Hi, Marc
On 2019/4/4 18:59, Marc Zyngier wrote:
> On Thu, 04 Apr 2019 11:07:59 +0100,
> "Tangnianyao (ICT)" <tangnianyao@xxxxxxxxxx> wrote:
>>
>>
>>
>> On 2019/3/30 17:43, Marc Zyngier wrote:
>>
>> Hi, Marc
>>
>>> On Sat, 30 Mar 2019 08:42:58 +0000,
>>> "Tangnianyao (ICT)" <tangnianyao@xxxxxxxxxx> wrote:
>>>>
>>>> Hi, Marc
>>>>
>>>> On 2019/3/21 1:02, Marc Zyngier Wrote:
>>>>> On Tue, 19 Mar 2019 21:25:47 +0800
>>>>> "Tangnianyao (ICT)" <tangnianyao@xxxxxxxxxx> wrote:
>>>>>
>>>>>> Hi, all
>>>>>>
>>>>>> Using gicv4, when guest is waiting for irq, it sends wfi and traps to kvm.
>>>>>> When vlpi is forwarded to PE after its_vpe_deschedule, before halt_poll in
>>>>>> kvm_vcpu_block, halt_poll may increase latency for this vlpi getting to guest.
>>>>>> In halt_poll process, it checks if there's pending irq for vcpu using pending_last.
>>>>>> However, doorbell is not enable at this moment and vlpi or doorbell can not set
>>>>>> pending_last true, to stop halt_poll. It will run until halt_poll time ends, if
>>>>>> there's no other physical irq coming in the meantime. And then vcpu is scheduled out.
>>>>>> This pending vlpi has to wait for vcpu getting schedule in next time.
>>>>>>
>>>>>> Should we enable doorbell before halt_poll process ?
>>>>>
>>>>> Enabling doorbells can be quite expensive. Depending on the HW, this is
>>>>> either:
>>>>>
>>>>> - a write to memory (+DSB, potential cache maintenance), a write to the
>>>>> INVLPI register, and a poll of the SYNC register
>>>>> - a write to memory (+DSB, potential cache maintenance), potentially
>>>>> a string of DISCARD+SYNC+MAPI+SYNC commands, and an INV+SYNC command
>>>>>
>>>> I have tested average cost of kvm_vgic_v4_enable_doorbell in our machine.
>>>> When gic_rdists->has_direct_lpi is 1, it costs 0.35 us.
>>>> When gic_rdists->has_direct_lpi is 0, it costs 1.4 us.
>>>
>>> This looks pretty low. Which HW is that on? How about on something
>>> like D05?
>>
>> I tested it on D06.
>> D05 doesn't not support gicv4 and I haven't tested on D05.
>
> I'm afraid you're mistaken. D05 does have GICv4. I have a pretty good
> idea it does because that's the system I used to write the SW
> support. So please dig one out of the cupboard and test on it. I'm
> pretty sure you're at the right place to do so.
I've learned that there's some implementation problem for the PCIe controller
of Hi1616, the processor of D05. The PCIe ACS was not implemented properly
and D05 doesn't support VM using pcie vf.
>
>>
>>>
>>>> Compared to default halt_poll_ns, 500000ns, enabling doorbells is not
>>>> large at all.
>>>
>>> Sure. But I'm not sure this is a universal figure.
>>>
>>>>
>>>>> Frankly, you want to be careful with that. I'd rather enable them late
>>>>> and have a chance of not blocking because of another (virtual)
>>>>> interrupt, which saves us the doorbell business.
>>>>>
>>>>> I wonder if you wouldn't be in a better position by drastically
>>>>> reducing halt_poll_ns for vcpu that can have directly injected
>>>>> interrupts.
>>>>>
>>>>
>>>> If we set halt_poll_ns to a small value, the chance of
>>>> not blocking and vcpu scheduled out becomes larger. The cost
>>>> of vcpu scheduled out is quite expensive.
>>>> In many cases, one pcpu is assigned to only
>>>> one vcpu, and halt_poll_ns is set quite large, to increase
>>>> chance of halt_poll process got terminated. This is the
>>>> reason we want to set halt_poll_ns large. And We hope vlpi
>>>> stop halt_poll process in time.
>>>
>>> Fair enough. It is certainly realistic to strictly partition the
>>> system when GICv4 is in use, so I can see some potential benefit.
>>>
>>>> Though it will check whether vcpu is runnable again by
>>>> kvm_vcpu_check_block. Vlpi can prevent scheduling vcpu out.
>>>> However it's somewhat later if halt_poll_ns is larger.
>>>>
>>>>> In any case, this is something that we should measure, not guess.
>>>
>>> Please post results of realistic benchmarks that we can reproduce,
>>> with and without this change. I'm willing to entertain the idea, but I
>>> need more than just a vague number.
>>>
>>> Thanks,
>>>
>>> M.
>>>
>>
>> I tested it with and without change (patch attached in the last).
>> halt_poll_ns is keep default, 500000ns.
>> I have merged the patch "arm64: KVM: Always set ICH_HCR_EL2.EN if GICv4 is enabled"
>> to my test kernel, to make sure ICH_HCR_EL2.EN=1 in guest.
>>
>> netperf result:
>> D06 as server, intel 8180 server as client
>> with change:
>> package 512 bytes - 5400 Mbits/s
>> package 64 bytes - 740 Mbits/s
>> without change:
>> package 512 bytes - 5000 Mbits/s
>> package 64 bytes - 710 Mbits/s
>>
>> Also I have tested D06 as client, intel machine as server,
>> with the change, the result remains the same.
>
> So anywhere between 4% and 8% improvement. Please post results for D05
> once you've found one.
>
>>
>>
>> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
>> index 55fe8e2..0f56904 100644
>> --- a/virt/kvm/kvm_main.c
>> +++ b/virt/kvm/kvm_main.c
>> @@ -2256,6 +2256,16 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
>> if (vcpu->halt_poll_ns) {
>> ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
>>
>> +#ifdef CONFIG_ARM64
>> + /*
>> + * When using gicv4, enable doorbell before halt pool wait
>> + * so that direct vlpi can stop halt poll.
>> + */
>> + if (vcpu->arch.vgic_cpu.vgic_v3.its_vpe.its_vm) {
>> + kvm_vgic_v4_enable_doorbell(vcpu);
>> + }
>> +#endif
>> +
>
> Irk. No. You're now leaving doorbells enabled when going back to the
> guest, and that's pretty bad as the whole logic relies on doorbells
> being disabled if the guest can run.
>
> So please try this instead on top of mainline. And it has to be on top
> of mainline as we've changed the way timer interrupts work in 5.1 --
> see accb99bcd0ca ("KVM: arm/arm64: Simplify bg_timer programming").
>
> diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
> index f25aa98a94df..0ae4ad5dcb12 100644
> --- a/virt/kvm/kvm_main.c
> +++ b/virt/kvm/kvm_main.c
> @@ -2252,6 +2252,8 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
> bool waited = false;
> u64 block_ns;
>
> + kvm_arch_vcpu_blocking(vcpu);
> +
> start = cur = ktime_get();
> if (vcpu->halt_poll_ns) {
> ktime_t stop = ktime_add_ns(ktime_get(), vcpu->halt_poll_ns);
> @@ -2272,8 +2274,6 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
> } while (single_task_running() && ktime_before(cur, stop));
> }
>
> - kvm_arch_vcpu_blocking(vcpu);
> -
> for (;;) {
> prepare_to_swait_exclusive(&vcpu->wq, &wait, TASK_INTERRUPTIBLE);
>
> @@ -2287,8 +2287,8 @@ void kvm_vcpu_block(struct kvm_vcpu *vcpu)
> finish_swait(&vcpu->wq, &wait);
> cur = ktime_get();
>
> - kvm_arch_vcpu_unblocking(vcpu);
> out:
> + kvm_arch_vcpu_unblocking(vcpu);
> block_ns = ktime_to_ns(cur) - ktime_to_ns(start);
>
> if (!vcpu_valid_wakeup(vcpu))
>
> Thanks,
>
> M.
>
I've tested your patch and the results showed as follow:
netperf result:
D06 as server, intel 8180 server as client
with change:
package 512 bytes - 5500 Mbits/s
package 64 bytes - 760 Mbits/s
without change:
package 512 bytes - 5000 Mbits/s
package 64 bytes - 710 Mbits/s
Thanks,
Tang
