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         if (trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 299                 flags |= JOURNAL_RES_GET_RESERVED;
 300
 301         ret = bch2_journal_res_get(&c->journal, &trans->journal_res,
 302                                    trans->journal_u64s, flags);
 303
 304         return ret == -EAGAIN ? BTREE_INSERT_NEED_JOURNAL_RES : ret;
 305 }
 306
 307 #define JSET_ENTRY_LOG_U64s             4
 308
 309 static noinline void journal_transaction_name(struct btree_trans *trans)
 310 {
 311         struct bch_fs *c = trans->c;
 312         struct jset_entry *entry = journal_res_entry(&c->journal, &trans->journal_res);
 313         struct jset_entry_log *l = container_of(entry, struct jset_entry_log, entry);
 314         unsigned u64s = JSET_ENTRY_LOG_U64s - 1;
 315         unsigned b, buflen = u64s * sizeof(u64);
 316
 317         l->entry.u64s           = cpu_to_le16(u64s);
 318         l->entry.btree_id       = 0;
 319         l->entry.level          = 0;
 320         l->entry.type           = BCH_JSET_ENTRY_log;
 321         l->entry.pad[0]         = 0;
 322         l->entry.pad[1]         = 0;
 323         l->entry.pad[2]         = 0;
 324         b = min_t(unsigned, strlen(trans->fn), buflen);
 325         memcpy(l->d, trans->fn, b);
 326         while (b < buflen)
 327                 l->d[b++] = '\0';
 328
 329         trans->journal_res.offset       += JSET_ENTRY_LOG_U64s;
 330         trans->journal_res.u64s         -= JSET_ENTRY_LOG_U64s;
 331 }
 332
 333 static inline enum btree_insert_ret
 334 btree_key_can_insert(struct btree_trans *trans,
 335                      struct btree *b,
 336                      unsigned u64s)
 337 {
 338         struct bch_fs *c = trans->c;
 339
 340         if (!bch2_btree_node_insert_fits(c, b, u64s))
 341                 return BTREE_INSERT_BTREE_NODE_FULL;
 342
 343         return BTREE_INSERT_OK;
 344 }
 345
 346 static enum btree_insert_ret
 347 btree_key_can_insert_cached(struct btree_trans *trans,
 348                             struct btree_path *path,
 349                             unsigned u64s)
 350 {
 351         struct bch_fs *c = trans->c;
 352         struct bkey_cached *ck = (void *) path->l[0].b;
 353         unsigned new_u64s;
 354         struct bkey_i *new_k;
 355
 356         EBUG_ON(path->level);
 357
 358         if (!test_bit(BKEY_CACHED_DIRTY, &ck->flags) &&
 359             bch2_btree_key_cache_must_wait(c) &&
 360             !(trans->flags & BTREE_INSERT_JOURNAL_RECLAIM))
 361                 return BTREE_INSERT_NEED_JOURNAL_RECLAIM;
 362
 363         /*
 364          * bch2_varint_decode can read past the end of the buffer by at most 7
 365          * bytes (it won't be used):
 366          */
 367         u64s += 1;
 368
 369         if (u64s <= ck->u64s)
 370                 return BTREE_INSERT_OK;
 371
 372         new_u64s        = roundup_pow_of_two(u64s);
 373         new_k           = krealloc(ck->k, new_u64s * sizeof(u64), GFP_NOFS);
 374         if (!new_k) {
 375                 bch_err(c, "error allocating memory for key cache key, btree %s u64s %u",
 376                         bch2_btree_ids[path->btree_id], new_u64s);
 377                 return -ENOMEM;
 378         }
 379
 380         ck->u64s        = new_u64s;
 381         ck->k           = new_k;
 382         /*
 383          * Keys returned by peek() are no longer valid pointers, so we need a
 384          * transaction restart:
 385          */
 386         trace_trans_restart_key_cache_key_realloced(trans->fn, _RET_IP_,
 387                                              path->btree_id, &path->pos);
 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         struct bkey_s_c old = { &i->old_k, i->old_v };
 463         int ret = 0;
 464
 465         if ((i->flags & BTREE_TRIGGER_NORUN) ||
 466             !(BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)))
 467                 return 0;
 468
 469         if (!overwrite) {
 470                 if (i->insert_trigger_run)
 471                         return 0;
 472
 473                 BUG_ON(i->overwrite_trigger_run);
 474                 i->insert_trigger_run = true;
 475         } else {
 476                 if (i->overwrite_trigger_run)
 477                         return 0;
 478
 479                 BUG_ON(!i->insert_trigger_run);
 480                 i->overwrite_trigger_run = true;
 481         }
 482
 483         if (overwrite) {
 484                 ret = bch2_trans_mark_old(trans, old, i->flags);
 485         } else if (bch2_bkey_ops[old.k->type].trans_trigger ==
 486                    bch2_bkey_ops[i->k->k.type].trans_trigger &&
 487             ((1U << old.k->type) & BTREE_TRIGGER_WANTS_OLD_AND_NEW)) {
 488                 i->overwrite_trigger_run = true;
 489                 ret = bch2_trans_mark_key(trans, old, i->k,
 490                                 BTREE_TRIGGER_INSERT|BTREE_TRIGGER_OVERWRITE|i->flags);
 491         } else {
 492                 ret = bch2_trans_mark_new(trans, i->k, i->flags);
 493         }
 494
 495         if (ret == -EINTR)
 496                 trace_trans_restart_mark(trans->fn, _RET_IP_,
 497                                          i->btree_id, &i->path->pos);
 498         return ret ?: 1;
 499 }
 500
 501 static int run_btree_triggers(struct btree_trans *trans, enum btree_id btree_id,
 502                               struct btree_insert_entry *btree_id_start)
 503 {
 504         struct btree_insert_entry *i;
 505         bool trans_trigger_run;
 506         int ret, overwrite;
 507
 508         for (overwrite = 0; overwrite < 2; overwrite++) {
 509
 510                 /*
 511                  * Running triggers will append more updates to the list of updates as
 512                  * we're walking it:
 513                  */
 514                 do {
 515                         trans_trigger_run = false;
 516
 517                         for (i = btree_id_start;
 518                              i < trans->updates + trans->nr_updates && i->btree_id <= btree_id;
 519                              i++) {
 520                                 if (i->btree_id != btree_id)
 521                                         continue;
 522
 523                                 ret = run_one_trans_trigger(trans, i, overwrite);
 524                                 if (ret < 0)
 525                                         return ret;
 526                                 if (ret)
 527                                         trans_trigger_run = true;
 528                         }
 529                 } while (trans_trigger_run);
 530         }
 531
 532         return 0;
 533 }
 534
 535 static int bch2_trans_commit_run_triggers(struct btree_trans *trans)
 536 {
 537         struct btree_insert_entry *i = NULL, *btree_id_start = trans->updates;
 538         unsigned btree_id = 0;
 539         int ret = 0;
 540
 541         /*
 542          *
 543          * For a given btree, this algorithm runs insert triggers before
 544          * overwrite triggers: this is so that when extents are being moved
 545          * (e.g. by FALLOCATE_FL_INSERT_RANGE), we don't drop references before
 546          * they are re-added.
 547          */
 548         for (btree_id = 0; btree_id < BTREE_ID_NR; btree_id++) {
 549                 if (btree_id == BTREE_ID_alloc)
 550                         continue;
 551
 552                 while (btree_id_start < trans->updates + trans->nr_updates &&
 553                        btree_id_start->btree_id < btree_id)
 554                         btree_id_start++;
 555
 556                 ret = run_btree_triggers(trans, btree_id, btree_id_start);
 557                 if (ret)
 558                         return ret;
 559         }
 560
 561         trans_for_each_update(trans, i) {
 562                 if (i->btree_id > BTREE_ID_alloc)
 563                         break;
 564                 if (i->btree_id == BTREE_ID_alloc) {
 565                         ret = run_btree_triggers(trans, BTREE_ID_alloc, i);
 566                         if (ret)
 567                                 return ret;
 568                         break;
 569                 }
 570         }
 571
 572         trans_for_each_update(trans, i)
 573                 BUG_ON(!(i->flags & BTREE_TRIGGER_NORUN) &&
 574                        (BTREE_NODE_TYPE_HAS_TRANS_TRIGGERS & (1U << i->bkey_type)) &&
 575                        (!i->insert_trigger_run || !i->overwrite_trigger_run));
 576
 577         return 0;
 578 }
 579
 580 static noinline int bch2_trans_commit_run_gc_triggers(struct btree_trans *trans)
 581 {
 582         struct bch_fs *c = trans->c;
 583         struct btree_insert_entry *i;
 584         int ret = 0;
 585
 586         trans_for_each_update(trans, i) {
 587                 /*
 588                  * XXX: synchronization of cached update triggers with gc
 589                  * XXX: synchronization of interior node updates with gc
 590                  */
 591                 BUG_ON(i->cached || i->level);
 592
 593                 if (gc_visited(c, gc_pos_btree_node(insert_l(i)->b))) {
 594                         ret = run_one_mem_trigger(trans, i, i->flags|BTREE_TRIGGER_GC);
 595                         if (ret)
 596                                 break;
 597                 }
 598         }
 599
 600         return ret;
 601 }
 602
 603 static inline int
 604 bch2_trans_commit_write_locked(struct btree_trans *trans,
 605                                struct btree_insert_entry **stopped_at,
 606                                unsigned long trace_ip)
 607 {
 608         struct bch_fs *c = trans->c;
 609         struct btree_insert_entry *i;
 610         struct btree_trans_commit_hook *h;
 611         unsigned u64s = 0;
 612         bool marking = false;
 613         int ret;
 614
 615         if (race_fault()) {
 616                 trace_trans_restart_fault_inject(trans->fn, trace_ip);
 617                 trans->restarted = true;
 618                 return -EINTR;
 619         }
 620
 621         /*
 622          * Check if the insert will fit in the leaf node with the write lock
 623          * held, otherwise another thread could write the node changing the
 624          * amount of space available:
 625          */
 626
 627         prefetch(&trans->c->journal.flags);
 628
 629         h = trans->hooks;
 630         while (h) {
 631                 ret = h->fn(trans, h);
 632                 if (ret)
 633                         return ret;
 634                 h = h->next;
 635         }
 636
 637         trans_for_each_update(trans, i) {
 638                 /* Multiple inserts might go to same leaf: */
 639                 if (!same_leaf_as_prev(trans, i))
 640                         u64s = 0;
 641
 642                 u64s += i->k->k.u64s;
 643                 ret = !i->cached
 644                         ? btree_key_can_insert(trans, insert_l(i)->b, u64s)
 645                         : btree_key_can_insert_cached(trans, i->path, u64s);
 646                 if (ret) {
 647                         *stopped_at = i;
 648                         return ret;
 649                 }
 650
 651                 if (btree_node_type_needs_gc(i->bkey_type))
 652                         marking = true;
 653
 654                 /*
 655                  * Revalidate before calling mem triggers - XXX, ugly:
 656                  *
 657                  * - successful btree node splits don't cause transaction
 658                  *   restarts and will have invalidated the pointer to the bkey
 659                  *   value
 660                  * - btree_node_lock_for_insert() -> btree_node_prep_for_write()
 661                  *   when it has to resort
 662                  * - btree_key_can_insert_cached() when it has to reallocate
 663                  *
 664                  *   Ugly because we currently have no way to tell if the
 665                  *   pointer's been invalidated, which means it's debatabale
 666                  *   whether we should be stashing the old key at all.
 667                  */
 668                 i->old_v = bch2_btree_path_peek_slot(i->path, &i->old_k).v;
 669
 670                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
 671                         struct bkey_i *j_k =
 672                                 bch2_journal_keys_peek(c, i->btree_id, i->level, i->k->k.p);
 673
 674                         if (j_k && !bpos_cmp(j_k->k.p, i->k->k.p)) {
 675                                 i->old_k = j_k->k;
 676                                 i->old_v = &j_k->v;
 677                         }
 678                 }
 679         }
 680
 681         /*
 682          * Don't get journal reservation until after we know insert will
 683          * succeed:
 684          */
 685         if (likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY))) {
 686                 ret = bch2_trans_journal_res_get(trans,
 687                                 JOURNAL_RES_GET_NONBLOCK);
 688                 if (ret)
 689                         return ret;
 690
 691                 if (unlikely(trans->journal_transaction_names))
 692                         journal_transaction_name(trans);
 693         } else {
 694                 trans->journal_res.seq = c->journal.replay_journal_seq;
 695         }
 696
 697         if (unlikely(trans->extra_journal_entry_u64s)) {
 698                 memcpy_u64s_small(journal_res_entry(&c->journal, &trans->journal_res),
 699                                   trans->extra_journal_entries,
 700                                   trans->extra_journal_entry_u64s);
 701
 702                 trans->journal_res.offset       += trans->extra_journal_entry_u64s;
 703                 trans->journal_res.u64s         -= trans->extra_journal_entry_u64s;
 704         }
 705
 706         /*
 707          * Not allowed to fail after we've gotten our journal reservation - we
 708          * have to use it:
 709          */
 710
 711         if (!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)) {
 712                 if (bch2_journal_seq_verify)
 713                         trans_for_each_update(trans, i)
 714                                 i->k->k.version.lo = trans->journal_res.seq;
 715                 else if (bch2_inject_invalid_keys)
 716                         trans_for_each_update(trans, i)
 717                                 i->k->k.version = MAX_VERSION;
 718         }
 719
 720         if (trans->fs_usage_deltas &&
 721             bch2_trans_fs_usage_apply(trans, trans->fs_usage_deltas))
 722                 return BTREE_INSERT_NEED_MARK_REPLICAS;
 723
 724         trans_for_each_update(trans, i)
 725                 if (BTREE_NODE_TYPE_HAS_MEM_TRIGGERS & (1U << i->bkey_type)) {
 726                         ret = run_one_mem_trigger(trans, i, i->flags);
 727                         if (ret)
 728                                 return ret;
 729                 }
 730
 731         if (unlikely(c->gc_pos.phase)) {
 732                 ret = bch2_trans_commit_run_gc_triggers(trans);
 733                 if  (ret)
 734                         return ret;
 735         }
 736
 737         trans_for_each_update(trans, i)
 738                 do_btree_insert_one(trans, i);
 739
 740         return ret;
 741 }
 742
 743 static inline void path_upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 744 {
 745         unsigned l;
 746
 747         for (l = 0; l < BTREE_MAX_DEPTH; l++)
 748                 if (btree_node_read_locked(path, l))
 749                         BUG_ON(!bch2_btree_node_upgrade(trans, path, l));
 750 }
 751
 752 static inline void upgrade_readers(struct btree_trans *trans, struct btree_path *path)
 753 {
 754         struct btree *b = path_l(path)->b;
 755
 756         do {
 757                 if (path->nodes_locked &&
 758                     path->nodes_locked != path->nodes_intent_locked)
 759                         path_upgrade_readers(trans, path);
 760         } while ((path = prev_btree_path(trans, path)) &&
 761                  path_l(path)->b == b);
 762 }
 763
 764 /*
 765  * Check for nodes that we have both read and intent locks on, and upgrade the
 766  * readers to intent:
 767  */
 768 static inline void normalize_read_intent_locks(struct btree_trans *trans)
 769 {
 770         struct btree_path *path;
 771         unsigned i, nr_read = 0, nr_intent = 0;
 772
 773         trans_for_each_path_inorder(trans, path, i) {
 774                 struct btree_path *next = i + 1 < trans->nr_sorted
 775                         ? trans->paths + trans->sorted[i + 1]
 776                         : NULL;
 777
 778                 if (path->nodes_locked) {
 779                         if (path->nodes_intent_locked)
 780                                 nr_intent++;
 781                         else
 782                                 nr_read++;
 783                 }
 784
 785                 if (!next || path_l(path)->b != path_l(next)->b) {
 786                         if (nr_read && nr_intent)
 787                                 upgrade_readers(trans, path);
 788
 789                         nr_read = nr_intent = 0;
 790                 }
 791         }
 792
 793         bch2_trans_verify_locks(trans);
 794 }
 795
 796 static inline bool have_conflicting_read_lock(struct btree_trans *trans, struct btree_path *pos)
 797 {
 798         struct btree_path *path;
 799         unsigned i;
 800
 801         trans_for_each_path_inorder(trans, path, i) {
 802                 //if (path == pos)
 803                 //      break;
 804
 805                 if (path->nodes_locked != path->nodes_intent_locked &&
 806                     !bch2_btree_path_upgrade(trans, path, path->level + 1))
 807                         return true;
 808         }
 809
 810         return false;
 811 }
 812
 813 static inline int trans_lock_write(struct btree_trans *trans)
 814 {
 815         struct btree_insert_entry *i;
 816
 817         trans_for_each_update(trans, i) {
 818                 if (same_leaf_as_prev(trans, i))
 819                         continue;
 820
 821                 if (!six_trylock_write(&insert_l(i)->b->c.lock)) {
 822                         if (have_conflicting_read_lock(trans, i->path))
 823                                 goto fail;
 824
 825                         btree_node_lock_type(trans, i->path,
 826                                              insert_l(i)->b,
 827                                              i->path->pos, i->level,
 828                                              SIX_LOCK_write, NULL, NULL);
 829                 }
 830
 831                 bch2_btree_node_prep_for_write(trans, i->path, insert_l(i)->b);
 832         }
 833
 834         return 0;
 835 fail:
 836         while (--i >= trans->updates) {
 837                 if (same_leaf_as_prev(trans, i))
 838                         continue;
 839
 840                 bch2_btree_node_unlock_write_inlined(trans, i->path, insert_l(i)->b);
 841         }
 842
 843         trace_trans_restart_would_deadlock_write(trans->fn);
 844         return btree_trans_restart(trans);
 845 }
 846
 847 static noinline void bch2_drop_overwrites_from_journal(struct btree_trans *trans)
 848 {
 849         struct btree_insert_entry *i;
 850
 851         trans_for_each_update(trans, i)
 852                 bch2_journal_key_overwritten(trans->c, i->btree_id, i->level, i->k->k.p);
 853 }
 854
 855 /*
 856  * Get journal reservation, take write locks, and attempt to do btree update(s):
 857  */
 858 static inline int do_bch2_trans_commit(struct btree_trans *trans,
 859                                        struct btree_insert_entry **stopped_at,
 860                                        unsigned long trace_ip)
 861 {
 862         struct bch_fs *c = trans->c;
 863         struct btree_insert_entry *i;
 864         int ret, u64s_delta = 0;
 865
 866         trans_for_each_update(trans, i) {
 867                 const char *invalid = bch2_bkey_invalid(c,
 868                                 bkey_i_to_s_c(i->k), i->bkey_type);
 869                 if (invalid) {
 870                         struct printbuf buf = PRINTBUF;
 871
 872                         bch2_bkey_val_to_text(&buf, c, bkey_i_to_s_c(i->k));
 873                         bch2_fs_fatal_error(c, "invalid bkey %s on insert from %s -> %ps: %s\n",
 874                                             buf.buf, trans->fn, (void *) i->ip_allocated, invalid);
 875                         printbuf_exit(&buf);
 876                         return -EINVAL;
 877                 }
 878                 btree_insert_entry_checks(trans, i);
 879         }
 880
 881         trans_for_each_update(trans, i) {
 882                 if (i->cached)
 883                         continue;
 884
 885                 u64s_delta += !bkey_deleted(&i->k->k) ? i->k->k.u64s : 0;
 886                 u64s_delta -= i->old_btree_u64s;
 887
 888                 if (!same_leaf_as_next(trans, i)) {
 889                         if (u64s_delta <= 0) {
 890                                 ret = bch2_foreground_maybe_merge(trans, i->path,
 891                                                         i->level, trans->flags);
 892                                 if (unlikely(ret))
 893                                         return ret;
 894                         }
 895
 896                         u64s_delta = 0;
 897                 }
 898         }
 899
 900         ret = bch2_journal_preres_get(&c->journal,
 901                         &trans->journal_preres, trans->journal_preres_u64s,
 902                         JOURNAL_RES_GET_NONBLOCK|
 903                         ((trans->flags & BTREE_INSERT_JOURNAL_RESERVED)
 904                          ? JOURNAL_RES_GET_RESERVED : 0));
 905         if (unlikely(ret == -EAGAIN))
 906                 ret = bch2_trans_journal_preres_get_cold(trans,
 907                                                 trans->journal_preres_u64s, trace_ip);
 908         if (unlikely(ret))
 909                 return ret;
 910
 911         normalize_read_intent_locks(trans);
 912
 913         ret = trans_lock_write(trans);
 914         if (unlikely(ret))
 915                 return ret;
 916
 917         ret = bch2_trans_commit_write_locked(trans, stopped_at, trace_ip);
 918
 919         if (!ret && unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags)))
 920                 bch2_drop_overwrites_from_journal(trans);
 921
 922         trans_for_each_update(trans, i)
 923                 if (!same_leaf_as_prev(trans, i))
 924                         bch2_btree_node_unlock_write_inlined(trans, i->path,
 925                                                         insert_l(i)->b);
 926
 927         if (!ret && trans->journal_pin)
 928                 bch2_journal_pin_add(&c->journal, trans->journal_res.seq,
 929                                      trans->journal_pin, NULL);
 930
 931         /*
 932          * Drop journal reservation after dropping write locks, since dropping
 933          * the journal reservation may kick off a journal write:
 934          */
 935         bch2_journal_res_put(&c->journal, &trans->journal_res);
 936
 937         if (unlikely(ret))
 938                 return ret;
 939
 940         bch2_trans_downgrade(trans);
 941
 942         return 0;
 943 }
 944
 945 static int journal_reclaim_wait_done(struct bch_fs *c)
 946 {
 947         int ret = bch2_journal_error(&c->journal) ?:
 948                 !bch2_btree_key_cache_must_wait(c);
 949
 950         if (!ret)
 951                 journal_reclaim_kick(&c->journal);
 952         return ret;
 953 }
 954
 955 static noinline
 956 int bch2_trans_commit_error(struct btree_trans *trans,
 957                             struct btree_insert_entry *i,
 958                             int ret, unsigned long trace_ip)
 959 {
 960         struct bch_fs *c = trans->c;
 961
 962         switch (ret) {
 963         case BTREE_INSERT_BTREE_NODE_FULL:
 964                 ret = bch2_btree_split_leaf(trans, i->path, trans->flags);
 965                 if (!ret)
 966                         return 0;
 967
 968                 if (ret == -EINTR)
 969                         trace_trans_restart_btree_node_split(trans->fn, trace_ip,
 970                                                 i->btree_id, &i->path->pos);
 971                 break;
 972         case BTREE_INSERT_NEED_MARK_REPLICAS:
 973                 bch2_trans_unlock(trans);
 974
 975                 ret = bch2_replicas_delta_list_mark(c, trans->fs_usage_deltas);
 976                 if (ret)
 977                         break;
 978
 979                 if (bch2_trans_relock(trans))
 980                         return 0;
 981
 982                 trace_trans_restart_mark_replicas(trans->fn, trace_ip);
 983                 ret = -EINTR;
 984                 break;
 985         case BTREE_INSERT_NEED_JOURNAL_RES:
 986                 bch2_trans_unlock(trans);
 987
 988                 if ((trans->flags & BTREE_INSERT_JOURNAL_RECLAIM) &&
 989                     !(trans->flags & BTREE_INSERT_JOURNAL_RESERVED)) {
 990                         trans->restarted = true;
 991                         ret = -EAGAIN;
 992                         break;
 993                 }
 994
 995                 ret = bch2_trans_journal_res_get(trans, JOURNAL_RES_GET_CHECK);
 996                 if (ret)
 997                         break;
 998
 999                 if (bch2_trans_relock(trans))
1000                         return 0;
1001
1002                 trace_trans_restart_journal_res_get(trans->fn, trace_ip);
1003                 ret = -EINTR;
1004                 break;
1005         case BTREE_INSERT_NEED_JOURNAL_RECLAIM:
1006                 bch2_trans_unlock(trans);
1007
1008                 trace_trans_blocked_journal_reclaim(trans->fn, trace_ip);
1009
1010                 wait_event_freezable(c->journal.reclaim_wait,
1011                                      (ret = journal_reclaim_wait_done(c)));
1012                 if (ret < 0)
1013                         break;
1014
1015                 if (bch2_trans_relock(trans))
1016                         return 0;
1017
1018                 trace_trans_restart_journal_reclaim(trans->fn, trace_ip);
1019                 ret = -EINTR;
1020                 break;
1021         default:
1022                 BUG_ON(ret >= 0);
1023                 break;
1024         }
1025
1026         BUG_ON((ret == EINTR || ret == -EAGAIN) && !trans->restarted);
1027         BUG_ON(ret == -ENOSPC &&
1028                !(trans->flags & BTREE_INSERT_NOWAIT) &&
1029                (trans->flags & BTREE_INSERT_NOFAIL));
1030
1031         return ret;
1032 }
1033
1034 static noinline int
1035 bch2_trans_commit_get_rw_cold(struct btree_trans *trans)
1036 {
1037         struct bch_fs *c = trans->c;
1038         int ret;
1039
1040         if (likely(!(trans->flags & BTREE_INSERT_LAZY_RW)) ||
1041             test_bit(BCH_FS_STARTED, &c->flags))
1042                 return -EROFS;
1043
1044         bch2_trans_unlock(trans);
1045
1046         ret = bch2_fs_read_write_early(c);
1047         if (ret)
1048                 return ret;
1049
1050         if (!bch2_trans_relock(trans))
1051                 return -EINTR;
1052
1053         percpu_ref_get(&c->writes);
1054         return 0;
1055 }
1056
1057 /*
1058  * This is for updates done in the early part of fsck - btree_gc - before we've
1059  * gone RW. we only add the new key to the list of keys for journal replay to
1060  * do.
1061  */
1062 static noinline int
1063 do_bch2_trans_commit_to_journal_replay(struct btree_trans *trans)
1064 {
1065         struct bch_fs *c = trans->c;
1066         struct btree_insert_entry *i;
1067         int ret = 0;
1068
1069         trans_for_each_update(trans, i) {
1070                 ret = bch2_journal_key_insert(c, i->btree_id, i->level, i->k);
1071                 if (ret)
1072                         break;
1073         }
1074
1075         return ret;
1076 }
1077
1078 int __bch2_trans_commit(struct btree_trans *trans)
1079 {
1080         struct bch_fs *c = trans->c;
1081         struct btree_insert_entry *i = NULL;
1082         unsigned u64s;
1083         int ret = 0;
1084
1085         if (!trans->nr_updates &&
1086             !trans->extra_journal_entry_u64s)
1087                 goto out_reset;
1088
1089         if (trans->flags & BTREE_INSERT_GC_LOCK_HELD)
1090                 lockdep_assert_held(&c->gc_lock);
1091
1092         ret = bch2_trans_commit_run_triggers(trans);
1093         if (ret)
1094                 goto out_reset;
1095
1096         if (unlikely(!test_bit(BCH_FS_MAY_GO_RW, &c->flags))) {
1097                 ret = do_bch2_trans_commit_to_journal_replay(trans);
1098                 goto out_reset;
1099         }
1100
1101         if (!(trans->flags & BTREE_INSERT_NOCHECK_RW) &&
1102             unlikely(!percpu_ref_tryget(&c->writes))) {
1103                 ret = bch2_trans_commit_get_rw_cold(trans);
1104                 if (ret)
1105                         goto out_reset;
1106         }
1107
1108         memset(&trans->journal_preres, 0, sizeof(trans->journal_preres));
1109
1110         trans->journal_u64s             = trans->extra_journal_entry_u64s;
1111         trans->journal_preres_u64s      = 0;
1112
1113         trans->journal_transaction_names = READ_ONCE(c->opts.journal_transaction_names);
1114
1115         if (trans->journal_transaction_names)
1116                 trans->journal_u64s += JSET_ENTRY_LOG_U64s;
1117
1118         trans_for_each_update(trans, i) {
1119                 BUG_ON(!i->path->should_be_locked);
1120
1121                 if (unlikely(!bch2_btree_path_upgrade(trans, i->path, i->level + 1))) {
1122                         trace_trans_restart_upgrade(trans->fn, _RET_IP_,
1123                                                     i->btree_id, &i->path->pos);
1124                         ret = btree_trans_restart(trans);
1125                         goto out;
1126                 }
1127
1128                 BUG_ON(!btree_node_intent_locked(i->path, i->level));
1129
1130                 u64s = jset_u64s(i->k->k.u64s);
1131                 if (i->cached &&
1132                     likely(!(trans->flags & BTREE_INSERT_JOURNAL_REPLAY)))
1133                         trans->journal_preres_u64s += u64s;
1134                 trans->journal_u64s += u64s;
1135         }
1136
1137         if (trans->extra_journal_res) {
1138                 ret = bch2_disk_reservation_add(c, trans->disk_res,
1139                                 trans->extra_journal_res,
1140                                 (trans->flags & BTREE_INSERT_NOFAIL)
1141                                 ? BCH_DISK_RESERVATION_NOFAIL : 0);
1142                 if (ret)
1143                         goto err;
1144         }
1145 retry:
1146         BUG_ON(trans->restarted);
1147         memset(&trans->journal_res, 0, sizeof(trans->journal_res));
1148
1149         ret = do_bch2_trans_commit(trans, &i, _RET_IP_);
1150
1151         /* make sure we didn't drop or screw up locks: */
1152         bch2_trans_verify_locks(trans);
1153
1154         if (ret)
1155                 goto err;
1156 out:
1157         bch2_journal_preres_put(&c->journal, &trans->journal_preres);
1158
1159         if (likely(!(trans->flags & BTREE_INSERT_NOCHECK_RW)))
1160                 percpu_ref_put(&c->writes);
1161 out_reset:
1162         trans_for_each_update(trans, i)
1163                 bch2_path_put(trans, i->path, true);
1164
1165         trans->extra_journal_res        = 0;
1166         trans->nr_updates               = 0;
1167         trans->hooks                    = NULL;
1168         trans->extra_journal_entries    = NULL;
1169         trans->extra_journal_entry_u64s = 0;
1170
1171         if (trans->fs_usage_deltas) {
1172                 trans->fs_usage_deltas->used = 0;
1173                 memset(&trans->fs_usage_deltas->memset_start, 0,
1174                        (void *) &trans->fs_usage_deltas->memset_end -
1175                        (void *) &trans->fs_usage_deltas->memset_start);
1176         }
1177
1178         return ret;
1179 err:
1180         ret = bch2_trans_commit_error(trans, i, ret, _RET_IP_);
1181         if (ret)
1182                 goto out;
1183
1184         goto retry;
1185 }
1186
1187 static int check_pos_snapshot_overwritten(struct btree_trans *trans,
1188                                           enum btree_id id,
1189                                           struct bpos pos)
1190 {
1191         struct bch_fs *c = trans->c;
1192         struct btree_iter iter;
1193         struct bkey_s_c k;
1194         int ret;
1195
1196         if (!btree_type_has_snapshots(id))
1197                 return 0;
1198
1199         if (!snapshot_t(c, pos.snapshot)->children[0])
1200                 return 0;
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_cmp(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 int bch2_trans_update_extent(struct btree_trans *trans,
1228                              struct btree_iter *orig_iter,
1229                              struct bkey_i *insert,
1230                              enum btree_update_flags flags)
1231 {
1232         struct bch_fs *c = trans->c;
1233         struct btree_iter iter, update_iter;
1234         struct bpos start = bkey_start_pos(&insert->k);
1235         struct bkey_i *update;
1236         struct bkey_s_c k;
1237         enum btree_id btree_id = orig_iter->btree_id;
1238         int ret = 0, compressed_sectors;
1239
1240         bch2_trans_iter_init(trans, &iter, btree_id, start,
1241                              BTREE_ITER_INTENT|
1242                              BTREE_ITER_WITH_UPDATES|
1243                              BTREE_ITER_NOT_EXTENTS);
1244         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1245         if ((ret = bkey_err(k)))
1246                 goto err;
1247         if (!k.k)
1248                 goto out;
1249
1250         if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
1251                 /*
1252                  * We can't merge extents if they belong to interior snapshot
1253                  * tree nodes, and there's a snapshot in which one extent is
1254                  * visible and the other is not - i.e. if visibility is
1255                  * different.
1256                  *
1257                  * Instead of checking if visibilitiy of the two extents is
1258                  * different, for now we just check if either has been
1259                  * overwritten:
1260                  */
1261                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1262                 if (ret < 0)
1263                         goto err;
1264                 if (ret)
1265                         goto nomerge1;
1266
1267                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1268                 if (ret < 0)
1269                         goto err;
1270                 if (ret)
1271                         goto nomerge1;
1272
1273                 update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1274                 if ((ret = PTR_ERR_OR_ZERO(update)))
1275                         goto err;
1276
1277                 bkey_reassemble(update, k);
1278
1279                 if (bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(insert))) {
1280                         ret = bch2_btree_delete_at(trans, &iter, flags);
1281                         if (ret)
1282                                 goto err;
1283
1284                         insert = update;
1285                         goto next;
1286                 }
1287         }
1288 nomerge1:
1289         ret = 0;
1290         if (!bkey_cmp(k.k->p, start))
1291                 goto next;
1292
1293         while (bkey_cmp(insert->k.p, bkey_start_pos(k.k)) > 0) {
1294                 bool front_split = bkey_cmp(bkey_start_pos(k.k), start) < 0;
1295                 bool back_split  = bkey_cmp(k.k->p, insert->k.p) > 0;
1296
1297                 /*
1298                  * If we're going to be splitting a compressed extent, note it
1299                  * so that __bch2_trans_commit() can increase our disk
1300                  * reservation:
1301                  */
1302                 if (((front_split && back_split) ||
1303                      ((front_split || back_split) && k.k->p.snapshot != insert->k.p.snapshot)) &&
1304                     (compressed_sectors = bch2_bkey_sectors_compressed(k)))
1305                         trans->extra_journal_res += compressed_sectors;
1306
1307                 if (front_split) {
1308                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1309                         if ((ret = PTR_ERR_OR_ZERO(update)))
1310                                 goto err;
1311
1312                         bkey_reassemble(update, k);
1313
1314                         bch2_cut_back(start, update);
1315
1316                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1317                                              BTREE_ITER_NOT_EXTENTS|
1318                                              BTREE_ITER_ALL_SNAPSHOTS|
1319                                              BTREE_ITER_INTENT);
1320                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1321                                 bch2_trans_update(trans, &update_iter, update,
1322                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1323                                                   flags);
1324                         bch2_trans_iter_exit(trans, &update_iter);
1325
1326                         if (ret)
1327                                 goto err;
1328                 }
1329
1330                 if (k.k->p.snapshot != insert->k.p.snapshot &&
1331                     (front_split || back_split)) {
1332                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1333                         if ((ret = PTR_ERR_OR_ZERO(update)))
1334                                 goto err;
1335
1336                         bkey_reassemble(update, k);
1337
1338                         bch2_cut_front(start, update);
1339                         bch2_cut_back(insert->k.p, update);
1340
1341                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1342                                              BTREE_ITER_NOT_EXTENTS|
1343                                              BTREE_ITER_ALL_SNAPSHOTS|
1344                                              BTREE_ITER_INTENT);
1345                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1346                                 bch2_trans_update(trans, &update_iter, update,
1347                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1348                                                   flags);
1349                         bch2_trans_iter_exit(trans, &update_iter);
1350                         if (ret)
1351                                 goto err;
1352                 }
1353
1354                 if (bkey_cmp(k.k->p, insert->k.p) <= 0) {
1355                         update = bch2_trans_kmalloc(trans, sizeof(*update));
1356                         if ((ret = PTR_ERR_OR_ZERO(update)))
1357                                 goto err;
1358
1359                         bkey_init(&update->k);
1360                         update->k.p = k.k->p;
1361
1362                         if (insert->k.p.snapshot != k.k->p.snapshot) {
1363                                 update->k.p.snapshot = insert->k.p.snapshot;
1364                                 update->k.type = KEY_TYPE_whiteout;
1365                         }
1366
1367                         bch2_trans_iter_init(trans, &update_iter, btree_id, update->k.p,
1368                                              BTREE_ITER_NOT_EXTENTS|
1369                                              BTREE_ITER_INTENT);
1370                         ret   = bch2_btree_iter_traverse(&update_iter) ?:
1371                                 bch2_trans_update(trans, &update_iter, update,
1372                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1373                                                   flags);
1374                         bch2_trans_iter_exit(trans, &update_iter);
1375
1376                         if (ret)
1377                                 goto err;
1378                 }
1379
1380                 if (back_split) {
1381                         update = bch2_trans_kmalloc(trans, bkey_bytes(k.k));
1382                         if ((ret = PTR_ERR_OR_ZERO(update)))
1383                                 goto err;
1384
1385                         bkey_reassemble(update, k);
1386                         bch2_cut_front(insert->k.p, update);
1387
1388                         ret = bch2_trans_update_by_path(trans, iter.path, update,
1389                                                   BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
1390                                                   flags);
1391                         if (ret)
1392                                 goto err;
1393                         goto out;
1394                 }
1395 next:
1396                 bch2_btree_iter_advance(&iter);
1397                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
1398                 if ((ret = bkey_err(k)))
1399                         goto err;
1400                 if (!k.k)
1401                         goto out;
1402         }
1403
1404         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
1405                 ret = check_pos_snapshot_overwritten(trans, btree_id, insert->k.p);
1406                 if (ret < 0)
1407                         goto err;
1408                 if (ret)
1409                         goto nomerge2;
1410
1411                 ret = check_pos_snapshot_overwritten(trans, btree_id, k.k->p);
1412                 if (ret < 0)
1413                         goto err;
1414                 if (ret)
1415                         goto nomerge2;
1416
1417                 bch2_bkey_merge(c, bkey_i_to_s(insert), k);
1418         }
1419 nomerge2:
1420         ret = 0;
1421 out:
1422         if (!bkey_deleted(&insert->k)) {
1423                 /*
1424                  * Rewinding iterators is expensive: get a new one and the one
1425                  * that points to the start of insert will be cloned from:
1426                  */
1427                 bch2_trans_iter_exit(trans, &iter);
1428                 bch2_trans_iter_init(trans, &iter, btree_id, insert->k.p,
1429                                      BTREE_ITER_NOT_EXTENTS|
1430                                      BTREE_ITER_INTENT);
1431                 ret   = bch2_btree_iter_traverse(&iter) ?:
1432                         bch2_trans_update(trans, &iter, insert, flags);
1433         }
1434 err:
1435         bch2_trans_iter_exit(trans, &iter);
1436
1437         return ret;
1438 }
1439
1440 /*
1441  * When deleting, check if we need to emit a whiteout (because we're overwriting
1442  * something in an ancestor snapshot)
1443  */
1444 static int need_whiteout_for_snapshot(struct btree_trans *trans,
1445                                       enum btree_id btree_id, struct bpos pos)
1446 {
1447         struct btree_iter iter;
1448         struct bkey_s_c k;
1449         u32 snapshot = pos.snapshot;
1450         int ret;
1451
1452         if (!bch2_snapshot_parent(trans->c, pos.snapshot))
1453                 return 0;
1454
1455         pos.snapshot++;
1456
1457         for_each_btree_key_norestart(trans, iter, btree_id, pos,
1458                            BTREE_ITER_ALL_SNAPSHOTS|
1459                            BTREE_ITER_NOPRESERVE, k, ret) {
1460                 if (bkey_cmp(k.k->p, pos))
1461                         break;
1462
1463                 if (bch2_snapshot_is_ancestor(trans->c, snapshot,
1464                                               k.k->p.snapshot)) {
1465                         ret = !bkey_whiteout(k.k);
1466                         break;
1467                 }
1468         }
1469         bch2_trans_iter_exit(trans, &iter);
1470
1471         return ret;
1472 }
1473
1474 static int __must_check
1475 bch2_trans_update_by_path(struct btree_trans *trans, struct btree_path *path,
1476                           struct bkey_i *k, enum btree_update_flags flags)
1477 {
1478         struct bch_fs *c = trans->c;
1479         struct btree_insert_entry *i, n;
1480
1481         BUG_ON(!path->should_be_locked);
1482
1483         BUG_ON(trans->nr_updates >= BTREE_ITER_MAX);
1484         BUG_ON(bpos_cmp(k->k.p, path->pos));
1485
1486         n = (struct btree_insert_entry) {
1487                 .flags          = flags,
1488                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
1489                 .btree_id       = path->btree_id,
1490                 .level          = path->level,
1491                 .cached         = path->cached,
1492                 .path           = path,
1493                 .k              = k,
1494                 .ip_allocated   = _RET_IP_,
1495         };
1496
1497 #ifdef CONFIG_BCACHEFS_DEBUG
1498         trans_for_each_update(trans, i)
1499                 BUG_ON(i != trans->updates &&
1500                        btree_insert_entry_cmp(i - 1, i) >= 0);
1501 #endif
1502
1503         /*
1504          * Pending updates are kept sorted: first, find position of new update,
1505          * then delete/trim any updates the new update overwrites:
1506          */
1507         trans_for_each_update(trans, i)
1508                 if (btree_insert_entry_cmp(&n, i) <= 0)
1509                         break;
1510
1511         if (i < trans->updates + trans->nr_updates &&
1512             !btree_insert_entry_cmp(&n, i)) {
1513                 BUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
1514
1515                 bch2_path_put(trans, i->path, true);
1516                 i->flags        = n.flags;
1517                 i->cached       = n.cached;
1518                 i->k            = n.k;
1519                 i->path         = n.path;
1520                 i->ip_allocated = n.ip_allocated;
1521         } else {
1522                 array_insert_item(trans->updates, trans->nr_updates,
1523                                   i - trans->updates, n);
1524
1525                 i->old_v = bch2_btree_path_peek_slot(path, &i->old_k).v;
1526                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
1527
1528                 if (unlikely(!test_bit(JOURNAL_REPLAY_DONE, &c->journal.flags))) {
1529                         struct bkey_i *j_k =
1530                                 bch2_journal_keys_peek(c, n.btree_id, n.level, k->k.p);
1531
1532                         if (j_k && !bpos_cmp(j_k->k.p, i->k->k.p)) {
1533                                 i->old_k = j_k->k;
1534                                 i->old_v = &j_k->v;
1535                         }
1536                 }
1537         }
1538
1539         __btree_path_get(n.path, true);
1540         return 0;
1541 }
1542
1543 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
1544                                    struct bkey_i *k, enum btree_update_flags flags)
1545 {
1546         struct btree_path *path = iter->update_path ?: iter->path;
1547         struct bkey_cached *ck;
1548         int ret;
1549
1550         if (iter->flags & BTREE_ITER_IS_EXTENTS)
1551                 return bch2_trans_update_extent(trans, iter, k, flags);
1552
1553         if (bkey_deleted(&k->k) &&
1554             !(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1555             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS)) {
1556                 ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
1557                 if (unlikely(ret < 0))
1558                         return ret;
1559
1560                 if (ret)
1561                         k->k.type = KEY_TYPE_whiteout;
1562         }
1563
1564         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
1565             !path->cached &&
1566             !path->level &&
1567             btree_id_cached(trans->c, path->btree_id)) {
1568                 if (!iter->key_cache_path ||
1569                     !iter->key_cache_path->should_be_locked ||
1570                     bpos_cmp(iter->key_cache_path->pos, k->k.p)) {
1571                         if (!iter->key_cache_path)
1572                                 iter->key_cache_path =
1573                                         bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
1574                                                       BTREE_ITER_INTENT|
1575                                                       BTREE_ITER_CACHED, _THIS_IP_);
1576
1577                         iter->key_cache_path =
1578                                 bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
1579                                                         iter->flags & BTREE_ITER_INTENT,
1580                                                         _THIS_IP_);
1581
1582                         ret = bch2_btree_path_traverse(trans, iter->key_cache_path,
1583                                                        BTREE_ITER_CACHED);
1584                         if (unlikely(ret))
1585                                 return ret;
1586
1587                         ck = (void *) iter->key_cache_path->l[0].b;
1588
1589                         if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
1590                                 trace_trans_restart_key_cache_raced(trans->fn, _RET_IP_);
1591                                 btree_trans_restart(trans);
1592                                 return -EINTR;
1593                         }
1594
1595                         iter->key_cache_path->should_be_locked = true;
1596                 }
1597
1598                 path = iter->key_cache_path;
1599         }
1600
1601         return bch2_trans_update_by_path(trans, path, k, flags);
1602 }
1603
1604 void bch2_trans_commit_hook(struct btree_trans *trans,
1605                             struct btree_trans_commit_hook *h)
1606 {
1607         h->next = trans->hooks;
1608         trans->hooks = h;
1609 }
1610
1611 int __bch2_btree_insert(struct btree_trans *trans,
1612                         enum btree_id id, struct bkey_i *k)
1613 {
1614         struct btree_iter iter;
1615         int ret;
1616
1617         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
1618                              BTREE_ITER_INTENT);
1619         ret   = bch2_btree_iter_traverse(&iter) ?:
1620                 bch2_trans_update(trans, &iter, k, 0);
1621         bch2_trans_iter_exit(trans, &iter);
1622         return ret;
1623 }
1624
1625 /**
1626  * bch2_btree_insert - insert keys into the extent btree
1627  * @c:                  pointer to struct bch_fs
1628  * @id:                 btree to insert into
1629  * @insert_keys:        list of keys to insert
1630  * @hook:               insert callback
1631  */
1632 int bch2_btree_insert(struct bch_fs *c, enum btree_id id,
1633                       struct bkey_i *k,
1634                       struct disk_reservation *disk_res,
1635                       u64 *journal_seq, int flags)
1636 {
1637         return bch2_trans_do(c, disk_res, journal_seq, flags,
1638                              __bch2_btree_insert(&trans, id, k));
1639 }
1640
1641 int bch2_btree_delete_at(struct btree_trans *trans,
1642                          struct btree_iter *iter, unsigned update_flags)
1643 {
1644         struct bkey_i *k;
1645
1646         k = bch2_trans_kmalloc(trans, sizeof(*k));
1647         if (IS_ERR(k))
1648                 return PTR_ERR(k);
1649
1650         bkey_init(&k->k);
1651         k->k.p = iter->pos;
1652         return bch2_trans_update(trans, iter, k, update_flags);
1653 }
1654
1655 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
1656                                   struct bpos start, struct bpos end,
1657                                   unsigned update_flags,
1658                                   u64 *journal_seq)
1659 {
1660         struct btree_iter iter;
1661         struct bkey_s_c k;
1662         int ret = 0;
1663
1664         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
1665 retry:
1666         while ((bch2_trans_begin(trans),
1667                (k = bch2_btree_iter_peek(&iter)).k) &&
1668                !(ret = bkey_err(k)) &&
1669                bkey_cmp(iter.pos, end) < 0) {
1670                 struct disk_reservation disk_res =
1671                         bch2_disk_reservation_init(trans->c, 0);
1672                 struct bkey_i delete;
1673
1674                 bkey_init(&delete.k);
1675
1676                 /*
1677                  * This could probably be more efficient for extents:
1678                  */
1679
1680                 /*
1681                  * For extents, iter.pos won't necessarily be the same as
1682                  * bkey_start_pos(k.k) (for non extents they always will be the
1683                  * same). It's important that we delete starting from iter.pos
1684                  * because the range we want to delete could start in the middle
1685                  * of k.
1686                  *
1687                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
1688                  * bkey_start_pos(k.k)).
1689                  */
1690                 delete.k.p = iter.pos;
1691
1692                 if (iter.flags & BTREE_ITER_IS_EXTENTS) {
1693                         unsigned max_sectors =
1694                                 KEY_SIZE_MAX & (~0 << trans->c->block_bits);
1695
1696                         /* create the biggest key we can */
1697                         bch2_key_resize(&delete.k, max_sectors);
1698                         bch2_cut_back(end, &delete);
1699
1700                         ret = bch2_extent_trim_atomic(trans, &iter, &delete);
1701                         if (ret)
1702                                 break;
1703                 }
1704
1705                 ret   = bch2_trans_update(trans, &iter, &delete, 0) ?:
1706                         bch2_trans_commit(trans, &disk_res, journal_seq,
1707                                           BTREE_INSERT_NOFAIL|
1708                                           update_flags);
1709                 bch2_disk_reservation_put(trans->c, &disk_res);
1710                 if (ret)
1711                         break;
1712         }
1713
1714         if (ret == -EINTR) {
1715                 ret = 0;
1716                 goto retry;
1717         }
1718
1719         bch2_trans_iter_exit(trans, &iter);
1720         return ret;
1721 }
1722
1723 /*
1724  * bch_btree_delete_range - delete everything within a given range
1725  *
1726  * Range is a half open interval - [start, end)
1727  */
1728 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
1729                             struct bpos start, struct bpos end,
1730                             unsigned update_flags,
1731                             u64 *journal_seq)
1732 {
1733         return bch2_trans_do(c, NULL, journal_seq, 0,
1734                              bch2_btree_delete_range_trans(&trans, id, start, end,
1735                                                            update_flags, journal_seq));
1736 }