|Main Archive Page > Month Archives > oss-security archives|
The libc' sigqueue() function allows to queue a signal, as well as some
accompanying data to a process.
The kernel's interface that is used to implement this function is known
as rt_sigqueueinfo(). It has been added in Linux 2.2.
This system call is interesting from a security perspective, because it
allows userland to compeletely specify the siginfo_t structure. This
structure is normally typically almost entirely written by the kernel
when a signal is delivered.
Since at least Linux 2.4.0, most abuses of the kernel interface have
been prevented with a simple check:
/* Not even root can pretend to send signals from the kernel.
Nor can they impersonate a kill(), which adds source info. */
if (info.si_code >= 0)
This check made sure that rt_sigqueueinfo() could not spoof a signal
whose SI_CODE would be SI_KERNEL or SI_USER. As the comment indicates, a
process receiving a signal should be able to trust its source pid or uid
if its si_code matches SI_USER.
Unfortunately, a couple of years later, when tgkill() and tkill() were
added, this check was forgotten and was not updated to prevent the
spoofing of a TGKILL si_code. Because of this, userland is unable to
trust the pid and uid information of a TKILL signal.
This is bad, because it is a useful feature in a scenario where a
process which cannot ptrace you can send you signals. This includes at
least the startup code of setuid binaries.
Meanwhile, userland and libc writers still assumed that they could trust
the origin of a SI_TKILL signal. Glibc authors too . Worse: they
even silently patched SI_TKILL with SI_USER , . So even a userland
application that (righfully so) only trusts SI_USER signals will be
A tentative patch for this vulnerability has been committed to Linus'
kernel tree .
In this patch, we prevent rt_sigqueueinfo() from specifying any si_code
!= SI_QUEUE. While we believe it to be very unlikley, this could in
theory break userland in some older Linux distributions, so we may
have to revert to a more concervative patch and prevent ( (si_code ==
SI_TKILL) || (si_code >= SI_QUEUE) ) instead.
Please credit "Julien Tinnes, Google security team" in any related advisory.