fs/pnode.c

   1 /*
   2  *  linux/fs/pnode.c
   3  *
   4  * (C) Copyright IBM Corporation 2005.
   5  *      Released under GPL v2.
   6  *      Author : Ram Pai (linuxram@us.ibm.com)
   7  *
   8  */
   9 #include <linux/mnt_namespace.h>
  10 #include <linux/mount.h>
  11 #include <linux/fs.h>
  12 #include <linux/nsproxy.h>
  13 #include "internal.h"
  14 #include "pnode.h"
  15
  16 /* return the next shared peer mount of @p */
  17 static inline struct mount *next_peer(struct mount *p)
  18 {
  19         return list_entry(p->mnt_share.next, struct mount, mnt_share);
  20 }
  21
  22 static inline struct mount *first_slave(struct mount *p)
  23 {
  24         return list_entry(p->mnt_slave_list.next, struct mount, mnt_slave);
  25 }
  26
  27 static inline struct mount *last_slave(struct mount *p)
  28 {
  29         return list_entry(p->mnt_slave_list.prev, struct mount, mnt_slave);
  30 }
  31
  32 static inline struct mount *next_slave(struct mount *p)
  33 {
  34         return list_entry(p->mnt_slave.next, struct mount, mnt_slave);
  35 }
  36
  37 static struct mount *get_peer_under_root(struct mount *mnt,
  38                                          struct mnt_namespace *ns,
  39                                          const struct path *root)
  40 {
  41         struct mount *m = mnt;
  42
  43         do {
  44                 /* Check the namespace first for optimization */
  45                 if (m->mnt_ns == ns && is_path_reachable(m, m->mnt.mnt_root, root))
  46                         return m;
  47
  48                 m = next_peer(m);
  49         } while (m != mnt);
  50
  51         return NULL;
  52 }
  53
  54 /*
  55  * Get ID of closest dominating peer group having a representative
  56  * under the given root.
  57  *
  58  * Caller must hold namespace_sem
  59  */
  60 int get_dominating_id(struct mount *mnt, const struct path *root)
  61 {
  62         struct mount *m;
  63
  64         for (m = mnt->mnt_master; m != NULL; m = m->mnt_master) {
  65                 struct mount *d = get_peer_under_root(m, mnt->mnt_ns, root);
  66                 if (d)
  67                         return d->mnt_group_id;
  68         }
  69
  70         return 0;
  71 }
  72
  73 static int do_make_slave(struct mount *mnt)
  74 {
  75         struct mount *peer_mnt = mnt, *master = mnt->mnt_master;
  76         struct mount *slave_mnt;
  77
  78         /*
  79          * slave 'mnt' to a peer mount that has the
  80          * same root dentry. If none is available then
  81          * slave it to anything that is available.
  82          */
  83         while ((peer_mnt = next_peer(peer_mnt)) != mnt &&
  84                peer_mnt->mnt.mnt_root != mnt->mnt.mnt_root) ;
  85
  86         if (peer_mnt == mnt) {
  87                 peer_mnt = next_peer(mnt);
  88                 if (peer_mnt == mnt)
  89                         peer_mnt = NULL;
  90         }
  91         if (mnt->mnt_group_id && IS_MNT_SHARED(mnt) &&
  92             list_empty(&mnt->mnt_share))
  93                 mnt_release_group_id(mnt);
  94
  95         list_del_init(&mnt->mnt_share);
  96         mnt->mnt_group_id = 0;
  97
  98         if (peer_mnt)
  99                 master = peer_mnt;
 100
 101         if (master) {
 102                 list_for_each_entry(slave_mnt, &mnt->mnt_slave_list, mnt_slave)
 103                         slave_mnt->mnt_master = master;
 104                 list_move(&mnt->mnt_slave, &master->mnt_slave_list);
 105                 list_splice(&mnt->mnt_slave_list, master->mnt_slave_list.prev);
 106                 INIT_LIST_HEAD(&mnt->mnt_slave_list);
 107         } else {
 108                 struct list_head *p = &mnt->mnt_slave_list;
 109                 while (!list_empty(p)) {
 110                         slave_mnt = list_first_entry(p,
 111                                         struct mount, mnt_slave);
 112                         list_del_init(&slave_mnt->mnt_slave);
 113                         slave_mnt->mnt_master = NULL;
 114                 }
 115         }
 116         mnt->mnt_master = master;
 117         CLEAR_MNT_SHARED(mnt);
 118         return 0;
 119 }
 120
 121 /*
 122  * vfsmount lock must be held for write
 123  */
 124 void change_mnt_propagation(struct mount *mnt, int type)
 125 {
 126         if (type == MS_SHARED) {
 127                 set_mnt_shared(mnt);
 128                 return;
 129         }
 130         do_make_slave(mnt);
 131         if (type != MS_SLAVE) {
 132                 list_del_init(&mnt->mnt_slave);
 133                 mnt->mnt_master = NULL;
 134                 if (type == MS_UNBINDABLE)
 135                         mnt->mnt.mnt_flags |= MNT_UNBINDABLE;
 136                 else
 137                         mnt->mnt.mnt_flags &= ~MNT_UNBINDABLE;
 138         }
 139 }
 140
 141 /*
 142  * get the next mount in the propagation tree.
 143  * @m: the mount seen last
 144  * @origin: the original mount from where the tree walk initiated
 145  *
 146  * Note that peer groups form contiguous segments of slave lists.
 147  * We rely on that in get_source() to be able to find out if
 148  * vfsmount found while iterating with propagation_next() is
 149  * a peer of one we'd found earlier.
 150  */
 151 static struct mount *propagation_next(struct mount *m,
 152                                          struct mount *origin)
 153 {
 154         /* are there any slaves of this mount? */
 155         if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
 156                 return first_slave(m);
 157
 158         while (1) {
 159                 struct mount *master = m->mnt_master;
 160
 161                 if (master == origin->mnt_master) {
 162                         struct mount *next = next_peer(m);
 163                         return (next == origin) ? NULL : next;
 164                 } else if (m->mnt_slave.next != &master->mnt_slave_list)
 165                         return next_slave(m);
 166
 167                 /* back at master */
 168                 m = master;
 169         }
 170 }
 171
 172 static struct mount *skip_propagation_subtree(struct mount *m,
 173                                                 struct mount *origin)
 174 {
 175         /*
 176          * Advance m such that propagation_next will not return
 177          * the slaves of m.
 178          */
 179         if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
 180                 m = last_slave(m);
 181
 182         return m;
 183 }
 184
 185 static struct mount *next_group(struct mount *m, struct mount *origin)
 186 {
 187         while (1) {
 188                 while (1) {
 189                         struct mount *next;
 190                         if (!IS_MNT_NEW(m) && !list_empty(&m->mnt_slave_list))
 191                                 return first_slave(m);
 192                         next = next_peer(m);
 193                         if (m->mnt_group_id == origin->mnt_group_id) {
 194                                 if (next == origin)
 195                                         return NULL;
 196                         } else if (m->mnt_slave.next != &next->mnt_slave)
 197                                 break;
 198                         m = next;
 199                 }
 200                 /* m is the last peer */
 201                 while (1) {
 202                         struct mount *master = m->mnt_master;
 203                         if (m->mnt_slave.next != &master->mnt_slave_list)
 204                                 return next_slave(m);
 205                         m = next_peer(master);
 206                         if (master->mnt_group_id == origin->mnt_group_id)
 207                                 break;
 208                         if (master->mnt_slave.next == &m->mnt_slave)
 209                                 break;
 210                         m = master;
 211                 }
 212                 if (m == origin)
 213                         return NULL;
 214         }
 215 }
 216
 217 /* all accesses are serialized by namespace_sem */
 218 static struct user_namespace *user_ns;
 219 static struct mount *last_dest, *first_source, *last_source, *dest_master;
 220 static struct mountpoint *mp;
 221 static struct hlist_head *list;
 222
 223 static inline bool peers(struct mount *m1, struct mount *m2)
 224 {
 225         return m1->mnt_group_id == m2->mnt_group_id && m1->mnt_group_id;
 226 }
 227
 228 static int propagate_one(struct mount *m)
 229 {
 230         struct mount *child;
 231         int type;
 232         /* skip ones added by this propagate_mnt() */
 233         if (IS_MNT_NEW(m))
 234                 return 0;
 235         /* skip if mountpoint isn't covered by it */
 236         if (!is_subdir(mp->m_dentry, m->mnt.mnt_root))
 237                 return 0;
 238         if (peers(m, last_dest)) {
 239                 type = CL_MAKE_SHARED;
 240         } else {
 241                 struct mount *n, *p;
 242                 bool done;
 243                 for (n = m; ; n = p) {
 244                         p = n->mnt_master;
 245                         if (p == dest_master || IS_MNT_MARKED(p))
 246                                 break;
 247                 }
 248                 do {
 249                         struct mount *parent = last_source->mnt_parent;
 250                         if (last_source == first_source)
 251                                 break;
 252                         done = parent->mnt_master == p;
 253                         if (done && peers(n, parent))
 254                                 break;
 255                         last_source = last_source->mnt_master;
 256                 } while (!done);
 257
 258                 type = CL_SLAVE;
 259                 /* beginning of peer group among the slaves? */
 260                 if (IS_MNT_SHARED(m))
 261                         type |= CL_MAKE_SHARED;
 262         }
 263
 264         /* Notice when we are propagating across user namespaces */
 265         if (m->mnt_ns->user_ns != user_ns)
 266                 type |= CL_UNPRIVILEGED;
 267         child = copy_tree(last_source, last_source->mnt.mnt_root, type);
 268         if (IS_ERR(child))
 269                 return PTR_ERR(child);
 270         child->mnt.mnt_flags &= ~MNT_LOCKED;
 271         mnt_set_mountpoint(m, mp, child);
 272         last_dest = m;
 273         last_source = child;
 274         if (m->mnt_master != dest_master) {
 275                 read_seqlock_excl(&mount_lock);
 276                 SET_MNT_MARK(m->mnt_master);
 277                 read_sequnlock_excl(&mount_lock);
 278         }
 279         hlist_add_head(&child->mnt_hash, list);
 280         return count_mounts(m->mnt_ns, child);
 281 }
 282
 283 /*
 284  * mount 'source_mnt' under the destination 'dest_mnt' at
 285  * dentry 'dest_dentry'. And propagate that mount to
 286  * all the peer and slave mounts of 'dest_mnt'.
 287  * Link all the new mounts into a propagation tree headed at
 288  * source_mnt. Also link all the new mounts using ->mnt_list
 289  * headed at source_mnt's ->mnt_list
 290  *
 291  * @dest_mnt: destination mount.
 292  * @dest_dentry: destination dentry.
 293  * @source_mnt: source mount.
 294  * @tree_list : list of heads of trees to be attached.
 295  */
 296 int propagate_mnt(struct mount *dest_mnt, struct mountpoint *dest_mp,
 297                     struct mount *source_mnt, struct hlist_head *tree_list)
 298 {
 299         struct mount *m, *n;
 300         int ret = 0;
 301
 302         /*
 303          * we don't want to bother passing tons of arguments to
 304          * propagate_one(); everything is serialized by namespace_sem,
 305          * so globals will do just fine.
 306          */
 307         user_ns = current->nsproxy->mnt_ns->user_ns;
 308         last_dest = dest_mnt;
 309         first_source = source_mnt;
 310         last_source = source_mnt;
 311         mp = dest_mp;
 312         list = tree_list;
 313         dest_master = dest_mnt->mnt_master;
 314
 315         /* all peers of dest_mnt, except dest_mnt itself */
 316         for (n = next_peer(dest_mnt); n != dest_mnt; n = next_peer(n)) {
 317                 ret = propagate_one(n);
 318                 if (ret)
 319                         goto out;
 320         }
 321
 322         /* all slave groups */
 323         for (m = next_group(dest_mnt, dest_mnt); m;
 324                         m = next_group(m, dest_mnt)) {
 325                 /* everything in that slave group */
 326                 n = m;
 327                 do {
 328                         ret = propagate_one(n);
 329                         if (ret)
 330                                 goto out;
 331                         n = next_peer(n);
 332                 } while (n != m);
 333         }
 334 out:
 335         read_seqlock_excl(&mount_lock);
 336         hlist_for_each_entry(n, tree_list, mnt_hash) {
 337                 m = n->mnt_parent;
 338                 if (m->mnt_master != dest_mnt->mnt_master)
 339                         CLEAR_MNT_MARK(m->mnt_master);
 340         }
 341         read_sequnlock_excl(&mount_lock);
 342         return ret;
 343 }
 344
 345 static struct mount *find_topper(struct mount *mnt)
 346 {
 347         /* If there is exactly one mount covering mnt completely return it. */
 348         struct mount *child;
 349
 350         if (!list_is_singular(&mnt->mnt_mounts))
 351                 return NULL;
 352
 353         child = list_first_entry(&mnt->mnt_mounts, struct mount, mnt_child);
 354         if (child->mnt_mountpoint != mnt->mnt.mnt_root)
 355                 return NULL;
 356
 357         return child;
 358 }
 359
 360 /*
 361  * return true if the refcount is greater than count
 362  */
 363 static inline int do_refcount_check(struct mount *mnt, int count)
 364 {
 365         return mnt_get_count(mnt) > count;
 366 }
 367
 368 /*
 369  * check if the mount 'mnt' can be unmounted successfully.
 370  * @mnt: the mount to be checked for unmount
 371  * NOTE: unmounting 'mnt' would naturally propagate to all
 372  * other mounts its parent propagates to.
 373  * Check if any of these mounts that **do not have submounts**
 374  * have more references than 'refcnt'. If so return busy.
 375  *
 376  * vfsmount lock must be held for write
 377  */
 378 int propagate_mount_busy(struct mount *mnt, int refcnt)
 379 {
 380         struct mount *m, *child, *topper;
 381         struct mount *parent = mnt->mnt_parent;
 382
 383         if (mnt == parent)
 384                 return do_refcount_check(mnt, refcnt);
 385
 386         /*
 387          * quickly check if the current mount can be unmounted.
 388          * If not, we don't have to go checking for all other
 389          * mounts
 390          */
 391         if (!list_empty(&mnt->mnt_mounts) || do_refcount_check(mnt, refcnt))
 392                 return 1;
 393
 394         for (m = propagation_next(parent, parent); m;
 395                         m = propagation_next(m, parent)) {
 396                 int count = 1;
 397                 child = __lookup_mnt(&m->mnt, mnt->mnt_mountpoint);
 398                 if (!child)
 399                         continue;
 400
 401                 /* Is there exactly one mount on the child that covers
 402                  * it completely whose reference should be ignored?
 403                  */
 404                 topper = find_topper(child);
 405                 if (topper)
 406                         count += 1;
 407                 else if (!list_empty(&child->mnt_mounts))
 408                         continue;
 409
 410                 if (do_refcount_check(child, count))
 411                         return 1;
 412         }
 413         return 0;
 414 }
 415
 416 /*
 417  * Clear MNT_LOCKED when it can be shown to be safe.
 418  *
 419  * mount_lock lock must be held for write
 420  */
 421 void propagate_mount_unlock(struct mount *mnt)
 422 {
 423         struct mount *parent = mnt->mnt_parent;
 424         struct mount *m, *child;
 425
 426         BUG_ON(parent == mnt);
 427
 428         for (m = propagation_next(parent, parent); m;
 429                         m = propagation_next(m, parent)) {
 430                 child = __lookup_mnt(&m->mnt, mnt->mnt_mountpoint);
 431                 if (child)
 432                         child->mnt.mnt_flags &= ~MNT_LOCKED;
 433         }
 434 }
 435
 436 static void umount_one(struct mount *mnt, struct list_head *to_umount)
 437 {
 438         CLEAR_MNT_MARK(mnt);
 439         mnt->mnt.mnt_flags |= MNT_UMOUNT;
 440         list_del_init(&mnt->mnt_child);
 441         list_del_init(&mnt->mnt_umounting);
 442         list_move_tail(&mnt->mnt_list, to_umount);
 443 }
 444
 445 /*
 446  * NOTE: unmounting 'mnt' naturally propagates to all other mounts its
 447  * parent propagates to.
 448  */
 449 static bool __propagate_umount(struct mount *mnt,
 450                                struct list_head *to_umount,
 451                                struct list_head *to_restore)
 452 {
 453         bool progress = false;
 454         struct mount *child;
 455
 456         /*
 457          * The state of the parent won't change if this mount is
 458          * already unmounted or marked as without children.
 459          */
 460         if (mnt->mnt.mnt_flags & (MNT_UMOUNT | MNT_MARKED))
 461                 goto out;
 462
 463         /* Verify topper is the only grandchild that has not been
 464          * speculatively unmounted.
 465          */
 466         list_for_each_entry(child, &mnt->mnt_mounts, mnt_child) {
 467                 if (child->mnt_mountpoint == mnt->mnt.mnt_root)
 468                         continue;
 469                 if (!list_empty(&child->mnt_umounting) && IS_MNT_MARKED(child))
 470                         continue;
 471                 /* Found a mounted child */
 472                 goto children;
 473         }
 474
 475         /* Mark mounts that can be unmounted if not locked */
 476         SET_MNT_MARK(mnt);
 477         progress = true;
 478
 479         /* If a mount is without children and not locked umount it. */
 480         if (!IS_MNT_LOCKED(mnt)) {
 481                 umount_one(mnt, to_umount);
 482         } else {
 483 children:
 484                 list_move_tail(&mnt->mnt_umounting, to_restore);
 485         }
 486 out:
 487         return progress;
 488 }
 489
 490 static void umount_list(struct list_head *to_umount,
 491                         struct list_head *to_restore)
 492 {
 493         struct mount *mnt, *child, *tmp;
 494         list_for_each_entry(mnt, to_umount, mnt_list) {
 495                 list_for_each_entry_safe(child, tmp, &mnt->mnt_mounts, mnt_child) {
 496                         /* topper? */
 497                         if (child->mnt_mountpoint == mnt->mnt.mnt_root)
 498                                 list_move_tail(&child->mnt_umounting, to_restore);
 499                         else
 500                                 umount_one(child, to_umount);
 501                 }
 502         }
 503 }
 504
 505 static void restore_mounts(struct list_head *to_restore)
 506 {
 507         /* Restore mounts to a clean working state */
 508         while (!list_empty(to_restore)) {
 509                 struct mount *mnt, *parent;
 510                 struct mountpoint *mp;
 511
 512                 mnt = list_first_entry(to_restore, struct mount, mnt_umounting);
 513                 CLEAR_MNT_MARK(mnt);
 514                 list_del_init(&mnt->mnt_umounting);
 515
 516                 /* Should this mount be reparented? */
 517                 mp = mnt->mnt_mp;
 518                 parent = mnt->mnt_parent;
 519                 while (parent->mnt.mnt_flags & MNT_UMOUNT) {
 520                         mp = parent->mnt_mp;
 521                         parent = parent->mnt_parent;
 522                 }
 523                 if (parent != mnt->mnt_parent)
 524                         mnt_change_mountpoint(parent, mp, mnt);
 525         }
 526 }
 527
 528 static void cleanup_umount_visitations(struct list_head *visited)
 529 {
 530         while (!list_empty(visited)) {
 531                 struct mount *mnt =
 532                         list_first_entry(visited, struct mount, mnt_umounting);
 533                 list_del_init(&mnt->mnt_umounting);
 534         }
 535 }
 536
 537 /*
 538  * collect all mounts that receive propagation from the mount in @list,
 539  * and return these additional mounts in the same list.
 540  * @list: the list of mounts to be unmounted.
 541  *
 542  * vfsmount lock must be held for write
 543  */
 544 int propagate_umount(struct list_head *list)
 545 {
 546         struct mount *mnt;
 547         LIST_HEAD(to_restore);
 548         LIST_HEAD(to_umount);
 549         LIST_HEAD(visited);
 550
 551         /* Find candidates for unmounting */
 552         list_for_each_entry_reverse(mnt, list, mnt_list) {
 553                 struct mount *parent = mnt->mnt_parent;
 554                 struct mount *m;
 555
 556                 /*
 557                  * If this mount has already been visited it is known that it's
 558                  * entire peer group and all of their slaves in the propagation
 559                  * tree for the mountpoint has already been visited and there is
 560                  * no need to visit them again.
 561                  */
 562                 if (!list_empty(&mnt->mnt_umounting))
 563                         continue;
 564
 565                 list_add_tail(&mnt->mnt_umounting, &visited);
 566                 for (m = propagation_next(parent, parent); m;
 567                      m = propagation_next(m, parent)) {
 568                         struct mount *child = __lookup_mnt(&m->mnt,
 569                                                            mnt->mnt_mountpoint);
 570                         if (!child)
 571                                 continue;
 572
 573                         if (!list_empty(&child->mnt_umounting)) {
 574                                 /*
 575                                  * If the child has already been visited it is
 576                                  * know that it's entire peer group and all of
 577                                  * their slaves in the propgation tree for the
 578                                  * mountpoint has already been visited and there
 579                                  * is no need to visit this subtree again.
 580                                  */
 581                                 m = skip_propagation_subtree(m, parent);
 582                                 continue;
 583                         } else if (child->mnt.mnt_flags & MNT_UMOUNT) {
 584                                 /*
 585                                  * We have come accross an partially unmounted
 586                                  * mount in list that has not been visited yet.
 587                                  * Remember it has been visited and continue
 588                                  * about our merry way.
 589                                  */
 590                                 list_add_tail(&child->mnt_umounting, &visited);
 591                                 continue;
 592                         }
 593
 594                         /* Check the child and parents while progress is made */
 595                         while (__propagate_umount(child,
 596                                                   &to_umount, &to_restore)) {
 597                                 /* Is the parent a umount candidate? */
 598                                 child = child->mnt_parent;
 599                                 if (list_empty(&child->mnt_umounting))
 600                                         break;
 601                         }
 602                 }
 603         }
 604
 605         umount_list(&to_umount, &to_restore);
 606         restore_mounts(&to_restore);
 607         cleanup_umount_visitations(&visited);
 608         list_splice_tail(&to_umount, list);
 609
 610         return 0;
 611 }