vmm svm: Refactor per-vCPU data.

- Allocate VMCBs separately to avoid excessive padding in struct
  svm_vcpu.

- Allocate APIC pages dynamically directly in struct vlapic.

- Move vm_mtrr into struct svm_vcpu rather than using a separate
  parallel array.

Reviewed by:	corvink, markj
Differential Revision:	https://reviews.freebsd.org/D37148

sys/amd64/vmm/amd/svm.c

@@ -565,8 +565,6 @@ svm_init(struct vm *vm, pmap_t pmap)
 	uint16_t maxcpus;
 
 	svm_sc = malloc(sizeof (*svm_sc), M_SVM, M_WAITOK | M_ZERO);
-	if (((uintptr_t)svm_sc & PAGE_MASK) != 0)
-		panic("malloc of svm_softc not aligned on page boundary");
 
 	svm_sc->msr_bitmap = contigmalloc(SVM_MSR_BITMAP_SIZE, M_SVM,
 	    M_WAITOK, 0, ~(vm_paddr_t)0, PAGE_SIZE, 0);
@@ -619,9 +617,11 @@ svm_init(struct vm *vm, pmap_t pmap)
 	maxcpus = vm_get_maxcpus(svm_sc->vm);
 	for (i = 0; i < maxcpus; i++) {
 		vcpu = svm_get_vcpu(svm_sc, i);
+		vcpu->vmcb = malloc_aligned(sizeof(struct vmcb), PAGE_SIZE,
+		    M_SVM, M_WAITOK | M_ZERO);
 		vcpu->nextrip = ~0;
 		vcpu->lastcpu = NOCPU;
-		vcpu->vmcb_pa = vtophys(&vcpu->vmcb);
+		vcpu->vmcb_pa = vtophys(vcpu->vmcb);
 		vmcb_init(svm_sc, i, iopm_pa, msrpm_pa, pml4_pa);
 		svm_msr_guest_init(svm_sc, i);
 	}
@@ -2149,7 +2149,14 @@ static void
 svm_cleanup(void *arg)
 {
 	struct svm_softc *sc = arg;
+	struct svm_vcpu *vcpu;
+	uint16_t i, maxcpus;
 
+	maxcpus = vm_get_maxcpus(sc->vm);
+	for (i = 0; i < maxcpus; i++) {
+		vcpu = svm_get_vcpu(sc, i);
+		free(vcpu->vmcb, M_SVM);
+	}
 	contigfree(sc->iopm_bitmap, SVM_IO_BITMAP_SIZE, M_SVM);
 	contigfree(sc->msr_bitmap, SVM_MSR_BITMAP_SIZE, M_SVM);
 	free(sc, M_SVM);
@@ -2400,7 +2407,8 @@ svm_vlapic_init(void *arg, int vcpuid)
 	vlapic = malloc(sizeof(struct vlapic), M_SVM_VLAPIC, M_WAITOK | M_ZERO);
 	vlapic->vm = svm_sc->vm;
 	vlapic->vcpuid = vcpuid;
-	vlapic->apic_page = (struct LAPIC *)&svm_sc->apic_page[vcpuid];
+	vlapic->apic_page = malloc_aligned(PAGE_SIZE, PAGE_SIZE, M_SVM_VLAPIC,
+	    M_WAITOK | M_ZERO);
 
 	vlapic_init(vlapic);
 
@@ -2412,6 +2420,7 @@ svm_vlapic_cleanup(void *arg, struct vlapic *vlapic)
 {
 
         vlapic_cleanup(vlapic);
+	free(vlapic->apic_page, M_SVM_VLAPIC);
         free(vlapic, M_SVM_VLAPIC);
 }
 

sys/amd64/vmm/amd/svm_msr.c

@@ -124,7 +124,7 @@ svm_rdmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t *result,
 	case MSR_MTRR16kBase ... MSR_MTRR16kBase + 1:
 	case MSR_MTRR64kBase:
 	case MSR_MTRRVarBase ... MSR_MTRRVarBase + (VMM_MTRR_VAR_MAX * 2) - 1:
-		if (vm_rdmtrr(&sc->mtrr[vcpu], num, result) != 0) {
+		if (vm_rdmtrr(&sc->vcpu[vcpu].mtrr, num, result) != 0) {
 			vm_inject_gp(sc->vm, vcpu);
 		}
 		break;
@@ -156,7 +156,7 @@ svm_wrmsr(struct svm_softc *sc, int vcpu, u_int num, uint64_t val, bool *retu)
 	case MSR_MTRR16kBase ... MSR_MTRR16kBase + 1:
 	case MSR_MTRR64kBase:
 	case MSR_MTRRVarBase ... MSR_MTRRVarBase + (VMM_MTRR_VAR_MAX * 2) - 1:
-		if (vm_wrmtrr(&sc->mtrr[vcpu], num, val) != 0) {
+		if (vm_wrmtrr(&sc->vcpu[vcpu].mtrr, num, val) != 0) {
 			vm_inject_gp(sc->vm, vcpu);
 		}
 		break;

sys/amd64/vmm/amd/svm_softc.h

@@ -41,12 +41,8 @@ struct asid {
 	uint32_t	num;	/* range is [1, nasid - 1] */
 };
 
-/*
- * XXX separate out 'struct vmcb' from 'svm_vcpu' to avoid wasting space
- * due to VMCB alignment requirements.
- */
 struct svm_vcpu {
-	struct vmcb	vmcb;	 /* hardware saved vcpu context */
+	struct vmcb	*vmcb;	 /* hardware saved vcpu context */
 	struct svm_regctx swctx; /* software saved vcpu context */
 	uint64_t	vmcb_pa; /* VMCB physical address */
 	uint64_t	nextrip; /* next instruction to be executed by guest */
@@ -54,23 +50,20 @@ struct svm_vcpu {
 	uint32_t	dirty;	 /* state cache bits that must be cleared */
 	long		eptgen;	 /* pmap->pm_eptgen when the vcpu last ran */
 	struct asid	asid;
-} __aligned(PAGE_SIZE);
+	struct vm_mtrr  mtrr;
+};
 
 /*
  * SVM softc, one per virtual machine.
  */
 struct svm_softc {
-	uint8_t apic_page[VM_MAXCPU][PAGE_SIZE];
 	struct svm_vcpu vcpu[VM_MAXCPU];
 	vm_offset_t 	nptp;			    /* nested page table */
 	uint8_t		*iopm_bitmap;    /* shared by all vcpus */
 	uint8_t		*msr_bitmap;    /* shared by all vcpus */
 	struct vm	*vm;
-	struct vm_mtrr  mtrr[VM_MAXCPU];
 };
 
-CTASSERT((offsetof(struct svm_softc, nptp) & PAGE_MASK) == 0);
-
 static __inline struct svm_vcpu *
 svm_get_vcpu(struct svm_softc *sc, int vcpu)
 {
@@ -82,21 +75,21 @@ static __inline struct vmcb *
 svm_get_vmcb(struct svm_softc *sc, int vcpu)
 {
 
-	return (&(sc->vcpu[vcpu].vmcb));
+	return ((sc->vcpu[vcpu].vmcb));
 }
 
 static __inline struct vmcb_state *
 svm_get_vmcb_state(struct svm_softc *sc, int vcpu)
 {
 
-	return (&(sc->vcpu[vcpu].vmcb.state));
+	return (&(sc->vcpu[vcpu].vmcb->state));
 }
 
 static __inline struct vmcb_ctrl *
 svm_get_vmcb_ctrl(struct svm_softc *sc, int vcpu)
 {
 
-	return (&(sc->vcpu[vcpu].vmcb.ctrl));
+	return (&(sc->vcpu[vcpu].vmcb->ctrl));
 }
 
 static __inline struct svm_regctx *