On Thu, May 4, 2017 at 9:39 PM, Al Viro <viro@xxxxxxxxxxxxxxxxxx> wrote:
> On Thu, May 04, 2017 at 08:46:49PM -0700, Linus Torvalds wrote:
>> On Thu, May 4, 2017 at 7:47 PM, Jann Horn <jannh@xxxxxxxxxx> wrote:
>> >
>> > Thread 1 starts an AT_BENEATH path walk using an O_PATH fd
>> > pointing to /srv/www/example.org/foo; the path given to the syscall is
>> > "bar/../../../../etc/passwd". The path walk enters the "bar" directory.
>> > Thread 2 moves /srv/www/example.org/foo/bar to
>> > /srv/www/example.org/bar.
>> > Thread 1 processes the rest of the path ("../../../../etc/passwd"), never
>> > hitting /srv/www/example.org/foo in the process.
>> >
>> > I'm not really familiar with the VFS internals, but from a coarse look
>> > at the patch, it seems like it wouldn't block this?
>>
>> I think you're right.
>>
>> I guess it would be safe for the RCU case due to the sequence number
>> check, but not the non-RCU case.
>
> Yes and no... FWIW, to exclude that it would suffice to have
> mount --rbind /src/www/example.org/foo /srv/www/example.org/foo done first.
> Then this kind of race will end up with -ENOENT due to path_connected()
> logics in follow_dotdot_rcu()/follow_dotdot(). I'm not sure about the
> intended applications, though - is that thing supposed to be used along with
> some horror like seccomp, or...?
How hard would it be for the kernel to prevent this on its own?
Asking users to do the mount --rbind seems like it's asking for users
to forget to do it.
