]> git.baikalelectronics.ru Git - kernel.git/commitdiff
namei: LOOKUP_{IN_ROOT,BENEATH}: permit limited ".." resolution
authorAleksa Sarai <cyphar@cyphar.com>
Fri, 6 Dec 2019 14:13:35 +0000 (01:13 +1100)
committerAl Viro <viro@zeniv.linux.org.uk>
Mon, 9 Dec 2019 00:09:44 +0000 (19:09 -0500)
Allow LOOKUP_BENEATH and LOOKUP_IN_ROOT to safely permit ".." resolution
(in the case of LOOKUP_BENEATH the resolution will still fail if ".."
resolution would resolve a path outside of the root -- while
LOOKUP_IN_ROOT will chroot(2)-style scope it). Magic-link jumps are
still disallowed entirely[*].

As Jann explains[1,2], the need for this patch (and the original no-".."
restriction) is explained by observing there is a fairly easy-to-exploit
race condition with chroot(2) (and thus by extension LOOKUP_IN_ROOT and
LOOKUP_BENEATH if ".." is allowed) where a rename(2) of a path can be
used to "skip over" nd->root and thus escape to the filesystem above
nd->root.

  thread1 [attacker]:
    for (;;)
      renameat2(AT_FDCWD, "/a/b/c", AT_FDCWD, "/a/d", RENAME_EXCHANGE);
  thread2 [victim]:
    for (;;)
      openat2(dirb, "b/c/../../etc/shadow",
              { .flags = O_PATH, .resolve = RESOLVE_IN_ROOT } );

With fairly significant regularity, thread2 will resolve to
"/etc/shadow" rather than "/a/b/etc/shadow". There is also a similar
(though somewhat more privileged) attack using MS_MOVE.

With this patch, such cases will be detected *during* ".." resolution
and will return -EAGAIN for userspace to decide to either retry or abort
the lookup. It should be noted that ".." is the weak point of chroot(2)
-- walking *into* a subdirectory tautologically cannot result in you
walking *outside* nd->root (except through a bind-mount or magic-link).
There is also no other way for a directory's parent to change (which is
the primary worry with ".." resolution here) other than a rename or
MS_MOVE.

The primary reason for deferring to userspace with -EAGAIN is that an
in-kernel retry loop (or doing a path_is_under() check after re-taking
the relevant seqlocks) can become unreasonably expensive on machines
with lots of VFS activity (nfsd can cause lots of rename_lock updates).
Thus it should be up to userspace how many times they wish to retry the
lookup -- the selftests for this attack indicate that there is a ~35%
chance of the lookup succeeding on the first try even with an attacker
thrashing rename_lock.

A variant of the above attack is included in the selftests for
openat2(2) later in this patch series. I've run this test on several
machines for several days and no instances of a breakout were detected.
While this is not concrete proof that this is safe, when combined with
the above argument it should lend some trustworthiness to this
construction.

[*] It may be acceptable in the future to do a path_is_under() check for
    magic-links after they are resolved. However this seems unlikely to
    be a feature that people *really* need -- it can be added later if
    it turns out a lot of people want it.

[1]: https://lore.kernel.org/lkml/CAG48ez1jzNvxB+bfOBnERFGp=oMM0vHWuLD6EULmne3R6xa53w@mail.gmail.com/
[2]: https://lore.kernel.org/lkml/CAG48ez30WJhbsro2HOc_DR7V91M+hNFzBP5ogRMZaxbAORvqzg@mail.gmail.com/

Cc: Christian Brauner <christian.brauner@ubuntu.com>
Suggested-by: Jann Horn <jannh@google.com>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Aleksa Sarai <cyphar@cyphar.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
fs/namei.c

index 53d815415810ef2273b3b3c64aa43deaaa911787..b367fdb9168262ffe5a219698a98239071070c71 100644 (file)
@@ -491,7 +491,7 @@ struct nameidata {
        struct path     root;
        struct inode    *inode; /* path.dentry.d_inode */
        unsigned int    flags;
-       unsigned        seq, m_seq;
+       unsigned        seq, m_seq, r_seq;
        int             last_type;
        unsigned        depth;
        int             total_link_count;
@@ -1793,22 +1793,31 @@ static inline int handle_dots(struct nameidata *nd, int type)
        if (type == LAST_DOTDOT) {
                int error = 0;
 
-               /*
-                * Scoped-lookup flags resolving ".." is not currently safe --
-                * races can cause our parent to have moved outside of the root
-                * and us to skip over it.
-                */
-               if (unlikely(nd->flags & LOOKUP_IS_SCOPED))
-                       return -EXDEV;
                if (!nd->root.mnt) {
                        error = set_root(nd);
                        if (error)
                                return error;
                }
-               if (nd->flags & LOOKUP_RCU) {
-                       return follow_dotdot_rcu(nd);
-               } else
-                       return follow_dotdot(nd);
+               if (nd->flags & LOOKUP_RCU)
+                       error = follow_dotdot_rcu(nd);
+               else
+                       error = follow_dotdot(nd);
+               if (error)
+                       return error;
+
+               if (unlikely(nd->flags & LOOKUP_IS_SCOPED)) {
+                       /*
+                        * If there was a racing rename or mount along our
+                        * path, then we can't be sure that ".." hasn't jumped
+                        * above nd->root (and so userspace should retry or use
+                        * some fallback).
+                        */
+                       smp_rmb();
+                       if (unlikely(__read_seqcount_retry(&mount_lock.seqcount, nd->m_seq)))
+                               return -EAGAIN;
+                       if (unlikely(__read_seqcount_retry(&rename_lock.seqcount, nd->r_seq)))
+                               return -EAGAIN;
+               }
        }
        return 0;
 }
@@ -2273,6 +2282,11 @@ static const char *path_init(struct nameidata *nd, unsigned flags)
        nd->last_type = LAST_ROOT; /* if there are only slashes... */
        nd->flags = flags | LOOKUP_JUMPED | LOOKUP_PARENT;
        nd->depth = 0;
+
+       nd->m_seq = __read_seqcount_begin(&mount_lock.seqcount);
+       nd->r_seq = __read_seqcount_begin(&rename_lock.seqcount);
+       smp_rmb();
+
        if (flags & LOOKUP_ROOT) {
                struct dentry *root = nd->root.dentry;
                struct inode *inode = root->d_inode;
@@ -2281,9 +2295,8 @@ static const char *path_init(struct nameidata *nd, unsigned flags)
                nd->path = nd->root;
                nd->inode = inode;
                if (flags & LOOKUP_RCU) {
-                       nd->seq = __read_seqcount_begin(&nd->path.dentry->d_seq);
+                       nd->seq = read_seqcount_begin(&nd->path.dentry->d_seq);
                        nd->root_seq = nd->seq;
-                       nd->m_seq = read_seqbegin(&mount_lock);
                } else {
                        path_get(&nd->path);
                }
@@ -2294,8 +2307,6 @@ static const char *path_init(struct nameidata *nd, unsigned flags)
        nd->path.mnt = NULL;
        nd->path.dentry = NULL;
 
-       nd->m_seq = read_seqbegin(&mount_lock);
-
        /* Absolute pathname -- fetch the root (LOOKUP_IN_ROOT uses nd->dfd). */
        if (*s == '/' && !(flags & LOOKUP_IN_ROOT)) {
                error = nd_jump_root(nd);