git.sesse.net Git - bcachefs-tools-debian/blob - libbcachefs/btree_update_leaf.c

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_update_interior.h"
   6 #include "btree_gc.h"
   7 #include "btree_io.h"
   8 #include "btree_iter.h"
   9 #include "btree_key_cache.h"
  10 #include "btree_locking.h"
  11 #include "buckets.h"
  12 #include "debug.h"
  13 #include "error.h"
  14 #include "extent_update.h"
  15 #include "journal.h"
  16 #include "journal_reclaim.h"
  17 #include "keylist.h"
  18 #include "recovery.h"
  19 #include "subvolume.h"
  20 #include "replicas.h"
  21
  22 #include <linux/prefetch.h>
  23 #include <linux/sort.h>
  24 #include <trace/events/bcachefs.h>
  25
  26 static int __must_check
  27 bch2_trans_update_by_path(struct btree_trans *, struct btree_path *,
  28                           struct bkey_i *, enum btree_update_flags);
  29
  30 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  31                                          const struct btree_insert_entry *r)
  32 {
  33         return   cmp_int(l->btree_id,   r->btree_id) ?:
  34                  -cmp_int(l->level,     r->level) ?:
  35                  bpos_cmp(l->k->k.p,    r->k->k.p);
  36 }
  37
  38 static inline struct btree_path_level *insert_l(struct btree_insert_entry *i)
  39 {
  40         return i->path->l + i->level;
  41 }
  42
  43 static inline bool same_leaf_as_prev(struct btree_trans *trans,
  44                                      struct btree_insert_entry *i)
  45 {
  46         return i != trans->updates &&
  47                 insert_l(&i[0])->b == insert_l(&i[-1])->b;
  48 }
  49
  50 static inline bool same_leaf_as_next(struct btree_trans *trans,
  51                                      struct btree_insert_entry *i)
  52 {
  53         return i + 1 < trans->updates + trans->nr_updates &&
  54                 insert_l(&i[0])->b == insert_l(&i[1])->b;
  55 }
  56
  57 static inline void bch2_btree_node_prep_for_write(struct btree_trans *trans,
  58                                                   struct btree_path *path,
  59                                                   struct btree *b)
  60 {
  61         struct bch_fs *c = trans->c;
  62
  63         if (path->cached)
  64                 return;
  65
  66         if (unlikely(btree_node_just_written(b)) &&
  67             bch2_btree_post_write_cleanup(c, b))
  68                 bch2_trans_node_reinit_iter(trans, b);
  69
  70         /*
  71          * If the last bset has been written, or if it's gotten too big - start
  72          * a new bset to insert into:
  73          */
  74         if (want_new_bset(c, b))
  75                 bch2_btree_init_next(trans, b);
  76 }
  77
  78 void bch2_btree_node_lock_for_insert(struct btree_trans *trans,
  79                                      struct btree_path *path,
  80                                      struct btree *b)
  81 {
  82         bch2_btree_node_lock_write(trans, path, b);
  83         bch2_btree_node_prep_for_write(trans, path, b);
  84 }
  85
  86 /* Inserting into a given leaf node (last stage of insert): */
  87
  88 /* Handle overwrites and do insert, for non extents: */
  89 bool bch2_btree_bset_insert_key(struct btree_trans *trans,
  90                                 struct btree_path *path,
  91                                 struct btree *b,
  92                                 struct btree_node_iter *node_iter,
  93                                 struct bkey_i *insert)
  94 {
  95         struct bkey_packed *k;
  96         unsigned clobber_u64s = 0, new_u64s = 0;
  97
  98         EBUG_ON(btree_node_just_written(b));
  99         EBUG_ON(bset_written(b, btree_bset_last(b)));
 100         EBUG_ON(bkey_deleted(&insert->k) && bkey_val_u64s(&insert->k));
 101         EBUG_ON(bpos_cmp(insert->k.p, b->data->min_key) < 0);
 102         EBUG_ON(bpos_cmp(insert->k.p, b->data->max_key) > 0);
 103         EBUG_ON(insert->k.u64s >
 104                 bch_btree_keys_u64s_remaining(trans->c, b));
 105
 106         k = bch2_btree_node_iter_peek_all(node_iter, b);
 107         if (k && bkey_cmp_left_packed(b, k, &insert->k.p))
 108                 k = NULL;
 109
 110         /* @k is the key being overwritten/deleted, if any: */
 111         EBUG_ON(k && bkey_deleted(k));
 112
 113         /* Deleting, but not found? nothing to do: */
 114         if (bkey_deleted(&insert->k) && !k)
 115                 return false;
 116
 117         if (bkey_deleted(&insert->k)) {
 118                 /* Deleting: */
 119                 btree_account_key_drop(b, k);
 120                 k->type = KEY_TYPE_deleted;
 121
 122                 if (k->needs_whiteout)
 123                         push_whiteout(trans->c, b, insert->k.p);
 124                 k->needs_whiteout = false;
 125
 126                 if (k >= btree_bset_last(b)->start) {
 127                         clobber_u64s = k->u64s;
 128                         bch2_bset_delete(b, k, clobber_u64s);
 129                         goto fix_iter;
 130                 } else {
 131                         bch2_btree_path_fix_key_modified(trans, b, k);
 132                 }
 133
 134                 return true;
 135         }
 136
 137         if (k) {
 138                 /* Overwriting: */
 139                 btree_account_key_drop(b, k);
 140                 k->type = KEY_TYPE_deleted;
 141
 142                 insert->k.needs_whiteout = k->needs_whiteout;
 143                 k->needs_whiteout = false;
 144
 145                 if (k >= btree_bset_last(b)->start) {
 146                         clobber_u64s = k->u64s;
 147                         goto overwrite;
 148                 } else {
 149                         bch2_btree_path_fix_key_modified(trans, b, k);
 150                 }
 151         }
 152
 153         k = bch2_btree_node_iter_bset_pos(node_iter, b, bset_tree_last(b));
 154 overwrite:
 155         bch2_bset_insert(b, node_iter, k, insert, clobber_u64s);
 156         new_u64s = k->u64s;
 157 fix_iter:
 158         if (clobber_u64s != new_u64s)
 159                 bch2_btree_node_iter_fix(trans, path, b, node_iter, k,
 160                                          clobber_u64s, new_u64s);
 161         return true;
 162 }
 163
 164 static int __btree_node_flush(struct journal *j, struct journal_entry_pin *pin,
 165                                unsigned i, u64 seq)
 166 {
 167         struct bch_fs *c = container_of(j, struct bch_fs, journal);
 168         struct btree_write *w = container_of(pin, struct btree_write, journal);
 169         struct btree *b = container_of(w, struct btree, writes[i]);
 170         unsigned long old, new, v;
 171         unsigned idx = w - b->writes;
 172
 173         six_lock_read(&b->c.lock, NULL, NULL);
 174         v = READ_ONCE(b->flags);
 175
 176         do {
 177                 old = new = v;
 178
 179                 if (!(old & (1 << BTREE_NODE_dirty)) ||
 180                     !!(old & (1 << BTREE_NODE_write_idx)) != idx ||
 181                     w->journal.seq != seq)
 182                         break;
 183
 184                 new |= 1 << BTREE_NODE_need_write;
 185         } while ((v = cmpxchg(&b->flags, old, new)) != old);
 186
 187         btree_node_write_if_need(c, b, SIX_LOCK_read);
 188         six_unlock_read(&b->c.lock);
 189         return 0;
 190 }
 191
 192 static int btree_node_flush0(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 193 {
 194         return __btree_node_flush(j, pin, 0, seq);
 195 }
 196
 197 static int btree_node_flush1(struct journal *j, struct journal_entry_pin *pin, u64 seq)
 198 {
 199         return __btree_node_flush(j, pin, 1, seq);
 200 }
 201
 202 inline void bch2_btree_add_journal_pin(struct bch_fs *c,
 203                                        struct btree *b, u64 seq)
 204 {
 205         struct btree_write *w = btree_current_write(b);
 206
 207         bch2_journal_pin_add(&c->journal, seq, &w->journal,
 208                              btree_node_write_idx(b) == 0
 209                              ? btree_node_flush0
 210                              : btree_node_flush1);
 211 }
 212
 213 /**
 214  * btree_insert_key - insert a key one key into a leaf node
 215  */
 216 static void btree_insert_key_leaf(struct btree_trans *trans,
 217                                   struct btree_insert_entry *insert)
 218 {
 219         struct bch_fs *c = trans->c;
 220         struct btree *b = insert_l(insert)->b;
 221         struct bset_tree *t = bset_tree_last(b);
 222         struct bset *i = bset(b, t);
 223         int old_u64s = bset_u64s(t);
 224         int old_live_u64s = b->nr.live_u64s;
 225         int live_u64s_added, u64s_added;
 226
 227         if (unlikely(!bch2_btree_bset_insert_key(trans, insert->path, b,
 228                                         &insert_l(insert)->iter, insert->k)))
 229                 return;
 230
 231         i->journal_seq = cpu_to_le64(max(trans->journal_res.seq,
 232                                          le64_to_cpu(i->journal_seq)));
 233
 234         bch2_btree_add_journal_pin(c, b, trans->journal_res.seq);
 235
 236         if (unlikely(!btree_node_dirty(b)))
 237                 set_btree_node_dirty_acct(c, b);
 238
 239         live_u64s_added = (int) b->nr.live_u64s - old_live_u64s;
 240         u64s_added = (int) bset_u64s(t) - old_u64s;
 241
 242         if (b->sib_u64s[0] != U16_MAX && live_u64s_added < 0)
 243                 b->sib_u64s[0] = max(0, (int) b->sib_u64s[0] + live_u64s_added);
 244         if (b->sib_u64s[1] != U16_MAX && live_u64s_added < 0)
 245                 b->sib_u64s[1] = max(0, (int) b->sib_u64s[1] + live_u64s_added);
 246
 247         if (u64s_added > live_u64s_added &&
 248             bch2_maybe_compact_whiteouts(c, b))
 249                 bch2_trans_node_reinit_iter(trans, b);
 250 }
 251
 252 /* Cached btree updates: */
 253
 254 /* Normal update interface: */
 255
 256 static inline void btree_insert_entry_checks(struct btree_trans *trans,
 257                                              struct btree_insert_entry *i)
 258 {
 259         BUG_ON(bpos_cmp(i->k->k.p, i->path->pos));
 260         BUG_ON(i->cached        != i->path->cached);
 261         BUG_ON(i->level         != i->path->level);
 262         BUG_ON(i->btree_id      != i->path->btree_id);
 263         EBUG_ON(!i->level &&
 264                 !(i->flags & BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE) &&
 265                 test_bit(JOURNAL_REPLAY_DONE, &trans->c->journal.flags) &&
 266                 i->k->k.p.snapshot &&
 267                 bch2_snapshot_internal_node(trans->c, i->k->k.p.snapshot));
 268 }
 269
 270 static noinline int
 271 bch2_trans_journal_preres_get_cold(struct btree_trans *trans, unsigned u64s,
 272                                    unsigned long trace_ip)
 273 {
 274         struct bch_fs *c = trans->c;
 275         int ret;
 276
 277         bch2_trans_unlock(trans);
 278
 279         ret = bch2_journal_preres_get(&c->journal,
 280                         &trans->journal_preres, u64s, 0);
 281         if (ret)
 282                 return ret;
 283
 284         if (!bch2_trans_relock(trans)) {
 285                 trace_trans_restart_journal_preres_get(trans->fn, trace_ip);
 286                 return -EINTR;
 287         }
 288
 289         return 0;
 290 }
 291
 292 static inline int bch2_trans_journal_res_get(struct btree_trans *trans,
 293                                              unsigned flags)
 294 {
 295         struct bch_fs *c = trans->c;
 296         int ret;
 297
 298         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 299                                    trans->journal_u64s,
 300                                    flags|
 301                                    (trans->flags & JOURNAL_WATERMARK_MASK));
 302
 303         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 304 }
 305
 306 #define JSET_ENTRY_LOG_U64s             4
 307
 308 static noinline void journal_transaction_name(struct btree_trans *trans)
 309 {
 310         struct bch_fs *c = trans->c;
 311         struct jset_entry *entry = journal_res_entry(&c->journal, &trans->journal_res);
 312         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 313         unsigned u64s = JSET_ENTRY_LOG_U64s - 1;
 314         unsigned b, buflen = u64s * sizeof(u64);
 315
 316         l->entry.u64s           = cpu_to_le16(u64s);
 317         l->entry.btree_id       = 0;
 318         l->entry.level          = 0;
 319         l->entry.type           = BCH_JSET_ENTRY_log;
 320         l->entry.pad[0]         = 0;
 321         l->entry.pad[1]         = 0;
 322         l->entry.pad[2]         = 0;
 323         b = min_t(unsigned, strlen(trans->fn), buflen);
 324         memcpy(l->d, trans->fn, b);
 325         while (b < buflen)
 326                 l->d[b++] = '\0';
 327
 328         trans->journal_res.offset       += JSET_ENTRY_LOG_U64s;
 329         trans->journal_res.u64s         -= JSET_ENTRY_LOG_U64s;
 330 }
 331
 332 static inline enum btree_insert_ret
 333 btree_key_can_insert(struct btree_trans *trans,
 334                      struct btree *b,
 335                      unsigned u64s)
 336 {
 337         struct bch_fs *c = trans->c;
 338
 339         if (!bch2_btree_node_insert_fits(c, b, u64s))
 340                 return BTREE_INSERT_BTREE_NODE_FULL;
 341
 342         return BTREE_INSERT_OK;
 343 }
 344
 345 static enum btree_insert_ret
 346 btree_key_can_insert_cached(struct btree_trans *trans,
 347                             struct btree_path *path,
 348                             unsigned u64s)
 349 {
 350         struct bch_fs *c = trans->c;
 351         struct bkey_cached *ck = (void *) path->l[0].b;
 352         unsigned old_u64s = ck->u64s, new_u64s;
 353         struct bkey_i *new_k;
 354
 355         EBUG_ON(path->level);
 356
 357         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 358             bch2_btree_key_cache_must_wait(c) &&
 359             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 360                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 361
 362         /*
 363          * bch2_varint_decode can read past the end of the buffer by at most 7
 364          * bytes (it won't be used):
 365          */
 366         u64s += 1;
 367
 368         if (u64s <= ck->u64s)
 369                 return BTREE_INSERT_OK;
 370
 371         new_u64s        = roundup_pow_of_two(u64s);
 372         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 373         if (!new_k) {
 374                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 375                         bch2_btree_ids[path->btree_id], new_u64s);
 376                 return -ENOMEM;
 377         }
 378
 379         ck->u64s        = new_u64s;
 380         ck->k           = new_k;
 381         /*
 382          * Keys returned by peek() are no longer valid pointers, so we need a
 383          * transaction restart:
 384          */
 385         trace_trans_restart_key_cache_key_realloced(trans->fn, _RET_IP_,
 386                                              path->btree_id, &path->pos,
 387                                              old_u64s, new_u64s);
 388         /*
 389          * Not using btree_trans_restart() because we can't unlock here, we have
 390          * write locks held:
 391          */
 392         trans->restarted = true;
 393         return -EINTR;
 394 }
 395
 396 static inline void do_btree_insert_one(struct btree_trans *trans,
 397                                        struct btree_insert_entry *i)
 398 {
 399         struct bch_fs *c = trans->c;
 400         struct journal *j = &c->journal;
 401
 402         EBUG_ON(trans->journal_res.ref !=
 403                 !(trans->flags & BTREE_INSERT_JOURNAL_REPLAY));
 404
 405         i->k->k.needs_whiteout = false;
 406
 407         if (!i->cached)
 408                 btree_insert_key_leaf(trans, i);
 409         else
 410                 bch2_btree_insert_key_cached(trans, i->path, i->k);
 411
 412         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 413                 bch2_journal_add_keys(j, &trans->journal_res,
 414                                       i->btree_id,
 415                                       i->level,
 416                                       i->k);
 417
 418                 if (trans->journal_seq)
 419                         *trans->journal_seq = trans->journal_res.seq;
 420         }
 421 }
 422
 423 /* Triggers: */
 424
 425 static int run_one_mem_trigger(struct btree_trans *trans,
 426                                struct btree_insert_entry *i,
 427                                unsigned flags)
 428 {
 429         struct bkey_s_c old = { &i->old_k, i->old_v };
 430         struct bkey_i *new = i->k;
 431         int ret;
 432
 433         if (unlikely(flags & BTREE_TRIGGER_NORUN))
 434                 return 0;
 435
 436         if (!btree_node_type_needs_gc(i->btree_id))
 437                 return 0;
 438
 439         if (bch2_bkey_ops[old.k->type].atomic_trigger ==
 440             bch2_bkey_ops[i->k->k.type].atomic_trigger &&
 441             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 442                 ret   = bch2_mark_key(trans, old, bkey_i_to_s_c(new),
 443                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|flags);
 444         } else {
 445                 struct bkey             _deleted = KEY(0, 0, 0);
 446                 struct bkey_s_c         deleted = (struct bkey_s_c) { &_deleted, NULL };
 447
 448                 _deleted.p = i->path->pos;
 449
 450                 ret   = bch2_mark_key(trans, deleted, bkey_i_to_s_c(new),
 451                                 BTREE_TRIGGER_INSERT|flags) ?:
 452                         bch2_mark_key(trans, old, deleted,
 453                                 BTREE_TRIGGER_OVERWRITE|flags);
 454         }
 455
 456         return ret;
 457 }
 458
 459 static int run_one_trans_trigger(struct btree_trans *trans, struct btree_insert_entry *i,
 460                                  bool overwrite)
 461 {
 462         /*
 463          * Transactional triggers create new btree_insert_entries, so we can't
 464          * pass them a pointer to a btree_insert_entry, that memory is going to
 465          * move:
 466          */
 467         struct bkey old_k = i->old_k;
 468         struct bkey_s_c old = { &old_k, i->old_v };
 469
 470         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 471             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 472                 return 0;
 473
 474         if (!i->insert_trigger_run &&
 475             !i->overwrite_trigger_run &&
 476             bch2_bkey_ops[old.k->type].trans_trigger ==
 477             bch2_bkey_ops[i->k->k.type].trans_trigger &&
 478             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 479                 i->overwrite_trigger_run = true;
 480                 i->insert_trigger_run = true;
 481                 return bch2_trans_mark_key(trans, i->btree_id, i->level, old, i->k,
 482                                            BTREE_TRIGGER_INSERT|
 483                                            BTREE_TRIGGER_OVERWRITE|
 484                                            i->flags) ?: 1;
 485         } else if (overwrite && !i->overwrite_trigger_run) {
 486                 i->overwrite_trigger_run = true;
 487                 return bch2_trans_mark_old(trans, i->btree_id, i->level, old, i->flags) ?: 1;
 488         } else if (!overwrite && !i->insert_trigger_run) {
 489                 i->insert_trigger_run = true;
 490                 return bch2_trans_mark_new(trans, i->btree_id, i->level, i->k, i->flags) ?: 1;
 491         } else {
 492                 return 0;
 493         }
 494 }
 495
 496 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 497                               struct btree_insert_entry *btree_id_start)
 498 {
 499         struct btree_insert_entry *i;
 500         bool trans_trigger_run;
 501         int ret, overwrite;
 502
 503         for (overwrite = 1; overwrite >= 0; --overwrite) {
 504
 505                 /*
 506                  * Running triggers will append more updates to the list of updates as
 507                  * we're walking it:
 508                  */
 509                 do {
 510                         trans_trigger_run = false;
 511
 512                         for (i = btree_id_start;
 513                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 514                              i++) {
 515                                 if (i->btree_id != btree_id)
 516                                         continue;
 517
 518                                 ret = run_one_trans_trigger(trans, i, overwrite);
 519                                 if (ret < 0)
 520                                         return ret;
 521                                 if (ret)
 522                                         trans_trigger_run = true;
 523                         }
 524                 } while (trans_trigger_run);
 525         }
 526
 527         return 0;
 528 }
 529
 530 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 531 {
 532         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 533         unsigned btree_id = 0;
 534         int ret = 0;
 535
 536         /*
 537          *
 538          * For a given btree, this algorithm runs insert triggers before
 539          * overwrite triggers: this is so that when extents are being moved
 540          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 541          * they are re-added.
 542          */
 543         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 544                 if (btree_id == BTREE_ID_alloc)
 545                         continue;
 546
 547                 while (btree_id_start < trans->updates + trans->nr_updates &&
 548                        btree_id_start->btree_id < btree_id)
 549                         btree_id_start++;
 550
 551                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 552                 if (ret)
 553                         return ret;
 554         }
 555
 556         trans_for_each_update(trans, i) {
 557                 if (i->btree_id > BTREE_ID_alloc)
 558                         break;
 559                 if (i->btree_id == BTREE_ID_alloc) {
 560                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 561                         if (ret)
 562                                 return ret;
 563                         break;
 564                 }
 565         }
 566
 567         trans_for_each_update(trans, i)
 568                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 569                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 570                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 571
 572         return 0;
 573 }
 574
 575 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 576 {
 577         struct bch_fs *c = trans->c;
 578         struct btree_insert_entry *i;
 579         int ret = 0;
 580
 581         trans_for_each_update(trans, i) {
 582                 /*
 583                  * XXX: synchronization of cached update triggers with gc
 584                  * XXX: synchronization of interior node updates with gc
 585                  */
 586                 BUG_ON(i->cached || i->level);
 587
 588                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 589                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 590                         if (ret)
 591                                 break;
 592                 }
 593         }
 594
 595         return ret;
 596 }
 597
 598 static inline int
 599 bch2_trans_commit_write_locked(struct btree_trans *trans,
 600                                struct btree_insert_entry **stopped_at,
 601                                unsigned long trace_ip)
 602 {
 603         struct bch_fs *c = trans->c;
 604         struct btree_insert_entry *i;
 605         struct btree_trans_commit_hook *h;
 606         unsigned u64s = 0;
 607         bool marking = false;
 608         int ret;
 609
 610         if (race_fault()) {
 611                 trace_trans_restart_fault_inject(trans->fn, trace_ip);
 612                 trans->restarted = true;
 613                 return -EINTR;
 614         }
 615
 616         /*
 617          * Check if the insert will fit in the leaf node with the write lock
 618          * held, otherwise another thread could write the node changing the
 619          * amount of space available:
 620          */
 621
 622         prefetch(&trans->c->journal.flags);
 623
 624         h = trans->hooks;
 625         while (h) {
 626                 ret = h->fn(trans, h);
 627                 if (ret)
 628                         return ret;
 629                 h = h->next;
 630         }
 631
 632         trans_for_each_update(trans, i) {
 633                 /* Multiple inserts might go to same leaf: */
 634                 if (!same_leaf_as_prev(trans, i))
 635                         u64s = 0;
 636
 637                 u64s += i->k->k.u64s;
 638                 ret = !i->cached
 639                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 640                         : btree_key_can_insert_cached(trans, i->path, u64s);
 641                 if (ret) {
 642                         *stopped_at = i;
 643                         return ret;
 644                 }
 645
 646                 if (btree_node_type_needs_gc(i->bkey_type))
 647                         marking = true;
 648
 649                 /*
 650                  * Revalidate before calling mem triggers - XXX, ugly:
 651                  *
 652                  * - successful btree node splits don't cause transaction
 653                  *   restarts and will have invalidated the pointer to the bkey
 654                  *   value
 655                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 656                  *   when it has to resort
 657                  * - btree_key_can_insert_cached() when it has to reallocate
 658                  *
 659                  *   Ugly because we currently have no way to tell if the
 660                  *   pointer's been invalidated, which means it's debatabale
 661                  *   whether we should be stashing the old key at all.
 662                  */
 663                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 664
 665                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 666                         struct bkey_i *j_k =
 667                                 bch2_journal_keys_peek_slot(c, i->btree_id, i->level,
 668                                                             i->k->k.p);
 669
 670                         if (j_k) {
 671                                 i->old_k = j_k->k;
 672                                 i->old_v = &j_k->v;
 673                         }
 674                 }
 675         }
 676
 677         /*
 678          * Don't get journal reservation until after we know insert will
 679          * succeed:
 680          */
 681         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 682                 ret = bch2_trans_journal_res_get(trans,
 683                                 JOURNAL_RES_GET_NONBLOCK);
 684                 if (ret)
 685                         return ret;
 686
 687                 if (unlikely(trans->journal_transaction_names))
 688                         journal_transaction_name(trans);
 689         } else {
 690                 trans->journal_res.seq = c->journal.replay_journal_seq;
 691         }
 692
 693         if (unlikely(trans->extra_journal_entries.nr)) {
 694                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 695                                   trans->extra_journal_entries.data,
 696                                   trans->extra_journal_entries.nr);
 697
 698                 trans->journal_res.offset       += trans->extra_journal_entries.nr;
 699                 trans->journal_res.u64s         -= trans->extra_journal_entries.nr;
 700         }
 701
 702         /*
 703          * Not allowed to fail after we've gotten our journal reservation - we
 704          * have to use it:
 705          */
 706
 707         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 708                 if (bch2_journal_seq_verify)
 709                         trans_for_each_update(trans, i)
 710                                 i->k->k.version.lo = trans->journal_res.seq;
 711                 else if (bch2_inject_invalid_keys)
 712                         trans_for_each_update(trans, i)
 713                                 i->k->k.version = MAX_VERSION;
 714         }
 715
 716         if (trans->fs_usage_deltas &&
 717             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 718                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 719
 720         trans_for_each_update(trans, i)
 721                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 722                         ret = run_one_mem_trigger(trans, i, i->flags);
 723                         if (ret)
 724                                 return ret;
 725                 }
 726
 727         if (unlikely(c->gc_pos.phase)) {
 728                 ret = bch2_trans_commit_run_gc_triggers(trans);
 729                 if  (ret)
 730                         return ret;
 731         }
 732
 733         trans_for_each_update(trans, i)
 734                 do_btree_insert_one(trans, i);
 735
 736         return ret;
 737 }
 738
 739 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 740 {
 741         unsigned l;
 742
 743         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 744                 if (btree_node_read_locked(path, l))
 745                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 746 }
 747
 748 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 749 {
 750         struct btree *b = path_l(path)->b;
 751
 752         do {
 753                 if (path->nodes_locked &&
 754                     path->nodes_locked != path->nodes_intent_locked)
 755                         path_upgrade_readers(trans, path);
 756         } while ((path = prev_btree_path(trans, path)) &&
 757                  path_l(path)->b == b);
 758 }
 759
 760 /*
 761  * Check for nodes that we have both read and intent locks on, and upgrade the
 762  * readers to intent:
 763  */
 764 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 765 {
 766         struct btree_path *path;
 767         unsigned i, nr_read = 0, nr_intent = 0;
 768
 769         trans_for_each_path_inorder(trans, path, i) {
 770                 struct btree_path *next = i + 1 < trans->nr_sorted
 771                         ? trans->paths + trans->sorted[i + 1]
 772                         : NULL;
 773
 774                 if (path->nodes_locked) {
 775                         if (path->nodes_intent_locked)
 776                                 nr_intent++;
 777                         else
 778                                 nr_read++;
 779                 }
 780
 781                 if (!next || path_l(path)->b != path_l(next)->b) {
 782                         if (nr_read && nr_intent)
 783                                 upgrade_readers(trans, path);
 784
 785                         nr_read = nr_intent = 0;
 786                 }
 787         }
 788
 789         bch2_trans_verify_locks(trans);
 790 }
 791
 792 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 793 {
 794         struct btree_path *path;
 795         unsigned i;
 796
 797         trans_for_each_path_inorder(trans, path, i) {
 798                 //if (path == pos)
 799                 //      break;
 800
 801                 if (path->nodes_locked != path->nodes_intent_locked &&
 802                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 803                         return true;
 804         }
 805
 806         return false;
 807 }
 808
 809 static inline int trans_lock_write(struct btree_trans *trans)
 810 {
 811         struct btree_insert_entry *i;
 812
 813         trans_for_each_update(trans, i) {
 814                 if (same_leaf_as_prev(trans, i))
 815                         continue;
 816
 817                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 818                         if (have_conflicting_read_lock(trans, i->path))
 819                                 goto fail;
 820
 821                         btree_node_lock_type(trans, i->path,
 822                                              insert_l(i)->b,
 823                                              i->path->pos, i->level,
 824                                              SIX_LOCK_write, NULL, NULL);
 825                 }
 826
 827                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 828         }
 829
 830         return 0;
 831 fail:
 832         while (--i >= trans->updates) {
 833                 if (same_leaf_as_prev(trans, i))
 834                         continue;
 835
 836                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 837         }
 838
 839         trace_trans_restart_would_deadlock_write(trans->fn);
 840         return btree_trans_restart(trans);
 841 }
 842
 843 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 844 {
 845         struct btree_insert_entry *i;
 846
 847         trans_for_each_update(trans, i)
 848                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 849 }
 850
 851 /*
 852  * Get journal reservation, take write locks, and attempt to do btree update(s):
 853  */
 854 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 855                                        struct btree_insert_entry **stopped_at,
 856                                        unsigned long trace_ip)
 857 {
 858         struct bch_fs *c = trans->c;
 859         struct btree_insert_entry *i;
 860         struct printbuf buf = PRINTBUF;
 861         int ret, u64s_delta = 0;
 862         int rw = (trans->flags & BTREE_INSERT_JOURNAL_REPLAY) ? READ : WRITE;
 863
 864         trans_for_each_update(trans, i) {
 865                 if (bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 866                                       i->bkey_type, rw, &buf)) {
 867                         printbuf_reset(&buf);
 868                         pr_buf(&buf, "invalid bkey on insert from %s -> %ps",
 869                                trans->fn, (void *) i->ip_allocated);
 870                         pr_newline(&buf);
 871                         pr_indent_push(&buf, 2);
 872
 873                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
 874                         pr_newline(&buf);
 875
 876                         bch2_bkey_invalid(c, bkey_i_to_s_c(i->k),
 877                                           i->bkey_type, rw, &buf);
 878
 879                         bch2_trans_inconsistent(trans, "%s", buf.buf);
 880                         printbuf_exit(&buf);
 881                         return -EINVAL;
 882                 }
 883                 btree_insert_entry_checks(trans, i);
 884         }
 885
 886         printbuf_exit(&buf);
 887
 888         trans_for_each_update(trans, i) {
 889                 if (i->cached)
 890                         continue;
 891
 892                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 893                 u64s_delta -= i->old_btree_u64s;
 894
 895                 if (!same_leaf_as_next(trans, i)) {
 896                         if (u64s_delta <= 0) {
 897                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 898                                                         i->level, trans->flags);
 899                                 if (unlikely(ret))
 900                                         return ret;
 901                         }
 902
 903                         u64s_delta = 0;
 904                 }
 905         }
 906
 907         ret = bch2_journal_preres_get(&c->journal,
 908                         &trans->journal_preres, trans->journal_preres_u64s,
 909                         JOURNAL_RES_GET_NONBLOCK|
 910                         (trans->flags & JOURNAL_WATERMARK_MASK));
 911         if (unlikely(ret == -EAGAIN))
 912                 ret = bch2_trans_journal_preres_get_cold(trans,
 913                                                 trans->journal_preres_u64s, trace_ip);
 914         if (unlikely(ret))
 915                 return ret;
 916
 917         normalize_read_intent_locks(trans);
 918
 919         ret = trans_lock_write(trans);
 920         if (unlikely(ret))
 921                 return ret;
 922
 923         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 924
 925         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 926                 bch2_drop_overwrites_from_journal(trans);
 927
 928         trans_for_each_update(trans, i)
 929                 if (!same_leaf_as_prev(trans, i))
 930                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 931                                                         insert_l(i)->b);
 932
 933         if (!ret && trans->journal_pin)
 934                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 935                                      trans->journal_pin, NULL);
 936
 937         /*
 938          * Drop journal reservation after dropping write locks, since dropping
 939          * the journal reservation may kick off a journal write:
 940          */
 941         bch2_journal_res_put(&c->journal, &trans->journal_res);
 942
 943         if (unlikely(ret))
 944                 return ret;
 945
 946         bch2_trans_downgrade(trans);
 947
 948         return 0;
 949 }
 950
 951 static int journal_reclaim_wait_done(struct bch_fs *c)
 952 {
 953         int ret = bch2_journal_error(&c->journal) ?:
 954                 !bch2_btree_key_cache_must_wait(c);
 955
 956         if (!ret)
 957                 journal_reclaim_kick(&c->journal);
 958         return ret;
 959 }
 960
 961 static noinline
 962 int bch2_trans_commit_error(struct btree_trans *trans,
 963                             struct btree_insert_entry *i,
 964                             int ret, unsigned long trace_ip)
 965 {
 966         struct bch_fs *c = trans->c;
 967
 968         switch (ret) {
 969         case BTREE_INSERT_BTREE_NODE_FULL:
 970                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 971                 if (!ret)
 972                         return 0;
 973
 974                 if (ret == -EINTR)
 975                         trace_trans_restart_btree_node_split(trans->fn, trace_ip,
 976                                                 i->btree_id, &i->path->pos);
 977                 break;
 978         case BTREE_INSERT_NEED_MARK_REPLICAS:
 979                 bch2_trans_unlock(trans);
 980
 981                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 982                 if (ret)
 983                         break;
 984
 985                 if (bch2_trans_relock(trans))
 986                         return 0;
 987
 988                 trace_trans_restart_mark_replicas(trans->fn, trace_ip);
 989                 ret = -EINTR;
 990                 break;
 991         case BTREE_INSERT_NEED_JOURNAL_RES:
 992                 bch2_trans_unlock(trans);
 993
 994                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 995                     !(trans->flags & JOURNAL_WATERMARK_reserved)) {
 996                         trans->restarted = true;
 997                         ret = -EAGAIN;
 998                         break;
 999                 }
1000
1001                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
1002                 if (ret)
1003                         break;
1004
1005                 if (bch2_trans_relock(trans))
1006                         return 0;
1007
1008                 trace_trans_restart_journal_res_get(trans->fn, trace_ip);
1009                 ret = -EINTR;
1010                 break;
1011         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
1012                 bch2_trans_unlock(trans);
1013
1014                 trace_trans_blocked_journal_reclaim(trans->fn, trace_ip);
1015
1016                 wait_event_freezable(c->journal.reclaim_wait,
1017                                      (ret = journal_reclaim_wait_done(c)));
1018                 if (ret < 0)
1019                         break;
1020
1021                 if (bch2_trans_relock(trans))
1022                         return 0;
1023
1024                 trace_trans_restart_journal_reclaim(trans->fn, trace_ip);
1025                 ret = -EINTR;
1026                 break;
1027         default:
1028                 BUG_ON(ret >= 0);
1029                 break;
1030         }
1031
1032         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
1033         BUG_ON(ret == -ENOSPC &&
1034                !(trans->flags & BTREE_INSERT_NOWAIT) &&
1035                (trans->flags & BTREE_INSERT_NOFAIL));
1036
1037         return ret;
1038 }
1039
1040 static noinline int
1041 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
1042 {
1043         struct bch_fs *c = trans->c;
1044         int ret;
1045
1046         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
1047             test_bit(BCH_FS_STARTED, &c->flags))
1048                 return -EROFS;
1049
1050         bch2_trans_unlock(trans);
1051
1052         ret = bch2_fs_read_write_early(c);
1053         if (ret)
1054                 return ret;
1055
1056         if (!bch2_trans_relock(trans))
1057                 return -EINTR;
1058
1059         percpu_ref_get(&c->writes);
1060         return 0;
1061 }
1062
1063 /*
1064  * This is for updates done in the early part of fsck - btree_gc - before we've
1065  * gone RW. we only add the new key to the list of keys for journal replay to
1066  * do.
1067  */
1068 static noinline int
1069 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1070 {
1071         struct bch_fs *c = trans->c;
1072         struct btree_insert_entry *i;
1073         int ret = 0;
1074
1075         trans_for_each_update(trans, i) {
1076                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1077                 if (ret)
1078                         break;
1079         }
1080
1081         return ret;
1082 }
1083
1084 int __bch2_trans_commit(struct btree_trans *trans)
1085 {
1086         struct bch_fs *c = trans->c;
1087         struct btree_insert_entry *i = NULL;
1088         unsigned u64s;
1089         int ret = 0;
1090
1091         if (!trans->nr_updates &&
1092             !trans->extra_journal_entries.nr)
1093                 goto out_reset;
1094
1095         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1096                 lockdep_assert_held(&c->gc_lock);
1097
1098         ret = bch2_trans_commit_run_triggers(trans);
1099         if (ret)
1100                 goto out_reset;
1101
1102         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1103                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1104                 goto out_reset;
1105         }
1106
1107         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1108             unlikely(!percpu_ref_tryget(&c->writes))) {
1109                 ret = bch2_trans_commit_get_rw_cold(trans);
1110                 if (ret)
1111                         goto out_reset;
1112         }
1113
1114         EBUG_ON(test_bit(BCH_FS_CLEAN_SHUTDOWN, &c->flags));
1115
1116         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1117
1118         trans->journal_u64s             = trans->extra_journal_entries.nr;
1119         trans->journal_preres_u64s      = 0;
1120
1121         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1122
1123         if (trans->journal_transaction_names)
1124                 trans->journal_u64s += JSET_ENTRY_LOG_U64s;
1125
1126         trans_for_each_update(trans, i) {
1127                 BUG_ON(!i->path->should_be_locked);
1128
1129                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1130                         trace_trans_restart_upgrade(trans->fn, _RET_IP_,
1131                                                     i->btree_id, &i->path->pos);
1132                         ret = btree_trans_restart(trans);
1133                         goto out;
1134                 }
1135
1136                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1137
1138                 u64s = jset_u64s(i->k->k.u64s);
1139                 if (i->cached &&
1140                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1141                         trans->journal_preres_u64s += u64s;
1142                 trans->journal_u64s += u64s;
1143         }
1144
1145         if (trans->extra_journal_res) {
1146                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1147                                 trans->extra_journal_res,
1148                                 (trans->flags & BTREE_INSERT_NOFAIL)
1149                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1150                 if (ret)
1151                         goto err;
1152         }
1153 retry:
1154         BUG_ON(trans->restarted);
1155         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1156
1157         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1158
1159         /* make sure we didn't drop or screw up locks: */
1160         bch2_trans_verify_locks(trans);
1161
1162         if (ret)
1163                 goto err;
1164
1165         trace_transaction_commit(trans->fn, _RET_IP_);
1166 out:
1167         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1168
1169         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1170                 percpu_ref_put(&c->writes);
1171 out_reset:
1172         trans_for_each_update(trans, i)
1173                 bch2_path_put(trans, i->path, true);
1174
1175         trans->extra_journal_res        = 0;
1176         trans->nr_updates               = 0;
1177         trans->hooks                    = NULL;
1178         trans->extra_journal_entries.nr = 0;
1179
1180         if (trans->fs_usage_deltas) {
1181                 trans->fs_usage_deltas->used = 0;
1182                 memset(&trans->fs_usage_deltas->memset_start, 0,
1183                        (void *) &trans->fs_usage_deltas->memset_end -
1184                        (void *) &trans->fs_usage_deltas->memset_start);
1185         }
1186
1187         return ret;
1188 err:
1189         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1190         if (ret)
1191                 goto out;
1192
1193         goto retry;
1194 }
1195
1196 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1197                                           enum btree_id id,
1198                                           struct bpos pos)
1199 {
1200         struct bch_fs *c = trans->c;
1201         struct btree_iter iter;
1202         struct bkey_s_c k;
1203         int ret;
1204
1205         if (!btree_type_has_snapshots(id))
1206                 return 0;
1207
1208         if (!snapshot_t(c, pos.snapshot)->children[0])
1209                 return 0;
1210
1211         bch2_trans_iter_init(trans, &iter, id, pos,
1212                              BTREE_ITER_NOT_EXTENTS|
1213                              BTREE_ITER_ALL_SNAPSHOTS);
1214         while (1) {
1215                 k = bch2_btree_iter_prev(&iter);
1216                 ret = bkey_err(k);
1217                 if (ret)
1218                         break;
1219
1220                 if (!k.k)
1221                         break;
1222
1223                 if (bkey_cmp(pos, k.k->p))
1224                         break;
1225
1226                 if (bch2_snapshot_is_ancestor(c, k.k->p.snapshot, pos.snapshot)) {
1227                         ret = 1;
1228                         break;
1229                 }
1230         }
1231         bch2_trans_iter_exit(trans, &iter);
1232
1233         return ret;
1234 }
1235
1236 int bch2_trans_update_extent(struct btree_trans *trans,
1237                              struct btree_iter *orig_iter,
1238                              struct bkey_i *insert,
1239                              enum btree_update_flags flags)
1240 {
1241         struct bch_fs *c = trans->c;
1242         struct btree_iter iter, update_iter;
1243         struct bpos start = bkey_start_pos(&insert->k);
1244         struct bkey_i *update;
1245         struct bkey_s_c k;
1246         enum btree_id btree_id = orig_iter->btree_id;
1247         int ret = 0, compressed_sectors;
1248
1249         bch2_trans_iter_init(trans, &iter, btree_id, start,
1250                              BTREE_ITER_INTENT|
1251                              BTREE_ITER_WITH_UPDATES|
1252                              BTREE_ITER_NOT_EXTENTS);
1253         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1254         if ((ret = bkey_err(k)))
1255                 goto err;
1256         if (!k.k)
1257                 goto out;
1258
1259         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1260                 /*
1261                  * We can't merge extents if they belong to interior snapshot
1262                  * tree nodes, and there's a snapshot in which one extent is
1263                  * visible and the other is not - i.e. if visibility is
1264                  * different.
1265                  *
1266                  * Instead of checking if visibilitiy of the two extents is
1267                  * different, for now we just check if either has been
1268                  * overwritten:
1269                  */
1270                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1271                 if (ret < 0)
1272                         goto err;
1273                 if (ret)
1274                         goto nomerge1;
1275
1276                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1277                 if (ret < 0)
1278                         goto err;
1279                 if (ret)
1280                         goto nomerge1;
1281
1282                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1283                 if ((ret = PTR_ERR_OR_ZERO(update)))
1284                         goto err;
1285
1286                 bkey_reassemble(update, k);
1287
1288                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1289                         ret = bch2_btree_delete_at(trans, &iter, flags);
1290                         if (ret)
1291                                 goto err;
1292
1293                         insert = update;
1294                         goto next;
1295                 }
1296         }
1297 nomerge1:
1298         ret = 0;
1299         if (!bkey_cmp(k.k->p, start))
1300                 goto next;
1301
1302         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1303                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1304                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1305
1306                 /*
1307                  * If we're going to be splitting a compressed extent, note it
1308                  * so that __bch2_trans_commit() can increase our disk
1309                  * reservation:
1310                  */
1311                 if (((front_split && back_split) ||
1312                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1313                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1314                         trans->extra_journal_res += compressed_sectors;
1315
1316                 if (front_split) {
1317                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1318                         if ((ret = PTR_ERR_OR_ZERO(update)))
1319                                 goto err;
1320
1321                         bkey_reassemble(update, k);
1322
1323                         bch2_cut_back(start, update);
1324
1325                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1326                                              BTREE_ITER_NOT_EXTENTS|
1327                                              BTREE_ITER_ALL_SNAPSHOTS|
1328                                              BTREE_ITER_INTENT);
1329                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1330                                 bch2_trans_update(trans, &update_iter, update,
1331                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1332                                                   flags);
1333                         bch2_trans_iter_exit(trans, &update_iter);
1334
1335                         if (ret)
1336                                 goto err;
1337                 }
1338
1339                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1340                     (front_split || back_split)) {
1341                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1342                         if ((ret = PTR_ERR_OR_ZERO(update)))
1343                                 goto err;
1344
1345                         bkey_reassemble(update, k);
1346
1347                         bch2_cut_front(start, update);
1348                         bch2_cut_back(insert->k.p, update);
1349
1350                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1351                                              BTREE_ITER_NOT_EXTENTS|
1352                                              BTREE_ITER_ALL_SNAPSHOTS|
1353                                              BTREE_ITER_INTENT);
1354                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1355                                 bch2_trans_update(trans, &update_iter, update,
1356                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1357                                                   flags);
1358                         bch2_trans_iter_exit(trans, &update_iter);
1359                         if (ret)
1360                                 goto err;
1361                 }
1362
1363                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1364                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1365                         if ((ret = PTR_ERR_OR_ZERO(update)))
1366                                 goto err;
1367
1368                         bkey_init(&update->k);
1369                         update->k.p = k.k->p;
1370
1371                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1372                                 update->k.p.snapshot = insert->k.p.snapshot;
1373                                 update->k.type = KEY_TYPE_whiteout;
1374                         }
1375
1376                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1377                                              BTREE_ITER_NOT_EXTENTS|
1378                                              BTREE_ITER_INTENT);
1379                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1380                                 bch2_trans_update(trans, &update_iter, update,
1381                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1382                                                   flags);
1383                         bch2_trans_iter_exit(trans, &update_iter);
1384
1385                         if (ret)
1386                                 goto err;
1387                 }
1388
1389                 if (back_split) {
1390                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1391                         if ((ret = PTR_ERR_OR_ZERO(update)))
1392                                 goto err;
1393
1394                         bkey_reassemble(update, k);
1395                         bch2_cut_front(insert->k.p, update);
1396
1397                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1398                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1399                                                   flags);
1400                         if (ret)
1401                                 goto err;
1402                         goto out;
1403                 }
1404 next:
1405                 bch2_btree_iter_advance(&iter);
1406                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1407                 if ((ret = bkey_err(k)))
1408                         goto err;
1409                 if (!k.k)
1410                         goto out;
1411         }
1412
1413         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1414                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1415                 if (ret < 0)
1416                         goto err;
1417                 if (ret)
1418                         goto nomerge2;
1419
1420                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1421                 if (ret < 0)
1422                         goto err;
1423                 if (ret)
1424                         goto nomerge2;
1425
1426                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1427         }
1428 nomerge2:
1429         ret = 0;
1430 out:
1431         if (!bkey_deleted(&insert->k)) {
1432                 /*
1433                  * Rewinding iterators is expensive: get a new one and the one
1434                  * that points to the start of insert will be cloned from:
1435                  */
1436                 bch2_trans_iter_exit(trans, &iter);
1437                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1438                                      BTREE_ITER_NOT_EXTENTS|
1439                                      BTREE_ITER_INTENT);
1440                 ret   = bch2_btree_iter_traverse(&iter) ?:
1441                         bch2_trans_update(trans, &iter, insert, flags);
1442         }
1443 err:
1444         bch2_trans_iter_exit(trans, &iter);
1445
1446         return ret;
1447 }
1448
1449 /*
1450  * When deleting, check if we need to emit a whiteout (because we're overwriting
1451  * something in an ancestor snapshot)
1452  */
1453 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1454                                       enum btree_id btree_id, struct bpos pos)
1455 {
1456         struct btree_iter iter;
1457         struct bkey_s_c k;
1458         u32 snapshot = pos.snapshot;
1459         int ret;
1460
1461         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1462                 return 0;
1463
1464         pos.snapshot++;
1465
1466         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1467                            BTREE_ITER_ALL_SNAPSHOTS|
1468                            BTREE_ITER_NOPRESERVE, k, ret) {
1469                 if (bkey_cmp(k.k->p, pos))
1470                         break;
1471
1472                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1473                                               k.k->p.snapshot)) {
1474                         ret = !bkey_whiteout(k.k);
1475                         break;
1476                 }
1477         }
1478         bch2_trans_iter_exit(trans, &iter);
1479
1480         return ret;
1481 }
1482
1483 static int __must_check
1484 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1485                           struct bkey_i *k, enum btree_update_flags flags)
1486 {
1487         struct bch_fs *c = trans->c;
1488         struct btree_insert_entry *i, n;
1489
1490         BUG_ON(!path->should_be_locked);
1491
1492         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1493         BUG_ON(bpos_cmp(k->k.p, path->pos));
1494
1495         n = (struct btree_insert_entry) {
1496                 .flags          = flags,
1497                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1498                 .btree_id       = path->btree_id,
1499                 .level          = path->level,
1500                 .cached         = path->cached,
1501                 .path           = path,
1502                 .k              = k,
1503                 .ip_allocated   = _RET_IP_,
1504         };
1505
1506 #ifdef CONFIG_BCACHEFS_DEBUG
1507         trans_for_each_update(trans, i)
1508                 BUG_ON(i != trans->updates &&
1509                        btree_insert_entry_cmp(i - 1, i) >= 0);
1510 #endif
1511
1512         /*
1513          * Pending updates are kept sorted: first, find position of new update,
1514          * then delete/trim any updates the new update overwrites:
1515          */
1516         trans_for_each_update(trans, i)
1517                 if (btree_insert_entry_cmp(&n, i) <= 0)
1518                         break;
1519
1520         if (i < trans->updates + trans->nr_updates &&
1521             !btree_insert_entry_cmp(&n, i)) {
1522                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1523
1524                 bch2_path_put(trans, i->path, true);
1525                 i->flags        = n.flags;
1526                 i->cached       = n.cached;
1527                 i->k            = n.k;
1528                 i->path         = n.path;
1529                 i->ip_allocated = n.ip_allocated;
1530         } else {
1531                 array_insert_item(trans->updates, trans->nr_updates,
1532                                   i - trans->updates, n);
1533
1534                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1535                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1536
1537                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1538                         struct bkey_i *j_k =
1539                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
1540
1541                         if (j_k) {
1542                                 i->old_k = j_k->k;
1543                                 i->old_v = &j_k->v;
1544                         }
1545                 }
1546         }
1547
1548         __btree_path_get(n.path, true);
1549         return 0;
1550 }
1551
1552 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1553                                    struct bkey_i *k, enum btree_update_flags flags)
1554 {
1555         struct btree_path *path = iter->update_path ?: iter->path;
1556         struct bkey_cached *ck;
1557         int ret;
1558
1559         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1560                 return bch2_trans_update_extent(trans, iter, k, flags);
1561
1562         if (bkey_deleted(&k->k) &&
1563             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1564             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1565                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1566                 if (unlikely(ret < 0))
1567                         return ret;
1568
1569                 if (ret)
1570                         k->k.type = KEY_TYPE_whiteout;
1571         }
1572
1573         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1574             !path->cached &&
1575             !path->level &&
1576             btree_id_cached(trans->c, path->btree_id)) {
1577                 if (!iter->key_cache_path ||
1578                     !iter->key_cache_path->should_be_locked ||
1579                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1580                         if (!iter->key_cache_path)
1581                                 iter->key_cache_path =
1582                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1583                                                       BTREE_ITER_INTENT|
1584                                                       BTREE_ITER_CACHED, _THIS_IP_);
1585
1586                         iter->key_cache_path =
1587                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1588                                                         iter->flags & BTREE_ITER_INTENT,
1589                                                         _THIS_IP_);
1590
1591                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1592                                                        BTREE_ITER_CACHED);
1593                         if (unlikely(ret))
1594                                 return ret;
1595
1596                         ck = (void *) iter->key_cache_path->l[0].b;
1597
1598                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1599                                 trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
1600                                 btree_trans_restart(trans);
1601                                 return -EINTR;
1602                         }
1603
1604                         iter->key_cache_path->should_be_locked = true;
1605                 }
1606
1607                 path = iter->key_cache_path;
1608         }
1609
1610         return bch2_trans_update_by_path(trans, path, k, flags);
1611 }
1612
1613 void bch2_trans_commit_hook(struct btree_trans *trans,
1614                             struct btree_trans_commit_hook *h)
1615 {
1616         h->next = trans->hooks;
1617         trans->hooks = h;
1618 }
1619
1620 int __bch2_btree_insert(struct btree_trans *trans,
1621                         enum btree_id id, struct bkey_i *k)
1622 {
1623         struct btree_iter iter;
1624         int ret;
1625
1626         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1627                              BTREE_ITER_INTENT);
1628         ret   = bch2_btree_iter_traverse(&iter) ?:
1629                 bch2_trans_update(trans, &iter, k, 0);
1630         bch2_trans_iter_exit(trans, &iter);
1631         return ret;
1632 }
1633
1634 /**
1635  * bch2_btree_insert - insert keys into the extent btree
1636  * @c:                  pointer to struct bch_fs
1637  * @id:                 btree to insert into
1638  * @insert_keys:        list of keys to insert
1639  * @hook:               insert callback
1640  */
1641 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1642                       struct bkey_i *k,
1643                       struct disk_reservation *disk_res,
1644                       u64 *journal_seq, int flags)
1645 {
1646         return bch2_trans_do(c, disk_res, journal_seq, flags,
1647                              __bch2_btree_insert(&trans, id, k));
1648 }
1649
1650 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
1651                                 unsigned len, unsigned update_flags)
1652 {
1653         struct bkey_i *k;
1654
1655         k = bch2_trans_kmalloc(trans, sizeof(*k));
1656         if (IS_ERR(k))
1657                 return PTR_ERR(k);
1658
1659         bkey_init(&k->k);
1660         k->k.p = iter->pos;
1661         bch2_key_resize(&k->k, len);
1662         return bch2_trans_update(trans, iter, k, update_flags);
1663 }
1664
1665 int bch2_btree_delete_at(struct btree_trans *trans,
1666                          struct btree_iter *iter, unsigned update_flags)
1667 {
1668         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
1669 }
1670
1671 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1672                                   struct bpos start, struct bpos end,
1673                                   unsigned update_flags,
1674                                   u64 *journal_seq)
1675 {
1676         struct btree_iter iter;
1677         struct bkey_s_c k;
1678         int ret = 0;
1679
1680         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1681 retry:
1682         while ((bch2_trans_begin(trans),
1683                (k = bch2_btree_iter_peek(&iter)).k) &&
1684                !(ret = bkey_err(k)) &&
1685                bkey_cmp(iter.pos, end) < 0) {
1686                 struct disk_reservation disk_res =
1687                         bch2_disk_reservation_init(trans->c, 0);
1688                 struct bkey_i delete;
1689
1690                 bkey_init(&delete.k);
1691
1692                 /*
1693                  * This could probably be more efficient for extents:
1694                  */
1695
1696                 /*
1697                  * For extents, iter.pos won't necessarily be the same as
1698                  * bkey_start_pos(k.k) (for non extents they always will be the
1699                  * same). It's important that we delete starting from iter.pos
1700                  * because the range we want to delete could start in the middle
1701                  * of k.
1702                  *
1703                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1704                  * bkey_start_pos(k.k)).
1705                  */
1706                 delete.k.p = iter.pos;
1707
1708                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1709                         unsigned max_sectors =
1710                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1711
1712                         /* create the biggest key we can */
1713                         bch2_key_resize(&delete.k, max_sectors);
1714                         bch2_cut_back(end, &delete);
1715
1716                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1717                         if (ret)
1718                                 break;
1719                 }
1720
1721                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
1722                         bch2_trans_commit(trans, &disk_res, journal_seq,
1723                                           BTREE_INSERT_NOFAIL);
1724                 bch2_disk_reservation_put(trans->c, &disk_res);
1725                 if (ret)
1726                         break;
1727         }
1728
1729         if (ret == -EINTR) {
1730                 ret = 0;
1731                 goto retry;
1732         }
1733
1734         bch2_trans_iter_exit(trans, &iter);
1735         return ret;
1736 }
1737
1738 /*
1739  * bch_btree_delete_range - delete everything within a given range
1740  *
1741  * Range is a half open interval - [start, end)
1742  */
1743 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1744                             struct bpos start, struct bpos end,
1745                             unsigned update_flags,
1746                             u64 *journal_seq)
1747 {
1748         return bch2_trans_do(c, NULL, journal_seq, 0,
1749                              bch2_btree_delete_range_trans(&trans, id, start, end,
1750                                                            update_flags, journal_seq));
1751 }
1752
1753 int bch2_trans_log_msg(struct btree_trans *trans, const char *msg)
1754 {
1755         unsigned len = strlen(msg);
1756         unsigned u64s = DIV_ROUND_UP(len, sizeof(u64));
1757         struct jset_entry_log *l;
1758         int ret;
1759
1760         ret = darray_make_room(&trans->extra_journal_entries, jset_u64s(u64s));
1761         if (ret)
1762                 return ret;
1763
1764         l = (void *) &darray_top(trans->extra_journal_entries);
1765         l->entry.u64s           = cpu_to_le16(u64s);
1766         l->entry.btree_id       = 0;
1767         l->entry.level          = 1;
1768         l->entry.type           = BCH_JSET_ENTRY_log;
1769         l->entry.pad[0]         = 0;
1770         l->entry.pad[1]         = 0;
1771         l->entry.pad[2]         = 0;
1772         memcpy(l->d, msg, len);
1773         while (len & 7)
1774                 l->d[len++] = '\0';
1775
1776         trans->extra_journal_entries.nr += jset_u64s(u64s);
1777         return 0;
1778 }