Re: [patch V2 25/37] rseq: Rework the TIF_NOTIFY handler

[Sean Christopherson <seanjc@google.com>]

On Tue, Sep 02, 2025, Thomas Gleixner wrote:
> On Tue, Aug 26 2025 at 11:12, Mathieu Desnoyers wrote:
> > On 2025-08-23 12:40, Thomas Gleixner wrote:
> >> +void __rseq_handle_notify_resume(struct pt_regs *regs)
> >> +{
> >> +	/*
> >> +	 * If invoked from hypervisors before entering the guest via
> >> +	 * resume_user_mode_work(), then @regs is a NULL pointer.
> >> +	 *
> >> +	 * resume_user_mode_work() clears TIF_NOTIFY_RESUME and re-raises
> >> +	 * it before returning from the ioctl() to user space when
> >> +	 * rseq_event.sched_switch is set.
> >> +	 *
> >> +	 * So it's safe to ignore here instead of pointlessly updating it
> >> +	 * in the vcpu_run() loop.
> >
> > I don't think any virt user should expect the userspace fields to be
> > updated on the host process while running in guest mode, but it's good
> > to clarify that we intend to change this user-visible behavior within
> > this series, to spare any unwelcome surprise.
> 
> Actually it is not really a user-visible change.

It's definitely a user-visible change in the sense that userspace, via the guest,
will see different behavior.

> TLS::rseq is thread local and any update to it becomes only visible to
> user space once the vCPU thread actually returns to user space. Arguably
> no guest has legitimately access to the hosts VCPU thread's TLS.
> 
> You might argue, that GDB might look at the thread's TLS::rseq while the
> task runs in VCPUs guest mode. But that's completely irrelevant because
> once a task enters the kernel the RSEQ CPU/NODE/MM ids have no meaning
> anymore. They are only valid as long as the task runs in user space.

Paravirt setups, e.g. hoisting host-controlled workloads into VMs, have explored
(ab)using rseq.  In such setups, host threads are often mapped 1:1 to vCPUs, in
which case the pCPU in particular becomes interesting.

> When a task hits a breakpoint GDB can only look at the state _before_
> that and that's all what it can see when it looks at the TLS of a
> thread, which voluntarily went into the kernel via the KVM ioctl.
> 
> That update is truly a kernel internal implementation detail and it got
> introduced way _after_ the initial RSEQ implementation.

Yes, but that doesn't change the fact that a user _could_ have come to depend on
the current behavior sometime in the last ~5 years.

I'm ok formally stating that exposing rseq directly to a KVM guest is unsupported,
but I would like to explicitly call out and document the change.