Date:   Mon, 22 Nov 2021 19:13:02 +0000
From:   Sean Christopherson <seanjc@google.com>
To:     Aili Yao <yaoaili126@gmail.com>
Cc:     pbonzini@redhat.com, vkuznets@redhat.com, wanpengli@tencent.com,
        jmattson@google.com, joro@8bytes.org, tglx@linutronix.de,
        mingo@redhat.com, bp@alien8.de, dave.hansen@linux.intel.com,
        x86@kernel.org, hpa@zytor.com, kvm@vger.kernel.org,
        linux-kernel@vger.kernel.org, yaoaili@kingsoft.com
Subject: Re: [PATCH] KVM: LAPIC: Per vCPU control over
 kvm_can_post_timer_interrupt
Message-ID: <YZvrvmRnuDc1e+gi@google.com>
References: <20211122095619.000060d2@gmail.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20211122095619.000060d2@gmail.com>
Precedence: bulk

On Mon, Nov 22, 2021, Aili Yao wrote:
> From: Aili Yao <yaoaili@kingsoft.com>
> 
> When we isolate some pyhiscal cores, We may not use them for kvm guests,
> We may use them for other purposes like DPDK, or we can make some kvm
> guests isolated and some not, the global judgement pi_inject_timer is
> not enough; We may make wrong decisions:
> 
> In such a scenario, the guests without isolated cores will not be
> permitted to use vmx preemption timer, and tscdeadline fastpath also be
> disabled, both will lead to performance penalty.
> 
> So check whether the vcpu->cpu is isolated, if not, don't post timer
> interrupt.
> 
> Signed-off-by: Aili Yao <yaoaili@kingsoft.com>
> ---
>  arch/x86/kvm/lapic.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kvm/lapic.c b/arch/x86/kvm/lapic.c
> index 759952dd1222..72dde5532101 100644
> --- a/arch/x86/kvm/lapic.c
> +++ b/arch/x86/kvm/lapic.c
> @@ -34,6 +34,7 @@
>  #include <asm/delay.h>
>  #include <linux/atomic.h>
>  #include <linux/jump_label.h>
> +#include <linux/sched/isolation.h>
>  #include "kvm_cache_regs.h"
>  #include "irq.h"
>  #include "ioapic.h"
> @@ -113,7 +114,8 @@ static inline u32 kvm_x2apic_id(struct kvm_lapic *apic)
>  
>  static bool kvm_can_post_timer_interrupt(struct kvm_vcpu *vcpu)
>  {
> -	return pi_inject_timer && kvm_vcpu_apicv_active(vcpu);
> +	return pi_inject_timer && kvm_vcpu_apicv_active(vcpu) &&
> +		!housekeeping_cpu(vcpu->cpu, HK_FLAG_TIMER);

I don't think this is safe, vcpu->cpu will be -1 if the vCPU isn't scheduled in.
This also doesn't play nice with the admin forcing pi_inject_timer=1.  Not saying
there's a reasonable use case for doing that, but it's supported today and this
would break that behavior.  It would also lead to weird behavior if a vCPU were
migrated on/off a housekeeping vCPU.  Again, probably not a reasonable use case,
but I don't see anything that would outright prevent that behavior.

The existing behavior also feels a bit unsafe as pi_inject_timer is writable while
KVM is running, though I supposed that's orthogonal to this discussion.

Rather than check vcpu->cpu, is there an existing vCPU flag that can be queried,
e.g. KVM_HINTS_REALTIME?

>  }
>  
>  bool kvm_can_use_hv_timer(struct kvm_vcpu *vcpu)
> -- 
> 2.25.1
>