038
1039         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1040
1041         trans->journal_u64s             = trans->extra_journal_entries.nr;
1042         trans->journal_preres_u64s      = 0;
1043
1044         /* For journalling transaction name: */
1045         trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1046
1047         trans_for_each_update(trans, i) {
1048                 EBUG_ON(!i->path->should_be_locked);
1049
1050                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1051                 if (unlikely(ret))
1052                         goto out;
1053
1054                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1055
1056                 if (i->key_cache_already_flushed)
1057                         continue;
1058
1059                 /* we're going to journal the key being updated: */
1060                 u64s = jset_u64s(i->k->k.u64s);
1061                 if (i->cached &&
1062                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1063                         trans->journal_preres_u64s += u64s;
1064                 trans->journal_u64s += u64s;
1065
1066                 /* and we're also going to log the overwrite: */
1067                 trans->journal_u64s += jset_u64s(i->old_k.u64s);
1068         }
1069
1070         if (trans->extra_journal_res) {
1071                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1072                                 trans->extra_journal_res,
1073                                 (trans->flags & BTREE_INSERT_NOFAIL)
1074                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1075                 if (ret)
1076                         goto err;
1077         }
1078 retry:
1079         EBUG_ON(trans->restarted);
1080         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1081
1082         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1083
1084         /* make sure we didn't drop or screw up locks: */
1085         bch2_trans_verify_locks(trans);
1086
1087         if (ret)
1088                 goto err;
1089
1090         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1091 out:
1092         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1093
1094         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1095                 percpu_ref_put(&c->writes);
1096 out_reset:
1097         bch2_trans_reset_updates(trans);
1098
1099         if (trans->fs_usage_deltas) {
1100                 trans->fs_usage_deltas->used = 0;
1101                 memset((void *) trans->fs_usage_deltas +
1102                        offsetof(struct replicas_delta_list, memset_start), 0,
1103                        (void *) &trans->fs_usage_deltas->memset_end -
1104                        (void *) &trans->fs_usage_deltas->memset_start);
1105         }
1106
1107         return ret;
1108 err:
1109         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1110         if (ret)
1111                 goto out;
1112
1113         goto retry;
1114 }
1115
1116 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1117                                           enum btree_id id,
1118                                           struct bpos pos)
1119 {
1120         struct bch_fs *c = trans->c;
1121         struct btree_iter iter;
1122         struct bkey_s_c k;
1123         int ret;
1124
1125         bch2_trans_iter_init(trans, &iter, id, pos,
1126                              BTREE_ITER_NOT_EXTENTS|
1127                              BTREE_ITER_ALL_SNAPSHOTS);
1128         while (1) {
1129                 k = bch2_btree_iter_prev(&iter);
1130                 ret = bkey_err(k);
1131                 if (ret)
1132                         break;
1133
1134                 if (!k.k)
1135                         break;
1136
1137                 if (!bkey_eq(pos, k.k->p))
1138                         break;
1139
1140                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1141                         ret = 1;
1142                         break;
1143                 }
1144         }
1145         bch2_trans_iter_exit(trans, &iter);
1146
1147         return ret;
1148 }
1149
1150 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1151                                           enum btree_id id,
1152                                           struct bpos pos)
1153 {
1154         if (!btree_type_has_snapshots(id) ||
1155             pos.snapshot == U32_MAX ||
1156             !snapshot_t(trans->c, pos.snapshot)->children[0])
1157                 return 0;
1158
1159         return __check_pos_snapshot_overwritten(trans, id, pos);
1160 }
1161
1162 int bch2_trans_update_extent(struct btree_trans *trans,
1163                              struct btree_iter *orig_iter,
1164                              struct bkey_i *insert,
1165                              enum btree_update_flags flags)
1166 {
1167         struct bch_fs *c = trans->c;
1168         struct btree_iter iter, update_iter;
1169         struct bpos start = bkey_start_pos(&insert->k);
1170         struct bkey_i *update;
1171         struct bkey_s_c k;
1172         enum btree_id btree_id = orig_iter->btree_id;
1173         int ret = 0, compressed_sectors;
1174
1175         bch2_trans_iter_init(trans, &iter, btree_id, start,
1176                              BTREE_ITER_INTENT|
1177                              BTREE_ITER_WITH_UPDATES|
1178                              BTREE_ITER_NOT_EXTENTS);
1179         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1180         if ((ret = bkey_err(k)))
1181                 goto err;
1182         if (!k.k)
1183                 goto out;
1184
1185         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1186                 /*
1187                  * We can't merge extents if they belong to interior snapshot
1188                  * tree nodes, and there's a snapshot in which one extent is
1189                  * visible and the other is not - i.e. if visibility is
1190                  * different.
1191                  *
1192                  * Instead of checking if visibilitiy of the two extents is
1193                  * different, for now we just check if either has been
1194                  * overwritten:
1195                  */
1196                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1197                 if (ret < 0)
1198                         goto err;
1199                 if (ret)
1200                         goto nomerge1;
1201
1202                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1203                 if (ret < 0)
1204                         goto err;
1205                 if (ret)
1206                         goto nomerge1;
1207
1208                 update = bch2_bkey_make_mut(trans, k);
1209                 if ((ret = PTR_ERR_OR_ZERO(update)))
1210                         goto err;
1211
1212                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1213                         ret = bch2_btree_delete_at(trans, &iter, flags);
1214                         if (ret)
1215                                 goto err;
1216
1217                         insert = update;
1218                         goto next;
1219                 }
1220         }
1221 nomerge1:
1222         ret = 0;
1223         if (bkey_eq(k.k->p, start))
1224                 goto next;
1225
1226         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1227                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1228                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1229
1230                 /*
1231                  * If we're going to be splitting a compressed extent, note it
1232                  * so that __bch2_trans_commit() can increase our disk
1233                  * reservation:
1234                  */
1235                 if (((front_split && back_split) ||
1236                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1237                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1238                         trans->extra_journal_res += compressed_sectors;
1239
1240                 if (front_split) {
1241                         update = bch2_bkey_make_mut(trans, k);
1242                         if ((ret = PTR_ERR_OR_ZERO(update)))
1243                                 goto err;
1244
1245                         bch2_cut_back(start, update);
1246
1247                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1248                                              BTREE_ITER_NOT_EXTENTS|
1249                                              BTREE_ITER_ALL_SNAPSHOTS|
1250                                              BTREE_ITER_INTENT);
1251                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1252                                 bch2_trans_update(trans, &update_iter, update,
1253                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1254                                                   flags);
1255                         bch2_trans_iter_exit(trans, &update_iter);
1256
1257                         if (ret)
1258                                 goto err;
1259                 }
1260
1261                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1262                     (front_split || back_split)) {
1263                         update = bch2_bkey_make_mut(trans, k);
1264                         if ((ret = PTR_ERR_OR_ZERO(update)))
1265                                 goto err;
1266
1267                         bch2_cut_front(start, update);
1268                         bch2_cut_back(insert->k.p, update);
1269
1270                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1271                                              BTREE_ITER_NOT_EXTENTS|
1272                                              BTREE_ITER_ALL_SNAPSHOTS|
1273                                              BTREE_ITER_INTENT);
1274                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1275                                 bch2_trans_update(trans, &update_iter, update,
1276                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1277                                                   flags);
1278                         bch2_trans_iter_exit(trans, &update_iter);
1279                         if (ret)
1280                                 goto err;
1281                 }
1282
1283                 if (bkey_le(k.k->p, insert->k.p)) {
1284                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1285                         if ((ret = PTR_ERR_OR_ZERO(update)))
1286                                 goto err;
1287
1288                         bkey_init(&update->k);
1289                         update->k.p = k.k->p;
1290
1291                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1292                                 update->k.p.snapshot = insert->k.p.snapshot;
1293                                 update->k.type = KEY_TYPE_whiteout;
1294                         }
1295
1296                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1297                                              BTREE_ITER_NOT_EXTENTS|
1298                                              BTREE_ITER_INTENT);
1299                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1300                                 bch2_trans_update(trans, &update_iter, update,
1301                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1302                                                   flags);
1303                         bch2_trans_iter_exit(trans, &update_iter);
1304
1305                         if (ret)
1306                                 goto err;
1307                 }
1308
1309                 if (back_split) {
1310                         update = bch2_bkey_make_mut(trans, k);
1311                         if ((ret = PTR_ERR_OR_ZERO(update)))
1312                                 goto err;
1313
1314                         bch2_cut_front(insert->k.p, update);
1315
1316                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1317                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1318                                                   flags);
1319                         if (ret)
1320                                 goto err;
1321                         goto out;
1322                 }
1323 next:
1324                 bch2_btree_iter_advance(&iter);
1325                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1326                 if ((ret = bkey_err(k)))
1327                         goto err;
1328                 if (!k.k)
1329                         goto out;
1330         }
1331
1332         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1333                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1334                 if (ret < 0)
1335                         goto err;
1336                 if (ret)
1337                         goto nomerge2;
1338
1339                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1340                 if (ret < 0)
1341                         goto err;
1342                 if (ret)
1343                         goto nomerge2;
1344
1345                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1346         }
1347 nomerge2:
1348         ret = 0;
1349 out:
1350         if (!bkey_deleted(&insert->k)) {
1351                 /*
1352                  * Rewinding iterators is expensive: get a new one and the one
1353                  * that points to the start of insert will be cloned from:
1354                  */
1355                 bch2_trans_iter_exit(trans, &iter);
1356                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1357                                      BTREE_ITER_NOT_EXTENTS|
1358                                      BTREE_ITER_INTENT);
1359                 ret   = bch2_btree_iter_traverse(&iter) ?:
1360                         bch2_trans_update(trans, &iter, insert, flags);
1361         }
1362 err:
1363         bch2_trans_iter_exit(trans, &iter);
1364
1365         return ret;
1366 }
1367
1368 /*
1369  * When deleting, check if we need to emit a whiteout (because we're overwriting
1370  * something in an ancestor snapshot)
1371  */
1372 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1373                                       enum btree_id btree_id, struct bpos pos)
1374 {
1375         struct btree_iter iter;
1376         struct bkey_s_c k;
1377         u32 snapshot = pos.snapshot;
1378         int ret;
1379
1380         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1381                 return 0;
1382
1383         pos.snapshot++;
1384
1385         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1386                            BTREE_ITER_ALL_SNAPSHOTS|
1387                            BTREE_ITER_NOPRESERVE, k, ret) {
1388                 if (!bkey_eq(k.k->p, pos))
1389                         break;
1390
1391                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1392                                               k.k->p.snapshot)) {
1393                         ret = !bkey_whiteout(k.k);
1394                         break;
1395                 }
1396         }
1397         bch2_trans_iter_exit(trans, &iter);
1398
1399         return ret;
1400 }
1401
1402 static int __must_check
1403 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1404                                 struct bkey_i *k, enum btree_update_flags flags,
1405                                 unsigned long ip);
1406
1407 static noinline int flush_new_cached_update(struct btree_trans *trans,
1408                                             struct btree_path *path,
1409                                             struct btree_insert_entry *i,
1410                                             enum btree_update_flags flags,
1411                                             unsigned long ip)
1412 {
1413         struct btree_path *btree_path;
1414         int ret;
1415
1416         i->key_cache_already_flushed = true;
1417         i->flags |= BTREE_TRIGGER_NORUN;
1418
1419         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1420                                    BTREE_ITER_INTENT, _THIS_IP_);
1421
1422         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1423         if (ret)
1424                 goto err;
1425
1426         btree_path_set_should_be_locked(btree_path);
1427         ret = bch2_trans_update_by_path_trace(trans, btree_path, i->k, flags, ip);
1428 err:
1429         bch2_path_put(trans, btree_path, true);
1430         return ret;
1431 }
1432
1433 static int __must_check
1434 bch2_trans_update_by_path_trace(struct btree_trans *trans, struct btree_path *path,
1435                                 struct bkey_i *k, enum btree_update_flags flags,
1436                                 unsigned long ip)
1437 {
1438         struct bch_fs *c = trans->c;
1439         struct btree_insert_entry *i, n;
1440         int cmp;
1441
1442         EBUG_ON(!path->should_be_locked);
1443         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1444         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1445
1446         n = (struct btree_insert_entry) {
1447                 .flags          = flags,
1448                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1449                 .btree_id       = path->btree_id,
1450                 .level          = path->level,
1451                 .cached         = path->cached,
1452                 .path           = path,
1453                 .k              = k,
1454                 .ip_allocated   = ip,
1455         };
1456
1457 #ifdef CONFIG_BCACHEFS_DEBUG
1458         trans_for_each_update(trans, i)
1459                 BUG_ON(i != trans->updates &&
1460                        btree_insert_entry_cmp(i - 1, i) >= 0);
1461 #endif
1462
1463         /*
1464          * Pending updates are kept sorted: first, find position of new update,
1465          * then delete/trim any updates the new update overwrites:
1466          */
1467         trans_for_each_update(trans, i) {
1468                 cmp = btree_insert_entry_cmp(&n, i);
1469                 if (cmp <= 0)
1470                         break;
1471         }
1472
1473         if (!cmp && i < trans->updates + trans->nr_updates) {
1474                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1475
1476                 bch2_path_put(trans, i->path, true);
1477                 i->flags        = n.flags;
1478                 i->cached       = n.cached;
1479                 i->k            = n.k;
1480                 i->path         = n.path;
1481                 i->ip_allocated = n.ip_allocated;
1482         } else {
1483                 array_insert_item(trans->updates, trans->nr_updates,
1484                                   i - trans->updates, n);
1485
1486                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1487                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1488
1489                 if (unlikely(trans->journal_replay_not_finished)) {
1490                         struct bkey_i *j_k =
1491                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1492
1493                         if (j_k) {
1494                                 i->old_k = j_k->k;
1495                                 i->old_v = &j_k->v;
1496                         }
1497                 }
1498         }
1499
1500         __btree_path_get(i->path, true);
1501
1502         /*
1503          * If a key is present in the key cache, it must also exist in the
1504          * btree - this is necessary for cache coherency. When iterating over
1505          * a btree that's cached in the key cache, the btree iter code checks
1506          * the key cache - but the key has to exist in the btree for that to
1507          * work:
1508          */
1509         if (unlikely(path->cached && bkey_deleted(&i->old_k)))
1510                 return flush_new_cached_update(trans, path, i, flags, ip);
1511
1512         return 0;
1513 }
1514
1515 static inline int __must_check
1516 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1517                           struct bkey_i *k, enum btree_update_flags flags)
1518 {
1519         return bch2_trans_update_by_path_trace(trans, path, k, flags, _RET_IP_);
1520 }
1521
1522 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1523                                    struct bkey_i *k, enum btree_update_flags flags)
1524 {
1525         struct btree_path *path = iter->update_path ?: iter->path;
1526         struct bkey_cached *ck;
1527         int ret;
1528
1529         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1530                 return bch2_trans_update_extent(trans, iter, k, flags);
1531
1532         if (bkey_deleted(&k->k) &&
1533             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1534             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1535                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1536                 if (unlikely(ret < 0))
1537                         return ret;
1538
1539                 if (ret)
1540                         k->k.type = KEY_TYPE_whiteout;
1541         }
1542
1543         /*
1544          * Ensure that updates to cached btrees go to the key cache:
1545          */
1546         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1547             !path->cached &&
1548             !path->level &&
1549             btree_id_cached(trans->c, path->btree_id)) {
1550                 if (!iter->key_cache_path ||
1551                     !iter->key_cache_path->should_be_locked ||
1552                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1553                         if (!iter->key_cache_path)
1554                                 iter->key_cache_path =
1555                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1556                                                       BTREE_ITER_INTENT|
1557                                                       BTREE_ITER_CACHED, _THIS_IP_);
1558
1559                         iter->key_cache_path =
1560                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1561                                                         iter->flags & BTREE_ITER_INTENT,
1562                                                         _THIS_IP_);
1563
1564                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1565                                                        BTREE_ITER_CACHED);
1566                         if (unlikely(ret))
1567                                 return ret;
1568
1569                         ck = (void *) iter->key_cache_path->l[0].b;
1570
1571                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1572                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1573                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1574                         }
1575
1576                         btree_path_set_should_be_locked(iter->key_cache_path);
1577                 }
1578
1579                 path = iter->key_cache_path;
1580         }
1581
1582         return bch2_trans_update_by_path(trans, path, k, flags);
1583 }
1584
1585 void bch2_trans_commit_hook(struct btree_trans *trans,
1586                             struct btree_trans_commit_hook *h)
1587 {
1588         h->next = trans->hooks;
1589         trans->hooks = h;
1590 }
1591
1592 int __bch2_btree_insert(struct btree_trans *trans,
1593                         enum btree_id id, struct bkey_i *k)
1594 {
1595         struct btree_iter iter;
1596         int ret;
1597
1598         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1599                              BTREE_ITER_INTENT);
1600         ret   = bch2_btree_iter_traverse(&iter) ?:
1601                 bch2_trans_update(trans, &iter, k, 0);
1602         bch2_trans_iter_exit(trans, &iter);
1603         return ret;
1604 }
1605
1606 /**
1607  * bch2_btree_insert - insert keys into the extent btree
1608  * @c:                  pointer to struct bch_fs
1609  * @id:                 btree to insert into
1610  * @insert_keys:        list of keys to insert
1611  * @hook:               insert callback
1612  */
1613 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1614                       struct bkey_i *k,
1615                       struct disk_reservation *disk_res,
1616                       u64 *journal_seq, int flags)
1617 {
1618         return bch2_trans_do(c, disk_res, journal_seq, flags,
1619                              __bch2_btree_insert(&trans, id, k));
1620 }
1621
1622 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1623                                 unsigned len, unsigned update_flags)
1624 {
1625         struct bkey_i *k;
1626
1627         k = bch2_trans_kmalloc(trans, sizeof(*k));
1628         if (IS_ERR(k))
1629                 return PTR_ERR(k);
1630
1631         bkey_init(&k->k);
1632         k->k.p = iter->pos;
1633         bch2_key_resize(&k->k, len);
1634         return bch2_trans_update(trans, iter, k, update_flags);
1635 }
1636
1637 int bch2_btree_delete_at(struct btree_trans *trans,
1638                          struct btree_iter *iter, unsigned update_flags)
1639 {
1640         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1641 }
1642
1643 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1644                                   struct bpos start, struct bpos end,
1645                                   unsigned update_flags,
1646                                   u64 *journal_seq)
1647 {
1648         u32 restart_count = trans->restart_count;
1649         struct btree_iter iter;
1650         struct bkey_s_c k;
1651         int ret = 0;
1652
1653         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1654         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1655                 struct disk_reservation disk_res =
1656                         bch2_disk_reservation_init(trans->c, 0);
1657                 struct bkey_i delete;
1658
1659                 ret = bkey_err(k);
1660                 if (ret)
1661                         goto err;
1662
1663                 bkey_init(&delete.k);
1664
1665                 /*
1666                  * This could probably be more efficient for extents:
1667                  */
1668
1669                 /*
1670                  * For extents, iter.pos won't necessarily be the same as
1671                  * bkey_start_pos(k.k) (for non extents they always will be the
1672                  * same). It's important that we delete starting from iter.pos
1673                  * because the range we want to delete could start in the middle
1674                  * of k.
1675                  *
1676                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1677                  * bkey_start_pos(k.k)).
1678                  */
1679                 delete.k.p = iter.pos;
1680
1681                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1682                         unsigned max_sectors =
1683                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1684
1685                         /* create the biggest key we can */
1686                         bch2_key_resize(&delete.k, max_sectors);
1687                         bch2_cut_back(end, &delete);
1688
1689                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1690                         if (ret)
1691                                 goto err;
1692                 }
1693
1694                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1695                         bch2_trans_commit(trans, &disk_res, journal_seq,
1696                                           BTREE_INSERT_NOFAIL);
1697                 bch2_disk_reservation_put(trans->c, &disk_res);
1698 err:
1699                 /*
1700                  * the bch2_trans_begin() call is in a weird place because we
1701                  * need to call it after every transaction commit, to avoid path
1702                  * overflow, but don't want to call it if the delete operation
1703                  * is a no-op and we have no work to do:
1704                  */
1705                 bch2_trans_begin(trans);
1706
1707                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1708                         ret = 0;
1709                 if (ret)
1710                         break;
1711         }
1712         bch2_trans_iter_exit(trans, &iter);
1713
1714         if (!ret && trans_was_restarted(trans, restart_count))
1715                 ret = -BCH_ERR_transaction_restart_nested;
1716         return ret;
1717 }
1718
1719 /*
1720  * bch_btree_delete_range - delete everything within a given range
1721  *
1722  * Range is a half open interval - [start, end)
1723  */
1724 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1725                             struct bpos start, struct bpos end,
1726                             unsigned update_flags,
1727                             u64 *journal_seq)
1728 {
1729         int ret = bch2_trans_run(c,
1730                         bch2_btree_delete_range_trans(&trans, id, start, end,
1731                                                       update_flags, journal_seq));
1732         if (ret == -BCH_ERR_transaction_restart_nested)
1733                 ret = 0;
1734         return ret;
1735 }
1736
1737 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1738 {
1739         unsigned len = strlen(msg);
1740         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1741         struct jset_entry_log *l;
1742         int ret;
1743
1744         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1745         if (ret)
1746                 return ret;
1747
1748         l = (void *) &darray_top(trans->extra_journal_entries);
1749         l->entry.u64s           = cpu_to_le16(u64s);
1750         l->entry.btree_id       = 0;
1751         l->entry.level          = 1;
1752         l->entry.type           = BCH_JSET_ENTRY_log;
1753         l->entry.pad[0]         = 0;
1754         l->entry.pad[1]         = 0;
1755         l->entry.pad[2]         = 0;
1756         memcpy(l->d, msg, len);
1757         while (len & 7)
1758                 l->d[len++] = '\0';
1759
1760         trans->extra_journal_entries.nr += jset_u64s(u64s);
1761         return 0;
1762 }