linux-kernel - Re: [PATCH v3 2/5] KVM: X86: Introduce need_cancel_enter

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives
Hash Suite: Windows password security audit tool. GUI, reports in PDF.
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <20200427183656.GO14870@linux.intel.com>
Date:   Mon, 27 Apr 2020 11:36:56 -0700
From:   Sean Christopherson <sean.j.christopherson@...el.com>
To:     Wanpeng Li <kernellwp@...il.com>
Cc:     LKML <linux-kernel@...r.kernel.org>, kvm <kvm@...r.kernel.org>,
        Paolo Bonzini <pbonzini@...hat.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Jim Mattson <jmattson@...gle.com>,
        Joerg Roedel <joro@...tes.org>,
        Haiwei Li <lihaiwei@...cent.com>
Subject: Re: [PATCH v3 2/5] KVM: X86: Introduce need_cancel_enter_guest helper

On Sun, Apr 26, 2020 at 10:05:00AM +0800, Wanpeng Li wrote:
> On Fri, 24 Apr 2020 at 14:23, Wanpeng Li <kernellwp@...il.com> wrote:
> >
> > From: Wanpeng Li <wanpengli@...cent.com>
> >
> > Introduce need_cancel_enter_guest() helper, we need to check some
> > conditions before doing CONT_RUN, in addition, it can also catch
> > the case vmexit occurred while another event was being delivered
> > to guest software since vmx_complete_interrupts() adds the request
> > bit.
> >
> > Tested-by: Haiwei Li <lihaiwei@...cent.com>
> > Cc: Haiwei Li <lihaiwei@...cent.com>
> > Signed-off-by: Wanpeng Li <wanpengli@...cent.com>
> > ---
> >  arch/x86/kvm/vmx/vmx.c | 12 +++++++-----
> >  arch/x86/kvm/x86.c     | 10 ++++++++--
> >  arch/x86/kvm/x86.h     |  1 +
> >  3 files changed, 16 insertions(+), 7 deletions(-)
> >
> > diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> > index f1f6638..5c21027 100644
> > --- a/arch/x86/kvm/vmx/vmx.c
> > +++ b/arch/x86/kvm/vmx/vmx.c
> > @@ -6577,7 +6577,7 @@ bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
> >
> >  static enum exit_fastpath_completion vmx_vcpu_run(struct kvm_vcpu *vcpu)
> >  {
> > -       enum exit_fastpath_completion exit_fastpath;
> > +       enum exit_fastpath_completion exit_fastpath = EXIT_FASTPATH_NONE;
> >         struct vcpu_vmx *vmx = to_vmx(vcpu);
> >         unsigned long cr3, cr4;
> >
> > @@ -6754,10 +6754,12 @@ static enum exit_fastpath_completion vmx_vcpu_run(struct kvm_vcpu *vcpu)
> >         vmx_recover_nmi_blocking(vmx);
> >         vmx_complete_interrupts(vmx);
> >
> > -       exit_fastpath = vmx_exit_handlers_fastpath(vcpu);
> > -       /* static call is better with retpolines */
> > -       if (exit_fastpath == EXIT_FASTPATH_CONT_RUN)
> > -               goto cont_run;
> > +       if (!kvm_need_cancel_enter_guest(vcpu)) {
> > +               exit_fastpath = vmx_exit_handlers_fastpath(vcpu);
> > +               /* static call is better with retpolines */
> > +               if (exit_fastpath == EXIT_FASTPATH_CONT_RUN)
> > +                       goto cont_run;
> > +       }
> 
> The kvm_need_cancel_enter_guest() should not before
> vmx_exit_handlers_fastpath() which will break IPI fastpath. How about
> applying something like below, otherwise, maybe introduce another
> EXIT_FASTPATH_CONT_FAIL to indicate fails due to
> kvm_need_cancel_enter_guest() if checking it after
> vmx_exit_handlers_fastpath(), then we return 1 in vmx_handle_exit()
> directly instead of kvm_skip_emulated_instruction(). VMX-preemption
> timer exit doesn't need to skip emulated instruction but wrmsr
> TSCDEADLINE MSR exit does which results in a little complex here.
> 
> Paolo, what do you think?
> 
> diff --git a/arch/x86/kvm/vmx/vmx.c b/arch/x86/kvm/vmx/vmx.c
> index 853d3af..9317924 100644
> --- a/arch/x86/kvm/vmx/vmx.c
> +++ b/arch/x86/kvm/vmx/vmx.c
> @@ -6564,6 +6564,9 @@ static enum exit_fastpath_completion
> handle_fastpath_preemption_timer(struct kvm
>  {
>      struct vcpu_vmx *vmx = to_vmx(vcpu);
> 
> +    if (kvm_need_cancel_enter_guest(vcpu))
> +        return EXIT_FASTPATH_NONE;
> +
>      if (!vmx->req_immediate_exit &&
>          !unlikely(vmx->loaded_vmcs->hv_timer_soft_disabled)) {
>              kvm_lapic_expired_hv_timer(vcpu);
> @@ -6771,12 +6774,10 @@ static enum exit_fastpath_completion
> vmx_vcpu_run(struct kvm_vcpu *vcpu)
>      vmx_recover_nmi_blocking(vmx);
>      vmx_complete_interrupts(vmx);
> 
> -    if (!(kvm_need_cancel_enter_guest(vcpu))) {
> -        exit_fastpath = vmx_exit_handlers_fastpath(vcpu);
> -        if (exit_fastpath == EXIT_FASTPATH_CONT_RUN) {
> -            vmx_sync_pir_to_irr(vcpu);
> -            goto cont_run;
> -        }
> +    exit_fastpath = vmx_exit_handlers_fastpath(vcpu);
> +    if (exit_fastpath == EXIT_FASTPATH_CONT_RUN) {

Relying on the handlers to check kvm_need_cancel_enter_guest() will be
error prone and costly to maintain.  I also don't like that it buries the
logic.

What about adding another flavor, e.g.:

	exit_fastpath = vmx_exit_handlers_fastpath(vcpu);
	if (exit_fastpath == EXIT_FASTPATH_CONT_RUN &&
	    kvm_need_cancel_enter_guest(vcpu))
		exit_fastpath = EXIT_FASTPATH_NOP;

That would also allow you to enable preemption timer without first having
to add CONT_RUN, which would be a very good thing for bisection.

> +        vmx_sync_pir_to_irr(vcpu);
> +        goto cont_run;
>      }
> 
>      return exit_fastpath;
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 99061ba..11b309c 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -1618,6 +1618,9 @@ static int
> handle_fastpath_set_x2apic_icr_irqoff(struct kvm_vcpu *vcpu, u64 data
> 
>  static int handle_fastpath_set_tscdeadline(struct kvm_vcpu *vcpu, u64 data)
>  {
> +    if (kvm_need_cancel_enter_guest(vcpu))
> +        return 1;
> +
>      if (!kvm_x86_ops.set_hv_timer ||
>          kvm_mwait_in_guest(vcpu->kvm) ||
>          kvm_can_post_timer_interrupt(vcpu))