Re: [PATCH v2] entry/kvm: Make vCPU tasks exit to userspace when a livepatch is pending

[Seth Forshee]

On Wed, May 04, 2022 at 04:28:09PM +0200, Petr Mladek wrote:
> On Wed 2022-05-04 08:50:22, Seth Forshee wrote:
> > On Wed, May 04, 2022 at 03:07:53PM +0200, Petr Mladek wrote:
> > > On Tue 2022-05-03 12:49:34, Seth Forshee wrote:
> > > > A task can be livepatched only when it is sleeping or it exits to
> > > > userspace. This may happen infrequently for a heavily loaded vCPU task,
> > > > leading to livepatch transition failures.
> > > 
> > > The problem was solved by sending a fake signal, see the commit
> > > 0b3d52790e1cfd6b80b826 ("livepatch: Remove signal sysfs attribute").
> > > It was achieved by calling signal_wake_up(). It set TIF_SIGPENDING
> > > and woke the task. It interrupted the syscall and the task was
> > > transitioned when leaving to the userspace.
> > > 
> > > signal_wake_up() was later replaced by set_notify_signal(),
> > > see the commit 8df1947c71ee53c7e21 ("livepatch: Replace
> > > the fake signal sending with TIF_NOTIFY_SIGNAL infrastructure").
> > > The difference is that set_notify_signal() uses TIF_NOTIFY_SIGNAL
> > > instead of TIF_SIGPENDING.
> > > 
> > > The effect is the same when running on a real hardware. The syscall
> > > gets interrupted and exit_to_user_mode_loop() is called where
> > > the livepatch state is updated (task migrated).
> > > 
> > > But it works a different way in kvm where the task works are
> > > called in the guest mode and the task does not return into
> > > the user space in the host mode.
> > 
> > > > --- a/kernel/entry/kvm.c
> > > > +++ b/kernel/entry/kvm.c
> > > > @@ -14,7 +14,12 @@ static int xfer_to_guest_mode_work(struct kvm_vcpu *vcpu, unsigned long ti_work)
> > > >  				task_work_run();
> > > >  		}
> > > >  
> > > > -		if (ti_work & _TIF_SIGPENDING) {
> > > > +		/*
> > > > +		 * When a livepatch is pending, force an exit to userspace
> > > > +		 * as though a signal is pending to allow the task to be
> > > > +		 * patched.
> > > > +		 */
> > > > +		if (ti_work & (_TIF_SIGPENDING | _TIF_PATCH_PENDING)) {
> > > >  			kvm_handle_signal_exit(vcpu);
> 
> Another problem. Is it safe to call kvm_handle_signal_exit(vcpu)
> for kthreads?
> 
> kthreads have _TIF_PATCH_PENDING when they need the livepatch transition.
> But kthreads never leave kernel so we do not send the fake signal
> signals to them.

xfer_to_guest_mode_handle_work() should only be getting called on user
threads running ioctl(KVM_RUN).

> 
> > > >  			return -EINTR;
> > > >  		}
> > > 
> > > Does xfer_to_guest_mode_work() interrupts the syscall running
> > > on the guest?
> > 
> > xfer_to_guest_mode_work() is called as part of a loop to execute kvm
> > guests (for example, on x86 see vcpu_run() in arch/x86/kvm/x86.c). When
> > guest execution is interrupted (in the livepatch case it is interrupted
> > when set_notify_signal() is called for the vCPU task)
> > xfer_to_guest_mode_work() is called if there is pending work, and if it
> > returns non-zero the loop does not immediately re-enter guest execution
> > but instead returns to userspace.
> 
> Thanks for the detailed explanation.
> 
> 
> > > If "yes" then we do not need to call kvm_handle_signal_exit(vcpu).
> > > It will be enough to call:
> > > 
> > > 		if (ti_work & _TIF_PATCH_PENDING)
> > > 			klp_update_patch_state(current);
> > 
> > What if the task's call stack contains a function being patched?
> 
> We do not need to check the stack when the syscall gets restarted.
> The task might be transitioned only when the syscall gets restarted.

I see. Thanks!

> > > If "no" then I do not understand why TIF_NOTIFY_SIGNAL interrupts
> > > the syscall on the real hardware and not in kvm.
> > 
> > It does interrupt, but xfer_to_guest_mode_handle_work() concludes it's
> > not necessary to return to userspace and resumes guest execution.
> 
> In this case, we should revert the commit 8df1947c71ee53c7e21
> ("livepatch: Replace the fake signal sending with TIF_NOTIFY_SIGNAL
> infrastructure"). The flag TIF_NOTIFY_SIGNAL clearly does not guarantee
> restarting the syscall or exiting to the user space with -EINTR.
> 
> It should solve this problem. And it looks like a cleaner solution
> to me.

It looks like that should fix the issue. I'll test to confirm.

Thanks,
Seth