Currently to support Intel->AMD migration, if CPU vendor is GenuineIntel,
we emulate the full 64 value for MSR_IA32_SYSENTER_{EIP|ESP}
msrs, and we also emulate the sysenter/sysexit instruction in long mode.
(Emulator does still refuse to emulate sysenter in 64 bit mode, on the
ground that the code for that wasn't tested and likely has no users)
However when virtual vmload/vmsave is enabled, the vmload instruction will
update these 32 bit msrs without triggering their msr intercept,
which will lead to having stale values in kvm's shadow copy of these msrs,
which relies on the intercept to be up to date.
Fix/optimize this by doing the following:
1. Enable the MSR intercepts for SYSENTER MSRs iff vendor=GenuineIntel
(This is both a tiny optimization and also ensures that in case
the guest cpu vendor is AMD, the msrs will be 32 bit wide as
AMD defined).
2. Store only high 32 bit part of these msrs on interception and combine
it with hardware msr value on intercepted read/writes
iff vendor=GenuineIntel.
3. Disable vmload/vmsave virtualization if vendor=GenuineIntel.
(It is somewhat insane to set vendor=GenuineIntel and still enable
SVM for the guest but well whatever).
Then zero the high 32 bit parts when kvm intercepts and emulates vmload.
Thanks a lot to Paulo Bonzini for helping me with fixing this in the most
correct way.
This patch fixes nested migration of 32 bit nested guests, that was
broken because incorrect cached values of SYSENTER msrs were stored in
the migration stream if L1 changed these msrs with
vmload prior to L2 entry.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210401111928.996871-3-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Set the PAE roots used as decrypted to play nice with SME when KVM is
using shadow paging. Explicitly skip setting the C-bit when loading
CR3 for PAE shadow paging, even though it's completely ignored by the
CPU. The extra documentation is nice to have.
Note, there are several subtleties at play with NPT. In addition to
legacy shadow paging, the PAE roots are used for SVM's NPT when either
KVM is 32-bit (uses PAE paging) or KVM is 64-bit and shadowing 32-bit
NPT. However, 32-bit Linux, and thus KVM, doesn't support SME. And
64-bit KVM can happily set the C-bit in CR3. This also means that
keeping __sme_set(root) for 32-bit KVM when NPT is enabled is
conceptually wrong, but functionally ok since SME is 64-bit only.
Leave it as is to avoid unnecessary pollution.
Fixes: d0ec49d4de ("kvm/x86/svm: Support Secure Memory Encryption within KVM")
Cc: stable@vger.kernel.org
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210309224207.1218275-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Retrieve the active PCID only when writing a guest CR3 value, i.e. don't
get the PCID when using EPT or NPT. The PCID is especially problematic
for EPT as the bits have different meaning, and so the PCID and must be
manually stripped, which is annoying and unnecessary. And on VMX,
getting the active PCID also involves reading the guest's CR3 and
CR4.PCIDE, i.e. may add pointless VMREADs.
Opportunistically rename the pgd/pgd_level params to root_hpa and
root_level to better reflect their new roles. Keep the function names,
as "load the guest PGD" is still accurate/correct.
Last, and probably least, pass root_hpa as a hpa_t/u64 instead of an
unsigned long. The EPTP holds a 64-bit value, even in 32-bit mode, so
in theory EPT could support HIGHMEM for 32-bit KVM. Never mind that
doing so would require changing the MMU page allocators and reworking
the MMU to use kmap().
Signed-off-by: Sean Christopherson <sean.j.christopherson@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305183123.3978098-2-seanjc@google.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Stop tagging MMIO SPTEs with specific available bits and instead detect
MMIO SPTEs by checking for their unique SPTE value. The value is
guaranteed to be unique on shadow paging and NPT as setting reserved
physical address bits on any other type of SPTE would consistute a KVM
bug. Ditto for EPT, as creating a WX non-MMIO would also be a bug.
Note, this approach is also future-compatibile with TDX, which will need
to reflect MMIO EPT violations as #VEs into the guest. To create an EPT
violation instead of a misconfig, TDX EPTs will need to have RWX=0, But,
MMIO SPTEs will also be the only case where KVM clears SUPPRESS_VE, so
MMIO SPTEs will still be guaranteed to have a unique value within a given
MMU context.
The main motivation is to make it easier to reason about which types of
SPTEs use which available bits. As a happy side effect, this frees up
two more bits for storing the MMIO generation.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210225204749.1512652-11-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Use the vmcb12 control clean field to determine which vmcb12.save
registers were marked dirty in order to minimize register copies
when switching from L1 to L2. Those vmcb12 registers marked as dirty need
to be copied to L0's vmcb02 as they will be used to update the vmcb
state cache for the L2 VMRUN. In the case where we have a different
vmcb12 from the last L2 VMRUN all vmcb12.save registers must be
copied over to L2.save.
Tested:
kvm-unit-tests
kvm selftests
Fedora L1 L2
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20210301200844.2000-1-cavery@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Newer AMD processors have a feature to virtualize the use of the
SPEC_CTRL MSR. Presence of this feature is indicated via CPUID
function 0x8000000A_EDX[20]: GuestSpecCtrl. Hypervisors are not
required to enable this feature since it is automatically enabled on
processors that support it.
A hypervisor may wish to impose speculation controls on guest
execution or a guest may want to impose its own speculation controls.
Therefore, the processor implements both host and guest
versions of SPEC_CTRL.
When in host mode, the host SPEC_CTRL value is in effect and writes
update only the host version of SPEC_CTRL. On a VMRUN, the processor
loads the guest version of SPEC_CTRL from the VMCB. When the guest
writes SPEC_CTRL, only the guest version is updated. On a VMEXIT,
the guest version is saved into the VMCB and the processor returns
to only using the host SPEC_CTRL for speculation control. The guest
SPEC_CTRL is located at offset 0x2E0 in the VMCB.
The effective SPEC_CTRL setting is the guest SPEC_CTRL setting or'ed
with the hypervisor SPEC_CTRL setting. This allows the hypervisor to
ensure a minimum SPEC_CTRL if desired.
This support also fixes an issue where a guest may sometimes see an
inconsistent value for the SPEC_CTRL MSR on processors that support
this feature. With the current SPEC_CTRL support, the first write to
SPEC_CTRL is intercepted and the virtualized version of the SPEC_CTRL
MSR is not updated. When the guest reads back the SPEC_CTRL MSR, it
will be 0x0, instead of the actual expected value. There isn’t a
security concern here, because the host SPEC_CTRL value is or’ed with
the Guest SPEC_CTRL value to generate the effective SPEC_CTRL value.
KVM writes with the guest's virtualized SPEC_CTRL value to SPEC_CTRL
MSR just before the VMRUN, so it will always have the actual value
even though it doesn’t appear that way in the guest. The guest will
only see the proper value for the SPEC_CTRL register if the guest was
to write to the SPEC_CTRL register again. With Virtual SPEC_CTRL
support, the save area spec_ctrl is properly saved and restored.
So, the guest will always see the proper value when it is read back.
Signed-off-by: Babu Moger <babu.moger@amd.com>
Message-Id: <161188100955.28787.11816849358413330720.stgit@bmoger-ubuntu>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This allows to avoid copying of these fields between vmcb01
and vmcb02 on nested guest entry/exit.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Thanks to the new macros that handle exception handling for SVM
instructions, it is easier to just do the VMLOAD/VMSAVE in C.
This is safe, as shown by the fact that the host reload is
already done outside the assembly source.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Skip PAUSE after interception to avoid unnecessarily re-executing the
instruction in the guest, e.g. after regaining control post-yield.
This is a benign bug as KVM disables PAUSE interception if filtering is
off, including the case where pause_filter_count is set to zero.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-10-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Remove bizarre code that causes KVM to run RDPMC through the emulator
when nrips is disabled. Accelerated emulation of RDPMC doesn't rely on
any additional data from the VMCB, and SVM has generic handling for
updating RIP to skip instructions when nrips is disabled.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-9-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the entirety of the accelerated RDPMC emulation to x86.c, and assign
the common handler directly to the exit handler array for VMX. SVM has
bizarre nrips behavior that prevents it from directly invoking the common
handler. The nrips goofiness will be addressed in a future patch.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the trivial exit handlers, e.g. for instructions that KVM
"emulates" as nops, to common x86 code. Assign the common handlers
directly to the exit handler arrays and drop the vendor trampolines.
Opportunistically use pr_warn_once() where appropriate.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-7-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Move the entirety of XSETBV emulation to x86.c, and assign the
function directly to both VMX's and SVM's exit handlers, i.e. drop the
unnecessary trampolines.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-6-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add another helper layer for VMLOAD+VMSAVE, the code is identical except
for the one line that determines which VMCB is the source and which is
the destination.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-5-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a helper to consolidate boilerplate for nested VM-Exits that don't
provide any data in exit_info_*.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210302174515.2812275-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Refactor the svm_exit_handlers API to pass @vcpu instead of @svm to
allow directly invoking common x86 exit handlers (in a future patch).
Opportunistically convert an absurd number of instances of 'svm->vcpu'
to direct uses of 'vcpu' to avoid pointless casting.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The logic of update_cr0_intercept is pointlessly complicated.
All svm_set_cr0 is compute the effective cr0 and compare it with
the guest value.
Inlining the function and simplifying the condition
clarifies what it is doing.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Now that SVM is using a separate vmcb01 and vmcb02 (and also uses the vmcb12
naming) we can give clearer names to functions that write to and read
from those VMCBs. Likewise, variables and parameters can be renamed
from nested_vmcb to vmcb12.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch moves the asid_generation from the vcpu to the vmcb
in order to track the ASID generation that was active the last
time the vmcb was run. If sd->asid_generation changes between
two runs, the old ASID is invalid and must be changed.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20210112164313.4204-3-cavery@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This patch moves the physical cpu tracking from the vcpu
to the vmcb in svm_switch_vmcb. If either vmcb01 or vmcb02
change physical cpus from one vmrun to the next the vmcb's
previous cpu is preserved for comparison with the current
cpu and the vmcb is marked dirty if different. This prevents
the processor from using old cached data for a vmcb that may
have been updated on a prior run on a different processor.
It also moves the physical cpu check from svm_vcpu_load
to pre_svm_run as the check only needs to be done at run.
Suggested-by: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20210112164313.4204-2-cavery@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
svm->vmcb will now point to a separate vmcb for L1 (not nested) or L2
(nested).
The main advantages are removing get_host_vmcb and hsave, in favor of
concepts that are shared with VMX.
We don't need anymore to stash the L1 registers in hsave while L2
runs, but we need to copy the VMLOAD/VMSAVE registers from VMCB01 to
VMCB02 and back. This more or less has the same cost, but code-wise
nested_svm_vmloadsave can be reused.
This patch omits several optimizations that are possible:
- for simplicity there is some wholesale copying of vmcb.control areas
which can go away.
- we should be able to better use the VMCB01 and VMCB02 clean bits.
- another possibility is to always use VMCB01 for VMLOAD and VMSAVE,
thus avoiding the copy of VMLOAD/VMSAVE registers from VMCB01 to
VMCB02 and back.
Tested:
kvm-unit-tests
kvm self tests
Loaded fedora nested guest on fedora
Signed-off-by: Cathy Avery <cavery@redhat.com>
Message-Id: <20201011184818.3609-3-cavery@redhat.com>
[Fix conflicts; keep VMCB02 G_PAT up to date whenever guest writes the
PAT MSR; do not copy CR4 over from VMCB01 as it is not needed anymore; add
a few more comments. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Don't strip the C-bit from the faulting address on an intercepted #PF,
the address is a virtual address, not a physical address.
Fixes: 0ede79e132 ("KVM: SVM: Clear C-bit from the page fault address")
Cc: stable@vger.kernel.org
Cc: Brijesh Singh <brijesh.singh@amd.com>
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305011101.3597423-13-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Directly connect the 'npt' param to the 'npt_enabled' variable so that
runtime adjustments to npt_enabled are reflected in sysfs. Move the
!PAE restriction to a runtime check to ensure NPT is forced off if the
host is using 2-level paging, and add a comment explicitly stating why
NPT requires a 64-bit kernel or a kernel with PAE enabled.
Opportunistically switch the param to octal permissions.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210305021637.3768573-1-seanjc@google.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This problem was reported on a SVM guest while executing kexec.
Kexec fails to load the new kernel when the PCID feature is enabled.
When kexec starts loading the new kernel, it starts the process by
resetting the vCPU's and then bringing each vCPU online one by one.
The vCPU reset is supposed to reset all the register states before the
vCPUs are brought online. However, the CR4 register is not reset during
this process. If this register is already setup during the last boot,
all the flags can remain intact. The X86_CR4_PCIDE bit can only be
enabled in long mode. So, it must be enabled much later in SMP
initialization. Having the X86_CR4_PCIDE bit set during SMP boot can
cause a boot failures.
Fix the issue by resetting the CR4 register in init_vmcb().
Signed-off-by: Babu Moger <babu.moger@amd.com>
Message-Id: <161471109108.30811.6392805173629704166.stgit@bmoger-ubuntu>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Fix the interpreation of nested_svm_vmexit()'s return value when
synthesizing a nested VM-Exit after intercepting an SVM instruction while
L2 was running. The helper returns '0' on success, whereas a return
value of '0' in the exit handler path means "exit to userspace". The
incorrect return value causes KVM to exit to userspace without filling
the run state, e.g. QEMU logs "KVM: unknown exit, hardware reason 0".
Fixes: 14c2bf81fc ("KVM: SVM: Fix #GP handling for doubly-nested virtualization")
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210224005627.657028-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Advertise INVPCID by default (if supported by the host kernel) instead
of having both SVM and VMX opt in. INVPCID was opt in when it was a
VMX only feature so that KVM wouldn't prematurely advertise support
if/when it showed up in the kernel on AMD hardware.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210212003411.1102677-3-seanjc@google.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Intercept INVPCID if it's disabled in the guest, even when using NPT,
as KVM needs to inject #UD in this case.
Fixes: 4407a797e9 ("KVM: SVM: Enable INVPCID feature on AMD")
Cc: Babu Moger <babu.moger@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210212003411.1102677-2-seanjc@google.com>
Reviewed-by: Jim Mattson <jmattson@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The sparse tool complains as follows:
arch/x86/kvm/svm/svm.c:204:6: warning:
symbol 'svm_gp_erratum_intercept' was not declared. Should it be static?
This symbol is not used outside of svm.c, so this
commit marks it static.
Fixes: 82a11e9c6f ("KVM: SVM: Add emulation support for #GP triggered by SVM instructions")
Reported-by: Hulk Robot <hulkci@huawei.com>
Signed-off-by: Wei Yongjun <weiyongjun1@huawei.com>
Message-Id: <20210210075958.1096317-1-weiyongjun1@huawei.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Push the injection of #GP up to the callers, so that they can just use
kvm_complete_insn_gp. __kvm_set_dr is pretty much what the callers can use
together with kvm_complete_insn_gp, so rename it to kvm_set_dr and drop
the old kvm_set_dr wrapper.
This also allows nested VMX code, which really wanted to use __kvm_set_dr,
to use the right function.
While at it, remove the kvm_require_dr() check from the SVM interception.
The APM states:
All normal exception checks take precedence over the SVM intercepts.
which includes the CR4.DE=1 #UD.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Drop a defunct forward declaration of svm_complete_interrupts().
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210205005750.3841462-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Rename cr3_lm_rsvd_bits to reserved_gpa_bits, and use it for all GPA
legality checks. AMD's APM states:
If the C-bit is an address bit, this bit is masked from the guest
physical address when it is translated through the nested page tables.
Thus, any access that can conceivably be run through NPT should ignore
the C-bit when checking for validity.
For features that KVM emulates in software, e.g. MTRRs, there is no
clear direction in the APM for how the C-bit should be handled. For
such cases, follow the SME behavior inasmuch as possible, since SEV is
is essentially a VM-specific variant of SME. For SME, the APM states:
In this case the upper physical address bits are treated as reserved
when the feature is enabled except where otherwise indicated.
Collecting the various relavant SME snippets in the APM and cross-
referencing the omissions with Linux kernel code, this leaves MTTRs and
APIC_BASE as the only flows that KVM emulates that should _not_ ignore
the C-bit.
Note, this means the reserved bit checks in the page tables are
technically broken. This will be remedied in a future patch.
Although the page table checks are technically broken, in practice, it's
all but guaranteed to be irrelevant. NPT is required for SEV, i.e.
shadowing page tables isn't needed in the common case. Theoretically,
the checks could be in play for nested NPT, but it's extremely unlikely
that anyone is running nested VMs on SEV, as doing so would require L1
to expose sensitive data to L0, e.g. the entire VMCB. And if anyone is
running nested VMs, L0 can't read the guest's encrypted memory, i.e. L1
would need to put its NPT in shared memory, in which case the C-bit will
never be set. Or, L1 could use shadow paging, but again, if L0 needs to
read page tables, e.g. to load PDPTRs, the memory can't be encrypted if
L1 has any expectation of L0 doing the right thing.
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Brijesh Singh <brijesh.singh@amd.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210204000117.3303214-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Replace the hard-coded value for bit# 1 in EFLAGS, with the available
#define.
Signed-off-by: Krish Sadhukhan <krish.sadhukhan@oracle.com>
Message-Id: <20210203012842.101447-2-krish.sadhukhan@oracle.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently we save host state like user-visible host MSRs, and do some
initial guest register setup for MSR_TSC_AUX and MSR_AMD64_TSC_RATIO
in svm_vcpu_load(). Defer this until just before we enter the guest by
moving the handling to kvm_x86_ops.prepare_guest_switch() similarly to
how it is done for the VMX implementation.
Additionally, since handling of saving/restoring host user MSRs is the
same both with/without SEV-ES enabled, move that handling to common
code.
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Michael Roth <michael.roth@amd.com>
Message-Id: <20210202190126.2185715-4-michael.roth@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Now that the set of host user MSRs that need to be individually
saved/restored are the same with/without SEV-ES, we can drop the
.sev_es_restored flag and just iterate through the list unconditionally
for both cases. A subsequent patch can then move these loops to a
common path.
Signed-off-by: Michael Roth <michael.roth@amd.com>
Message-Id: <20210202190126.2185715-3-michael.roth@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Using a guest workload which simply issues 'hlt' in a tight loop to
generate VMEXITs, it was observed (on a recent EPYC processor) that a
significant amount of the VMEXIT overhead measured on the host was the
result of MSR reads/writes in svm_vcpu_load/svm_vcpu_put according to
perf:
67.49%--kvm_arch_vcpu_ioctl_run
|
|--23.13%--vcpu_put
| kvm_arch_vcpu_put
| |
| |--21.31%--native_write_msr
| |
| --1.27%--svm_set_cr4
|
|--16.11%--vcpu_load
| |
| --15.58%--kvm_arch_vcpu_load
| |
| |--13.97%--svm_set_cr4
| | |
| | |--12.64%--native_read_msr
Most of these MSRs relate to 'syscall'/'sysenter' and segment bases, and
can be saved/restored using 'vmsave'/'vmload' instructions rather than
explicit MSR reads/writes. In doing so there is a significant reduction
in the svm_vcpu_load/svm_vcpu_put overhead measured for the above
workload:
50.92%--kvm_arch_vcpu_ioctl_run
|
|--19.28%--disable_nmi_singlestep
|
|--13.68%--vcpu_load
| kvm_arch_vcpu_load
| |
| |--9.19%--svm_set_cr4
| | |
| | --6.44%--native_read_msr
| |
| --3.55%--native_write_msr
|
|--6.05%--kvm_inject_nmi
|--2.80%--kvm_sev_es_mmio_read
|--2.19%--vcpu_put
| |
| --1.25%--kvm_arch_vcpu_put
| native_write_msr
Quantifying this further, if we look at the raw cycle counts for a
normal iteration of the above workload (according to 'rdtscp'),
kvm_arch_vcpu_ioctl_run() takes ~4600 cycles from start to finish with
the current behavior. Using 'vmsave'/'vmload', this is reduced to
~2800 cycles, a savings of 39%.
While this approach doesn't seem to manifest in any noticeable
improvement for more realistic workloads like UnixBench, netperf, and
kernel builds, likely due to their exit paths generally involving IO
with comparatively high latencies, it does improve overall overhead
of KVM_RUN significantly, which may still be noticeable for certain
situations. It also simplifies some aspects of the code.
With this change, explicit save/restore is no longer needed for the
following host MSRs, since they are documented[1] as being part of the
VMCB State Save Area:
MSR_STAR, MSR_LSTAR, MSR_CSTAR,
MSR_SYSCALL_MASK, MSR_KERNEL_GS_BASE,
MSR_IA32_SYSENTER_CS,
MSR_IA32_SYSENTER_ESP,
MSR_IA32_SYSENTER_EIP,
MSR_FS_BASE, MSR_GS_BASE
and only the following MSR needs individual handling in
svm_vcpu_put/svm_vcpu_load:
MSR_TSC_AUX
We could drop the host_save_user_msrs array/loop and instead handle
MSR read/write of MSR_TSC_AUX directly, but we leave that for now as
a potential follow-up.
Since 'vmsave'/'vmload' also handles the LDTR and FS/GS segment
registers (and associated hidden state)[2], some of the code
previously used to handle this is no longer needed, so we drop it
as well.
The first public release of the SVM spec[3] also documents the same
handling for the host state in question, so we make these changes
unconditionally.
Also worth noting is that we 'vmsave' to the same page that is
subsequently used by 'vmrun' to record some host additional state. This
is okay, since, in accordance with the spec[2], the additional state
written to the page by 'vmrun' does not overwrite any fields written by
'vmsave'. This has also been confirmed through testing (for the above
CPU, at least).
[1] AMD64 Architecture Programmer's Manual, Rev 3.33, Volume 2, Appendix B, Table B-2
[2] AMD64 Architecture Programmer's Manual, Rev 3.31, Volume 3, Chapter 4, VMSAVE/VMLOAD
[3] Secure Virtual Machine Architecture Reference Manual, Rev 3.01
Suggested-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Michael Roth <michael.roth@amd.com>
Message-Id: <20210202190126.2185715-2-michael.roth@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add svm_asm*() macros, a la the existing vmx_asm*() macros, to handle
faults on SVM instructions instead of using the generic __ex(), a.k.a.
__kvm_handle_fault_on_reboot(). Using asm goto generates slightly
better code as it eliminates the in-line JMP+CALL sequences that are
needed by __kvm_handle_fault_on_reboot() to avoid triggering BUG()
from fixup (which generates bad stack traces).
Using SVM specific macros also drops the last user of __ex() and the
the last asm linkage to kvm_spurious_fault(), and adds a helper for
VMSAVE, which may gain an addition call site in the future (as part
of optimizing the SVM context switching).
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20201231002702.2223707-8-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
A subsequent patch introduces macros in preparation for simplifying the
definition for vmx_x86_ops and svm_x86_ops. Making the naming more uniform
expands the coverage of the macros. Add vmx/svm prefix to the following
functions: update_exception_bitmap(), enable_nmi_window(),
enable_irq_window(), update_cr8_intercept and enable_smi_window().
Cc: Paolo Bonzini <pbonzini@redhat.com>
Cc: Sean Christopherson <seanjc@google.com>
Cc: Thomas Gleixner <tglx@linutronix.de>
Cc: Ingo Molnar <mingo@redhat.com>
Cc: Borislav Petkov <bp@alien8.de>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Andrea Arcangeli <aarcange@redhat.com>
Signed-off-by: Jason Baron <jbaron@akamai.com>
Message-Id: <ed594696f8e2c2b2bfc747504cee9bbb2a269300.1610680941.git.jbaron@akamai.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Under the case of nested on nested (L0, L1, L2 are all hypervisors),
we do not support emulation of the vVMLOAD/VMSAVE feature, the
L0 hypervisor can inject the proper #VMEXIT to inform L1 of what is
happening and L1 can avoid invoking the #GP workaround. For this
reason we turns on guest VM's X86_FEATURE_SVME_ADDR_CHK bit for KVM
running inside VM to receive the notification and change behavior.
Similarly we check if vcpu is under guest mode before emulating the
vmware-backdoor instructions. For the case of nested on nested, we
let the guest handle it.
Co-developed-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Wei Huang <wei.huang2@amd.com>
Tested-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210126081831.570253-5-wei.huang2@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
New AMD CPUs have a change that checks #VMEXIT intercept on special SVM
instructions before checking their EAX against reserved memory region.
This change is indicated by CPUID_0x8000000A_EDX[28]. If it is 1, #VMEXIT
is triggered before #GP. KVM doesn't need to intercept and emulate #GP
faults as #GP is supposed to be triggered.
Co-developed-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Bandan Das <bsd@redhat.com>
Signed-off-by: Wei Huang <wei.huang2@amd.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20210126081831.570253-4-wei.huang2@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
While running SVM related instructions (VMRUN/VMSAVE/VMLOAD), some AMD
CPUs check EAX against reserved memory regions (e.g. SMM memory on host)
before checking VMCB's instruction intercept. If EAX falls into such
memory areas, #GP is triggered before VMEXIT. This causes problem under
nested virtualization. To solve this problem, KVM needs to trap #GP and
check the instructions triggering #GP. For VM execution instructions,
KVM emulates these instructions.
Co-developed-by: Wei Huang <wei.huang2@amd.com>
Signed-off-by: Wei Huang <wei.huang2@amd.com>
Signed-off-by: Bandan Das <bsd@redhat.com>
Message-Id: <20210126081831.570253-3-wei.huang2@amd.com>
[Conditionally enable #GP intercept. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
DR6_INIT contains the 1-reserved bits as well as the bit that is cleared
to 0 when the condition (e.g. RTM) happens. The value can be used to
initialize dr6 and also be the XOR mask between the #DB exit
qualification (or payload) and DR6.
Concerning that DR6_INIT is used as initial value only once, rename it
to DR6_ACTIVE_LOW and apply it in other places, which would make the
incoming changes for bus lock debug exception more simple.
Signed-off-by: Chenyi Qiang <chenyi.qiang@intel.com>
Message-Id: <20210202090433.13441-2-chenyi.qiang@intel.com>
[Define DR6_FIXED_1 from DR6_ACTIVE_LOW and DR6_VOLATILE. - Paolo]
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Don't let KVM load when running as an SEV guest, regardless of what
CPUID says. Memory is encrypted with a key that is not accessible to
the host (L0), thus it's impossible for L0 to emulate SVM, e.g. it'll
see garbage when reading the VMCB.
Technically, KVM could decrypt all memory that needs to be accessible to
the L0 and use shadow paging so that L0 does not need to shadow NPT, but
exposing such information to L0 largely defeats the purpose of running as
an SEV guest. This can always be revisited if someone comes up with a
use case for running VMs inside SEV guests.
Note, VMLOAD, VMRUN, etc... will also #GP on GPAs with C-bit set, i.e. KVM
is doomed even if the SEV guest is debuggable and the hypervisor is willing
to decrypt the VMCB. This may or may not be fixed on CPUs that have the
SVME_ADDR_CHK fix.
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20210202212017.2486595-1-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On VMX, if we exit and then re-enter immediately without leaving
the vmx_vcpu_run() function, the kvm_entry event is not logged.
That means we will see one (or more) kvm_exit, without its (their)
corresponding kvm_entry, as shown here:
CPU-1979 [002] 89.871187: kvm_entry: vcpu 1
CPU-1979 [002] 89.871218: kvm_exit: reason MSR_WRITE
CPU-1979 [002] 89.871259: kvm_exit: reason MSR_WRITE
It also seems possible for a kvm_entry event to be logged, but then
we leave vmx_vcpu_run() right away (if vmx->emulation_required is
true). In this case, we will have a spurious kvm_entry event in the
trace.
Fix these situations by moving trace_kvm_entry() inside vmx_vcpu_run()
(where trace_kvm_exit() already is).
A trace obtained with this patch applied looks like this:
CPU-14295 [000] 8388.395387: kvm_entry: vcpu 0
CPU-14295 [000] 8388.395392: kvm_exit: reason MSR_WRITE
CPU-14295 [000] 8388.395393: kvm_entry: vcpu 0
CPU-14295 [000] 8388.395503: kvm_exit: reason EXTERNAL_INTERRUPT
Of course, not calling trace_kvm_entry() in common x86 code any
longer means that we need to adjust the SVM side of things too.
Signed-off-by: Lorenzo Brescia <lorenzo.brescia@edu.unito.it>
Signed-off-by: Dario Faggioli <dfaggioli@suse.com>
Message-Id: <160873470698.11652.13483635328769030605.stgit@Wayrath>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Typically under KVM, an AP is booted using the INIT-SIPI-SIPI sequence,
where the guest vCPU register state is updated and then the vCPU is VMRUN
to begin execution of the AP. For an SEV-ES guest, this won't work because
the guest register state is encrypted.
Following the GHCB specification, the hypervisor must not alter the guest
register state, so KVM must track an AP/vCPU boot. Should the guest want
to park the AP, it must use the AP Reset Hold exit event in place of, for
example, a HLT loop.
First AP boot (first INIT-SIPI-SIPI sequence):
Execute the AP (vCPU) as it was initialized and measured by the SEV-ES
support. It is up to the guest to transfer control of the AP to the
proper location.
Subsequent AP boot:
KVM will expect to receive an AP Reset Hold exit event indicating that
the vCPU is being parked and will require an INIT-SIPI-SIPI sequence to
awaken it. When the AP Reset Hold exit event is received, KVM will place
the vCPU into a simulated HLT mode. Upon receiving the INIT-SIPI-SIPI
sequence, KVM will make the vCPU runnable. It is again up to the guest
to then transfer control of the AP to the proper location.
To differentiate between an actual HLT and an AP Reset Hold, a new MP
state is introduced, KVM_MP_STATE_AP_RESET_HOLD, which the vCPU is
placed in upon receiving the AP Reset Hold exit event. Additionally, to
communicate the AP Reset Hold exit event up to userspace (if needed), a
new exit reason is introduced, KVM_EXIT_AP_RESET_HOLD.
A new x86 ops function is introduced, vcpu_deliver_sipi_vector, in order
to accomplish AP booting. For VMX, vcpu_deliver_sipi_vector is set to the
original SIPI delivery function, kvm_vcpu_deliver_sipi_vector(). SVM adds
a new function that, for non SEV-ES guests, invokes the original SIPI
delivery function, kvm_vcpu_deliver_sipi_vector(), but for SEV-ES guests,
implements the logic above.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Message-Id: <e8fbebe8eb161ceaabdad7c01a5859a78b424d5e.1609791600.git.thomas.lendacky@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Commit 16809ecdc1 moved __svm_vcpu_run the prototype to svm.h,
but forgot to remove the original from svm.c.
Fixes: 16809ecdc1 ("KVM: SVM: Provide an updated VMRUN invocation for SEV-ES guests")
Cc: Tom Lendacky <thomas.lendacky@amd.com>
Cc: Paolo Bonzini <pbonzini@redhat.com>
Signed-off-by: Uros Bizjak <ubizjak@gmail.com>
Message-Id: <20201220200339.65115-1-ubizjak@gmail.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The run sequence is different for an SEV-ES guest compared to a legacy or
even an SEV guest. The guest vCPU register state of an SEV-ES guest will
be restored on VMRUN and saved on VMEXIT. There is no need to restore the
guest registers directly and through VMLOAD before VMRUN and no need to
save the guest registers directly and through VMSAVE on VMEXIT.
Update the svm_vcpu_run() function to skip register state saving and
restoring and provide an alternative function for running an SEV-ES guest
in vmenter.S
Additionally, certain host state is restored across an SEV-ES VMRUN. As
a result certain register states are not required to be restored upon
VMEXIT (e.g. FS, GS, etc.), so only do that if the guest is not an SEV-ES
guest.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Message-Id: <fb1c66d32f2194e171b95fc1a8affd6d326e10c1.1607620209.git.thomas.lendacky@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
An SEV-ES vCPU requires additional VMCB vCPU load/put requirements. SEV-ES
hardware will restore certain registers on VMEXIT, but not save them on
VMRUN (see Table B-3 and Table B-4 of the AMD64 APM Volume 2), so make the
following changes:
General vCPU load changes:
- During vCPU loading, perform a VMSAVE to the per-CPU SVM save area and
save the current values of XCR0, XSS and PKRU to the per-CPU SVM save
area as these registers will be restored on VMEXIT.
General vCPU put changes:
- Do not attempt to restore registers that SEV-ES hardware has already
restored on VMEXIT.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Message-Id: <019390e9cb5e93cd73014fa5a040c17d42588733.1607620209.git.thomas.lendacky@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
An SEV-ES vCPU requires additional VMCB initialization requirements for
vCPU creation and vCPU load/put requirements. This includes:
General VMCB initialization changes:
- Set a VMCB control bit to enable SEV-ES support on the vCPU.
- Set the VMCB encrypted VM save area address.
- CRx registers are part of the encrypted register state and cannot be
updated. Remove the CRx register read and write intercepts and replace
them with CRx register write traps to track the CRx register values.
- Certain MSR values are part of the encrypted register state and cannot
be updated. Remove certain MSR intercepts (EFER, CR_PAT, etc.).
- Remove the #GP intercept (no support for "enable_vmware_backdoor").
- Remove the XSETBV intercept since the hypervisor cannot modify XCR0.
General vCPU creation changes:
- Set the initial GHCB gpa value as per the GHCB specification.
Signed-off-by: Tom Lendacky <thomas.lendacky@amd.com>
Message-Id: <3a8aef366416eddd5556dfa3fdc212aafa1ad0a2.1607620209.git.thomas.lendacky@amd.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
