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                 bch2_fs_lazy_rw(c);
  41
  42                 ret = bch2_subvolume_delete(trans, iter->pos.offset);
  43                 if (ret)
  44                         bch_err(c, "error deleting subvolume %llu: %s",
  45                                 iter->pos.offset, bch2_err_str(ret));
  46                 return ret ?: -BCH_ERR_transaction_restart_nested;
  47         }
  48
  49         if (!BCH_SUBVOLUME_SNAP(subvol.v)) {
  50                 u32 snapshot_root = bch2_snapshot_root(c, le32_to_cpu(subvol.v->snapshot));
  51                 u32 snapshot_tree;
  52                 struct bch_snapshot_tree st;
  53
  54                 rcu_read_lock();
  55                 snapshot_tree = snapshot_t(c, snapshot_root)->tree;
  56                 rcu_read_unlock();
  57
  58                 ret = bch2_snapshot_tree_lookup(trans, snapshot_tree, &st);
  59
  60                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
  61                                 "%s: snapshot tree %u not found", __func__, snapshot_tree);
  62
  63                 if (ret)
  64                         return ret;
  65
  66                 if (fsck_err_on(le32_to_cpu(st.master_subvol) != subvol.k->p.offset, c,
  67                                 "subvolume %llu is not set as snapshot but is not master subvolume",
  68                                 k.k->p.offset)) {
  69                         struct bkey_i_subvolume *s =
  70                                 bch2_bkey_make_mut_typed(trans, iter, &subvol.s_c, 0, subvolume);
  71                         ret = PTR_ERR_OR_ZERO(s);
  72                         if (ret)
  73                                 return ret;
  74
  75                         SET_BCH_SUBVOLUME_SNAP(&s->v, true);
  76                 }
  77         }
  78
  79 fsck_err:
  80         return ret;
  81 }
  82
  83 int bch2_check_subvols(struct bch_fs *c)
  84 {
  85         struct btree_iter iter;
  86         struct bkey_s_c k;
  87         int ret;
  88
  89         ret = bch2_trans_run(c,
  90                 for_each_btree_key_commit(&trans, iter,
  91                         BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
  92                         NULL, NULL, BTREE_INSERT_LAZY_RW|BTREE_INSERT_NOFAIL,
  93                 check_subvol(&trans, &iter, k)));
  94         if (ret)
  95                 bch_err_fn(c, ret);
  96         return ret;
  97 }
  98
  99 /* Subvolumes: */
 100
 101 int bch2_subvolume_invalid(const struct bch_fs *c, struct bkey_s_c k,
 102                            unsigned flags, struct printbuf *err)
 103 {
 104         if (bkey_lt(k.k->p, SUBVOL_POS_MIN) ||
 105             bkey_gt(k.k->p, SUBVOL_POS_MAX)) {
 106                 prt_printf(err, "invalid pos");
 107                 return -BCH_ERR_invalid_bkey;
 108         }
 109
 110         return 0;
 111 }
 112
 113 void bch2_subvolume_to_text(struct printbuf *out, struct bch_fs *c,
 114                             struct bkey_s_c k)
 115 {
 116         struct bkey_s_c_subvolume s = bkey_s_c_to_subvolume(k);
 117
 118         prt_printf(out, "root %llu snapshot id %u",
 119                    le64_to_cpu(s.v->inode),
 120                    le32_to_cpu(s.v->snapshot));
 121
 122         if (bkey_val_bytes(s.k) > offsetof(struct bch_subvolume, parent))
 123                 prt_printf(out, " parent %u", le32_to_cpu(s.v->parent));
 124 }
 125
 126 static __always_inline int
 127 bch2_subvolume_get_inlined(struct btree_trans *trans, unsigned subvol,
 128                            bool inconsistent_if_not_found,
 129                            int iter_flags,
 130                            struct bch_subvolume *s)
 131 {
 132         int ret = bch2_bkey_get_val_typed(trans, BTREE_ID_subvolumes, POS(0, subvol),
 133                                           iter_flags, subvolume, s);
 134         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT) &&
 135                                 inconsistent_if_not_found,
 136                                 trans->c, "missing subvolume %u", subvol);
 137         return ret;
 138 }
 139
 140 int bch2_subvolume_get(struct btree_trans *trans, unsigned subvol,
 141                        bool inconsistent_if_not_found,
 142                        int iter_flags,
 143                        struct bch_subvolume *s)
 144 {
 145         return bch2_subvolume_get_inlined(trans, subvol, inconsistent_if_not_found, iter_flags, s);
 146 }
 147
 148 int bch2_snapshot_get_subvol(struct btree_trans *trans, u32 snapshot,
 149                              struct bch_subvolume *subvol)
 150 {
 151         struct bch_snapshot snap;
 152
 153         return  bch2_snapshot_lookup(trans, snapshot, &snap) ?:
 154                 bch2_subvolume_get(trans, le32_to_cpu(snap.subvol), true, 0, subvol);
 155 }
 156
 157 int bch2_subvolume_get_snapshot(struct btree_trans *trans, u32 subvolid,
 158                                 u32 *snapid)
 159 {
 160         struct btree_iter iter;
 161         struct bkey_s_c_subvolume subvol;
 162         int ret;
 163
 164         subvol = bch2_bkey_get_iter_typed(trans, &iter,
 165                                           BTREE_ID_subvolumes, POS(0, subvolid),
 166                                           BTREE_ITER_CACHED|BTREE_ITER_WITH_UPDATES,
 167                                           subvolume);
 168         ret = bkey_err(subvol);
 169         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 170                                 "missing subvolume %u", subvolid);
 171
 172         if (likely(!ret))
 173                 *snapid = le32_to_cpu(subvol.v->snapshot);
 174         bch2_trans_iter_exit(trans, &iter);
 175         return ret;
 176 }
 177
 178 static int bch2_subvolume_reparent(struct btree_trans *trans,
 179                                    struct btree_iter *iter,
 180                                    struct bkey_s_c k,
 181                                    u32 old_parent, u32 new_parent)
 182 {
 183         struct bkey_i_subvolume *s;
 184         int ret;
 185
 186         if (k.k->type != KEY_TYPE_subvolume)
 187                 return 0;
 188
 189         if (bkey_val_bytes(k.k) > offsetof(struct bch_subvolume, parent) &&
 190             le32_to_cpu(bkey_s_c_to_subvolume(k).v->parent) != old_parent)
 191                 return 0;
 192
 193         s = bch2_bkey_make_mut_typed(trans, iter, &k, 0, subvolume);
 194         ret = PTR_ERR_OR_ZERO(s);
 195         if (ret)
 196                 return ret;
 197
 198         s->v.parent = cpu_to_le32(new_parent);
 199         return 0;
 200 }
 201
 202 /*
 203  * Separate from the snapshot tree in the snapshots btree, we record the tree
 204  * structure of how snapshot subvolumes were created - the parent subvolume of
 205  * each snapshot subvolume.
 206  *
 207  * When a subvolume is deleted, we scan for child subvolumes and reparant them,
 208  * to avoid dangling references:
 209  */
 210 static int bch2_subvolumes_reparent(struct btree_trans *trans, u32 subvolid_to_delete)
 211 {
 212         struct btree_iter iter;
 213         struct bkey_s_c k;
 214         struct bch_subvolume s;
 215
 216         return lockrestart_do(trans,
 217                         bch2_subvolume_get(trans, subvolid_to_delete, true,
 218                                    BTREE_ITER_CACHED, &s)) ?:
 219                 for_each_btree_key_commit(trans, iter,
 220                                 BTREE_ID_subvolumes, POS_MIN, BTREE_ITER_PREFETCH, k,
 221                                 NULL, NULL, BTREE_INSERT_NOFAIL,
 222                         bch2_subvolume_reparent(trans, &iter, k,
 223                                         subvolid_to_delete, le32_to_cpu(s.parent)));
 224 }
 225
 226 /*
 227  * Delete subvolume, mark snapshot ID as deleted, queue up snapshot
 228  * deletion/cleanup:
 229  */
 230 static int __bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
 231 {
 232         struct btree_iter iter;
 233         struct bkey_s_c_subvolume subvol;
 234         struct btree_trans_commit_hook *h;
 235         u32 snapid;
 236         int ret = 0;
 237
 238         subvol = bch2_bkey_get_iter_typed(trans, &iter,
 239                                 BTREE_ID_subvolumes, POS(0, subvolid),
 240                                 BTREE_ITER_CACHED|BTREE_ITER_INTENT,
 241                                 subvolume);
 242         ret = bkey_err(subvol);
 243         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 244                                 "missing subvolume %u", subvolid);
 245         if (ret)
 246                 return ret;
 247
 248         snapid = le32_to_cpu(subvol.v->snapshot);
 249
 250         ret = bch2_btree_delete_at(trans, &iter, 0);
 251         if (ret)
 252                 goto err;
 253
 254         ret = bch2_snapshot_node_set_deleted(trans, snapid);
 255         if (ret)
 256                 goto err;
 257
 258         h = bch2_trans_kmalloc(trans, sizeof(*h));
 259         ret = PTR_ERR_OR_ZERO(h);
 260         if (ret)
 261                 goto err;
 262
 263         h->fn = bch2_delete_dead_snapshots_hook;
 264         bch2_trans_commit_hook(trans, h);
 265 err:
 266         bch2_trans_iter_exit(trans, &iter);
 267         return ret;
 268 }
 269
 270 static int bch2_subvolume_delete(struct btree_trans *trans, u32 subvolid)
 271 {
 272         return bch2_subvolumes_reparent(trans, subvolid) ?:
 273                 commit_do(trans, NULL, NULL, BTREE_INSERT_NOFAIL,
 274                           __bch2_subvolume_delete(trans, subvolid));
 275 }
 276
 277 static void bch2_subvolume_wait_for_pagecache_and_delete(struct work_struct *work)
 278 {
 279         struct bch_fs *c = container_of(work, struct bch_fs,
 280                                 snapshot_wait_for_pagecache_and_delete_work);
 281         snapshot_id_list s;
 282         u32 *id;
 283         int ret = 0;
 284
 285         while (!ret) {
 286                 mutex_lock(&c->snapshots_unlinked_lock);
 287                 s = c->snapshots_unlinked;
 288                 darray_init(&c->snapshots_unlinked);
 289                 mutex_unlock(&c->snapshots_unlinked_lock);
 290
 291                 if (!s.nr)
 292                         break;
 293
 294                 bch2_evict_subvolume_inodes(c, &s);
 295
 296                 for (id = s.data; id < s.data + s.nr; id++) {
 297                         ret = bch2_trans_run(c, bch2_subvolume_delete(&trans, *id));
 298                         if (ret) {
 299                                 bch_err(c, "error deleting subvolume %u: %s", *id, bch2_err_str(ret));
 300                                 break;
 301                         }
 302                 }
 303
 304                 darray_exit(&s);
 305         }
 306
 307         bch2_write_ref_put(c, BCH_WRITE_REF_snapshot_delete_pagecache);
 308 }
 309
 310 struct subvolume_unlink_hook {
 311         struct btree_trans_commit_hook  h;
 312         u32                             subvol;
 313 };
 314
 315 static int bch2_subvolume_wait_for_pagecache_and_delete_hook(struct btree_trans *trans,
 316                                                       struct btree_trans_commit_hook *_h)
 317 {
 318         struct subvolume_unlink_hook *h = container_of(_h, struct subvolume_unlink_hook, h);
 319         struct bch_fs *c = trans->c;
 320         int ret = 0;
 321
 322         mutex_lock(&c->snapshots_unlinked_lock);
 323         if (!snapshot_list_has_id(&c->snapshots_unlinked, h->subvol))
 324                 ret = snapshot_list_add(c, &c->snapshots_unlinked, h->subvol);
 325         mutex_unlock(&c->snapshots_unlinked_lock);
 326
 327         if (ret)
 328                 return ret;
 329
 330         if (!bch2_write_ref_tryget(c, BCH_WRITE_REF_snapshot_delete_pagecache))
 331                 return -EROFS;
 332
 333         if (!queue_work(c->write_ref_wq, &c->snapshot_wait_for_pagecache_and_delete_work))
 334                 bch2_write_ref_put(c, BCH_WRITE_REF_snapshot_delete_pagecache);
 335         return 0;
 336 }
 337
 338 int bch2_subvolume_unlink(struct btree_trans *trans, u32 subvolid)
 339 {
 340         struct btree_iter iter;
 341         struct bkey_i_subvolume *n;
 342         struct subvolume_unlink_hook *h;
 343         int ret = 0;
 344
 345         h = bch2_trans_kmalloc(trans, sizeof(*h));
 346         ret = PTR_ERR_OR_ZERO(h);
 347         if (ret)
 348                 return ret;
 349
 350         h->h.fn         = bch2_subvolume_wait_for_pagecache_and_delete_hook;
 351         h->subvol       = subvolid;
 352         bch2_trans_commit_hook(trans, &h->h);
 353
 354         n = bch2_bkey_get_mut_typed(trans, &iter,
 355                         BTREE_ID_subvolumes, POS(0, subvolid),
 356                         BTREE_ITER_CACHED, subvolume);
 357         ret = PTR_ERR_OR_ZERO(n);
 358         if (unlikely(ret)) {
 359                 bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), trans->c,
 360                                         "missing subvolume %u", subvolid);
 361                 return ret;
 362         }
 363
 364         SET_BCH_SUBVOLUME_UNLINKED(&n->v, true);
 365         bch2_trans_iter_exit(trans, &iter);
 366         return ret;
 367 }
 368
 369 int bch2_subvolume_create(struct btree_trans *trans, u64 inode,
 370                           u32 src_subvolid,
 371                           u32 *new_subvolid,
 372                           u32 *new_snapshotid,
 373                           bool ro)
 374 {
 375         struct bch_fs *c = trans->c;
 376         struct btree_iter dst_iter, src_iter = (struct btree_iter) { NULL };
 377         struct bkey_i_subvolume *new_subvol = NULL;
 378         struct bkey_i_subvolume *src_subvol = NULL;
 379         u32 parent = 0, new_nodes[2], snapshot_subvols[2];
 380         int ret = 0;
 381
 382         ret = bch2_bkey_get_empty_slot(trans, &dst_iter,
 383                                 BTREE_ID_subvolumes, POS(0, U32_MAX));
 384         if (ret == -BCH_ERR_ENOSPC_btree_slot)
 385                 ret = -BCH_ERR_ENOSPC_subvolume_create;
 386         if (ret)
 387                 return ret;
 388
 389         snapshot_subvols[0] = dst_iter.pos.offset;
 390         snapshot_subvols[1] = src_subvolid;
 391
 392         if (src_subvolid) {
 393                 /* Creating a snapshot: */
 394
 395                 src_subvol = bch2_bkey_get_mut_typed(trans, &src_iter,
 396                                 BTREE_ID_subvolumes, POS(0, src_subvolid),
 397                                 BTREE_ITER_CACHED, subvolume);
 398                 ret = PTR_ERR_OR_ZERO(src_subvol);
 399                 if (unlikely(ret)) {
 400                         bch2_fs_inconsistent_on(bch2_err_matches(ret, ENOENT), c,
 401                                                 "subvolume %u not found", src_subvolid);
 402                         goto err;
 403                 }
 404
 405                 parent = le32_to_cpu(src_subvol->v.snapshot);
 406         }
 407
 408         ret = bch2_snapshot_node_create(trans, parent, new_nodes,
 409                                         snapshot_subvols,
 410                                         src_subvolid ? 2 : 1);
 411         if (ret)
 412                 goto err;
 413
 414         if (src_subvolid) {
 415                 src_subvol->v.snapshot = cpu_to_le32(new_nodes[1]);
 416                 ret = bch2_trans_update(trans, &src_iter, &src_subvol->k_i, 0);
 417                 if (ret)
 418                         goto err;
 419         }
 420
 421         new_subvol = bch2_bkey_alloc(trans, &dst_iter, 0, subvolume);
 422         ret = PTR_ERR_OR_ZERO(new_subvol);
 423         if (ret)
 424                 goto err;
 425
 426         new_subvol->v.flags     = 0;
 427         new_subvol->v.snapshot  = cpu_to_le32(new_nodes[0]);
 428         new_subvol->v.inode     = cpu_to_le64(inode);
 429         new_subvol->v.parent    = cpu_to_le32(src_subvolid);
 430         new_subvol->v.otime.lo  = cpu_to_le64(bch2_current_time(c));
 431         new_subvol->v.otime.hi  = 0;
 432
 433         SET_BCH_SUBVOLUME_RO(&new_subvol->v, ro);
 434         SET_BCH_SUBVOLUME_SNAP(&new_subvol->v, src_subvolid != 0);
 435
 436         *new_subvolid   = new_subvol->k.p.offset;
 437         *new_snapshotid = new_nodes[0];
 438 err:
 439         bch2_trans_iter_exit(trans, &src_iter);
 440         bch2_trans_iter_exit(trans, &dst_iter);
 441         return ret;
 442 }
 443
 444 int bch2_fs_subvolumes_init(struct bch_fs *c)
 445 {
 446         INIT_WORK(&c->snapshot_delete_work, bch2_delete_dead_snapshots_work);
 447         INIT_WORK(&c->snapshot_wait_for_pagecache_and_delete_work,
 448                   bch2_subvolume_wait_for_pagecache_and_delete);
 449         mutex_init(&c->snapshots_unlinked_lock);
 450         return 0;
 451 }