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 "btree_write_buffer.h"
  12 #include "buckets.h"
  13 #include "debug.h"
  14 #include "errcode.h"
  15 #include "error.h"
  16 #include "extent_update.h"
  17 #include "journal.h"
  18 #include "journal_reclaim.h"
  19 #include "keylist.h"
  20 #include "recovery.h"
  21 #include "subvolume.h"
  22 #include "replicas.h"
  23 #include "trace.h"
  24
  25 #include <linux/prefetch.h>
  26 #include <linux/sort.h>
  27
  28 /*
  29  * bch2_btree_path_peek_slot() for a cached iterator might return a key in a
  30  * different snapshot:
  31  */
  32 static struct bkey_s_c bch2_btree_path_peek_slot_exact(struct btree_path *path, struct bkey *u)
  33 {
  34         struct bkey_s_c k = bch2_btree_path_peek_slot(path, u);
  35
  36         if (k.k && bpos_eq(path->pos, k.k->p))
  37                 return k;
  38
  39         bkey_init(u);
  40         u->p = path->pos;
  41         return (struct bkey_s_c) { u, NULL };
  42 }
  43
  44 static void verify_update_old_key(struct btree_trans *trans, struct btree_insert_entry *i)
  45 {
  46 #ifdef CONFIG_BCACHEFS_DEBUG
  47         struct bch_fs *c = trans->c;
  48         struct bkey u;
  49         struct bkey_s_c k = bch2_btree_path_peek_slot_exact(i->path, &u);
  50
  51         if (unlikely(trans->journal_replay_not_finished)) {
  52                 struct bkey_i *j_k =
  53                         bch2_journal_keys_peek_slot(c, i->btree_id, i->level, i->k->k.p);
  54
  55                 if (j_k)
  56                         k = bkey_i_to_s_c(j_k);
  57         }
  58
  59         u = *k.k;
  60         u.needs_whiteout = i->old_k.needs_whiteout;
  61
  62         BUG_ON(memcmp(&i->old_k, &u, sizeof(struct bkey)));
  63         BUG_ON(i->old_v != k.v);
  64 #endif
  65 }
  66
  67 static int __must_check
  68 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  69                           struct bkey_i *, enum btree_update_flags,
  70                           unsigned long ip);
  71
  72 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  73                                          const struct btree_insert_entry *r)
  74 {
  75         return   cmp_int(l->btree_id,   r->btree_id) ?:
  76                  cmp_int(l->cached,     r->cached) ?:
  77                  -cmp_int(l->level,     r->level) ?:
  78                  bpos_cmp(l->k->k.p,    r->k->k.p);
  79 }
  80
  81 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  82 {
  83         return i->path->l + i->level;
  84 }
  85
  86 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  87                                      struct btree_insert_entry *i)
  88 {
  89         return i != trans->updates &&
  90                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  91 }
  92
  93 static inline bool same_leaf_as_next(struct btree_trans *trans,
  94                                      struct btree_insert_entry *i)
  95 {
  96         return i + 1 < trans->updates + trans->nr_updates &&
  97                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  98 }
  99
 100 inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
 101                                            struct btree_path *path,
 102                                            struct btree *b)
 103 {
 104         struct bch_fs *c = trans->c;
 105
 106         if (unlikely(btree_node_just_written(b)) &&
 107             bch2_btree_post_write_cleanup(c, b))
 108                 bch2_trans_node_reinit_iter(trans, b);
 109
 110         /*
 111          * If the last bset has been written, or if it's gotten too big - start
 112          * a new bset to insert into:
 113          */
 114         if (want_new_bset(c, b))
 115                 bch2_btree_init_next(trans, b);
 116 }
 117
 118 /* Inserting into a given leaf node (last stage of insert): */
 119
 120 /* Handle overwrites and do insert, for non extents: */
 121 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
 122                                 struct btree_path *path,
 123                                 struct btree *b,
 124                                 struct btree_node_iter *node_iter,
 125                                 struct bkey_i *insert)
 126 {
 127         struct bkey_packed *k;
 128         unsigned clobber_u64s = 0, new_u64s = 0;
 129
 130         EBUG_ON(btree_node_just_written(b));
 131         EBUG_ON(bset_written(b, btree_bset_last(b)));
 132         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 133         EBUG_ON(bpos_lt(insert->k.p, b->data->min_key));
 134         EBUG_ON(bpos_gt(insert->k.p, b->data->max_key));
 135         EBUG_ON(insert->k.u64s >
 136                 bch_btree_keys_u64s_remaining(trans->c, b));
 137
 138         k = bch2_btree_node_iter_peek_all(node_iter, b);
 139         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 140                 k = NULL;
 141
 142         /* @k is the key being overwritten/deleted, if any: */
 143         EBUG_ON(k && bkey_deleted(k));
 144
 145         /* Deleting, but not found? nothing to do: */
 146         if (bkey_deleted(&insert->k) && !k)
 147                 return false;
 148
 149         if (bkey_deleted(&insert->k)) {
 150                 /* Deleting: */
 151                 btree_account_key_drop(b, k);
 152                 k->type = KEY_TYPE_deleted;
 153
 154                 if (k->needs_whiteout)
 155                         push_whiteout(trans->c, b, insert->k.p);
 156                 k->needs_whiteout = false;
 157
 158                 if (k >= btree_bset_last(b)->start) {
 159                         clobber_u64s = k->u64s;
 160                         bch2_bset_delete(b, k, clobber_u64s);
 161                         goto fix_iter;
 162                 } else {
 163                         bch2_btree_path_fix_key_modified(trans, b, k);
 164                 }
 165
 166                 return true;
 167         }
 168
 169         if (k) {
 170                 /* Overwriting: */
 171                 btree_account_key_drop(b, k);
 172                 k->type = KEY_TYPE_deleted;
 173
 174                 insert->k.needs_whiteout = k->needs_whiteout;
 175                 k->needs_whiteout = false;
 176
 177                 if (k >= btree_bset_last(b)->start) {
 178                         clobber_u64s = k->u64s;
 179                         goto overwrite;
 180                 } else {
 181                         bch2_btree_path_fix_key_modified(trans, b, k);
 182                 }
 183         }
 184
 185         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 186 overwrite:
 187         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 188         new_u64s = k->u64s;
 189 fix_iter:
 190         if (clobber_u64s != new_u64s)
 191                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 192                                          clobber_u64s, new_u64s);
 193         return true;
 194 }
 195
 196 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 197                                unsigned i, u64 seq)
 198 {
 199         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 200         struct btree_write *w = container_of(pin, struct btree_write, journal);
 201         struct btree *b = container_of(w, struct btree, writes[i]);
 202         struct btree_trans trans;
 203         unsigned long old, new, v;
 204         unsigned idx = w - b->writes;
 205
 206         bch2_trans_init(&trans, c, 0, 0);
 207
 208         btree_node_lock_nopath_nofail(&trans, &b->c, SIX_LOCK_read);
 209         v = READ_ONCE(b->flags);
 210
 211         do {
 212                 old = new = v;
 213
 214                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 215                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 216                     w->journal.seq != seq)
 217                         break;
 218
 219                 new &= ~BTREE_WRITE_TYPE_MASK;
 220                 new |= BTREE_WRITE_journal_reclaim;
 221                 new |= 1 << BTREE_NODE_need_write;
 222         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 223
 224         btree_node_write_if_need(c, b, SIX_LOCK_read);
 225         six_unlock_read(&b->c.lock);
 226
 227         bch2_trans_exit(&trans);
 228         return 0;
 229 }
 230
 231 int bch2_btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 232 {
 233         return __btree_node_flush(j, pin, 0, seq);
 234 }
 235
 236 int bch2_btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 237 {
 238         return __btree_node_flush(j, pin, 1, seq);
 239 }
 240
 241 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 242                                        struct btree *b, u64 seq)
 243 {
 244         struct btree_write *w = btree_current_write(b);
 245
 246         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 247                              btree_node_write_idx(b) == 0
 248                              ? bch2_btree_node_flush0
 249                              : bch2_btree_node_flush1);
 250 }
 251
 252 /**
 253  * btree_insert_key - insert a key one key into a leaf node
 254  */
 255 inline void bch2_btree_insert_key_leaf(struct btree_trans *trans,
 256                                        struct btree_path *path,
 257                                        struct bkey_i *insert,
 258                                        u64 journal_seq)
 259 {
 260         struct bch_fs *c = trans->c;
 261         struct btree *b = path_l(path)->b;
 262         struct bset_tree *t = bset_tree_last(b);
 263         struct bset *i = bset(b, t);
 264         int old_u64s = bset_u64s(t);
 265         int old_live_u64s = b->nr.live_u64s;
 266         int live_u64s_added, u64s_added;
 267
 268         if (unlikely(!bch2_btree_bset_insert_key(trans, path, b,
 269                                         &path_l(path)->iter, insert)))
 270                 return;
 271
 272         i->journal_seq = cpu_to_le64(max(journal_seq, le64_to_cpu(i->journal_seq)));
 273
 274         bch2_btree_add_journal_pin(c, b, journal_seq);
 275
 276         if (unlikely(!btree_node_dirty(b))) {
 277                 EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
 278                 set_btree_node_dirty_acct(c, b);
 279         }
 280
 281         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 282         u64s_added = (int) bset_u64s(t) - old_u64s;
 283
 284         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 285                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 286         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 287                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 288
 289         if (u64s_added > live_u64s_added &&
 290             bch2_maybe_compact_whiteouts(c, b))
 291                 bch2_trans_node_reinit_iter(trans, b);
 292 }
 293
 294 /* Cached btree updates: */
 295
 296 /* Normal update interface: */
 297
 298 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 299                                              struct btree_insert_entry *i)
 300 {
 301         BUG_ON(!bpos_eq(i->k->k.p, i->path->pos));
 302         BUG_ON(i->cached        != i->path->cached);
 303         BUG_ON(i->level         != i->path->level);
 304         BUG_ON(i->btree_id      != i->path->btree_id);
 305         EBUG_ON(!i->level &&
 306                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 307                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 308                 i->k->k.p.snapshot &&
 309                 bch2_snapshot_is_internal_node(trans->c, i->k->k.p.snapshot));
 310 }
 311
 312 static noinline int
 313 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned flags,
 314                                    unsigned long trace_ip)
 315 {
 316         return drop_locks_do(trans,
 317                 bch2_journal_preres_get(&trans->c->journal,
 318                         &trans->journal_preres,
 319                         trans->journal_preres_u64s,
 320                         (flags & BCH_WATERMARK_MASK)));
 321 }
 322
 323 static __always_inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 324                                                       unsigned flags)
 325 {
 326         return bch2_journal_res_get(&trans->c->journal, &trans->journal_res,
 327                                     trans->journal_u64s, flags);
 328 }
 329
 330 #define JSET_ENTRY_LOG_U64s             4
 331
 332 static noinline 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, unsigned flags,
 358                                        struct btree_path *path, unsigned u64s)
 359 {
 360         struct bch_fs *c = trans->c;
 361         struct bkey_cached *ck = (void *) path->l[0].b;
 362         struct btree_insert_entry *i;
 363         unsigned new_u64s;
 364         struct bkey_i *new_k;
 365
 366         EBUG_ON(path->level);
 367
 368         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 369             bch2_btree_key_cache_must_wait(c) &&
 370             !(flags & BTREE_INSERT_JOURNAL_RECLAIM))
 371                 return -BCH_ERR_btree_insert_need_journal_reclaim;
 372
 373         /*
 374          * bch2_varint_decode can read past the end of the buffer by at most 7
 375          * bytes (it won't be used):
 376          */
 377         u64s += 1;
 378
 379         if (u64s <= ck->u64s)
 380                 return 0;
 381
 382         new_u64s        = roundup_pow_of_two(u64s);
 383         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 384         if (!new_k) {
 385                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 386                         bch2_btree_ids[path->btree_id], new_u64s);
 387                 return -BCH_ERR_ENOMEM_btree_key_cache_insert;
 388         }
 389
 390         trans_for_each_update(trans, i)
 391                 if (i->old_v == &ck->k->v)
 392                         i->old_v = &new_k->v;
 393
 394         ck->u64s        = new_u64s;
 395         ck->k           = new_k;
 396         return 0;
 397 }
 398
 399 /* Triggers: */
 400
 401 static int run_one_mem_trigger(struct btree_trans *trans,
 402                                struct btree_insert_entry *i,
 403                                unsigned flags)
 404 {
 405         struct bkey_s_c old = { &i->old_k, i->old_v };
 406         struct bkey_i *new = i->k;
 407         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 408         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 409         int ret;
 410
 411         verify_update_old_key(trans, i);
 412
 413         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 414                 return 0;
 415
 416         if (!btree_node_type_needs_gc(i->btree_id))
 417                 return 0;
 418
 419         if (old_ops->atomic_trigger == new_ops->atomic_trigger &&
 420             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 421                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 422                                 old, bkey_i_to_s_c(new),
 423                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 424         } else {
 425                 struct bkey             _deleted = KEY(0, 0, 0);
 426                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 427
 428                 _deleted.p = i->path->pos;
 429
 430                 ret   = bch2_mark_key(trans, i->btree_id, i->level,
 431                                 deleted, bkey_i_to_s_c(new),
 432                                 BTREE_TRIGGER_INSERT|flags) ?:
 433                         bch2_mark_key(trans, i->btree_id, i->level,
 434                                 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         const struct bkey_ops *old_ops = bch2_bkey_type_ops(old.k->type);
 452         const struct bkey_ops *new_ops = bch2_bkey_type_ops(i->k->k.type);
 453
 454         verify_update_old_key(trans, i);
 455
 456         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 457             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 458                 return 0;
 459
 460         if (!i->insert_trigger_run &&
 461             !i->overwrite_trigger_run &&
 462             old_ops->trans_trigger == new_ops->trans_trigger &&
 463             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 464                 i->overwrite_trigger_run = true;
 465                 i->insert_trigger_run = true;
 466                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 467                                            BTREE_TRIGGER_INSERT|
 468                                            BTREE_TRIGGER_OVERWRITE|
 469                                            i->flags) ?: 1;
 470         } else if (overwrite && !i->overwrite_trigger_run) {
 471                 i->overwrite_trigger_run = true;
 472                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 473         } else if (!overwrite && !i->insert_trigger_run) {
 474                 i->insert_trigger_run = true;
 475                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 476         } else {
 477                 return 0;
 478         }
 479 }
 480
 481 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 482                               struct btree_insert_entry *btree_id_start)
 483 {
 484         struct btree_insert_entry *i;
 485         bool trans_trigger_run;
 486         int ret, overwrite;
 487
 488         for (overwrite = 1; overwrite >= 0; --overwrite) {
 489
 490                 /*
 491                  * Running triggers will append more updates to the list of updates as
 492                  * we're walking it:
 493                  */
 494                 do {
 495                         trans_trigger_run = false;
 496
 497                         for (i = btree_id_start;
 498                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 499                              i++) {
 500                                 if (i->btree_id != btree_id)
 501                                         continue;
 502
 503                                 ret = run_one_trans_trigger(trans, i, overwrite);
 504                                 if (ret < 0)
 505                                         return ret;
 506                                 if (ret)
 507                                         trans_trigger_run = true;
 508                         }
 509                 } while (trans_trigger_run);
 510         }
 511
 512         return 0;
 513 }
 514
 515 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 516 {
 517         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 518         unsigned btree_id = 0;
 519         int ret = 0;
 520
 521         /*
 522          *
 523          * For a given btree, this algorithm runs insert triggers before
 524          * overwrite triggers: this is so that when extents are being moved
 525          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 526          * they are re-added.
 527          */
 528         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 529                 if (btree_id == BTREE_ID_alloc)
 530                         continue;
 531
 532                 while (btree_id_start < trans->updates + trans->nr_updates &&
 533                        btree_id_start->btree_id < btree_id)
 534                         btree_id_start++;
 535
 536                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 537                 if (ret)
 538                         return ret;
 539         }
 540
 541         trans_for_each_update(trans, i) {
 542                 if (i->btree_id > BTREE_ID_alloc)
 543                         break;
 544                 if (i->btree_id == BTREE_ID_alloc) {
 545                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 546                         if (ret)
 547                                 return ret;
 548                         break;
 549                 }
 550         }
 551
 552 #ifdef CONFIG_BCACHEFS_DEBUG
 553         trans_for_each_update(trans, i)
 554                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 555                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 556                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 557 #endif
 558         return 0;
 559 }
 560
 561 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 562 {
 563         struct bch_fs *c = trans->c;
 564         struct btree_insert_entry *i;
 565         int ret = 0;
 566
 567         trans_for_each_update(trans, i) {
 568                 /*
 569                  * XXX: synchronization of cached update triggers with gc
 570                  * XXX: synchronization of interior node updates with gc
 571                  */
 572                 BUG_ON(i->cached || i->level);
 573
 574                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 575                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 576                         if (ret)
 577                                 break;
 578                 }
 579         }
 580
 581         return ret;
 582 }
 583
 584 static inline int
 585 bch2_trans_commit_write_locked(struct btree_trans *trans, unsigned flags,
 586                                struct btree_insert_entry **stopped_at,
 587                                unsigned long trace_ip)
 588 {
 589         struct bch_fs *c = trans->c;
 590         struct btree_insert_entry *i;
 591         struct btree_write_buffered_key *wb;
 592         struct btree_trans_commit_hook *h;
 593         unsigned u64s = 0;
 594         bool marking = false;
 595         int ret;
 596
 597         if (race_fault()) {
 598                 trace_and_count(c, trans_restart_fault_inject, trans, trace_ip);
 599                 return btree_trans_restart_nounlock(trans, BCH_ERR_transaction_restart_fault_inject);
 600         }
 601
 602         /*
 603          * Check if the insert will fit in the leaf node with the write lock
 604          * held, otherwise another thread could write the node changing the
 605          * amount of space available:
 606          */
 607
 608         prefetch(&trans->c->journal.flags);
 609
 610         trans_for_each_update(trans, i) {
 611                 /* Multiple inserts might go to same leaf: */
 612                 if (!same_leaf_as_prev(trans, i))
 613                         u64s = 0;
 614
 615                 u64s += i->k->k.u64s;
 616                 ret = !i->cached
 617                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 618                         : btree_key_can_insert_cached(trans, flags, i->path, u64s);
 619                 if (ret) {
 620                         *stopped_at = i;
 621                         return ret;
 622                 }
 623
 624                 if (btree_node_type_needs_gc(i->bkey_type))
 625                         marking = true;
 626         }
 627
 628         if (trans->nr_wb_updates &&
 629             trans->nr_wb_updates + c->btree_write_buffer.state.nr > c->btree_write_buffer.size)
 630                 return -BCH_ERR_btree_insert_need_flush_buffer;
 631
 632         /*
 633          * Don't get journal reservation until after we know insert will
 634          * succeed:
 635          */
 636         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 637                 ret = bch2_trans_journal_res_get(trans,
 638                                 (flags & BCH_WATERMARK_MASK)|
 639                                 JOURNAL_RES_GET_NONBLOCK);
 640                 if (ret)
 641                         return ret;
 642
 643                 if (unlikely(trans->journal_transaction_names))
 644                         journal_transaction_name(trans);
 645         } else {
 646                 trans->journal_res.seq = c->journal.replay_journal_seq;
 647         }
 648
 649         /*
 650          * Not allowed to fail after we've gotten our journal reservation - we
 651          * have to use it:
 652          */
 653
 654         if (IS_ENABLED(CONFIG_BCACHEFS_DEBUG) &&
 655             !(flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 656                 if (bch2_journal_seq_verify)
 657                         trans_for_each_update(trans, i)
 658                                 i->k->k.version.lo = trans->journal_res.seq;
 659                 else if (bch2_inject_invalid_keys)
 660                         trans_for_each_update(trans, i)
 661                                 i->k->k.version = MAX_VERSION;
 662         }
 663
 664         if (trans->fs_usage_deltas &&
 665             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 666                 return -BCH_ERR_btree_insert_need_mark_replicas;
 667
 668         if (trans->nr_wb_updates) {
 669                 EBUG_ON(flags & BTREE_INSERT_JOURNAL_REPLAY);
 670
 671                 ret = bch2_btree_insert_keys_write_buffer(trans);
 672                 if (ret)
 673                         goto revert_fs_usage;
 674         }
 675
 676         h = trans->hooks;
 677         while (h) {
 678                 ret = h->fn(trans, h);
 679                 if (ret)
 680                         goto revert_fs_usage;
 681                 h = h->next;
 682         }
 683
 684         trans_for_each_update(trans, i)
 685                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 686                         ret = run_one_mem_trigger(trans, i, i->flags);
 687                         if (ret)
 688                                 goto fatal_err;
 689                 }
 690
 691         if (unlikely(c->gc_pos.phase)) {
 692                 ret = bch2_trans_commit_run_gc_triggers(trans);
 693                 if  (ret)
 694                         goto fatal_err;
 695         }
 696
 697         if (unlikely(trans->extra_journal_entries.nr)) {
 698                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 699                                   trans->extra_journal_entries.data,
 700                                   trans->extra_journal_entries.nr);
 701
 702                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 703                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 704         }
 705
 706         if (likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 707                 struct journal *j = &c->journal;
 708                 struct jset_entry *entry;
 709
 710                 trans_for_each_update(trans, i) {
 711                         if (i->key_cache_already_flushed)
 712                                 continue;
 713
 714                         if (i->flags & BTREE_UPDATE_NOJOURNAL)
 715                                 continue;
 716
 717                         verify_update_old_key(trans, i);
 718
 719                         if (trans->journal_transaction_names) {
 720                                 entry = bch2_journal_add_entry(j, &trans->journal_res,
 721                                                        BCH_JSET_ENTRY_overwrite,
 722                                                        i->btree_id, i->level,
 723                                                        i->old_k.u64s);
 724                                 bkey_reassemble(&entry->start[0],
 725                                                 (struct bkey_s_c) { &i->old_k, i->old_v });
 726                         }
 727
 728                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 729                                                BCH_JSET_ENTRY_btree_keys,
 730                                                i->btree_id, i->level,
 731                                                i->k->k.u64s);
 732                         bkey_copy(&entry->start[0], i->k);
 733                 }
 734
 735                 trans_for_each_wb_update(trans, wb) {
 736                         entry = bch2_journal_add_entry(j, &trans->journal_res,
 737                                                BCH_JSET_ENTRY_btree_keys,
 738                                                wb->btree, 0,
 739                                                wb->k.k.u64s);
 740                         bkey_copy(&entry->start[0], &wb->k);
 741                 }
 742
 743                 if (trans->journal_seq)
 744                         *trans->journal_seq = trans->journal_res.seq;
 745         }
 746
 747         trans_for_each_update(trans, i) {
 748                 i->k->k.needs_whiteout = false;
 749
 750                 if (!i->cached) {
 751                         u64 seq = trans->journal_res.seq;
 752
 753                         if (i->flags & BTREE_UPDATE_PREJOURNAL)
 754                                 seq = i->seq;
 755
 756                         bch2_btree_insert_key_leaf(trans, i->path, i->k, seq);
 757                 } else if (!i->key_cache_already_flushed)
 758                         bch2_btree_insert_key_cached(trans, flags, i);
 759                 else {
 760                         bch2_btree_key_cache_drop(trans, i->path);
 761                         btree_path_set_dirty(i->path, BTREE_ITER_NEED_TRAVERSE);
 762                 }
 763         }
 764
 765         return 0;
 766 fatal_err:
 767         bch2_fatal_error(c);
 768 revert_fs_usage:
 769         if (trans->fs_usage_deltas)
 770                 bch2_trans_fs_usage_revert(trans, trans->fs_usage_deltas);
 771         return ret;
 772 }
 773
 774 static noinline int trans_lock_write_fail(struct btree_trans *trans, struct btree_insert_entry *i)
 775 {
 776         while (--i >= trans->updates) {
 777                 if (same_leaf_as_prev(trans, i))
 778                         continue;
 779
 780                 bch2_btree_node_unlock_write(trans, i->path, insert_l(i)->b);
 781         }
 782
 783         trace_and_count(trans->c, trans_restart_would_deadlock_write, trans);
 784         return btree_trans_restart(trans, BCH_ERR_transaction_restart_would_deadlock_write);
 785 }
 786
 787 static inline int trans_lock_write(struct btree_trans *trans)
 788 {
 789         struct btree_insert_entry *i;
 790
 791         trans_for_each_update(trans, i) {
 792                 if (same_leaf_as_prev(trans, i))
 793                         continue;
 794
 795                 if (bch2_btree_node_lock_write(trans, i->path, &insert_l(i)->b->c))
 796                         return trans_lock_write_fail(trans, i);
 797
 798                 if (!i->cached)
 799                         bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 800         }
 801
 802         return 0;
 803 }
 804
 805 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 806 {
 807         struct btree_insert_entry *i;
 808         struct btree_write_buffered_key *wb;
 809
 810         trans_for_each_update(trans, i)
 811                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 812
 813         trans_for_each_wb_update(trans, wb)
 814                 bch2_journal_key_overwritten(trans->c, wb->btree, 0, wb->k.k.p);
 815 }
 816
 817 #ifdef CONFIG_BCACHEFS_DEBUG
 818 static noinline int bch2_trans_commit_bkey_invalid(struct btree_trans *trans, unsigned flags,
 819                                                    struct btree_insert_entry *i,
 820                                                    struct printbuf *err)
 821 {
 822         struct bch_fs *c = trans->c;
 823         int rw = (flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 824
 825         printbuf_reset(err);
 826         prt_printf(err, "invalid bkey on insert from %s -> %ps",
 827                    trans->fn, (void *) i->ip_allocated);
 828         prt_newline(err);
 829         printbuf_indent_add(err, 2);
 830
 831         bch2_bkey_val_to_text(err, c, bkey_i_to_s_c(i->k));
 832         prt_newline(err);
 833
 834         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 835                           i->bkey_type, rw, err);
 836         bch2_print_string_as_lines(KERN_ERR, err->buf);
 837
 838         bch2_inconsistent_error(c);
 839         bch2_dump_trans_updates(trans);
 840         printbuf_exit(err);
 841
 842         return -EINVAL;
 843 }
 844 #endif
 845
 846 /*
 847  * Get journal reservation, take write locks, and attempt to do btree update(s):
 848  */
 849 static inline int do_bch2_trans_commit(struct btree_trans *trans, unsigned flags,
 850                                        struct btree_insert_entry **stopped_at,
 851                                        unsigned long trace_ip)
 852 {
 853         struct bch_fs *c = trans->c;
 854         struct btree_insert_entry *i;
 855         int ret, u64s_delta = 0;
 856
 857 #ifdef CONFIG_BCACHEFS_DEBUG
 858         struct printbuf buf = PRINTBUF;
 859
 860         trans_for_each_update(trans, i) {
 861                 enum bkey_invalid_flags invalid_flags = 0;
 862
 863                 if (!(flags & BTREE_INSERT_JOURNAL_REPLAY))
 864                         invalid_flags |= BKEY_INVALID_WRITE|BKEY_INVALID_COMMIT;
 865
 866                 if (unlikely(bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 867                                                i->bkey_type, invalid_flags, &buf)))
 868                         return bch2_trans_commit_bkey_invalid(trans, flags, i, &buf);
 869                 btree_insert_entry_checks(trans, i);
 870         }
 871         printbuf_exit(&buf);
 872 #endif
 873
 874         trans_for_each_update(trans, i) {
 875                 if (i->cached)
 876                         continue;
 877
 878                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 879                 u64s_delta -= i->old_btree_u64s;
 880
 881                 if (!same_leaf_as_next(trans, i)) {
 882                         if (u64s_delta <= 0) {
 883                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 884                                                         i->level, flags);
 885                                 if (unlikely(ret))
 886                                         return ret;
 887                         }
 888
 889                         u64s_delta = 0;
 890                 }
 891         }
 892
 893         ret = bch2_journal_preres_get(&c->journal,
 894                         &trans->journal_preres, trans->journal_preres_u64s,
 895                         (flags & BCH_WATERMARK_MASK)|JOURNAL_RES_GET_NONBLOCK);
 896         if (unlikely(ret == -BCH_ERR_journal_preres_get_blocked))
 897                 ret = bch2_trans_journal_preres_get_cold(trans, flags, trace_ip);
 898         if (unlikely(ret))
 899                 return ret;
 900
 901         ret = trans_lock_write(trans);
 902         if (unlikely(ret))
 903                 return ret;
 904
 905         ret = bch2_trans_commit_write_locked(trans, flags, stopped_at, trace_ip);
 906
 907         if (!ret && unlikely(trans->journal_replay_not_finished))
 908                 bch2_drop_overwrites_from_journal(trans);
 909
 910         trans_for_each_update(trans, i)
 911                 if (!same_leaf_as_prev(trans, i))
 912                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 913                                                         insert_l(i)->b);
 914
 915         if (!ret && trans->journal_pin)
 916                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 917                                      trans->journal_pin, NULL);
 918
 919         /*
 920          * Drop journal reservation after dropping write locks, since dropping
 921          * the journal reservation may kick off a journal write:
 922          */
 923         bch2_journal_res_put(&c->journal, &trans->journal_res);
 924
 925         if (unlikely(ret))
 926                 return ret;
 927
 928         bch2_trans_downgrade(trans);
 929
 930         return 0;
 931 }
 932
 933 static int journal_reclaim_wait_done(struct bch_fs *c)
 934 {
 935         int ret = bch2_journal_error(&c->journal) ?:
 936                 !bch2_btree_key_cache_must_wait(c);
 937
 938         if (!ret)
 939                 journal_reclaim_kick(&c->journal);
 940         return ret;
 941 }
 942
 943 static noinline
 944 int bch2_trans_commit_error(struct btree_trans *trans, unsigned flags,
 945                             struct btree_insert_entry *i,
 946                             int ret, unsigned long trace_ip)
 947 {
 948         struct bch_fs *c = trans->c;
 949
 950         switch (ret) {
 951         case -BCH_ERR_btree_insert_btree_node_full:
 952                 ret = bch2_btree_split_leaf(trans, i->path, flags);
 953                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 954                         trace_and_count(c, trans_restart_btree_node_split, trans, trace_ip, i->path);
 955                 break;
 956         case -BCH_ERR_btree_insert_need_mark_replicas:
 957                 ret = drop_locks_do(trans,
 958                         bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas));
 959                 break;
 960         case -BCH_ERR_journal_res_get_blocked:
 961                 /*
 962                  * XXX: this should probably be a separate BTREE_INSERT_NONBLOCK
 963                  * flag
 964                  */
 965                 if ((flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 966                     (flags & BCH_WATERMARK_MASK) != BCH_WATERMARK_reclaim) {
 967                         ret = -BCH_ERR_journal_reclaim_would_deadlock;
 968                         break;
 969                 }
 970
 971                 ret = drop_locks_do(trans,
 972                         bch2_trans_journal_res_get(trans,
 973                                         (flags & BCH_WATERMARK_MASK)|
 974                                         JOURNAL_RES_GET_CHECK));
 975                 break;
 976         case -BCH_ERR_btree_insert_need_journal_reclaim:
 977                 bch2_trans_unlock(trans);
 978
 979                 trace_and_count(c, trans_blocked_journal_reclaim, trans, trace_ip);
 980
 981                 wait_event_freezable(c->journal.reclaim_wait,
 982                                      (ret = journal_reclaim_wait_done(c)));
 983                 if (ret < 0)
 984                         break;
 985
 986                 ret = bch2_trans_relock(trans);
 987                 break;
 988         case -BCH_ERR_btree_insert_need_flush_buffer: {
 989                 struct btree_write_buffer *wb = &c->btree_write_buffer;
 990
 991                 ret = 0;
 992
 993                 if (wb->state.nr > wb->size * 3 / 4) {
 994                         bch2_trans_unlock(trans);
 995                         mutex_lock(&wb->flush_lock);
 996
 997                         if (wb->state.nr > wb->size * 3 / 4) {
 998                                 bch2_trans_begin(trans);
 999                                 ret = __bch2_btree_write_buffer_flush(trans,
1000                                                 flags|BTREE_INSERT_NOCHECK_RW, true);
1001                                 if (!ret) {
1002                                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1003                                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1004                                 }
1005                         } else {
1006                                 mutex_unlock(&wb->flush_lock);
1007                                 ret = bch2_trans_relock(trans);
1008                         }
1009                 }
1010                 break;
1011         }
1012         default:
1013                 BUG_ON(ret >= 0);
1014                 break;
1015         }
1016
1017         BUG_ON(bch2_err_matches(ret, BCH_ERR_transaction_restart) != !!trans->restarted);
1018
1019         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOSPC) &&
1020                                 !(flags & BTREE_INSERT_NOWAIT) &&
1021                                 (flags & BTREE_INSERT_NOFAIL), c,
1022                 "%s: incorrectly got %s\n", __func__, bch2_err_str(ret));
1023
1024         return ret;
1025 }
1026
1027 static noinline int
1028 bch2_trans_commit_get_rw_cold(struct btree_trans *trans, unsigned flags)
1029 {
1030         struct bch_fs *c = trans->c;
1031         int ret;
1032
1033         if (likely(!(flags & BTREE_INSERT_LAZY_RW)) ||
1034             test_bit(BCH_FS_STARTED, &c->flags))
1035                 return -BCH_ERR_erofs_trans_commit;
1036
1037         ret = drop_locks_do(trans, bch2_fs_read_write_early(c));
1038         if (ret)
1039                 return ret;
1040
1041         bch2_write_ref_get(c, BCH_WRITE_REF_trans);
1042         return 0;
1043 }
1044
1045 /*
1046  * This is for updates done in the early part of fsck - btree_gc - before we've
1047  * gone RW. we only add the new key to the list of keys for journal replay to
1048  * do.
1049  */
1050 static noinline int
1051 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1052 {
1053         struct bch_fs *c = trans->c;
1054         struct btree_insert_entry *i;
1055         int ret = 0;
1056
1057         trans_for_each_update(trans, i) {
1058                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1059                 if (ret)
1060                         break;
1061         }
1062
1063         return ret;
1064 }
1065
1066 int __bch2_trans_commit(struct btree_trans *trans, unsigned flags)
1067 {
1068         struct bch_fs *c = trans->c;
1069         struct btree_insert_entry *i = NULL;
1070         struct btree_write_buffered_key *wb;
1071         unsigned u64s;
1072         int ret = 0;
1073
1074         if (!trans->nr_updates &&
1075             !trans->nr_wb_updates &&
1076             !trans->extra_journal_entries.nr)
1077                 goto out_reset;
1078
1079         if (flags & BTREE_INSERT_GC_LOCK_HELD)
1080                 lockdep_assert_held(&c->gc_lock);
1081
1082         ret = bch2_trans_commit_run_triggers(trans);
1083         if (ret)
1084                 goto out_reset;
1085
1086         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1087                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1088                 goto out_reset;
1089         }
1090
1091         if (!(flags & BTREE_INSERT_NOCHECK_RW) &&
1092             unlikely(!bch2_write_ref_tryget(c, BCH_WRITE_REF_trans))) {
1093                 ret = bch2_trans_commit_get_rw_cold(trans, flags);
1094                 if (ret)
1095                         goto out_reset;
1096         }
1097
1098         if (c->btree_write_buffer.state.nr > c->btree_write_buffer.size / 2 &&
1099             mutex_trylock(&c->btree_write_buffer.flush_lock)) {
1100                 bch2_trans_begin(trans);
1101                 bch2_trans_unlock(trans);
1102
1103                 ret = __bch2_btree_write_buffer_flush(trans,
1104                                         flags|BTREE_INSERT_NOCHECK_RW, true);
1105                 if (!ret) {
1106                         trace_and_count(c, trans_restart_write_buffer_flush, trans, _THIS_IP_);
1107                         ret = btree_trans_restart(trans, BCH_ERR_transaction_restart_write_buffer_flush);
1108                 }
1109                 goto out;
1110         }
1111
1112         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1113
1114         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1115
1116         trans->journal_u64s             = trans->extra_journal_entries.nr;
1117         trans->journal_preres_u64s      = 0;
1118
1119         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1120
1121         if (trans->journal_transaction_names)
1122                 trans->journal_u64s += jset_u64s(JSET_ENTRY_LOG_U64s);
1123
1124         trans_for_each_update(trans, i) {
1125                 EBUG_ON(!i->path->should_be_locked);
1126
1127                 ret = bch2_btree_path_upgrade(trans, i->path, i->level + 1);
1128                 if (unlikely(ret))
1129                         goto out;
1130
1131                 EBUG_ON(!btree_node_intent_locked(i->path, i->level));
1132
1133                 if (i->key_cache_already_flushed)
1134                         continue;
1135
1136                 /* we're going to journal the key being updated: */
1137                 u64s = jset_u64s(i->k->k.u64s);
1138                 if (i->cached &&
1139                     likely(!(flags & BTREE_INSERT_JOURNAL_REPLAY)))
1140                         trans->journal_preres_u64s += u64s;
1141
1142                 if (i->flags & BTREE_UPDATE_NOJOURNAL)
1143                         continue;
1144
1145                 trans->journal_u64s += u64s;
1146
1147                 /* and we're also going to log the overwrite: */
1148                 if (trans->journal_transaction_names)
1149                         trans->journal_u64s += jset_u64s(i->old_k.u64s);
1150         }
1151
1152         trans_for_each_wb_update(trans, wb)
1153                 trans->journal_u64s += jset_u64s(wb->k.k.u64s);
1154
1155         if (trans->extra_journal_res) {
1156                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1157                                 trans->extra_journal_res,
1158                                 (flags & BTREE_INSERT_NOFAIL)
1159                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1160                 if (ret)
1161                         goto err;
1162         }
1163 retry:
1164         bch2_trans_verify_not_in_restart(trans);
1165         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1166
1167         ret = do_bch2_trans_commit(trans, flags, &i, _RET_IP_);
1168
1169         /* make sure we didn't drop or screw up locks: */
1170         bch2_trans_verify_locks(trans);
1171
1172         if (ret)
1173                 goto err;
1174
1175         trace_and_count(c, transaction_commit, trans, _RET_IP_);
1176 out:
1177         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1178
1179         if (likely(!(flags & BTREE_INSERT_NOCHECK_RW)))
1180                 bch2_write_ref_put(c, BCH_WRITE_REF_trans);
1181 out_reset:
1182         bch2_trans_reset_updates(trans);
1183
1184         return ret;
1185 err:
1186         ret = bch2_trans_commit_error(trans, flags, i, ret, _RET_IP_);
1187         if (ret)
1188                 goto out;
1189
1190         goto retry;
1191 }
1192
1193 static noinline int __check_pos_snapshot_overwritten(struct btree_trans *trans,
1194                                           enum btree_id id,
1195                                           struct bpos pos)
1196 {
1197         struct bch_fs *c = trans->c;
1198         struct btree_iter iter;
1199         struct bkey_s_c k;
1200         int ret;
1201
1202         bch2_trans_iter_init(trans, &iter, id, pos,
1203                              BTREE_ITER_NOT_EXTENTS|
1204                              BTREE_ITER_ALL_SNAPSHOTS);
1205         while (1) {
1206                 k = bch2_btree_iter_prev(&iter);
1207                 ret = bkey_err(k);
1208                 if (ret)
1209                         break;
1210
1211                 if (!k.k)
1212                         break;
1213
1214                 if (!bkey_eq(pos, k.k->p))
1215                         break;
1216
1217                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1218                         ret = 1;
1219                         break;
1220                 }
1221         }
1222         bch2_trans_iter_exit(trans, &iter);
1223
1224         return ret;
1225 }
1226
1227 static inline int check_pos_snapshot_overwritten(struct btree_trans *trans,
1228                                           enum btree_id id,
1229                                           struct bpos pos)
1230 {
1231         if (!btree_type_has_snapshots(id) ||
1232             bch2_snapshot_is_leaf(trans->c, pos.snapshot))
1233                 return 0;
1234
1235         return __check_pos_snapshot_overwritten(trans, id, pos);
1236 }
1237
1238 static noinline int extent_front_merge(struct btree_trans *trans,
1239                                        struct btree_iter *iter,
1240                                        struct bkey_s_c k,
1241                                        struct bkey_i **insert,
1242                                        enum btree_update_flags flags)
1243 {
1244         struct bch_fs *c = trans->c;
1245         struct bkey_i *update;
1246         int ret;
1247
1248         update = bch2_bkey_make_mut_noupdate(trans, k);
1249         ret = PTR_ERR_OR_ZERO(update);
1250         if (ret)
1251                 return ret;
1252
1253         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
1254                 return 0;
1255
1256         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p) ?:
1257                 check_pos_snapshot_overwritten(trans, iter->btree_id, (*insert)->k.p);
1258         if (ret < 0)
1259                 return ret;
1260         if (ret)
1261                 return 0;
1262
1263         ret = bch2_btree_delete_at(trans, iter, flags);
1264         if (ret)
1265                 return ret;
1266
1267         *insert = update;
1268         return 0;
1269 }
1270
1271 static noinline int extent_back_merge(struct btree_trans *trans,
1272                                       struct btree_iter *iter,
1273                                       struct bkey_i *insert,
1274                                       struct bkey_s_c k)
1275 {
1276         struct bch_fs *c = trans->c;
1277         int ret;
1278
1279         ret =   check_pos_snapshot_overwritten(trans, iter->btree_id, insert->k.p) ?:
1280                 check_pos_snapshot_overwritten(trans, iter->btree_id, k.k->p);
1281         if (ret < 0)
1282                 return ret;
1283         if (ret)
1284                 return 0;
1285
1286         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1287         return 0;
1288 }
1289
1290 /*
1291  * When deleting, check if we need to emit a whiteout (because we're overwriting
1292  * something in an ancestor snapshot)
1293  */
1294 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1295                                       enum btree_id btree_id, struct bpos pos)
1296 {
1297         struct btree_iter iter;
1298         struct bkey_s_c k;
1299         u32 snapshot = pos.snapshot;
1300         int ret;
1301
1302         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1303                 return 0;
1304
1305         pos.snapshot++;
1306
1307         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1308                            BTREE_ITER_ALL_SNAPSHOTS|
1309                            BTREE_ITER_NOPRESERVE, k, ret) {
1310                 if (!bkey_eq(k.k->p, pos))
1311                         break;
1312
1313                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1314                                               k.k->p.snapshot)) {
1315                         ret = !bkey_whiteout(k.k);
1316                         break;
1317                 }
1318         }
1319         bch2_trans_iter_exit(trans, &iter);
1320
1321         return ret;
1322 }
1323
1324 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
1325                                    enum btree_id id,
1326                                    struct bpos old_pos,
1327                                    struct bpos new_pos)
1328 {
1329         struct bch_fs *c = trans->c;
1330         struct btree_iter old_iter, new_iter;
1331         struct bkey_s_c old_k, new_k;
1332         snapshot_id_list s;
1333         struct bkey_i *update;
1334         int ret;
1335
1336         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
1337                 return 0;
1338
1339         darray_init(&s);
1340
1341         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
1342                              BTREE_ITER_NOT_EXTENTS|
1343                              BTREE_ITER_ALL_SNAPSHOTS);
1344         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
1345                !(ret = bkey_err(old_k)) &&
1346                bkey_eq(old_pos, old_k.k->p)) {
1347                 struct bpos whiteout_pos =
1348                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
1349
1350                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
1351                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
1352                         continue;
1353
1354                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
1355                                            BTREE_ITER_NOT_EXTENTS|
1356                                            BTREE_ITER_INTENT);
1357                 ret = bkey_err(new_k);
1358                 if (ret)
1359                         break;
1360
1361                 if (new_k.k->type == KEY_TYPE_deleted) {
1362                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
1363                         ret = PTR_ERR_OR_ZERO(update);
1364                         if (ret)
1365                                 break;
1366
1367                         bkey_init(&update->k);
1368                         update->k.p             = whiteout_pos;
1369                         update->k.type          = KEY_TYPE_whiteout;
1370
1371                         ret = bch2_trans_update(trans, &new_iter, update,
1372                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
1373                 }
1374                 bch2_trans_iter_exit(trans, &new_iter);
1375
1376                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
1377                 if (ret)
1378                         break;
1379         }
1380         bch2_trans_iter_exit(trans, &old_iter);
1381         darray_exit(&s);
1382
1383         return ret;
1384 }
1385
1386 int bch2_trans_update_extent(struct btree_trans *trans,
1387                              struct btree_iter *orig_iter,
1388                              struct bkey_i *insert,
1389                              enum btree_update_flags flags)
1390 {
1391         struct btree_iter iter;
1392         struct bpos start = bkey_start_pos(&insert->k);
1393         struct bkey_i *update;
1394         struct bkey_s_c k;
1395         enum btree_id btree_id = orig_iter->btree_id;
1396         int ret = 0, compressed_sectors;
1397
1398         bch2_trans_iter_init(trans, &iter, btree_id, start,
1399                              BTREE_ITER_INTENT|
1400                              BTREE_ITER_WITH_UPDATES|
1401                              BTREE_ITER_NOT_EXTENTS);
1402         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1403         if ((ret = bkey_err(k)))
1404                 goto err;
1405         if (!k.k)
1406                 goto out;
1407
1408         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
1409                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1410                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
1411                         if (ret)
1412                                 goto err;
1413                 }
1414
1415                 goto next;
1416         }
1417
1418         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
1419                 bool front_split = bkey_lt(bkey_start_pos(k.k), start);
1420                 bool back_split  = bkey_gt(k.k->p, insert->k.p);
1421
1422                 /*
1423                  * If we're going to be splitting a compressed extent, note it
1424                  * so that __bch2_trans_commit() can increase our disk
1425                  * reservation:
1426                  */
1427                 if (((front_split && back_split) ||
1428                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1429                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1430                         trans->extra_journal_res += compressed_sectors;
1431
1432                 if (front_split) {
1433                         update = bch2_bkey_make_mut_noupdate(trans, k);
1434                         if ((ret = PTR_ERR_OR_ZERO(update)))
1435                                 goto err;
1436
1437                         bch2_cut_back(start, update);
1438
1439                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1440                                                 k.k->p, update->k.p) ?:
1441                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1442                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1443                         if (ret)
1444                                 goto err;
1445                 }
1446
1447                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1448                     (front_split || back_split)) {
1449                         update = bch2_bkey_make_mut_noupdate(trans, k);
1450                         if ((ret = PTR_ERR_OR_ZERO(update)))
1451                                 goto err;
1452
1453                         bch2_cut_front(start, update);
1454                         bch2_cut_back(insert->k.p, update);
1455
1456                         ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
1457                                                 k.k->p, update->k.p) ?:
1458                                 bch2_btree_insert_nonextent(trans, btree_id, update,
1459                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1460                         if (ret)
1461                                 goto err;
1462                 }
1463
1464                 if (bkey_le(k.k->p, insert->k.p)) {
1465                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1466                         if ((ret = PTR_ERR_OR_ZERO(update)))
1467                                 goto err;
1468
1469                         bkey_init(&update->k);
1470                         update->k.p = k.k->p;
1471                         update->k.p.snapshot = insert->k.p.snapshot;
1472
1473                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1474                                 update->k.type = KEY_TYPE_whiteout;
1475                         } else if (btree_type_has_snapshots(btree_id)) {
1476                                 ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
1477                                 if (ret < 0)
1478                                         goto err;
1479                                 if (ret)
1480                                         update->k.type = KEY_TYPE_whiteout;
1481                         }
1482
1483                         ret = bch2_btree_insert_nonextent(trans, btree_id, update,
1484                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
1485                         if (ret)
1486                                 goto err;
1487                 }
1488
1489                 if (back_split) {
1490                         update = bch2_bkey_make_mut_noupdate(trans, k);
1491                         if ((ret = PTR_ERR_OR_ZERO(update)))
1492                                 goto err;
1493
1494                         bch2_cut_front(insert->k.p, update);
1495
1496                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1497                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1498                                                   flags, _RET_IP_);
1499                         if (ret)
1500                                 goto err;
1501                         goto out;
1502                 }
1503 next:
1504                 bch2_btree_iter_advance(&iter);
1505                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1506                 if ((ret = bkey_err(k)))
1507                         goto err;
1508                 if (!k.k)
1509                         goto out;
1510         }
1511
1512         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1513                 ret = extent_back_merge(trans, &iter, insert, k);
1514                 if (ret)
1515                         goto err;
1516         }
1517 out:
1518         if (!bkey_deleted(&insert->k)) {
1519                 /*
1520                  * Rewinding iterators is expensive: get a new one and the one
1521                  * that points to the start of insert will be cloned from:
1522                  */
1523                 bch2_trans_iter_exit(trans, &iter);
1524                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1525                                      BTREE_ITER_NOT_EXTENTS|
1526                                      BTREE_ITER_INTENT);
1527                 ret   = bch2_btree_iter_traverse(&iter) ?:
1528                         bch2_trans_update(trans, &iter, insert, flags);
1529         }
1530 err:
1531         bch2_trans_iter_exit(trans, &iter);
1532
1533         return ret;
1534 }
1535
1536 static noinline int flush_new_cached_update(struct btree_trans *trans,
1537                                             struct btree_path *path,
1538                                             struct btree_insert_entry *i,
1539                                             enum btree_update_flags flags,
1540                                             unsigned long ip)
1541 {
1542         struct btree_path *btree_path;
1543         struct bkey k;
1544         int ret;
1545
1546         btree_path = bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1547                                    BTREE_ITER_INTENT, _THIS_IP_);
1548         ret = bch2_btree_path_traverse(trans, btree_path, 0);
1549         if (ret)
1550                 goto out;
1551
1552         /*
1553          * The old key in the insert entry might actually refer to an existing
1554          * key in the btree that has been deleted from cache and not yet
1555          * flushed. Check for this and skip the flush so we don't run triggers
1556          * against a stale key.
1557          */
1558         bch2_btree_path_peek_slot_exact(btree_path, &k);
1559         if (!bkey_deleted(&k))
1560                 goto out;
1561
1562         i->key_cache_already_flushed = true;
1563         i->flags |= BTREE_TRIGGER_NORUN;
1564
1565         btree_path_set_should_be_locked(btree_path);
1566         ret = bch2_trans_update_by_path(trans, btree_path, i->k, flags, ip);
1567 out:
1568         bch2_path_put(trans, btree_path, true);
1569         return ret;
1570 }
1571
1572 static int __must_check
1573 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1574                           struct bkey_i *k, enum btree_update_flags flags,
1575                           unsigned long ip)
1576 {
1577         struct bch_fs *c = trans->c;
1578         struct btree_insert_entry *i, n;
1579         u64 seq = 0;
1580         int cmp;
1581
1582         EBUG_ON(!path->should_be_locked);
1583         EBUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1584         EBUG_ON(!bpos_eq(k->k.p, path->pos));
1585
1586         /*
1587          * The transaction journal res hasn't been allocated at this point.
1588          * That occurs at commit time. Reuse the seq field to pass in the seq
1589          * of a prejournaled key.
1590          */
1591         if (flags & BTREE_UPDATE_PREJOURNAL)
1592                 seq = trans->journal_res.seq;
1593
1594         n = (struct btree_insert_entry) {
1595                 .flags          = flags,
1596                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1597                 .btree_id       = path->btree_id,
1598                 .level          = path->level,
1599                 .cached         = path->cached,
1600                 .path           = path,
1601                 .k              = k,
1602                 .seq            = seq,
1603                 .ip_allocated   = ip,
1604         };
1605
1606 #ifdef CONFIG_BCACHEFS_DEBUG
1607         trans_for_each_update(trans, i)
1608                 BUG_ON(i != trans->updates &&
1609                        btree_insert_entry_cmp(i - 1, i) >= 0);
1610 #endif
1611
1612         /*
1613          * Pending updates are kept sorted: first, find position of new update,
1614          * then delete/trim any updates the new update overwrites:
1615          */
1616         trans_for_each_update(trans, i) {
1617                 cmp = btree_insert_entry_cmp(&n, i);
1618                 if (cmp <= 0)
1619                         break;
1620         }
1621
1622         if (!cmp && i < trans->updates + trans->nr_updates) {
1623                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1624
1625                 bch2_path_put(trans, i->path, true);
1626                 i->flags        = n.flags;
1627                 i->cached       = n.cached;
1628                 i->k            = n.k;
1629                 i->path         = n.path;
1630                 i->seq          = n.seq;
1631                 i->ip_allocated = n.ip_allocated;
1632         } else {
1633                 array_insert_item(trans->updates, trans->nr_updates,
1634                                   i - trans->updates, n);
1635
1636                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
1637                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1638
1639                 if (unlikely(trans->journal_replay_not_finished)) {
1640                         struct bkey_i *j_k =
1641                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1642
1643                         if (j_k) {
1644                                 i->old_k = j_k->k;
1645                                 i->old_v = &j_k->v;
1646                         }
1647                 }
1648         }
1649
1650         __btree_path_get(i->path, true);
1651
1652         /*
1653          * If a key is present in the key cache, it must also exist in the
1654          * btree - this is necessary for cache coherency. When iterating over
1655          * a btree that's cached in the key cache, the btree iter code checks
1656          * the key cache - but the key has to exist in the btree for that to
1657          * work:
1658          */
1659         if (path->cached && bkey_deleted(&i->old_k))
1660                 return flush_new_cached_update(trans, path, i, flags, ip);
1661
1662         return 0;
1663 }
1664
1665 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1666                                    struct bkey_i *k, enum btree_update_flags flags)
1667 {
1668         struct btree_path *path = iter->update_path ?: iter->path;
1669         struct bkey_cached *ck;
1670         int ret;
1671
1672         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1673                 return bch2_trans_update_extent(trans, iter, k, flags);
1674
1675         if (bkey_deleted(&k->k) &&
1676             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1677             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1678                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1679                 if (unlikely(ret < 0))
1680                         return ret;
1681
1682                 if (ret)
1683                         k->k.type = KEY_TYPE_whiteout;
1684         }
1685
1686         /*
1687          * Ensure that updates to cached btrees go to the key cache:
1688          */
1689         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1690             !path->cached &&
1691             !path->level &&
1692             btree_id_cached(trans->c, path->btree_id)) {
1693                 if (!iter->key_cache_path ||
1694                     !iter->key_cache_path->should_be_locked ||
1695                     !bpos_eq(iter->key_cache_path->pos, k->k.p)) {
1696                         if (!iter->key_cache_path)
1697                                 iter->key_cache_path =
1698                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1699                                                       BTREE_ITER_INTENT|
1700                                                       BTREE_ITER_CACHED, _THIS_IP_);
1701
1702                         iter->key_cache_path =
1703                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1704                                                         iter->flags & BTREE_ITER_INTENT,
1705                                                         _THIS_IP_);
1706
1707                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1708                                                        BTREE_ITER_CACHED);
1709                         if (unlikely(ret))
1710                                 return ret;
1711
1712                         ck = (void *) iter->key_cache_path->l[0].b;
1713
1714                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1715                                 trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
1716                                 return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
1717                         }
1718
1719                         btree_path_set_should_be_locked(iter->key_cache_path);
1720                 }
1721
1722                 path = iter->key_cache_path;
1723         }
1724
1725         return bch2_trans_update_by_path(trans, path, k, flags, _RET_IP_);
1726 }
1727
1728 /*
1729  * Add a transaction update for a key that has already been journaled.
1730  */
1731 int __must_check bch2_trans_update_seq(struct btree_trans *trans, u64 seq,
1732                                        struct btree_iter *iter, struct bkey_i *k,
1733                                        enum btree_update_flags flags)
1734 {
1735         trans->journal_res.seq = seq;
1736         return bch2_trans_update(trans, iter, k, flags|BTREE_UPDATE_NOJOURNAL|
1737                                                  BTREE_UPDATE_PREJOURNAL);
1738 }
1739
1740 int __must_check bch2_trans_update_buffered(struct btree_trans *trans,
1741                                             enum btree_id btree,
1742                                             struct bkey_i *k)
1743 {
1744         struct btree_write_buffered_key *i;
1745         int ret;
1746
1747         EBUG_ON(trans->nr_wb_updates > trans->wb_updates_size);
1748         EBUG_ON(k->k.u64s > BTREE_WRITE_BUFERED_U64s_MAX);
1749
1750         trans_for_each_wb_update(trans, i) {
1751                 if (i->btree == btree && bpos_eq(i->k.k.p, k->k.p)) {
1752                         bkey_copy(&i->k, k);
1753                         return 0;
1754                 }
1755         }
1756
1757         if (!trans->wb_updates ||
1758             trans->nr_wb_updates == trans->wb_updates_size) {
1759                 struct btree_write_buffered_key *u;
1760
1761                 if (trans->nr_wb_updates == trans->wb_updates_size) {
1762                         struct btree_transaction_stats *s = btree_trans_stats(trans);
1763
1764                         BUG_ON(trans->wb_updates_size > U8_MAX / 2);
1765                         trans->wb_updates_size = max(1, trans->wb_updates_size * 2);
1766                         if (s)
1767                                 s->wb_updates_size = trans->wb_updates_size;
1768                 }
1769
1770                 u = bch2_trans_kmalloc_nomemzero(trans,
1771                                         trans->wb_updates_size *
1772                                         sizeof(struct btree_write_buffered_key));
1773                 ret = PTR_ERR_OR_ZERO(u);
1774                 if (ret)
1775                         return ret;
1776
1777                 if (trans->nr_wb_updates)
1778                         memcpy(u, trans->wb_updates, trans->nr_wb_updates *
1779                                sizeof(struct btree_write_buffered_key));
1780                 trans->wb_updates = u;
1781         }
1782
1783         trans->wb_updates[trans->nr_wb_updates] = (struct btree_write_buffered_key) {
1784                 .btree  = btree,
1785         };
1786
1787         bkey_copy(&trans->wb_updates[trans->nr_wb_updates].k, k);
1788         trans->nr_wb_updates++;
1789
1790         return 0;
1791 }
1792
1793 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
1794                              enum btree_id btree, struct bpos end)
1795 {
1796         struct bkey_s_c k;
1797         int ret = 0;
1798
1799         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
1800         k = bch2_btree_iter_prev(iter);
1801         ret = bkey_err(k);
1802         if (ret)
1803                 goto err;
1804
1805         bch2_btree_iter_advance(iter);
1806         k = bch2_btree_iter_peek_slot(iter);
1807         ret = bkey_err(k);
1808         if (ret)
1809                 goto err;
1810
1811         BUG_ON(k.k->type != KEY_TYPE_deleted);
1812
1813         if (bkey_gt(k.k->p, end)) {
1814                 ret = -BCH_ERR_ENOSPC_btree_slot;
1815                 goto err;
1816         }
1817
1818         return 0;
1819 err:
1820         bch2_trans_iter_exit(trans, iter);
1821         return ret;
1822 }
1823
1824 void bch2_trans_commit_hook(struct btree_trans *trans,
1825                             struct btree_trans_commit_hook *h)
1826 {
1827         h->next = trans->hooks;
1828         trans->hooks = h;
1829 }
1830
1831 int bch2_btree_insert_nonextent(struct btree_trans *trans,
1832                                 enum btree_id btree, struct bkey_i *k,
1833                                 enum btree_update_flags flags)
1834 {
1835         struct btree_iter iter;
1836         int ret;
1837
1838         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
1839                              BTREE_ITER_NOT_EXTENTS|
1840                              BTREE_ITER_INTENT);
1841         ret   = bch2_btree_iter_traverse(&iter) ?:
1842                 bch2_trans_update(trans, &iter, k, flags);
1843         bch2_trans_iter_exit(trans, &iter);
1844         return ret;
1845 }
1846
1847 int __bch2_btree_insert(struct btree_trans *trans, enum btree_id id,
1848                         struct bkey_i *k, enum btree_update_flags flags)
1849 {
1850         struct btree_iter iter;
1851         int ret;
1852
1853         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1854                              BTREE_ITER_CACHED|
1855                              BTREE_ITER_INTENT);
1856         ret   = bch2_btree_iter_traverse(&iter) ?:
1857                 bch2_trans_update(trans, &iter, k, flags);
1858         bch2_trans_iter_exit(trans, &iter);
1859         return ret;
1860 }
1861
1862 /**
1863  * bch2_btree_insert - insert keys into the extent btree
1864  * @c:                  pointer to struct bch_fs
1865  * @id:                 btree to insert into
1866  * @insert_keys:        list of keys to insert
1867  * @hook:               insert callback
1868  */
1869 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1870                       struct bkey_i *k,
1871                       struct disk_reservation *disk_res,
1872                       u64 *journal_seq, int flags)
1873 {
1874         return bch2_trans_do(c, disk_res, journal_seq, flags,
1875                              __bch2_btree_insert(&trans, id, k, 0));
1876 }
1877
1878 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1879                                 unsigned len, unsigned update_flags)
1880 {
1881         struct bkey_i *k;
1882
1883         k = bch2_trans_kmalloc(trans, sizeof(*k));
1884         if (IS_ERR(k))
1885                 return PTR_ERR(k);
1886
1887         bkey_init(&k->k);
1888         k->k.p = iter->pos;
1889         bch2_key_resize(&k->k, len);
1890         return bch2_trans_update(trans, iter, k, update_flags);
1891 }
1892
1893 int bch2_btree_delete_at(struct btree_trans *trans,
1894                          struct btree_iter *iter, unsigned update_flags)
1895 {
1896         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1897 }
1898
1899 int bch2_btree_delete_at_buffered(struct btree_trans *trans,
1900                                   enum btree_id btree, struct bpos pos)
1901 {
1902         struct bkey_i *k;
1903
1904         k = bch2_trans_kmalloc(trans, sizeof(*k));
1905         if (IS_ERR(k))
1906                 return PTR_ERR(k);
1907
1908         bkey_init(&k->k);
1909         k->k.p = pos;
1910         return bch2_trans_update_buffered(trans, btree, k);
1911 }
1912
1913 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1914                                   struct bpos start, struct bpos end,
1915                                   unsigned update_flags,
1916                                   u64 *journal_seq)
1917 {
1918         u32 restart_count = trans->restart_count;
1919         struct btree_iter iter;
1920         struct bkey_s_c k;
1921         int ret = 0;
1922
1923         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1924         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
1925                 struct disk_reservation disk_res =
1926                         bch2_disk_reservation_init(trans->c, 0);
1927                 struct bkey_i delete;
1928
1929                 ret = bkey_err(k);
1930                 if (ret)
1931                         goto err;
1932
1933                 bkey_init(&delete.k);
1934
1935                 /*
1936                  * This could probably be more efficient for extents:
1937                  */
1938
1939                 /*
1940                  * For extents, iter.pos won't necessarily be the same as
1941                  * bkey_start_pos(k.k) (for non extents they always will be the
1942                  * same). It's important that we delete starting from iter.pos
1943                  * because the range we want to delete could start in the middle
1944                  * of k.
1945                  *
1946                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1947                  * bkey_start_pos(k.k)).
1948                  */
1949                 delete.k.p = iter.pos;
1950
1951                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
1952                         bch2_key_resize(&delete.k,
1953                                         bpos_min(end, k.k->p).offset -
1954                                         iter.pos.offset);
1955
1956                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1957                         bch2_trans_commit(trans, &disk_res, journal_seq,
1958                                           BTREE_INSERT_NOFAIL);
1959                 bch2_disk_reservation_put(trans->c, &disk_res);
1960 err:
1961                 /*
1962                  * the bch2_trans_begin() call is in a weird place because we
1963                  * need to call it after every transaction commit, to avoid path
1964                  * overflow, but don't want to call it if the delete operation
1965                  * is a no-op and we have no work to do:
1966                  */
1967                 bch2_trans_begin(trans);
1968
1969                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
1970                         ret = 0;
1971                 if (ret)
1972                         break;
1973         }
1974         bch2_trans_iter_exit(trans, &iter);
1975
1976         if (!ret && trans_was_restarted(trans, restart_count))
1977                 ret = -BCH_ERR_transaction_restart_nested;
1978         return ret;
1979 }
1980
1981 /*
1982  * bch_btree_delete_range - delete everything within a given range
1983  *
1984  * Range is a half open interval - [start, end)
1985  */
1986 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1987                             struct bpos start, struct bpos end,
1988                             unsigned update_flags,
1989                             u64 *journal_seq)
1990 {
1991         int ret = bch2_trans_run(c,
1992                         bch2_btree_delete_range_trans(&trans, id, start, end,
1993                                                       update_flags, journal_seq));
1994         if (ret == -BCH_ERR_transaction_restart_nested)
1995                 ret = 0;
1996         return ret;
1997 }
1998
1999 int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
2000                        struct bpos pos, bool set)
2001 {
2002         struct bkey_i *k;
2003         int ret = 0;
2004
2005         k = bch2_trans_kmalloc_nomemzero(trans, sizeof(*k));
2006         ret = PTR_ERR_OR_ZERO(k);
2007         if (unlikely(ret))
2008                 return ret;
2009
2010         bkey_init(&k->k);
2011         k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
2012         k->k.p = pos;
2013
2014         return bch2_trans_update_buffered(trans, btree, k);
2015 }
2016
2017 static int __bch2_trans_log_msg(darray_u64 *entries, const char *fmt, va_list args)
2018 {
2019         struct printbuf buf = PRINTBUF;
2020         struct jset_entry_log *l;
2021         unsigned u64s;
2022         int ret;
2023
2024         prt_vprintf(&buf, fmt, args);
2025         ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
2026         if (ret)
2027                 goto err;
2028
2029         u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
2030
2031         ret = darray_make_room(entries, jset_u64s(u64s));
2032         if (ret)
2033                 goto err;
2034
2035         l = (void *) &darray_top(*entries);
2036         l->entry.u64s           = cpu_to_le16(u64s);
2037         l->entry.btree_id       = 0;
2038         l->entry.level          = 1;
2039         l->entry.type           = BCH_JSET_ENTRY_log;
2040         l->entry.pad[0]         = 0;
2041         l->entry.pad[1]         = 0;
2042         l->entry.pad[2]         = 0;
2043         memcpy(l->d, buf.buf, buf.pos);
2044         while (buf.pos & 7)
2045                 l->d[buf.pos++] = '\0';
2046
2047         entries->nr += jset_u64s(u64s);
2048 err:
2049         printbuf_exit(&buf);
2050         return ret;
2051 }
2052
2053 static int
2054 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
2055                   va_list args)
2056 {
2057         int ret;
2058
2059         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
2060                 ret = __bch2_trans_log_msg(&c->journal.early_journal_entries, fmt, args);
2061         } else {
2062                 ret = bch2_trans_do(c, NULL, NULL,
2063                         BTREE_INSERT_LAZY_RW|commit_flags,
2064                         __bch2_trans_log_msg(&trans.extra_journal_entries, fmt, args));
2065         }
2066
2067         return ret;
2068 }
2069
2070 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
2071 {
2072         va_list args;
2073         int ret;
2074
2075         va_start(args, fmt);
2076         ret = __bch2_fs_log_msg(c, 0, fmt, args);
2077         va_end(args);
2078         return ret;
2079 }
2080
2081 /*
2082  * Use for logging messages during recovery to enable reserved space and avoid
2083  * blocking.
2084  */
2085 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
2086 {
2087         va_list args;
2088         int ret;
2089
2090         va_start(args, fmt);
2091         ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
2092         va_end(args);
2093         return ret;
2094 }