Skip to contentSkip to conversation

KVM/nVMX: Use __vmx_vcpu_run in nested_vmx_check_vmentry_hw

PATCH In review Uros Bizjak 3 participants 6 messages Oct 29, 2020 – Dec 17, 2020

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

PATCH 1/1 KVM/nVMX: Use __vmx_vcpu_run in nested_vmx_check_vmentry_hw

Replace inline assembly in nested_vmx_check_vmentry_hw with a call to __vmx_vcpu_run. The function is not performance critical, so (double) GPR save/restore in __vmx_vcpu_run can be tolerated, as far as performance effects are concerned.

Cc Paolo Bonzini <[email protected]>
Cc Sean Christopherson <[email protected]>
---
 arch/x86/kvm/vmx/nested.c | 32 +++-----------------------------
 arch/x86/kvm/vmx/vmx.c    |  2 --
 arch/x86/kvm/vmx/vmx.h    |  1 +
 3 files changed, 4 insertions(+), 31 deletions(-)
arch/x86/kvm/vmx/nested.c +3 -29
index 89af692deb7e..6ab62bf277c4 100644
@@ -12,6 +12,7 @@
 #include "nested.h"
 #include "pmu.h"
 #include "trace.h"
+#include "vmx.h"
 #include "x86.h"
 
 static bool __read_mostly enable_shadow_vmcs = 1;
@@ -3056,35 +3057,8 @@ static int nested_vmx_check_vmentry_hw(struct kvm_vcpu *vcpu)
 		vmx->loaded_vmcs->host_state.cr4 = cr4;
 	}
 
-	asm(
-		"sub $%c[wordsize], %%" _ASM_SP "\n\t" /* temporarily adjust RSP for CALL */
-		"cmp %%" _ASM_SP ", %c[host_state_rsp](%[loaded_vmcs]) \n\t"
-		"je 1f \n\t"
-		__ex("vmwrite %%" _ASM_SP ", %[HOST_RSP]") "\n\t"
-		"mov %%" _ASM_SP ", %c[host_state_rsp](%[loaded_vmcs]) \n\t"
-		"1: \n\t"
-		"add $%c[wordsize], %%" _ASM_SP "\n\t" /* un-adjust RSP */
-
-		/* Check if vmlaunch or vmresume is needed */
-		"cmpb $0, %c[launched](%[loaded_vmcs])\n\t"
-
-		/*
-		 * VMLAUNCH and VMRESUME clear RFLAGS.{CF,ZF} on VM-Exit, set
-		 * RFLAGS.CF on VM-Fail Invalid and set RFLAGS.ZF on VM-Fail
-		 * Valid.  vmx_vmenter() directly "returns" RFLAGS, and so the
-		 * results of VM-Enter is captured via CC_{SET,OUT} to vm_fail.
-		 */
-		"call vmx_vmenter\n\t"
-
-		CC_SET(be)
-	      : ASM_CALL_CONSTRAINT, CC_OUT(be) (vm_fail)
-	      :	[HOST_RSP]"r"((unsigned long)HOST_RSP),
-		[loaded_vmcs]"r"(vmx->loaded_vmcs),
-		[launched]"i"(offsetof(struct loaded_vmcs, launched)),
-		[host_state_rsp]"i"(offsetof(struct loaded_vmcs, host_state.rsp)),
-		[wordsize]"i"(sizeof(ulong))
-	      : "memory"
-	);
+	vm_fail = __vmx_vcpu_run(vmx, (unsigned long *)&vcpu->arch.regs,
+				 vmx->loaded_vmcs->launched);
 
 	if (vmx->msr_autoload.host.nr)
 		vmcs_write32(VM_EXIT_MSR_LOAD_COUNT, vmx->msr_autoload.host.nr);
arch/x86/kvm/vmx/vmx.c +0 -2
index d14c94d0aff1..0f390c748b18 100644
@@ -6591,8 +6591,6 @@ static fastpath_t vmx_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
 	}
 }
 
-bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
-
 static noinstr void vmx_vcpu_enter_exit(struct kvm_vcpu *vcpu,
 					struct vcpu_vmx *vmx)
 {
arch/x86/kvm/vmx/vmx.h +1 -0
index f6f66e5c6510..32db3b033e9b 100644
@@ -339,6 +339,7 @@ void vmx_set_virtual_apic_mode(struct kvm_vcpu *vcpu);
 struct vmx_uret_msr *vmx_find_uret_msr(struct vcpu_vmx *vmx, u32 msr);
 void pt_update_intercept_for_msr(struct kvm_vcpu *vcpu);
 void vmx_update_host_rsp(struct vcpu_vmx *vmx, unsigned long host_rsp);
+bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
 int vmx_find_loadstore_msr_slot(struct vmx_msrs *m, u32 msr);
 void vmx_ept_load_pdptrs(struct kvm_vcpu *vcpu);
 
Signature · 1 line
2.26.2
#1 lore raw

Ping. This patch didn't receive any feedback.

Thanks,
Uros.

Quoted 94 lines · Uros Bizjak

Replace inline assembly in nested_vmx_check_vmentry_hw with a call to __vmx_vcpu_run. The function is not performance critical, so (double) GPR save/restore in __vmx_vcpu_run can be tolerated, as far as performance effects are concerned.

Cc Paolo Bonzini <[email protected]>
Cc Sean Christopherson <[email protected]>
---
 arch/x86/kvm/vmx/nested.c | 32 +++-----------------------------
 arch/x86/kvm/vmx/vmx.c    |  2 --
 arch/x86/kvm/vmx/vmx.h    |  1 +
 3 files changed, 4 insertions(+), 31 deletions(-)
arch/x86/kvm/vmx/nested.c +3 -29
index 89af692deb7e..6ab62bf277c4 100644
@@ -12,6 +12,7 @@
 #include "nested.h"
 #include "pmu.h"
 #include "trace.h"
+#include "vmx.h"
 #include "x86.h"
 
 static bool __read_mostly enable_shadow_vmcs = 1;
@@ -3056,35 +3057,8 @@ static int nested_vmx_check_vmentry_hw(struct kvm_vcpu *vcpu)
                vmx->loaded_vmcs->host_state.cr4 = cr4;
        }
 
-       asm(
-               "sub $%c[wordsize], %%" _ASM_SP "\n\t" /* temporarily adjust RSP for CALL */
-               "cmp %%" _ASM_SP ", %c[host_state_rsp](%[loaded_vmcs]) \n\t"
-               "je 1f \n\t"
-               __ex("vmwrite %%" _ASM_SP ", %[HOST_RSP]") "\n\t"
-               "mov %%" _ASM_SP ", %c[host_state_rsp](%[loaded_vmcs]) \n\t"
-               "1: \n\t"
-               "add $%c[wordsize], %%" _ASM_SP "\n\t" /* un-adjust RSP */
-
-               /* Check if vmlaunch or vmresume is needed */
-               "cmpb $0, %c[launched](%[loaded_vmcs])\n\t"
-
-               /*
-                * VMLAUNCH and VMRESUME clear RFLAGS.{CF,ZF} on VM-Exit, set
-                * RFLAGS.CF on VM-Fail Invalid and set RFLAGS.ZF on VM-Fail
-                * Valid.  vmx_vmenter() directly "returns" RFLAGS, and so the
-                * results of VM-Enter is captured via CC_{SET,OUT} to vm_fail.
-                */
-               "call vmx_vmenter\n\t"
-
-               CC_SET(be)
-             : ASM_CALL_CONSTRAINT, CC_OUT(be) (vm_fail)
-             : [HOST_RSP]"r"((unsigned long)HOST_RSP),
-               [loaded_vmcs]"r"(vmx->loaded_vmcs),
-               [launched]"i"(offsetof(struct loaded_vmcs, launched)),
-               [host_state_rsp]"i"(offsetof(struct loaded_vmcs, host_state.rsp)),
-               [wordsize]"i"(sizeof(ulong))
-             : "memory"
-       );
+       vm_fail = __vmx_vcpu_run(vmx, (unsigned long *)&vcpu->arch.regs,
+                                vmx->loaded_vmcs->launched);
 
        if (vmx->msr_autoload.host.nr)
                vmcs_write32(VM_EXIT_MSR_LOAD_COUNT, vmx->msr_autoload.host.nr);
arch/x86/kvm/vmx/vmx.c +0 -2
index d14c94d0aff1..0f390c748b18 100644
@@ -6591,8 +6591,6 @@ static fastpath_t vmx_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
        }
 }
 
-bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
-
 static noinstr void vmx_vcpu_enter_exit(struct kvm_vcpu *vcpu,
                                        struct vcpu_vmx *vmx)
 {
arch/x86/kvm/vmx/vmx.h +1 -1
index f6f66e5c6510..32db3b033e9b 100644
@@ -339,6 +339,7 @@ void vmx_set_virtual_apic_mode(struct kvm_vcpu *vcpu);
 struct vmx_uret_msr *vmx_find_uret_msr(struct vcpu_vmx *vmx, u32 msr);
 void pt_update_intercept_for_msr(struct kvm_vcpu *vcpu);
 void vmx_update_host_rsp(struct vcpu_vmx *vmx, unsigned long host_rsp);
+bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
 int vmx_find_loadstore_msr_slot(struct vmx_msrs *m, u32 msr);
 void vmx_ept_load_pdptrs(struct kvm_vcpu *vcpu);
 
--

2.26.2

#2 lore raw
Quoted 96 lines · Uros Bizjak

Ping. This patch didn't receive any feedback.

Thanks,
Uros.

Quoted 90 lines · Uros Bizjak

Replace inline assembly in nested_vmx_check_vmentry_hw with a call to __vmx_vcpu_run. The function is not performance critical, so (double) GPR save/restore in __vmx_vcpu_run can be tolerated, as far as performance effects are concerned.

Cc Paolo Bonzini <[email protected]>
Cc Sean Christopherson <[email protected]>
---
  arch/x86/kvm/vmx/nested.c | 32 +++-----------------------------
  arch/x86/kvm/vmx/vmx.c    |  2 --
  arch/x86/kvm/vmx/vmx.h    |  1 +
  3 files changed, 4 insertions(+), 31 deletions(-)
arch/x86/kvm/vmx/nested.c +3 -29
index 89af692deb7e..6ab62bf277c4 100644
@@ -12,6 +12,7 @@
  #include "nested.h"
  #include "pmu.h"
  #include "trace.h"
+#include "vmx.h"
  #include "x86.h"
 
  static bool __read_mostly enable_shadow_vmcs = 1;
@@ -3056,35 +3057,8 @@ static int nested_vmx_check_vmentry_hw(struct kvm_vcpu *vcpu)
                 vmx->loaded_vmcs->host_state.cr4 = cr4;
         }
 
-       asm(
-               "sub $%c[wordsize], %%" _ASM_SP "\n\t" /* temporarily adjust RSP for CALL */
-               "cmp %%" _ASM_SP ", %c[host_state_rsp](%[loaded_vmcs]) \n\t"
-               "je 1f \n\t"
-               __ex("vmwrite %%" _ASM_SP ", %[HOST_RSP]") "\n\t"
-               "mov %%" _ASM_SP ", %c[host_state_rsp](%[loaded_vmcs]) \n\t"
-               "1: \n\t"
-               "add $%c[wordsize], %%" _ASM_SP "\n\t" /* un-adjust RSP */
-
-               /* Check if vmlaunch or vmresume is needed */
-               "cmpb $0, %c[launched](%[loaded_vmcs])\n\t"
-
-               /*
-                * VMLAUNCH and VMRESUME clear RFLAGS.{CF,ZF} on VM-Exit, set
-                * RFLAGS.CF on VM-Fail Invalid and set RFLAGS.ZF on VM-Fail
-                * Valid.  vmx_vmenter() directly "returns" RFLAGS, and so the
-                * results of VM-Enter is captured via CC_{SET,OUT} to vm_fail.
-                */
-               "call vmx_vmenter\n\t"
-
-               CC_SET(be)
-             : ASM_CALL_CONSTRAINT, CC_OUT(be) (vm_fail)
-             : [HOST_RSP]"r"((unsigned long)HOST_RSP),
-               [loaded_vmcs]"r"(vmx->loaded_vmcs),
-               [launched]"i"(offsetof(struct loaded_vmcs, launched)),
-               [host_state_rsp]"i"(offsetof(struct loaded_vmcs, host_state.rsp)),
-               [wordsize]"i"(sizeof(ulong))
-             : "memory"
-       );
+       vm_fail = __vmx_vcpu_run(vmx, (unsigned long *)&vcpu->arch.regs,
+                                vmx->loaded_vmcs->launched);
 
         if (vmx->msr_autoload.host.nr)
                 vmcs_write32(VM_EXIT_MSR_LOAD_COUNT, vmx->msr_autoload.host.nr);
arch/x86/kvm/vmx/vmx.c +0 -2
index d14c94d0aff1..0f390c748b18 100644
@@ -6591,8 +6591,6 @@ static fastpath_t vmx_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
         }
  }
 
-bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
-
  static noinstr void vmx_vcpu_enter_exit(struct kvm_vcpu *vcpu,
                                         struct vcpu_vmx *vmx)
  {
arch/x86/kvm/vmx/vmx.h +1 -0
index f6f66e5c6510..32db3b033e9b 100644
@@ -339,6 +339,7 @@ void vmx_set_virtual_apic_mode(struct kvm_vcpu *vcpu);
  struct vmx_uret_msr *vmx_find_uret_msr(struct vcpu_vmx *vmx, u32 msr);
  void pt_update_intercept_for_msr(struct kvm_vcpu *vcpu);
  void vmx_update_host_rsp(struct vcpu_vmx *vmx, unsigned long host_rsp);
+bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
  int vmx_find_loadstore_msr_slot(struct vmx_msrs *m, u32 msr);
  void vmx_ept_load_pdptrs(struct kvm_vcpu *vcpu);
 
Quoted 3 lines
Quoted 3 lines

--
2.26.2

Semantically __vmx_vcpu_run() is called to enter guest mode. In nested_vmx_check_vmentry_hw(), we are not entering guest mode. Guest mode is entered when nested_vmx_enter_non_root_mode() calls enter_guest_mode().

Secondly, why not just replace the first half of the assembly block with a call to vmx_update_host_rsp() and leave the rest as is ?

#3 lore raw
Quoted 25 lines · Krish Sadhukhan
Quoted 25 lines
Quoted 25 lines
arch/x86/kvm/vmx/vmx.c +0 -2
index d14c94d0aff1..0f390c748b18 100644
@@ -6591,8 +6591,6 @@ static fastpath_t vmx_exit_handlers_fastpath(struct kvm_vcpu *vcpu)
         }
  }
 
-bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
-
  static noinstr void vmx_vcpu_enter_exit(struct kvm_vcpu *vcpu,
                                         struct vcpu_vmx *vmx)
  {
arch/x86/kvm/vmx/vmx.h +1 -0
index f6f66e5c6510..32db3b033e9b 100644
@@ -339,6 +339,7 @@ void vmx_set_virtual_apic_mode(struct kvm_vcpu *vcpu);
  struct vmx_uret_msr *vmx_find_uret_msr(struct vcpu_vmx *vmx, u32 msr);
  void pt_update_intercept_for_msr(struct kvm_vcpu *vcpu);
  void vmx_update_host_rsp(struct vcpu_vmx *vmx, unsigned long host_rsp);
+bool __vmx_vcpu_run(struct vcpu_vmx *vmx, unsigned long *regs, bool launched);
  int vmx_find_loadstore_msr_slot(struct vmx_msrs *m, u32 msr);
  void vmx_ept_load_pdptrs(struct kvm_vcpu *vcpu);
 
Quoted 6 lines
Quoted 3 lines
Quoted 3 lines

--
2.26.2

Semantically __vmx_vcpu_run() is called to enter guest mode. In nested_vmx_check_vmentry_hw(), we are not entering guest mode. Guest mode is entered when nested_vmx_enter_non_root_mode() calls enter_guest_mode().

Naming aside, this patch intentionally redefines the semantics to mean "execute VM-Enter that may or may not succeed". And as called out in the changelog, the overhead of the GPR save/load/restore is tolerable; reusing code and avoiding ugly inline asm is more important.

Quoted 2 lines

Secondly, why not just replace the first half of the assembly block with a call to vmx_update_host_rsp() and leave the rest as is ?

As above, though not called out in the changelog, the goal is to move away from the inline asm without introducing another asm subroutine.

Uros, I'll try to double check and review this later today.

#4 lore raw
In reply to Uros Bizjak

For future patches, please Cc LKML (in additional to KVM) so that the automatic archiving and patchwork stuff kicks in. Thanks!

Quoted 14 lines · Uros Bizjak

Ping. This patch didn't receive any feedback.

Quoted 11 lines · Uros Bizjak

Replace inline assembly in nested_vmx_check_vmentry_hw with a call to __vmx_vcpu_run. The function is not performance critical, so (double) GPR save/restore in __vmx_vcpu_run can be tolerated, as far as performance effects are concerned.

Cc Paolo Bonzini <[email protected]>
Cc Sean Christopherson <[email protected]>

---

vmx_vmenter() in vmx/vmenter.S can and should now use SYM_FUNC_START_LOCAL instead of SYM_FUNC_LOCAL. Other than that nit:

#5 lore raw
Quoted 20 lines · Sean Christopherson

For future patches, please Cc LKML (in additional to KVM) so that the automatic archiving and patchwork stuff kicks in. Thanks!

Quoted 14 lines · Uros Bizjak

Ping. This patch didn't receive any feedback.

Quoted 11 lines · Uros Bizjak

Replace inline assembly in nested_vmx_check_vmentry_hw with a call to __vmx_vcpu_run. The function is not performance critical, so (double) GPR save/restore in __vmx_vcpu_run can be tolerated, as far as performance effects are concerned.

Cc Paolo Bonzini <[email protected]>
Cc Sean Christopherson <[email protected]>

---

Quoted 4 lines

vmx_vmenter() in vmx/vmenter.S can and should now use SYM_FUNC_START_LOCAL instead of SYM_FUNC_LOCAL. Other than that nit:

Thanks!

I'll prepare a v2 (and added LKML, as you suggested).

Uros.