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;
MIME-Version: 1.0
References: <f51251cc-885e-2f7a-b18d-faa76db15b87@redhat.com> <20200410174703.1138-1-sean.j.christopherson@intel.com>
In-Reply-To: <20200410174703.1138-1-sean.j.christopherson@intel.com>
From:   Wanpeng Li <kernellwp@gmail.com>
Date:   Sat, 11 Apr 2020 08:34:22 +0800
Message-ID: <CANRm+CxGsBTeVy6-2-DhMYhB3J64jstaDgotEem01bpeiCY-7g@mail.gmail.com>
Subject: Re: [PATCH v2] KVM: X86: Ultra fast single target IPI fastpath
To:     Sean Christopherson <sean.j.christopherson@intel.com>
Cc:     Paolo Bonzini <pbonzini@redhat.com>,
        Vitaly Kuznetsov <vkuznets@redhat.com>,
        Wanpeng Li <wanpengli@tencent.com>,
        Jim Mattson <jmattson@google.com>,
        Joerg Roedel <joro@8bytes.org>, kvm <kvm@vger.kernel.org>,
        LKML <linux-kernel@vger.kernel.org>,
        Haiwei Li <lihaiwei@tencent.com>
Content-Type: text/plain; charset="UTF-8"
Sender: linux-kernel-owner@vger.kernel.org
Precedence: bulk

On Sat, 11 Apr 2020 at 01:47, Sean Christopherson
<sean.j.christopherson@intel.com> wrote:
>
> On Fri, Apr 10, 2020 at 05:50:35PM +0200, Paolo Bonzini wrote:
> > On 10/04/20 17:35, Sean Christopherson wrote:
> > > IMO, this should come at the very end of vmx_vcpu_run().  At a minimum, it
> > > needs to be moved below the #MC handling and below
> > >
> > >     if (vmx->fail || (vmx->exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY))
> > >             return;
> >
> > Why?  It cannot run in any of those cases, since the vmx->exit_reason
> > won't match.
>
> #MC and consistency checks should have "priority" over everything else.
> That there isn't actually a conflict is irrelevant IMO.  And it's something
> that will likely confuse newbies (to VMX and/or KVM) as it won't be obvious
> that the motivation was to shave a few cycles, e.g. versus some corner case
> where the fastpath handling does something meaningful even on failure.
>
> > > KVM more or less assumes vmx->idt_vectoring_info is always valid, and it's
> > > not obvious that a generic fastpath call can safely run before
> > > vmx_complete_interrupts(), e.g. the kvm_clear_interrupt_queue() call.
> >
> > Not KVM, rather vmx.c.  You're right about a generic fastpath, but in
> > this case kvm_irq_delivery_to_apic_fast is not touching VMX state; even
> > if you have a self-IPI, the modification of vCPU state is only scheduled
> > here and will happen later via either kvm_x86_ops.sync_pir_to_irr or
> > KVM_REQ_EVENT.
>
> I think what I don't like is that the fast-IPI code is buried in a helper
> that masquerades as a generic fastpath handler.  If that's open-coded in
> vmx_vcpu_run(), I'm ok with doing the fast-IPI handler immediately after
> the failure checks.
>
> And fast-IPI aside, the code could use a bit of optimization to prioritize
> successful VM-Enter, which would slot in nicely as a prep patch.  Patches
> (should be) following.

Thanks for v3. :)

>
> IMO, this is more logically correct:
>
>         vmx->exit_reason = vmcs_read32(VM_EXIT_REASON);
>         if (unlikely((u16)vmx->exit_reason == EXIT_REASON_MCE_DURING_VMENTRY))
>                 kvm_machine_check();
>
>         if (unlikely(vmx->exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY))
>                 return EXIT_FASTPATH_NONE;
>
>         if (!is_guest_mode(vcpu) && vmx->exit_reason == EXIT_REASON_MSR_WRITE)
>                 exit_fastpath = handle_fastpath_set_msr_irqoff(vcpu);
>         else
>                 exit_fastpath = EXIT_FASTPATH_NONE;
>
> And on my system, the compiler hoists fast-IPI above the #MC, e.g. moving
> the fast-IPI down only adds a single macrofused uop, testb+jne for
> FAILED_VMENTERY, to the code path.
>
>    0xffffffff81067d1d <+701>:   vmread %rax,%rax
>    0xffffffff81067d20 <+704>:   ja,pt  0xffffffff81067d2d <vmx_vcpu_run+717>
>    0xffffffff81067d23 <+707>:   pushq  $0x0
>    0xffffffff81067d25 <+709>:   push   %rax
>    0xffffffff81067d26 <+710>:   callq  0xffffffff81071790 <vmread_error_trampoline>
>    0xffffffff81067d2b <+715>:   pop    %rax
>    0xffffffff81067d2c <+716>:   pop    %rax
>    0xffffffff81067d2d <+717>:   test   %eax,%eax
>    0xffffffff81067d2f <+719>:   mov    %eax,0x32b0(%rbp)
>    0xffffffff81067d35 <+725>:   js     0xffffffff81067d5a <vmx_vcpu_run+762>
>    0xffffffff81067d37 <+727>:   testb  $0x20,0x2dc(%rbp)
>    0xffffffff81067d3e <+734>:   jne    0xffffffff81067d49 <vmx_vcpu_run+745>
>    0xffffffff81067d40 <+736>:   cmp    $0x20,%eax
>    0xffffffff81067d43 <+739>:   je     0xffffffff810686d4 <vmx_vcpu_run+3188> <-- fastpath handler
>    0xffffffff81067d49 <+745>:   xor    %ebx,%ebx
>    0xffffffff81067d4b <+747>:   jmpq   0xffffffff81067e65 <vmx_vcpu_run+1029>