Skip to contentSkip to conversation

x86: svm: make wbinvd faster

PATCH In review Joel Schopp 4 participants 10 messages Feb 28, 2015 – Mar 10, 2015

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 x86: svm: make wbinvd faster

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

[extracted from larger unlrelated patch, forward ported, tested]
---
 arch/x86/kvm/svm.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
arch/x86/kvm/svm.c +9 -1
index d319e0c..86ecd21 100644
@@ -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,
#1 lore raw
Quoted 16 lines · Joel Schopp

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

[extracted from larger unlrelated patch, forward ported, tested]
---
 arch/x86/kvm/svm.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
arch/x86/kvm/svm.c +0 -0
index d319e0c..86ecd21 100644
@@ -2776,6 +2776,14 @@ static int skinit_interception(struct vcpu_svm *svm)
 	return 1;

arch/x86/kvm/svm.c static int skinit_interception(struct vcpu_svm *svm) · line 2786

Quoted 10 lines
arch/x86/kvm/svm.c +8 -0
 }
 
+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 ?

arch/x86/kvm/svm.c static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = { · line 3379

Quoted 8 lines
arch/x86/kvm/svm.c +0 -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,

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

Quoted 9 lines
arch/x86/kvm/svm.c +1 -1
+	[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 [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#2 lore raw

2015-03-01 21:29-0500, Bandan Das:

Quoted 5 lines
Quoted 3 lines · Joel Schopp

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

Quoted 15 lines

arch/x86/kvm/svm.c static int skinit_interception(struct vcpu_svm *svm) · line 2786

Quoted 13 lines
[extracted from larger unlrelated patch, forward ported, tested]

---

arch/x86/kvm/svm.c +8 -0
+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.

Quoted 10 lines

arch/x86/kvm/svm.c static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = { · line 3379

Quoted 8 lines

static int xsetbv_interception(struct vcpu_svm *svm) {

arch/x86/kvm/svm.c +0 -1
 	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.

#3 lore raw
Quoted 22 lines · Radim Krčmář

2015-03-01 21:29-0500, Bandan Das:

Quoted 5 lines
Quoted 3 lines · Joel Schopp

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

Quoted 15 lines

arch/x86/kvm/svm.c static int skinit_interception(struct vcpu_svm *svm) · line 2786

Quoted 13 lines
[extracted from larger unlrelated patch, forward ported, tested]

---

arch/x86/kvm/svm.c +8 -0
+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 ?

Quoted 3 lines

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

Quoted 12 lines
Quoted 10 lines

arch/x86/kvm/svm.c static int (*const svm_exit_handlers[])(struct vcpu_svm *svm) = { · line 3379

Quoted 8 lines

static int xsetbv_interception(struct vcpu_svm *svm) {

arch/x86/kvm/svm.c +0 -1
 	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.

#4 lore raw

2015-03-02 10:25-0500, Bandan Das:

Quoted 12 lines
Quoted 11 lines · Radim Krčmář

2015-03-01 21:29-0500, Bandan Das:

Quoted 9 lines

arch/x86/kvm/svm.c static int skinit_interception(struct vcpu_svm *svm) · line 2784

Quoted 6 lines · Joel Schopp
arch/x86/kvm/svm.c +6 -0
+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 ?

Quoted 5 lines
Quoted 3 lines

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.

#5 lore raw
Quoted 28 lines · Radim Krčmář

2015-03-02 10:25-0500, Bandan Das:

Quoted 11 lines
Quoted 10 lines · Radim Krčmář

2015-03-01 21:29-0500, Bandan Das:

Quoted 9 lines

arch/x86/kvm/svm.c static int skinit_interception(struct vcpu_svm *svm) · line 2784

Quoted 6 lines · Joel Schopp
arch/x86/kvm/svm.c +6 -0
+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 ?

Quoted 4 lines
Quoted 3 lines

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}()

Quoted 7 lines

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.

Reviewed-by on v1 1/1
#6 lore raw
In reply to Joel Schopp

2015-02-27 18:19-0600, Joel Schopp:

Quoted 7 lines

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

[extracted from larger unlrelated patch, forward ported, tested]

---

arch/x86/kvm/svm.c static int skinit_interception(struct vcpu_svm *svm) · line 2786

Quoted 9 lines
arch/x86/kvm/svm.c +8 -0
+static int wbinvd_interception(struct vcpu_svm *svm)
+{
+	kvm_emulate_wbinvd(&svm->vcpu);
+	skip_emulated_instruction(&svm->vcpu);
+	return 1;
+}
+
+

(One line is optimal.)

#7 lore raw
In reply to Joel Schopp
Quoted 6 lines · Joel Schopp

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

[extracted from larger unlrelated patch, forward ported, tested]

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

Quoted 37 lines
---
 arch/x86/kvm/svm.c |   10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)
arch/x86/kvm/svm.c +9 -2
index d319e0c..86ecd21 100644
@@ -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 [email protected]
More majordomo info at http://vger.kernel.org/majordomo-info.html

#8 lore raw

2015-03-09 20:28-0300, Marcelo Tosatti:

Quoted 9 lines · Joel Schopp
Quoted 6 lines

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

[extracted from larger unlrelated patch, forward ported, tested]

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.

#9 lore raw
Quoted 9 lines · Radim Krčmář

2015-03-09 20:28-0300, Marcelo Tosatti:

Quoted 8 lines · Joel Schopp
Quoted 6 lines

From: David Kaplan <[email protected]>
No need to re-decode WBINVD since we know what it is from the intercept.

[extracted from larger unlrelated patch, forward ported, tested]
Quoted 5 lines
Quoted 1 line

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.