Date:   Tue, 1 Dec 2020 09:25:29 +0800
From:   "Xu, Like" <like.xu@...el.com>
To:     Peter Zijlstra <peterz@...radead.org>
Cc:     Like Xu <like.xu@...ux.intel.com>,
        "Kleen, Andi" <andi.kleen@...el.com>,
        Paolo Bonzini <pbonzini@...hat.com>, kvm@...r.kernel.org,
        Sean Christopherson <sean.j.christopherson@...el.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>,
        Joerg Roedel <joro@...tes.org>,
        Kan Liang <kan.liang@...ux.intel.com>, luwei.kang@...el.com,
        Thomas Gleixner <tglx@...utronix.de>, wei.w.wang@...el.com,
        Tony Luck <tony.luck@...el.com>,
        Stephane Eranian <eranian@...gle.com>,
        Mark Gross <mgross@...ux.intel.com>,
        Srinivas Pandruvada <srinivas.pandruvada@...ux.intel.com>,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v2 04/17] perf: x86/ds: Handle guest PEBS overflow PMI and
 inject it to guest

Hi Peter,

On 2020/11/30 18:49, Peter Zijlstra wrote:
> On Fri, Nov 27, 2020 at 10:14:49AM +0800, Xu, Like wrote:
>
>>> OK, but the code here wanted to inspect the guest DS from the host. It
>>> states this is somehow complicated/expensive. But surely we can at the
>>> very least map the first guest DS page somewhere so we can at least
>>> access the control bits without too much magic.
>> We note that the SDM has a contiguous present memory mapping
>> assumption about the DS save area and the PEBS buffer area.
>>
>> Therefore, we revisit your suggestion here and move it a bit forward:
>>
>> When the PEBS is enabled, KVM will cache the following values:
>> - gva ds_area (kvm msr trap)
>> - hva1 for "gva ds_area" (walk guest page table)
>> - hva2 for "gva pebs_buffer_base" via hva1 (walk guest page table)
> What this [gh]va? Guest/Host Virtual Address? I think you're assuming I
> know about all this virt crap,.. I don't.
Oh, my bad and let me add it:

gva: guest virtual address
gpa: guest physical address
gfn: guest frame number
hva: host virtual adderss
hpa: host physical address

In the KVM, we get hva from gva in the following way:

gpa = kvm_mmu_gva_to_gpa_system(vcpu, gva, NULL);
gfn = gpa >> PAGE_SHIFT;
slot = gfn_to_memslot(kvm, gfn);
hva = gfn_to_hva_memslot_prot(slot, gfn, NULL);

>
>> if the "gva ds_area" cache hits,
> what?
Sorry, it looks a misuse of terminology.

I mean KVM will save the last used "gva ds_area" value and its hva in the 
extra fields,
if the "gva ds_area" does not change this time, we will not walk the guest 
page table
to get its hva again.

I think it's the main point in your suggestion, and I try to elaborate it.
>> - access PEBS "interrupt threshold" and "Counter Reset[]" via hva1
>> - get "gva2 pebs_buffer_base" via __copy_from_user(hva1)
> But you already had hva2, so what's the point?
hva1 is for for "gva ds_area"
hva2 is for "gva pebs_buffer_base"

The point is before using the last save hva2, we need to
make sure that "gva pebs_buffer_base" is not changed to avoid
that some malicious drivers may change it without changing ds_area.

>
>> if the "gva2 pebs_buffer_base" cache hits,
> What?
>
>> - we get "gva2 pebs_index" via __copy_from_user(hva2),
> pebs_index is in ds_are, which would be hva1
Yes, we get "gva2 pebs_index" via __copy_from_user(hva1).
>
>> - rewrite the guest PEBS records via hva2 and pebs_index
>>
>> If any cache misses, setup the cache values via walking tables again.
>>
>> I wonder if you would agree with this optimization idea,
>> we look forward to your confirmation for the next step.
> I'm utterly confused. I really can't follow.
Generally, KVM will save hva1 (gva1 ds_area) and hva2 (for gva2 
pebs_buffer_base)
in the first round of the guest page table walking and reuse them
if they're not changed in subsequent use.

I think this approach is feasible, and please complain if you are still 
confused or disagree.

Thanks,
Like Xu

Powered by blists - more mailing lists