Hi Like,
On 11/20/22 22:42, Like Xu wrote:
> On 19/11/2022 8:28 pm, Dongli Zhang wrote:
>> This patchset is to fix two svm pmu virtualization bugs.
>>
>> 1. The 1st bug is that "-cpu,-pmu" cannot disable svm pmu virtualization.
>>
>> To use "-cpu EPYC" or "-cpu host,-pmu" cannot disable the pmu
>> virtualization. There is still below at the VM linux side ...
>
> Many QEMU vendor forks already have similar fixes, and
> thanks for bringing this issue back to the mainline.
Would you mind helping point to if there used to be any prior patchset for
mainline to resolve the issue?
>
>>
>> [ 0.510611] Performance Events: Fam17h+ core perfctr, AMD PMU driver.
>>
>> ... although we expect something like below.
>>
>> [ 0.596381] Performance Events: PMU not available due to virtualization,
>> using software events only.
>> [ 0.600972] NMI watchdog: Perf NMI watchdog permanently disabled
>>
>> The patch 1-2 is to disable the pmu virtualization via KVM_PMU_CAP_DISABLE
>> if the per-vcpu "pmu" property is disabled.
>>
>> I considered 'KVM_X86_SET_MSR_FILTER' initially.
>> Since both KVM_X86_SET_MSR_FILTER and KVM_PMU_CAP_DISABLE are VM ioctl. I
>> finally used the latter because it is easier to use.
>>
>>
>> 2. The 2nd bug is that un-reclaimed perf events (after QEMU system_reset)
>> at the KVM side may inject random unwanted/unknown NMIs to the VM.
>>
>> The svm pmu registers are not reset during QEMU system_reset.
>>
>> (1). The VM resets (e.g., via QEMU system_reset or VM kdump/kexec) while it
>> is running "perf top". The pmu registers are not disabled gracefully.
>>
>> (2). Although the x86_cpu_reset() resets many registers to zero, the
>> kvm_put_msrs() does not puts AMD pmu registers to KVM side. As a result,
>> some pmu events are still enabled at the KVM side.
>>
>> (3). The KVM pmc_speculative_in_use() always returns true so that the events
>> will not be reclaimed. The kvm_pmc->perf_event is still active.
>
> I'm not sure if you're saying KVM doing something wrong, I don't think so
> because KVM doesn't sense the system_reset defined by QEME or other user space,
> AMD's vPMC will continue to be enabled (if it was enabled before), generating pmi
> injection into the guest, and the newly started guest doesn't realize the
> counter is still
> enabled and blowing up the error log.
I were not saying KVM was buggy.
I was trying to explain how the issue impacts KVM side and VM side.
>
>>
>> (4). After the reboot, the VM kernel reports below error:
>>
>> [ 0.092011] Performance Events: Fam17h+ core perfctr, Broken BIOS detected,
>> complain to your hardware vendor.
>> [ 0.092023] [Firmware Bug]: the BIOS has corrupted hw-PMU resources (MSR
>> c0010200 is 530076)
>>
>> (5). In a worse case, the active kvm_pmc->perf_event is still able to
>> inject unknown NMIs randomly to the VM kernel.
>>
>> [...] Uhhuh. NMI received for unknown reason 30 on CPU 0.
>>
>> The patch 3 is to fix the issue by resetting AMD pmu registers as well as
>> Intel registers.
>
> This fix idea looks good, it does require syncing the new changed device state
> of QEMU to KVM.
Thank you very much!
Dongli Zhang
>
>>
>>
>> This patchset does cover does not cover PerfMonV2, until the below patchset
>> is merged into the KVM side.
>>
>> [PATCH v3 0/8] KVM: x86: Add AMD Guest PerfMonV2 PMU support
>> https://lore.kernel.org/all/20221111102645.82001-1-likexu@tencent.com/
>>
>>
>> Dongli Zhang (3):
>> kvm: introduce a helper before creating the 1st vcpu
>> i386: kvm: disable KVM_CAP_PMU_CAPABILITY if "pmu" is disabled
>> target/i386/kvm: get and put AMD pmu registers
>>
>> accel/kvm/kvm-all.c | 7 ++-
>> include/sysemu/kvm.h | 2 +
>> target/arm/kvm64.c | 4 ++
>> target/i386/cpu.h | 5 +++
>> target/i386/kvm/kvm.c | 104 +++++++++++++++++++++++++++++++++++++++++++-
>> target/mips/kvm.c | 4 ++
>> target/ppc/kvm.c | 4 ++
>> target/riscv/kvm.c | 4 ++
>> target/s390x/kvm/kvm.c | 4 ++
>> 9 files changed, 134 insertions(+), 4 deletions(-)
>>
>> Thank you very much!
>>
>> Dongli Zhang
>>
>>
>>