KVM: x86: nSVM: Improve virtualization of VMCB12 G_PAT
cover KVM: x86: nSVM: Improve virtualization of VMCB12 G_PAT
There are several problems with KVM's virtualization of the G_PAT field when nested paging is enabled in VMCB12.
* The VMCB12 G_PAT field is not checked for validity when emulating VMRUN. (APM volume 2, section 15.25.4: Nested Paging and VMRUN/#VMEXIT)
* RDMSR(PAT) and WRMSR(PAT) from L2 access L1's PAT MSR rather than L2's Guest PAT register. (APM volume 2, section 15.25.2: Replicated State)
* The L2 Guest PAT register is not written back to VMCB12 on #VMEXIT from L2 to L1. (APM volume 3, Section 4: "VMRUN")
* The value of L2's Guest PAT register is not serialized for save/restore when a checkpoint is taken while L2 is active.
Commit 4995a3685f1b ("KVM: SVM: Use a separate vmcb for the nested L2 guest") left this comment in nested_vmcb02_compute_g_pat():
/* FIXME: merge g_pat from vmcb01 and vmcb12. */
This comment makes no sense. It is true that there are now three different PATs to consider: L2's PAT for guest page tables, L1's PAT for the nested page tables mapping L2 guest physical addresses to L1 guest physical addresses, and L0's PAT for the nested page tables mapping L1 guest physical addresses to host physical addresses. However, if there is any "merging" to be done, it would involve the latter two, and would happen during shadow nested page table construction. (For the record, I don't think "merging" the two nested page table PATs is feasible.) In any case, the VMCB12 G_PAT should be copied unmodified into VMCB02.
Maybe the rest of the current implementation is a consistent quirk based on the existing nested_vmcb02_compute_g_pat() code that bypasses L1's request in VMCB12 and copies L1's PAT MSR into vmcb02 instead. However, an L1 hypervisor that does not intercept accesses to the PAT MSR would legitimately be surprised to find that its L2 guest can modify the hypervisor's own PAT!
The commits in this series are in an awkward order, because I didn't want to change nested_vmcb02_compute_g_pat() until I had removed the call site from svm_set_msr().
The first two commits should arguably be one, but I tried to deal with the serialization issue separately from the RDMSR/WRMSR issue, despite the two being intertwined.
I don't like the ugliness of KVM_GET_MSRS saving the L2 Guest PAT register during a checkpoint, but KVM_SET_MSRS restoring the architectural PAT MSR on restore (because when KVM_SET_MSRS is called, L2 is not active). The APM section on replicated state offers a possible out:
While nested paging is enabled, all (guest) references to the state of the paging registers by x86 code (MOV to/from CRn, etc.) read and write the guest copy of the registers
If we consider KVM_{GET,SET}_MSRS not to be "guest" references, we could always access the architected PAT MSR from userspace, and we could grab 64 bits from the SVM nested state header to serialize L2's G_PAT. In some ways, that seems cleaner, but it does mean that KVM_{GET,SET}_MSR will access L1's PAT, which is irrelevant while L2 is active.
Hence, I am posting this series as an RFC.
Jim Mattson (6):
KVM: x86: nSVM: Shuffle guest PAT and PAT MSR in
svm_set_nested_state()
KVM: x86: nSVM: Redirect PAT MSR accesses to gPAT when NPT is enabled
in vmcb12
KVM: x86: nSVM: Copy current vmcb02 g_pat to vmcb12 g_pat on #VMEXIT
KVM: x86: nSVM: Cache g_pat in vmcb_ctrl_area_cached
KVM: x86: nSVM: Add validity check for the VMCB12 g_pat
KVM: x86: nSVM: Use cached VMCB12 g_pat in VMCB02 when using NPT
arch/x86/include/uapi/asm/kvm.h | 2 ++ arch/x86/kvm/svm/nested.c | 35 +++++++++++++++++++++++++++++++-- arch/x86/kvm/svm/svm.c | 25 +++++++++++++++-------- arch/x86/kvm/svm/svm.h | 1 + 4 files changed, 53 insertions(+), 10 deletions(-)
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
RFC PATCH 1/6 KVM: x86: nSVM: Shuffle guest PAT and PAT MSR in svm_set_nested_state()
When L2 is active and using nested paging, accesses to the PAT MSR should be redirected to the Guest PAT register. As a result, KVM_GET_MSRS will save the Guest PAT register rather than the PAT MSR. However, on restore, KVM_SET_MSRS is called before KVM_SET_NESTED_STATE, so the Guest PAT register will be restored to the PAT MSR.
To fix the serialization of the Guest PAT register and the PAT MSR, copy the PAT MSR to the Guest PAT register (vmcb02->save.g_pat) and copy vmcb01->save.g_pat to the PAT MSR in svm_set_nested_state() under the right conditions. One of these conditions is a new SVM nested state flag, which will be set in the commit that modifies the KVM_{GET,SET}_MSRS semantics.
--- arch/x86/include/uapi/asm/kvm.h | 2 ++ arch/x86/kvm/svm/nested.c | 15 +++++++++++++++ 2 files changed, 17 insertions(+)
arch/x86/include/uapi/asm/kvm.h +2 -0
@@ -494,6 +494,7 @@ struct kvm_sync_regs {#define KVM_STATE_NESTED_SVM_VMCB_SIZE 0x1000#define KVM_STATE_VMX_PREEMPTION_TIMER_DEADLINE 0x00000001 +#define KVM_STATE_SVM_VALID_GPAT 0x00000001/* vendor-independent attributes for system fd (group 0) */ #define KVM_X86_GRP_SYSTEM 0
@@ -529,6 +530,7 @@ struct kvm_svm_nested_state_data {struct kvm_svm_nested_state_hdr {__u64 vmcb_pa; +__u32 flags; };/* for KVM_CAP_NESTED_STATE */
arch/x86/kvm/svm/nested.c +15 -0
@@ -1052,6 +1052,7 @@ void svm_copy_vmrun_state(struct vmcb_save_area *to_save, to_save->rsp = from_save->rsp; to_save->rip = from_save->rip; to_save->cpl =0 ; + to_save->g_pat = from_save->g_pat;if (kvm_cpu_cap_has (X86_FEATURE_SHSTK )) { to_save->s_cet = from_save->s_cet;
@@ -1890,6 +1891,20 @@ static int svm_set_nested_state(struct kvm_vcpu *vcpu,if (WARN_ON_ONCE (ret)) goto out_free; +/* +* If nested paging is enabled in vmcb12, then KVM_SET_MSRS restored +* the guest PAT register to the PAT MSR. Move this to the guest PAT +* register (svm->vmcb->save.g_pat) and restore the PAT MSR from +* svm->vmcb01.ptr->save.g_pat). +*/ +if ((kvm_state->hdr.svm.flags &KVM_STATE_SVM_VALID_GPAT ) && +nested_npt_enabled (svm)) { + ret = -EINVAL ; + svm->vmcb->save.g_pat = vcpu->arch.pat; +if (!kvm_pat_valid (svm->vmcb01.ptr->save.g_pat)) + goto out_free; + vcpu->arch.pat = svm->vmcb01.ptr->save.g_pat; + } svm->nested.force_msr_bitmap_recalc = true;kvm_make_request (KVM_REQ_GET_NESTED_STATE_PAGES , vcpu);
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
RFC PATCH 2/6 KVM: x86: nSVM: Redirect PAT MSR accesses to gPAT when NPT is enabled in vmcb12
In reply to Jim MattsonAccording to the APM volume 2, section 15.25.2: "Replicated State,"
While nested paging is enabled, all (guest) references to the state of the paging registers by x86 code (MOV to/from CRn, etc.) read and write the guest copy of the registers.
The PAT MSR is explicitly enumerated as a "paging register" in that section of the APM.
Implement the architected behavior by redirecting PAT MSR accesses from vcpu->arch.pat to svm->vmcb->save.g_pat when L2 is active and nested paging is enabled in vmcb12.
Note that the change in KVM_{GET,SET}_MSRS semantics breaks serialization. Trigger a fixup in svm_set_nested_state() by setting the VALID_GPAT flag in the SVM nested state header.
--- arch/x86/kvm/svm/nested.c | 1 + arch/x86/kvm/svm/svm.c | 25 +++++++++++++++++-------- 2 files changed, 18 insertions(+), 8 deletions(-)
arch/x86/kvm/svm/nested.c +1 -0
@@ -1728,6 +1728,7 @@ static int svm_get_nested_state(struct kvm_vcpu *vcpu,/* First fill in the header and copy it out. */ if (is_guest_mode (vcpu)) { kvm_state.hdr.svm.vmcb_pa = svm->nested.vmcb12_gpa; + kvm_state.hdr.svm.flags =KVM_STATE_SVM_VALID_GPAT ; kvm_state.size +=KVM_STATE_NESTED_SVM_VMCB_SIZE ; kvm_state.flags |=KVM_STATE_NESTED_GUEST_MODE ;
arch/x86/kvm/svm/svm.c +17 -8
@@ -2675,6 +2675,12 @@ static int svm_get_msr(struct kvm_vcpu *vcpu, struct msr_data *msr_info)return 1 ; msr_info->data = svm->tsc_ratio_msr;break ; +case MSR_IA32_CR_PAT : +if (!(is_guest_mode (vcpu) &&nested_npt_enabled (svm))) + msr_info->data = vcpu->arch.pat; +else + msr_info->data = svm->vmcb->save.g_pat; +break ;case MSR_STAR : msr_info->data = svm->vmcb01.ptr->save.star;break ;
@@ -2864,14 +2870,17 @@ static int svm_set_msr(struct kvm_vcpu *vcpu, struct msr_data *msr)break ;case MSR_IA32_CR_PAT : - ret =kvm_set_msr_common (vcpu, msr); -if (ret) -break ; - - svm->vmcb01.ptr->save.g_pat = data; -if (is_guest_mode (vcpu)) -nested_vmcb02_compute_g_pat (svm); -vmcb_mark_dirty (svm->vmcb,VMCB_NPT ); +if (!kvm_pat_valid (data)) +return 1 ; +if (!(is_guest_mode (vcpu) &&nested_npt_enabled (svm))) { + vcpu->arch.pat = data; + svm->vmcb01.ptr->save.g_pat = data; +vmcb_mark_dirty (svm->vmcb01.ptr,VMCB_NPT ); + } +if (is_guest_mode (vcpu)) { + svm->vmcb->save.g_pat = data; +vmcb_mark_dirty (svm->vmcb,VMCB_NPT ); + }break ;case MSR_IA32_SPEC_CTRL : if (!msr->host_initiated &&
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
RFC PATCH 3/6 KVM: x86: nSVM: Copy current vmcb02 g_pat to vmcb12 g_pat on #VMEXIT
In reply to Jim MattsonAccording to the APM volume 3 pseudo-code for "VMRUN," when nested paging is enabled in the VMCB, the guest PAT register is copied into the g_pat field of the VMCB on #VMEXIT.
In KVM, the guest PAT register is the g_pat field of vmcb02. When nested paging is enabled in vmcb12, copy the g_pat field of the vmcb02 to the g_pat field of the vmcb12 in nested_svm_vmexit().
--- arch/x86/kvm/svm/nested.c | 3 +++ 1 file changed, 3 insertions(+)
arch/x86/kvm/svm/nested.c +3 -0
@@ -1126,6 +1126,9 @@ int nested_svm_vmexit(struct vcpu_svm *svm) vmcb12->save.dr6 = svm->vcpu.arch.dr6; vmcb12->save.cpl = vmcb02->save.cpl; +if (nested_npt_enabled (svm)) + vmcb12->save.g_pat = vmcb02->save.g_pat; +if (guest_cpu_cap_has (vcpu,X86_FEATURE_SHSTK )) { vmcb12->save.s_cet = vmcb02->save.s_cet; vmcb12->save.isst_addr = vmcb02->save.isst_addr;
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
RFC PATCH 4/6 KVM: x86: nSVM: Cache g_pat in vmcb_ctrl_area_cached
In reply to Jim MattsonAdd a g_pat field to the vmcb_ctrl_area_cached struct for caching the VMCB12 g_pat at emulated VMRUN. This is a preliminary step to allow for proper validation and handling of the VMCB12 g_pat when nested paging is enabled in VMCB12.
--- arch/x86/kvm/svm/nested.c | 6 ++++++ arch/x86/kvm/svm/svm.h | 1 + 2 files changed, 7 insertions(+)
arch/x86/kvm/svm/nested.c +6 -0
@@ -984,6 +984,12 @@ int nested_svm_vmrun(struct kvm_vcpu *vcpu)nested_copy_vmcb_control_to_cache (svm, &vmcb12->control);nested_copy_vmcb_save_to_cache (svm, &vmcb12->save); +/* +* To facilitate independent validation of the cached state +* save area and the cached control area, we cache the vmcb12 +* g_pat with the cached controls. +*/ + svm->nested.ctl.g_pat = vmcb12->save.g_pat;if (!nested_vmcb_check_save (vcpu) || !nested_vmcb_check_controls (vcpu)) {
arch/x86/kvm/svm/svm.h +1 -0
@@ -176,6 +176,7 @@ struct vmcb_ctrl_area_cached {u64 virt_ext;u32 clean;u64 bus_lock_rip; +u64 g_pat;union {#if IS_ENABLED (CONFIG_HYPERV ) ||IS_ENABLED (CONFIG_KVM_HYPERV )struct hv_vmcb_enlightenments hv_enlightenments;
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
RFC PATCH 5/6 KVM: x86: nSVM: Add validity check for the VMCB12 g_pat
In reply to Jim MattsonWhen nested paging is enabled for VMCB12, an invalid g_pat causes an immediate #VMEXIT with exit code VMEXIT_INVALID, as specified in the APM.
--- arch/x86/kvm/svm/nested.c | 4 ++++ 1 file changed, 4 insertions(+)
arch/x86/kvm/svm/nested.c +4 -0
@@ -336,6 +336,10 @@ static bool __nested_vmcb_check_controls(struct kvm_vcpu *vcpu,if (CC ((control->nested_ctl &SVM_NESTED_CTL_NP_ENABLE ) && !npt_enabled))return false; +if (CC ((control->nested_ctl &SVM_NESTED_CTL_NP_ENABLE ) && + npt_enabled && !kvm_pat_valid (control->g_pat))) +return false; +if (CC (!nested_svm_check_bitmap_pa (vcpu, control->msrpm_base_pa,MSRPM_SIZE )))return false;
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
RFC PATCH 6/6 KVM: x86: nSVM: Use cached VMCB12 g_pat in VMCB02 when using NPT
In reply to Jim MattsonWhen nested paging is enabled in VMCB12, copy the (cached and validated) VMCB12 g_pat to VMCB02.
--- arch/x86/kvm/svm/nested.c | 6 ++++-- 1 file changed, 4 insertions(+), 2 deletions(-)
arch/x86/kvm/svm/nested.c +4 -2
@@ -605,8 +605,10 @@ void nested_vmcb02_compute_g_pat(struct vcpu_svm *svm)if (!svm->nested.vmcb02.ptr)return ; -/* FIXME: merge g_pat from vmcb01 and vmcb12. */ - svm->nested.vmcb02.ptr->save.g_pat = svm->vmcb01.ptr->save.g_pat; +if (nested_npt_enabled (svm)) + svm->nested.vmcb02.ptr->save.g_pat = svm->nested.ctl.g_pat; +else + svm->nested.vmcb02.ptr->save.g_pat = svm->vmcb01.ptr->save.g_pat; }static void nested_vmcb02_prepare_save (struct vcpu_svm *svm,struct vmcb *vmcb12)
Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog
Quoted 58 lines · Jim Mattson
There are several problems with KVM's virtualization of the
G_PATfield when nested paging is enabled in VMCB12.* The VMCB12
G_PATfield is not checked for validity when emulating VMRUN. (APM volume 2, section 15.25.4: Nested Paging and VMRUN/#VMEXIT)* RDMSR(PAT) and WRMSR(PAT) from L2 access L1's PAT MSR rather than L2's Guest PAT register. (APM volume 2, section 15.25.2: Replicated State)
* The L2 Guest PAT register is not written back to VMCB12 on #VMEXIT from L2 to L1. (APM volume 3, Section 4: "VMRUN")
* The value of L2's Guest PAT register is not serialized for save/restore when a checkpoint is taken while L2 is active.
Commit
4995a3685f1b("KVM: SVM: Use a separate vmcb for the nested L2 guest") left this comment innested_vmcb02_compute_g_pat():/* FIXME: merge g_pat from vmcb01 and vmcb12. */This comment makes no sense. It is true that there are now three different PATs to consider: L2's PAT for guest page tables, L1's PAT for the nested page tables mapping L2 guest physical addresses to L1 guest physical addresses, and L0's PAT for the nested page tables mapping L1 guest physical addresses to host physical addresses. However, if there is any "merging" to be done, it would involve the latter two, and would happen during shadow nested page table construction. (For the record, I don't think "merging" the two nested page table PATs is feasible.) In any case, the VMCB12
G_PATshould be copied unmodified into VMCB02.Maybe the rest of the current implementation is a consistent quirk based on the existing
nested_vmcb02_compute_g_pat()code that bypasses L1's request in VMCB12 and copies L1's PAT MSR into vmcb02 instead. However, an L1 hypervisor that does not intercept accesses to the PAT MSR would legitimately be surprised to find that its L2 guest can modify the hypervisor's own PAT!The commits in this series are in an awkward order, because I didn't want to change
nested_vmcb02_compute_g_pat()until I had removed the call site fromsvm_set_msr().The first two commits should arguably be one, but I tried to deal with the serialization issue separately from the RDMSR/WRMSR issue, despite the two being intertwined.
I don't like the ugliness of
KVM_GET_MSRSsaving the L2 Guest PAT register during a checkpoint, butKVM_SET_MSRSrestoring the architectural PAT MSR on restore (because whenKVM_SET_MSRSis called, L2 is not active). The APM section on replicated state offers a possible out:While nested paging is enabled, all (guest) references to the state of the paging registers by x86 code (MOV to/from CRn, etc.) read and write the guest copy of the registers
Quoted 8 lines
If we consider KVM_{GET,SET}_MSRS not to be "guest" references, we could always access the architected PAT MSR from userspace, and we could grab 64 bits from the SVM nested state header to serialize L2's
G_PAT. In some ways, that seems cleaner, but it does mean that KVM_{GET,SET}_MSR will access L1's PAT, which is irrelevant while L2 is active.Hence, I am posting this series as an RFC.
A little bit more context here, the APM is a bit unclear about how PAT/gPAT are actually handled. Specifically, whether or not gPAT is context switched with the PAT MSR on VMRUN and #VMEXIT or not.
On one hand, the APM mentions that with nested paging, paging registers are replicated and are loaded on VMRUN (in section 15.25.2 in Vol. 2):
Most processor state affecting paging is replicated for host and guest. This includes the paging registers CR0, CR3, CR4, EFER and PAT. CR2 is not replicated but is loaded by VMRUN. The MTRRs are not replicated.
While nested paging is enabled, all (guest) references to the state of the paging registers by x86 code (MOV to/from CRn, etc.) read and write the guest copy of the registers; the VMM's versions of the registers are untouched and continue to control the second level translations from guest physical to system physical addresses. In contrast, when nested paging is disabled, the VMM's paging control registers are stored in the host state save area and the paging control registers from the guest VMCB are the only active versions of those registers.
This gives the impression that gPAT is loaded into the PAT MSR on VMRUN, and switched back on #VMEXIT.
However, the APM also have multiple hints about PAT being different from the other paging registers, and that gPAT might be a hidden register different from PAT MSR:
- Looking at the pseudocode for VMRUN in Vol. 3 in the APM:
* The part with "save host state to physical memory indicated in the VM_HSAVE_PA MSR:" does not include host PAT.
* The part with "from the VMCB at physical address rAX, load guest state:" has:
IF (NP_ENABLE == 1)
gPAT // Leaves host hPAT register unchanged.
- Similarly, the pseudocode for #VMEXIT does not include restoring PAT as part of the host state. Also, the part with "save guest state to VMCB:" specifies "gPAT" not "PAT" being saved to the VMCB.
- In Vol 2, Table B-4: VMSA Layout, State Save Area
for SEV-ES, EFER, CR0, CR3, and CR4 are swap type A, but gPAT has no swap type and a note: "Swapped for guest, not used in host mode."
So it's unclear how the hardware actually handles gPAT. This RFC implements the second interpretation, that gPAT is not loaded into the architerctural PAT MSR on VMRUN, and that guest accesses are redirect to this hidden gPAT register. This complicates things like save/restore, as Jim mentioned.
The alternative would be loading L2's PAT (or the gPAT set by L1 for L2) into the architectural state when entering guest mode, similar to EFER and other paging registers, which would considerably simplify things, especially save/restore. This, however, may not be the architectural behavior.
From the guest prespective it should be more-or-less the same thing, which makes me inclined to the simpler approach. However, I am not sure if there could be any repercussions from not following the architecture, assuming the architecture is really not loading gPAT into the PAT MSR.
It would be great if someone from AMD could clarify the architectural behavior here and voice their opinion about which approach we should take here.
Side-note: If we follow the simple approach, we can probably disable intercepts to PAT completely when NPT is enabled, as guest accesses to PAT are redirected to the gPAT (whatever that is) anyway. It probably won't buy us much tho.
Quoted 20 lines
Jim Mattson (6):
KVM: x86: nSVM: Shuffle guest PAT and PAT MSR in
svm_set_nested_state()
KVM: x86: nSVM: Redirect PAT MSR accesses to gPAT when NPT is enabled
in vmcb12
KVM: x86: nSVM: Copy current vmcb02g_patto vmcb12g_paton #VMEXIT
KVM: x86: nSVM: Cacheg_patinvmcb_ctrl_area_cached
KVM: x86: nSVM: Add validity check for the VMCB12g_pat
KVM: x86: nSVM: Use cached VMCB12g_patin VMCB02 when using NPTarch/x86/include/uapi/asm/kvm.h | 2 ++ arch/x86/kvm/svm/nested.c | 35 +++++++++++++++++++++++++++++++-- arch/x86/kvm/svm/svm.c | 25 +++++++++++++++-------- arch/x86/kvm/svm/svm.h | 1 + 4 files changed, 53 insertions(+), 10 deletions(-)Signature · 1 line
2.51.2.1041.gc1ab5b90ca-goog