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 "snapshot.h"
  10 #include "subvolume.h"
  11
  12 #include <linux/random.h>
  13
  14 static int bch2_subvolume_delete(struct btree_trans *, u32);
  15
  16 static int check_subvol(struct btree_trans *trans,
  17                         struct btree_iter *iter,
  18                         struct bkey_s_c k)
  19 {
  20         struct bch_fs *c = trans->c;
  21         struct bkey_s_c_subvolume subvol;
  22         struct bch_snapshot snapshot;
  23         unsigned snapid;
  24         int ret = 0;
  25
  26         if (k.k->type != KEY_TYPE_subvolume)
  27                 return 0;
  28
  29         subvol = bkey_s_c_to_subvolume(k);
  30         snapid = le32_to_cpu(subvol.v->snapshot);
  31         ret = bch2_snapshot_lookup(trans, snapid, &snapshot);
  32
  33         if (bch2_err_matches(ret, ENOENT))
  34                 bch_err(c, "subvolume %llu points to nonexistent snapshot %u",
  35                         k.k->p.offset, snapid);
  36         if (ret)
  37                 return ret;
  38
  39         if (BCH_SUBVOLUME_UNLINKED(subvol.v)) {
  40                 ret = bch2_subvolume_delete(trans, iter->pos.offset);
  41                 bch_err_msg(c, ret, "deleting subvolume %llu", iter->pos.offset);
  42                 return ret ?: -BCH_ERR_transaction_restart_nested;
  43         }
  44
  45         struct bch_inode_unpacked inode;
  46         struct btree_iter inode_iter = {};
  47         ret = bch2_inode_peek_nowarn(trans, &inode_iter, &inode,
  48                                     (subvol_inum) { k.k->p.offset, le64_to_cpu(subvol.v->inode) },
  49                                     0);
  50         bch2_trans_iter_exit(trans, &inode_iter);
  51
  52         if (ret && !bch2_err_matches(ret, ENOENT))
  53                 return ret;
  54
  55         if (fsck_err_on(ret, c, subvol_to_missing_root,
  56                         "subvolume %llu points to missing subvolume root %llu:%u",
  57                         k.k->p.offset, le64_to_cpu(subvol.v->inode),
  58                         le32_to_cpu(subvol.v->snapshot))) {
  59                 ret = bch2_subvolume_delete(trans, iter->pos.offset);
  60                 bch_err_msg(c, ret, "deleting subvolume %llu", iter->pos.offset);
  61                 return ret ?: -BCH_ERR_transaction_restart_nested;
  62         }
  63
  64         if (fsck_err_on(inode.bi_subvol != subvol.k->p.offset,
  65                         c, subvol_root_wrong_bi_subvol,
  66                         "subvol root %llu:%u has wrong bi_subvol field: got %u, should be %llu",
  67                         inode.bi_inum, inode_iter.k.p.snapshot,
  68                         inode.bi_subvol, subvol.k->p.offset)) {
  69                 inode.bi_subvol = subvol.k->p.offset;
  70                 ret = __bch2_fsck_write_inode(trans, &inode, le32_to_cpu(subvol.v->snapshot));
  71                 if (ret)
  72                         goto err;
  73         }
  74
  75         if (!BCH_SUBVOLUME_SNAP(subvol.v)) {
  76                 u32 snapshot_root = bch2_snapshot_root(c, le32_to_cpu(subvol.v->snapshot));
  77                 u32 snapshot_tree;
  78                 struct bch_snapshot_tree st;
  79
  80                 rcu_read_lock();
  81                 snapshot_tree = snapshot_t(c, snapshot_root)->tree;
  82                 rcu_read_unlock();
  83
  84                 ret = bch2_snapshot_tree_lookup(trans, snapshot_tree, &st);
  85
  86                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
  87                                 "%s: snapshot tree %u not found", __func__, snapshot_tree);
  88
  89                 if (ret)
  90                         return ret;
  91
  92                 if (fsck_err_on(le32_to_cpu(st.master_subvol) != subvol.k->p.offset,
  93                                 c, subvol_not_master_and_not_snapshot,
  94                                 "subvolume %llu is not set as snapshot but is not master subvolume",
  95                                 k.k->p.offset)) {
  96                         struct bkey_i_subvolume *s =
  97                                 bch2_bkey_make_mut_typed(trans, iter, &subvol.s_c, 0, subvolume);
  98                         ret = PTR_ERR_OR_ZERO(s);
  99                         if (ret)
 100                                 return ret;
 101
 102                         SET_BCH_SUBVOLUME_SNAP(&s->v, true);
 103                 }
 104         }
 105
 106 err:
 107 fsck_err:
 108         return ret;
 109 }
 110
 111 int bch2_check_subvols(struct bch_fs *c)
 112 {
 113         int ret = bch2_trans_run(c,
 114                 for_each_btree_key_commit(trans, iter,
 115                                 BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
 116                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 117                         check_subvol(trans, &iter, k)));
 118         bch_err_fn(c, ret);
 119         return ret;
 120 }
 121
 122 /* Subvolumes: */
 123
 124 int bch2_subvolume_invalid(struct bch_fs *c, struct bkey_s_c k,
 125                            enum bkey_invalid_flags flags, struct printbuf *err)
 126 {
 127         int ret = 0;
 128
 129         bkey_fsck_err_on(bkey_lt(k.k->p, SUBVOL_POS_MIN) ||
 130                          bkey_gt(k.k->p, SUBVOL_POS_MAX), c, err,
 131                          subvol_pos_bad,
 132                          "invalid pos");
 133 fsck_err:
 134         return ret;
 135 }
 136
 137 void bch2_subvolume_to_text(struct printbuf *out, struct bch_fs *c,
 138                             struct bkey_s_c k)
 139 {
 140         struct bkey_s_c_subvolume s = bkey_s_c_to_subvolume(k);
 141
 142         prt_printf(out, "root %llu snapshot id %u",
 143                    le64_to_cpu(s.v->inode),
 144                    le32_to_cpu(s.v->snapshot));
 145
 146         if (bkey_val_bytes(s.k) > offsetof(struct bch_subvolume, parent))
 147                 prt_printf(out, " parent %u", le32_to_cpu(s.v->parent));
 148 }
 149
 150 static __always_inline int
 151 bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
 152                            bool inconsistent_if_not_found,
 153                            int iter_flags,
 154                            struct bch_subvolume *s)
 155 {
 156         int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_subvolumes, POS(0, subvol),
 157                                           iter_flags, subvolume, s);
 158         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT) &&
 159                                 inconsistent_if_not_found,
 160                                 trans->c, "missing subvolume %u", subvol);
 161         return ret;
 162 }
 163
 164 int bch2_subvolume_get(struct btree_trans *trans, unsigned subvol,
 165                        bool inconsistent_if_not_found,
 166                        int iter_flags,
 167                        struct bch_subvolume *s)
 168 {
 169         return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, iter_flags, s);
 170 }
 171
 172 int bch2_subvol_is_ro_trans(struct btree_trans *trans, u32 subvol)
 173 {
 174         struct bch_subvolume s;
 175         int ret = bch2_subvolume_get_inlined(trans, subvol, true, 0, &s);
 176         if (ret)
 177                 return ret;
 178
 179         if (BCH_SUBVOLUME_RO(&s))
 180                 return -EROFS;
 181         return 0;
 182 }
 183
 184 int bch2_subvol_is_ro(struct bch_fs *c, u32 subvol)
 185 {
 186         return bch2_trans_do(c, NULL, NULL, 0,
 187                 bch2_subvol_is_ro_trans(trans, subvol));
 188 }
 189
 190 int bch2_snapshot_get_subvol(struct btree_trans *trans, u32 snapshot,
 191                              struct bch_subvolume *subvol)
 192 {
 193         struct bch_snapshot snap;
 194
 195         return  bch2_snapshot_lookup(trans, snapshot, &snap) ?:
 196                 bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, 0, subvol);
 197 }
 198
 199 int bch2_subvolume_get_snapshot(struct btree_trans *trans, u32 subvolid,
 200                                 u32 *snapid)
 201 {
 202         struct btree_iter iter;
 203         struct bkey_s_c_subvolume subvol;
 204         int ret;
 205
 206         subvol = bch2_bkey_get_iter_typed(trans, &iter,
 207                                           BTREE_ID_subvolumes, POS(0, subvolid),
 208                                           BTREE_ITER_CACHED|BTREE_ITER_WITH_UPDATES,
 209                                           subvolume);
 210         ret = bkey_err(subvol);
 211         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 212                                 "missing subvolume %u", subvolid);
 213
 214         if (likely(!ret))
 215                 *snapid = le32_to_cpu(subvol.v->snapshot);
 216         bch2_trans_iter_exit(trans, &iter);
 217         return ret;
 218 }
 219
 220 static int bch2_subvolume_reparent(struct btree_trans *trans,
 221                                    struct btree_iter *iter,
 222                                    struct bkey_s_c k,
 223                                    u32 old_parent, u32 new_parent)
 224 {
 225         struct bkey_i_subvolume *s;
 226         int ret;
 227
 228         if (k.k->type != KEY_TYPE_subvolume)
 229                 return 0;
 230
 231         if (bkey_val_bytes(k.k) > offsetof(struct bch_subvolume, parent) &&
 232             le32_to_cpu(bkey_s_c_to_subvolume(k).v->parent) != old_parent)
 233                 return 0;
 234
 235         s = bch2_bkey_make_mut_typed(trans, iter, &k, 0, subvolume);
 236         ret = PTR_ERR_OR_ZERO(s);
 237         if (ret)
 238                 return ret;
 239
 240         s->v.parent = cpu_to_le32(new_parent);
 241         return 0;
 242 }
 243
 244 /*
 245  * Separate from the snapshot tree in the snapshots btree, we record the tree
 246  * structure of how snapshot subvolumes were created - the parent subvolume of
 247  * each snapshot subvolume.
 248  *
 249  * When a subvolume is deleted, we scan for child subvolumes and reparant them,
 250  * to avoid dangling references:
 251  */
 252 static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_delete)
 253 {
 254         struct bch_subvolume s;
 255
 256         return lockrestart_do(trans,
 257                         bch2_subvolume_get(trans, subvolid_to_delete, true,
 258                                    BTREE_ITER_CACHED, &s)) ?:
 259                 for_each_btree_key_commit(trans, iter,
 260                                 BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
 261                                 NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 262                         bch2_subvolume_reparent(trans, &iter, k,
 263                                         subvolid_to_delete, le32_to_cpu(s.parent)));
 264 }
 265
 266 /*
 267  * Delete subvolume, mark snapshot ID as deleted, queue up snapshot
 268  * deletion/cleanup:
 269  */
 270 static int __bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
 271 {
 272         struct btree_iter iter;
 273         struct bkey_s_c_subvolume subvol;
 274         u32 snapid;
 275         int ret = 0;
 276
 277         subvol = bch2_bkey_get_iter_typed(trans, &iter,
 278                                 BTREE_ID_subvolumes, POS(0, subvolid),
 279                                 BTREE_ITER_CACHED|BTREE_ITER_INTENT,
 280                                 subvolume);
 281         ret = bkey_err(subvol);
 282         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 283                                 "missing subvolume %u", subvolid);
 284         if (ret)
 285                 return ret;
 286
 287         snapid = le32_to_cpu(subvol.v->snapshot);
 288
 289         ret =   bch2_btree_delete_at(trans, &iter, 0) ?:
 290                 bch2_snapshot_node_set_deleted(trans, snapid);
 291         bch2_trans_iter_exit(trans, &iter);
 292         return ret;
 293 }
 294
 295 static int bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
 296 {
 297         return bch2_subvolumes_reparent(trans, subvolid) ?:
 298                 commit_do(trans, NULL, NULL, BCH_TRANS_COMMIT_no_enospc,
 299                           __bch2_subvolume_delete(trans, subvolid));
 300 }
 301
 302 static void bch2_subvolume_wait_for_pagecache_and_delete(struct work_struct *work)
 303 {
 304         struct bch_fs *c = container_of(work, struct bch_fs,
 305                                 snapshot_wait_for_pagecache_and_delete_work);
 306         snapshot_id_list s;
 307         u32 *id;
 308         int ret = 0;
 309
 310         while (!ret) {
 311                 mutex_lock(&c->snapshots_unlinked_lock);
 312                 s = c->snapshots_unlinked;
 313                 darray_init(&c->snapshots_unlinked);
 314                 mutex_unlock(&c->snapshots_unlinked_lock);
 315
 316                 if (!s.nr)
 317                         break;
 318
 319                 bch2_evict_subvolume_inodes(c, &s);
 320
 321                 for (id = s.data; id < s.data + s.nr; id++) {
 322                         ret = bch2_trans_run(c, bch2_subvolume_delete(trans, *id));
 323                         bch_err_msg(c, ret, "deleting subvolume %u", *id);
 324                         if (ret)
 325                                 break;
 326                 }
 327
 328                 darray_exit(&s);
 329         }
 330
 331         bch2_write_ref_put(c, BCH_WRITE_REF_snapshot_delete_pagecache);
 332 }
 333
 334 struct subvolume_unlink_hook {
 335         struct btree_trans_commit_hook  h;
 336         u32                             subvol;
 337 };
 338
 339 static int bch2_subvolume_wait_for_pagecache_and_delete_hook(struct btree_trans *trans,
 340                                                       struct btree_trans_commit_hook *_h)
 341 {
 342         struct subvolume_unlink_hook *h = container_of(_h, struct subvolume_unlink_hook, h);
 343         struct bch_fs *c = trans->c;
 344         int ret = 0;
 345
 346         mutex_lock(&c->snapshots_unlinked_lock);
 347         if (!snapshot_list_has_id(&c->snapshots_unlinked, h->subvol))
 348                 ret = snapshot_list_add(c, &c->snapshots_unlinked, h->subvol);
 349         mutex_unlock(&c->snapshots_unlinked_lock);
 350
 351         if (ret)
 352                 return ret;
 353
 354         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_snapshot_delete_pagecache))
 355                 return -EROFS;
 356
 357         if (!queue_work(c->write_ref_wq, &c->snapshot_wait_for_pagecache_and_delete_work))
 358                 bch2_write_ref_put(c, BCH_WRITE_REF_snapshot_delete_pagecache);
 359         return 0;
 360 }
 361
 362 int bch2_subvolume_unlink(struct btree_trans *trans, u32 subvolid)
 363 {
 364         struct btree_iter iter;
 365         struct bkey_i_subvolume *n;
 366         struct subvolume_unlink_hook *h;
 367         int ret = 0;
 368
 369         h = bch2_trans_kmalloc(trans, sizeof(*h));
 370         ret = PTR_ERR_OR_ZERO(h);
 371         if (ret)
 372                 return ret;
 373
 374         h->h.fn         = bch2_subvolume_wait_for_pagecache_and_delete_hook;
 375         h->subvol       = subvolid;
 376         bch2_trans_commit_hook(trans, &h->h);
 377
 378         n = bch2_bkey_get_mut_typed(trans, &iter,
 379                         BTREE_ID_subvolumes, POS(0, subvolid),
 380                         BTREE_ITER_CACHED, subvolume);
 381         ret = PTR_ERR_OR_ZERO(n);
 382         if (unlikely(ret)) {
 383                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 384                                         "missing subvolume %u", subvolid);
 385                 return ret;
 386         }
 387
 388         SET_BCH_SUBVOLUME_UNLINKED(&n->v, true);
 389         bch2_trans_iter_exit(trans, &iter);
 390         return ret;
 391 }
 392
 393 int bch2_subvolume_create(struct btree_trans *trans, u64 inode,
 394                           u32 src_subvolid,
 395                           u32 *new_subvolid,
 396                           u32 *new_snapshotid,
 397                           bool ro)
 398 {
 399         struct bch_fs *c = trans->c;
 400         struct btree_iter dst_iter, src_iter = (struct btree_iter) { NULL };
 401         struct bkey_i_subvolume *new_subvol = NULL;
 402         struct bkey_i_subvolume *src_subvol = NULL;
 403         u32 parent = 0, new_nodes[2], snapshot_subvols[2];
 404         int ret = 0;
 405
 406         ret = bch2_bkey_get_empty_slot(trans, &dst_iter,
 407                                 BTREE_ID_subvolumes, POS(0, U32_MAX));
 408         if (ret == -BCH_ERR_ENOSPC_btree_slot)
 409                 ret = -BCH_ERR_ENOSPC_subvolume_create;
 410         if (ret)
 411                 return ret;
 412
 413         snapshot_subvols[0] = dst_iter.pos.offset;
 414         snapshot_subvols[1] = src_subvolid;
 415
 416         if (src_subvolid) {
 417                 /* Creating a snapshot: */
 418
 419                 src_subvol = bch2_bkey_get_mut_typed(trans, &src_iter,
 420                                 BTREE_ID_subvolumes, POS(0, src_subvolid),
 421                                 BTREE_ITER_CACHED, subvolume);
 422                 ret = PTR_ERR_OR_ZERO(src_subvol);
 423                 if (unlikely(ret)) {
 424                         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
 425                                                 "subvolume %u not found", src_subvolid);
 426                         goto err;
 427                 }
 428
 429                 parent = le32_to_cpu(src_subvol->v.snapshot);
 430         }
 431
 432         ret = bch2_snapshot_node_create(trans, parent, new_nodes,
 433                                         snapshot_subvols,
 434                                         src_subvolid ? 2 : 1);
 435         if (ret)
 436                 goto err;
 437
 438         if (src_subvolid) {
 439                 src_subvol->v.snapshot = cpu_to_le32(new_nodes[1]);
 440                 ret = bch2_trans_update(trans, &src_iter, &src_subvol->k_i, 0);
 441                 if (ret)
 442                         goto err;
 443         }
 444
 445         new_subvol = bch2_bkey_alloc(trans, &dst_iter, 0, subvolume);
 446         ret = PTR_ERR_OR_ZERO(new_subvol);
 447         if (ret)
 448                 goto err;
 449
 450         new_subvol->v.flags     = 0;
 451         new_subvol->v.snapshot  = cpu_to_le32(new_nodes[0]);
 452         new_subvol->v.inode     = cpu_to_le64(inode);
 453         new_subvol->v.parent    = cpu_to_le32(src_subvolid);
 454         new_subvol->v.otime.lo  = cpu_to_le64(bch2_current_time(c));
 455         new_subvol->v.otime.hi  = 0;
 456
 457         SET_BCH_SUBVOLUME_RO(&new_subvol->v, ro);
 458         SET_BCH_SUBVOLUME_SNAP(&new_subvol->v, src_subvolid != 0);
 459
 460         *new_subvolid   = new_subvol->k.p.offset;
 461         *new_snapshotid = new_nodes[0];
 462 err:
 463         bch2_trans_iter_exit(trans, &src_iter);
 464         bch2_trans_iter_exit(trans, &dst_iter);
 465         return ret;
 466 }
 467
 468 int bch2_fs_subvolumes_init(struct bch_fs *c)
 469 {
 470         INIT_WORK(&c->snapshot_delete_work, bch2_delete_dead_snapshots_work);
 471         INIT_WORK(&c->snapshot_wait_for_pagecache_and_delete_work,
 472                   bch2_subvolume_wait_for_pagecache_and_delete);
 473         mutex_init(&c->snapshots_unlinked_lock);
 474         return 0;
 475 }