Re: [PATCH] x86_64: test that userspace stack is in fact NX

[Alexey Dobriyan]

On Mon, Oct 02, 2023 at 07:23:10AM -0700, Dave Hansen wrote:
> On 10/1/23 09:31, Alexey Dobriyan wrote:
> > Here is how it works:
> > 
> > * fault and fill the stack from rsp with int3 down until rlimit allows,
> > * fill upwards with int3 too, overwrite libc stuff, argv, envp,
> > * try to exec int3 on each page and catch it with either SIGSEGV or
> >   SIGTRAP handler.
> > 
> > Note: trying to execute _every_ int3 takes 30-40 seconds even on fast
> > machine, so only 1 int3 per page is tried.
> > 
> > Tested on F37 kernel and on custom kernel which did
> > 
> > 	vm_flags |= VM_EXEC;
> > 
> > to stack VMA.
> 
> I guess the subject implies it, but it's probably worth a sentence or
> two in the changelog about this being 64-bit only.
> 
> IIRC, there _are_ x86_64 CPUs that don't support NX.  It's also entirely
> possible for a hypervisor to disable NX enumeration for a guest.  Those
> two are (probably) rare enough that they can be ignored for now.  But it
> might mean adding a CPUID check at some point.
> 
> Basically, could you spend a moment in the changelog to talk about:
> 
> 1. 32-bit kernels on NX hardware
> and
> 2. 64-bit kernels on non-NX hardware

Sure. My logic whas that i386 is dead arch, but this test is easy to
port to i386, only 2 simple functions.

I don't want to parse /proc/cpuinfo. If someone knows they're shipping
NX-incapable hardware, just let them disable the test.