Re: [PATCH 1/2] ARM: entry-common: fix forgotten set of thread_info->syscall

[Roman Peniaev <r.peniaev@gmail.com>]

On Thu, Jan 22, 2015 at 8:32 AM, Kees Cook <keescook@chromium.org> wrote:
> On Tue, Jan 20, 2015 at 3:04 PM, Russell King - ARM Linux
> <linux@arm.linux.org.uk> wrote:
>> On Tue, Jan 20, 2015 at 10:45:19PM +0000, Russell King - ARM Linux wrote:
>>> Well, the whole question is this: is restarting a system call like
>>> usleep() really a separate system call, or is it a kernel implementation
>>> detail?
>>>
>>> If you wanted seccomp to see this, what would be the use case?  Why
>>> would seccomp want to block this syscall?  Does it make sense for
>>> seccomp to block this syscall when it doesn't block something like
>>> usleep() and then have usleep() fail just because the thread received
>>> a signal?
>>>
>>> I personally regard the whole restart system call thing as a purely
>>> kernel internal thing which should not be exposed to userland.  If
>>> we decide that it should be exposed to userland, then it becomes part
>>> of the user ABI, and it /could/ become difficult if we needed to
>>> change it in future - and I'd rather not get into the "oh shit, we
>>> can't change it because that would break app X" crap.
>>
>> Here's a scenario where it could become a problem:
>>
>> Let's say that we want to use seccomp to secure some code which issues
>> system calls.  We determine that the app uses system calls which don't
>> result in the restart system call being issued, so we decide to ask
>> seccomp to block the restart system call.  Some of these system calls
>> that the app was using are restartable system calls.
>>
>> When these system calls are restarted, what we see via ptrace etc is
>> that the system call simply gets re-issued as its own system call.
>>
>> In a future kernel version, we decide that we could really do with one
>> of those system calls using the restart block feature, so we arrange
>> for it to set up the restart block, and return -ERESTART_BLOCK.  That's
>> fine for most applications, but this app now breaks.
>>
>> The side effect of that breakage is that we have to revert that kernel
>> change - because we've broken userland, and that's simply not allowed.
>>
>> Now look at the alternative: we don't make the restart syscall visible.
>> This means that we hide that detail, and we actually reflect the
>> behaviour that we've had for the other system call restart mechanisms,
>> and we don't have to fear userspace breakage as a result of switching
>> from one restart mechanism to another.
>>
>> I am very much of the opinion that we should be trying to limit the
>> exposure of inappropriate kernel internal details to userland, because
>> userland has a habbit of becoming reliant on them, and when it does,
>> it makes kernel maintanence unnecessarily harder.
>
> I totally agree with you. :) My question here is more about what we
> should do with what we currently have since we have some unexpected
> combinations.
>
> There is already an __NR_restart_syscall syscall and it seems like
> it's already part of the userspace ABI:
>  - it is possible to call it from userspace directly
>  - seccomp "sees" it
>  - ptrace doesn't see it
>
> Native ARM64 hides the restart from both seccomp and ptrace, and this
> seems like the right idea, except that restart_syscall is still
> callable from userspace. I don't think there's a way to make that
> vanish, which means we'll always have an exposed syscall. If anything
> goes wrong with it (which we've been quite close to recently[1]),
> there would be no way to have seccomp filter it.
>
> So, at the least, I'd like arm64 to NOT hide restart_syscall from
> seccomp, and at best I'd like both arm and arm64 to (somehow) entirely
> remove restart_syscall from the userspace ABI so it wouldn't need to
> be filtered, and wouldn't become a weird ABI hiccup as you've
> described.
>
> I fail to imagine a way to remove restart_syscall from userspace, so
> I'm left with wanting parity of behavior between ARM and ARM64 (and
> x86). What's the right way forward?

My sufferings are an opposite of what seccompt expects: currently I do
not see any possible way (from userspace) to get syscall number which was
restarted, because at some given time userspace checks the procfs
syscall file and sees NR_restart there, without any chance to understand
what exactly was restarted (I am talking about some kind of debugging
tool which does dead-lock analysis of stuck tasks).

I totally agree with Russell not to provide kernel guts to userspace,
but it is already done.  Too late.

So probably there is a need to split syscall on two numbers:
real and effective.  Real number we have right now on x86.

And this should be done for both ptrace and procfs syscall file.
(am I right that only for ARM we already have PTRACE_SET_SYSCALL?
 seems we can add also real/effective getter)

--
Roman