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

   1 // SPDX-License-Identifier: GPL-2.0
   2
   3 #include "bcachefs.h"
   4 #include "btree_key_cache.h"
   5 #include "btree_update.h"
   6 #include "errcode.h"
   7 #include "error.h"
   8 #include "fs.h"
   9 #include "subvolume.h"
  10
  11 static int bch2_subvolume_delete(struct btree_trans *, u32);
  12
  13 /* Snapshot tree: */
  14
  15 void bch2_snapshot_tree_to_text(struct printbuf *out, struct bch_fs *c,
  16                                 struct bkey_s_c k)
  17 {
  18         struct bkey_s_c_snapshot_tree t = bkey_s_c_to_snapshot_tree(k);
  19
  20         prt_printf(out, "subvol %u root snapshot %u",
  21                    le32_to_cpu(t.v->master_subvol),
  22                    le32_to_cpu(t.v->root_snapshot));
  23 }
  24
  25 int bch2_snapshot_tree_invalid(const struct bch_fs *c, struct bkey_s_c k,
  26                                unsigned flags, struct printbuf *err)
  27 {
  28         if (bkey_gt(k.k->p, POS(0, U32_MAX)) ||
  29             bkey_lt(k.k->p, POS(0, 1))) {
  30                 prt_printf(err, "bad pos");
  31                 return -BCH_ERR_invalid_bkey;
  32         }
  33
  34         return 0;
  35 }
  36
  37 int bch2_snapshot_tree_lookup(struct btree_trans *trans, u32 id,
  38                               struct bch_snapshot_tree *s)
  39 {
  40         int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_snapshot_trees, POS(0, id),
  41                                           BTREE_ITER_WITH_UPDATES, snapshot_tree, s);
  42
  43         if (bch2_err_matches(ret, ENOENT))
  44                 ret = -BCH_ERR_ENOENT_snapshot_tree;
  45         return ret;
  46 }
  47
  48 static struct bkey_i_snapshot_tree *
  49 __snapshot_tree_create(struct btree_trans *trans)
  50 {
  51         struct btree_iter iter;
  52         int ret = bch2_bkey_get_empty_slot(trans, &iter,
  53                         BTREE_ID_snapshot_trees, POS(0, U32_MAX));
  54         struct bkey_i_snapshot_tree *s_t;
  55
  56         if (ret == -BCH_ERR_ENOSPC_btree_slot)
  57                 ret = -BCH_ERR_ENOSPC_snapshot_tree;
  58         if (ret)
  59                 return ERR_PTR(ret);
  60
  61         s_t = bch2_bkey_alloc(trans, &iter, 0, snapshot_tree);
  62         ret = PTR_ERR_OR_ZERO(s_t);
  63         bch2_trans_iter_exit(trans, &iter);
  64         return ret ? ERR_PTR(ret) : s_t;
  65 }
  66
  67 static int snapshot_tree_create(struct btree_trans *trans,
  68                                 u32 root_id, u32 subvol_id, u32 *tree_id)
  69 {
  70         struct bkey_i_snapshot_tree *n_tree =
  71                 __snapshot_tree_create(trans);
  72
  73         if (IS_ERR(n_tree))
  74                 return PTR_ERR(n_tree);
  75
  76         n_tree->v.master_subvol = cpu_to_le32(subvol_id);
  77         n_tree->v.root_snapshot = cpu_to_le32(root_id);
  78         *tree_id = n_tree->k.p.offset;
  79         return 0;
  80 }
  81
  82 /* Snapshot nodes: */
  83
  84 void bch2_snapshot_to_text(struct printbuf *out, struct bch_fs *c,
  85                            struct bkey_s_c k)
  86 {
  87         struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(k);
  88
  89         prt_printf(out, "is_subvol %llu deleted %llu parent %10u children %10u %10u subvol %u",
  90                BCH_SNAPSHOT_SUBVOL(s.v),
  91                BCH_SNAPSHOT_DELETED(s.v),
  92                le32_to_cpu(s.v->parent),
  93                le32_to_cpu(s.v->children[0]),
  94                le32_to_cpu(s.v->children[1]),
  95                le32_to_cpu(s.v->subvol));
  96 }
  97
  98 int bch2_snapshot_invalid(const struct bch_fs *c, struct bkey_s_c k,
  99                           unsigned flags, struct printbuf *err)
 100 {
 101         struct bkey_s_c_snapshot s;
 102         u32 i, id;
 103
 104         if (bkey_gt(k.k->p, POS(0, U32_MAX)) ||
 105             bkey_lt(k.k->p, POS(0, 1))) {
 106                 prt_printf(err, "bad pos");
 107                 return -BCH_ERR_invalid_bkey;
 108         }
 109
 110         s = bkey_s_c_to_snapshot(k);
 111
 112         id = le32_to_cpu(s.v->parent);
 113         if (id && id <= k.k->p.offset) {
 114                 prt_printf(err, "bad parent node (%u <= %llu)",
 115                        id, k.k->p.offset);
 116                 return -BCH_ERR_invalid_bkey;
 117         }
 118
 119         if (le32_to_cpu(s.v->children[0]) < le32_to_cpu(s.v->children[1])) {
 120                 prt_printf(err, "children not normalized");
 121                 return -BCH_ERR_invalid_bkey;
 122         }
 123
 124         if (s.v->children[0] &&
 125             s.v->children[0] == s.v->children[1]) {
 126                 prt_printf(err, "duplicate child nodes");
 127                 return -BCH_ERR_invalid_bkey;
 128         }
 129
 130         for (i = 0; i < 2; i++) {
 131                 id = le32_to_cpu(s.v->children[i]);
 132
 133                 if (id >= k.k->p.offset) {
 134                         prt_printf(err, "bad child node (%u >= %llu)",
 135                                id, k.k->p.offset);
 136                         return -BCH_ERR_invalid_bkey;
 137                 }
 138         }
 139
 140         return 0;
 141 }
 142
 143 int bch2_mark_snapshot(struct btree_trans *trans,
 144                        enum btree_id btree, unsigned level,
 145                        struct bkey_s_c old, struct bkey_s_c new,
 146                        unsigned flags)
 147 {
 148         struct bch_fs *c = trans->c;
 149         struct snapshot_t *t;
 150
 151         t = genradix_ptr_alloc(&c->snapshots,
 152                                U32_MAX - new.k->p.offset,
 153                                GFP_KERNEL);
 154         if (!t)
 155                 return -BCH_ERR_ENOMEM_mark_snapshot;
 156
 157         if (new.k->type == KEY_TYPE_snapshot) {
 158                 struct bkey_s_c_snapshot s = bkey_s_c_to_snapshot(new);
 159
 160                 t->parent       = le32_to_cpu(s.v->parent);
 161                 t->children[0]  = le32_to_cpu(s.v->children[0]);
 162                 t->children[1]  = le32_to_cpu(s.v->children[1]);
 163                 t->subvol       = BCH_SNAPSHOT_SUBVOL(s.v) ? le32_to_cpu(s.v->subvol) : 0;
 164                 t->tree         = le32_to_cpu(s.v->tree);
 165         } else {
 166                 t->parent       = 0;
 167                 t->children[0]  = 0;
 168                 t->children[1]  = 0;
 169                 t->subvol       = 0;
 170                 t->tree         = 0;
 171         }
 172
 173         return 0;
 174 }
 175
 176 static int snapshot_lookup(struct btree_trans *trans, u32 id,
 177                            struct bch_snapshot *s)
 178 {
 179         return bch2_bkey_get_val_typed(trans, BTREE_ID_snapshots, POS(0, id),
 180                                        BTREE_ITER_WITH_UPDATES, snapshot, s);
 181 }
 182
 183 static int snapshot_live(struct btree_trans *trans, u32 id)
 184 {
 185         struct bch_snapshot v;
 186         int ret;
 187
 188         if (!id)
 189                 return 0;
 190
 191         ret = snapshot_lookup(trans, id, &v);
 192         if (bch2_err_matches(ret, ENOENT))
 193                 bch_err(trans->c, "snapshot node %u not found", id);
 194         if (ret)
 195                 return ret;
 196
 197         return !BCH_SNAPSHOT_DELETED(&v);
 198 }
 199
 200 static int bch2_snapshot_set_equiv(struct btree_trans *trans, struct bkey_s_c k)
 201 {
 202         struct bch_fs *c = trans->c;
 203         unsigned i, nr_live = 0, live_idx = 0;
 204         struct bkey_s_c_snapshot snap;
 205         u32 id = k.k->p.offset, child[2];
 206
 207         if (k.k->type != KEY_TYPE_snapshot)
 208                 return 0;
 209
 210         snap = bkey_s_c_to_snapshot(k);
 211
 212         child[0] = le32_to_cpu(snap.v->children[0]);
 213         child[1] = le32_to_cpu(snap.v->children[1]);
 214
 215         for (i = 0; i < 2; i++) {
 216                 int ret = snapshot_live(trans, child[i]);
 217
 218                 if (ret < 0)
 219                         return ret;
 220
 221                 if (ret)
 222                         live_idx = i;
 223                 nr_live += ret;
 224         }
 225
 226         snapshot_t(c, id)->equiv = nr_live == 1
 227                 ? snapshot_t(c, child[live_idx])->equiv
 228                 : id;
 229         return 0;
 230 }
 231
 232 /* fsck: */
 233
 234 static u32 bch2_snapshot_child(struct bch_fs *c, u32 id, unsigned child)
 235 {
 236         return snapshot_t(c, id)->children[child];
 237 }
 238
 239 static u32 bch2_snapshot_left_child(struct bch_fs *c, u32 id)
 240 {
 241         return bch2_snapshot_child(c, id, 0);
 242 }
 243
 244 static u32 bch2_snapshot_right_child(struct bch_fs *c, u32 id)
 245 {
 246         return bch2_snapshot_child(c, id, 1);
 247 }
 248
 249 static u32 bch2_snapshot_tree_next(struct bch_fs *c, u32 id)
 250 {
 251         u32 n, parent;
 252
 253         n = bch2_snapshot_left_child(c, id);
 254         if (n)
 255                 return n;
 256
 257         while ((parent = bch2_snapshot_parent(c, id))) {
 258                 n = bch2_snapshot_right_child(c, parent);
 259                 if (n && n != id)
 260                         return n;
 261                 id = parent;
 262         }
 263
 264         return 0;
 265 }
 266
 267 static u32 bch2_snapshot_tree_oldest_subvol(struct bch_fs *c, u32 snapshot_root)
 268 {
 269         u32 id = snapshot_root;
 270         u32 subvol = 0, s;
 271
 272         while (id) {
 273                 s = snapshot_t(c, id)->subvol;
 274
 275                 if (s && (!subvol || s < subvol))
 276                         subvol = s;
 277
 278                 id = bch2_snapshot_tree_next(c, id);
 279         }
 280
 281         return subvol;
 282 }
 283
 284 static int bch2_snapshot_tree_master_subvol(struct btree_trans *trans,
 285                                             u32 snapshot_root, u32 *subvol_id)
 286 {
 287         struct bch_fs *c = trans->c;
 288         struct btree_iter iter;
 289         struct bkey_s_c k;
 290         struct bkey_s_c_subvolume s;
 291         bool found = false;
 292         int ret;
 293
 294         for_each_btree_key_norestart(trans, iter, BTREE_ID_subvolumes, POS_MIN,
 295                                      0, k, ret) {
 296                 if (k.k->type != KEY_TYPE_subvolume)
 297                         continue;
 298
 299                 s = bkey_s_c_to_subvolume(k);
 300                 if (!bch2_snapshot_is_ancestor(c, le32_to_cpu(s.v->snapshot), snapshot_root))
 301                         continue;
 302                 if (!BCH_SUBVOLUME_SNAP(s.v)) {
 303                         *subvol_id = s.k->p.offset;
 304                         found = true;
 305                         break;
 306                 }
 307         }
 308
 309         bch2_trans_iter_exit(trans, &iter);
 310
 311         if (!ret && !found) {
 312                 struct bkey_i_subvolume *s;
 313
 314                 *subvol_id = bch2_snapshot_tree_oldest_subvol(c, snapshot_root);
 315
 316                 s = bch2_bkey_get_mut_typed(trans, &iter,
 317                                             BTREE_ID_subvolumes, POS(0, *subvol_id),
 318                                             0, subvolume);
 319                 ret = PTR_ERR_OR_ZERO(s);
 320                 if (ret)
 321                         return ret;
 322
 323                 SET_BCH_SUBVOLUME_SNAP(&s->v, false);
 324         }
 325
 326         return ret;
 327 }
 328
 329 static int check_snapshot_tree(struct btree_trans *trans,
 330                                struct btree_iter *iter,
 331                                struct bkey_s_c k)
 332 {
 333         struct bch_fs *c = trans->c;
 334         struct bkey_s_c_snapshot_tree st;
 335         struct bch_snapshot s;
 336         struct bch_subvolume subvol;
 337         struct printbuf buf = PRINTBUF;
 338         u32 root_id;
 339         int ret;
 340
 341         if (k.k->type != KEY_TYPE_snapshot_tree)
 342                 return 0;
 343
 344         st = bkey_s_c_to_snapshot_tree(k);
 345         root_id = le32_to_cpu(st.v->root_snapshot);
 346
 347         ret = snapshot_lookup(trans, root_id, &s);
 348         if (ret && !bch2_err_matches(ret, ENOENT))
 349                 goto err;
 350
 351         if (fsck_err_on(ret ||
 352                         root_id != bch2_snapshot_root(c, root_id) ||
 353                         st.k->p.offset != le32_to_cpu(s.tree),
 354                         c,
 355                         "snapshot tree points to missing/incorrect snapshot:\n  %s",
 356                         (bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf))) {
 357                 ret = bch2_btree_delete_at(trans, iter, 0);
 358                 goto err;
 359         }
 360
 361         ret = bch2_subvolume_get(trans, le32_to_cpu(st.v->master_subvol),
 362                                  false, 0, &subvol);
 363         if (ret && !bch2_err_matches(ret, ENOENT))
 364                 goto err;
 365
 366         if (fsck_err_on(ret, c,
 367                         "snapshot tree points to missing subvolume:\n  %s",
 368                         (printbuf_reset(&buf),
 369                          bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf)) ||
 370             fsck_err_on(!bch2_snapshot_is_ancestor(c,
 371                                                    le32_to_cpu(subvol.snapshot),
 372                                                    root_id), c,
 373                         "snapshot tree points to subvolume that does not point to snapshot in this tree:\n  %s",
 374                         (printbuf_reset(&buf),
 375                          bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf)) ||
 376             fsck_err_on(BCH_SUBVOLUME_SNAP(&subvol), c,
 377                         "snapshot tree points to snapshot subvolume:\n  %s",
 378                         (printbuf_reset(&buf),
 379                          bch2_bkey_val_to_text(&buf, c, st.s_c), buf.buf))) {
 380                 struct bkey_i_snapshot_tree *u;
 381                 u32 subvol_id;
 382
 383                 ret = bch2_snapshot_tree_master_subvol(trans, root_id, &subvol_id);
 384                 if (ret)
 385                         goto err;
 386
 387                 u = bch2_bkey_make_mut_typed(trans, iter, k, 0, snapshot_tree);
 388                 ret = PTR_ERR_OR_ZERO(u);
 389                 if (ret)
 390                         goto err;
 391
 392                 u->v.master_subvol = cpu_to_le32(subvol_id);
 393                 st = snapshot_tree_i_to_s_c(u);
 394         }
 395 err:
 396 fsck_err:
 397         printbuf_exit(&buf);
 398         return ret;
 399 }
 400
 401 /*
 402  * For each snapshot_tree, make sure it points to the root of a snapshot tree
 403  * and that snapshot entry points back to it, or delete it.
 404  *
 405  * And, make sure it points to a subvolume within that snapshot tree, or correct
 406  * it to point to the oldest subvolume within that snapshot tree.
 407  */
 408 int bch2_fs_check_snapshot_trees(struct bch_fs *c)
 409 {
 410         struct btree_iter iter;
 411         struct bkey_s_c k;
 412         int ret;
 413
 414         ret = bch2_trans_run(c,
 415                 for_each_btree_key_commit(&trans, iter,
 416                         BTREE_ID_snapshot_trees, POS_MIN,
 417                         BTREE_ITER_PREFETCH, k,
 418                         NULL, NULL, BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
 419                 check_snapshot_tree(&trans, &iter, k)));
 420
 421         if (ret)
 422                 bch_err(c, "error %i checking snapshot trees", ret);
 423         return ret;
 424 }
 425
 426 /*
 427  * Look up snapshot tree for @tree_id and find root,
 428  * make sure @snap_id is a descendent:
 429  */
 430 static int snapshot_tree_ptr_good(struct btree_trans *trans,
 431                                   u32 snap_id, u32 tree_id)
 432 {
 433         struct bch_snapshot_tree s_t;
 434         int ret = bch2_snapshot_tree_lookup(trans, tree_id, &s_t);
 435
 436         if (bch2_err_matches(ret, ENOENT))
 437                 return 0;
 438         if (ret)
 439                 return ret;
 440
 441         return bch2_snapshot_is_ancestor(trans->c, snap_id, le32_to_cpu(s_t.root_snapshot));
 442 }
 443
 444 /*
 445  * snapshot_tree pointer was incorrect: look up root snapshot node, make sure
 446  * its snapshot_tree pointer is correct (allocate new one if necessary), then
 447  * update this node's pointer to root node's pointer:
 448  */
 449 static int snapshot_tree_ptr_repair(struct btree_trans *trans,
 450                                     struct btree_iter *iter,
 451                                     struct bkey_s_c_snapshot *s)
 452 {
 453         struct bch_fs *c = trans->c;
 454         struct btree_iter root_iter;
 455         struct bch_snapshot_tree s_t;
 456         struct bkey_s_c_snapshot root;
 457         struct bkey_i_snapshot *u;
 458         u32 root_id = bch2_snapshot_root(c, s->k->p.offset), tree_id;
 459         int ret;
 460
 461         root = bch2_bkey_get_iter_typed(trans, &root_iter,
 462                                BTREE_ID_snapshots, POS(0, root_id),
 463                                BTREE_ITER_WITH_UPDATES, snapshot);
 464         ret = bkey_err(root);
 465         if (ret)
 466                 goto err;
 467
 468         tree_id = le32_to_cpu(root.v->tree);
 469
 470         ret = bch2_snapshot_tree_lookup(trans, tree_id, &s_t);
 471         if (ret && !bch2_err_matches(ret, ENOENT))
 472                 return ret;
 473
 474         if (ret || le32_to_cpu(s_t.root_snapshot) != root_id) {
 475                 u = bch2_bkey_make_mut_typed(trans, &root_iter, root.s_c, 0, snapshot);
 476                 ret =   PTR_ERR_OR_ZERO(u) ?:
 477                         snapshot_tree_create(trans, root_id,
 478                                 bch2_snapshot_tree_oldest_subvol(c, root_id),
 479                                 &tree_id);
 480                 if (ret)
 481                         goto err;
 482
 483                 u->v.tree = cpu_to_le32(tree_id);
 484                 if (s->k->p.snapshot == root_id)
 485                         *s = snapshot_i_to_s_c(u);
 486         }
 487
 488         if (s->k->p.snapshot != root_id) {
 489                 u = bch2_bkey_make_mut_typed(trans, iter, s->s_c, 0, snapshot);
 490                 ret = PTR_ERR_OR_ZERO(u);
 491                 if (ret)
 492                         goto err;
 493
 494                 u->v.tree = cpu_to_le32(tree_id);
 495                 *s = snapshot_i_to_s_c(u);
 496         }
 497 err:
 498         bch2_trans_iter_exit(trans, &root_iter);
 499         return ret;
 500 }
 501
 502 static int check_snapshot(struct btree_trans *trans,
 503                           struct btree_iter *iter,
 504                           struct bkey_s_c k)
 505 {
 506         struct bch_fs *c = trans->c;
 507         struct bkey_s_c_snapshot s;
 508         struct bch_subvolume subvol;
 509         struct bch_snapshot v;
 510         struct printbuf buf = PRINTBUF;
 511         bool should_have_subvol;
 512         u32 i, id;
 513         int ret = 0;
 514
 515         if (k.k->type != KEY_TYPE_snapshot)
 516                 return 0;
 517
 518         s = bkey_s_c_to_snapshot(k);
 519         id = le32_to_cpu(s.v->parent);
 520         if (id) {
 521                 ret = snapshot_lookup(trans, id, &v);
 522                 if (bch2_err_matches(ret, ENOENT))
 523                         bch_err(c, "snapshot with nonexistent parent:\n  %s",
 524                                 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf));
 525                 if (ret)
 526                         goto err;
 527
 528                 if (le32_to_cpu(v.children[0]) != s.k->p.offset &&
 529                     le32_to_cpu(v.children[1]) != s.k->p.offset) {
 530                         bch_err(c, "snapshot parent %u missing pointer to child %llu",
 531                                 id, s.k->p.offset);
 532                         ret = -EINVAL;
 533                         goto err;
 534                 }
 535         }
 536
 537         for (i = 0; i < 2 && s.v->children[i]; i++) {
 538                 id = le32_to_cpu(s.v->children[i]);
 539
 540                 ret = snapshot_lookup(trans, id, &v);
 541                 if (bch2_err_matches(ret, ENOENT))
 542                         bch_err(c, "snapshot node %llu has nonexistent child %u",
 543                                 s.k->p.offset, id);
 544                 if (ret)
 545                         goto err;
 546
 547                 if (le32_to_cpu(v.parent) != s.k->p.offset) {
 548                         bch_err(c, "snapshot child %u has wrong parent (got %u should be %llu)",
 549                                 id, le32_to_cpu(v.parent), s.k->p.offset);
 550                         ret = -EINVAL;
 551                         goto err;
 552                 }
 553         }
 554
 555         should_have_subvol = BCH_SNAPSHOT_SUBVOL(s.v) &&
 556                 !BCH_SNAPSHOT_DELETED(s.v);
 557
 558         if (should_have_subvol) {
 559                 id = le32_to_cpu(s.v->subvol);
 560                 ret = bch2_subvolume_get(trans, id, 0, false, &subvol);
 561                 if (bch2_err_matches(ret, ENOENT))
 562                         bch_err(c, "snapshot points to nonexistent subvolume:\n  %s",
 563                                 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf));
 564                 if (ret)
 565                         goto err;
 566
 567                 if (BCH_SNAPSHOT_SUBVOL(s.v) != (le32_to_cpu(subvol.snapshot) == s.k->p.offset)) {
 568                         bch_err(c, "snapshot node %llu has wrong BCH_SNAPSHOT_SUBVOL",
 569                                 s.k->p.offset);
 570                         ret = -EINVAL;
 571                         goto err;
 572                 }
 573         } else {
 574                 if (fsck_err_on(s.v->subvol, c, "snapshot should not point to subvol:\n  %s",
 575                                 (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
 576                         struct bkey_i_snapshot *u = bch2_trans_kmalloc(trans, sizeof(*u));
 577
 578                         ret = PTR_ERR_OR_ZERO(u);
 579                         if (ret)
 580                                 goto err;
 581
 582                         bkey_reassemble(&u->k_i, s.s_c);
 583                         u->v.subvol = 0;
 584                         ret = bch2_trans_update(trans, iter, &u->k_i, 0);
 585                         if (ret)
 586                                 goto err;
 587
 588                         s = snapshot_i_to_s_c(u);
 589                 }
 590         }
 591
 592         ret = snapshot_tree_ptr_good(trans, s.k->p.offset, le32_to_cpu(s.v->tree));
 593         if (ret < 0)
 594                 goto err;
 595
 596         if (fsck_err_on(!ret, c, "snapshot points to missing/incorrect tree:\n  %s",
 597                         (bch2_bkey_val_to_text(&buf, c, s.s_c), buf.buf))) {
 598                 ret = snapshot_tree_ptr_repair(trans, iter, &s);
 599                 if (ret)
 600                         goto err;
 601         }
 602         ret = 0;
 603
 604         if (BCH_SNAPSHOT_DELETED(s.v))
 605                 set_bit(BCH_FS_HAVE_DELETED_SNAPSHOTS, &c->flags);
 606 err:
 607 fsck_err:
 608         printbuf_exit(&buf);
 609         return ret;
 610 }
 611
 612 int bch2_fs_check_snapshots(struct bch_fs *c)
 613 {
 614         struct btree_iter iter;
 615         struct bkey_s_c k;
 616         int ret;
 617
 618         ret = bch2_trans_run(c,
 619                 for_each_btree_key_commit(&trans, iter,
 620                         BTREE_ID_snapshots, POS_MIN,
 621                         BTREE_ITER_PREFETCH, k,
 622                         NULL, NULL, BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
 623                 check_snapshot(&trans, &iter, k)));
 624         if (ret)
 625                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 626         return ret;
 627 }
 628
 629 static int check_subvol(struct btree_trans *trans,
 630                         struct btree_iter *iter,
 631                         struct bkey_s_c k)
 632 {
 633         struct bch_fs *c = trans->c;
 634         struct bkey_s_c_subvolume subvol;
 635         struct bch_snapshot snapshot;
 636         unsigned snapid;
 637         int ret = 0;
 638
 639         if (k.k->type != KEY_TYPE_subvolume)
 640                 return 0;
 641
 642         subvol = bkey_s_c_to_subvolume(k);
 643         snapid = le32_to_cpu(subvol.v->snapshot);
 644         ret = snapshot_lookup(trans, snapid, &snapshot);
 645
 646         if (bch2_err_matches(ret, ENOENT))
 647                 bch_err(c, "subvolume %llu points to nonexistent snapshot %u",
 648                         k.k->p.offset, snapid);
 649         if (ret)
 650                 return ret;
 651
 652         if (BCH_SUBVOLUME_UNLINKED(subvol.v)) {
 653                 bch2_fs_lazy_rw(c);
 654
 655                 ret = bch2_subvolume_delete(trans, iter->pos.offset);
 656                 if (ret)
 657                         bch_err(c, "error deleting subvolume %llu: %s",
 658                                 iter->pos.offset, bch2_err_str(ret));
 659                 return ret ?: -BCH_ERR_transaction_restart_nested;
 660         }
 661
 662         if (!BCH_SUBVOLUME_SNAP(subvol.v)) {
 663                 u32 snapshot_root = bch2_snapshot_root(c, le32_to_cpu(subvol.v->snapshot));
 664                 u32 snapshot_tree = snapshot_t(c, snapshot_root)->tree;
 665                 struct bch_snapshot_tree st;
 666
 667                 ret = bch2_snapshot_tree_lookup(trans, snapshot_tree, &st);
 668
 669                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
 670                                 "%s: snapshot tree %u not found", __func__, snapshot_tree);
 671
 672                 if (ret)
 673                         return ret;
 674
 675                 if (fsck_err_on(le32_to_cpu(st.master_subvol) != subvol.k->p.offset, c,
 676                                 "subvolume %llu is not set as snapshot but is not master subvolume",
 677                                 k.k->p.offset)) {
 678                         struct bkey_i_subvolume *s =
 679                                 bch2_bkey_make_mut_typed(trans, iter, subvol.s_c, 0, subvolume);
 680                         ret = PTR_ERR_OR_ZERO(s);
 681                         if (ret)
 682                                 return ret;
 683
 684                         SET_BCH_SUBVOLUME_SNAP(&s->v, true);
 685                 }
 686         }
 687
 688 fsck_err:
 689         return ret;
 690 }
 691
 692 int bch2_fs_check_subvols(struct bch_fs *c)
 693 {
 694         struct btree_iter iter;
 695         struct bkey_s_c k;
 696         int ret;
 697
 698         ret = bch2_trans_run(c,
 699                 for_each_btree_key_commit(&trans, iter,
 700                         BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
 701                         NULL, NULL, BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
 702                 check_subvol(&trans, &iter, k)));
 703         if (ret)
 704                 bch_err(c, "%s: error %s", __func__, bch2_err_str(ret));
 705
 706         return ret;
 707 }
 708
 709 void bch2_fs_snapshots_exit(struct bch_fs *c)
 710 {
 711         genradix_free(&c->snapshots);
 712 }
 713
 714 int bch2_fs_snapshots_start(struct bch_fs *c)
 715 {
 716         struct btree_iter iter;
 717         struct bkey_s_c k;
 718         int ret = 0;
 719
 720         ret = bch2_trans_run(c,
 721                 for_each_btree_key2(&trans, iter, BTREE_ID_snapshots,
 722                            POS_MIN, 0, k,
 723                         bch2_mark_snapshot(&trans, BTREE_ID_snapshots, 0, bkey_s_c_null, k, 0) ?:
 724                         bch2_snapshot_set_equiv(&trans, k)));
 725         if (ret)
 726                 bch_err(c, "error starting snapshots: %s", bch2_err_str(ret));
 727         return ret;
 728 }
 729
 730 /*
 731  * Mark a snapshot as deleted, for future cleanup:
 732  */
 733 static int bch2_snapshot_node_set_deleted(struct btree_trans *trans, u32 id)
 734 {
 735         struct btree_iter iter;
 736         struct bkey_i_snapshot *s;
 737         int ret = 0;
 738
 739         s = bch2_bkey_get_mut_typed(trans, &iter,
 740                                     BTREE_ID_snapshots, POS(0, id),
 741                                     0, snapshot);
 742         ret = PTR_ERR_OR_ZERO(s);
 743         if (unlikely(ret)) {
 744                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT),
 745                                         trans->c, "missing snapshot %u", id);
 746                 return ret;
 747         }
 748
 749         /* already deleted? */
 750         if (BCH_SNAPSHOT_DELETED(&s->v))
 751                 goto err;
 752
 753         SET_BCH_SNAPSHOT_DELETED(&s->v, true);
 754         SET_BCH_SNAPSHOT_SUBVOL(&s->v, false);
 755         s->v.subvol = 0;
 756 err:
 757         bch2_trans_iter_exit(trans, &iter);
 758         return ret;
 759 }
 760
 761 static int bch2_snapshot_node_delete(struct btree_trans *trans, u32 id)
 762 {
 763         struct bch_fs *c = trans->c;
 764         struct btree_iter iter, p_iter = (struct btree_iter) { NULL };
 765         struct btree_iter tree_iter = (struct btree_iter) { NULL };
 766         struct bkey_s_c_snapshot s;
 767         u32 parent_id;
 768         unsigned i;
 769         int ret = 0;
 770
 771         s = bch2_bkey_get_iter_typed(trans, &iter, BTREE_ID_snapshots, POS(0, id),
 772                                      BTREE_ITER_INTENT, snapshot);
 773         ret = bkey_err(s);
 774         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
 775                                 "missing snapshot %u", id);
 776
 777         if (ret)
 778                 goto err;
 779
 780         BUG_ON(!BCH_SNAPSHOT_DELETED(s.v));
 781         parent_id = le32_to_cpu(s.v->parent);
 782
 783         if (parent_id) {
 784                 struct bkey_i_snapshot *parent;
 785
 786                 parent = bch2_bkey_get_mut_typed(trans, &p_iter,
 787                                      BTREE_ID_snapshots, POS(0, parent_id),
 788                                      0, snapshot);
 789                 ret = PTR_ERR_OR_ZERO(parent);
 790                 if (unlikely(ret)) {
 791                         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
 792                                                 "missing snapshot %u", parent_id);
 793                         goto err;
 794                 }
 795
 796                 for (i = 0; i < 2; i++)
 797                         if (le32_to_cpu(parent->v.children[i]) == id)
 798                                 break;
 799
 800                 if (i == 2)
 801                         bch_err(c, "snapshot %u missing child pointer to %u",
 802                                 parent_id, id);
 803                 else
 804                         parent->v.children[i] = 0;
 805
 806                 if (le32_to_cpu(parent->v.children[0]) <
 807                     le32_to_cpu(parent->v.children[1]))
 808                         swap(parent->v.children[0],
 809                              parent->v.children[1]);
 810         } else {
 811                 /*
 812                  * We're deleting the root of a snapshot tree: update the
 813                  * snapshot_tree entry to point to the new root, or delete it if
 814                  * this is the last snapshot ID in this tree:
 815                  */
 816                 struct bkey_i_snapshot_tree *s_t;
 817
 818                 BUG_ON(s.v->children[1]);
 819
 820                 s_t = bch2_bkey_get_mut_typed(trans, &tree_iter,
 821                                 BTREE_ID_snapshot_trees, POS(0, le32_to_cpu(s.v->tree)),
 822                                 0, snapshot_tree);
 823                 ret = PTR_ERR_OR_ZERO(s_t);
 824                 if (ret)
 825                         goto err;
 826
 827                 if (s.v->children[0]) {
 828                         s_t->v.root_snapshot = cpu_to_le32(s.v->children[0]);
 829                 } else {
 830                         s_t->k.type = KEY_TYPE_deleted;
 831                         set_bkey_val_u64s(&s_t->k, 0);
 832                 }
 833         }
 834
 835         ret = bch2_btree_delete_at(trans, &iter, 0);
 836 err:
 837         bch2_trans_iter_exit(trans, &tree_iter);
 838         bch2_trans_iter_exit(trans, &p_iter);
 839         bch2_trans_iter_exit(trans, &iter);
 840         return ret;
 841 }
 842
 843 static int create_snapids(struct btree_trans *trans, u32 parent, u32 tree,
 844                           u32 *new_snapids,
 845                           u32 *snapshot_subvols,
 846                           unsigned nr_snapids)
 847 {
 848         struct btree_iter iter;
 849         struct bkey_i_snapshot *n;
 850         struct bkey_s_c k;
 851         unsigned i;
 852         int ret;
 853
 854         bch2_trans_iter_init(trans, &iter, BTREE_ID_snapshots,
 855                              POS_MIN, BTREE_ITER_INTENT);
 856         k = bch2_btree_iter_peek(&iter);
 857         ret = bkey_err(k);
 858         if (ret)
 859                 goto err;
 860
 861         for (i = 0; i < nr_snapids; i++) {
 862                 k = bch2_btree_iter_prev_slot(&iter);
 863                 ret = bkey_err(k);
 864                 if (ret)
 865                         goto err;
 866
 867                 if (!k.k || !k.k->p.offset) {
 868                         ret = -BCH_ERR_ENOSPC_snapshot_create;
 869                         goto err;
 870                 }
 871
 872                 n = bch2_bkey_alloc(trans, &iter, 0, snapshot);
 873                 ret = PTR_ERR_OR_ZERO(n);
 874                 if (ret)
 875                         goto err;
 876
 877                 n->v.flags      = 0;
 878                 n->v.parent     = cpu_to_le32(parent);
 879                 n->v.subvol     = cpu_to_le32(snapshot_subvols[i]);
 880                 n->v.tree       = cpu_to_le32(tree);
 881                 SET_BCH_SNAPSHOT_SUBVOL(&n->v, true);
 882
 883                 ret = bch2_mark_snapshot(trans, BTREE_ID_snapshots, 0,
 884                                          bkey_s_c_null, bkey_i_to_s_c(&n->k_i), 0);
 885                 if (ret)
 886                         goto err;
 887
 888                 new_snapids[i]  = iter.pos.offset;
 889         }
 890 err:
 891         bch2_trans_iter_exit(trans, &iter);
 892         return ret;
 893 }
 894
 895 /*
 896  * Create new snapshot IDs as children of an existing snapshot ID:
 897  */
 898 static int bch2_snapshot_node_create_children(struct btree_trans *trans, u32 parent,
 899                               u32 *new_snapids,
 900                               u32 *snapshot_subvols,
 901                               unsigned nr_snapids)
 902 {
 903         struct btree_iter iter;
 904         struct bkey_i_snapshot *n_parent;
 905         int ret = 0;
 906
 907         n_parent = bch2_bkey_get_mut_typed(trans, &iter,
 908                         BTREE_ID_snapshots, POS(0, parent),
 909                         0, snapshot);
 910         ret = PTR_ERR_OR_ZERO(n_parent);
 911         if (unlikely(ret)) {
 912                 if (bch2_err_matches(ret, ENOENT))
 913                         bch_err(trans->c, "snapshot %u not found", parent);
 914                 return ret;
 915         }
 916
 917         if (n_parent->v.children[0] || n_parent->v.children[1]) {
 918                 bch_err(trans->c, "Trying to add child snapshot nodes to parent that already has children");
 919                 ret = -EINVAL;
 920                 goto err;
 921         }
 922
 923         ret = create_snapids(trans, parent, le32_to_cpu(n_parent->v.tree),
 924                              new_snapids, snapshot_subvols, nr_snapids);
 925         if (ret)
 926                 goto err;
 927
 928         n_parent->v.children[0] = cpu_to_le32(new_snapids[0]);
 929         n_parent->v.children[1] = cpu_to_le32(new_snapids[1]);
 930         n_parent->v.subvol = 0;
 931         SET_BCH_SNAPSHOT_SUBVOL(&n_parent->v, false);
 932 err:
 933         bch2_trans_iter_exit(trans, &iter);
 934         return ret;
 935 }
 936
 937 /*
 938  * Create a snapshot node that is the root of a new tree:
 939  */
 940 static int bch2_snapshot_node_create_tree(struct btree_trans *trans,
 941                               u32 *new_snapids,
 942                               u32 *snapshot_subvols,
 943                               unsigned nr_snapids)
 944 {
 945         struct bkey_i_snapshot_tree *n_tree;
 946         int ret;
 947
 948         n_tree = __snapshot_tree_create(trans);
 949         ret =   PTR_ERR_OR_ZERO(n_tree) ?:
 950                 create_snapids(trans, 0, n_tree->k.p.offset,
 951                              new_snapids, snapshot_subvols, nr_snapids);
 952         if (ret)
 953                 return ret;
 954
 955         n_tree->v.master_subvol = cpu_to_le32(snapshot_subvols[0]);
 956         n_tree->v.root_snapshot = cpu_to_le32(new_snapids[0]);
 957         return 0;
 958 }
 959
 960 int bch2_snapshot_node_create(struct btree_trans *trans, u32 parent,
 961                               u32 *new_snapids,
 962                               u32 *snapshot_subvols,
 963                               unsigned nr_snapids)
 964 {
 965         BUG_ON((parent == 0) != (nr_snapids == 1));
 966         BUG_ON((parent != 0) != (nr_snapids == 2));
 967
 968         return parent
 969                 ? bch2_snapshot_node_create_children(trans, parent,
 970                                 new_snapids, snapshot_subvols, nr_snapids)
 971                 : bch2_snapshot_node_create_tree(trans,
 972                                 new_snapids, snapshot_subvols, nr_snapids);
 973
 974 }
 975
 976 static int snapshot_delete_key(struct btree_trans *trans,
 977                                struct btree_iter *iter,
 978                                struct bkey_s_c k,
 979                                snapshot_id_list *deleted,
 980                                snapshot_id_list *equiv_seen,
 981                                struct bpos *last_pos)
 982 {
 983         struct bch_fs *c = trans->c;
 984         u32 equiv = snapshot_t(c, k.k->p.snapshot)->equiv;
 985
 986         if (!bkey_eq(k.k->p, *last_pos))
 987                 equiv_seen->nr = 0;
 988         *last_pos = k.k->p;
 989
 990         if (snapshot_list_has_id(deleted, k.k->p.snapshot) ||
 991             snapshot_list_has_id(equiv_seen, equiv)) {
 992                 return bch2_btree_delete_at(trans, iter,
 993                                             BTREE_UPDATE_INTERNAL_SNAPSHOT_NODE);
 994         } else {
 995                 return snapshot_list_add(c, equiv_seen, equiv);
 996         }
 997 }
 998
 999 static int bch2_delete_redundant_snapshot(struct btree_trans *trans, struct btree_iter *iter,
1000                                           struct bkey_s_c k)
1001 {
1002         struct bkey_s_c_snapshot snap;
1003         u32 children[2];
1004         int ret;
1005
1006         if (k.k->type != KEY_TYPE_snapshot)
1007                 return 0;
1008
1009         snap = bkey_s_c_to_snapshot(k);
1010         if (BCH_SNAPSHOT_DELETED(snap.v) ||
1011             BCH_SNAPSHOT_SUBVOL(snap.v))
1012                 return 0;
1013
1014         children[0] = le32_to_cpu(snap.v->children[0]);
1015         children[1] = le32_to_cpu(snap.v->children[1]);
1016
1017         ret   = snapshot_live(trans, children[0]) ?:
1018                 snapshot_live(trans, children[1]);
1019         if (ret < 0)
1020                 return ret;
1021
1022         if (!ret)
1023                 return bch2_snapshot_node_set_deleted(trans, k.k->p.offset);
1024         return 0;
1025 }
1026
1027 int bch2_delete_dead_snapshots(struct bch_fs *c)
1028 {
1029         struct btree_trans trans;
1030         struct btree_iter iter;
1031         struct bkey_s_c k;
1032         struct bkey_s_c_snapshot snap;
1033         snapshot_id_list deleted = { 0 };
1034         u32 i, id;
1035         int ret = 0;
1036
1037         if (!test_bit(BCH_FS_HAVE_DELETED_SNAPSHOTS, &c->flags))
1038                 return 0;
1039
1040         if (!test_bit(BCH_FS_STARTED, &c->flags)) {
1041                 ret = bch2_fs_read_write_early(c);
1042                 if (ret) {
1043                         bch_err(c, "error deleleting dead snapshots: error going rw: %s", bch2_err_str(ret));
1044                         return ret;
1045                 }
1046         }
1047
1048         bch2_trans_init(&trans, c, 0, 0);
1049
1050         /*
1051          * For every snapshot node: If we have no live children and it's not
1052          * pointed to by a subvolume, delete it:
1053          */
1054         ret = for_each_btree_key_commit(&trans, iter, BTREE_ID_snapshots,
1055                         POS_MIN, 0, k,
1056                         NULL, NULL, 0,
1057                 bch2_delete_redundant_snapshot(&trans, &iter, k));
1058         if (ret) {
1059                 bch_err(c, "error deleting redundant snapshots: %s", bch2_err_str(ret));
1060                 goto err;
1061         }
1062
1063         for_each_btree_key2(&trans, iter, BTREE_ID_snapshots,
1064                            POS_MIN, 0, k,
1065                 bch2_snapshot_set_equiv(&trans, k));
1066         if (ret) {
1067                 bch_err(c, "error in bch2_snapshots_set_equiv: %s", bch2_err_str(ret));
1068                 goto err;
1069         }
1070
1071         for_each_btree_key(&trans, iter, BTREE_ID_snapshots,
1072                            POS_MIN, 0, k, ret) {
1073                 if (k.k->type != KEY_TYPE_snapshot)
1074                         continue;
1075
1076                 snap = bkey_s_c_to_snapshot(k);
1077                 if (BCH_SNAPSHOT_DELETED(snap.v)) {
1078                         ret = snapshot_list_add(c, &deleted, k.k->p.offset);
1079                         if (ret)
1080                                 break;
1081                 }
1082         }
1083         bch2_trans_iter_exit(&trans, &iter);
1084
1085         if (ret) {
1086                 bch_err(c, "error walking snapshots: %s", bch2_err_str(ret));
1087                 goto err;
1088         }
1089
1090         for (id = 0; id < BTREE_ID_NR; id++) {
1091                 struct bpos last_pos = POS_MIN;
1092                 snapshot_id_list equiv_seen = { 0 };
1093
1094                 if (!btree_type_has_snapshots(id))
1095                         continue;
1096
1097                 ret = for_each_btree_key_commit(&trans, iter,
1098                                 id, POS_MIN,
1099                                 BTREE_ITER_PREFETCH|BTREE_ITER_ALL_SNAPSHOTS, k,
1100                                 NULL, NULL, BTREE_INSERT_NOFAIL,
1101                         snapshot_delete_key(&trans, &iter, k, &deleted, &equiv_seen, &last_pos));
1102
1103                 darray_exit(&equiv_seen);
1104
1105                 if (ret) {
1106                         bch_err(c, "error deleting snapshot keys: %s", bch2_err_str(ret));
1107                         goto err;
1108                 }
1109         }
1110
1111         for (i = 0; i < deleted.nr; i++) {
1112                 ret = commit_do(&trans, NULL, NULL, 0,
1113                         bch2_snapshot_node_delete(&trans, deleted.data[i]));
1114                 if (ret) {
1115                         bch_err(c, "error deleting snapshot %u: %s",
1116                                 deleted.data[i], bch2_err_str(ret));
1117                         goto err;
1118                 }
1119         }
1120
1121         clear_bit(BCH_FS_HAVE_DELETED_SNAPSHOTS, &c->flags);
1122 err:
1123         darray_exit(&deleted);
1124         bch2_trans_exit(&trans);
1125         return ret;
1126 }
1127
1128 static void bch2_delete_dead_snapshots_work(struct work_struct *work)
1129 {
1130         struct bch_fs *c = container_of(work, struct bch_fs, snapshot_delete_work);
1131
1132         bch2_delete_dead_snapshots(c);
1133         bch2_write_ref_put(c, BCH_WRITE_REF_delete_dead_snapshots);
1134 }
1135
1136 void bch2_delete_dead_snapshots_async(struct bch_fs *c)
1137 {
1138         if (bch2_write_ref_tryget(c, BCH_WRITE_REF_delete_dead_snapshots) &&
1139             !queue_work(c->write_ref_wq, &c->snapshot_delete_work))
1140                 bch2_write_ref_put(c, BCH_WRITE_REF_delete_dead_snapshots);
1141 }
1142
1143 static int bch2_delete_dead_snapshots_hook(struct btree_trans *trans,
1144                                            struct btree_trans_commit_hook *h)
1145 {
1146         struct bch_fs *c = trans->c;
1147
1148         set_bit(BCH_FS_HAVE_DELETED_SNAPSHOTS, &c->flags);
1149
1150         if (!test_bit(BCH_FS_FSCK_DONE, &c->flags))
1151                 return 0;
1152
1153         bch2_delete_dead_snapshots_async(c);
1154         return 0;
1155 }
1156
1157 /* Subvolumes: */
1158
1159 int bch2_subvolume_invalid(const struct bch_fs *c, struct bkey_s_c k,
1160                            unsigned flags, struct printbuf *err)
1161 {
1162         if (bkey_lt(k.k->p, SUBVOL_POS_MIN) ||
1163             bkey_gt(k.k->p, SUBVOL_POS_MAX)) {
1164                 prt_printf(err, "invalid pos");
1165                 return -BCH_ERR_invalid_bkey;
1166         }
1167
1168         return 0;
1169 }
1170
1171 void bch2_subvolume_to_text(struct printbuf *out, struct bch_fs *c,
1172                             struct bkey_s_c k)
1173 {
1174         struct bkey_s_c_subvolume s = bkey_s_c_to_subvolume(k);
1175
1176         prt_printf(out, "root %llu snapshot id %u",
1177                    le64_to_cpu(s.v->inode),
1178                    le32_to_cpu(s.v->snapshot));
1179
1180         if (bkey_val_bytes(s.k) > offsetof(struct bch_subvolume, parent))
1181                 prt_printf(out, " parent %u", le32_to_cpu(s.v->parent));
1182 }
1183
1184 static __always_inline int
1185 bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
1186                            bool inconsistent_if_not_found,
1187                            int iter_flags,
1188                            struct bch_subvolume *s)
1189 {
1190         int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_subvolumes, POS(0, subvol),
1191                                           iter_flags, subvolume, s);
1192         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT) &&
1193                                 inconsistent_if_not_found,
1194                                 trans->c, "missing subvolume %u", subvol);
1195         return ret;
1196 }
1197
1198 int bch2_subvolume_get(struct btree_trans *trans, unsigned subvol,
1199                        bool inconsistent_if_not_found,
1200                        int iter_flags,
1201                        struct bch_subvolume *s)
1202 {
1203         return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, iter_flags, s);
1204 }
1205
1206 int bch2_snapshot_get_subvol(struct btree_trans *trans, u32 snapshot,
1207                              struct bch_subvolume *subvol)
1208 {
1209         struct bch_snapshot snap;
1210
1211         return  snapshot_lookup(trans, snapshot, &snap) ?:
1212                 bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, 0, subvol);
1213 }
1214
1215 int bch2_subvolume_get_snapshot(struct btree_trans *trans, u32 subvol,
1216                                 u32 *snapid)
1217 {
1218         struct btree_iter iter;
1219         struct bkey_s_c k;
1220         int ret;
1221
1222         k = bch2_bkey_get_iter(trans, &iter, BTREE_ID_subvolumes, POS(0, subvol),
1223                                BTREE_ITER_CACHED|
1224                                BTREE_ITER_WITH_UPDATES);
1225         ret = bkey_err(k) ?: k.k->type == KEY_TYPE_subvolume ? 0 : -BCH_ERR_ENOENT_subvolume;
1226
1227         if (likely(!ret))
1228                 *snapid = le32_to_cpu(bkey_s_c_to_subvolume(k).v->snapshot);
1229         else if (bch2_err_matches(ret, ENOENT))
1230                 bch2_fs_inconsistent(trans->c, "missing subvolume %u", subvol);
1231         bch2_trans_iter_exit(trans, &iter);
1232         return ret;
1233 }
1234
1235 static int bch2_subvolume_reparent(struct btree_trans *trans,
1236                                    struct btree_iter *iter,
1237                                    struct bkey_s_c k,
1238                                    u32 old_parent, u32 new_parent)
1239 {
1240         struct bkey_i_subvolume *s;
1241         int ret;
1242
1243         if (k.k->type != KEY_TYPE_subvolume)
1244                 return 0;
1245
1246         if (bkey_val_bytes(k.k) > offsetof(struct bch_subvolume, parent) &&
1247             le32_to_cpu(bkey_s_c_to_subvolume(k).v->parent) != old_parent)
1248                 return 0;
1249
1250         s = bch2_bkey_make_mut_typed(trans, iter, k, 0, subvolume);
1251         ret = PTR_ERR_OR_ZERO(s);
1252         if (ret)
1253                 return ret;
1254
1255         s->v.parent = cpu_to_le32(new_parent);
1256         return 0;
1257 }
1258
1259 /*
1260  * Scan for subvolumes with parent @subvolid_to_delete, reparent:
1261  */
1262 static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_delete)
1263 {
1264         struct btree_iter iter;
1265         struct bkey_s_c k;
1266         struct bch_subvolume s;
1267
1268         return lockrestart_do(trans,
1269                         bch2_subvolume_get(trans, subvolid_to_delete, true,
1270                                    BTREE_ITER_CACHED, &s)) ?:
1271                 for_each_btree_key_commit(trans, iter,
1272                                 BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
1273                                 NULL, NULL, BTREE_INSERT_NOFAIL,
1274                         bch2_subvolume_reparent(trans, &iter, k,
1275                                         subvolid_to_delete, le32_to_cpu(s.parent)));
1276 }
1277
1278 /*
1279  * Delete subvolume, mark snapshot ID as deleted, queue up snapshot
1280  * deletion/cleanup:
1281  */
1282 static int __bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
1283 {
1284         struct btree_iter iter;
1285         struct bkey_s_c_subvolume subvol;
1286         struct btree_trans_commit_hook *h;
1287         u32 snapid;
1288         int ret = 0;
1289
1290         subvol = bch2_bkey_get_iter_typed(trans, &iter,
1291                                 BTREE_ID_subvolumes, POS(0, subvolid),
1292                                 BTREE_ITER_CACHED|BTREE_ITER_INTENT,
1293                                 subvolume);
1294         ret = bkey_err(subvol);
1295         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
1296                                 "missing subvolume %u", subvolid);
1297         if (ret)
1298                 return ret;
1299
1300         snapid = le32_to_cpu(subvol.v->snapshot);
1301
1302         ret = bch2_btree_delete_at(trans, &iter, 0);
1303         if (ret)
1304                 goto err;
1305
1306         ret = bch2_snapshot_node_set_deleted(trans, snapid);
1307         if (ret)
1308                 goto err;
1309
1310         h = bch2_trans_kmalloc(trans, sizeof(*h));
1311         ret = PTR_ERR_OR_ZERO(h);
1312         if (ret)
1313                 goto err;
1314
1315         h->fn = bch2_delete_dead_snapshots_hook;
1316         bch2_trans_commit_hook(trans, h);
1317 err:
1318         bch2_trans_iter_exit(trans, &iter);
1319         return ret;
1320 }
1321
1322 static int bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
1323 {
1324         return bch2_subvolumes_reparent(trans, subvolid) ?:
1325                 commit_do(trans, NULL, NULL, BTREE_INSERT_NOFAIL,
1326                           __bch2_subvolume_delete(trans, subvolid));
1327 }
1328
1329 void bch2_subvolume_wait_for_pagecache_and_delete(struct work_struct *work)
1330 {
1331         struct bch_fs *c = container_of(work, struct bch_fs,
1332                                 snapshot_wait_for_pagecache_and_delete_work);
1333         snapshot_id_list s;
1334         u32 *id;
1335         int ret = 0;
1336
1337         while (!ret) {
1338                 mutex_lock(&c->snapshots_unlinked_lock);
1339                 s = c->snapshots_unlinked;
1340                 darray_init(&c->snapshots_unlinked);
1341                 mutex_unlock(&c->snapshots_unlinked_lock);
1342
1343                 if (!s.nr)
1344                         break;
1345
1346                 bch2_evict_subvolume_inodes(c, &s);
1347
1348                 for (id = s.data; id < s.data + s.nr; id++) {
1349                         ret = bch2_trans_run(c, bch2_subvolume_delete(&trans, *id));
1350                         if (ret) {
1351                                 bch_err(c, "error deleting subvolume %u: %s", *id, bch2_err_str(ret));
1352                                 break;
1353                         }
1354                 }
1355
1356                 darray_exit(&s);
1357         }
1358
1359         bch2_write_ref_put(c, BCH_WRITE_REF_snapshot_delete_pagecache);
1360 }
1361
1362 struct subvolume_unlink_hook {
1363         struct btree_trans_commit_hook  h;
1364         u32                             subvol;
1365 };
1366
1367 int bch2_subvolume_wait_for_pagecache_and_delete_hook(struct btree_trans *trans,
1368                                                       struct btree_trans_commit_hook *_h)
1369 {
1370         struct subvolume_unlink_hook *h = container_of(_h, struct subvolume_unlink_hook, h);
1371         struct bch_fs *c = trans->c;
1372         int ret = 0;
1373
1374         mutex_lock(&c->snapshots_unlinked_lock);
1375         if (!snapshot_list_has_id(&c->snapshots_unlinked, h->subvol))
1376                 ret = snapshot_list_add(c, &c->snapshots_unlinked, h->subvol);
1377         mutex_unlock(&c->snapshots_unlinked_lock);
1378
1379         if (ret)
1380                 return ret;
1381
1382         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_snapshot_delete_pagecache))
1383                 return -EROFS;
1384
1385         if (!queue_work(c->write_ref_wq, &c->snapshot_wait_for_pagecache_and_delete_work))
1386                 bch2_write_ref_put(c, BCH_WRITE_REF_snapshot_delete_pagecache);
1387         return 0;
1388 }
1389
1390 int bch2_subvolume_unlink(struct btree_trans *trans, u32 subvolid)
1391 {
1392         struct btree_iter iter;
1393         struct bkey_i_subvolume *n;
1394         struct subvolume_unlink_hook *h;
1395         int ret = 0;
1396
1397         h = bch2_trans_kmalloc(trans, sizeof(*h));
1398         ret = PTR_ERR_OR_ZERO(h);
1399         if (ret)
1400                 return ret;
1401
1402         h->h.fn         = bch2_subvolume_wait_for_pagecache_and_delete_hook;
1403         h->subvol       = subvolid;
1404         bch2_trans_commit_hook(trans, &h->h);
1405
1406         n = bch2_bkey_get_mut_typed(trans, &iter,
1407                         BTREE_ID_subvolumes, POS(0, subvolid),
1408                         BTREE_ITER_CACHED, subvolume);
1409         ret = PTR_ERR_OR_ZERO(n);
1410         if (unlikely(ret)) {
1411                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
1412                                         "missing subvolume %u", subvolid);
1413                 return ret;
1414         }
1415
1416         SET_BCH_SUBVOLUME_UNLINKED(&n->v, true);
1417         bch2_trans_iter_exit(trans, &iter);
1418         return ret;
1419 }
1420
1421 int bch2_subvolume_create(struct btree_trans *trans, u64 inode,
1422                           u32 src_subvolid,
1423                           u32 *new_subvolid,
1424                           u32 *new_snapshotid,
1425                           bool ro)
1426 {
1427         struct bch_fs *c = trans->c;
1428         struct btree_iter dst_iter, src_iter = (struct btree_iter) { NULL };
1429         struct bkey_i_subvolume *new_subvol = NULL;
1430         struct bkey_i_subvolume *src_subvol = NULL;
1431         u32 parent = 0, new_nodes[2], snapshot_subvols[2];
1432         int ret = 0;
1433
1434         ret = bch2_bkey_get_empty_slot(trans, &dst_iter,
1435                                 BTREE_ID_subvolumes, POS(0, U32_MAX));
1436         if (ret == -BCH_ERR_ENOSPC_btree_slot)
1437                 ret = -BCH_ERR_ENOSPC_subvolume_create;
1438         if (ret)
1439                 return ret;
1440
1441         snapshot_subvols[0] = dst_iter.pos.offset;
1442         snapshot_subvols[1] = src_subvolid;
1443
1444         if (src_subvolid) {
1445                 /* Creating a snapshot: */
1446
1447                 src_subvol = bch2_bkey_get_mut_typed(trans, &src_iter,
1448                                 BTREE_ID_subvolumes, POS(0, src_subvolid),
1449                                 BTREE_ITER_CACHED, subvolume);
1450                 ret = PTR_ERR_OR_ZERO(src_subvol);
1451                 if (unlikely(ret)) {
1452                         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
1453                                                 "subvolume %u not found", src_subvolid);
1454                         goto err;
1455                 }
1456
1457                 parent = le32_to_cpu(src_subvol->v.snapshot);
1458         }
1459
1460         ret = bch2_snapshot_node_create(trans, parent, new_nodes,
1461                                         snapshot_subvols,
1462                                         src_subvolid ? 2 : 1);
1463         if (ret)
1464                 goto err;
1465
1466         if (src_subvolid) {
1467                 src_subvol->v.snapshot = cpu_to_le32(new_nodes[1]);
1468                 ret = bch2_trans_update(trans, &src_iter, &src_subvol->k_i, 0);
1469                 if (ret)
1470                         goto err;
1471         }
1472
1473         new_subvol = bch2_bkey_alloc(trans, &dst_iter, 0, subvolume);
1474         ret = PTR_ERR_OR_ZERO(new_subvol);
1475         if (ret)
1476                 goto err;
1477
1478         new_subvol->v.flags     = 0;
1479         new_subvol->v.snapshot  = cpu_to_le32(new_nodes[0]);
1480         new_subvol->v.inode     = cpu_to_le64(inode);
1481         new_subvol->v.parent    = cpu_to_le32(src_subvolid);
1482         new_subvol->v.otime.lo  = cpu_to_le64(bch2_current_time(c));
1483         new_subvol->v.otime.hi  = 0;
1484
1485         SET_BCH_SUBVOLUME_RO(&new_subvol->v, ro);
1486         SET_BCH_SUBVOLUME_SNAP(&new_subvol->v, src_subvolid != 0);
1487
1488         *new_subvolid   = new_subvol->k.p.offset;
1489         *new_snapshotid = new_nodes[0];
1490 err:
1491         bch2_trans_iter_exit(trans, &src_iter);
1492         bch2_trans_iter_exit(trans, &dst_iter);
1493         return ret;
1494 }
1495
1496 int bch2_fs_subvolumes_init(struct bch_fs *c)
1497 {
1498         INIT_WORK(&c->snapshot_delete_work, bch2_delete_dead_snapshots_work);
1499         INIT_WORK(&c->snapshot_wait_for_pagecache_and_delete_work,
1500                   bch2_subvolume_wait_for_pagecache_and_delete);
1501         mutex_init(&c->snapshots_unlinked_lock);
1502         return 0;
1503 }