Re: [PATCH v3 35/37] x86/cet: Add PTRACE interface for CET

[Mike Rapoport]

On Mon, Nov 21, 2022 at 03:52:57PM +0000, Edgecombe, Rick P wrote:
> On Mon, 2022-11-21 at 09:40 +0200, Mike Rapoport wrote:
> > On Thu, Nov 17, 2022 at 07:57:59PM +0000, Edgecombe, Rick P wrote:
> > > On Thu, 2022-11-17 at 12:25 +0000, Schimpe, Christina wrote:
> > > > > Hmm, we definitely need to be able to set the SSP. Christina,
> > > > > does
> > > > > GDB need
> > > > > anything else? I thought maybe toggling SHSTK_EN?
> > > > 
> > > > In addition to the SSP, we want to write the CET state. For
> > > > instance
> > > > for inferior calls,
> > > > we want to reset the IBT bits.
> > > > However, we won't write states that are disallowed by HW.
> > > 
> > > Sorry, I should have given more background. Peter is saying we
> > > should
> > > split the ptrace interface so that shadow stack and IBT are
> > > separate. 
> > > They would also no longer necessarily mirror the CET_U MSR format.
> > > Instead the kernel would expose a kernel specific format that has
> > > the
> > > needed bits of shadow stack support. And a separate one later for
> > > IBT.
> > > 
> > > So the question is what does shadow stack need to support for
> > > ptrace
> > > besides SSP? Is it only SSP? The other features are SHSTK_EN and
> > > WRSS_EN. It might actually be nice to keep how these bits get
> > > flipped
> > > more controlled (remove them from ptrace). It looks like CRIU
> > > didn't
> > > need them.
> > 
> >  
> > CRIU reads CET_U with ptrace(PTRACE_GETREGSET, NT_X86_CET). It's done
> > before the injection of the parasite. The value of SHSTK_EN is used
> > then to
> > detect if shadow stack is enabled and to setup victim's shadow stack
> > for
> > sigreturn.
> 
> Hmm, can it read /proc/pid/status? It has some lines like this:
> x86_Thread_features: shstk wrss
> x86_Thread_features_locked: shstk wrss

It could, but that would be much more intrusive than GETREGSET because
currently /proc parsing and parasite injection don't really interact.
If anything, arch_prctl(ARCH_CET_GET) via ptrace would be much nicer than
/proc. 

-- 
Sincerely yours,
Mike.