Re: [Qemu-devel] [PATCH, RFC] trace: implement guest tracepoint passthrough

[Blue Swirl <blauwirbel@gmail.com>]

On Sat, Sep 3, 2011 at 9:26 AM, Dhaval Giani <dhaval.giani@gmail.com> wrote:
> On Sat, Sep 3, 2011 at 1:53 AM, Blue Swirl <blauwirbel@gmail.com> wrote:
>> On Wed, Aug 31, 2011 at 6:00 PM, Dhaval Giani <dhaval.giani@gmail.com> wrote:
>>> On Wed, Aug 31, 2011 at 10:58 AM, Blue Swirl <blauwirbel@gmail.com> wrote:
>>>> On Wed, Aug 31, 2011 at 8:38 AM, Avi Kivity <avi@redhat.com> wrote:
>>>>> On 08/26/2011 10:06 PM, Blue Swirl wrote:
>>>>>>
>>>>>> Let guests inject tracepoint data via fw_cfg device.
>>>>>>
>>>>>>
>>>>>
>>>>> At least on x86, fw_cfg is pretty slow, involving multiple exits.  IMO, for
>>>>> kvm, even one exit per tracepoint is too high.  We need to use a shared
>>>>> memory transport with a way to order guest/host events later on (by using a
>>>>> clock).
>>>>
>>>> This could be an easy way, if the guest always had access to an
>>>> accurate clock, but that may not be the case.
>>>>
>>>
>>> From what I understand, kvmclock should be good enoguh for this
>>> purpose. That is what I am using.
>>
>> It's only available for KVM on x86, that is most certainly not enough.
>>
>
> From what I understood it is available on other architectures as well,
> but let us just confirm with glommer.

This line in Makefile.target limits the device to KVM on x86:
obj-i386-$(CONFIG_KVM) += kvmclock.o

Moreover, the code assumes that CPUState structure contains a field
cpuid_kvm_features but that is only available on x86.

Perhaps the device can be made generic but I don't see how it should
work with TCG. Actually I don't understand it at all, I think key
functionality must be inside KVM module.