Re: [PATCH 3/3] KVM: x86: correct mwait and monitor emulation

[Nadav Amit <nadav.amit@gmail.com>]

On 6/19/14, 3:07 PM, Gleb Natapov wrote:
> On Thu, Jun 19, 2014 at 02:52:20PM +0300, Nadav Amit wrote:
>> On 6/19/14, 2:23 PM, Gleb Natapov wrote:
>>> On Thu, Jun 19, 2014 at 01:53:36PM +0300, Nadav Amit wrote:
>>>>
>>>> On Jun 19, 2014, at 1:18 PM, Michael S. Tsirkin <mst@redhat.com> wrote:
>>>>
>>>>> On Wed, Jun 18, 2014 at 02:46:01PM -0400, Gabriel L. Somlo wrote:
>>>>>> On Wed, Jun 18, 2014 at 10:59:14AM -0700, Eric Northup wrote:
>>>>>>> On Wed, Jun 18, 2014 at 7:19 AM, Nadav Amit <namit@cs.technion.ac.il> wrote:
>>>>>>>> mwait and monitor are currently handled as nop. Considering this behavior, they
>>>>>>>> should still be handled correctly, i.e., check execution conditions and generate
>>>>>>>> exceptions when required. mwait and monitor may also be executed in real-mode
>>>>>>>> and are not handled in that case.  This patch performs the emulation of
>>>>>>>> monitor-mwait according to Intel SDM (other than checking whether interrupt can
>>>>>>>> be used as a break event).
>>>>>>>>
>>>>>>>> Signed-off-by: Nadav Amit <namit@cs.technion.ac.il>
>>>>>>
>>>>>> How about this instead (details in the commit log below) ? Please let
>>>>>> me know what you think, and if you'd prefer me to send it out as a
>>>>>> separate patch rather than a reply to this thread.
>>>>>>
>>>>>> Thanks,
>>>>>> --Gabriel
>>>>>
>>>>> If there's an easy workaround, I'm inclined to agree.
>>>>> We can always go back to Gabriel's patch (and then we'll need
>>>>> Nadav's one too) but if we release a kernel with this
>>>>> support it becomes an ABI and we can't go back.
>>>>>
>>>>> So let's be careful here, and revert the hack for 3.16.
>>>>>
>>>>>
>>>>> Acked-by: Michael S. Tsirkin <mst@redhat.com>
>>>>>
>>>> Personally, I got a custom guest which requires mwait for executing correctly.
>>> Can you elaborate on this guest a little bit. With nop implementation
>>> for mwait the guest will hog a host cpu. Do you consider this to be
>>> "executing correctly?"
>>>
>>> --
>>
>> mwait is not as "clean" as it may appear. It encounters false wake-ups due
>> to a variety of reasons, and any code need to recheck the wake-up condition
>> afterwards. Actually, some CPUs had bugs that caused excessive wake-ups that
>> degraded performance considerably (Nehalem, if I am not mistaken).
>> Therefore, handling mwait as nop is logically correct (although it may
>> degrade performance).
>>
>> For the reference, if you look at the SDM 8.10.4, you'll see:
>> "Multiple events other than a write to the triggering address range can
>> cause a processor that executed MWAIT to wake up. These include events that
>> would lead to voluntary or involuntary context switches, such as..."
>>
>> Note the words "include" in the sentence "These include events". Software
>> has no way of controlling whether it gets false wake-ups and cannot rely on
>> the wake-up as indication to anything.
>>
> That's all well and good and I didn't say that nop is not a valid
> mwait implementation, it is, though there is a big difference between
> "encounters false wake-ups" and never sleeps.  What I asked is do you
> consider your guest hogging host cpu to be "executing correctly?". What
> this guest is doing that such behaviour is tolerated and shouldn't it
> be better to just poll for a condition you are waiting for instead of
> executing expensive vmexits. This will also hog 100% host cpu, but will
> be actually faster.
>
You are correct, but unfortunately I have no control over the guest 
workload. In this specific workload I do not care about performance but 
only about correctness.

Nadav