lkml.org 
[lkml]   [2018]   [Oct]   [31]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
SubjectRe: [PATCH v13 08/12] KVM: x86: Add Intel PT context switch for each vcpu
From
Date
On 30/10/2018 11:00, Thomas Gleixner wrote:
> On Mon, 29 Oct 2018, Paolo Bonzini wrote:
>> On 24/10/2018 12:13, Alexander Shishkin wrote:
>>> Luwei Kang <luwei.kang@intel.com> writes:
>>>> + /*
>>>> + * Set guest state of MSR_IA32_RTIT_CTL MSR (PT will be disabled
>>>> + * on VM entry when it has been disabled in guest before).
>>>> + */
>>>> + vmcs_write64(GUEST_IA32_RTIT_CTL, vmx->pt_desc.guest.ctl);
>>>> +
>>>> + if (vmx->pt_desc.guest.ctl & RTIT_CTL_TRACEEN) {
>>>> + wrmsrl(MSR_IA32_RTIT_CTL, 0);
>>>> + pt_save_msr(&vmx->pt_desc.host, vmx->pt_desc.addr_range);
>>>> + pt_load_msr(&vmx->pt_desc.guest, vmx->pt_desc.addr_range);
>>>> + }
>>>> +}
>>>
>>> From my side this is still a NAK, because [1].
>>>
>>> [1] https://marc.info/?l=kvm&m=153847567226248&w=2
>>
>> Then you should have replied to
>> https://marc.info/?l=kvm&m=153865386015249&w=2 instead of having Luwei
>> do the work for nothing.
>>
>> Quoting from there:
>>
>>>> One shouldn't have to enable or disable anything in KVM to stop it from
>>>> breaking one's existing workflow. That makes no sense.
>>>
>>> If you "have to enable or disable anything" it means you have to
>>> override the default. But the default in this patches is "no change
>>> compared to before the patches", leaving tracing of both host and guest
>>> entirely to the host, so I don't understand your remark. What workflow
>>> is broken?
>>>
>>>> There already are controls in perf that enable/disable guest tracing.
>>>
>>> You are confusing "tracing guest from the host" and "the guest can trace
>>> itself". This patchset is adding support for the latter, and that
>>> affects directly whether the tracing CPUID leaf can be added to the
>>> guest. Therefore it's not perf that can decide whether to turn it on;
>>> KVM must know it when /dev/kvm is opened, which is why it is a module
>>> parameter.
>>
>> I'd be happier if we found an agreement, but without discussion that
>> just won't happen.
>
> So at least we need a way for perf on the host to programmatically detect,
> that 'guest traces itself' is enabled, so it can inject that information
> into the host data and post processing can tell that. W/o something like
> that it's going to be a FAQ.

In guest-tracing mode there will be already a TIP.PGD and TIP.PGE packet
respectively before vmentry and after vmexit, caused by the RTIT_CTL
WRMSRs in pt_guest_enter and pt_guest_exit. The target IP of the
packets will come from kvm-intel.ko.

In system mode instead you get a Paging Information Packet on
vmentry/vmexit, with bit 0 set in the third byte. You won't get it if
guest-side tracing is on (because tracing has been disabled by
pt_guest_enter and won't be re-enabled until pt_guest_exit). I don't
think it's correct to "fake" the PIP in guest-tracing mode, because
TIP.PGD should be followed immediately by TIP.PGE.

Is this okay for perf users?

Paolo

\
 
 \ /
  Last update: 2018-10-31 11:44    [W:0.163 / U:0.552 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site