KVM: SVM: Add VMRUN handler
authorAlexander Graf <agraf@suse.de>
Tue, 25 Nov 2008 19:17:07 +0000 (20:17 +0100)
committerAvi Kivity <avi@redhat.com>
Tue, 24 Mar 2009 09:02:47 +0000 (11:02 +0200)
This patch implements VMRUN. VMRUN enters a virtual CPU and runs that
in the same context as the normal guest CPU would run.
So basically it is implemented the same way, a normal CPU would do it.

We also prepare all intercepts that get OR'ed with the original
intercepts, as we do not allow a level 2 guest to be intercepted less
than the first level guest.

v2 implements the following improvements:

- fixes the CPL check
- does not allocate iopm when not used
- remembers the host's IF in the HIF bit in the hflags

v3:

- make use of the new permission checking
- add support for V_INTR_MASKING_MASK

v4:

- use host page backed hsave

v5:

- remove IOPM merging code

v6:

- save cr4 so PAE l1 guests work

v7:

- return 0 on vmrun so we check the MSRs too
- fix MSR check to use the correct variable

Acked-by: Joerg Roedel <joro@8bytes.org>
Signed-off-by: Alexander Graf <agraf@suse.de>
Signed-off-by: Avi Kivity <avi@redhat.com>
arch/x86/include/asm/kvm_host.h
arch/x86/kvm/kvm_svm.h
arch/x86/kvm/svm.c

index 29e4157732db6d16e578f5baf47777db0935142d..53779309514abbcd3e75639ce1010fb86df5e8a9 100644 (file)
@@ -740,6 +740,8 @@ enum {
 };
 
 #define HF_GIF_MASK            (1 << 0)
+#define HF_HIF_MASK            (1 << 1)
+#define HF_VINTR_MASK          (1 << 2)
 
 /*
  * Hardware virtualization extension instructions may fault if a
index a0877cac7b9c192abd73a22897f2d266b712e1ea..91673413d8f740b416595b6b451fe709d66456d4 100644 (file)
@@ -43,6 +43,14 @@ struct vcpu_svm {
        u32 *msrpm;
        struct vmcb *hsave;
        u64 hsave_msr;
+
+       u64 nested_vmcb;
+
+       /* These are the merged vectors */
+       u32 *nested_msrpm;
+
+       /* gpa pointers to the real vectors */
+       u64 nested_vmcb_msrpm;
 };
 
 #endif
index a83c94eb577186343eff129c581bc303156389e7..fad187cbfabe1c7c7df002c8070329de84e15e79 100644 (file)
@@ -77,6 +77,11 @@ static inline struct vcpu_svm *to_svm(struct kvm_vcpu *vcpu)
        return container_of(vcpu, struct vcpu_svm, vcpu);
 }
 
+static inline bool is_nested(struct vcpu_svm *svm)
+{
+       return svm->nested_vmcb;
+}
+
 static unsigned long iopm_base;
 
 struct kvm_ldttss_desc {
@@ -601,6 +606,7 @@ static void init_vmcb(struct vcpu_svm *svm)
        }
        force_new_asid(&svm->vcpu);
 
+       svm->nested_vmcb = 0;
        svm->vcpu.arch.hflags = HF_GIF_MASK;
 }
 
@@ -627,6 +633,7 @@ static struct kvm_vcpu *svm_create_vcpu(struct kvm *kvm, unsigned int id)
        struct page *page;
        struct page *msrpm_pages;
        struct page *hsave_page;
+       struct page *nested_msrpm_pages;
        int err;
 
        svm = kmem_cache_zalloc(kvm_vcpu_cache, GFP_KERNEL);
@@ -649,6 +656,11 @@ static struct kvm_vcpu *svm_create_vcpu(struct kvm *kvm, unsigned int id)
        msrpm_pages = alloc_pages(GFP_KERNEL, MSRPM_ALLOC_ORDER);
        if (!msrpm_pages)
                goto uninit;
+
+       nested_msrpm_pages = alloc_pages(GFP_KERNEL, MSRPM_ALLOC_ORDER);
+       if (!nested_msrpm_pages)
+               goto uninit;
+
        svm->msrpm = page_address(msrpm_pages);
        svm_vcpu_init_msrpm(svm->msrpm);
 
@@ -657,6 +669,8 @@ static struct kvm_vcpu *svm_create_vcpu(struct kvm *kvm, unsigned int id)
                goto uninit;
        svm->hsave = page_address(hsave_page);
 
+       svm->nested_msrpm = page_address(nested_msrpm_pages);
+
        svm->vmcb = page_address(page);
        clear_page(svm->vmcb);
        svm->vmcb_pa = page_to_pfn(page) << PAGE_SHIFT;
@@ -687,6 +701,7 @@ static void svm_free_vcpu(struct kvm_vcpu *vcpu)
        __free_page(pfn_to_page(svm->vmcb_pa >> PAGE_SHIFT));
        __free_pages(virt_to_page(svm->msrpm), MSRPM_ALLOC_ORDER);
        __free_page(virt_to_page(svm->hsave));
+       __free_pages(virt_to_page(svm->nested_msrpm), MSRPM_ALLOC_ORDER);
        kvm_vcpu_uninit(vcpu);
        kmem_cache_free(kvm_vcpu_cache, svm);
 }
@@ -1243,6 +1258,123 @@ static int nested_svm_do(struct vcpu_svm *svm,
        return retval;
 }
 
+
+static int nested_svm_vmrun_msrpm(struct vcpu_svm *svm, void *arg1,
+                                 void *arg2, void *opaque)
+{
+       int i;
+       u32 *nested_msrpm = (u32*)arg1;
+       for (i=0; i< PAGE_SIZE * (1 << MSRPM_ALLOC_ORDER) / 4; i++)
+               svm->nested_msrpm[i] = svm->msrpm[i] | nested_msrpm[i];
+       svm->vmcb->control.msrpm_base_pa = __pa(svm->nested_msrpm);
+
+       return 0;
+}
+
+static int nested_svm_vmrun(struct vcpu_svm *svm, void *arg1,
+                           void *arg2, void *opaque)
+{
+       struct vmcb *nested_vmcb = (struct vmcb *)arg1;
+       struct vmcb *hsave = svm->hsave;
+
+       /* nested_vmcb is our indicator if nested SVM is activated */
+       svm->nested_vmcb = svm->vmcb->save.rax;
+
+       /* Clear internal status */
+       svm->vcpu.arch.exception.pending = false;
+
+       /* Save the old vmcb, so we don't need to pick what we save, but
+          can restore everything when a VMEXIT occurs */
+       memcpy(hsave, svm->vmcb, sizeof(struct vmcb));
+       /* We need to remember the original CR3 in the SPT case */
+       if (!npt_enabled)
+               hsave->save.cr3 = svm->vcpu.arch.cr3;
+       hsave->save.cr4 = svm->vcpu.arch.cr4;
+       hsave->save.rip = svm->next_rip;
+
+       if (svm->vmcb->save.rflags & X86_EFLAGS_IF)
+               svm->vcpu.arch.hflags |= HF_HIF_MASK;
+       else
+               svm->vcpu.arch.hflags &= ~HF_HIF_MASK;
+
+       /* Load the nested guest state */
+       svm->vmcb->save.es = nested_vmcb->save.es;
+       svm->vmcb->save.cs = nested_vmcb->save.cs;
+       svm->vmcb->save.ss = nested_vmcb->save.ss;
+       svm->vmcb->save.ds = nested_vmcb->save.ds;
+       svm->vmcb->save.gdtr = nested_vmcb->save.gdtr;
+       svm->vmcb->save.idtr = nested_vmcb->save.idtr;
+       svm->vmcb->save.rflags = nested_vmcb->save.rflags;
+       svm_set_efer(&svm->vcpu, nested_vmcb->save.efer);
+       svm_set_cr0(&svm->vcpu, nested_vmcb->save.cr0);
+       svm_set_cr4(&svm->vcpu, nested_vmcb->save.cr4);
+       if (npt_enabled) {
+               svm->vmcb->save.cr3 = nested_vmcb->save.cr3;
+               svm->vcpu.arch.cr3 = nested_vmcb->save.cr3;
+       } else {
+               kvm_set_cr3(&svm->vcpu, nested_vmcb->save.cr3);
+               kvm_mmu_reset_context(&svm->vcpu);
+       }
+       svm->vmcb->save.cr2 = nested_vmcb->save.cr2;
+       kvm_register_write(&svm->vcpu, VCPU_REGS_RAX, nested_vmcb->save.rax);
+       kvm_register_write(&svm->vcpu, VCPU_REGS_RSP, nested_vmcb->save.rsp);
+       kvm_register_write(&svm->vcpu, VCPU_REGS_RIP, nested_vmcb->save.rip);
+       /* In case we don't even reach vcpu_run, the fields are not updated */
+       svm->vmcb->save.rax = nested_vmcb->save.rax;
+       svm->vmcb->save.rsp = nested_vmcb->save.rsp;
+       svm->vmcb->save.rip = nested_vmcb->save.rip;
+       svm->vmcb->save.dr7 = nested_vmcb->save.dr7;
+       svm->vmcb->save.dr6 = nested_vmcb->save.dr6;
+       svm->vmcb->save.cpl = nested_vmcb->save.cpl;
+
+       /* We don't want a nested guest to be more powerful than the guest,
+          so all intercepts are ORed */
+       svm->vmcb->control.intercept_cr_read |=
+               nested_vmcb->control.intercept_cr_read;
+       svm->vmcb->control.intercept_cr_write |=
+               nested_vmcb->control.intercept_cr_write;
+       svm->vmcb->control.intercept_dr_read |=
+               nested_vmcb->control.intercept_dr_read;
+       svm->vmcb->control.intercept_dr_write |=
+               nested_vmcb->control.intercept_dr_write;
+       svm->vmcb->control.intercept_exceptions |=
+               nested_vmcb->control.intercept_exceptions;
+
+       svm->vmcb->control.intercept |= nested_vmcb->control.intercept;
+
+       svm->nested_vmcb_msrpm = nested_vmcb->control.msrpm_base_pa;
+
+       force_new_asid(&svm->vcpu);
+       svm->vmcb->control.exit_int_info = nested_vmcb->control.exit_int_info;
+       svm->vmcb->control.exit_int_info_err = nested_vmcb->control.exit_int_info_err;
+       svm->vmcb->control.int_ctl = nested_vmcb->control.int_ctl | V_INTR_MASKING_MASK;
+       if (nested_vmcb->control.int_ctl & V_IRQ_MASK) {
+               nsvm_printk("nSVM Injecting Interrupt: 0x%x\n",
+                               nested_vmcb->control.int_ctl);
+       }
+       if (nested_vmcb->control.int_ctl & V_INTR_MASKING_MASK)
+               svm->vcpu.arch.hflags |= HF_VINTR_MASK;
+       else
+               svm->vcpu.arch.hflags &= ~HF_VINTR_MASK;
+
+       nsvm_printk("nSVM exit_int_info: 0x%x | int_state: 0x%x\n",
+                       nested_vmcb->control.exit_int_info,
+                       nested_vmcb->control.int_state);
+
+       svm->vmcb->control.int_vector = nested_vmcb->control.int_vector;
+       svm->vmcb->control.int_state = nested_vmcb->control.int_state;
+       svm->vmcb->control.tsc_offset += nested_vmcb->control.tsc_offset;
+       if (nested_vmcb->control.event_inj & SVM_EVTINJ_VALID)
+               nsvm_printk("Injecting Event: 0x%x\n",
+                               nested_vmcb->control.event_inj);
+       svm->vmcb->control.event_inj = nested_vmcb->control.event_inj;
+       svm->vmcb->control.event_inj_err = nested_vmcb->control.event_inj_err;
+
+       svm->vcpu.arch.hflags |= HF_GIF_MASK;
+
+       return 0;
+}
+
 static int nested_svm_vmloadsave(struct vmcb *from_vmcb, struct vmcb *to_vmcb)
 {
        to_vmcb->save.fs = from_vmcb->save.fs;
@@ -1299,6 +1431,26 @@ static int vmsave_interception(struct vcpu_svm *svm, struct kvm_run *kvm_run)
        return 1;
 }
 
+static int vmrun_interception(struct vcpu_svm *svm, struct kvm_run *kvm_run)
+{
+       nsvm_printk("VMrun\n");
+       if (nested_svm_check_permissions(svm))
+               return 1;
+
+       svm->next_rip = kvm_rip_read(&svm->vcpu) + 3;
+       skip_emulated_instruction(&svm->vcpu);
+
+       if (nested_svm_do(svm, svm->vmcb->save.rax, 0,
+                         NULL, nested_svm_vmrun))
+               return 1;
+
+       if (nested_svm_do(svm, svm->nested_vmcb_msrpm, 0,
+                     NULL, nested_svm_vmrun_msrpm))
+               return 1;
+
+       return 1;
+}
+
 static int stgi_interception(struct vcpu_svm *svm, struct kvm_run *kvm_run)
 {
        if (nested_svm_check_permissions(svm))
@@ -1632,7 +1784,7 @@ static int (*svm_exit_handlers[])(struct vcpu_svm *svm,
        [SVM_EXIT_MSR]                          = msr_interception,
        [SVM_EXIT_TASK_SWITCH]                  = task_switch_interception,
        [SVM_EXIT_SHUTDOWN]                     = shutdown_interception,
-       [SVM_EXIT_VMRUN]                        = invalid_op_interception,
+       [SVM_EXIT_VMRUN]                        = vmrun_interception,
        [SVM_EXIT_VMMCALL]                      = vmmcall_interception,
        [SVM_EXIT_VMLOAD]                       = vmload_interception,
        [SVM_EXIT_VMSAVE]                       = vmsave_interception,
@@ -1939,7 +2091,8 @@ static void svm_vcpu_run(struct kvm_vcpu *vcpu, struct kvm_run *kvm_run)
        svm->host_cr2 = kvm_read_cr2();
        svm->host_dr6 = read_dr6();
        svm->host_dr7 = read_dr7();
-       svm->vmcb->save.cr2 = vcpu->arch.cr2;
+       if (!is_nested(svm))
+               svm->vmcb->save.cr2 = vcpu->arch.cr2;
        /* required for live migration with NPT */
        if (npt_enabled)
                svm->vmcb->save.cr3 = vcpu->arch.cr3;