KVM: nVMX: Exiting from L2 to L1
authorNadav Har'El <nyh@il.ibm.com>
Wed, 25 May 2011 20:11:34 +0000 (23:11 +0300)
committerAvi Kivity <avi@redhat.com>
Tue, 12 Jul 2011 08:45:16 +0000 (11:45 +0300)
This patch implements nested_vmx_vmexit(), called when the nested L2 guest
exits and we want to run its L1 parent and let it handle this exit.

Note that this will not necessarily be called on every L2 exit. L0 may decide
to handle a particular exit on its own, without L1's involvement; In that
case, L0 will handle the exit, and resume running L2, without running L1 and
without calling nested_vmx_vmexit(). The logic for deciding whether to handle
a particular exit in L1 or in L0, i.e., whether to call nested_vmx_vmexit(),
will appear in a separate patch below.

Signed-off-by: Nadav Har'El <nyh@il.ibm.com>
Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
arch/x86/include/asm/vmx.h
arch/x86/kvm/vmx.c

index 37690bd580c83a7278278074da0b9f00f2067aa0..b747773cf83bc44707c5df1d4e59d83dbe3d747a 100644 (file)
@@ -132,6 +132,8 @@ enum vmcs_field {
        GUEST_IA32_PAT_HIGH             = 0x00002805,
        GUEST_IA32_EFER                 = 0x00002806,
        GUEST_IA32_EFER_HIGH            = 0x00002807,
+       GUEST_IA32_PERF_GLOBAL_CTRL     = 0x00002808,
+       GUEST_IA32_PERF_GLOBAL_CTRL_HIGH= 0x00002809,
        GUEST_PDPTR0                    = 0x0000280a,
        GUEST_PDPTR0_HIGH               = 0x0000280b,
        GUEST_PDPTR1                    = 0x0000280c,
@@ -144,6 +146,8 @@ enum vmcs_field {
        HOST_IA32_PAT_HIGH              = 0x00002c01,
        HOST_IA32_EFER                  = 0x00002c02,
        HOST_IA32_EFER_HIGH             = 0x00002c03,
+       HOST_IA32_PERF_GLOBAL_CTRL      = 0x00002c04,
+       HOST_IA32_PERF_GLOBAL_CTRL_HIGH = 0x00002c05,
        PIN_BASED_VM_EXEC_CONTROL       = 0x00004000,
        CPU_BASED_VM_EXEC_CONTROL       = 0x00004002,
        EXCEPTION_BITMAP                = 0x00004004,
index f8dccb9df34bcea35f9b8bb84ff6a4ee34e573dc..7d778e4976ea07638cf68ef4662f6a34b1497191 100644 (file)
@@ -6144,6 +6144,268 @@ static int nested_vmx_run(struct kvm_vcpu *vcpu, bool launch)
        return 1;
 }
 
+/*
+ * On a nested exit from L2 to L1, vmcs12.guest_cr0 might not be up-to-date
+ * because L2 may have changed some cr0 bits directly (CRO_GUEST_HOST_MASK).
+ * This function returns the new value we should put in vmcs12.guest_cr0.
+ * It's not enough to just return the vmcs02 GUEST_CR0. Rather,
+ *  1. Bits that neither L0 nor L1 trapped, were set directly by L2 and are now
+ *     available in vmcs02 GUEST_CR0. (Note: It's enough to check that L0
+ *     didn't trap the bit, because if L1 did, so would L0).
+ *  2. Bits that L1 asked to trap (and therefore L0 also did) could not have
+ *     been modified by L2, and L1 knows it. So just leave the old value of
+ *     the bit from vmcs12.guest_cr0. Note that the bit from vmcs02 GUEST_CR0
+ *     isn't relevant, because if L0 traps this bit it can set it to anything.
+ *  3. Bits that L1 didn't trap, but L0 did. L1 believes the guest could have
+ *     changed these bits, and therefore they need to be updated, but L0
+ *     didn't necessarily allow them to be changed in GUEST_CR0 - and rather
+ *     put them in vmcs02 CR0_READ_SHADOW. So take these bits from there.
+ */
+static inline unsigned long
+vmcs12_guest_cr0(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
+{
+       return
+       /*1*/   (vmcs_readl(GUEST_CR0) & vcpu->arch.cr0_guest_owned_bits) |
+       /*2*/   (vmcs12->guest_cr0 & vmcs12->cr0_guest_host_mask) |
+       /*3*/   (vmcs_readl(CR0_READ_SHADOW) & ~(vmcs12->cr0_guest_host_mask |
+                       vcpu->arch.cr0_guest_owned_bits));
+}
+
+static inline unsigned long
+vmcs12_guest_cr4(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
+{
+       return
+       /*1*/   (vmcs_readl(GUEST_CR4) & vcpu->arch.cr4_guest_owned_bits) |
+       /*2*/   (vmcs12->guest_cr4 & vmcs12->cr4_guest_host_mask) |
+       /*3*/   (vmcs_readl(CR4_READ_SHADOW) & ~(vmcs12->cr4_guest_host_mask |
+                       vcpu->arch.cr4_guest_owned_bits));
+}
+
+/*
+ * prepare_vmcs12 is part of what we need to do when the nested L2 guest exits
+ * and we want to prepare to run its L1 parent. L1 keeps a vmcs for L2 (vmcs12),
+ * and this function updates it to reflect the changes to the guest state while
+ * L2 was running (and perhaps made some exits which were handled directly by L0
+ * without going back to L1), and to reflect the exit reason.
+ * Note that we do not have to copy here all VMCS fields, just those that
+ * could have changed by the L2 guest or the exit - i.e., the guest-state and
+ * exit-information fields only. Other fields are modified by L1 with VMWRITE,
+ * which already writes to vmcs12 directly.
+ */
+void prepare_vmcs12(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
+{
+       /* update guest state fields: */
+       vmcs12->guest_cr0 = vmcs12_guest_cr0(vcpu, vmcs12);
+       vmcs12->guest_cr4 = vmcs12_guest_cr4(vcpu, vmcs12);
+
+       kvm_get_dr(vcpu, 7, (unsigned long *)&vmcs12->guest_dr7);
+       vmcs12->guest_rsp = kvm_register_read(vcpu, VCPU_REGS_RSP);
+       vmcs12->guest_rip = kvm_register_read(vcpu, VCPU_REGS_RIP);
+       vmcs12->guest_rflags = vmcs_readl(GUEST_RFLAGS);
+
+       vmcs12->guest_es_selector = vmcs_read16(GUEST_ES_SELECTOR);
+       vmcs12->guest_cs_selector = vmcs_read16(GUEST_CS_SELECTOR);
+       vmcs12->guest_ss_selector = vmcs_read16(GUEST_SS_SELECTOR);
+       vmcs12->guest_ds_selector = vmcs_read16(GUEST_DS_SELECTOR);
+       vmcs12->guest_fs_selector = vmcs_read16(GUEST_FS_SELECTOR);
+       vmcs12->guest_gs_selector = vmcs_read16(GUEST_GS_SELECTOR);
+       vmcs12->guest_ldtr_selector = vmcs_read16(GUEST_LDTR_SELECTOR);
+       vmcs12->guest_tr_selector = vmcs_read16(GUEST_TR_SELECTOR);
+       vmcs12->guest_es_limit = vmcs_read32(GUEST_ES_LIMIT);
+       vmcs12->guest_cs_limit = vmcs_read32(GUEST_CS_LIMIT);
+       vmcs12->guest_ss_limit = vmcs_read32(GUEST_SS_LIMIT);
+       vmcs12->guest_ds_limit = vmcs_read32(GUEST_DS_LIMIT);
+       vmcs12->guest_fs_limit = vmcs_read32(GUEST_FS_LIMIT);
+       vmcs12->guest_gs_limit = vmcs_read32(GUEST_GS_LIMIT);
+       vmcs12->guest_ldtr_limit = vmcs_read32(GUEST_LDTR_LIMIT);
+       vmcs12->guest_tr_limit = vmcs_read32(GUEST_TR_LIMIT);
+       vmcs12->guest_gdtr_limit = vmcs_read32(GUEST_GDTR_LIMIT);
+       vmcs12->guest_idtr_limit = vmcs_read32(GUEST_IDTR_LIMIT);
+       vmcs12->guest_es_ar_bytes = vmcs_read32(GUEST_ES_AR_BYTES);
+       vmcs12->guest_cs_ar_bytes = vmcs_read32(GUEST_CS_AR_BYTES);
+       vmcs12->guest_ss_ar_bytes = vmcs_read32(GUEST_SS_AR_BYTES);
+       vmcs12->guest_ds_ar_bytes = vmcs_read32(GUEST_DS_AR_BYTES);
+       vmcs12->guest_fs_ar_bytes = vmcs_read32(GUEST_FS_AR_BYTES);
+       vmcs12->guest_gs_ar_bytes = vmcs_read32(GUEST_GS_AR_BYTES);
+       vmcs12->guest_ldtr_ar_bytes = vmcs_read32(GUEST_LDTR_AR_BYTES);
+       vmcs12->guest_tr_ar_bytes = vmcs_read32(GUEST_TR_AR_BYTES);
+       vmcs12->guest_es_base = vmcs_readl(GUEST_ES_BASE);
+       vmcs12->guest_cs_base = vmcs_readl(GUEST_CS_BASE);
+       vmcs12->guest_ss_base = vmcs_readl(GUEST_SS_BASE);
+       vmcs12->guest_ds_base = vmcs_readl(GUEST_DS_BASE);
+       vmcs12->guest_fs_base = vmcs_readl(GUEST_FS_BASE);
+       vmcs12->guest_gs_base = vmcs_readl(GUEST_GS_BASE);
+       vmcs12->guest_ldtr_base = vmcs_readl(GUEST_LDTR_BASE);
+       vmcs12->guest_tr_base = vmcs_readl(GUEST_TR_BASE);
+       vmcs12->guest_gdtr_base = vmcs_readl(GUEST_GDTR_BASE);
+       vmcs12->guest_idtr_base = vmcs_readl(GUEST_IDTR_BASE);
+
+       vmcs12->guest_activity_state = vmcs_read32(GUEST_ACTIVITY_STATE);
+       vmcs12->guest_interruptibility_info =
+               vmcs_read32(GUEST_INTERRUPTIBILITY_INFO);
+       vmcs12->guest_pending_dbg_exceptions =
+               vmcs_readl(GUEST_PENDING_DBG_EXCEPTIONS);
+
+       /* TODO: These cannot have changed unless we have MSR bitmaps and
+        * the relevant bit asks not to trap the change */
+       vmcs12->guest_ia32_debugctl = vmcs_read64(GUEST_IA32_DEBUGCTL);
+       if (vmcs12->vm_entry_controls & VM_EXIT_SAVE_IA32_PAT)
+               vmcs12->guest_ia32_pat = vmcs_read64(GUEST_IA32_PAT);
+       vmcs12->guest_sysenter_cs = vmcs_read32(GUEST_SYSENTER_CS);
+       vmcs12->guest_sysenter_esp = vmcs_readl(GUEST_SYSENTER_ESP);
+       vmcs12->guest_sysenter_eip = vmcs_readl(GUEST_SYSENTER_EIP);
+
+       /* update exit information fields: */
+
+       vmcs12->vm_exit_reason  = vmcs_read32(VM_EXIT_REASON);
+       vmcs12->exit_qualification = vmcs_readl(EXIT_QUALIFICATION);
+
+       vmcs12->vm_exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
+       vmcs12->vm_exit_intr_error_code = vmcs_read32(VM_EXIT_INTR_ERROR_CODE);
+       vmcs12->idt_vectoring_info_field =
+               vmcs_read32(IDT_VECTORING_INFO_FIELD);
+       vmcs12->idt_vectoring_error_code =
+               vmcs_read32(IDT_VECTORING_ERROR_CODE);
+       vmcs12->vm_exit_instruction_len = vmcs_read32(VM_EXIT_INSTRUCTION_LEN);
+       vmcs12->vmx_instruction_info = vmcs_read32(VMX_INSTRUCTION_INFO);
+
+       /* clear vm-entry fields which are to be cleared on exit */
+       if (!(vmcs12->vm_exit_reason & VMX_EXIT_REASONS_FAILED_VMENTRY))
+               vmcs12->vm_entry_intr_info_field &= ~INTR_INFO_VALID_MASK;
+}
+
+/*
+ * A part of what we need to when the nested L2 guest exits and we want to
+ * run its L1 parent, is to reset L1's guest state to the host state specified
+ * in vmcs12.
+ * This function is to be called not only on normal nested exit, but also on
+ * a nested entry failure, as explained in Intel's spec, 3B.23.7 ("VM-Entry
+ * Failures During or After Loading Guest State").
+ * This function should be called when the active VMCS is L1's (vmcs01).
+ */
+void load_vmcs12_host_state(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
+{
+       if (vmcs12->vm_exit_controls & VM_EXIT_LOAD_IA32_EFER)
+               vcpu->arch.efer = vmcs12->host_ia32_efer;
+       if (vmcs12->vm_exit_controls & VM_EXIT_HOST_ADDR_SPACE_SIZE)
+               vcpu->arch.efer |= (EFER_LMA | EFER_LME);
+       else
+               vcpu->arch.efer &= ~(EFER_LMA | EFER_LME);
+       vmx_set_efer(vcpu, vcpu->arch.efer);
+
+       kvm_register_write(vcpu, VCPU_REGS_RSP, vmcs12->host_rsp);
+       kvm_register_write(vcpu, VCPU_REGS_RIP, vmcs12->host_rip);
+       /*
+        * Note that calling vmx_set_cr0 is important, even if cr0 hasn't
+        * actually changed, because it depends on the current state of
+        * fpu_active (which may have changed).
+        * Note that vmx_set_cr0 refers to efer set above.
+        */
+       kvm_set_cr0(vcpu, vmcs12->host_cr0);
+       /*
+        * If we did fpu_activate()/fpu_deactivate() during L2's run, we need
+        * to apply the same changes to L1's vmcs. We just set cr0 correctly,
+        * but we also need to update cr0_guest_host_mask and exception_bitmap.
+        */
+       update_exception_bitmap(vcpu);
+       vcpu->arch.cr0_guest_owned_bits = (vcpu->fpu_active ? X86_CR0_TS : 0);
+       vmcs_writel(CR0_GUEST_HOST_MASK, ~vcpu->arch.cr0_guest_owned_bits);
+
+       /*
+        * Note that CR4_GUEST_HOST_MASK is already set in the original vmcs01
+        * (KVM doesn't change it)- no reason to call set_cr4_guest_host_mask();
+        */
+       vcpu->arch.cr4_guest_owned_bits = ~vmcs_readl(CR4_GUEST_HOST_MASK);
+       kvm_set_cr4(vcpu, vmcs12->host_cr4);
+
+       /* shadow page tables on either EPT or shadow page tables */
+       kvm_set_cr3(vcpu, vmcs12->host_cr3);
+       kvm_mmu_reset_context(vcpu);
+
+       if (enable_vpid) {
+               /*
+                * Trivially support vpid by letting L2s share their parent
+                * L1's vpid. TODO: move to a more elaborate solution, giving
+                * each L2 its own vpid and exposing the vpid feature to L1.
+                */
+               vmx_flush_tlb(vcpu);
+       }
+
+
+       vmcs_write32(GUEST_SYSENTER_CS, vmcs12->host_ia32_sysenter_cs);
+       vmcs_writel(GUEST_SYSENTER_ESP, vmcs12->host_ia32_sysenter_esp);
+       vmcs_writel(GUEST_SYSENTER_EIP, vmcs12->host_ia32_sysenter_eip);
+       vmcs_writel(GUEST_IDTR_BASE, vmcs12->host_idtr_base);
+       vmcs_writel(GUEST_GDTR_BASE, vmcs12->host_gdtr_base);
+       vmcs_writel(GUEST_TR_BASE, vmcs12->host_tr_base);
+       vmcs_writel(GUEST_GS_BASE, vmcs12->host_gs_base);
+       vmcs_writel(GUEST_FS_BASE, vmcs12->host_fs_base);
+       vmcs_write16(GUEST_ES_SELECTOR, vmcs12->host_es_selector);
+       vmcs_write16(GUEST_CS_SELECTOR, vmcs12->host_cs_selector);
+       vmcs_write16(GUEST_SS_SELECTOR, vmcs12->host_ss_selector);
+       vmcs_write16(GUEST_DS_SELECTOR, vmcs12->host_ds_selector);
+       vmcs_write16(GUEST_FS_SELECTOR, vmcs12->host_fs_selector);
+       vmcs_write16(GUEST_GS_SELECTOR, vmcs12->host_gs_selector);
+       vmcs_write16(GUEST_TR_SELECTOR, vmcs12->host_tr_selector);
+
+       if (vmcs12->vm_exit_controls & VM_EXIT_LOAD_IA32_PAT)
+               vmcs_write64(GUEST_IA32_PAT, vmcs12->host_ia32_pat);
+       if (vmcs12->vm_exit_controls & VM_EXIT_LOAD_IA32_PERF_GLOBAL_CTRL)
+               vmcs_write64(GUEST_IA32_PERF_GLOBAL_CTRL,
+                       vmcs12->host_ia32_perf_global_ctrl);
+}
+
+/*
+ * Emulate an exit from nested guest (L2) to L1, i.e., prepare to run L1
+ * and modify vmcs12 to make it see what it would expect to see there if
+ * L2 was its real guest. Must only be called when in L2 (is_guest_mode())
+ */
+static void nested_vmx_vmexit(struct kvm_vcpu *vcpu)
+{
+       struct vcpu_vmx *vmx = to_vmx(vcpu);
+       int cpu;
+       struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
+
+       leave_guest_mode(vcpu);
+       prepare_vmcs12(vcpu, vmcs12);
+
+       cpu = get_cpu();
+       vmx->loaded_vmcs = &vmx->vmcs01;
+       vmx_vcpu_put(vcpu);
+       vmx_vcpu_load(vcpu, cpu);
+       vcpu->cpu = cpu;
+       put_cpu();
+
+       /* if no vmcs02 cache requested, remove the one we used */
+       if (VMCS02_POOL_SIZE == 0)
+               nested_free_vmcs02(vmx, vmx->nested.current_vmptr);
+
+       load_vmcs12_host_state(vcpu, vmcs12);
+
+       /* Update TSC_OFFSET if vmx_adjust_tsc_offset() was used while L2 ran */
+       vmcs_write64(TSC_OFFSET, vmx->nested.vmcs01_tsc_offset);
+
+       /* This is needed for same reason as it was needed in prepare_vmcs02 */
+       vmx->host_rsp = 0;
+
+       /* Unpin physical memory we referred to in vmcs02 */
+       if (vmx->nested.apic_access_page) {
+               nested_release_page(vmx->nested.apic_access_page);
+               vmx->nested.apic_access_page = 0;
+       }
+
+       /*
+        * Exiting from L2 to L1, we're now back to L1 which thinks it just
+        * finished a VMLAUNCH or VMRESUME instruction, so we need to set the
+        * success or failure flag accordingly.
+        */
+       if (unlikely(vmx->fail)) {
+               vmx->fail = 0;
+               nested_vmx_failValid(vcpu, vmcs_read32(VM_INSTRUCTION_ERROR));
+       } else
+               nested_vmx_succeed(vcpu);
+}
+
 static int vmx_check_intercept(struct kvm_vcpu *vcpu,
                               struct x86_instruction_info *info,
                               enum x86_intercept_stage stage)