Re: [PATCH v4 3/4] Introduce the NVMM impl

[Maxime Villard]

Le 02/03/2020 à 20:35, Paolo Bonzini a écrit :
> 
> 
> Il lun 2 mar 2020, 20:28 Maxime Villard <address@hidden 
> <mailto:address@hidden>> ha scritto:
> 
> 
>     >> +        nvmm_vcpu_pre_run(cpu);
>     >> +
>     >> +        if (atomic_read(&cpu->exit_request)) {
>     >> +            qemu_cpu_kick_self();
>     >> +        }
>     >> +
>     >
>     > This is racy without something like KVM's immediate_exit mechanism.
>     > This should be fixed in NVMM.
> 
>     I don't immediately see how this is racy.
> 
> 
> You can get an IPI signal immediately after reading cpu->exit_request.
> 
>     It reproduces the existing
>     logic found in whpx-all.c, and if there is a real problem it can be
>     fixed in a future commit along with WHPX.
> 
> 
> It's buggy there too and it has to be fixed in the hypervisor so it can't be 
> done at the same time I'm both. KVM does it right by having a flag 
> ("immediate_exit") that is set by the signal handler and checked by the 
> hypervisor.
> 
> An earlier version of KVM instead atomically unblocked the signal while 
> executing the guest, and then ate it with a sigwaitinfo after exiting back to 
> userspace.
> 
> You don't have to fix it immediately, but adding a FIXME would be a good idea.
> 
> Paolo

Kamil, please add /* FIXME: possible race here */ before the atomic_read().

Thanks