Re: [PATCH 0/2] Fix and re-enable vsyscall=emulate

["H. Peter Anvin" <hpa@kernel.org>]

On 12/02/2011 02:47 PM, Andy Lutomirski wrote:
> On Mon, Nov 7, 2011 at 4:33 PM, Andy Lutomirski <luto@amacapital.net> wrote:
>> The really nice fix (wiring up access_ok failures to be able to raise
>> signals) won't be ready on time for 3.2, so let's try the simpler fix
>> for now.
> 
> I spoke to hpa about this a couple days ago, and he pointed out a
> problem with making access_ok send signals.  Userspace expects signals
> that come with full context information to be restartable, and many
> system calls are not restartable.  read() and write() are the obvious
> examples: once they're processed the beginning of the buffer, unless
> they adjust their parameters, they can't safely be restarted.  So
> without massive changes, I think allowing access_ok to raise a signal
> with full context is asking for trouble.
> 
> I can still do the patch with two modes: signals without context via
> arch_prctl and signals with context via vsyscall emulation, but that's
> probably overkill for fixing this bug.  I'd say just apply these
> patches as is (for 3.3).
> 

It's somewhat questionable if the "return -EFAULT and deliver SIGSEGV"
semantic resolves the problem; obviously the signal handler isn't
restartable, but returning from the signal handler will at least cause
the application to see the EFAULT and not try to restart a system call
in a way that is likely to cause massive failure.  If the handler is
aware about what needs to be done then it can correct the situation and
restart the system call -- but it would have to have detailed
information about the state before the system call.

I am also concerned about information leaks from the kernel.  The
existing kernel paths are not necessarily designed to be robust against
giving out additional error information.  This may be a theoretical
concern, but there have been real security holes in the past from these
kinds of changes.

	-hpa