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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_update.h"
   5 #include "btree_iter.h"
   6 #include "btree_journal_iter.h"
   7 #include "btree_locking.h"
   8 #include "buckets.h"
   9 #include "debug.h"
  10 #include "errcode.h"
  11 #include "error.h"
  12 #include "extents.h"
  13 #include "keylist.h"
  14 #include "snapshot.h"
  15 #include "trace.h"
  16
  17 #include <linux/darray.h>
  18
  19 static inline int btree_insert_entry_cmp(const struct btree_insert_entry *l,
  20                                          const struct btree_insert_entry *r)
  21 {
  22         return   cmp_int(l->btree_id,   r->btree_id) ?:
  23                  cmp_int(l->cached,     r->cached) ?:
  24                  -cmp_int(l->level,     r->level) ?:
  25                  bpos_cmp(l->k->k.p,    r->k->k.p);
  26 }
  27
  28 static int __must_check
  29 bch2_trans_update_by_path(struct btree_trans *, btree_path_idx_t,
  30                           struct bkey_i *, enum btree_update_flags,
  31                           unsigned long ip);
  32
  33 static noinline int extent_front_merge(struct btree_trans *trans,
  34                                        struct btree_iter *iter,
  35                                        struct bkey_s_c k,
  36                                        struct bkey_i **insert,
  37                                        enum btree_update_flags flags)
  38 {
  39         struct bch_fs *c = trans->c;
  40         struct bkey_i *update;
  41         int ret;
  42
  43         update = bch2_bkey_make_mut_noupdate(trans, k);
  44         ret = PTR_ERR_OR_ZERO(update);
  45         if (ret)
  46                 return ret;
  47
  48         if (!bch2_bkey_merge(c, bkey_i_to_s(update), bkey_i_to_s_c(*insert)))
  49                 return 0;
  50
  51         ret =   bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p) ?:
  52                 bch2_key_has_snapshot_overwrites(trans, iter->btree_id, (*insert)->k.p);
  53         if (ret < 0)
  54                 return ret;
  55         if (ret)
  56                 return 0;
  57
  58         ret = bch2_btree_delete_at(trans, iter, flags);
  59         if (ret)
  60                 return ret;
  61
  62         *insert = update;
  63         return 0;
  64 }
  65
  66 static noinline int extent_back_merge(struct btree_trans *trans,
  67                                       struct btree_iter *iter,
  68                                       struct bkey_i *insert,
  69                                       struct bkey_s_c k)
  70 {
  71         struct bch_fs *c = trans->c;
  72         int ret;
  73
  74         ret =   bch2_key_has_snapshot_overwrites(trans, iter->btree_id, insert->k.p) ?:
  75                 bch2_key_has_snapshot_overwrites(trans, iter->btree_id, k.k->p);
  76         if (ret < 0)
  77                 return ret;
  78         if (ret)
  79                 return 0;
  80
  81         bch2_bkey_merge(c, bkey_i_to_s(insert), k);
  82         return 0;
  83 }
  84
  85 static struct bkey_s_c peek_slot_including_whiteouts(struct btree_trans *trans, struct btree_iter *iter,
  86                                                      enum btree_id btree, struct bpos pos)
  87 {
  88         struct bkey_s_c k;
  89         int ret;
  90
  91         for_each_btree_key_norestart(trans, *iter, btree, pos,
  92                            BTREE_ITER_ALL_SNAPSHOTS|
  93                            BTREE_ITER_NOPRESERVE, k, ret) {
  94                 if (!bkey_eq(k.k->p, pos))
  95                         break;
  96                 if (bch2_snapshot_is_ancestor(trans->c, pos.snapshot, k.k->p.snapshot))
  97                         return k;
  98         }
  99         bch2_trans_iter_exit(trans, iter);
 100
 101         return ret ? bkey_s_c_err(ret) : bkey_s_c_null;
 102 }
 103
 104 /*
 105  * When deleting, check if we need to emit a whiteout (because we're overwriting
 106  * something in an ancestor snapshot)
 107  */
 108 static int need_whiteout_for_snapshot(struct btree_trans *trans, enum btree_id btree, struct bpos pos)
 109 {
 110         pos.snapshot = bch2_snapshot_parent(trans->c, pos.snapshot);
 111         if (!pos.snapshot)
 112                 return 0;
 113
 114         struct btree_iter iter;
 115         struct bkey_s_c k = peek_slot_including_whiteouts(trans, &iter, btree, pos);
 116         int ret = bkey_err(k) ?: k.k && !bkey_whiteout(k.k);
 117         bch2_trans_iter_exit(trans, &iter);
 118
 119         return ret;
 120 }
 121
 122 /*
 123  * We're overwriting a key at @pos in snapshot @snapshot, so we need to insert a
 124  * whiteout: that might be in @snapshot, or if there are overwites in sibling
 125  * snapshots, find the common ancestor where @pos is overwritten in every
 126  * descendent and insert the whiteout there - which might be at @pos.
 127  */
 128 static int delete_interior_snapshot_key(struct btree_trans *trans,
 129                                         enum btree_id btree,
 130                                         struct bpos whiteout, bool deleting,
 131                                         struct bpos overwrite, bool old_is_whiteout)
 132 {
 133         struct bch_fs *c = trans->c;
 134         struct bpos orig_whiteout = whiteout, sib = whiteout;
 135         struct btree_iter iter;
 136         struct bkey_s_c k;
 137         int ret;
 138
 139         sib.snapshot = bch2_snapshot_sibling(c, sib.snapshot);
 140
 141         for_each_btree_key_norestart(trans, iter, btree, sib,
 142                                      BTREE_ITER_ALL_SNAPSHOTS|BTREE_ITER_INTENT, k, ret) {
 143                 BUG_ON(bpos_gt(k.k->p, overwrite));
 144
 145                 if (bpos_lt(k.k->p, sib)) /* unrelated branch - skip */
 146                         continue;
 147                 if (bpos_gt(k.k->p, sib)) /* did not find @sib */
 148                         break;
 149
 150                 /* @overwrite is also written in @sib, now check parent */
 151                 whiteout.snapshot = bch2_snapshot_parent(c, whiteout.snapshot);
 152                 if (bpos_eq(whiteout, overwrite))
 153                         break;
 154
 155                 sib = whiteout;
 156                 sib.snapshot = bch2_snapshot_sibling(c, sib.snapshot);
 157         }
 158
 159         if (ret)
 160                 goto err;
 161
 162         if (!deleting && bpos_eq(whiteout, orig_whiteout))
 163                 goto out;
 164
 165         if (!bpos_eq(iter.pos, whiteout)) {
 166                 bch2_trans_iter_exit(trans, &iter);
 167                 bch2_trans_iter_init(trans, &iter, btree, whiteout, BTREE_ITER_INTENT);
 168                 k = bch2_btree_iter_peek_slot(&iter);
 169                 ret = bkey_err(k);
 170                 if (ret)
 171                         goto err;
 172         }
 173
 174         iter.flags &= ~BTREE_ITER_ALL_SNAPSHOTS;
 175         iter.flags |= BTREE_ITER_FILTER_SNAPSHOTS;
 176
 177         struct bkey_i *delete = bch2_trans_kmalloc(trans, sizeof(*delete));
 178         ret = PTR_ERR_OR_ZERO(delete);
 179         if (ret)
 180                 goto err;
 181
 182         bkey_init(&delete->k);
 183         delete->k.p = whiteout;
 184
 185         ret = !bpos_eq(whiteout, overwrite)
 186                 ? !old_is_whiteout
 187                 : need_whiteout_for_snapshot(trans, btree, whiteout);
 188         if (ret < 0)
 189                 goto err;
 190         if (ret)
 191                 delete->k.type = KEY_TYPE_whiteout;
 192
 193         ret = bch2_trans_update(trans, &iter, delete,
 194                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 195                                 BTREE_UPDATE_SNAPSHOT_WHITEOUT_CHECKS_DONE);
 196 out:
 197 err:
 198         bch2_trans_iter_exit(trans, &iter);
 199         return ret;
 200 }
 201
 202 /*
 203  * We're overwriting a key in a snapshot that has ancestors: if we're
 204  * overwriting a key in a different snapshot, we need to check if it is now
 205  * fully overritten and can be deleted, and if we're deleting a key in the
 206  * current snapshot we need to check if we need to leave a whiteout.
 207  */
 208 static noinline int
 209 overwrite_interior_snapshot_key(struct btree_trans *trans,
 210                                 struct btree_iter *iter,
 211                                 struct bkey_i *k)
 212 {
 213         struct bkey_s_c old = bch2_btree_iter_peek_slot(iter);
 214
 215         int ret = bkey_err(old);
 216         if (ret)
 217                 return ret;
 218
 219         if (!bkey_deleted(old.k)) {
 220                 if (btree_type_snapshots_unreffed(iter->btree_id) &&
 221                     old.k->p.snapshot != k->k.p.snapshot) {
 222                         /*
 223                          * We're overwriting a key in a different snapshot:
 224                          * check if it's also been overwritten in siblings
 225                          */
 226                         ret = delete_interior_snapshot_key(trans, iter->btree_id,
 227                                                            k->k.p,   bkey_deleted(&k->k),
 228                                                            old.k->p, bkey_whiteout(old.k));
 229                         if (ret)
 230                                 return ret;
 231                         if (bkey_deleted(&k->k))
 232                                 return 1;
 233                 } else if (bkey_deleted(&k->k)) {
 234                         /*
 235                          * We're deleting a key in the current snapshot:
 236                          * check if we need to leave a whiteout
 237                          */
 238                         ret = need_whiteout_for_snapshot(trans, iter->btree_id, k->k.p);
 239                         if (unlikely(ret < 0))
 240                                 return ret;
 241                         if (ret)
 242                                 k->k.type = KEY_TYPE_whiteout;
 243                 }
 244         }
 245
 246         return 0;
 247 }
 248
 249 int __bch2_insert_snapshot_whiteouts(struct btree_trans *trans,
 250                                    enum btree_id id,
 251                                    struct bpos old_pos,
 252                                    struct bpos new_pos)
 253 {
 254         struct bch_fs *c = trans->c;
 255         struct btree_iter old_iter, new_iter = { NULL };
 256         struct bkey_s_c old_k, new_k;
 257         snapshot_id_list s;
 258         struct bkey_i *update;
 259         int ret = 0;
 260
 261         if (!bch2_snapshot_has_children(c, old_pos.snapshot))
 262                 return 0;
 263
 264         darray_init(&s);
 265
 266         bch2_trans_iter_init(trans, &old_iter, id, old_pos,
 267                              BTREE_ITER_NOT_EXTENTS|
 268                              BTREE_ITER_ALL_SNAPSHOTS);
 269         while ((old_k = bch2_btree_iter_prev(&old_iter)).k &&
 270                !(ret = bkey_err(old_k)) &&
 271                bkey_eq(old_pos, old_k.k->p)) {
 272                 struct bpos whiteout_pos =
 273                         SPOS(new_pos.inode, new_pos.offset, old_k.k->p.snapshot);;
 274
 275                 if (!bch2_snapshot_is_ancestor(c, old_k.k->p.snapshot, old_pos.snapshot) ||
 276                     snapshot_list_has_ancestor(c, &s, old_k.k->p.snapshot))
 277                         continue;
 278
 279                 new_k = bch2_bkey_get_iter(trans, &new_iter, id, whiteout_pos,
 280                                            BTREE_ITER_NOT_EXTENTS|
 281                                            BTREE_ITER_INTENT);
 282                 ret = bkey_err(new_k);
 283                 if (ret)
 284                         break;
 285
 286                 if (new_k.k->type == KEY_TYPE_deleted) {
 287                         update = bch2_trans_kmalloc(trans, sizeof(struct bkey_i));
 288                         ret = PTR_ERR_OR_ZERO(update);
 289                         if (ret)
 290                                 break;
 291
 292                         bkey_init(&update->k);
 293                         update->k.p             = whiteout_pos;
 294                         update->k.type          = KEY_TYPE_whiteout;
 295
 296                         ret = bch2_trans_update(trans, &new_iter, update,
 297                                                 BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 298                 }
 299                 bch2_trans_iter_exit(trans, &new_iter);
 300
 301                 ret = snapshot_list_add(c, &s, old_k.k->p.snapshot);
 302                 if (ret)
 303                         break;
 304         }
 305         bch2_trans_iter_exit(trans, &new_iter);
 306         bch2_trans_iter_exit(trans, &old_iter);
 307         darray_exit(&s);
 308
 309         return ret;
 310 }
 311
 312 int bch2_trans_update_extent_overwrite(struct btree_trans *trans,
 313                                        struct btree_iter *iter,
 314                                        enum btree_update_flags flags,
 315                                        struct bkey_s_c old,
 316                                        struct bkey_s_c new)
 317 {
 318         enum btree_id btree_id = iter->btree_id;
 319         struct bkey_i *update;
 320         struct bpos new_start = bkey_start_pos(new.k);
 321         unsigned front_split = bkey_lt(bkey_start_pos(old.k), new_start);
 322         unsigned back_split  = bkey_gt(old.k->p, new.k->p);
 323         unsigned middle_split = (front_split || back_split) &&
 324                 old.k->p.snapshot != new.k->p.snapshot;
 325         unsigned nr_splits = front_split + back_split + middle_split;
 326         int ret = 0, compressed_sectors;
 327
 328         /*
 329          * If we're going to be splitting a compressed extent, note it
 330          * so that __bch2_trans_commit() can increase our disk
 331          * reservation:
 332          */
 333         if (nr_splits > 1 &&
 334             (compressed_sectors = bch2_bkey_sectors_compressed(old)))
 335                 trans->extra_disk_res += compressed_sectors * (nr_splits - 1);
 336
 337         if (front_split) {
 338                 update = bch2_bkey_make_mut_noupdate(trans, old);
 339                 if ((ret = PTR_ERR_OR_ZERO(update)))
 340                         return ret;
 341
 342                 bch2_cut_back(new_start, update);
 343
 344                 ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
 345                                         old.k->p, update->k.p) ?:
 346                         bch2_btree_insert_nonextent(trans, btree_id, update,
 347                                         BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 348                 if (ret)
 349                         return ret;
 350         }
 351
 352         /* If we're overwriting in a different snapshot - middle split: */
 353         if (middle_split) {
 354                 update = bch2_bkey_make_mut_noupdate(trans, old);
 355                 if ((ret = PTR_ERR_OR_ZERO(update)))
 356                         return ret;
 357
 358                 bch2_cut_front(new_start, update);
 359                 bch2_cut_back(new.k->p, update);
 360
 361                 ret =   bch2_insert_snapshot_whiteouts(trans, btree_id,
 362                                         old.k->p, update->k.p) ?:
 363                         bch2_btree_insert_nonextent(trans, btree_id, update,
 364                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 365                 if (ret)
 366                         return ret;
 367         }
 368
 369         if (bkey_le(old.k->p, new.k->p)) {
 370                 update = bch2_trans_kmalloc(trans, sizeof(*update));
 371                 if ((ret = PTR_ERR_OR_ZERO(update)))
 372                         return ret;
 373
 374                 bkey_init(&update->k);
 375                 update->k.p = old.k->p;
 376                 update->k.p.snapshot = new.k->p.snapshot;
 377
 378                 if (new.k->p.snapshot != old.k->p.snapshot) {
 379                         update->k.type = KEY_TYPE_whiteout;
 380                 } else if (btree_type_has_snapshots(btree_id)) {
 381                         ret = need_whiteout_for_snapshot(trans, btree_id, update->k.p);
 382                         if (ret < 0)
 383                                 return ret;
 384                         if (ret)
 385                                 update->k.type = KEY_TYPE_whiteout;
 386                 }
 387
 388                 ret = bch2_btree_insert_nonextent(trans, btree_id, update,
 389                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|flags);
 390                 if (ret)
 391                         return ret;
 392         }
 393
 394         if (back_split) {
 395                 update = bch2_bkey_make_mut_noupdate(trans, old);
 396                 if ((ret = PTR_ERR_OR_ZERO(update)))
 397                         return ret;
 398
 399                 bch2_cut_front(new.k->p, update);
 400
 401                 ret = bch2_trans_update_by_path(trans, iter->path, update,
 402                                           BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE|
 403                                           flags, _RET_IP_);
 404                 if (ret)
 405                         return ret;
 406         }
 407
 408         return 0;
 409 }
 410
 411 static int bch2_trans_update_extent(struct btree_trans *trans,
 412                                     struct btree_iter *orig_iter,
 413                                     struct bkey_i *insert,
 414                                     enum btree_update_flags flags)
 415 {
 416         struct btree_iter iter;
 417         struct bkey_s_c k;
 418         enum btree_id btree_id = orig_iter->btree_id;
 419         int ret = 0;
 420
 421         bch2_trans_iter_init(trans, &iter, btree_id, bkey_start_pos(&insert->k),
 422                              BTREE_ITER_INTENT|
 423                              BTREE_ITER_WITH_UPDATES|
 424                              BTREE_ITER_NOT_EXTENTS);
 425         k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
 426         if ((ret = bkey_err(k)))
 427                 goto err;
 428         if (!k.k)
 429                 goto out;
 430
 431         if (bkey_eq(k.k->p, bkey_start_pos(&insert->k))) {
 432                 if (bch2_bkey_maybe_mergable(k.k, &insert->k)) {
 433                         ret = extent_front_merge(trans, &iter, k, &insert, flags);
 434                         if (ret)
 435                                 goto err;
 436                 }
 437
 438                 goto next;
 439         }
 440
 441         while (bkey_gt(insert->k.p, bkey_start_pos(k.k))) {
 442                 bool done = bkey_lt(insert->k.p, k.k->p);
 443
 444                 ret = bch2_trans_update_extent_overwrite(trans, &iter, flags, k, bkey_i_to_s_c(insert));
 445                 if (ret)
 446                         goto err;
 447
 448                 if (done)
 449                         goto out;
 450 next:
 451                 bch2_btree_iter_advance(&iter);
 452                 k = bch2_btree_iter_peek_upto(&iter, POS(insert->k.p.inode, U64_MAX));
 453                 if ((ret = bkey_err(k)))
 454                         goto err;
 455                 if (!k.k)
 456                         goto out;
 457         }
 458
 459         if (bch2_bkey_maybe_mergable(&insert->k, k.k)) {
 460                 ret = extent_back_merge(trans, &iter, insert, k);
 461                 if (ret)
 462                         goto err;
 463         }
 464 out:
 465         if (!bkey_deleted(&insert->k))
 466                 ret = bch2_btree_insert_nonextent(trans, btree_id, insert, flags);
 467 err:
 468         bch2_trans_iter_exit(trans, &iter);
 469
 470         return ret;
 471 }
 472
 473 static noinline int flush_new_cached_update(struct btree_trans *trans,
 474                                             struct btree_insert_entry *i,
 475                                             enum btree_update_flags flags,
 476                                             unsigned long ip)
 477 {
 478         struct bkey k;
 479         int ret;
 480
 481         btree_path_idx_t path_idx =
 482                 bch2_path_get(trans, i->btree_id, i->old_k.p, 1, 0,
 483                               BTREE_ITER_INTENT, _THIS_IP_);
 484         ret = bch2_btree_path_traverse(trans, path_idx, 0);
 485         if (ret)
 486                 goto out;
 487
 488         struct btree_path *btree_path = trans->paths + path_idx;
 489
 490         /*
 491          * The old key in the insert entry might actually refer to an existing
 492          * key in the btree that has been deleted from cache and not yet
 493          * flushed. Check for this and skip the flush so we don't run triggers
 494          * against a stale key.
 495          */
 496         bch2_btree_path_peek_slot_exact(btree_path, &k);
 497         if (!bkey_deleted(&k))
 498                 goto out;
 499
 500         i->key_cache_already_flushed = true;
 501         i->flags |= BTREE_TRIGGER_NORUN;
 502
 503         btree_path_set_should_be_locked(btree_path);
 504         ret = bch2_trans_update_by_path(trans, path_idx, i->k, flags, ip);
 505 out:
 506         bch2_path_put(trans, path_idx, true);
 507         return ret;
 508 }
 509
 510 static int __must_check
 511 bch2_trans_update_by_path(struct btree_trans *trans, btree_path_idx_t path_idx,
 512                           struct bkey_i *k, enum btree_update_flags flags,
 513                           unsigned long ip)
 514 {
 515         struct bch_fs *c = trans->c;
 516         struct btree_insert_entry *i, n;
 517         int cmp;
 518
 519         struct btree_path *path = trans->paths + path_idx;
 520         EBUG_ON(!path->should_be_locked);
 521         EBUG_ON(trans->nr_updates >= trans->nr_paths);
 522         EBUG_ON(!bpos_eq(k->k.p, path->pos));
 523
 524         n = (struct btree_insert_entry) {
 525                 .flags          = flags,
 526                 .bkey_type      = __btree_node_type(path->level, path->btree_id),
 527                 .btree_id       = path->btree_id,
 528                 .level          = path->level,
 529                 .cached         = path->cached,
 530                 .path           = path_idx,
 531                 .k              = k,
 532                 .ip_allocated   = ip,
 533         };
 534
 535 #ifdef CONFIG_BCACHEFS_DEBUG
 536         trans_for_each_update(trans, i)
 537                 BUG_ON(i != trans->updates &&
 538                        btree_insert_entry_cmp(i - 1, i) >= 0);
 539 #endif
 540
 541         /*
 542          * Pending updates are kept sorted: first, find position of new update,
 543          * then delete/trim any updates the new update overwrites:
 544          */
 545         for (i = trans->updates; i < trans->updates + trans->nr_updates; i++) {
 546                 cmp = btree_insert_entry_cmp(&n, i);
 547                 if (cmp <= 0)
 548                         break;
 549         }
 550
 551         if (!cmp && i < trans->updates + trans->nr_updates) {
 552                 EBUG_ON(i->insert_trigger_run || i->overwrite_trigger_run);
 553
 554                 bch2_path_put(trans, i->path, true);
 555                 i->flags        = n.flags;
 556                 i->cached       = n.cached;
 557                 i->k            = n.k;
 558                 i->path         = n.path;
 559                 i->ip_allocated = n.ip_allocated;
 560         } else {
 561                 array_insert_item(trans->updates, trans->nr_updates,
 562                                   i - trans->updates, n);
 563
 564                 i->old_v = bch2_btree_path_peek_slot_exact(path, &i->old_k).v;
 565                 i->old_btree_u64s = !bkey_deleted(&i->old_k) ? i->old_k.u64s : 0;
 566
 567                 if (unlikely(trans->journal_replay_not_finished)) {
 568                         struct bkey_i *j_k =
 569                                 bch2_journal_keys_peek_slot(c, n.btree_id, n.level, k->k.p);
 570
 571                         if (j_k) {
 572                                 i->old_k = j_k->k;
 573                                 i->old_v = &j_k->v;
 574                         }
 575                 }
 576         }
 577
 578         __btree_path_get(trans->paths + i->path, true);
 579
 580         /*
 581          * If a key is present in the key cache, it must also exist in the
 582          * btree - this is necessary for cache coherency. When iterating over
 583          * a btree that's cached in the key cache, the btree iter code checks
 584          * the key cache - but the key has to exist in the btree for that to
 585          * work:
 586          */
 587         if (path->cached && bkey_deleted(&i->old_k))
 588                 return flush_new_cached_update(trans, i, flags, ip);
 589
 590         return 0;
 591 }
 592
 593 static noinline int bch2_trans_update_get_key_cache(struct btree_trans *trans,
 594                                                     struct btree_iter *iter,
 595                                                     struct btree_path *path)
 596 {
 597         struct btree_path *key_cache_path = btree_iter_key_cache_path(trans, iter);
 598
 599         if (!key_cache_path ||
 600             !key_cache_path->should_be_locked ||
 601             !bpos_eq(key_cache_path->pos, iter->pos)) {
 602                 struct bkey_cached *ck;
 603                 int ret;
 604
 605                 if (!iter->key_cache_path)
 606                         iter->key_cache_path =
 607                                 bch2_path_get(trans, path->btree_id, path->pos, 1, 0,
 608                                               BTREE_ITER_INTENT|
 609                                               BTREE_ITER_CACHED, _THIS_IP_);
 610
 611                 iter->key_cache_path =
 612                         bch2_btree_path_set_pos(trans, iter->key_cache_path, path->pos,
 613                                                 iter->flags & BTREE_ITER_INTENT,
 614                                                 _THIS_IP_);
 615
 616                 ret = bch2_btree_path_traverse(trans, iter->key_cache_path, BTREE_ITER_CACHED);
 617                 if (unlikely(ret))
 618                         return ret;
 619
 620                 ck = (void *) trans->paths[iter->key_cache_path].l[0].b;
 621
 622                 if (test_bit(BKEY_CACHED_DIRTY, &ck->flags)) {
 623                         trace_and_count(trans->c, trans_restart_key_cache_raced, trans, _RET_IP_);
 624                         return btree_trans_restart(trans, BCH_ERR_transaction_restart_key_cache_raced);
 625                 }
 626
 627                 btree_path_set_should_be_locked(trans->paths + iter->key_cache_path);
 628         }
 629
 630         return 0;
 631 }
 632
 633 int __must_check bch2_trans_update(struct btree_trans *trans, struct btree_iter *iter,
 634                                    struct bkey_i *k, enum btree_update_flags flags)
 635 {
 636         if (iter->flags & BTREE_ITER_IS_EXTENTS)
 637                 return bch2_trans_update_extent(trans, iter, k, flags);
 638
 639         if (!(flags & (BTREE_UPDATE_SNAPSHOT_WHITEOUT_CHECKS_DONE|
 640                        BTREE_UPDATE_KEY_CACHE_RECLAIM)) &&
 641             (iter->flags & BTREE_ITER_FILTER_SNAPSHOTS) &&
 642             bch2_snapshot_parent(trans->c, k->k.p.snapshot)) {
 643                 int ret = overwrite_interior_snapshot_key(trans, iter, k);
 644                 if (ret)
 645                         return ret < 0 ? ret : 0;
 646         }
 647
 648         /*
 649          * Ensure that updates to cached btrees go to the key cache:
 650          */
 651         btree_path_idx_t path_idx = iter->update_path ?: iter->path;
 652         struct btree_path *path = trans->paths + path_idx;
 653
 654         if (!(flags & BTREE_UPDATE_KEY_CACHE_RECLAIM) &&
 655             !path->cached &&
 656             !path->level &&
 657             btree_id_cached(trans->c, path->btree_id)) {
 658                 int ret = bch2_trans_update_get_key_cache(trans, iter, path);
 659                 if (ret)
 660                         return ret;
 661
 662                 path_idx = iter->key_cache_path;
 663         }
 664
 665         return bch2_trans_update_by_path(trans, path_idx, k, flags, _RET_IP_);
 666 }
 667
 668 int bch2_btree_insert_clone_trans(struct btree_trans *trans,
 669                                   enum btree_id btree,
 670                                   struct bkey_i *k)
 671 {
 672         struct bkey_i *n = bch2_trans_kmalloc(trans, bkey_bytes(&k->k));
 673         int ret = PTR_ERR_OR_ZERO(n);
 674         if (ret)
 675                 return ret;
 676
 677         bkey_copy(n, k);
 678         return bch2_btree_insert_trans(trans, btree, n, 0);
 679 }
 680
 681 struct jset_entry *__bch2_trans_jset_entry_alloc(struct btree_trans *trans, unsigned u64s)
 682 {
 683         unsigned new_top = trans->journal_entries_u64s + u64s;
 684         unsigned old_size = trans->journal_entries_size;
 685
 686         if (new_top > trans->journal_entries_size) {
 687                 trans->journal_entries_size = roundup_pow_of_two(new_top);
 688
 689                 btree_trans_stats(trans)->journal_entries_size = trans->journal_entries_size;
 690         }
 691
 692         struct jset_entry *n =
 693                 bch2_trans_kmalloc_nomemzero(trans,
 694                                 trans->journal_entries_size * sizeof(u64));
 695         if (IS_ERR(n))
 696                 return ERR_CAST(n);
 697
 698         if (trans->journal_entries)
 699                 memcpy(n, trans->journal_entries, old_size * sizeof(u64));
 700         trans->journal_entries = n;
 701
 702         struct jset_entry *e = btree_trans_journal_entries_top(trans);
 703         trans->journal_entries_u64s = new_top;
 704         return e;
 705 }
 706
 707 int bch2_bkey_get_empty_slot(struct btree_trans *trans, struct btree_iter *iter,
 708                              enum btree_id btree, struct bpos end)
 709 {
 710         struct bkey_s_c k;
 711         int ret = 0;
 712
 713         bch2_trans_iter_init(trans, iter, btree, POS_MAX, BTREE_ITER_INTENT);
 714         k = bch2_btree_iter_prev(iter);
 715         ret = bkey_err(k);
 716         if (ret)
 717                 goto err;
 718
 719         bch2_btree_iter_advance(iter);
 720         k = bch2_btree_iter_peek_slot(iter);
 721         ret = bkey_err(k);
 722         if (ret)
 723                 goto err;
 724
 725         BUG_ON(k.k->type != KEY_TYPE_deleted);
 726
 727         if (bkey_gt(k.k->p, end)) {
 728                 ret = -BCH_ERR_ENOSPC_btree_slot;
 729                 goto err;
 730         }
 731
 732         return 0;
 733 err:
 734         bch2_trans_iter_exit(trans, iter);
 735         return ret;
 736 }
 737
 738 void bch2_trans_commit_hook(struct btree_trans *trans,
 739                             struct btree_trans_commit_hook *h)
 740 {
 741         h->next = trans->hooks;
 742         trans->hooks = h;
 743 }
 744
 745 int bch2_btree_insert_nonextent(struct btree_trans *trans,
 746                                 enum btree_id btree, struct bkey_i *k,
 747                                 enum btree_update_flags flags)
 748 {
 749         struct btree_iter iter;
 750         int ret;
 751
 752         bch2_trans_iter_init(trans, &iter, btree, k->k.p,
 753                              BTREE_ITER_CACHED|
 754                              BTREE_ITER_NOT_EXTENTS|
 755                              BTREE_ITER_INTENT);
 756         ret   = bch2_btree_iter_traverse(&iter) ?:
 757                 bch2_trans_update(trans, &iter, k, flags);
 758         bch2_trans_iter_exit(trans, &iter);
 759         return ret;
 760 }
 761
 762 int bch2_btree_insert_trans(struct btree_trans *trans, enum btree_id id,
 763                             struct bkey_i *k, enum btree_update_flags flags)
 764 {
 765         struct btree_iter iter;
 766         int ret;
 767
 768         bch2_trans_iter_init(trans, &iter, id, bkey_start_pos(&k->k),
 769                              BTREE_ITER_CACHED|
 770                              BTREE_ITER_INTENT);
 771         ret   = bch2_btree_iter_traverse(&iter) ?:
 772                 bch2_trans_update(trans, &iter, k, flags);
 773         bch2_trans_iter_exit(trans, &iter);
 774         return ret;
 775 }
 776
 777 /**
 778  * bch2_btree_insert - insert keys into the extent btree
 779  * @c:                  pointer to struct bch_fs
 780  * @id:                 btree to insert into
 781  * @k:                  key to insert
 782  * @disk_res:           must be non-NULL whenever inserting or potentially
 783  *                      splitting data extents
 784  * @flags:              transaction commit flags
 785  *
 786  * Returns:             0 on success, error code on failure
 787  */
 788 int bch2_btree_insert(struct bch_fs *c, enum btree_id id, struct bkey_i *k,
 789                       struct disk_reservation *disk_res, int flags)
 790 {
 791         return bch2_trans_do(c, disk_res, NULL, flags,
 792                              bch2_btree_insert_trans(trans, id, k, 0));
 793 }
 794
 795 int bch2_btree_delete_extent_at(struct btree_trans *trans, struct btree_iter *iter,
 796                                 unsigned len, unsigned update_flags)
 797 {
 798         struct bkey_i *k;
 799
 800         k = bch2_trans_kmalloc(trans, sizeof(*k));
 801         if (IS_ERR(k))
 802                 return PTR_ERR(k);
 803
 804         bkey_init(&k->k);
 805         k->k.p = iter->pos;
 806         bch2_key_resize(&k->k, len);
 807         return bch2_trans_update(trans, iter, k, update_flags);
 808 }
 809
 810 int bch2_btree_delete_at(struct btree_trans *trans,
 811                          struct btree_iter *iter, unsigned update_flags)
 812 {
 813         return bch2_btree_delete_extent_at(trans, iter, 0, update_flags);
 814 }
 815
 816 int bch2_btree_delete(struct btree_trans *trans,
 817                       enum btree_id btree, struct bpos pos,
 818                       unsigned update_flags)
 819 {
 820         struct btree_iter iter;
 821         int ret;
 822
 823         bch2_trans_iter_init(trans, &iter, btree, pos,
 824                              BTREE_ITER_CACHED|
 825                              BTREE_ITER_INTENT);
 826         ret   = bch2_btree_iter_traverse(&iter) ?:
 827                 bch2_btree_delete_at(trans, &iter, update_flags);
 828         bch2_trans_iter_exit(trans, &iter);
 829
 830         return ret;
 831 }
 832
 833 int bch2_btree_delete_range_trans(struct btree_trans *trans, enum btree_id id,
 834                                   struct bpos start, struct bpos end,
 835                                   unsigned update_flags,
 836                                   u64 *journal_seq)
 837 {
 838         u32 restart_count = trans->restart_count;
 839         struct btree_iter iter;
 840         struct bkey_s_c k;
 841         int ret = 0;
 842
 843         bch2_trans_iter_init(trans, &iter, id, start, BTREE_ITER_INTENT);
 844         while ((k = bch2_btree_iter_peek_upto(&iter, end)).k) {
 845                 struct disk_reservation disk_res =
 846                         bch2_disk_reservation_init(trans->c, 0);
 847                 struct bkey_i delete;
 848
 849                 ret = bkey_err(k);
 850                 if (ret)
 851                         goto err;
 852
 853                 bkey_init(&delete.k);
 854
 855                 /*
 856                  * This could probably be more efficient for extents:
 857                  */
 858
 859                 /*
 860                  * For extents, iter.pos won't necessarily be the same as
 861                  * bkey_start_pos(k.k) (for non extents they always will be the
 862                  * same). It's important that we delete starting from iter.pos
 863                  * because the range we want to delete could start in the middle
 864                  * of k.
 865                  *
 866                  * (bch2_btree_iter_peek() does guarantee that iter.pos >=
 867                  * bkey_start_pos(k.k)).
 868                  */
 869                 delete.k.p = iter.pos;
 870
 871                 if (iter.flags & BTREE_ITER_IS_EXTENTS)
 872                         bch2_key_resize(&delete.k,
 873                                         bpos_min(end, k.k->p).offset -
 874                                         iter.pos.offset);
 875
 876                 ret   = bch2_trans_update(trans, &iter, &delete, update_flags) ?:
 877                         bch2_trans_commit(trans, &disk_res, journal_seq,
 878                                           BCH_TRANS_COMMIT_no_enospc);
 879                 bch2_disk_reservation_put(trans->c, &disk_res);
 880 err:
 881                 /*
 882                  * the bch2_trans_begin() call is in a weird place because we
 883                  * need to call it after every transaction commit, to avoid path
 884                  * overflow, but don't want to call it if the delete operation
 885                  * is a no-op and we have no work to do:
 886                  */
 887                 bch2_trans_begin(trans);
 888
 889                 if (bch2_err_matches(ret, BCH_ERR_transaction_restart))
 890                         ret = 0;
 891                 if (ret)
 892                         break;
 893         }
 894         bch2_trans_iter_exit(trans, &iter);
 895
 896         return ret ?: trans_was_restarted(trans, restart_count);
 897 }
 898
 899 /*
 900  * bch_btree_delete_range - delete everything within a given range
 901  *
 902  * Range is a half open interval - [start, end)
 903  */
 904 int bch2_btree_delete_range(struct bch_fs *c, enum btree_id id,
 905                             struct bpos start, struct bpos end,
 906                             unsigned update_flags,
 907                             u64 *journal_seq)
 908 {
 909         int ret = bch2_trans_run(c,
 910                         bch2_btree_delete_range_trans(trans, id, start, end,
 911                                                       update_flags, journal_seq));
 912         if (ret == -BCH_ERR_transaction_restart_nested)
 913                 ret = 0;
 914         return ret;
 915 }
 916
 917 int bch2_btree_bit_mod(struct btree_trans *trans, enum btree_id btree,
 918                        struct bpos pos, bool set)
 919 {
 920         struct bkey_i *k = bch2_trans_kmalloc(trans, sizeof(*k));
 921         int ret = PTR_ERR_OR_ZERO(k);
 922         if (ret)
 923                 return ret;
 924
 925         bkey_init(&k->k);
 926         k->k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
 927         k->k.p = pos;
 928
 929         struct btree_iter iter;
 930         bch2_trans_iter_init(trans, &iter, btree, pos, BTREE_ITER_INTENT);
 931
 932         ret   = bch2_btree_iter_traverse(&iter) ?:
 933                 bch2_trans_update(trans, &iter, k, 0);
 934         bch2_trans_iter_exit(trans, &iter);
 935         return ret;
 936 }
 937
 938 int bch2_btree_bit_mod_buffered(struct btree_trans *trans, enum btree_id btree,
 939                                 struct bpos pos, bool set)
 940 {
 941         struct bkey_i k;
 942
 943         bkey_init(&k.k);
 944         k.k.type = set ? KEY_TYPE_set : KEY_TYPE_deleted;
 945         k.k.p = pos;
 946
 947         return bch2_trans_update_buffered(trans, btree, &k);
 948 }
 949
 950 static int __bch2_trans_log_msg(struct btree_trans *trans, struct printbuf *buf, unsigned u64s)
 951 {
 952         struct jset_entry *e = bch2_trans_jset_entry_alloc(trans, jset_u64s(u64s));
 953         int ret = PTR_ERR_OR_ZERO(e);
 954         if (ret)
 955                 return ret;
 956
 957         struct jset_entry_log *l = container_of(e, struct jset_entry_log, entry);
 958         journal_entry_init(e, BCH_JSET_ENTRY_log, 0, 1, u64s);
 959         memcpy(l->d, buf->buf, buf->pos);
 960         return 0;
 961 }
 962
 963 __printf(3, 0)
 964 static int
 965 __bch2_fs_log_msg(struct bch_fs *c, unsigned commit_flags, const char *fmt,
 966                   va_list args)
 967 {
 968         struct printbuf buf = PRINTBUF;
 969         prt_vprintf(&buf, fmt, args);
 970
 971         unsigned u64s = DIV_ROUND_UP(buf.pos, sizeof(u64));
 972         prt_chars(&buf, '\0', u64s * sizeof(u64) - buf.pos);
 973
 974         int ret = buf.allocation_failure ? -BCH_ERR_ENOMEM_trans_log_msg : 0;
 975         if (ret)
 976                 goto err;
 977
 978         if (!test_bit(JOURNAL_STARTED, &c->journal.flags)) {
 979                 ret = darray_make_room(&c->journal.early_journal_entries, jset_u64s(u64s));
 980                 if (ret)
 981                         goto err;
 982
 983                 struct jset_entry_log *l = (void *) &darray_top(c->journal.early_journal_entries);
 984                 journal_entry_init(&l->entry, BCH_JSET_ENTRY_log, 0, 1, u64s);
 985                 memcpy(l->d, buf.buf, buf.pos);
 986                 c->journal.early_journal_entries.nr += jset_u64s(u64s);
 987         } else {
 988                 ret = bch2_trans_do(c, NULL, NULL,
 989                         BCH_TRANS_COMMIT_lazy_rw|commit_flags,
 990                         __bch2_trans_log_msg(trans, &buf, u64s));
 991         }
 992 err:
 993         printbuf_exit(&buf);
 994         return ret;
 995 }
 996
 997 __printf(2, 3)
 998 int bch2_fs_log_msg(struct bch_fs *c, const char *fmt, ...)
 999 {
1000         va_list args;
1001         int ret;
1002
1003         va_start(args, fmt);
1004         ret = __bch2_fs_log_msg(c, 0, fmt, args);
1005         va_end(args);
1006         return ret;
1007 }
1008
1009 /*
1010  * Use for logging messages during recovery to enable reserved space and avoid
1011  * blocking.
1012  */
1013 __printf(2, 3)
1014 int bch2_journal_log_msg(struct bch_fs *c, const char *fmt, ...)
1015 {
1016         va_list args;
1017         int ret;
1018
1019         va_start(args, fmt);
1020         ret = __bch2_fs_log_msg(c, BCH_WATERMARK_reclaim, fmt, args);
1021         va_end(args);
1022         return ret;
1023 }