Paul Mackerras's on June 17, 2019 11:45 am:
> On Mon, May 20, 2019 at 10:56:57AM +1000, Nicholas Piggin wrote:
>> Signed-off-by: Nicholas Piggin <npiggin@xxxxxxxxx>
>
> Comment below...
>
>> diff --git a/arch/powerpc/kvm/book3s_hv_builtin.c b/arch/powerpc/kvm/book3s_hv_builtin.c
>> index 6035d24f1d1d..5ae7f8359368 100644
>> --- a/arch/powerpc/kvm/book3s_hv_builtin.c
>> +++ b/arch/powerpc/kvm/book3s_hv_builtin.c
>> @@ -758,6 +758,53 @@ void kvmhv_p9_restore_lpcr(struct kvm_split_mode *sip)
>> local_paca->kvm_hstate.kvm_split_mode = NULL;
>> }
>>
>> +static void kvmppc_end_cede(struct kvm_vcpu *vcpu)
>> +{
>> + vcpu->arch.ceded = 0;
>> + if (vcpu->arch.timer_running) {
>> + hrtimer_try_to_cancel(&vcpu->arch.dec_timer);
>
> So now we're potentially calling hrtimer_try_to_cancel in real mode.
> Are you absolutely sure that nothing in the hrtimer code accesses
> anything that is vmalloc'd? I'm not. Maybe you can prove that when
> called in real mode, vcpu->arch.timer_running will always be false,
> but it seems fragile to me.
Good point, no we shouldn't do this.
Is the guest always going to be out of cede at this point? Possibly
just a variant of the function that doesn't end cede would be the
go.
Thanks,
Nick
