2020-03-07 02:17:02

by Andy Lutomirski

Subject: [PATCH] x86/kvm: Disable KVM_ASYNC_PF_SEND_ALWAYS

The ABI is broken and we cannot support it properly. Turn it off.

If this causes a meaningful performance regression for someone, KVM
can introduce an improved ABI that is supportable.

Cc: [email protected]
Signed-off-by: Andy Lutomirski <[email protected]>
---
arch/x86/kernel/kvm.c | 11 ++++++++---
1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
index 93ab0cbd304e..71f9f39f93da 100644
--- a/arch/x86/kernel/kvm.c
+++ b/arch/x86/kernel/kvm.c
@@ -318,11 +318,16 @@ static void kvm_guest_cpu_init(void)

pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));

-#ifdef CONFIG_PREEMPTION
- pa |= KVM_ASYNC_PF_SEND_ALWAYS;
-#endif
pa |= KVM_ASYNC_PF_ENABLED;

+ /*
+ * We do not set KVM_ASYNC_PF_SEND_ALWAYS. With the current
+ * KVM paravirt ABI, if an async page fault occurs on an early
+ * memory access in the normal (sync) #PF path or in an NMI
+ * that happens early in the #PF code, the combination of CR2
+ * and the APF reason field will be corrupted.
+ */
+
if (kvm_para_has_feature(KVM_FEATURE_ASYNC_PF_VMEXIT))
pa |= KVM_ASYNC_PF_DELIVERY_AS_PF_VMEXIT;

--
2.24.1

2020-03-07 02:21:50

by Andy Lutomirski

Subject: Re: [PATCH] x86/kvm: Disable KVM_ASYNC_PF_SEND_ALWAYS

On Fri, Mar 6, 2020 at 6:16 PM Andy Lutomirski <[email protected]> wrote:
>
> The ABI is broken and we cannot support it properly. Turn it off.
>
> If this causes a meaningful performance regression for someone, KVM
> can introduce an improved ABI that is supportable.

Ugh, disregard this patch. I flubbed the comment, and I'll send v2 shortly.

2020-03-07 10:09:42

by Thomas Gleixner

Subject: Re: [PATCH] x86/kvm: Disable KVM_ASYNC_PF_SEND_ALWAYS

Andy Lutomirski <[email protected]> writes:

> The ABI is broken and we cannot support it properly. Turn it off.
>
> If this causes a meaningful performance regression for someone, KVM
> can introduce an improved ABI that is supportable.
>
> Cc: [email protected]
> Signed-off-by: Andy Lutomirski <[email protected]>
> ---
> arch/x86/kernel/kvm.c | 11 ++++++++---
> 1 file changed, 8 insertions(+), 3 deletions(-)
>
> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
> index 93ab0cbd304e..71f9f39f93da 100644
> --- a/arch/x86/kernel/kvm.c
> +++ b/arch/x86/kernel/kvm.c
> @@ -318,11 +318,16 @@ static void kvm_guest_cpu_init(void)
>
> pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
>
> -#ifdef CONFIG_PREEMPTION
> - pa |= KVM_ASYNC_PF_SEND_ALWAYS;
> -#endif
> pa |= KVM_ASYNC_PF_ENABLED;
>
> + /*
> + * We do not set KVM_ASYNC_PF_SEND_ALWAYS. With the current
> + * KVM paravirt ABI, if an async page fault occurs on an early
> + * memory access in the normal (sync) #PF path or in an NMI
> + * that happens early in the #PF code, the combination of CR2
> + * and the APF reason field will be corrupted.

I don't think this can happen. In both cases IF == 0 and that async
(think host side) page fault will be completely handled on the
host. There is no injection happening in such a case ever. If it does,
then yes the host side implementation is buggered, but AFAICT this is
not the case.

See also my reply in the other thread:

https://lore.kernel.org/r/[email protected]

Thanks,

tglx

2020-03-07 14:37:41

by Andy Lutomirski

Subject: Re: [PATCH] x86/kvm: Disable KVM_ASYNC_PF_SEND_ALWAYS

> On Mar 7, 2020, at 2:09 AM, Thomas Gleixner <[email protected]> wrote:
>
> Andy Lutomirski <[email protected]> writes:
>
>> The ABI is broken and we cannot support it properly. Turn it off.
>>
>> If this causes a meaningful performance regression for someone, KVM
>> can introduce an improved ABI that is supportable.
>>
>> Cc: [email protected]
>> Signed-off-by: Andy Lutomirski <[email protected]>
>> ---
>> arch/x86/kernel/kvm.c | 11 ++++++++---
>> 1 file changed, 8 insertions(+), 3 deletions(-)
>>
>> diff --git a/arch/x86/kernel/kvm.c b/arch/x86/kernel/kvm.c
>> index 93ab0cbd304e..71f9f39f93da 100644
>> --- a/arch/x86/kernel/kvm.c
>> +++ b/arch/x86/kernel/kvm.c
>> @@ -318,11 +318,16 @@ static void kvm_guest_cpu_init(void)
>>
>> pa = slow_virt_to_phys(this_cpu_ptr(&apf_reason));
>>
>> -#ifdef CONFIG_PREEMPTION
>> - pa |= KVM_ASYNC_PF_SEND_ALWAYS;
>> -#endif
>> pa |= KVM_ASYNC_PF_ENABLED;
>>
>> + /*
>> + * We do not set KVM_ASYNC_PF_SEND_ALWAYS. With the current
>> + * KVM paravirt ABI, if an async page fault occurs on an early
>> + * memory access in the normal (sync) #PF path or in an NMI
>> + * that happens early in the #PF code, the combination of CR2
>> + * and the APF reason field will be corrupted.
>
> I don't think this can happen. In both cases IF == 0 and that async
> (think host side) page fault will be completely handled on the
> host. There is no injection happening in such a case ever. If it does,
> then yes the host side implementation is buggered, but AFAICT this is
> not the case.

Indeed. But read v2 please.

>
> See also my reply in the other thread:
>
> https://lore.kernel.org/r/[email protected]
>
> Thanks,
>
> tglx