KVM: nVMX: Prepare vmcs02 from vmcs01 and vmcs12
authorNadav Har'El <nyh@il.ibm.com>
Wed, 25 May 2011 20:10:02 +0000 (23:10 +0300)
committerAvi Kivity <avi@redhat.com>
Tue, 12 Jul 2011 08:45:14 +0000 (11:45 +0300)
This patch contains code to prepare the VMCS which can be used to actually
run the L2 guest, vmcs02. prepare_vmcs02 appropriately merges the information
in vmcs12 (the vmcs that L1 built for L2) and in vmcs01 (our desires for our
own guests).

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

index 3134638e9e906f0b36ba65577eebae87b098f1de..506c9144074a0ad3efe46a91d153d7b1328a0806 100644 (file)
@@ -345,6 +345,12 @@ struct nested_vmx {
        /* vmcs02_list cache of VMCSs recently used to run L2 guests */
        struct list_head vmcs02_pool;
        int vmcs02_num;
+       u64 vmcs01_tsc_offset;
+       /*
+        * Guest pages referred to in vmcs02 with host-physical pointers, so
+        * we must keep them pinned while L2 runs.
+        */
+       struct page *apic_access_page;
 };
 
 struct vcpu_vmx {
@@ -847,6 +853,18 @@ static inline bool report_flexpriority(void)
        return flexpriority_enabled;
 }
 
+static inline bool nested_cpu_has(struct vmcs12 *vmcs12, u32 bit)
+{
+       return vmcs12->cpu_based_vm_exec_control & bit;
+}
+
+static inline bool nested_cpu_has2(struct vmcs12 *vmcs12, u32 bit)
+{
+       return (vmcs12->cpu_based_vm_exec_control &
+                       CPU_BASED_ACTIVATE_SECONDARY_CONTROLS) &&
+               (vmcs12->secondary_vm_exec_control & bit);
+}
+
 static int __find_msr_index(struct vcpu_vmx *vmx, u32 msr)
 {
        int i;
@@ -1441,6 +1459,22 @@ static void vmx_fpu_activate(struct kvm_vcpu *vcpu)
 
 static void vmx_decache_cr0_guest_bits(struct kvm_vcpu *vcpu);
 
+/*
+ * Return the cr0 value that a nested guest would read. This is a combination
+ * of the real cr0 used to run the guest (guest_cr0), and the bits shadowed by
+ * its hypervisor (cr0_read_shadow).
+ */
+static inline unsigned long nested_read_cr0(struct vmcs12 *fields)
+{
+       return (fields->guest_cr0 & ~fields->cr0_guest_host_mask) |
+               (fields->cr0_read_shadow & fields->cr0_guest_host_mask);
+}
+static inline unsigned long nested_read_cr4(struct vmcs12 *fields)
+{
+       return (fields->guest_cr4 & ~fields->cr4_guest_host_mask) |
+               (fields->cr4_read_shadow & fields->cr4_guest_host_mask);
+}
+
 static void vmx_fpu_deactivate(struct kvm_vcpu *vcpu)
 {
        vmx_decache_cr0_guest_bits(vcpu);
@@ -3438,6 +3472,9 @@ static void set_cr4_guest_host_mask(struct vcpu_vmx *vmx)
        vmx->vcpu.arch.cr4_guest_owned_bits = KVM_CR4_GUEST_OWNED_BITS;
        if (enable_ept)
                vmx->vcpu.arch.cr4_guest_owned_bits |= X86_CR4_PGE;
+       if (is_guest_mode(&vmx->vcpu))
+               vmx->vcpu.arch.cr4_guest_owned_bits &=
+                       ~get_vmcs12(&vmx->vcpu)->cr4_guest_host_mask;
        vmcs_writel(CR4_GUEST_HOST_MASK, ~vmx->vcpu.arch.cr4_guest_owned_bits);
 }
 
@@ -4736,6 +4773,11 @@ static void free_nested(struct vcpu_vmx *vmx)
                vmx->nested.current_vmptr = -1ull;
                vmx->nested.current_vmcs12 = NULL;
        }
+       /* Unpin physical memory we referred to in current vmcs02 */
+       if (vmx->nested.apic_access_page) {
+               nested_release_page(vmx->nested.apic_access_page);
+               vmx->nested.apic_access_page = 0;
+       }
 
        nested_free_all_saved_vmcss(vmx);
 }
@@ -5810,6 +5852,243 @@ static void vmx_set_supported_cpuid(u32 func, struct kvm_cpuid_entry2 *entry)
 {
 }
 
+/*
+ * prepare_vmcs02 is called when the L1 guest hypervisor runs its nested
+ * L2 guest. L1 has a vmcs for L2 (vmcs12), and this function "merges" it
+ * with L0's requirements for its guest (a.k.a. vmsc01), so we can run the L2
+ * guest in a way that will both be appropriate to L1's requests, and our
+ * needs. In addition to modifying the active vmcs (which is vmcs02), this
+ * function also has additional necessary side-effects, like setting various
+ * vcpu->arch fields.
+ */
+static void prepare_vmcs02(struct kvm_vcpu *vcpu, struct vmcs12 *vmcs12)
+{
+       struct vcpu_vmx *vmx = to_vmx(vcpu);
+       u32 exec_control;
+
+       vmcs_write16(GUEST_ES_SELECTOR, vmcs12->guest_es_selector);
+       vmcs_write16(GUEST_CS_SELECTOR, vmcs12->guest_cs_selector);
+       vmcs_write16(GUEST_SS_SELECTOR, vmcs12->guest_ss_selector);
+       vmcs_write16(GUEST_DS_SELECTOR, vmcs12->guest_ds_selector);
+       vmcs_write16(GUEST_FS_SELECTOR, vmcs12->guest_fs_selector);
+       vmcs_write16(GUEST_GS_SELECTOR, vmcs12->guest_gs_selector);
+       vmcs_write16(GUEST_LDTR_SELECTOR, vmcs12->guest_ldtr_selector);
+       vmcs_write16(GUEST_TR_SELECTOR, vmcs12->guest_tr_selector);
+       vmcs_write32(GUEST_ES_LIMIT, vmcs12->guest_es_limit);
+       vmcs_write32(GUEST_CS_LIMIT, vmcs12->guest_cs_limit);
+       vmcs_write32(GUEST_SS_LIMIT, vmcs12->guest_ss_limit);
+       vmcs_write32(GUEST_DS_LIMIT, vmcs12->guest_ds_limit);
+       vmcs_write32(GUEST_FS_LIMIT, vmcs12->guest_fs_limit);
+       vmcs_write32(GUEST_GS_LIMIT, vmcs12->guest_gs_limit);
+       vmcs_write32(GUEST_LDTR_LIMIT, vmcs12->guest_ldtr_limit);
+       vmcs_write32(GUEST_TR_LIMIT, vmcs12->guest_tr_limit);
+       vmcs_write32(GUEST_GDTR_LIMIT, vmcs12->guest_gdtr_limit);
+       vmcs_write32(GUEST_IDTR_LIMIT, vmcs12->guest_idtr_limit);
+       vmcs_write32(GUEST_ES_AR_BYTES, vmcs12->guest_es_ar_bytes);
+       vmcs_write32(GUEST_CS_AR_BYTES, vmcs12->guest_cs_ar_bytes);
+       vmcs_write32(GUEST_SS_AR_BYTES, vmcs12->guest_ss_ar_bytes);
+       vmcs_write32(GUEST_DS_AR_BYTES, vmcs12->guest_ds_ar_bytes);
+       vmcs_write32(GUEST_FS_AR_BYTES, vmcs12->guest_fs_ar_bytes);
+       vmcs_write32(GUEST_GS_AR_BYTES, vmcs12->guest_gs_ar_bytes);
+       vmcs_write32(GUEST_LDTR_AR_BYTES, vmcs12->guest_ldtr_ar_bytes);
+       vmcs_write32(GUEST_TR_AR_BYTES, vmcs12->guest_tr_ar_bytes);
+       vmcs_writel(GUEST_ES_BASE, vmcs12->guest_es_base);
+       vmcs_writel(GUEST_CS_BASE, vmcs12->guest_cs_base);
+       vmcs_writel(GUEST_SS_BASE, vmcs12->guest_ss_base);
+       vmcs_writel(GUEST_DS_BASE, vmcs12->guest_ds_base);
+       vmcs_writel(GUEST_FS_BASE, vmcs12->guest_fs_base);
+       vmcs_writel(GUEST_GS_BASE, vmcs12->guest_gs_base);
+       vmcs_writel(GUEST_LDTR_BASE, vmcs12->guest_ldtr_base);
+       vmcs_writel(GUEST_TR_BASE, vmcs12->guest_tr_base);
+       vmcs_writel(GUEST_GDTR_BASE, vmcs12->guest_gdtr_base);
+       vmcs_writel(GUEST_IDTR_BASE, vmcs12->guest_idtr_base);
+
+       vmcs_write64(GUEST_IA32_DEBUGCTL, vmcs12->guest_ia32_debugctl);
+       vmcs_write32(VM_ENTRY_INTR_INFO_FIELD,
+               vmcs12->vm_entry_intr_info_field);
+       vmcs_write32(VM_ENTRY_EXCEPTION_ERROR_CODE,
+               vmcs12->vm_entry_exception_error_code);
+       vmcs_write32(VM_ENTRY_INSTRUCTION_LEN,
+               vmcs12->vm_entry_instruction_len);
+       vmcs_write32(GUEST_INTERRUPTIBILITY_INFO,
+               vmcs12->guest_interruptibility_info);
+       vmcs_write32(GUEST_ACTIVITY_STATE, vmcs12->guest_activity_state);
+       vmcs_write32(GUEST_SYSENTER_CS, vmcs12->guest_sysenter_cs);
+       vmcs_writel(GUEST_DR7, vmcs12->guest_dr7);
+       vmcs_writel(GUEST_RFLAGS, vmcs12->guest_rflags);
+       vmcs_writel(GUEST_PENDING_DBG_EXCEPTIONS,
+               vmcs12->guest_pending_dbg_exceptions);
+       vmcs_writel(GUEST_SYSENTER_ESP, vmcs12->guest_sysenter_esp);
+       vmcs_writel(GUEST_SYSENTER_EIP, vmcs12->guest_sysenter_eip);
+
+       vmcs_write64(VMCS_LINK_POINTER, -1ull);
+
+       vmcs_write32(PIN_BASED_VM_EXEC_CONTROL,
+               (vmcs_config.pin_based_exec_ctrl |
+                vmcs12->pin_based_vm_exec_control));
+
+       /*
+        * Whether page-faults are trapped is determined by a combination of
+        * 3 settings: PFEC_MASK, PFEC_MATCH and EXCEPTION_BITMAP.PF.
+        * If enable_ept, L0 doesn't care about page faults and we should
+        * set all of these to L1's desires. However, if !enable_ept, L0 does
+        * care about (at least some) page faults, and because it is not easy
+        * (if at all possible?) to merge L0 and L1's desires, we simply ask
+        * to exit on each and every L2 page fault. This is done by setting
+        * MASK=MATCH=0 and (see below) EB.PF=1.
+        * Note that below we don't need special code to set EB.PF beyond the
+        * "or"ing of the EB of vmcs01 and vmcs12, because when enable_ept,
+        * vmcs01's EB.PF is 0 so the "or" will take vmcs12's value, and when
+        * !enable_ept, EB.PF is 1, so the "or" will always be 1.
+        *
+        * A problem with this approach (when !enable_ept) is that L1 may be
+        * injected with more page faults than it asked for. This could have
+        * caused problems, but in practice existing hypervisors don't care.
+        * To fix this, we will need to emulate the PFEC checking (on the L1
+        * page tables), using walk_addr(), when injecting PFs to L1.
+        */
+       vmcs_write32(PAGE_FAULT_ERROR_CODE_MASK,
+               enable_ept ? vmcs12->page_fault_error_code_mask : 0);
+       vmcs_write32(PAGE_FAULT_ERROR_CODE_MATCH,
+               enable_ept ? vmcs12->page_fault_error_code_match : 0);
+
+       if (cpu_has_secondary_exec_ctrls()) {
+               u32 exec_control = vmx_secondary_exec_control(vmx);
+               if (!vmx->rdtscp_enabled)
+                       exec_control &= ~SECONDARY_EXEC_RDTSCP;
+               /* Take the following fields only from vmcs12 */
+               exec_control &= ~SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES;
+               if (nested_cpu_has(vmcs12,
+                               CPU_BASED_ACTIVATE_SECONDARY_CONTROLS))
+                       exec_control |= vmcs12->secondary_vm_exec_control;
+
+               if (exec_control & SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES) {
+                       /*
+                        * Translate L1 physical address to host physical
+                        * address for vmcs02. Keep the page pinned, so this
+                        * physical address remains valid. We keep a reference
+                        * to it so we can release it later.
+                        */
+                       if (vmx->nested.apic_access_page) /* shouldn't happen */
+                               nested_release_page(vmx->nested.apic_access_page);
+                       vmx->nested.apic_access_page =
+                               nested_get_page(vcpu, vmcs12->apic_access_addr);
+                       /*
+                        * If translation failed, no matter: This feature asks
+                        * to exit when accessing the given address, and if it
+                        * can never be accessed, this feature won't do
+                        * anything anyway.
+                        */
+                       if (!vmx->nested.apic_access_page)
+                               exec_control &=
+                                 ~SECONDARY_EXEC_VIRTUALIZE_APIC_ACCESSES;
+                       else
+                               vmcs_write64(APIC_ACCESS_ADDR,
+                                 page_to_phys(vmx->nested.apic_access_page));
+               }
+
+               vmcs_write32(SECONDARY_VM_EXEC_CONTROL, exec_control);
+       }
+
+
+       /*
+        * Set host-state according to L0's settings (vmcs12 is irrelevant here)
+        * Some constant fields are set here by vmx_set_constant_host_state().
+        * Other fields are different per CPU, and will be set later when
+        * vmx_vcpu_load() is called, and when vmx_save_host_state() is called.
+        */
+       vmx_set_constant_host_state();
+
+       /*
+        * HOST_RSP is normally set correctly in vmx_vcpu_run() just before
+        * entry, but only if the current (host) sp changed from the value
+        * we wrote last (vmx->host_rsp). This cache is no longer relevant
+        * if we switch vmcs, and rather than hold a separate cache per vmcs,
+        * here we just force the write to happen on entry.
+        */
+       vmx->host_rsp = 0;
+
+       exec_control = vmx_exec_control(vmx); /* L0's desires */
+       exec_control &= ~CPU_BASED_VIRTUAL_INTR_PENDING;
+       exec_control &= ~CPU_BASED_VIRTUAL_NMI_PENDING;
+       exec_control &= ~CPU_BASED_TPR_SHADOW;
+       exec_control |= vmcs12->cpu_based_vm_exec_control;
+       /*
+        * Merging of IO and MSR bitmaps not currently supported.
+        * Rather, exit every time.
+        */
+       exec_control &= ~CPU_BASED_USE_MSR_BITMAPS;
+       exec_control &= ~CPU_BASED_USE_IO_BITMAPS;
+       exec_control |= CPU_BASED_UNCOND_IO_EXITING;
+
+       vmcs_write32(CPU_BASED_VM_EXEC_CONTROL, exec_control);
+
+       /* EXCEPTION_BITMAP and CR0_GUEST_HOST_MASK should basically be the
+        * bitwise-or of what L1 wants to trap for L2, and what we want to
+        * trap. Note that CR0.TS also needs updating - we do this later.
+        */
+       update_exception_bitmap(vcpu);
+       vcpu->arch.cr0_guest_owned_bits &= ~vmcs12->cr0_guest_host_mask;
+       vmcs_writel(CR0_GUEST_HOST_MASK, ~vcpu->arch.cr0_guest_owned_bits);
+
+       /* Note: IA32_MODE, LOAD_IA32_EFER are modified by vmx_set_efer below */
+       vmcs_write32(VM_EXIT_CONTROLS,
+               vmcs12->vm_exit_controls | vmcs_config.vmexit_ctrl);
+       vmcs_write32(VM_ENTRY_CONTROLS, vmcs12->vm_entry_controls |
+               (vmcs_config.vmentry_ctrl & ~VM_ENTRY_IA32E_MODE));
+
+       if (vmcs12->vm_entry_controls & VM_ENTRY_LOAD_IA32_PAT)
+               vmcs_write64(GUEST_IA32_PAT, vmcs12->guest_ia32_pat);
+       else if (vmcs_config.vmentry_ctrl & VM_ENTRY_LOAD_IA32_PAT)
+               vmcs_write64(GUEST_IA32_PAT, vmx->vcpu.arch.pat);
+
+
+       set_cr4_guest_host_mask(vmx);
+
+       vmcs_write64(TSC_OFFSET,
+               vmx->nested.vmcs01_tsc_offset + vmcs12->tsc_offset);
+
+       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.
+                */
+               vmcs_write16(VIRTUAL_PROCESSOR_ID, vmx->vpid);
+               vmx_flush_tlb(vcpu);
+       }
+
+       if (vmcs12->vm_entry_controls & VM_ENTRY_LOAD_IA32_EFER)
+               vcpu->arch.efer = vmcs12->guest_ia32_efer;
+       if (vmcs12->vm_entry_controls & VM_ENTRY_IA32E_MODE)
+               vcpu->arch.efer |= (EFER_LMA | EFER_LME);
+       else
+               vcpu->arch.efer &= ~(EFER_LMA | EFER_LME);
+       /* Note: modifies VM_ENTRY/EXIT_CONTROLS and GUEST/HOST_IA32_EFER */
+       vmx_set_efer(vcpu, vcpu->arch.efer);
+
+       /*
+        * This sets GUEST_CR0 to vmcs12->guest_cr0, with possibly a modified
+        * TS bit (for lazy fpu) and bits which we consider mandatory enabled.
+        * The CR0_READ_SHADOW is what L2 should have expected to read given
+        * the specifications by L1; It's not enough to take
+        * vmcs12->cr0_read_shadow because on our cr0_guest_host_mask we we
+        * have more bits than L1 expected.
+        */
+       vmx_set_cr0(vcpu, vmcs12->guest_cr0);
+       vmcs_writel(CR0_READ_SHADOW, nested_read_cr0(vmcs12));
+
+       vmx_set_cr4(vcpu, vmcs12->guest_cr4);
+       vmcs_writel(CR4_READ_SHADOW, nested_read_cr4(vmcs12));
+
+       /* shadow page tables on either EPT or shadow page tables */
+       kvm_set_cr3(vcpu, vmcs12->guest_cr3);
+       kvm_mmu_reset_context(vcpu);
+
+       kvm_register_write(vcpu, VCPU_REGS_RSP, vmcs12->guest_rsp);
+       kvm_register_write(vcpu, VCPU_REGS_RIP, vmcs12->guest_rip);
+}
+
 static int vmx_check_intercept(struct kvm_vcpu *vcpu,
                               struct x86_instruction_info *info,
                               enum x86_intercept_stage stage)