Only save/restore MSRs when needed

I've tested this with 32-bit and 64-bit guests on a 64-bit host and with 32-bit guests on a 32-bit host.

I *think* it's doing the right thing wrt to DEBUGCTL but an extra set of eyes would be helpful.

Depending on my machine's mood, we're at somewhere between 3200 and 2700 cycles per VMEXIT.

Regards,

Anthony Liguori
------------------------------------------------------------------------

From: Anthony Liguori <aliguori-r/[email protected]>
Subject: [PATCH][SVM] Only save/restore MSRs when needed

We only have to save/restore MSR_GS_BASE on every VMEXIT. The rest can be saved/restored when we leave the VCPU. As a special case, MSR_IA32_DEBUGCTL only needs to be saved/restored on every exit if debugging is enabled in either the host or the guest.

Signed-off-by Anthony Liguori <aliguori-r/[email protected]>

Index: kernel/drivers/kvm/svm.c
===================================================================

drivers/kvm/svm.c +31 -9

@@ -611,7 +611,7 @@
 
 static void svm_vcpu_load(struct kvm_vcpu *vcpu)
 {
-	int cpu;
+	int cpu, i;
 
 	cpu = get_cpu();
 	if (unlikely(cpu != vcpu->cpu)) {

@@ -626,10 +626,24 @@
 		vcpu->svm->vmcb->control.tsc_offset += delta;
 		vcpu->cpu = cpu;
 	}
+
+	for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
+		rdmsrl(host_save_user_msrs[i], vcpu->svm->host_user_msrs[i]);
+
+	rdmsrl(MSR_IA32_DEBUGCTLMSR, vcpu->svm->host_ia32_debugctl);
 }
 
 static void svm_vcpu_put(struct kvm_vcpu *vcpu)
 {
+	int i;
+
+	for (i = 0; i < NR_HOST_SAVE_USER_MSRS; i++)
+		wrmsrl(host_save_user_msrs[i], vcpu->svm->host_user_msrs[i]);
+
+	wrmsrl(MSR_IA32_DEBUGCTLMSR, vcpu->svm->host_ia32_debugctl);
+	/* FIXME: LASTBRANCHFROMIP, LASTBRANCHTOIP, LASTINTFROMIP,
+	   LASTINTTOIP */
+
 	rdtscll(vcpu->host_tsc);
 	put_cpu();
 }

@@ -815,18 +829,16 @@
 
 static void load_host_msrs(struct kvm_vcpu *vcpu)
 {
-	int i;
-
-	for ( i = 0; i < NR_HOST_SAVE_MSRS; i++)
-		wrmsrl(host_save_msrs[i], vcpu->svm->host_msrs[i]);
+#ifdef CONFIG_X86_64
+	wrmsrl(MSR_GS_BASE, vcpu->svm->host_gs_base);
+#endif
 }
 
 static void save_host_msrs(struct kvm_vcpu *vcpu)
 {
-	int i;
-
-	for ( i = 0; i < NR_HOST_SAVE_MSRS; i++)
-		rdmsrl(host_save_msrs[i], vcpu->svm->host_msrs[i]);
+#ifdef CONFIG_X86_64
+	rdmsrl(MSR_GS_BASE, vcpu->svm->host_gs_base);
+#endif
 }
 
 static void new_asid(struct kvm_vcpu *vcpu, struct svm_cpu_data *svm_data)

@@ -1498,6 +1510,11 @@
 		load_db_regs(vcpu->svm->db_regs);
 	}
 
+	if ((vcpu->svm->vmcb->save.dr7 & 0xff) ||
+	    (vcpu->svm->host_dr7 & 0xff)) {
+		wrmsrl(MSR_IA32_DEBUGCTLMSR, vcpu->svm->host_ia32_debugctl);
+	}
+
 	if (vcpu->fpu_active) {
 		fx_save(vcpu->host_fx_image);
 		fx_restore(vcpu->guest_fx_image);

@@ -1617,6 +1634,11 @@
 		fx_restore(vcpu->host_fx_image);
 	}
 
+	if ((vcpu->svm->vmcb->save.dr7 & 0xff) ||
+	    (vcpu->svm->host_dr7 & 0xff)) {
+		rdmsrl(MSR_IA32_DEBUGCTLMSR, vcpu->svm->host_ia32_debugctl);
+	}
+
 	if ((vcpu->svm->vmcb->save.dr7 & 0xff))
 		load_db_regs(vcpu->svm->host_db_regs);
 

Index: kernel/drivers/kvm/kvm_svm.h
===================================================================

drivers/kvm/kvm_svm.h +6 -8

@@ -9,17 +9,15 @@
 #include "svm.h"
 #include "kvm.h"
 
-static const u32 host_save_msrs[] = {
+static const u32 host_save_user_msrs[] = {
 #ifdef CONFIG_X86_64
 	MSR_STAR, MSR_LSTAR, MSR_CSTAR, MSR_SYSCALL_MASK, MSR_KERNEL_GS_BASE,
-	MSR_FS_BASE, MSR_GS_BASE,
+	MSR_FS_BASE,
 #endif
 	MSR_IA32_SYSENTER_CS, MSR_IA32_SYSENTER_ESP, MSR_IA32_SYSENTER_EIP,
-	MSR_IA32_DEBUGCTLMSR, /*MSR_IA32_LASTBRANCHFROMIP,
-	MSR_IA32_LASTBRANCHTOIP, MSR_IA32_LASTINTFROMIP,MSR_IA32_LASTINTTOIP,*/
 };
 
-#define NR_HOST_SAVE_MSRS ARRAY_SIZE(host_save_msrs)
+#define NR_HOST_SAVE_USER_MSRS ARRAY_SIZE(host_save_user_msrs)
 #define NUM_DB_REGS 4
 
 struct vcpu_svm {

@@ -32,7 +30,9 @@
 
 	u64 next_rip;
 
-	u64 host_msrs[NR_HOST_SAVE_MSRS];
+	u64 host_user_msrs[NR_HOST_SAVE_USER_MSRS];
+	u64 host_ia32_debugctl;
+	u64 host_gs_base;
 	unsigned long host_cr2;
 	unsigned long host_db_regs[NUM_DB_REGS];
 	unsigned long host_dr6;
 
------------------------------------------------------------------------
 
-------------------------------------------------------------------------

This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/
------------------------------------------------------------------------

_______________________________________________
kvm-devel mailing list [email protected]
https://lists.sourceforge.net/lists/listinfo/kvm-devel

Quoted 5 lines · Anthony Liguori

I've tested this with 32-bit and 64-bit guests on a 64-bit host and with 32-bit guests on a 32-bit host.

I *think* it's doing the right thing wrt to DEBUGCTL but an extra set of eyes would be helpful.

I think the follow patch preserves the current behavior but I'm starting to believe that the current behavior is wrong.

The current code assumes that the guest DEBUGCTL gets loaded/saved during vmrun/vmexit. By my reading of the spec (I finally printed out the new one :-)), DEBUGCTL only gets saved if LBR virtualization is enabled. LBR virtualization also triggers the save/restore of the other 4 debug registers. The spec seems to suggest that the host version is saved before the guest is loaded too.

I think the right behavior is to do nothing if LBR is available and enabled; hw will do it all for us. If LBR isn't available, I think we should be saving/restoring not just DEBUGCTL but the other 4. We can further optimize that path though by doing the same sort of tricks I have below.

If that all seems sane, I'll submit another patch to fix that up.

Quoted 18 lines · Anthony Liguori

Quoted 6 lines · Anthony Liguori

I've tested this with 32-bit and 64-bit guests on a 64-bit host and with 32-bit guests on a 32-bit host.

I *think* it's doing the right thing wrt to DEBUGCTL but an extra set of eyes would be helpful.

I think the follow patch preserves the current behavior but I'm starting to believe that the current behavior is wrong.

The current code assumes that the guest DEBUGCTL gets loaded/saved during vmrun/vmexit. By my reading of the spec (I finally printed out the new one :-)), DEBUGCTL only gets saved if LBR virtualization is enabled. LBR virtualization also triggers the save/restore of the other 4 debug registers. The spec seems to suggest that the host version is saved before the guest is loaded too.

Quoted 9 lines

I think the right behavior is to do nothing if LBR is available and enabled; hw will do it all for us. If LBR isn't available, I think we should be saving/restoring not just DEBUGCTL but the other 4. We can further optimize that path though by doing the same sort of tricks I have below.

If that all seems sane, I'll submit another patch to fix that up.

Some observations from grepping the sources:

- The host never uses the DEBUGCTL msr. So we only need to restore it (to zero) if the guest changes it.
- The guest never uses the DEBUGCTL msr, as kvm doesn't handle wrmsr to DEBUGCTL.

So, doing nothing (in hardware and software) would probably work well and be quite fast. If we ever emulate DEBUGCTL, we can conditionally enable hardware or software save/restore (this is fairly similar to lazy fpu).

If Linux ever gets DEBUGCTL support, we'd just ask for a hook so we can conditionally save and restore on the host side.