Skip to contentSkip to conversation

seccomp: add code to disable TSC when enabling seccomp

PATCH Chuck Ebbert 2 participants 2 messages Jul 14, 2006

lore
Share & tools
lore
#0 lore raw

In-Reply-To: <[email protected]>

Quoted 5 lines · On Fri, 14 Jul 2006 04:02:57 +0200, [email protected] wrote:
Quoted 4 lines

Also, what prevents this flag from being set on a running process? If that happens the CPU state and flag could get out of sync and this could cause problems because of the way the current code tests the flag.

Quoted 8 lines

Yes, there could be a tiny race where if the controller and seccomp tasks run on two different CPUs: the seccomp task may write to the pipe, and then read, but the read may not actually stop anywhere, because the second CPU may have enabled seccomp and answered faster than the first cpu. So there's tiny window for the TSC not to be disabled synchronously at the start of the seccomp computations (and if there are multiple seccomp tasks running the new ones could let the old ones run a timeslice with the tsc enabled).

But it looks like the mismatch could persist indefinitely: if a seccomp task inherits the wrong cr4 flag it could pass it on to another, or back to the original one and so on. I think this is the only safe way:

	if (test_tsk_thread_flag(next_p, TIF_NOTSC) ||
	    test_tsk_thread_flag(prev_p, TIF_NOTSC)) {

		/* Flip TSC disable bit if necessary. */
		unsigned int cr4 = read_cr4();

		if (test_tsk_thread_flag(next_p, TIF_NOTSC)) {
			if (!(cr4 & X86_CR4_TSD))
				write_cr4(cr4 | X86_CR4_TSD);
		} else
			write_cr4(cr4 & ~X86_CR4_TSD);
	}

(Testing TSD in the 'else' path is not worth the trouble.)

Quoted 9 lines

To fix the tiny window if it's the current task writing to self, we should also update the cr4 before returning from base.c. If it was a different task it's more complicated (we would need to send a forced sigstop, and wait the task->state to change, but then we go into the ptrace parallelism I truly don't want to deal with in any way in seccomp context). The whole point of seccomp is to be simple. So my suggestion is either we ignore the tiny window, or we do it only from the current task. If I've to deal with any sigstop then I could use ptrace or utrace in the first place ;).

The tiny window shouldn't be a problem, should it? Just what is the risk to begin with, and how much harder is it to exploit in such a small window?

Signature · 2 lines
Chuck
 "You can't read a newspaper if you can't read."  --George W. Bush
#1 lore raw
Quoted 17 lines · Chuck Ebbert

But it looks like the mismatch could persist indefinitely: if a seccomp task inherits the wrong cr4 flag it could pass it on to another, or back to the original one and so on. I think this is the only safe way:

	if (test_tsk_thread_flag(next_p, TIF_NOTSC) ||
	    test_tsk_thread_flag(prev_p, TIF_NOTSC)) {

		/* Flip TSC disable bit if necessary. */
		unsigned int cr4 = read_cr4();

		if (test_tsk_thread_flag(next_p, TIF_NOTSC)) {
			if (!(cr4 & X86_CR4_TSD))
				write_cr4(cr4 | X86_CR4_TSD);
		} else
			write_cr4(cr4 & ~X86_CR4_TSD);
	}
Quoted 1 line

(Testing TSD in the 'else' path is not worth the trouble.)

Yes that solves the problem of the first seccomp task propagating it to the other seccomp tasks.

Quoted 3 lines

The tiny window shouldn't be a problem, should it? Just what is the risk to begin with, and how much harder is it to exploit in such a small window?

Even the original patch of yours was perfectly fine in practice. Eventually one of the per-cpu daemons would also be scheduled even if there are more seccomp tasks per cpu.

But in my opinion if we care about this window and we change something, it worth to close it completely, mostly for code-style reasons (i.e. strict code). Note that the NOTSC will never be checked unless there's a second task, so it's not a single timeslice.

As far as I can tell, the only way to close it completely, is to require that the task that fires up seccomp is the current task. That is a very fine requirement. That guarantees that no other cpu could require any update in-core. So it's enough to always flip the NOTSC bit in the current cpu with test_and_set_bit and to synchronously update the cr4 accordingly to the result of test_and_set_bit, then we can go back to the nicer (and faster) xor way in the scheduler slow path ;).

This is incidentally exactly what I implemented and tested with my last patch ;) (which is of course a modification and extension of yours)

Thanks.