Some code refactor surround CR4.UMIP virtualization
cover Some code refactor surround CR4.UMIP virtualization
This is some small code refactor/cleanup surround CR4.UMIP virtualization.
vmx_umip_emulated() is misleading, it is actually cpu_has_vmx_desc(). E.g., if host has both UMIP cap and VMX_FEATURE_DESC_EXITING feature, vmx_umip_emulated() returns true, but UMIP in fact isn't emulated. Rename it.
Also, vCPU can have UMIP feature (i.e. CR4.UMIP is valid) if and only if host has UMIP and/or VMX_FEATURE_DESC_EXITING cap, (see vmx_set_cpu_caps()); then checking cpu_has_vmx_desc() for emulating UMIP is redundant, because boot_cpu_has(X86_FEATURE_UMIP) and cpu_has_vmx_desc() must at least one be true.
===Test===
kvm-unit-test umip cases, enabled and disabled (unittests.cfg patch below), pass. But currently I have no machine with UMIP feature, so actually only emulation path is covered.
In January, test was carried out on Simics platform with UMIP. Passed.
x86/unittests.cfg +4 -0
@@ -323,6 +323,10 @@ arch = x86_64 file = umip.flat extra_params = -cpu qemu64,+umip +[umip-disabled] +file = umip.flat +extra_params = -cpu qemu64,-umip +
Robert Hoo (3):
KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in
vmx_set_cr4()
KVM: VMX: Use the canonical interface to read CR4.UMIP bit
arch/x86/kvm/vmx/capabilities.h | 2 +- arch/x86/kvm/vmx/nested.c | 4 ++-- arch/x86/kvm/vmx/vmx.c | 12 +++++++++--- 3 files changed, 12 insertions(+), 6 deletions(-)
base-commit: 45dd9bc75d9adc9483f0c7d662ba6e73ed698a0b
Signature · 1 line
2.31.1
PATCH 1/3 KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
Just rename, no functional changes intended.
vmx_umip_emulated() comes from the ancient time when there was a kvm_x86_ops::umip_emulated(), which originally simply returned false. (66336cab3531d "KVM: x86: add support for emulating UMIP"). Afterwards, its body changed and moved from vmx.c to the dedicated capabilities.h, but kept its old name which looks weired among cpu_has_vmx_XXX() group.
Rename it to align with other analogous functions, the new name is more accurate for what it does. And, vmcs_config.cpu_based_2nd_exec_ctrl & SECONDARY_EXEC_DESC just means it has the capability of emulating UMIP, not *umip-being-emulated*, e.g. if host has caps of UMIP, it's actually not emulated. On the other hand, UMIP concerned instructions are just subset of those SECONDARY_EXEC_DESC intercepts [1][2].
[1] SDM. Vol.3 Table 25-7. Definitions of Secondary Processor-Based VM-Execution Controls SECONDARY_EXEC_DESC "determines whether executions of LGDT, LIDT, LLDT, LTR, SGDT, SIDT, SLDT, and STR cause VM exits."
[2] SDM. Vol.3 2.5 Control Registers CR4.UMIP is about SGDT, SIDT, SLDT, SMSW, and STR.
--- arch/x86/kvm/vmx/capabilities.h | 2 +- arch/x86/kvm/vmx/nested.c | 4 ++-- arch/x86/kvm/vmx/vmx.c | 10 ++++++++-- 3 files changed, 11 insertions(+), 5 deletions(-)
arch/x86/kvm/vmx/capabilities.h +1 -1
@@ -150,7 +150,7 @@ static inline bool cpu_has_vmx_ept(void)SECONDARY_EXEC_ENABLE_EPT ; } -static inline bool vmx_umip_emulated (void ) +static inline bool cpu_has_vmx_desc (void ) {return vmcs_config.cpu_based_2nd_exec_ctrl &SECONDARY_EXEC_DESC ;
arch/x86/kvm/vmx/nested.c +2 -2
@@ -2322,7 +2322,7 @@ static void prepare_vmcs02_early(struct vcpu_vmx *vmx, struct loaded_vmcs *vmcs0* Preset *DT exiting when emulating UMIP, so that vmx_set_cr4() * will not have to rewrite the controls just for this bit. */ -if (!boot_cpu_has (X86_FEATURE_UMIP ) &&vmx_umip_emulated () && +if (!boot_cpu_has (X86_FEATURE_UMIP ) &&cpu_has_vmx_desc () && (vmcs12->guest_cr4 &X86_CR4_UMIP )) exec_control |=SECONDARY_EXEC_DESC ;
@@ -6984,7 +6984,7 @@ void nested_vmx_setup_ctls_msrs(struct vmcs_config *vmcs_conf, u32 ept_caps)rdmsrl (MSR_IA32_VMX_CR0_FIXED1 , msrs->cr0_fixed1);rdmsrl (MSR_IA32_VMX_CR4_FIXED1 , msrs->cr4_fixed1); -if (vmx_umip_emulated ()) +if (cpu_has_vmx_desc ()) msrs->cr4_fixed1 |=X86_CR4_UMIP ; msrs->vmcs_enum =nested_vmx_calc_vmcs_enum_msr ();
arch/x86/kvm/vmx/vmx.c +8 -2
@@ -3431,7 +3431,13 @@ void vmx_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)else hw_cr4 |=KVM_PMODE_VM_CR4_ALWAYS_ON ; -if (!boot_cpu_has (X86_FEATURE_UMIP ) &&vmx_umip_emulated ()) { +/* +* Emulate UMIP via enable secondary_exec_control.DESC +* It can get here means it has passed valid_cr4() check, i.e. +* guest been exposed with UMIP feature, i.e. either host has cap +* of UMIP or vmx_set_cpu_caps() set it because of cpu_has_vmx_desc() +*/ +if (!boot_cpu_has (X86_FEATURE_UMIP ) &&cpu_has_vmx_desc ()) {if (cr4 &X86_CR4_UMIP ) {secondary_exec_controls_setbit (vmx,SECONDARY_EXEC_DESC ); hw_cr4 &= ~X86_CR4_UMIP ;
@@ -7820,7 +7826,7 @@ static __init void vmx_set_cpu_caps(void)kvm_cpu_cap_clear (X86_FEATURE_SGX2 ); } -if (vmx_umip_emulated ()) +if (cpu_has_vmx_desc ())kvm_cpu_cap_set (X86_FEATURE_UMIP );/* CPUID 0xD.1 */
Signature · 1 line
2.31.1
Sean Christopherson
maintainerRe: [PATCH 1/3] KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
Quoted 3 lines · Robert Hoo
Just rename, no functional changes intended.
vmx_umip_emulated()comes from the ancient time when there was a
No, vmx_umip_emulated() comes from the fact that "cpu_has_vmx_desc()" is inscrutable for the relevant code. There is zero reason to require that readers have a priori knowledge of why intercepting descriptor table access instructions is relevant to handing CR4.UMIP changes.
If it really bothers someone, we could do
static inline bool cpu_has_vmx_desc(void)
{
return vmcs_config.cpu_based_2nd_exec_ctrl &
SECONDARY_EXEC_DESC;
}
static inline bool vmx_umip_emulated(void)
{
return cpu_has_vmx_desc();
}but I don't see the point since there is no usage for SECONDARY_EXEC_DESC outside of UMIP emulation.
Re: [PATCH 1/3] KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
Quoted 10 lines · Sean Christopherson <[email protected]> 于2023年3月10日周五 23:59写道:
Quoted 3 lines · Robert Hoo
Just rename, no functional changes intended.
vmx_umip_emulated()comes from the ancient time when there was aNo,
vmx_umip_emulated()comes from the fact that "cpu_has_vmx_desc()" is inscrutable for the relevant code. There is zero reason to require that readers have a priori knowledge of why intercepting descriptor table access instructions is relevant to handing CR4.UMIP changes.
I think this is where comments can play its role.
Quoted 2 lines
If it really bothers someone, we could do
Yeah, below also came to my mind, as one option.
Quoted 12 lines
static inline bool cpu_has_vmx_desc(void)
{
return vmcs_config.cpu_based_2nd_exec_ctrl &
SECONDARY_EXEC_DESC;
}
static inline bool vmx_umip_emulated(void)
{
return cpu_has_vmx_desc();
}Quoted 2 lines
but I don't see the point since there is no usage for
SECONDARY_EXEC_DESCoutside of UMIP emulation.
SECONDARY_EXEC_DESC concerns more than UMIP does. UMIP emulation just leverages part of it.
Also, vmx_umip_emulated() == true doesn't necessarily mean, as its name indicates,
UMIP-being-emulated, e.g. Host has UMIP capability, then UMIP isn't emulated thoughvmx_umip_emulated() indicates true.
Sean Christopherson
maintainerRe: [PATCH 1/3] KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
Please fix your editor or whatever it is that is resulting your emails being wrapped at very bizarre boundaries.
Quoted 3 lines · Robert Hoo
Also,
vmx_umip_emulated()== true doesn't necessarily mean, as its name indicates, UMIP-being-emulated, e.g. Host has UMIP capability, then UMIP isn't emulated thoughvmx_umip_emulated()indicates true.
True. I was going to say "there's no perfect solution" since KVM needs to check both "is UMIP emulated" and "can UMIP be emulated", but that's not actually the case. vmx_emulate_umip() _should_ check for native support, as there's no legitimate use for checking if UMIP _can_ be emulated.
Functionally, this should be a glorified nop, but I agree it's worth changing. I'll formally post this after testing.
From: Sean Christopherson <[email protected]>
Date: Wed, 15 Mar 2023 10:27:40 -0700
Subject: [PATCH] KVM: VMX: Treat UMIP as emulated if and only if the host
doesn't have UMIP
Advertise UMIP as emulated if and only if the host doesn't natively support UMIP, otherwise vmx_umip_emulated() is misleading when the host _does_ support UMIP. Of the four users of vmx_umip_emulated(), two already check for native support, and the logic in vmx_set_cpu_caps() is relevant if and only if UMIP isn't natively supported as UMIP is set in KVM's caps by kvm_set_cpu_caps() when UMIP is present in hardware.
That leaves KVM's stuffing of X86_CR4_UMIP into the default cr4_fixed1 value enumerated for nested VMX. In that case, checking for (lack of) host support is actually a bug fix of sorts, as enumerating UMIP support based solely on descriptor table existing works only because KVM doesn't sanity check MSR_IA32_VMX_CR4_FIXED1. E.g. if a (very theoretical) host supported UMIP in hardware but didn't allow UMIP+VMX, KVM would advertise UMIP but not actually emulate UMIP. Of course, KVM would explode long before it could run a nested VM on said theoretical CPU, as KVM doesn't modify host CR4 when enabling VMX.
--- arch/x86/kvm/vmx/capabilities.h | 4 ++-- arch/x86/kvm/vmx/nested.c | 3 +-- arch/x86/kvm/vmx/vmx.c | 2 +- 3 files changed, 4 insertions(+), 5 deletions(-)
arch/x86/kvm/vmx/capabilities.h +2 -2
@@ -152,8 +152,8 @@ static inline bool cpu_has_vmx_ept(void)static inline bool vmx_umip_emulated (void ) { -return vmcs_config.cpu_based_2nd_exec_ctrl & -SECONDARY_EXEC_DESC ; +return !boot_cpu_has (X86_FEATURE_UMIP ) && + (vmcs_config.cpu_based_2nd_exec_ctrl &SECONDARY_EXEC_DESC ); }static inline bool cpu_has_vmx_rdtscp (void )
arch/x86/kvm/vmx/nested.c +1 -2
@@ -2322,8 +2322,7 @@ static void prepare_vmcs02_early(struct vcpu_vmx *vmx, struct loaded_vmcs *vmcs0* Preset *DT exiting when emulating UMIP, so that vmx_set_cr4() * will not have to rewrite the controls just for this bit. */ -if (!boot_cpu_has (X86_FEATURE_UMIP ) &&vmx_umip_emulated () && - (vmcs12->guest_cr4 &X86_CR4_UMIP )) +if (vmx_umip_emulated () && (vmcs12->guest_cr4 &X86_CR4_UMIP )) exec_control |=SECONDARY_EXEC_DESC ;if (exec_control &SECONDARY_EXEC_VIRTUAL_INTR_DELIVERY )
arch/x86/kvm/vmx/vmx.c +1 -1
@@ -3467,7 +3467,7 @@ void vmx_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)else hw_cr4 |=KVM_PMODE_VM_CR4_ALWAYS_ON ; -if (!boot_cpu_has (X86_FEATURE_UMIP ) &&vmx_umip_emulated ()) { +if (vmx_umip_emulated ()) {if (cr4 &X86_CR4_UMIP ) {secondary_exec_controls_setbit (vmx,SECONDARY_EXEC_DESC ); hw_cr4 &= ~X86_CR4_UMIP ;
base-commit: 0da2a674e56a7fb429eb1f96a3da04d56ec167fd
Re: [PATCH 1/3] KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
Quoted 4 lines · Sean Christopherson <[email protected]> 于2023年3月16日周四 01:50写道:
Please fix your editor or whatever it is that is resulting your emails being wrapped at very bizarre boundaries.
(Sorry for late reply.)
Yes, I also noticed this. Just began using Gmail web portal for community mails. I worried that it has no auto wrapping (didn't find the setting), so manually wrapped; but now looks like it has some.
Give me some time, I'm going to switch to some mail client.
Welcome suggestions of mail clients which is suited for community communications, on Windows platform.🙂
Quoted 5 lines · Robert Hoo
Quoted 3 lines
Also,
vmx_umip_emulated()== true doesn't necessarily mean, as its name indicates, UMIP-being-emulated, e.g. Host has UMIP capability, then UMIP isn't emulated thoughvmx_umip_emulated()indicates true.
[...]
Quoted 2 lines
there's no
legitimate use for checking if UMIP _can_ be emulated.
Agree.
Quoted 2 lines
Functionally, this should be a glorified nop,
Personally, I think it is more than this; good naming, not misleading its user, saving their time to look into, is always good.
Quoted 2 lines
but I agree it's worth changing.
I'll formally post this after testing.
Thanks, looks good to me, it enriches the original patch, i.e. the cr4_fixed1 part is beyond my original findings. Just don't quite understand the last paragraph of the description. I ask inline below.
Quoted 6 lines
From: Sean Christopherson <[email protected]>
Date: Wed, 15 Mar 2023 10:27:40 -0700
Subject: [PATCH] KVM: VMX: Treat UMIP as emulated if and only if the host
doesn't have UMIP
Quoted 7 lines
Advertise UMIP as emulated if and only if the host doesn't natively support UMIP, otherwise
vmx_umip_emulated()is misleading when the host _does_ support UMIP. Of the four users ofvmx_umip_emulated(), two already check for native support, and the logic invmx_set_cpu_caps()is relevant if and only if UMIP isn't natively supported as UMIP is set in KVM's caps bykvm_set_cpu_caps()when UMIP is present in hardware.
Sorry I don't fully understand the paragraph below, though I believe you're right.:)
Quoted 3 lines
That leaves KVM's stuffing of
X86_CR4_UMIPinto the defaultcr4_fixed1value enumerated for nested VMX. In that case, checking for (lack of) host support is actually a bug fix of sorts,
What bug?
By your assumption below:
* host supports UMIP, host doesn't allow (nested?) vmx
* UMIP enumerated to L1, L1 thinks it has UMIP capability and
enumerates to L2
* L1 MSR_IA32_VMX_CR4_FIXED1.UMIP is set (meaning allow 1, not fixed to 1)
L2 can use UMIP, no matter L1's UMIP capability is backed by L0 host's HW UMIP or L0's vmx emulation, I don't see any problem. Shed more light?
Quoted 2 lines
as enumerating UMIP support
based solely on descriptor table
What "descriptor table" do you mean here?
Quoted 2 lines
existing works only because KVM doesn't
sanity checkMSR_IA32_VMX_CR4_FIXED1.
Emm, nested_cr4_valid() should be applied to vmx_set_cr4()?
Quoted 5 lines
E.g. if a (very theoretical) host
supported UMIP in hardware but didn't allow UMIP+VMX, KVM would advertise UMIP but not actually emulate UMIP. Of course, KVM would explode long before it could run a nested VM on said theoretical CPU, as KVM doesn't modify host CR4 when enabling VMX.
Sean Christopherson
maintainerRe: [PATCH 1/3] KVM: VMX: Rename vmx_umip_emulated() to cpu_has_vmx_desc()
Quoted 5 lines · Robert Hoo
Quoted 4 lines · Sean Christopherson <[email protected]> 于2023年3月16日周四 01:50写道:
Please fix your editor or whatever it is that is resulting your emails being wrapped at very bizarre boundaries.
Quoted 7 lines
(Sorry for late reply.)
Yes, I also noticed this. Just began using Gmail web portal for community mails. I worried that it has no auto wrapping (didn't find the setting), so manually wrapped; but now looks like it has some.
Give me some time, I'm going to switch to some mail client.
Welcome suggestions of mail clients which is suited for community communications, on Windows platform.🙂
Heh, none? The "on Windows" is a bit problematic. Sorry I can't help on this front.
Quoted 11 lines
Quoted 3 lines
That leaves KVM's stuffing of
X86_CR4_UMIPinto the defaultcr4_fixed1value enumerated for nested VMX. In that case, checking for (lack of) host support is actually a bug fix of sorts,What bug?
By your assumption below:
* host supports UMIP, host doesn't allow (nested?) vmx
* UMIP enumerated to L1, L1 thinks it has UMIP capability and
enumerates to L2
* L1 MSR_IA32_VMX_CR4_FIXED1.UMIP is set (meaning allow 1, not fixed to 1)
Quoted 8 lines
L2 can use UMIP, no matter L1's UMIP capability is backed by L0 host's HW UMIP or L0's vmx emulation, I don't see any problem. Shed more light?
Quoted 2 lines
as enumerating UMIP support
based solely on descriptor tableWhat "descriptor table" do you mean here?
There's a typo below. It should be "exiting", not "existing". As in "descriptor table exiting", i.e. the feature that allows KVM to intercept LGDT and friends.
Quoted 4 lines
Quoted 2 lines
existing works only because KVM doesn't
sanity checkMSR_IA32_VMX_CR4_FIXED1.Emm,
nested_cr4_valid()should be applied tovmx_set_cr4()?
No, what this is saying is that if a (virtual) CPU does support UMIP for bare metal (from the host's perspective), but does NOT allow UMIP to be set in a VMX guest's CR4, then KVM would end up advertising UMIP to L1 but would neither virtualize (set in hardware) nor emulate UMIP for L1.
The blurb about KVM exploding is calling out that in this very, very theoretical scenario, KVM will fail on the very first VM-Entry (if not before) as KVM uses the host kernel's CR4 verbatim when setting vmcs.HOST_CR4, i.e. will fail the consistency check on a "cannot be 1" bits being set in HOST_CR4.
Quoted 4 lines
Quoted 4 lines
E.g. if a (very theoretical) host supported UMIP in hardware but didn't allow UMIP+VMX, KVM would advertise UMIP but not actually emulate UMIP. Of course, KVM would explode long before it could run a nested VM on said theoretical CPU, as KVM doesn't modify host CR4 when enabling VMX.
PATCH 2/3 KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
In reply to Robert HooRemove the unnecessary cpu_has_vmx_desc() check for emulating UMIP.
When it arrives vmx_set_cr4(), it has passed kvm_is_valid_cr4(), meaning vmx_set_cpu_caps() has already assign X86_FEATURE_UMIP with this vcpu, because of either host CPU has the capability or it can be emulated,
vmx_set_cpu_caps():
if (cpu_has_vmx_desc()) kvm_cpu_cap_set(X86_FEATURE_UMIP);
that is to say, if !boot_cpu_has(X86_FEATURE_UMIP) here, cpu_has_vmx_desc() must be true, it should be a meaningless check here.
--- arch/x86/kvm/vmx/vmx.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kvm/vmx/vmx.c +1 -1
@@ -3437,7 +3437,7 @@ void vmx_set_cr4(struct kvm_vcpu *vcpu, unsigned long cr4)* guest been exposed with UMIP feature, i.e. either host has cap * of UMIP or vmx_set_cpu_caps() set it because of cpu_has_vmx_desc() */ -if (!boot_cpu_has (X86_FEATURE_UMIP ) &&cpu_has_vmx_desc ()) { +if (!boot_cpu_has (X86_FEATURE_UMIP )) {if (cr4 &X86_CR4_UMIP ) {secondary_exec_controls_setbit (vmx,SECONDARY_EXEC_DESC ); hw_cr4 &= ~X86_CR4_UMIP ;
Signature · 1 line
2.31.1
Sean Christopherson
maintainerRe: [PATCH 2/3] KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
Quoted 1 line · Robert Hoo
Remove the unnecessary
cpu_has_vmx_desc()check for emulating UMIP.
It's not unnecessary. See commit 64f7a11586ab ("KVM: vmx: update sec exec controls for UMIP iff emulating UMIP"). Dropping the check will cause KVM to execute
secondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC);
on CPUs that don't have SECONDARY_VM_EXEC_CONTROL.
Re: [PATCH 2/3] KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
Quoted 10 lines · Sean Christopherson <[email protected]> 于2023年3月11日周六 00:12写道:
Quoted 1 line · Robert Hoo
Remove the unnecessary
cpu_has_vmx_desc()check for emulating UMIP.It's not unnecessary. See commit
64f7a11586ab("KVM: vmx: update sec exec controls for UMIP iff emulating UMIP"). Dropping the check will cause KVM to executesecondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC);on CPUs that don't have
SECONDARY_VM_EXEC_CONTROL.
Sorry I don't follow you.
My point is that, given it has passed kvm_is_valid_cr4() (in kvm_set_cr4()), we can assertboot_cpu_has(X86_FEATURE_UMIP) and vmx_umip_emulated() must be at least one true. Therefore when !boot_cpu_has(X86_FEATURE_UMIP), vmx_umip_emulated() must be
true, therefore checking it is redundant.
Do you mean other call path other than kvm_set_cr4() --> vmx_set_cr4()? i.e. vmx_set_cr0() --> vmx_set_cr4()?nested_... --> vmx_set_cr4()?
Emm, they seem don't go through kvm_is_valid_cr4() firstly.
Sean Christopherson
maintainerRe: [PATCH 2/3] KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
Quoted 8 lines · Robert Hoo
Quoted 7 lines · Sean Christopherson <[email protected]> 于2023年3月11日周六 00:12写道:
Quoted 1 line · Robert Hoo
Remove the unnecessary
cpu_has_vmx_desc()check for emulating UMIP.It's not unnecessary. See commit
64f7a11586ab("KVM: vmx: update sec exec controls for UMIP iff emulating UMIP"). Dropping the check will cause KVM to execute
Quoted 8 lines
Quoted 3 lines
secondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC);on CPUs that don't have
SECONDARY_VM_EXEC_CONTROL.Sorry I don't follow you.
My point is that, given it has passedkvm_is_valid_cr4()(inkvm_set_cr4()), we can assertboot_cpu_has(X86_FEATURE_UMIP)andvmx_umip_emulated()must be at least one true.
This assertion is wrong for the case where guest.CR4.UMIP=0. The below code is not guarded with a check on guest.CR4.UMIP. If the vmx_umip_emulated() check goes away and guest.CR4.UMIP=0, KVM will attempt to write secondary controls.
Technically, now that controls_shadow exists, KVM won't actually do a VMWRITE, but I most definitely don't want to rely on controls_shadow for functional correctness. And controls_shadow aside, the "vmx_umip_emulated()" effectively serves as documentation for why KVM is mucking with UMIP when it's obviously not supported in hardware.
if (!boot_cpu_has(X86_FEATURE_UMIP) && vmx_umip_emulated()) {
if (cr4 & X86_CR4_UMIP) {
secondary_exec_controls_setbit(vmx, SECONDARY_EXEC_DESC);
hw_cr4 &= ~X86_CR4_UMIP;
} else if (!is_guest_mode(vcpu) ||
!nested_cpu_has2(get_vmcs12(vcpu), SECONDARY_EXEC_DESC)) {
secondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC);
}
}Re: [PATCH 2/3] KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
Quoted 9 lines · Sean Christopherson <[email protected]> 于2023年3月16日周四 00:36写道:
Quoted 4 lines
Sorry I don't follow you.
My point is that, given it has passedkvm_is_valid_cr4()(inkvm_set_cr4()), we can assertboot_cpu_has(X86_FEATURE_UMIP)andvmx_umip_emulated()must be at least one true.This assertion is wrong for the case where guest.CR4.UMIP=0. The below code is not guarded with a check on guest.CR4.UMIP. If the
vmx_umip_emulated()check goes away and guest.CR4.UMIP=0, KVM will attempt to write secondary controls.
Sorry still don't follow you. Do you mean in nested case? the "guest" above is L1?
Quoted 15 lines
Technically, now that
controls_shadowexists, KVM won't actually do a VMWRITE, but I most definitely don't want to rely oncontrols_shadowfor functional correctness. Andcontrols_shadowaside, the "vmx_umip_emulated()" effectively serves as documentation for why KVM is mucking with UMIP when it's obviously not supported in hardware.if (!boot_cpu_has(X86_FEATURE_UMIP) && vmx_umip_emulated()) { if (cr4 & X86_CR4_UMIP) { secondary_exec_controls_setbit(vmx, SECONDARY_EXEC_DESC); hw_cr4 &= ~X86_CR4_UMIP; } else if (!is_guest_mode(vcpu) || !nested_cpu_has2(get_vmcs12(vcpu), SECONDARY_EXEC_DESC)) { secondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC); } }
Sean Christopherson
maintainerRe: [PATCH 2/3] KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
Quoted 6 lines · Robert Hoo
Quoted 5 lines · Sean Christopherson <[email protected]> 于2023年3月16日周四 00:36写道:
Quoted 4 lines
Sorry I don't follow you.
My point is that, given it has passedkvm_is_valid_cr4()(inkvm_set_cr4()), we can assertboot_cpu_has(X86_FEATURE_UMIP)andvmx_umip_emulated()must be at least one true.
Quoted 7 lines
Quoted 4 lines
This assertion is wrong for the case where guest.CR4.UMIP=0. The below code is not guarded with a check on guest.CR4.UMIP. If the
vmx_umip_emulated()check goes away and guest.CR4.UMIP=0, KVM will attempt to write secondary controls.Sorry still don't follow you. Do you mean in nested case? the "guest" above is L1?
Please take the time to walk through the code with possible inputs/scenarios before asking these types of questions, e.g. if necessary use a whiteboard, pen+paper, etc. I'm happy to explain subtleties and add answer specific questions, but as evidenced by my delayed response, I simply do not have the bandwidth to answer questions where the answer is literally a trace-through of a small, fully contained section of code.
if (!boot_cpu_has(X86_FEATURE_UMIP)) { <= evaluates true when UMIP is NOT supported
if (cr4 & X86_CR4_UMIP) { <= evaluates false when guest.CR4.UMIP == 0
secondary_exec_controls_setbit(vmx, SECONDARY_EXEC_DESC);
hw_cr4 &= ~X86_CR4_UMIP;
} else if (!is_guest_mode(vcpu) || <= evalutes true when L2 is NOT active
!nested_cpu_has2(get_vmcs12(vcpu), SECONDARY_EXEC_DESC)) {
secondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC); <= KVM "blindly" writes secondary controls
}
}Re: [PATCH 2/3] KVM: VMX: Remove a unnecessary cpu_has_vmx_desc() check in vmx_set_cr4()
Quoted 12 lines · Sean Christopherson
Quoted 11 lines · Robert Hoo
Quoted 9 lines · Sean Christopherson <[email protected]> 于2023年3月16日周四 00:36写道:
Quoted 4 lines
Sorry I don't follow you.
My point is that, given it has passedkvm_is_valid_cr4()(inkvm_set_cr4()), we can assertboot_cpu_has(X86_FEATURE_UMIP)andvmx_umip_emulated()must be at least one true.This assertion is wrong for the case where guest.CR4.UMIP=0. The below code is not guarded with a check on guest.CR4.UMIP. If the
vmx_umip_emulated()check goes away and guest.CR4.UMIP=0, KVM will attempt to write secondary controls.
Quoted 8 lines
Quoted 2 lines
Sorry still don't follow you. Do you mean in nested case? the "guest" above is L1?
Please take the time to walk through the code with possible inputs/scenarios before asking these types of questions, e.g. if necessary use a whiteboard, pen+paper, etc. I'm happy to explain subtleties and add answer specific questions, but as evidenced by my delayed response, I simply do not have the bandwidth to answer questions where the answer is literally a trace-through of a small, fully contained section of code.
Sure, fully understand your busyness, thanks for still providing detailed explanation. Sorry for the annoyance.
Given that you've merged host UMIP cap check into vmx_umip_emulated(), I might have not tangled in this point.
We can close this thread of patch set now.
Quoted 10 lines
if (!boot_cpu_has(X86_FEATURE_UMIP)) { <= evaluates true when UMIP is NOT supported
if (cr4 & X86_CR4_UMIP) { <= evaluates false when guest.CR4.UMIP == 0
secondary_exec_controls_setbit(vmx, SECONDARY_EXEC_DESC);
hw_cr4 &= ~X86_CR4_UMIP;
} else if (!is_guest_mode(vcpu) || <= evalutes true when L2 is NOT active
!nested_cpu_has2(get_vmcs12(vcpu), SECONDARY_EXEC_DESC)) {
secondary_exec_controls_clearbit(vmx, SECONDARY_EXEC_DESC); <= KVM "blindly" writes secondary controls
}
}PATCH 3/3 KVM: VMX: Use the canonical interface to read CR4.UMIP bit
In reply to Robert HooUse kvm_read_cr4_bits() rather than directly read vcpu->arch.cr4, now that we have reg cache layer and defined this wrapper.
Although, effectively for CR4.UMIP, it's the same, at present, as it's not guest owned, in case of future changes, here better to use the canonical interface.
--- arch/x86/kvm/vmx/vmx.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
arch/x86/kvm/vmx/vmx.c +1 -1
@@ -5435,7 +5435,7 @@ static int handle_set_cr4(struct kvm_vcpu *vcpu, unsigned long val)static int handle_desc (struct kvm_vcpu *vcpu) { -WARN_ON (!(vcpu->arch.cr4 &X86_CR4_UMIP )); +WARN_ON (!kvm_read_cr4_bits (vcpu,X86_CR4_UMIP ));return kvm_emulate_instruction (vcpu,0 ); }
Signature · 1 line
2.31.1
Sean Christopherson
maintainerRe: [PATCH 3/3] KVM: VMX: Use the canonical interface to read CR4.UMIP bit
Quoted 2 lines · Robert Hoo
Use
kvm_read_cr4_bits()rather than directly readvcpu->arch.cr4, now that we have reg cache layer and defined this wrapper.
kvm_read_cr4_bits() predates this code by ~7 years.
Quoted 3 lines
Although, effectively for CR4.UMIP, it's the same, at present, as it's not guest owned, in case of future changes, here better to use the canonical interface.
Practically speaking, UMIP _can't_ be guest owned without breaking UMIP emulation. I do like not open coding vcpu->arch.cr4, but I don't particuarly like the changelog.
This would also be a good time to opportunistically convert the WARN_ON() to a WARN_ON_ONCE() (when it fires, it fires a _lot).
This, with a reworded changelog?
/* * UMIP emulation relies on intercepting writes to CR4.UMIP, i.e. this * and other code needs to be updated if UMIP can be guest owned. */ BUILD_BUG_ON(KVM_POSSIBLE_CR4_GUEST_BITS & X86_CR4_UMIP); WARN_ON_ONCE(!kvm_read_cr4_bits(vcpu, X86_CR4_UMIP)); return kvm_emulate_instruction(vcpu, 0);
A message that is not in the archive (sent off-list or not yet received): CA+wubQBsiaH_==UJ-JUi7hwS8W1i5MLZ-dPuw2smVH8Z0sqXsw@mail.gmail.com
Sean Christopherson
maintainerRe: [PATCH 3/3] KVM: VMX: Use the canonical interface to read CR4.UMIP bit
Quoted 7 lines · Robert Hoo
Quoted 1 line · Sean Christopherson <[email protected]> 于2023年3月11日周六 00:27写道:
Practically speaking, UMIP _can't_ be guest owned without breaking UMIP
emulation.
Agree.
Can we expect that in the future (not near future) that emulating UMIP isn't needed at all? I mean, when all x86 CPUs on the market have UMIP in native.
Heh, define "all x86s on the market". UMIP didn't come along until Broadwell, and we still get bug reports for Core2. It's going to be a long, long time before KVM can require UMIP support on Intel CPUs.
Re: [PATCH 3/3] KVM: VMX: Use the canonical interface to read CR4.UMIP bit
In reply to Sean ChristophersonQuoted 19 lines · Sean Christopherson <[email protected]> 于2023年3月11日周六 00:27写道:
Quoted 2 lines · Robert Hoo
Use
kvm_read_cr4_bits()rather than directly readvcpu->arch.cr4, now that we have reg cache layer and defined this wrapper.
kvm_read_cr4_bits()predates this code by ~7 years.Quoted 3 lines
Although, effectively for CR4.UMIP, it's the same, at present, as it's not guest owned, in case of future changes, here better to use the canonical interface.
Practically speaking, UMIP _can't_ be guest owned without breaking UMIP emulation. I do like not open coding
vcpu->arch.cr4, but I don't particuarly like the changelog.This would also be a good time to opportunistically convert the
WARN_ON()to aWARN_ON_ONCE()(when it fires, it fires a_lot).This, with a reworded changelog?
Quoted 8 lines
/*
* UMIP emulation relies on intercepting writes to CR4.UMIP, i.e. this
* and other code needs to be updated if UMIP can be guest owned.
*/
BUILD_BUG_ON(KVM_POSSIBLE_CR4_GUEST_BITS & X86_CR4_UMIP);
WARN_ON_ONCE(!kvm_read_cr4_bits(vcpu, X86_CR4_UMIP));
return kvm_emulate_instruction(vcpu, 0);Are you going to have this along with your "[PATCH] KVM: VMX: Treat UMIP as emulated if and only if the host doesn't have UMIP"?
Sean Christopherson
maintainerRe: [PATCH 3/3] KVM: VMX: Use the canonical interface to read CR4.UMIP bit
Quoted 26 lines · Robert Hoo
Quoted 25 lines · Sean Christopherson <[email protected]> 于2023年3月11日周六 00:27写道:
Quoted 2 lines · Robert Hoo
Use
kvm_read_cr4_bits()rather than directly readvcpu->arch.cr4, now that we have reg cache layer and defined this wrapper.
kvm_read_cr4_bits()predates this code by ~7 years.Quoted 3 lines
Although, effectively for CR4.UMIP, it's the same, at present, as it's not guest owned, in case of future changes, here better to use the canonical interface.
Practically speaking, UMIP _can't_ be guest owned without breaking UMIP emulation. I do like not open coding
vcpu->arch.cr4, but I don't particuarly like the changelog.This would also be a good time to opportunistically convert the
WARN_ON()to aWARN_ON_ONCE()(when it fires, it fires a_lot).This, with a reworded changelog?
/* * UMIP emulation relies on intercepting writes to CR4.UMIP, i.e. this * and other code needs to be updated if UMIP can be guest owned. */ BUILD_BUG_ON(KVM_POSSIBLE_CR4_GUEST_BITS & X86_CR4_UMIP);
Quoted 5 lines
Quoted 2 lines
WARN_ON_ONCE(!kvm_read_cr4_bits(vcpu, X86_CR4_UMIP)); return kvm_emulate_instruction(vcpu, 0);Are you going to have this along with your "[PATCH] KVM: VMX: Treat UMIP as emulated if and only if the host doesn't have UMIP"?
Sure, I'll add a patch for that.