Re: REGRESSION on linux-next (next-20250919)

[Sean Christopherson <seanjc@google.com>]

On Tue, Sep 30, 2025, Dapeng Mi wrote:
> 
> On 9/30/2025 1:30 PM, Borah, Chaitanya Kumar wrote:
> > Hello Sean,
> >
> > Hope you are doing well. I am Chaitanya from the linux graphics team in 
> > Intel.
> >
> > This mail is regarding a regression we are seeing in our CI runs[1] on
> > linux-next repository.
> >
> > Since the version next-20250919 [2], we are seeing the following regression
> >
> > `````````````````````````````````````````````````````````````````````````````````
> > <4>[   10.973827] ------------[ cut here ]------------
> > <4>[   10.973841] WARNING: arch/x86/events/core.c:3089 at 
> > perf_get_x86_pmu_capability+0xd/0xc0, CPU#15: (udev-worker)/386
> > ...
> > <4>[   10.974028] Call Trace:
> > <4>[   10.974030]  <TASK>
> > <4>[   10.974033]  ? kvm_init_pmu_capability+0x2b/0x190 [kvm]
> > <4>[   10.974154]  kvm_x86_vendor_init+0x1b0/0x1a40 [kvm]
> > <4>[   10.974248]  vmx_init+0xdb/0x260 [kvm_intel]
> > <4>[   10.974278]  ? __pfx_vt_init+0x10/0x10 [kvm_intel]
> > <4>[   10.974296]  vt_init+0x12/0x9d0 [kvm_intel]
> > <4>[   10.974309]  ? __pfx_vt_init+0x10/0x10 [kvm_intel]
> > <4>[   10.974322]  do_one_initcall+0x60/0x3f0
> > <4>[   10.974335]  do_init_module+0x97/0x2b0
> > <4>[   10.974345]  load_module+0x2d08/0x2e30
> > <4>[   10.974349]  ? __kernel_read+0x158/0x2f0
> > <4>[   10.974370]  ? kernel_read_file+0x2b1/0x320
> > <4>[   10.974381]  init_module_from_file+0x96/0xe0
> > <4>[   10.974384]  ? init_module_from_file+0x96/0xe0
> > <4>[   10.974399]  idempotent_init_module+0x117/0x330
> > <4>[   10.974415]  __x64_sys_finit_module+0x73/0xe0
> > ...
> > `````````````````````````````````````````````````````````````````````````````````
> > Details log can be found in [3].
> >
> > After bisecting the tree, the following patch [4] seems to be the first 
> > "bad" commit
> >
> > `````````````````````````````````````````````````````````````````````````````````````````````````````````
> >  From 51f34b1e650fc5843530266cea4341750bd1ae37 Mon Sep 17 00:00:00 2001
> >
> > From: Sean Christopherson <seanjc@google.com>
> >
> > Date: Wed, 6 Aug 2025 12:56:39 -0700
> >
> > Subject: KVM: x86/pmu: Snapshot host (i.e. perf's) reported PMU capabilities
> >
> > Take a snapshot of the unadulterated PMU capabilities provided by perf so
> > that KVM can compare guest vPMU capabilities against hardware capabilities
> > when determining whether or not to intercept PMU MSRs (and RDPMC).
> > `````````````````````````````````````````````````````````````````````````````````````````````````````````
> >
> > We also verified that if we revert the patch the issue is not seen.
> >
> > Could you please check why the patch causes this regression and provide 
> > a fix if necessary?
> 
> Hi Chaitanya,
> 
> I suppose you found this warning on a hybrid client platform, right? It
> looks the warning is triggered by the below WARN_ON_ONCE() in
> perf_get_x86_pmu_capability() function.
> 
>   if (WARN_ON_ONCE(cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) ||
>         !x86_pmu_initialized()) {
>         memset(cap, 0, sizeof(*cap));
>         return;
>     }
> 
> The below change should fix it (just building, not test it). I would run a
> full scope vPMU test after I come back from China national day's holiday.

I have access to a hybrid system, I'll also double check there (though I'm 99.9%
certain you've got it right).

> Thanks.
> 
> diff --git a/arch/x86/kvm/pmu.c b/arch/x86/kvm/pmu.c
> index cebce7094de8..6d87c25226d8 100644
> --- a/arch/x86/kvm/pmu.c
> +++ b/arch/x86/kvm/pmu.c
> @@ -108,8 +108,6 @@ void kvm_init_pmu_capability(struct kvm_pmu_ops *pmu_ops)
>         bool is_intel = boot_cpu_data.x86_vendor == X86_VENDOR_INTEL;
>         int min_nr_gp_ctrs = pmu_ops->MIN_NR_GP_COUNTERS;
> 
> -       perf_get_x86_pmu_capability(&kvm_host_pmu);
> -
>         /*
>          * Hybrid PMUs don't play nice with virtualization without careful
>          * configuration by userspace, and KVM's APIs for reporting supported
> @@ -120,6 +118,8 @@ void kvm_init_pmu_capability(struct kvm_pmu_ops *pmu_ops)
>                 enable_pmu = false;
> 
>         if (enable_pmu) {
> +               perf_get_x86_pmu_capability(&kvm_host_pmu);
> +
>                 /*
>                  * WARN if perf did NOT disable hardware PMU if the number of
>                  * architecturally required GP counters aren't present, i.e. if

If we go this route, then the !enable_pmu path should explicitly zero kvm_host_pmu
so that the behavior is consistent userspace loads kvm.ko with enable_pmu=0,
versus enable_pmu being cleared because of lack of support.

	if (!enable_pmu) {
		memset(&kvm_host_pmu, 0, sizeof(kvm_host_pmu));
		memset(&kvm_pmu_cap, 0, sizeof(kvm_pmu_cap));
		return;
	}

The alternative would be keep kvm_host_pmu valid at all times for !HYBRID, which
is what I intended with the bad patch, but that too would lead to inconsistent
behavior.  So I think it makes sense to go with Dapeng's approach; we can always
revisit this if some future thing in KVM _needs_ kvm_host_pmu even with enable_pmu=0. 

	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU)) {
		enable_pmu = false;
		memset(&kvm_host_pmu, 0, sizeof(kvm_host_pmu));
	} else {
		perf_get_x86_pmu_capability(&kvm_host_pmu);
	}