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