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