Omit RIP from the emulator's _regs array, which is used only for GPRs,
i.e. registers that can be referenced via ModRM and/or SIB bytes. The
emulator uses the dedicated _eip field for RIP, and manually reads from
_eip to handle RIP-relative addressing.
To avoid an even bigger, slightly more dangerous change, hardcode the
number of GPRs to 16 for the time being even though 32-bit KVM's emulator
technically should only have 8 GPRs. Add a TODO to address that in a
future commit.
See also the comments above the read_gpr() and write_gpr() declarations,
and obviously the handling in writeback_registers().
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Message-Id: <20220526210817.3428868-4-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
WARN and truncate the incoming GPR number/index when reading/writing GPRs
in the emulator to guard against KVM bugs, e.g. to avoid out-of-bounds
accesses to ctxt->_regs[] if KVM generates a bogus index. Truncate the
index instead of returning e.g. zero, as reg_write() returns a pointer
to the register, i.e. returning zero would result in a NULL pointer
dereference. KVM could also force the index to any arbitrary GPR, but
that's no better or worse, just different.
Open code the restriction to 16 registers; RIP is handled via _eip and
should never be accessed through reg_read() or reg_write(). See the
comments above the declarations of reg_read() and reg_write(), and the
behavior of writeback_registers(). The horrific open coded mess will be
cleaned up in a future commit.
There are no such bugs known to exist in the emulator, but determining
that KVM is bug-free is not at all simple and requires a deep dive into
the emulator. The code is so convoluted that GCC-12 with the recently
enable -Warray-bounds spits out a false-positive due to a GCC bug:
arch/x86/kvm/emulate.c:254:27: warning: array subscript 32 is above array
bounds of 'long unsigned int[17]' [-Warray-bounds]
254 | return ctxt->_regs[nr];
| ~~~~~~~~~~~^~~~
In file included from arch/x86/kvm/emulate.c:23:
arch/x86/kvm/kvm_emulate.h: In function 'reg_rmw':
arch/x86/kvm/kvm_emulate.h:366:23: note: while referencing '_regs'
366 | unsigned long _regs[NR_VCPU_REGS];
| ^~~~~
Link: https://lore.kernel.org/all/YofQlBrlx18J7h9Y@google.com
Link: https://bugzilla.kernel.org/show_bug.cgi?id=216026
Link: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105679
Reported-and-tested-by: Robert Dinse <nanook@eskimo.com>
Reported-by: Kees Cook <keescook@chromium.org>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220526210817.3428868-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Capture ctxt->regs_dirty in a local 'unsigned long' instead of casting it
to an 'unsigned long *' for use in for_each_set_bit(). The bitops helpers
really do read the entire 'unsigned long', even though the walking of the
read value is capped at the specified size. I.e. 64-bit KVM is reading
memory beyond ctxt->regs_dirty, which is a u32 and thus 4 bytes, whereas
an unsigned long is 8 bytes. Functionally it's not an issue because
regs_dirty is in the middle of x86_emulate_ctxt, i.e. KVM is just reading
its own memory, but relying on that coincidence is gross and unsafe.
Reviewed-by: Vitaly Kuznetsov <vkuznets@redhat.com>
Reviewed-by: Kees Cook <keescook@chromium.org>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220526210817.3428868-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
s390:
* add an interface to provide a hypervisor dump for secure guests
* improve selftests to show tests
x86:
* Intel IPI virtualization
* Allow getting/setting pending triple fault with KVM_GET/SET_VCPU_EVENTS
* PEBS virtualization
* Simplify PMU emulation by just using PERF_TYPE_RAW events
* More accurate event reinjection on SVM (avoid retrying instructions)
* Allow getting/setting the state of the speaker port data bit
* Rewrite gfn-pfn cache refresh
* Refuse starting the module if VM-Entry/VM-Exit controls are inconsistent
* "Notify" VM exit
The selftests nested code only supports 4-level paging at the moment.
This means it cannot map nested guest physical addresses with more than
48 bits. Allow perf_test_util nested mode to work on hosts with more
than 48 physical addresses by restricting the guest test region to
48-bits.
While here, opportunistically fix an off-by-one error when dealing with
vm_get_max_gfn(). perf_test_util.c was treating this as the maximum
number of GFNs, rather than the maximum allowed GFN. This didn't result
in any correctness issues, but it did end up shifting the test region
down slightly when using huge pages.
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220520233249.3776001-12-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add an option to dirty_log_perf_test that configures the vCPUs to run in
L2 instead of L1. This makes it possible to benchmark the dirty logging
performance of nested virtualization, which is particularly interesting
because KVM must shadow L1's EPT/NPT tables.
For now this support only works on x86_64 CPUs with VMX. Otherwise
passing -n results in the test being skipped.
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220520233249.3776001-11-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The linker does obey strong/weak symbols when linking static libraries,
it simply resolves an undefined symbol to the first-encountered symbol.
This means that defining __weak arch-generic functions and then defining
arch-specific strong functions to override them in libkvm will not
always work.
More specifically, if we have:
lib/generic.c:
void __weak foo(void)
{
pr_info("weak\n");
}
void bar(void)
{
foo();
}
lib/x86_64/arch.c:
void foo(void)
{
pr_info("strong\n");
}
And a selftest that calls bar(), it will print "weak". Now if you make
generic.o explicitly depend on arch.o (e.g. add function to arch.c that
is called directly from generic.c) it will print "strong". In other
words, it seems that the linker is free to throw out arch.o when linking
because generic.o does not explicitly depend on it, which causes the
linker to lose the strong symbol.
One solution is to link libkvm.a with --whole-archive so that the linker
doesn't throw away object files it thinks are unnecessary. However that
is a bit difficult to plumb since we are using the common selftests
makefile rules. An easier solution is to drop libkvm.a just link
selftests with all the .o files that were originally in libkvm.a.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220520233249.3776001-9-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Refactor nested_map() to specify that it explicityl wants 4K mappings
(the existing behavior) and push the implementation down into
__nested_map(), which can be used in subsequent commits to create huge
page mappings.
No function change intended.
Reviewed-by: Peter Xu <peterx@redhat.com>
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220520233249.3776001-5-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The current EPT mapping code in the selftests only supports mapping 4K
pages. This commit extends that support with an option to map at 2M or
1G. This will be used in a future commit to create large page mappings
to test eager page splitting.
No functional change intended.
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220520233249.3776001-3-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
x86_page_size is an enum used to communicate the desired page size with
which to map a range of memory. Under the hood they just encode the
desired level at which to map the page. This ends up being clunky in a
few ways:
- The name suggests it encodes the size of the page rather than the
level.
- In other places in x86_64/processor.c we just use a raw int to encode
the level.
Simplify this by adopting the kernel style of PG_LEVEL_XX enums and pass
around raw ints when referring to the level. This makes the code easier
to understand since these macros are very common in KVM MMU code.
Signed-off-by: David Matlack <dmatlack@google.com>
Message-Id: <20220520233249.3776001-2-dmatlack@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Commit 74fd41ed16 ("KVM: x86: nSVM: support PAUSE filtering when L0
doesn't intercept PAUSE") introduced passthrough support for nested pause
filtering, (when the host doesn't intercept PAUSE) (either disabled with
kvm module param, or disabled with '-overcommit cpu-pm=on')
Before this commit, L1 KVM didn't intercept PAUSE at all; afterwards,
the feature was exposed as supported by KVM cpuid unconditionally, thus
if L1 could try to use it even when the L0 KVM can't really support it.
In this case the fallback caused KVM to intercept each PAUSE instruction;
in some cases, such intercept can slow down the nested guest so much
that it can fail to boot. Instead, before the problematic commit KVM
was already setting both thresholds to 0 in vmcb02, but after the first
userspace VM exit shrink_ple_window was called and would reset the
pause_filter_count to the default value.
To fix this, change the fallback strategy - ignore the guest threshold
values, but use/update the host threshold values unless the guest
specifically requests disabling PAUSE filtering (either simple or
advanced).
Also fix a minor bug: on nested VM exit, when PAUSE filter counter
were copied back to vmcb01, a dirty bit was not set.
Thanks a lot to Suravee Suthikulpanit for debugging this!
Fixes: 74fd41ed16 ("KVM: x86: nSVM: support PAUSE filtering when L0 doesn't intercept PAUSE")
Reported-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Tested-by: Suravee Suthikulpanit <suravee.suthikulpanit@amd.com>
Co-developed-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220518072709.730031-1-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Now that these functions are always called with preemption disabled,
remove the preempt_disable()/preempt_enable() pair inside them.
No functional change intended.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220606180829.102503-8-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On SVM, if preemption happens right after the call to finish_rcuwait
but before call to kvm_arch_vcpu_unblocking on SVM/AVIC, it itself
will re-enable AVIC, and then we will try to re-enable it again
in kvm_arch_vcpu_unblocking which will lead to a warning
in __avic_vcpu_load.
The same problem can happen if the vCPU is preempted right after the call
to kvm_arch_vcpu_blocking but before the call to prepare_to_rcuwait
and in this case, we will end up with AVIC enabled during sleep -
Ooops.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220606180829.102503-7-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Currently nothing prevents preemption in kvm_vcpu_update_apicv.
On SVM, If the preemption happens after we update the
vcpu->arch.apicv_active, the preemption itself will
'update' the inhibition since the AVIC will be first disabled
on vCPU unload and then enabled, when the current task
is loaded again.
Then we will try to update it again, which will lead to a warning
in __avic_vcpu_load, that the AVIC is already enabled.
Fix this by disabling preemption in this code.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220606180829.102503-6-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There are two issues in avic_kick_target_vcpus_fast
1. It is legal to issue an IPI request with APIC_DEST_NOSHORT
and a physical destination of 0xFF (or 0xFFFFFFFF in case of x2apic),
which must be treated as a broadcast destination.
Fix this by explicitly checking for it.
Also don’t use ‘index’ in this case as it gives no new information.
2. It is legal to issue a logical IPI request to more than one target.
Index field only provides index in physical id table of first
such target and therefore can't be used before we are sure
that only a single target was addressed.
Instead, parse the ICRL/ICRH, double check that a unicast interrupt
was requested, and use that info to figure out the physical id
of the target vCPU.
At that point there is no need to use the index field as well.
In addition to fixing the above issues, also skip the call to
kvm_apic_match_dest.
It is possible to do this now, because now as long as AVIC is not
inhibited, it is guaranteed that none of the vCPUs changed their
apic id from its default value.
This fixes boot of windows guest with AVIC enabled because it uses
IPI with 0xFF destination and no destination shorthand.
Fixes: 7223fd2d53 ("KVM: SVM: Use target APIC ID to complete AVIC IRQs when possible")
Cc: stable@vger.kernel.org
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220606180829.102503-5-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
AVIC is now inhibited if the guest changes the apic id,
and therefore this code is no longer needed.
There are several ways this code was broken, including:
1. a vCPU was only allowed to change its apic id to an apic id
of an existing vCPU.
2. After such change, the vCPU whose apic id entry was overwritten,
could not correctly change its own apic id, because its own
entry is already overwritten.
Signed-off-by: Maxim Levitsky <mlevitsk@redhat.com>
Message-Id: <20220606180829.102503-4-mlevitsk@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Assign shadow_me_value, not shadow_me_mask, to PAE root entries,
a.k.a. shadow PDPTRs, when host memory encryption is supported. The
"mask" is the set of all possible memory encryption bits, e.g. MKTME
KeyIDs, whereas "value" holds the actual value that needs to be
stuffed into host page tables.
Using shadow_me_mask results in a failed VM-Entry due to setting
reserved PA bits in the PDPTRs, and ultimately causes an OOPS due to
physical addresses with non-zero MKTME bits sending to_shadow_page()
into the weeds:
set kvm_intel.dump_invalid_vmcs=1 to dump internal KVM state.
BUG: unable to handle page fault for address: ffd43f00063049e8
PGD 86dfd8067 P4D 0
Oops: 0000 [#1] PREEMPT SMP
RIP: 0010:mmu_free_root_page+0x3c/0x90 [kvm]
kvm_mmu_free_roots+0xd1/0x200 [kvm]
__kvm_mmu_unload+0x29/0x70 [kvm]
kvm_mmu_unload+0x13/0x20 [kvm]
kvm_arch_destroy_vm+0x8a/0x190 [kvm]
kvm_put_kvm+0x197/0x2d0 [kvm]
kvm_vm_release+0x21/0x30 [kvm]
__fput+0x8e/0x260
____fput+0xe/0x10
task_work_run+0x6f/0xb0
do_exit+0x327/0xa90
do_group_exit+0x35/0xa0
get_signal+0x911/0x930
arch_do_signal_or_restart+0x37/0x720
exit_to_user_mode_prepare+0xb2/0x140
syscall_exit_to_user_mode+0x16/0x30
do_syscall_64+0x4e/0x90
entry_SYSCALL_64_after_hwframe+0x44/0xae
Fixes: e54f1ff244 ("KVM: x86/mmu: Add shadow_me_value and repurpose shadow_me_mask")
Signed-off-by: Yuan Yao <yuan.yao@intel.com>
Reviewed-by: Kai Huang <kai.huang@intel.com>
Message-Id: <20220608012015.19566-1-yuan.yao@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
KVM/arm64 fixes for 5.19, take #1
- Properly reset the SVE/SME flags on vcpu load
- Fix a vgic-v2 regression regarding accessing the pending
state of a HW interrupt from userspace (and make the code
common with vgic-v3)
- Fix access to the idreg range for protected guests
- Ignore 'kvm-arm.mode=protected' when using VHE
- Return an error from kvm_arch_init_vm() on allocation failure
- A bunch of small cleanups (comments, annotations, indentation)
The layout of 'struct kvm_vcpu_arch' has evolved significantly since
the initial port of KVM/arm64, so remove the stale comment suggesting
that a prefix of the structure is used exclusively from assembly code.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220609121223.2551-7-will@kernel.org
has_vhe() expands to a compile-time constant when evaluated from the VHE
or nVHE code, alternatively checking a static key when called from
elsewhere in the kernel. On face value, this looks like a case of
premature optimization, but in fact this allows symbol references on
VHE-specific code paths to be dropped from the nVHE object.
Expand the comment in has_vhe() to make this clearer, hopefully
discouraging anybody from simplifying the code.
Cc: David Brazdil <dbrazdil@google.com>
Acked-by: Mark Rutland <mark.rutland@arm.com>
Signed-off-by: Will Deacon <will@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220609121223.2551-5-will@kernel.org
A protected VM accessing ID_AA64ISAR2_EL1 gets punished with an UNDEF,
while it really should only get a zero back if the register is not
handled by the hypervisor emulation (as mandated by the architecture).
Introduce all the missing ID registers (including the unallocated ones),
and have them to return 0.
Reported-by: Will Deacon <will@kernel.org>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220609121223.2551-3-will@kernel.org
Commit fed9b26b25 ("MAINTAINERS: Update KVM RISC-V entry to cover
selftests support") optimistically adds a file entry for
tools/testing/selftests/kvm/riscv/, but this directory does not exist.
Hence, ./scripts/get_maintainer.pl --self-test=patterns complains about a
broken reference. The script is very useful to keep MAINTAINERS up to date
and MAINTAINERS can be kept in a state where the script emits no warning.
So, just drop the non-matching file entry rather than starting to collect
exceptions of entries that may match in some close or distant future.
Fixes: fed9b26b25 ("MAINTAINERS: Update KVM RISC-V entry to cover selftests support")
Signed-off-by: Lukas Bulwahn <lukas.bulwahn@gmail.com>
Signed-off-by: Anup Patel <anup@brainfault.org>
Various spelling mistakes in comments.
Detected with the help of Coccinelle.
Signed-off-by: Julia Lawall <Julia.Lawall@inria.fr>
Signed-off-by: Anup Patel <anup@brainfault.org>
Currently the state of the speaker port (0x61) data bit (bit 1) is not
saved in the exported state (kvm_pit_state2) and hence is lost when
re-constructing guest state.
This patch removes the 'speaker_data_port' field from kvm_kpit_state and
instead tracks the state using a new KVM_PIT_FLAGS_SPEAKER_DATA_ON flag
defined in the API.
Signed-off-by: Paul Durrant <pdurrant@amazon.com>
Message-Id: <20220531124421.1427-1-pdurrant@amazon.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add an on-by-default module param, error_on_inconsistent_vmcs_config, to
allow rejecting the load of kvm_intel if an inconsistent VMCS config is
detected. Continuing on with an inconsistent, degraded config is
undesirable in the vast majority of use cases, e.g. may result in a
misconfigured VM, poor performance due to lack of fast MSR switching, or
even security issues in the unlikely event the guest is relying on MPX.
Practically speaking, an inconsistent VMCS config should never be
encountered in a production quality environment, e.g. on bare metal it
indicates a silicon defect (or a disturbing lack of validation by the
hardware vendor), and in a virtualized machine (KVM as L1) it indicates a
buggy/misconfigured L0 VMM/hypervisor.
Provide a module param to override the behavior for testing purposes, or
in the unlikely scenario that KVM is deployed on a flawed-but-usable CPU
or virtual machine.
Note, what is or isn't an inconsistency is somewhat subjective, e.g. one
might argue that LOAD_EFER without SAVE_EFER is an inconsistency. KVM's
unofficial guideline for an "inconsistency" is either scenarios that are
completely nonsensical, e.g. the existing checks on having EPT/VPID knobs
without EPT/VPID, and/or scenarios that prevent KVM from virtualizing or
utilizing a feature, e.g. the unpaired entry/exit controls checks. Other
checks that fall into one or both of the covered scenarios could be added
in the future, e.g. asserting that a VMCS control exists available if and
only if the associated feature is supported in bare metal.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220527170658.3571367-3-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Sanitize the VM-Entry/VM-Exit control pairs (load+load or load+clear)
during setup instead of checking both controls in a pair at runtime. If
only one control is supported, KVM will report the associated feature as
not available, but will leave the supported control bit set in the VMCS
config, which could lead to corruption of host state. E.g. if only the
VM-Entry control is supported and the feature is not dynamically toggled,
KVM will set the control in all VMCSes and load zeros without restoring
host state.
Note, while this is technically a bug fix, practically speaking no sane
CPU or VMM would support only one control. KVM's behavior of checking
both controls is mostly pedantry.
Cc: Chenyi Qiang <chenyi.qiang@intel.com>
Cc: Lei Wang <lei4.wang@intel.com>
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220527170658.3571367-2-seanjc@google.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Whenever an MSR is part of KVM_GET_MSR_INDEX_LIST, as is the case for
MSR_K7_EVNTSEL0 or MSR_F15H_PERF_CTL0, it has to be always retrievable
and settable with KVM_GET_MSR and KVM_SET_MSR.
Accept a zero value for these MSRs to obey the contract.
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220601031925.59693-1-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Once vPMU is disabled, the KVM would not expose features like:
PEBS (via clear kvm_pmu_cap.pebs_ept), legacy LBR and ARCH_LBR,
CPUID 0xA leaf, PDCM bit and MSR_IA32_PERF_CAPABILITIES, plus
PT_MODE_HOST_GUEST mode.
What this group of features has in common is that their use
relies on the underlying PMU counter and the host perf_event as a
back-end resource requester or sharing part of the irq delivery path.
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220601031925.59693-2-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
The BTS feature (including the ability to set the BTS and BTINT
bits in the DEBUGCTL MSR) is currently unsupported on KVM.
But we may try using the BTS facility on a PEBS enabled guest like this:
perf record -e branches:u -c 1 -d ls
and then we would encounter the following call trace:
[] unchecked MSR access error: WRMSR to 0x1d9 (tried to write 0x00000000000003c0)
at rIP: 0xffffffff810745e4 (native_write_msr+0x4/0x20)
[] Call Trace:
[] intel_pmu_enable_bts+0x5d/0x70
[] bts_event_add+0x54/0x70
[] event_sched_in+0xee/0x290
As it lacks any CPUID indicator or perf_capabilities valid bit
fields to prompt for this information, the platform would hint
the Intel BTS feature unavailable to guest by setting the
BTS_UNAVAIL bit in the IA32_MISC_ENABLE.
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220601031925.59693-3-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
On some virt platforms (L1 guest w/o PMU), the value of module parameter
'enable_pmu' for nested L2 guests should be updated at initialisation.
Considering that there is no concept of "architecture pmu" in AMD or Hygon
and that the versions (prior to Zen 4) are all 0, but that the theoretical
available counters are at least AMD64_NUM_COUNTERS, the utility
check_hw_exists() is reused in the initialisation call path.
Opportunistically update Intel specific comments.
Fixes: 8eeac7e999e8 ("KVM: x86/pmu: Add kvm_pmu_cap to optimize perf_get_x86_pmu_capability")
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220518170118.66263-3-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
If the PMU is broken due to firmware issues, check_hw_exists() will return
false but perf_get_x86_pmu_capability() will still return data from x86_pmu.
Likewise if some of the hotplug callbacks cannot be installed the contents
of x86_pmu will not be reverted.
Handle the failure in both cases by clearing x86_pmu if init_hw_perf_events()
or reverts to software events only.
Co-developed-by: Like Xu <likexu@tencent.com>
Signed-off-by: Like Xu <likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
There are cases that malicious virtual machines can cause CPU stuck (due
to event windows don't open up), e.g., infinite loop in microcode when
nested #AC (CVE-2015-5307). No event window means no event (NMI, SMI and
IRQ) can be delivered. It leads the CPU to be unavailable to host or
other VMs.
VMM can enable notify VM exit that a VM exit generated if no event
window occurs in VM non-root mode for a specified amount of time (notify
window).
Feature enabling:
- The new vmcs field SECONDARY_EXEC_NOTIFY_VM_EXITING is introduced to
enable this feature. VMM can set NOTIFY_WINDOW vmcs field to adjust
the expected notify window.
- Add a new KVM capability KVM_CAP_X86_NOTIFY_VMEXIT so that user space
can query and enable this feature in per-VM scope. The argument is a
64bit value: bits 63:32 are used for notify window, and bits 31:0 are
for flags. Current supported flags:
- KVM_X86_NOTIFY_VMEXIT_ENABLED: enable the feature with the notify
window provided.
- KVM_X86_NOTIFY_VMEXIT_USER: exit to userspace once the exits happen.
- It's safe to even set notify window to zero since an internal hardware
threshold is added to vmcs.notify_window.
VM exit handling:
- Introduce a vcpu state notify_window_exits to records the count of
notify VM exits and expose it through the debugfs.
- Notify VM exit can happen incident to delivery of a vector event.
Allow it in KVM.
- Exit to userspace unconditionally for handling when VM_CONTEXT_INVALID
bit is set.
Nested handling
- Nested notify VM exits are not supported yet. Keep the same notify
window control in vmcs02 as vmcs01, so that L1 can't escape the
restriction of notify VM exits through launching L2 VM.
Notify VM exit is defined in latest Intel Architecture Instruction Set
Extensions Programming Reference, chapter 9.2.
Co-developed-by: Xiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: Xiaoyao Li <xiaoyao.li@intel.com>
Signed-off-by: Tao Xu <tao3.xu@intel.com>
Co-developed-by: Chenyi Qiang <chenyi.qiang@intel.com>
Signed-off-by: Chenyi Qiang <chenyi.qiang@intel.com>
Message-Id: <20220524135624.22988-5-chenyi.qiang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add kvm_caps to hold a variety of capabilites and defaults that aren't
handled by kvm_cpu_caps because they aren't CPUID bits in order to reduce
the amount of boilerplate code required to add a new feature. The vast
majority (all?) of the caps interact with vendor code and are written
only during initialization, i.e. should be tagged __read_mostly, declared
extern in x86.h, and exported.
No functional change intended.
Signed-off-by: Sean Christopherson <seanjc@google.com>
Message-Id: <20220524135624.22988-4-chenyi.qiang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
Add a selftest for triple fault event:
- launch the L2 and exit to userspace via I/O.
- using KVM_SET_VCPU_EVENTS to pend a triple fault event.
- with the immediate_exit, check the triple fault is pending.
- run for real with pending triple fault and L1 can see the triple
fault.
Suggested-by: Sean Christopherson <seanjc@google.com>
Signed-off-by: Chenyi Qiang <chenyi.qiang@intel.com>
Message-Id: <20220524135624.22988-3-chenyi.qiang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
For the triple fault sythesized by KVM, e.g. the RSM path or
nested_vmx_abort(), if KVM exits to userspace before the request is
serviced, userspace could migrate the VM and lose the triple fault.
Extend KVM_{G,S}ET_VCPU_EVENTS to support pending triple fault with a
new event KVM_VCPUEVENT_VALID_FAULT_FAULT so that userspace can save and
restore the triple fault event. This extension is guarded by a new KVM
capability KVM_CAP_TRIPLE_FAULT_EVENT.
Note that in the set_vcpu_events path, userspace is able to set/clear
the triple fault request through triple_fault.pending field.
Signed-off-by: Chenyi Qiang <chenyi.qiang@intel.com>
Message-Id: <20220524135624.22988-2-chenyi.qiang@intel.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
A recurrent bug in the KVM/arm64 code base consists in trying to
access the timer pending state outside of the vcpu context, which
makes zero sense (the pending state only exists when the vcpu
is loaded).
In order to avoid more embarassing crashes and catch the offenders
red-handed, add a warning to kvm_arch_timer_get_input_level() and
return the state as non-pending. This avoids taking the system down,
and still helps tracking down silly bugs.
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Signed-off-by: Marc Zyngier <maz@kernel.org>
Link: https://lore.kernel.org/r/20220607131427.1164881-4-maz@kernel.org
Now that GICv2 has a proper userspace accessor for the pending state,
switch GICv3 over to it, dropping the local version, moving over the
specific behaviours that CGIv3 requires (such as the distinction
between pending latch and line level which were never enforced
with GICv2).
We also gain extra locking that isn't really necessary for userspace,
but that's a small price to pay for getting rid of superfluous code.
Signed-off-by: Marc Zyngier <maz@kernel.org>
Reviewed-by: Eric Auger <eric.auger@redhat.com>
Link: https://lore.kernel.org/r/20220607131427.1164881-3-maz@kernel.org
All gp or fixed counters have been reprogrammed using PERF_TYPE_RAW,
which means that the table that maps perf_hw_id to event select values is
no longer useful, at least for AMD.
For Intel, the logic to check if the pmu event reported by Intel cpuid is
not available is still required, in which case pmc_perf_hw_id() could be
renamed to hw_event_is_unavail() and a bool value is returned to replace
the semantics of "PERF_COUNT_HW_MAX+1".
Signed-off-by: Like Xu <likexu@tencent.com>
Message-Id: <20220518132512.37864-12-likexu@tencent.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
