SYNTHESIZED_F() generally is used together with setup_force_cpu_cap(),
i.e. when it makes sense to present the feature even if cpuid does not
have it *and* the VM is not able to see the difference. For example,
it can be used when mitigations on the host automatically protect
the guest as well.
The "SYNTHESIZED_F(SRSO_USER_KERNEL_NO)" line came in as a conflict
resolution between the CPUID overhaul from the KVM tree and support
for the feature in the x86 tree. Using it right now does not hurt,
or make a difference for that matter, because there is no
setup_force_cpu_cap(X86_FEATURE_SRSO_USER_KERNEL_NO). However, it
is a little less future proof in case such a setup_force_cpu_cap()
appears later, for a case where the kernel somehow is not vulnerable
but the guest would have to apply the mitigation.
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
* Clear LLBCTL if secondary mmu mapping changes.
* Add hypercall service support for usermode VMM.
x86:
* Add a comment to kvm_mmu_do_page_fault() to explain why KVM performs a
direct call to kvm_tdp_page_fault() when RETPOLINE is enabled.
* Ensure that all SEV code is compiled out when disabled in Kconfig, even
if building with less brilliant compilers.
* Remove a redundant TLB flush on AMD processors when guest CR4.PGE changes.
* Use str_enabled_disabled() to replace open coded strings.
* Drop kvm_x86_ops.hwapic_irr_update() as KVM updates hardware's APICv cache
prior to every VM-Enter.
* Overhaul KVM's CPUID feature infrastructure to track all vCPU capabilities
instead of just those where KVM needs to manage state and/or explicitly
enable the feature in hardware. Along the way, refactor the code to make
it easier to add features, and to make it more self-documenting how KVM
is handling each feature.
* Rework KVM's handling of VM-Exits during event vectoring; this plugs holes
where KVM unintentionally puts the vCPU into infinite loops in some scenarios
(e.g. if emulation is triggered by the exit), and brings parity between VMX
and SVM.
* Add pending request and interrupt injection information to the kvm_exit and
kvm_entry tracepoints respectively.
* Fix a relatively benign flaw where KVM would end up redoing RDPKRU when
loading guest/host PKRU, due to a refactoring of the kernel helpers that
didn't account for KVM's pre-checking of the need to do WRPKRU.
* Make the completion of hypercalls go through the complete_hypercall
function pointer argument, no matter if the hypercall exits to
userspace or not. Previously, the code assumed that KVM_HC_MAP_GPA_RANGE
specifically went to userspace, and all the others did not; the new code
need not special case KVM_HC_MAP_GPA_RANGE and in fact does not care at
all whether there was an exit to userspace or not.
* As part of enabling TDX virtual machines, support support separation of
private/shared EPT into separate roots. When TDX will be enabled, operations
on private pages will need to go through the privileged TDX Module via SEAMCALLs;
as a result, they are limited and relatively slow compared to reading a PTE.
The patches included in 6.14 allow KVM to keep a mirror of the private EPT in
host memory, and define entries in kvm_x86_ops to operate on external page
tables such as the TDX private EPT.
* The recently introduced conversion of the NX-page reclamation kthread to
vhost_task moved the task under the main process. The task is created as
soon as KVM_CREATE_VM was invoked and this, of course, broke userspace that
didn't expect to see any child task of the VM process until it started
creating its own userspace threads. In particular crosvm refuses to fork()
if procfs shows any child task, so unbreak it by creating the task lazily.
This is arguably a userspace bug, as there can be other kinds of legitimate
worker tasks and they wouldn't impede fork(); but it's not like userspace
has a way to distinguish kernel worker tasks right now. Should they show
as "Kthread: 1" in proc/.../status?
x86 - Intel:
* Fix a bug where KVM updates hardware's APICv cache of the highest ISR bit
while L2 is active, while ultimately results in a hardware-accelerated L1
EOI effectively being lost.
* Honor event priority when emulating Posted Interrupt delivery during nested
VM-Enter by queueing KVM_REQ_EVENT instead of immediately handling the
interrupt.
* Rework KVM's processing of the Page-Modification Logging buffer to reap
entries in the same order they were created, i.e. to mark gfns dirty in the
same order that hardware marked the page/PTE dirty.
* Misc cleanups.
Generic:
* Cleanup and harden kvm_set_memory_region(); add proper lockdep assertions when
setting memory regions and add a dedicated API for setting KVM-internal
memory regions. The API can then explicitly disallow all flags for
KVM-internal memory regions.
* Explicitly verify the target vCPU is online in kvm_get_vcpu() to fix a bug
where KVM would return a pointer to a vCPU prior to it being fully online,
and give kvm_for_each_vcpu() similar treatment to fix a similar flaw.
* Wait for a vCPU to come online prior to executing a vCPU ioctl, to fix a
bug where userspace could coerce KVM into handling the ioctl on a vCPU that
isn't yet onlined.
* Gracefully handle xarray insertion failures; even though such failures are
impossible in practice after xa_reserve(), reserving an entry is always followed
by xa_store() which does not know (or differentiate) whether there was an
xa_reserve() before or not.
RISC-V:
* Zabha, Svvptc, and Ziccrse extension support for guests. None of them
require anything in KVM except for detecting them and marking them
as supported; Zabha adds byte and halfword atomic operations, while the
others are markers for specific operation of the TLB and of LL/SC
instructions respectively.
* Virtualize SBI system suspend extension for Guest/VM
* Support firmware counters which can be used by the guests to collect
statistics about traps that occur in the host.
Selftests:
* Rework vcpu_get_reg() to return a value instead of using an out-param, and
update all affected arch code accordingly.
* Convert the max_guest_memory_test into a more generic mmu_stress_test.
The basic gist of the "conversion" is to have the test do mprotect() on
guest memory while vCPUs are accessing said memory, e.g. to verify KVM
and mmu_notifiers are working as intended.
* Play nice with treewrite builds of unsupported architectures, e.g. arm
(32-bit), as KVM selftests' Makefile doesn't do anything to ensure the
target architecture is actually one KVM selftests supports.
* Use the kernel's $(ARCH) definition instead of the target triple for arch
specific directories, e.g. arm64 instead of aarch64, mainly so as not to
be different from the rest of the kernel.
* Ensure that format strings for logging statements are checked by the
compiler even when the logging statement itself is disabled.
* Attempt to whack the last LLC references/misses mole in the Intel PMU
counters test by adding a data load and doing CLFLUSH{OPT} on the data
instead of the code being executed. It seems that modern Intel CPUs
have learned new code prefetching tricks that bypass the PMU counters.
* Fix a flaw in the Intel PMU counters test where it asserts that events
are counting correctly without actually knowing what the events count
given the underlying hardware; this can happen if Intel reuses a
formerly microarchitecture-specific event encoding as an architectural
event, as was the case for Top-Down Slots.
-----BEGIN PGP SIGNATURE-----
iQFIBAABCAAyFiEE8TM4V0tmI4mGbHaCv/vSX3jHroMFAmeTuzoUHHBib256aW5p
QHJlZGhhdC5jb20ACgkQv/vSX3jHroOkBwf8CRNExYaM3j9y2E7mmo6AiL2ug6+J
Uy5Hai1poY48pPwKC6ke3EWT8WVsgj/Py5pCeHvLojQchWNjCCYNfSQluJdkRxwG
DgP3QUljSxEJWBeSwyTRcKM+IySi5hZd1IFo3gePFRB829Jpnj05vjbvCyv8gIwU
y3HXxSYDsViaaFoNg4OlZFsIGis7mtknsZzk++QjuCXmxNa6UCbv3qvE/UkVLhVg
WH65RTRdjk+EsdwaOMHKuUvQoGa+iM4o39b6bqmw8+ZMK39+y33WeTX/y5RXsp1N
tUUBRfS+MuuYgC/6LmTr66EkMzoChxk3Dp3kKUaCBcfqRC8PxQag5reZhw==
=NEaO
-----END PGP SIGNATURE-----
Merge tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm
Pull kvm updates from Paolo Bonzini:
"Loongarch:
- Clear LLBCTL if secondary mmu mapping changes
- Add hypercall service support for usermode VMM
x86:
- Add a comment to kvm_mmu_do_page_fault() to explain why KVM
performs a direct call to kvm_tdp_page_fault() when RETPOLINE is
enabled
- Ensure that all SEV code is compiled out when disabled in Kconfig,
even if building with less brilliant compilers
- Remove a redundant TLB flush on AMD processors when guest CR4.PGE
changes
- Use str_enabled_disabled() to replace open coded strings
- Drop kvm_x86_ops.hwapic_irr_update() as KVM updates hardware's
APICv cache prior to every VM-Enter
- Overhaul KVM's CPUID feature infrastructure to track all vCPU
capabilities instead of just those where KVM needs to manage state
and/or explicitly enable the feature in hardware. Along the way,
refactor the code to make it easier to add features, and to make it
more self-documenting how KVM is handling each feature
- Rework KVM's handling of VM-Exits during event vectoring; this
plugs holes where KVM unintentionally puts the vCPU into infinite
loops in some scenarios (e.g. if emulation is triggered by the
exit), and brings parity between VMX and SVM
- Add pending request and interrupt injection information to the
kvm_exit and kvm_entry tracepoints respectively
- Fix a relatively benign flaw where KVM would end up redoing RDPKRU
when loading guest/host PKRU, due to a refactoring of the kernel
helpers that didn't account for KVM's pre-checking of the need to
do WRPKRU
- Make the completion of hypercalls go through the complete_hypercall
function pointer argument, no matter if the hypercall exits to
userspace or not.
Previously, the code assumed that KVM_HC_MAP_GPA_RANGE specifically
went to userspace, and all the others did not; the new code need
not special case KVM_HC_MAP_GPA_RANGE and in fact does not care at
all whether there was an exit to userspace or not
- As part of enabling TDX virtual machines, support support
separation of private/shared EPT into separate roots.
When TDX will be enabled, operations on private pages will need to
go through the privileged TDX Module via SEAMCALLs; as a result,
they are limited and relatively slow compared to reading a PTE.
The patches included in 6.14 allow KVM to keep a mirror of the
private EPT in host memory, and define entries in kvm_x86_ops to
operate on external page tables such as the TDX private EPT
- The recently introduced conversion of the NX-page reclamation
kthread to vhost_task moved the task under the main process. The
task is created as soon as KVM_CREATE_VM was invoked and this, of
course, broke userspace that didn't expect to see any child task of
the VM process until it started creating its own userspace threads.
In particular crosvm refuses to fork() if procfs shows any child
task, so unbreak it by creating the task lazily. This is arguably a
userspace bug, as there can be other kinds of legitimate worker
tasks and they wouldn't impede fork(); but it's not like userspace
has a way to distinguish kernel worker tasks right now. Should they
show as "Kthread: 1" in proc/.../status?
x86 - Intel:
- Fix a bug where KVM updates hardware's APICv cache of the highest
ISR bit while L2 is active, while ultimately results in a
hardware-accelerated L1 EOI effectively being lost
- Honor event priority when emulating Posted Interrupt delivery
during nested VM-Enter by queueing KVM_REQ_EVENT instead of
immediately handling the interrupt
- Rework KVM's processing of the Page-Modification Logging buffer to
reap entries in the same order they were created, i.e. to mark gfns
dirty in the same order that hardware marked the page/PTE dirty
- Misc cleanups
Generic:
- Cleanup and harden kvm_set_memory_region(); add proper lockdep
assertions when setting memory regions and add a dedicated API for
setting KVM-internal memory regions. The API can then explicitly
disallow all flags for KVM-internal memory regions
- Explicitly verify the target vCPU is online in kvm_get_vcpu() to
fix a bug where KVM would return a pointer to a vCPU prior to it
being fully online, and give kvm_for_each_vcpu() similar treatment
to fix a similar flaw
- Wait for a vCPU to come online prior to executing a vCPU ioctl, to
fix a bug where userspace could coerce KVM into handling the ioctl
on a vCPU that isn't yet onlined
- Gracefully handle xarray insertion failures; even though such
failures are impossible in practice after xa_reserve(), reserving
an entry is always followed by xa_store() which does not know (or
differentiate) whether there was an xa_reserve() before or not
RISC-V:
- Zabha, Svvptc, and Ziccrse extension support for guests. None of
them require anything in KVM except for detecting them and marking
them as supported; Zabha adds byte and halfword atomic operations,
while the others are markers for specific operation of the TLB and
of LL/SC instructions respectively
- Virtualize SBI system suspend extension for Guest/VM
- Support firmware counters which can be used by the guests to
collect statistics about traps that occur in the host
Selftests:
- Rework vcpu_get_reg() to return a value instead of using an
out-param, and update all affected arch code accordingly
- Convert the max_guest_memory_test into a more generic
mmu_stress_test. The basic gist of the "conversion" is to have the
test do mprotect() on guest memory while vCPUs are accessing said
memory, e.g. to verify KVM and mmu_notifiers are working as
intended
- Play nice with treewrite builds of unsupported architectures, e.g.
arm (32-bit), as KVM selftests' Makefile doesn't do anything to
ensure the target architecture is actually one KVM selftests
supports
- Use the kernel's $(ARCH) definition instead of the target triple
for arch specific directories, e.g. arm64 instead of aarch64,
mainly so as not to be different from the rest of the kernel
- Ensure that format strings for logging statements are checked by
the compiler even when the logging statement itself is disabled
- Attempt to whack the last LLC references/misses mole in the Intel
PMU counters test by adding a data load and doing CLFLUSH{OPT} on
the data instead of the code being executed. It seems that modern
Intel CPUs have learned new code prefetching tricks that bypass the
PMU counters
- Fix a flaw in the Intel PMU counters test where it asserts that
events are counting correctly without actually knowing what the
events count given the underlying hardware; this can happen if
Intel reuses a formerly microarchitecture-specific event encoding
as an architectural event, as was the case for Top-Down Slots"
* tag 'for-linus' of git://git.kernel.org/pub/scm/virt/kvm/kvm: (151 commits)
kvm: defer huge page recovery vhost task to later
KVM: x86/mmu: Return RET_PF* instead of 1 in kvm_mmu_page_fault()
KVM: Disallow all flags for KVM-internal memslots
KVM: x86: Drop double-underscores from __kvm_set_memory_region()
KVM: Add a dedicated API for setting KVM-internal memslots
KVM: Assert slots_lock is held when setting memory regions
KVM: Open code kvm_set_memory_region() into its sole caller (ioctl() API)
LoongArch: KVM: Add hypercall service support for usermode VMM
LoongArch: KVM: Clear LLBCTL if secondary mmu mapping is changed
KVM: SVM: Use str_enabled_disabled() helper in svm_hardware_setup()
KVM: VMX: read the PML log in the same order as it was written
KVM: VMX: refactor PML terminology
KVM: VMX: Fix comment of handle_vmx_instruction()
KVM: VMX: Reinstate __exit attribute for vmx_exit()
KVM: SVM: Use str_enabled_disabled() helper in sev_hardware_setup()
KVM: x86: Avoid double RDPKRU when loading host/guest PKRU
KVM: x86: Use LVT_TIMER instead of an open coded literal
RISC-V: KVM: Add new exit statstics for redirected traps
RISC-V: KVM: Update firmware counters for various events
RISC-V: KVM: Redirect instruction access fault trap to guest
...
SRSO_USER_KERNEL_NO denotes whether the CPU is affected by SRSO across
user/kernel boundaries. Advertise it to guest userspace.
Signed-off-by: Borislav Petkov (AMD) <bp@alien8.de>
Reviewed-by: Nikolay Borisov <nik.borisov@suse.com>
Link: https://lore.kernel.org/r/20241202120416.6054-3-bp@kernel.org
Refactor the kvm_cpu_cap_init() macro magic to collect supported features
in a local variable instead of passing them to the macro as a "mask". As
pointed out by Maxim, relying on macros to "return" a value and set local
variables is surprising, as the bitwise-OR logic suggests the macros are
pure, i.e. have no side effects.
Ideally, the feature initializers would have zero side effects, e.g. would
take local variables as params, but there isn't a sane way to do so
without either sacrificing the various compile-time assertions (basically
a non-starter), or passing at least one variable, e.g. a struct, to each
macro usage (adds a lot of noise and boilerplate code).
Opportunistically force callers to emit a trailing comma by intentionally
omitting a semicolon after invoking the feature initializers. Forcing a
trailing comma isotales futures changes to a single line, i.e. doesn't
cause churn for unrelated features/lines when adding/removing/modifying a
feature.
No functional change intended.
Suggested-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-58-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add one last (hopefully) CPUID feature macro, RUNTIME_F(), and use it
to track features that KVM supports, but that are only set at runtime
(in response to other state), and aren't advertised to userspace via
KVM_GET_SUPPORTED_CPUID.
Currently, RUNTIME_F() is mostly just documentation, but tracking all
KVM-supported features will allow for asserting, at build time, take),
that all features that are set, cleared, *or* checked by KVM are known to
kvm_set_cpu_caps().
No functional change intended.
Link: https://lore.kernel.org/r/20241128013424.4096668-57-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add another CPUID feature macro, VENDOR_F(), and use it to track features
that KVM supports, but that need additional vendor support and so are
conditionally enabled in vendor code.
Currently, VENDOR_F() is mostly just documentation, but tracking all
KVM-supported features will allow for asserting, at build time, take),
that all features that are set, cleared, *or* checked by KVM are known to
kvm_set_cpu_caps().
To fudge around a macro collision on 32-bit kernels, #undef DS to be able
to get at X86_FEATURE_DS.
No functional change intended.
Link: https://lore.kernel.org/r/20241128013424.4096668-56-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Now that each feature flag is on its own line, i.e. brevity isn't a major
concern, drop the "SF" acronym and use the (almost) full name, SCATTERED_F.
No functional change intended.
Link: https://lore.kernel.org/r/20241128013424.4096668-55-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Don't memcpy() all of boot_cpu_data.x86_capability, and instead explicitly
fill each kvm_cpu_cap_init leaf during kvm_cpu_cap_init(). While clever,
copying all kernel capabilities risks over-reporting KVM capabilities,
e.g. if KVM added support in __do_cpuid_func(), but neglected to init the
supported set of capabilities.
Note, explicitly grabbing leafs deliberately keeps Linux-defined leafs as
0! KVM should never advertise Linux-defined leafs; any relevant features
that are "real", but scattered, must be gathered in their correct hardware-
defined leaf.
Link: https://lore.kernel.org/r/20241128013424.4096668-54-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add yet another CPUID macro, this time for features that the host kernel
synthesizes into boot_cpu_data, i.e. that the kernel force sets even in
situations where the feature isn't reported by CPUID. Thanks to the
macro shenanigans of kvm_cpu_cap_init(), such features can now be handled
in the core CPUID framework, i.e. don't need to be handled out-of-band and
thus without as many guardrails.
Adding a dedicated macro also helps document what's going on, e.g. the
calls to kvm_cpu_cap_check_and_set() are very confusing unless the reader
knows exactly how kvm_cpu_cap_init() generates kvm_cpu_caps (and even
then, it's far from obvious).
Link: https://lore.kernel.org/r/20241128013424.4096668-53-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Switch all queries (except XSAVES) of guest features from guest CPUID to
guest capabilities, i.e. replace all calls to guest_cpuid_has() with calls
to guest_cpu_cap_has().
Keep guest_cpuid_has() around for XSAVES, but subsume its helper
guest_cpuid_get_register() and add a compile-time assertion to prevent
using guest_cpuid_has() for any other feature. Add yet another comment
for XSAVE to explain why KVM is allowed to query its raw guest CPUID.
Opportunistically drop the unused guest_cpuid_clear(), as there should be
no circumstance in which KVM needs to _clear_ a guest CPUID feature now
that everything is tracked via cpu_caps. E.g. KVM may need to _change_
a feature to emulate dynamic CPUID flags, but KVM should never need to
clear a feature in guest CPUID to prevent it from being used by the guest.
Delete the last remnants of the governed features framework, as the lone
holdout was vmx_adjust_secondary_exec_control()'s divergent behavior for
governed vs. ungoverned features.
Note, replacing guest_cpuid_has() checks with guest_cpu_cap_has() when
computing reserved CR4 bits is a nop when viewed as a whole, as KVM's
capabilities are already incorporated into the calculation, i.e. if a
feature is present in guest CPUID but unsupported by KVM, its CR4 bit
was already being marked as reserved, checking guest_cpu_cap_has() simply
double-stamps that it's a reserved bit.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-51-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
When updating guest CPUID entries to emulate runtime behavior, e.g. when
the guest enables a CR4-based feature that is tied to a CPUID flag, also
update the vCPU's cpu_caps accordingly. This will allow replacing all
usage of guest_cpuid_has() with guest_cpu_cap_has().
Note, this relies on kvm_set_cpuid() taking a snapshot of cpu_caps before
invoking kvm_update_cpuid_runtime(), i.e. when KVM is updating CPUID
entries that *may* become the vCPU's CPUID, so that unwinding to the old
cpu_caps is possible if userspace tries to set bogus CPUID information.
Note #2, none of the features in question use guest_cpu_cap_has() at this
time, i.e. aside from settings bits in cpu_caps, this is a glorified nop.
Cc: Yang Weijiang <weijiang.yang@intel.com>
Cc: Robert Hoo <robert.hoo.linux@gmail.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-49-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
When making runtime CPUID updates, change OSXSAVE and OSPKE even if their
respective base features (XSAVE, PKU) are not supported by the host. KVM
already incorporates host support in the vCPU's effective reserved CR4 bits.
I.e. OSXSAVE and OSPKE can be set if and only if the host supports them.
And conversely, since KVM's ABI is that KVM owns the dynamic OS feature
flags, clearing them when they obviously aren't supported and thus can't
be enabled is arguably a fix.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-48-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Drop an unnecessary check that kvm_find_cpuid_entry_index(), i.e.
cpuid_entry2_find(), returns the correct leaf when getting CPUID.0x7.0x0
to update X86_FEATURE_OSPKE. cpuid_entry2_find() never returns an entry
for the wrong function. And not that it matters, but cpuid_entry2_find()
will always return a precise match for CPUID.0x7.0x0 since the index is
significant.
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-47-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Move the handling of X86_FEATURE_MWAIT during CPUID runtime updates to
utilize the lookup done for other CPUID.0x1 features.
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-46-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Constrain all guest cpu_caps based on KVM support instead of constraining
only the few features that KVM _currently_ needs to verify are actually
supported by KVM. The intent of cpu_caps is to track what the guest is
actually capable of using, not the raw, unfiltered CPUID values that the
guest sees.
I.e. KVM should always consult it's only support when making decisions
based on guest CPUID, and the only reason KVM has historically made the
checks opt-in was due to lack of centralized tracking.
Suggested-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-45-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Enumerate MWAIT in cpuid_func_emulated(), but only if the caller wants to
include "partially emulated" features, i.e. features that KVM kinda sorta
emulates, but with major caveats. This will allow initializing the guest
cpu_caps based on the set of features that KVM virtualizes and/or emulates,
without needing to handle things like MONITOR/MWAIT as one-off exceptions.
Adding one-off handling for individual features is quite painful,
especially when considering future hardening. It's very doable to verify,
at compile time, that every CPUID-based feature that KVM queries when
emulating guest behavior is actually known to KVM, e.g. to prevent KVM
bugs where KVM emulates some feature but fails to advertise support to
userspace. In other words, any features that are special cased, i.e. not
handled generically in the CPUID framework, would also need to be special
cased for any hardening efforts that build on said framework.
Link: https://lore.kernel.org/r/20241128013424.4096668-44-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Extract the meat of __do_cpuid_func_emulated() into a separate helper,
cpuid_func_emulated(), so that cpuid_func_emulated() can be used with a
single CPUID entry. This will allow marking emulated features as fully
supported in the guest cpu_caps without needing to hardcode the set of
emulated features in multiple locations.
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-43-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Initialize a vCPU's capabilities based on the guest CPUID provided by
userspace instead of simply zeroing the entire array. This is the first
step toward using cpu_caps to query *all* CPUID-based guest capabilities,
i.e. will allow converting all usage of guest_cpuid_has() to
guest_cpu_cap_has().
Zeroing the array was the logical choice when using cpu_caps was opt-in,
e.g. "unsupported" was generally a safer default, and the whole point of
governed features is that KVM would need to check host and guest support,
i.e. making everything unsupported by default didn't require more code.
But requiring KVM to manually "enable" every CPUID-based feature in
cpu_caps would require an absurd amount of boilerplate code.
Follow existing CPUID/kvm_cpu_caps nomenclature where possible, e.g. for
the change() and clear() APIs. Replace check_and_set() with constrain()
to try and capture that KVM is constraining userspace's desired guest
feature set based on KVM's capabilities.
This is intended to be gigantic nop, i.e. should not have any impact on
guest or KVM functionality.
This is also an intermediate step; a future commit will also incorporate
KVM support into the vCPU's cpu_caps before converting guest_cpuid_has()
to guest_cpu_cap_has().
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-42-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Replace the internals of the governed features framework with a more
comprehensive "guest CPU capabilities" implementation, i.e. with a guest
version of kvm_cpu_caps. Keep the skeleton of governed features around
for now as vmx_adjust_sec_exec_control() relies on detecting governed
features to do the right thing for XSAVES, and switching all guest feature
queries to guest_cpu_cap_has() requires subtle and non-trivial changes,
i.e. is best done as a standalone change.
Tracking *all* guest capabilities that KVM cares will allow excising the
poorly named "governed features" framework, and effectively optimizes all
KVM queries of guest capabilities, i.e. doesn't require making a
subjective decision as to whether or not a feature is worth "governing",
and doesn't require adding the code to do so.
The cost of tracking all features is currently 92 bytes per vCPU on 64-bit
kernels: 100 bytes for cpu_caps versus 8 bytes for governed_features.
That cost is well worth paying even if the only benefit was eliminating
the "governed features" terminology. And practically speaking, the real
cost is zero unless those 92 bytes pushes the size of vcpu_vmx or vcpu_svm
into a new order-N allocation, and if that happens there are better ways
to reduce the footprint of kvm_vcpu_arch, e.g. making the PMU and/or MTRR
state separate allocations.
Suggested-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-41-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
As the first step toward replacing KVM's so-called "governed features"
framework with a more comprehensive, less poorly named implementation,
replace the "kvm_governed_feature" function prefix with "guest_cpu_cap"
and rename guest_can_use() to guest_cpu_cap_has().
The "guest_cpu_cap" naming scheme mirrors that of "kvm_cpu_cap", and
provides a more clear distinction between guest capabilities, which are
KVM controlled (heh, or one might say "governed"), and guest CPUID, which
with few exceptions is fully userspace controlled.
Opportunistically rewrite the comment about XSS passthrough for SEV-ES
guests to avoid referencing so many functions, as such comments are prone
to becoming stale (case in point...).
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-40-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Unconditionally advertise "support" for the HYPERVISOR feature in CPUID,
as the flag simply communicates to the guest that's it's running under a
hypervisor.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-39-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Unconditionally advertise TSC_DEADLINE_TIMER via KVM_GET_SUPPORTED_CPUID,
as KVM always emulates deadline mode, *if* the VM has an in-kernel local
APIC. The odds of a VMM emulating the local APIC in userspace, not
emulating the TSC deadline timer, _and_ reflecting
KVM_GET_SUPPORTED_CPUID back into KVM_SET_CPUID2, i.e. the risk of
over-advertising and breaking any setups, is extremely low.
KVM has _unconditionally_ advertised X2APIC via CPUID since commit
0d1de2d901 ("KVM: Always report x2apic as supported feature"), and it
is completely impossible for userspace to emulate X2APIC as KVM doesn't
support forwarding the MSR accesses to userspace. I.e. KVM has relied on
userspace VMMs to not misreport local APIC capabilities for nearly 13
years.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-38-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Convert all use of cpuid_entry2_find() to kvm_find_cpuid_entry{,index}()
now that cpuid_entry2_find() operates on the vCPU state, i.e. now that
there is no need to use cpuid_entry2_find() directly in order to pass in
non-vCPU state.
To help prevent unwanted usage of cpuid_entry2_find(), #undef
KVM_CPUID_INDEX_NOT_SIGNIFICANT, i.e. force KVM to use
kvm_find_cpuid_entry().
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-37-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Move kvm_find_cpuid_entry{,_index}() "up" in cpuid.c so that they are
colocated with cpuid_entry2_find(), e.g. to make it easier to see the
effective guts of the helpers without having to bounce around cpuid.c.
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-36-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Now that KVM sets vcpu->arch.cpuid_{entries,nent} before processing the
incoming CPUID entries during KVM_SET_CPUID{,2}, drop the @entries and
@nent params from cpuid_entry2_find() and unconditionally operate on the
vCPU state.
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-35-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Now that KVM only searches for KVM's PV CPUID base when userspace sets
guest CPUID, drop the cache and simply do the search every time.
Practically speaking, this is a nop except for situations where userspace
sets CPUID _after_ running the vCPU, which is anything but a hot path,
e.g. QEMU does so only when hotplugging a vCPU. And on the flip side,
caching guest CPUID information, especially information that is used to
query/modify _other_ CPUID state, is inherently dangerous as it's all too
easy to use stale information, i.e. KVM should only cache CPUID state when
the performance and/or programming benefits justify it.
Link: https://lore.kernel.org/r/20241128013424.4096668-34-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Now that KVM disallows disabling HLT-exiting after vCPUs have been created,
i.e. now that it's impossible for kvm_hlt_in_guest() to change while vCPUs
are running, apply KVM's PV_UNHALT quirk only when userspace is setting
guest CPUID.
Opportunistically rename the helper to make it clear that KVM's behavior
is a quirk that should never have been added. KVM's documentation
explicitly states that userspace should not advertise PV_UNHALT if
HLT-exiting is disabled, but for unknown reasons, commit caa057a2ca
("KVM: X86: Provide a capability to disable HLT intercepts") didn't stop
at documenting the requirement and also massaged the incoming guest CPUID.
Unfortunately, it's quite likely that userspace has come to rely on KVM's
behavior, i.e. the code can't simply be deleted. The only reason KVM
doesn't have an "official" quirk is that there is no known use case where
disabling the quirk would make sense, i.e. letting userspace disable the
quirk would further increase KVM's burden without any benefit.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-33-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
When handling KVM_SET_CPUID{,2}, swap the old and new CPUID arrays and
lengths before processing the new CPUID, and simply undo the swap if
setting the new CPUID fails for whatever reason.
To keep the diff reasonable, continue passing the entry array and length
to most helpers, and defer the more complete cleanup to future commits.
For any sane VMM, setting "bad" CPUID state is not a hot path (or even
something that is surviable), and setting guest CPUID before it's known
good will allow removing all of KVM's infrastructure for processing CPUID
entries directly (as opposed to operating on vcpu->arch.cpuid_entries).
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-32-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Now that kvm_cpu_cap_init() is a macro with its own scope, add EMUL_F() to
OR-in features that KVM emulates in software, i.e. that don't depend on
the feature being available in hardware. The contained scope
of kvm_cpu_cap_init() allows using a local variable to track the set of
emulated leaves, which in addition to avoiding confusing and/or
unnecessary variables, helps prevent misuse of EMUL_F().
Link: https://lore.kernel.org/r/20241128013424.4096668-31-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add a macro for use in kvm_set_cpu_caps() to automagically initialize
features that KVM wants to support based solely on the CPU's capabilities,
e.g. KVM advertises LA57 support if it's available in hardware, even if
the host kernel isn't utilizing 57-bit virtual addresses.
Track a features that are passed through to userspace (from hardware) in
a local variable, and simply OR them in *after* adjusting the capabilities
that came from boot_cpu_data.
Note, eliminating the open-coded call to cpuid_ecx() also fixes a largely
benign bug where KVM could incorrectly report LA57 support on Intel CPUs
whose max supported CPUID is less than 7, i.e. if the max supported leaf
(<7) happened to have bit 16 set. In practice, barring a funky virtual
machine setup, the bug is benign as all known CPUs that support VMX also
support leaf 7.
Link: https://lore.kernel.org/r/20241128013424.4096668-30-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add compile-time assertions to verify that usage of F() and friends in
kvm_set_cpu_caps() is scoped to the correct CPUID word, e.g. to detect
bugs where KVM passes a feature bit from word X into word y.
Add a one-off assertion in the aliased feature macro to ensure that only
word 0x8000_0001.EDX aliased the features defined for 0x1.EDX.
To do so, convert kvm_cpu_cap_init() to a macro and have it define a
local variable to track which CPUID word is being initialized that is
then used to validate usage of F() (all of the inputs are compile-time
constants and thus can be fed into BUILD_BUG_ON()).
Redefine KVM_VALIDATE_CPU_CAP_USAGE after kvm_set_cpu_caps() to be a nop
so that F() can be used in other flows that aren't as easily hardened,
e.g. __do_cpuid_func_emulated() and __do_cpuid_func().
Invoke KVM_VALIDATE_CPU_CAP_USAGE() in SF() and X86_64_F() to ensure the
validation occurs, e.g. if the usage of F() is completely compiled out
(which shouldn't happen for boot_cpu_has(), but could happen in the future,
e.g. if KVM were to use cpu_feature_enabled()).
Link: https://lore.kernel.org/r/20241128013424.4096668-29-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Undefine SPEC_CTRL_SSBD, which is #defined by msr-index.h to represent the
enable flag in MSR_IA32_SPEC_CTRL, to avoid issues with the macro being
unpacked into its raw value when passed to KVM's F() macro. This will
allow using multiple layers of macros in F() and friends, e.g. to harden
against incorrect usage of F().
No functional change intended (cpuid.c doesn't consume SPEC_CTRL_SSBD).
Link: https://lore.kernel.org/r/20241128013424.4096668-28-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Merge kvm_cpu_cap_init() and kvm_cpu_cap_init_kvm_defined() into a single
helper. The only advantage of separating the two was to make it somewhat
obvious that KVM directly initializes the KVM-defined words, whereas using
a common helper will allow for hardening both kernel- and KVM-defined
CPUID words without needing copy+paste.
No functional change intended.
Link: https://lore.kernel.org/r/20241128013424.4096668-27-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add a macro to precisely handle CPUID features that AMD duplicated from
CPUID.0x1.EDX into CPUID.0x8000_0001.EDX. This will allow adding an
assert that all features passed to kvm_cpu_cap_init() match the word being
processed, e.g. to prevent passing a feature from CPUID 0x7 to CPUID 0x1.
Because the kernel simply reuses the X86_FEATURE_* definitions from
CPUID.0x1.EDX, KVM's use of the aliased features would result in false
positives from such an assert.
No functional change intended.
Link: https://lore.kernel.org/r/20241128013424.4096668-26-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Add a macro to mask-in feature flags that are supported only on 64-bit
kernels/KVM. In addition to reducing overall #ifdeffery, using a macro
will allow hardening the kvm_cpu_cap initialization sequences to assert
that the features being advertised are indeed included in the word being
initialized. And arguably using *F() macros through is more readable.
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-25-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Rename kvm_cpu_cap_mask() to kvm_cpu_cap_init() in anticipation of merging
it with kvm_cpu_cap_init_kvm_defined(), and in anticipation of _setting_
bits in the helper (a future commit will play macro games to set emulated
feature flags via kvm_cpu_cap_init()).
No functional change intended.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-24-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Refactor kvm_set_cpu_caps() to express each supported (or not) feature
flag on a separate line, modulo a handful of cases where KVM does not, and
likely will not, support a sequence of flags. This will allow adding
fancier macros with longer, more descriptive names without resulting in
absurd line lengths and/or weird code. Isolating each flag also makes it
far easier to review changes, reduces code conflicts, and generally makes
it easier to resolve conflicts. Lastly, it allows co-locating comments
for notable flags, e.g. MONITOR, precisely with the relevant flag.
No functional change intended.
Suggested-by: Maxim Levitsky <mlevitsk@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-23-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Explicitly zero out the feature word in kvm_cpu_caps if the word's
associated CPUID function is greater than the max leaf supported by the
CPU. For such unsupported functions, Intel CPUs return the output from
the last supported leaf, not all zeros.
Practically speaking, this is likely a benign bug, as KVM uses the raw
host CPUID to mask the kernel's computed capabilities, and the kernel does
perform max leaf checks when populating boot_cpu_data. The only way KVM's
goof could be problematic is if the kernel force-set a feature in a leaf
that is completely unsupported, _and_ the max supported leaf happened to
return a value with '1' the same bit position. Which is theoretically
possible, but extremely unlikely. And even if that did happen, it's
entirely possible that KVM would still provide the correct functionality;
the kernel did set the capability after all.
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-22-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Revert the chunk of commit 01b4f510b9 ("kvm: x86: ensure pv_cpuid.features
is initialized when enabling cap") that forced a PV features cache refresh
during KVM_CAP_ENFORCE_PV_FEATURE_CPUID, as whatever ioctl() ordering
issue it alleged to have fixed never existed upstream, and likely never
existed in any kernel.
At the time of the commit, there was a tangentially related ioctl()
ordering issue, as toggling KVM_X86_DISABLE_EXITS_HLT after KVM_SET_CPUID2
would have resulted in KVM potentially leaving KVM_FEATURE_PV_UNHALT set.
But (a) that bug affected the entire guest CPUID, not just the cache, (b)
commit 01b4f510b9 didn't address that bug, it only refreshed the cache
(with the bad CPUID), and (c) setting KVM_X86_DISABLE_EXITS_HLT after vCPU
creation is completely broken as KVM configures HLT-exiting only during
vCPU creation, which is why KVM_CAP_X86_DISABLE_EXITS is now disallowed if
vCPUs have been created.
Another tangentially related bug was KVM's failure to clear the cache when
handling KVM_SET_CPUID2, but again commit 01b4f510b9 did nothing to fix
that bug.
The most plausible explanation for the what commit 01b4f510b9 was trying
to fix is a bug that existed in Google's internal kernel that was the
source of commit 01b4f510b9. At the time, Google's internal kernel had
not yet picked up commit 0d3b2ba16b ("KVM: X86: Go on updating other
CPUID leaves when leaf 1 is absent"), i.e. KVM would not initialize the
PV features cache if KVM_SET_CPUID2 was called without a CPUID.0x1 entry.
Of course, no sane real world VMM would omit CPUID.0x1, including the KVM
selftest added by commit ac4a4d6de2 ("selftests: kvm: test enforcement
of paravirtual cpuid features"). And the test didn't actually try to
verify multiple orderings, nor did the selftest enter the guest without
doing KVM_SET_CPUID2, so who knows what motivated the change.
Regardless of why commit 01b4f510b9 ("kvm: x86: ensure pv_cpuid.features
is initialized when enabling cap") was added, refreshing the cache during
KVM_CAP_ENFORCE_PV_FEATURE_CPUID isn't necessary.
Cc: Oliver Upton <oliver.upton@linux.dev>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-20-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Clear KVM's PV feature cache prior when processing a new guest CPUID so
that KVM doesn't keep a stale cache entry if userspace does KVM_SET_CPUID2
multiple times, once with a PV features entry, and a second time without.
Fixes: 66570e966d ("kvm: x86: only provide PV features if enabled in guest's CPUID")
Cc: Oliver Upton <oliver.upton@linux.dev>
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-19-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Drop x86.c's local pre-computed cr4_reserved bits and instead fold KVM's
reserved bits into the guest's reserved bits. This fixes a bug where VMX's
set_cr4_guest_host_mask() fails to account for KVM-reserved bits when
deciding which bits can be passed through to the guest. In most cases,
letting the guest directly write reserved CR4 bits is ok, i.e. attempting
to set the bit(s) will still #GP, but not if a feature is available in
hardware but explicitly disabled by the host, e.g. if FSGSBASE support is
disabled via "nofsgsbase".
Note, the extra overhead of computing host reserved bits every time
userspace sets guest CPUID is negligible. The feature bits that are
queried are packed nicely into a handful of words, and so checking and
setting each reserved bit costs in the neighborhood of ~5 cycles, i.e. the
total cost will be in the noise even if the number of checked CR4 bits
doubles over the next few years. In other words, x86 will run out of CR4
bits long before the overhead becomes problematic.
Note #2, __cr4_reserved_bits() starts from CR4_RESERVED_BITS, which is
why the existing __kvm_cpu_cap_has() processing doesn't explicitly OR in
CR4_RESERVED_BITS (and why the new code doesn't do so either).
Fixes: 2ed41aa631 ("KVM: VMX: Intercept guest reserved CR4 bits to inject #GP fault")
Reviewed-by: Maxim Levitsky <mlevitsk@redhat.com>
Reviewed-by: Chao Gao <chao.gao@intel.com>
Reviewed-by: Binbin Wu <binbin.wu@linux.intel.com>
Reviewed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-6-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Explicitly perform runtime CPUID adjustments as part of the "after set
CPUID" flow to guard against bugs where KVM consumes stale vCPU/CPUID
state during kvm_update_cpuid_runtime(). E.g. see commit 4736d85f0d
("KVM: x86: Use actual kvm_cpuid.base for clearing KVM_FEATURE_PV_UNHALT").
Whacking each mole individually is not sustainable or robust, e.g. while
the aforemention commit fixed KVM's PV features, the same issue lurks for
Xen and Hyper-V features, Xen and Hyper-V simply don't have any runtime
features (though spoiler alert, neither should KVM).
Updating runtime features in the "full" path will also simplify adding a
snapshot of the guest's capabilities, i.e. of caching the intersection of
guest CPUID and kvm_cpu_caps (modulo a few edge cases).
Link: https://lore.kernel.org/r/20241128013424.4096668-5-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
During vCPU creation, process KVM's default, empty CPUID as if userspace
set an empty CPUID to ensure consistent and correct behavior with respect
to guest CPUID. E.g. if userspace never sets guest CPUID, KVM will never
configure cr4_guest_rsvd_bits, and thus create divergent, incorrect, guest-
visible behavior due to letting the guest set any KVM-supported CR4 bits
despite the features not being allowed per guest CPUID.
Note! This changes KVM's ABI, as lack of full CPUID processing allowed
userspace to stuff garbage vCPU state, e.g. userspace could set CR4 to a
guest-unsupported value via KVM_SET_SREGS. But it's extremely unlikely
that this is a breaking change, as KVM already has many flows that require
userspace to set guest CPUID before loading vCPU state. E.g. multiple MSR
flows consult guest CPUID on host writes, and KVM_SET_SREGS itself already
relies on guest CPUID being up-to-date, as KVM's validity check on CR3
consumes CPUID.0x7.1 (for LAM) and CPUID.0x80000008 (for MAXPHYADDR).
Furthermore, the plan is to commit to enforcing guest CPUID for userspace
writes to MSRs, at which point bypassing sregs CPUID checks is even more
nonsensical.
Link: https://lore.kernel.org/r/20241128013424.4096668-4-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Define and undefine the F() and SF() macros precisely around
kvm_set_cpu_caps() to make it all but impossible to use the macros outside
of kvm_cpu_cap_{mask,init_kvm_defined}(). Currently, F() is a simple
passthrough, but SF() is actively dangerous as it checks that the scattered
feature is supported by the host kernel.
And usage outside of the aforementioned helpers will run afoul of future
changes to harden KVM's CPUID management.
Opportunistically switch to feature_bit() when stuffing LA57 based on raw
hardware support.
No functional change intended.
Link: https://lore.kernel.org/r/20241128013424.4096668-3-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
When clearing CONSTANT_TSC during CPUID emulation due to a Hyper-V quirk,
use feature_bit() instead of SF() to ensure the bit is actually cleared.
SF() evaluates to zero if the _host_ doesn't support the feature. I.e.
KVM could keep the bit set if userspace advertised CONSTANT_TSC despite
it not being supported in hardware.
Note, translating from a scattered feature to a the hardware version is
done by __feature_translate(), not SF(). The sole purpose of SF() is to
check kernel support for the scattered feature, *before* translation.
Cc: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Link: https://lore.kernel.org/r/20241128013424.4096668-2-seanjc@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Snapshot the output of CPUID.0xD.[1..n] during kvm.ko initiliaization to
avoid the overead of CPUID during runtime. The offset, size, and metadata
for CPUID.0xD.[1..n] sub-leaves does not depend on XCR0 or XSS values, i.e.
is constant for a given CPU, and thus can be cached during module load.
On Intel's Emerald Rapids, CPUID is *wildly* expensive, to the point where
recomputing XSAVE offsets and sizes results in a 4x increase in latency of
nested VM-Enter and VM-Exit (nested transitions can trigger
xstate_required_size() multiple times per transition), relative to using
cached values. The issue is easily visible by running `perf top` while
triggering nested transitions: kvm_update_cpuid_runtime() shows up at a
whopping 50%.
As measured via RDTSC from L2 (using KVM-Unit-Test's CPUID VM-Exit test
and a slightly modified L1 KVM to handle CPUID in the fastpath), a nested
roundtrip to emulate CPUID on Skylake (SKX), Icelake (ICX), and Emerald
Rapids (EMR) takes:
SKX 11650
ICX 22350
EMR 28850
Using cached values, the latency drops to:
SKX 6850
ICX 9000
EMR 7900
The underlying issue is that CPUID itself is slow on ICX, and comically
slow on EMR. The problem is exacerbated on CPUs which support XSAVES
and/or XSAVEC, as KVM invokes xstate_required_size() twice on each
runtime CPUID update, and because there are more supported XSAVE features
(CPUID for supported XSAVE feature sub-leafs is significantly slower).
SKX:
CPUID.0xD.2 = 348 cycles
CPUID.0xD.3 = 400 cycles
CPUID.0xD.4 = 276 cycles
CPUID.0xD.5 = 236 cycles
<other sub-leaves are similar>
EMR:
CPUID.0xD.2 = 1138 cycles
CPUID.0xD.3 = 1362 cycles
CPUID.0xD.4 = 1068 cycles
CPUID.0xD.5 = 910 cycles
CPUID.0xD.6 = 914 cycles
CPUID.0xD.7 = 1350 cycles
CPUID.0xD.8 = 734 cycles
CPUID.0xD.9 = 766 cycles
CPUID.0xD.10 = 732 cycles
CPUID.0xD.11 = 718 cycles
CPUID.0xD.12 = 734 cycles
CPUID.0xD.13 = 1700 cycles
CPUID.0xD.14 = 1126 cycles
CPUID.0xD.15 = 898 cycles
CPUID.0xD.16 = 716 cycles
CPUID.0xD.17 = 748 cycles
CPUID.0xD.18 = 776 cycles
Note, updating runtime CPUID information multiple times per nested
transition is itself a flaw, especially since CPUID is a mandotory
intercept on both Intel and AMD. E.g. KVM doesn't need to ensure emulated
CPUID state is up-to-date while running L2. That flaw will be fixed in a
future patch, as deferring runtime CPUID updates is more subtle than it
appears at first glance, the benefits aren't super critical to have once
the XSAVE issue is resolved, and caching CPUID output is desirable even if
KVM's updates are deferred.
Cc: Jim Mattson <jmattson@google.com>
Cc: stable@vger.kernel.org
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-ID: <20241211013302.1347853-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Latest Intel platform Clearwater Forest has introduced new instructions
enumerated by CPUIDs of SHA512, SM3, SM4 and AVX-VNNI-INT16. Advertise
these CPUIDs to userspace so that guests can query them directly.
SHA512, SM3 and SM4 are on an expected-dense CPUID leaf and some other
bits on this leaf have kernel usages. Considering they have not truly
kernel usages, hide them in /proc/cpuinfo.
These new instructions only operate in xmm, ymm registers and have no new
VMX controls, so there is no additional host enabling required for guests
to use these instructions, i.e. advertising these CPUIDs to userspace is
safe.
Tested-by: Jiaan Lu <jiaan.lu@intel.com>
Tested-by: Xuelian Guo <xuelian.guo@intel.com>
Signed-off-by: Tao Su <tao1.su@linux.intel.com>
Message-ID: <20241105054825.870939-1-tao1.su@linux.intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
This is an inherent feature of IA32_PRED_CMD[0], so it is trivially
virtualizable (as long as IA32_PRED_CMD[0] is virtualized).
Suggested-by: Tom Lendacky <thomas.lendacky@amd.com>
Signed-off-by: Jim Mattson <jmattson@google.com>
Reviewed-by: Tom Lendacky <thomas.lendacky@amd.com>
Reviewed-by: Thomas Gleixner <tglx@linutronix.de>
Link: https://lore.kernel.org/r/20241011214353.1625057-4-jmattson@google.com
Signed-off-by: Sean Christopherson <seanjc@google.com>
Advertise AVX10.1 related CPUIDs, i.e. report AVX10 support bit via
CPUID.(EAX=07H, ECX=01H):EDX[bit 19] and new CPUID leaf 0x24H so that
guest OS and applications can query the AVX10.1 CPUIDs directly. Intel
AVX10 represents the first major new vector ISA since the introduction of
Intel AVX512, which will establish a common, converged vector instruction
set across all Intel architectures[1].
AVX10.1 is an early version of AVX10, that enumerates the Intel AVX512
instruction set at 128, 256, and 512 bits which is enabled on
Granite Rapids. I.e., AVX10.1 is only a new CPUID enumeration with no
new functionality. New features, e.g. Embedded Rounding and Suppress
All Exceptions (SAE) will be introduced in AVX10.2.
Advertising AVX10.1 is safe because there is nothing to enable for AVX10.1,
i.e. it's purely a new way to enumerate support, thus there will never be
anything for the kernel to enable. Note just the CPUID checking is changed
when using AVX512 related instructions, e.g. if using one AVX512
instruction needs to check (AVX512 AND AVX512DQ), it can check
((AVX512 AND AVX512DQ) OR AVX10.1) after checking XCR0[7:5].
The versions of AVX10 are expected to be inclusive, e.g. version N+1 is
a superset of version N. Per the spec, the version can never be 0, just
advertise AVX10.1 if it's supported in hardware. Moreover, advertising
AVX10_{128,256,512} needs to land in the same commit as advertising basic
AVX10.1 support, otherwise KVM would advertise an impossible CPU model.
E.g. a CPU with AVX512 but not AVX10.1/512 is impossible per the SDM.
As more and more AVX related CPUIDs are added (it would have resulted in
around 40-50 CPUID flags when developing AVX10), the versioning approach
is introduced. But incrementing version numbers are bad for virtualization.
E.g. if AVX10.2 has a feature that shouldn't be enumerated to guests for
whatever reason, then KVM can't enumerate any "later" features either,
because the only way to hide the problematic AVX10.2 feature is to set the
version to AVX10.1 or lower[2]. But most AVX features are just passed
through and don't have virtualization controls, so AVX10 should not be
problematic in practice, so long as Intel honors their promise that future
versions will be supersets of past versions.
[1] https://cdrdv2.intel.com/v1/dl/getContent/784267
[2] https://lore.kernel.org/all/Zkz5Ak0PQlAN8DxK@google.com/
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Tao Su <tao1.su@linux.intel.com>
Link: https://lore.kernel.org/r/20240819062327.3269720-1-tao1.su@linux.intel.com
[sean: minor changelog tweaks]
Signed-off-by: Sean Christopherson <seanjc@google.com>
