KVM: SVM: add support for Nested Paging
authorJoerg Roedel <joerg.roedel@amd.com>
Thu, 7 Feb 2008 12:47:45 +0000 (13:47 +0100)
committerAvi Kivity <avi@qumranet.com>
Sun, 27 Apr 2008 08:53:21 +0000 (11:53 +0300)
This patch contains the SVM architecture dependent changes for KVM to enable
support for the Nested Paging feature of AMD Barcelona and Phenom processors.

Signed-off-by: Joerg Roedel <joerg.roedel@amd.com>
Signed-off-by: Avi Kivity <avi@qumranet.com>
arch/x86/kvm/svm.c

index 9e29a13136c44a0fd3f13ba0b1e0a6e2debcd1eb..8e9d4a5dacda35101d79c5f1a4ae570f4734ac1d 100644 (file)
@@ -47,7 +47,12 @@ MODULE_LICENSE("GPL");
 #define SVM_FEATURE_LBRV (1 << 1)
 #define SVM_DEATURE_SVML (1 << 2)
 
+/* enable NPT for AMD64 and X86 with PAE */
+#if defined(CONFIG_X86_64) || defined(CONFIG_X86_PAE)
+static bool npt_enabled = true;
+#else
 static bool npt_enabled = false;
+#endif
 static int npt = 1;
 
 module_param(npt, int, S_IRUGO);
@@ -187,7 +192,7 @@ static inline void flush_guest_tlb(struct kvm_vcpu *vcpu)
 
 static void svm_set_efer(struct kvm_vcpu *vcpu, u64 efer)
 {
-       if (!(efer & EFER_LMA))
+       if (!npt_enabled && !(efer & EFER_LMA))
                efer &= ~EFER_LME;
 
        to_svm(vcpu)->vmcb->save.efer = efer | MSR_EFER_SVME_MASK;
@@ -573,6 +578,22 @@ static void init_vmcb(struct vmcb *vmcb)
        save->cr0 = 0x00000010 | X86_CR0_PG | X86_CR0_WP;
        save->cr4 = X86_CR4_PAE;
        /* rdx = ?? */
+
+       if (npt_enabled) {
+               /* Setup VMCB for Nested Paging */
+               control->nested_ctl = 1;
+               control->intercept_exceptions &= ~(1 << PF_VECTOR);
+               control->intercept_cr_read &= ~(INTERCEPT_CR0_MASK|
+                                               INTERCEPT_CR3_MASK);
+               control->intercept_cr_write &= ~(INTERCEPT_CR0_MASK|
+                                                INTERCEPT_CR3_MASK);
+               save->g_pat = 0x0007040600070406ULL;
+               /* enable caching because the QEMU Bios doesn't enable it */
+               save->cr0 = X86_CR0_ET;
+               save->cr3 = 0;
+               save->cr4 = 0;
+       }
+
 }
 
 static int svm_vcpu_reset(struct kvm_vcpu *vcpu)
@@ -807,6 +828,9 @@ static void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
                }
        }
 #endif
+       if (npt_enabled)
+               goto set;
+
        if ((vcpu->arch.cr0 & X86_CR0_TS) && !(cr0 & X86_CR0_TS)) {
                svm->vmcb->control.intercept_exceptions &= ~(1 << NM_VECTOR);
                vcpu->fpu_active = 1;
@@ -814,18 +838,26 @@ static void svm_set_cr0(struct kvm_vcpu *vcpu, unsigned long cr0)
 
        vcpu->arch.cr0 = cr0;
        cr0 |= X86_CR0_PG | X86_CR0_WP;
-       cr0 &= ~(X86_CR0_CD | X86_CR0_NW);
        if (!vcpu->fpu_active) {
                svm->vmcb->control.intercept_exceptions |= (1 << NM_VECTOR);
                cr0 |= X86_CR0_TS;
        }
+set:
+       /*
+        * re-enable caching here because the QEMU bios
+        * does not do it - this results in some delay at
+        * reboot
+        */
+       cr0 &= ~(X86_CR0_CD | X86_CR0_NW);
        svm->vmcb->save.cr0 = cr0;
 }
 
 static void svm_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)
 {
        vcpu->arch.cr4 = cr4;
-       to_svm(vcpu)->vmcb->save.cr4 = cr4 | X86_CR4_PAE;
+       if (!npt_enabled)
+              cr4 |= X86_CR4_PAE;
+       to_svm(vcpu)->vmcb->save.cr4 = cr4;
 }
 
 static void svm_set_segment(struct kvm_vcpu *vcpu,
@@ -1313,14 +1345,34 @@ static int (*svm_exit_handlers[])(struct vcpu_svm *svm,
        [SVM_EXIT_WBINVD]                       = emulate_on_interception,
        [SVM_EXIT_MONITOR]                      = invalid_op_interception,
        [SVM_EXIT_MWAIT]                        = invalid_op_interception,
+       [SVM_EXIT_NPF]                          = pf_interception,
 };
 
-
 static int handle_exit(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
 {
        struct vcpu_svm *svm = to_svm(vcpu);
        u32 exit_code = svm->vmcb->control.exit_code;
 
+       if (npt_enabled) {
+               int mmu_reload = 0;
+               if ((vcpu->arch.cr0 ^ svm->vmcb->save.cr0) & X86_CR0_PG) {
+                       svm_set_cr0(vcpu, svm->vmcb->save.cr0);
+                       mmu_reload = 1;
+               }
+               vcpu->arch.cr0 = svm->vmcb->save.cr0;
+               vcpu->arch.cr3 = svm->vmcb->save.cr3;
+               if (is_paging(vcpu) && is_pae(vcpu) && !is_long_mode(vcpu)) {
+                       if (!load_pdptrs(vcpu, vcpu->arch.cr3)) {
+                               kvm_inject_gp(vcpu, 0);
+                               return 1;
+                       }
+               }
+               if (mmu_reload) {
+                       kvm_mmu_reset_context(vcpu);
+                       kvm_mmu_load(vcpu);
+               }
+       }
+
        kvm_reput_irq(svm);
 
        if (svm->vmcb->control.exit_code == SVM_EXIT_ERR) {
@@ -1331,7 +1383,8 @@ static int handle_exit(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
        }
 
        if (is_external_interrupt(svm->vmcb->control.exit_int_info) &&
-           exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR)
+           exit_code != SVM_EXIT_EXCP_BASE + PF_VECTOR &&
+           exit_code != SVM_EXIT_NPF)
                printk(KERN_ERR "%s: unexpected exit_ini_info 0x%x "
                       "exit_code 0x%x\n",
                       __FUNCTION__, svm->vmcb->control.exit_int_info,
@@ -1522,6 +1575,9 @@ static void svm_vcpu_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
        svm->host_dr6 = read_dr6();
        svm->host_dr7 = read_dr7();
        svm->vmcb->save.cr2 = vcpu->arch.cr2;
+       /* required for live migration with NPT */
+       if (npt_enabled)
+               svm->vmcb->save.cr3 = vcpu->arch.cr3;
 
        if (svm->vmcb->save.dr7 & 0xff) {
                write_dr7(0);
@@ -1665,6 +1721,12 @@ static void svm_set_cr3(struct kvm_vcpu *vcpu, unsigned long root)
 {
        struct vcpu_svm *svm = to_svm(vcpu);
 
+       if (npt_enabled) {
+               svm->vmcb->control.nested_cr3 = root;
+               force_new_asid(vcpu);
+               return;
+       }
+
        svm->vmcb->save.cr3 = root;
        force_new_asid(vcpu);