Message-ID: <870aaaf3-7a52-f91a-c5f3-fd3c7276a5d9@redhat.com>
Date:   Tue, 15 Oct 2019 21:46:58 +0200
From:   Paolo Bonzini <pbonzini@...hat.com>
To:     Andrea Arcangeli <aarcange@...hat.com>
Cc:     kvm@...r.kernel.org, linux-kernel@...r.kernel.org,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Sean Christopherson <sean.j.christopherson@...el.com>
Subject: Re: [PATCH 12/14] KVM: retpolines: x86: eliminate retpoline from
 vmx.c exit handlers

On 15/10/19 18:49, Andrea Arcangeli wrote:
> On Tue, Oct 15, 2019 at 10:28:39AM +0200, Paolo Bonzini wrote:
>> If you're including EXIT_REASON_EPT_MISCONFIG (MMIO access) then you
>> should include EXIT_REASON_IO_INSTRUCTION too.  Depending on the devices
>> that are in the guest, the doorbell register might be MMIO or PIO.
> 
> The fact outb/inb devices exists isn't the question here. The question
> you should clarify is: which of the PIO devices is performance
> critical as much as MMIO with virtio/vhost?

virtio 0.9 uses PIO.

> I mean even on real hardware those devices aren't performance critical.

On virtual machines they're actually faster than MMIO because they don't
need to go through page table walks.

>> So, the difference between my suggested list (which I admit is just
>> based on conjecture, not benchmarking) is that you add
>> EXIT_REASON_PAUSE_INSTRUCTION, EXIT_REASON_PENDING_INTERRUPT,
>> EXIT_REASON_EXTERNAL_INTERRUPT, EXIT_REASON_HLT, EXIT_REASON_MSR_READ,
>> EXIT_REASON_CPUID.
>>
>> Which of these make a difference for the hrtimer testcase?  It's of
>> course totally fine to use benchmarks to prove that my intuition was
>> bad---but you must also use them to show why your intuition is right. :)
> 
> The hrtimer flood hits on this:
> 
>            MSR_WRITE     338793    56.54%     5.51%      0.33us     34.44us      0.44us ( +-   0.20% )
>    PENDING_INTERRUPT     168431    28.11%     2.52%      0.36us     32.06us      0.40us ( +-   0.28% )
>     PREEMPTION_TIMER      91723    15.31%     1.32%      0.34us     30.51us      0.39us ( +-   0.41% )
>   EXTERNAL_INTERRUPT        234     0.04%     0.00%      0.25us      5.53us      0.43us ( +-   5.67% )
>                  HLT         65     0.01%    90.64%      0.49us 319933.79us  37562.71us ( +-  21.68% )
>             MSR_READ          6     0.00%     0.00%      0.67us      1.96us      1.06us ( +-  17.97% )
>        EPT_MISCONFIG          6     0.00%     0.01%      3.09us    105.50us     26.76us ( +-  62.10% )
> 
> PENDING_INTERRUPT is the big missing thing in your list. It probably
> accounts for the bulk of slowdown from your list.

Makes sense.

> However I could imagine other loads with higher external
> interrupt/hlt/rdmsr than the hrtimer one so I didn't drop those.
External interrupts should only tick at 1 Hz on nohz_full kernels,
and even at 1000 Hz (if physical CPUs are not isolated) it should not
really matter.  We can include it since it has such a short handler so
the cost of retpolines is in % much more than other exits.

HLT is certainly a slow path, the guest only invokes if things such as
NAPI interrupt mitigation have failed.  As long as the guest stays in
halted state for a microsecond or so, the cost of retpoline will all but
disappear.

RDMSR again shouldn't be there, guests sometimes read the PMTimer (which
is an I/O port) or TSC but for example do not really ever read the APIC
TMCCT.

> I'm pretty sure HLT/EXTERNAL_INTERRUPT/PENDING_INTERRUPT should be
> included.
> I also wonder if VMCALL should be added, certain loads hit on fairly
> frequent VMCALL, but none of the one I benchmarked.

I agree for external interrupt and pending interrupt, and VMCALL is fine
too.  In addition I'd add I/O instructions which are useful for some
guests and also for benchmarking (e.g. vmexit.flat has both IN and OUT
tests).

Paolo

Powered by blists - more mailing lists