Received-SPF: pass (google.com: best guess record for domain of linux-kernel-owner@vger.kernel.org designates 209.132.180.67 as permitted sender) client-ip=209.132.180.67;
From:   Vitaly Kuznetsov <vkuznets@redhat.com>
To:     Wanpeng Li <kernellwp@gmail.com>
Cc:     LKML <linux-kernel@vger.kernel.org>, kvm <kvm@vger.kernel.org>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Sean Christopherson <sean.j.christopherson@intel.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>,
        Thadeu Lima de Souza Cascardo <cascardo@canonical.com>
Subject: Re: [PATCH] KVM: Pre-allocate 1 cpumask variable per cpu for both pv tlb and pv ipis
In-Reply-To: <CANRm+CzkN9oYf4UqWYp2SHFii02=pvVLbW4oNkLmPan7ZroDZA@mail.gmail.com>
References: <CANRm+CwwYoSLeA3Squp-_fVZpmYmxEfqOB+DGoQN4Y_iMT347w@mail.gmail.com> <878slio6hp.fsf@vitty.brq.redhat.com> <CANRm+CzkN9oYf4UqWYp2SHFii02=pvVLbW4oNkLmPan7ZroDZA@mail.gmail.com>
Date:   Tue, 04 Feb 2020 15:36:53 +0100
Message-ID: <874kw6o1ve.fsf@vitty.brq.redhat.com>
MIME-Version: 1.0
Content-Type: text/plain
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

Wanpeng Li <kernellwp@gmail.com> writes:

>>
>> Honestly, I'd simplify the check in kvm_alloc_cpumask() as
>>
>> if (!kvm_para_available())
>>         return;
>>
>> and allocated masks for all other cases.
>
> This will waste the memory if pv tlb and pv ipis are not exposed which
> are the only users currently.
>

My assumption is that the number of cases where we a) expose KVM b)
don't expose IPIs and PV-TLB and c) care about 1 cpumask per cpu is
relatively low. Ok, let's at least have a function for

	if (kvm_para_has_feature(KVM_FEATURE_PV_TLB_FLUSH) &&
	    !kvm_para_has_hint(KVM_HINTS_REALTIME) &&
	    kvm_para_has_feature(KVM_FEATURE_STEAL_TIME))

as we now check it twice: in kvm_alloc_cpumask() and kvm_guest_init(),
something like pv_tlb_flush_supported(). We can also do
pv_ipi_supported() and probably others for consistency.

Also, probably not for this patch but it all makes me wonder why there's
no per-cpu 'scratch' cpumask for the whole kernel to use. We definitely
need it for hypervisor support but I also see
arch/x86/kernel/apic/x2apic_cluster.c has similar needs.

-- 
Vitaly