On 8/27/2020 11:56 AM, Andy Lutomirski wrote:

quoted

On Aug 27, 2020, at 11:13 AM, Yu, Yu-cheng [off-list ref] wrote:

On 8/27/2020 6:36 AM, Florian Weimer wrote:

quoted

* H. J. Lu:

quoted

quoted

On Thu, Aug 27, 2020 at 6:19 AM Florian Weimer [off-list ref] wrote:

quoted

* Dave Martin:

quoted

You're right that this has implications: for i386, libc probably pulls
more arguments off the stack than are really there in some situations.
This isn't a new problem though.  There are already generic prctls with
fewer than 4 args that are used on x86.

As originally posted, glibc prctl would have to know that it has to pull
an u64 argument off the argument list for ARCH_X86_CET_DISABLE.  But
then the u64 argument is a problem for arch_prctl as well.

Argument of ARCH_X86_CET_DISABLE is int and passed in register.

The commit message and the C source say otherwise, I think (not sure
about the C source, not a kernel hacker).

H.J. Lu suggested that we fix x86 arch_prctl() to take four arguments, and then keep MMAP_SHSTK as an arch_prctl().  Because now the map flags and size are all in registers, this also solves problems being pointed out earlier.  Without a wrapper, the shadow stack mmap call (from user space) will be:

syscall(_NR_arch_prctl, ARCH_X86_CET_MMAP_SHSTK, size, MAP_32BIT).

I admit I don’t see a show stopping technical reason we can’t add arguments to an existing syscall, but I’m pretty sure it’s unprecedented, and it doesn’t seem like a good idea.

There are nine existing arch_prctl calls now.  If the concern is the 
extra new arguments getting misused, we can mask them out for the 
existing calls.  Otherwise, I have not seen anything that can break.

Yu-cheng