[PATCH] x86: svm: make wbinvd faster

[PATCH] x86: svm: make wbinvd faster

[Joel Schopp]

From: David Kaplan <David.Kaplan@amd.com>
No need to re-decode WBINVD since we know what it is from the intercept.

Signed-off-by: David Kaplan <David.Kaplan@amd.com>
[extracted from larger unlrelated patch, forward ported, tested]
Signed-off-by: Joel Schopp <joel.schopp@amd.com>
---
 arch/x86/kvm/svm.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
index d319e0c..86ecd21 100644
--- a/arch/x86/kvm/svm.c
+++ b/arch/x86/kvm/svm.c
@@ -2776,6 +2776,14 @@ static int skinit_interception(struct vcpu_svm *svm)
 	return 1;
 }
 
+static int wbinvd_interception(struct vcpu_svm *svm)
+{
+	kvm_emulate_wbinvd(&svm->vcpu);
+	skip_emulated_instruction(&svm->vcpu);
+	return 1;
+}
+
+
 static int xsetbv_interception(struct vcpu_svm *svm)
 {
 	u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
@@ -3376,7 +3384,7 @@ static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
 	[SVM_EXIT_STGI]				= stgi_interception,
 	[SVM_EXIT_CLGI]				= clgi_interception,
 	[SVM_EXIT_SKINIT]			= skinit_interception,
-	[SVM_EXIT_WBINVD]                       = emulate_on_interception,
+	[SVM_EXIT_WBINVD]                       = wbinvd_interception,
 	[SVM_EXIT_MONITOR]			= monitor_interception,
 	[SVM_EXIT_MWAIT]			= mwait_interception,
 	[SVM_EXIT_XSETBV]			= xsetbv_interception,

Re: [PATCH] x86: svm: make wbinvd faster

[Bandan Das]

Joel Schopp <joel.schopp@amd.com> writes:

> From: David Kaplan <David.Kaplan@amd.com>
> No need to re-decode WBINVD since we know what it is from the intercept.
>
> Signed-off-by: David Kaplan <David.Kaplan@amd.com>
> [extracted from larger unlrelated patch, forward ported, tested]
> Signed-off-by: Joel Schopp <joel.schopp@amd.com>
> ---
>  arch/x86/kvm/svm.c |   10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
>
> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> index d319e0c..86ecd21 100644
> --- a/arch/x86/kvm/svm.c
> +++ b/arch/x86/kvm/svm.c
> @@ -2776,6 +2776,14 @@ static int skinit_interception(struct vcpu_svm *svm)
>  	return 1;
>  }
>  
> +static int wbinvd_interception(struct vcpu_svm *svm)
> +{
> +	kvm_emulate_wbinvd(&svm->vcpu);
> +	skip_emulated_instruction(&svm->vcpu);
> +	return 1;
> +}
> +
> +
Can't we merge this to kvm_emulate_wbinvd, and just call that function
directly for both vmx and svm ?

>  static int xsetbv_interception(struct vcpu_svm *svm)
>  {
>  	u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
> @@ -3376,7 +3384,7 @@ static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
>  	[SVM_EXIT_STGI]				= stgi_interception,
>  	[SVM_EXIT_CLGI]				= clgi_interception,
>  	[SVM_EXIT_SKINIT]			= skinit_interception,
> -	[SVM_EXIT_WBINVD]                       = emulate_on_interception,
So, this means x86_emulate_insn() in emulate.c has no callers left for the
wbinvd case ? vmx calls kvm_emulate_wbinvd directly too..

Bandan
> +	[SVM_EXIT_WBINVD]                       = wbinvd_interception,
>  	[SVM_EXIT_MONITOR]			= monitor_interception,
>  	[SVM_EXIT_MWAIT]			= mwait_interception,
>  	[SVM_EXIT_XSETBV]			= xsetbv_interception,
>
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH] x86: svm: make wbinvd faster

[Radim Krčmář]

2015-03-01 21:29-0500, Bandan Das:
> Joel Schopp <joel.schopp@amd.com> writes:
> 
> > From: David Kaplan <David.Kaplan@amd.com>
> > No need to re-decode WBINVD since we know what it is from the intercept.
> >
> > Signed-off-by: David Kaplan <David.Kaplan@amd.com>
> > [extracted from larger unlrelated patch, forward ported, tested]
> > Signed-off-by: Joel Schopp <joel.schopp@amd.com>
> > ---
> > diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> > +static int wbinvd_interception(struct vcpu_svm *svm)
> > +{
> > +	kvm_emulate_wbinvd(&svm->vcpu);
> > +	skip_emulated_instruction(&svm->vcpu);
> > +	return 1;
> > +}
> > +
> > +
> Can't we merge this to kvm_emulate_wbinvd, and just call that function
> directly for both vmx and svm ?

kvm_emulate_wbinvd() lives in x86.c and skip_emulated_instruction() is
from svm.c/vmx.c:  so we'd have to create a new x86 op and change the
emulator code as well ... it's probably better like this.

> >  static int xsetbv_interception(struct vcpu_svm *svm)
> >  {
> >  	u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
> > @@ -3376,7 +3384,7 @@ static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
> >  	[SVM_EXIT_STGI]				= stgi_interception,
> >  	[SVM_EXIT_CLGI]				= clgi_interception,
> >  	[SVM_EXIT_SKINIT]			= skinit_interception,
> > -	[SVM_EXIT_WBINVD]                       = emulate_on_interception,
> So, this means x86_emulate_insn() in emulate.c has no callers left for the
> wbinvd case ? vmx calls kvm_emulate_wbinvd directly too..

I think that invalid state emulation might still hit wbinvd.

Re: [PATCH] x86: svm: make wbinvd faster

[Bandan Das]

Radim Krčmář <rkrcmar@redhat.com> writes:

> 2015-03-01 21:29-0500, Bandan Das:
>> Joel Schopp <joel.schopp@amd.com> writes:
>> 
>> > From: David Kaplan <David.Kaplan@amd.com>
>> > No need to re-decode WBINVD since we know what it is from the intercept.
>> >
>> > Signed-off-by: David Kaplan <David.Kaplan@amd.com>
>> > [extracted from larger unlrelated patch, forward ported, tested]
>> > Signed-off-by: Joel Schopp <joel.schopp@amd.com>
>> > ---
>> > diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
>> > +static int wbinvd_interception(struct vcpu_svm *svm)
>> > +{
>> > +	kvm_emulate_wbinvd(&svm->vcpu);
>> > +	skip_emulated_instruction(&svm->vcpu);
>> > +	return 1;
>> > +}
>> > +
>> > +
>> Can't we merge this to kvm_emulate_wbinvd, and just call that function
>> directly for both vmx and svm ?
>
> kvm_emulate_wbinvd() lives in x86.c and skip_emulated_instruction() is
> from svm.c/vmx.c:  so we'd have to create a new x86 op and change the
> emulator code as well ... it's probably better like this.

There's already one - kvm_x86_ops->skip_emulated_instruction

>> >  static int xsetbv_interception(struct vcpu_svm *svm)
>> >  {
>> >  	u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
>> > @@ -3376,7 +3384,7 @@ static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
>> >  	[SVM_EXIT_STGI]				= stgi_interception,
>> >  	[SVM_EXIT_CLGI]				= clgi_interception,
>> >  	[SVM_EXIT_SKINIT]			= skinit_interception,
>> > -	[SVM_EXIT_WBINVD]                       = emulate_on_interception,
>> So, this means x86_emulate_insn() in emulate.c has no callers left for the
>> wbinvd case ? vmx calls kvm_emulate_wbinvd directly too..
>
> I think that invalid state emulation might still hit wbinvd.

Re: [PATCH] x86: svm: make wbinvd faster

[Radim Krčmář]

2015-03-02 10:25-0500, Bandan Das:
> Radim Krčmář <rkrcmar@redhat.com> writes:
> > 2015-03-01 21:29-0500, Bandan Das:
> >> Joel Schopp <joel.schopp@amd.com> writes:
> >> > +static int wbinvd_interception(struct vcpu_svm *svm)
> >> > +{
> >> > +	kvm_emulate_wbinvd(&svm->vcpu);
> >> > +	skip_emulated_instruction(&svm->vcpu);
> >> > +	return 1;
> >> > +}
> >> Can't we merge this to kvm_emulate_wbinvd, and just call that function
> >> directly for both vmx and svm ?
> >
> > kvm_emulate_wbinvd() lives in x86.c and skip_emulated_instruction() is
> > from svm.c/vmx.c:  so we'd have to create a new x86 op and change the
> > emulator code as well ... it's probably better like this.
> 
> There's already one - kvm_x86_ops->skip_emulated_instruction

My bad, its usage is inconsistent and I only looked at two close
interceptions where it was used ... kvm_emulate_cpuid() calls
kvm_x86_ops->skip_emulated_instruction(), while kvm_emulate_halt() and
kvm_emulate_hypercall() need an external skip.

We do "skip" the instruction with kvm_emulate(), so automatically
skipping the instruction on kvm_emulate_*() makes sense:
 1. rename kvm_emulate_halt() and kvm_emulate_wbinvd() to accommodate
    callers that don't want to skip
 2. introduce kvm_emulate_{halt,wbinvd}() and move the skip to to
    kvm_emulate_{halt,wbinvd,hypercall}()

The alternative is to remove kvm_x86_ops->skip_emulated_instruction():
 1. remove skip from kvm_emulate_cpuid() and modify callers
 2. move kvm_complete_insn_gp to a header file and use
    skip_emulated_instruction directly
 3. remove unused kvm_x86_ops->skip_emulated_instruction()

Which one do you prefer?

Thanks.

Re: [PATCH] x86: svm: make wbinvd faster

[Joel Schopp]

On 03/02/2015 10:03 AM, Radim Krčmář wrote:
> 2015-03-02 10:25-0500, Bandan Das:
>> Radim Krčmář <rkrcmar@redhat.com> writes:
>>> 2015-03-01 21:29-0500, Bandan Das:
>>>> Joel Schopp <joel.schopp@amd.com> writes:
>>>>> +static int wbinvd_interception(struct vcpu_svm *svm)
>>>>> +{
>>>>> +	kvm_emulate_wbinvd(&svm->vcpu);
>>>>> +	skip_emulated_instruction(&svm->vcpu);
>>>>> +	return 1;
>>>>> +}
>>>> Can't we merge this to kvm_emulate_wbinvd, and just call that function
>>>> directly for both vmx and svm ?
>>> kvm_emulate_wbinvd() lives in x86.c and skip_emulated_instruction() is
>>> from svm.c/vmx.c:  so we'd have to create a new x86 op and change the
>>> emulator code as well ... it's probably better like this.
>> There's already one - kvm_x86_ops->skip_emulated_instruction
> My bad, its usage is inconsistent and I only looked at two close
> interceptions where it was used ... kvm_emulate_cpuid() calls
> kvm_x86_ops->skip_emulated_instruction(), while kvm_emulate_halt() and
> kvm_emulate_hypercall() need an external skip.
>
> We do "skip" the instruction with kvm_emulate(), so automatically
> skipping the instruction on kvm_emulate_*() makes sense:
>   1. rename kvm_emulate_halt() and kvm_emulate_wbinvd() to accommodate
>      callers that don't want to skip
>   2. introduce kvm_emulate_{halt,wbinvd}() and move the skip to to
>      kvm_emulate_{halt,wbinvd,hypercall}()
>
> The alternative is to remove kvm_x86_ops->skip_emulated_instruction():
>   1. remove skip from kvm_emulate_cpuid() and modify callers
>   2. move kvm_complete_insn_gp to a header file and use
>      skip_emulated_instruction directly
>   3. remove unused kvm_x86_ops->skip_emulated_instruction()
>
> Which one do you prefer?
I prefer renaming them,  ie kvm_emulate_wbinvd_noskip(), and making the 
existing ones, ie kvm_emulate_wbinvd() call the noskip verion and add a 
skip similar to how wbinvd_interception above does.  I can send out a 
patch later today with that rework.

Re: [PATCH] x86: svm: make wbinvd faster

[Radim Krčmář]

2015-02-27 18:19-0600, Joel Schopp:
> From: David Kaplan <David.Kaplan@amd.com>
> No need to re-decode WBINVD since we know what it is from the intercept.
> 
> Signed-off-by: David Kaplan <David.Kaplan@amd.com>
> [extracted from larger unlrelated patch, forward ported, tested]
> Signed-off-by: Joel Schopp <joel.schopp@amd.com>
> ---

Reviewed-by: Radim Krčmář <rkrcmar@redhat.com>

> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> +static int wbinvd_interception(struct vcpu_svm *svm)
> +{
> +	kvm_emulate_wbinvd(&svm->vcpu);
> +	skip_emulated_instruction(&svm->vcpu);
> +	return 1;
> +}
> +
> +

(One line is optimal.)

Re: [PATCH] x86: svm: make wbinvd faster

[Marcelo Tosatti]

On Fri, Feb 27, 2015 at 06:19:18PM -0600, Joel Schopp wrote:
> From: David Kaplan <David.Kaplan@amd.com>
> No need to re-decode WBINVD since we know what it is from the intercept.
> 
> Signed-off-by: David Kaplan <David.Kaplan@amd.com>
> [extracted from larger unlrelated patch, forward ported, tested]
> Signed-off-by: Joel Schopp <joel.schopp@amd.com>

Can't you disable the intercept if need_emulate_wbinvd(vcpu) == false? 

> ---
>  arch/x86/kvm/svm.c |   10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/arch/x86/kvm/svm.c b/arch/x86/kvm/svm.c
> index d319e0c..86ecd21 100644
> --- a/arch/x86/kvm/svm.c
> +++ b/arch/x86/kvm/svm.c
> @@ -2776,6 +2776,14 @@ static int skinit_interception(struct vcpu_svm *svm)
>  	return 1;
>  }
>  
> +static int wbinvd_interception(struct vcpu_svm *svm)
> +{
> +	kvm_emulate_wbinvd(&svm->vcpu);
> +	skip_emulated_instruction(&svm->vcpu);
> +	return 1;
> +}
> +
> +
>  static int xsetbv_interception(struct vcpu_svm *svm)
>  {
>  	u64 new_bv = kvm_read_edx_eax(&svm->vcpu);
> @@ -3376,7 +3384,7 @@ static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = {
>  	[SVM_EXIT_STGI]				= stgi_interception,
>  	[SVM_EXIT_CLGI]				= clgi_interception,
>  	[SVM_EXIT_SKINIT]			= skinit_interception,
> -	[SVM_EXIT_WBINVD]                       = emulate_on_interception,
> +	[SVM_EXIT_WBINVD]                       = wbinvd_interception,
>  	[SVM_EXIT_MONITOR]			= monitor_interception,
>  	[SVM_EXIT_MWAIT]			= mwait_interception,
>  	[SVM_EXIT_XSETBV]			= xsetbv_interception,
> 
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: [PATCH] x86: svm: make wbinvd faster

[Radim Krčmář]

2015-03-09 20:28-0300, Marcelo Tosatti:
> On Fri, Feb 27, 2015 at 06:19:18PM -0600, Joel Schopp wrote:
> > From: David Kaplan <David.Kaplan@amd.com>
> > No need to re-decode WBINVD since we know what it is from the intercept.
> > 
> > Signed-off-by: David Kaplan <David.Kaplan@amd.com>
> > [extracted from larger unlrelated patch, forward ported, tested]
> > Signed-off-by: Joel Schopp <joel.schopp@amd.com>
> 
> Can't you disable the intercept if need_emulate_wbinvd(vcpu) == false? 

I don't think we want to:  it should be faster to intercept and ignore
than to invalidate all caches.  The exit doesn't affect other physical
cores and costs just about 10(?) L3 cache misses.

Re: [PATCH] x86: svm: make wbinvd faster

[Marcelo Tosatti]

On Tue, Mar 10, 2015 at 12:01:31PM +0100, Radim Krčmář wrote:
> 2015-03-09 20:28-0300, Marcelo Tosatti:
> > On Fri, Feb 27, 2015 at 06:19:18PM -0600, Joel Schopp wrote:
> > > From: David Kaplan <David.Kaplan@amd.com>
> > > No need to re-decode WBINVD since we know what it is from the intercept.
> > > 
> > > Signed-off-by: David Kaplan <David.Kaplan@amd.com>
> > > [extracted from larger unlrelated patch, forward ported, tested]
> > > Signed-off-by: Joel Schopp <joel.schopp@amd.com>
> > 
> > Can't you disable the intercept if need_emulate_wbinvd(vcpu) == false? 
> 
> I don't think we want to:  it should be faster to intercept and ignore
> than to invalidate all caches.  The exit doesn't affect other physical
> cores and costs just about 10(?) L3 cache misses.

Yes, right.