Yes, of course the subject line is silly... but in fact, the vulnerable
combination actually occurs quite often. Still, I'm posting it here not
because it's a very serious flaw, but because I find it amusing and
unique. It's a CPU/BIOS/OS vulnerability, of sorts, and nobody's at fault,
of course.

Thanks to Bulba for wasting his time on helping me figure out what's going
on, and for a number of people for risking their lives testing this
problem on their systems.

To the point. If your machine:

  - is equipped with Pentium II or better,

  - has a certain type of BIOS - tested and confirmed vulnerable (the list
    is definitely open and incomplete):

      IBM ThinkPad X IZET9AWW 2.22 (09/2002)
      Dell Latitude CPx H* revision A09
      Dell Latitude CPi A* revision A15
      Compaq 686T2 v08.22.1999

    Tested but not vulnerable:

      Dell Latitude C800 revision A17
      Dell OptiPlex GX150 revision A10
      Dell Latitude C640 revision A08

...and either...

  - dual boots between a fairly recent system that supports fast syscalls
    via SYSENTER (say, Windows XP) and a system that does not (say, Linux
    2.4),

...or...

  - had run a newer SYSENTER-enabled unstable/patched kernel, later
    downgraded to a stable version...

...then your system can be DoSed in a fairly ugly way by any of your
users.

Pentium II introduced SYSENTER/SYSEXIT, a new, fast system call interface
that is considerably more effective than the traditional entry method via
INT or LCALL.

When you boot to a system that supports this mechanism, the system will
configure certain MSRs (model-specific registers) of the CPU - primarily
0x174 (CS) and 0x176 (EIP) - to point to a specific handler code.

Once 0x174 is set, an invocation of SYSENTER opcode will cause the CPU to
attempt to switch to the segment and address described in those registers.
When 0x174 is zeroed, SYSENTER will simply fail, raising GPF.

Quite unfortunately, certain BIOSes do not zero those MSRs on reboot. It
is not clear to me why the CPU does not reset those registers itself, even
after a triple fault, but it does not. There seems to be no reasonable
explanation for persistence of this setting, yet this behavior has been
confirmed with several chips - Pentium II, Pentium III Katmai and
Coppermine and others.

As a result, when a SYSENTER-enabled system is shut down and the machine
is rebooted - but not powered down - the old setup is preserved. If a
system that does not have a working SYSENTER support - as it is the case
with all stable releases of Linux - is then booted up, the new system will
continue to run with the "inherited" MSR settings. At this point, any user
can issue a SYSENTER opcode to crash the system.

Note that those MSRs remain persistent on those boxes over subsequent warm
boots, so the attack can be successful even after a very long period of
time since the other system was last booted up.

Well, that's the story.

If you're concerned, you don't have to rewire your CPU or update your BIOS
- the fix is to compile the following code and invoke it from your rc
scripts after '/sbin/insmod msr':

-- sysleave.c --

#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <fcntl.h>

int main(void) {
  unsigned long long w = 0;
  int i = open("/dev/cpu/0/msr",O_WRONLY);
  if (i<0) { printf("Cannot open MSR device (no module?).\n"); exit(1); }
  lseek(i,0x174,SEEK_SET);
  if (write(i,&w,8) < 0) { printf("MSR write error.\n"); exit(2); }
  printf("SYSENTER disabled.\n");
  return close(i);
}

-- EOF --

If you want to test your system, you can follow the guidelines posted at
http://lcamtuf.coredump.cx/bioses.txt .

Cheers!

-- 
------------------------- bash$ :(){ :|:&};: --
 Michal Zalewski * [http://lcamtuf.coredump.cx]
    Did you know that clones never use mirrors?
--------------------------- 2003-07-24 16:48 --