linux-kernel - Re: [kvm-unit-tests PATCH 2/2] x86/pmu: Add AMD Guest PerfMonV2 testcases

lists.openwall.net		lists / announce owl-users owl-dev john-users john-dev passwdqc-users yescrypt popa3d-users / oss-security kernel-hardening musl sabotage tlsify passwords / crypt-dev xvendor / Bugtraq Full-Disclosure linux-kernel linux-netdev linux-ext4 linux-hardening linux-cve-announce PHC
Open Source and information security mailing list archives

Hash Suite: Windows password security audit tool. GUI, reports in PDF.

[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]

Message-ID: <Yz4AZZWE1/wWhXFP@google.com>
Date:   Wed, 5 Oct 2022 22:08:37 +0000
From:   Sean Christopherson <seanjc@...gle.com>
To:     Like Xu <like.xu.linux@...il.com>
Cc:     Paolo Bonzini <pbonzini@...hat.com>,
        Sandipan Das <sandipan.das@....com>, kvm@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [kvm-unit-tests PATCH 2/2] x86/pmu: Add AMD Guest PerfMonV2
 testcases

Can you provide a single series for all of the KVM-Unit-Tests PMU work (separate
from the KVM patches)?  Ya, it'll be big and is a blatant violation of "do one
thing", but trying to manually handle the dependencies on the review side is time
consuming.

One thought to help keep track of dependencies between KVM and KUT would be to
add dummy commits between each sub-series, with the dummy commit containing a lore
link to the relevant KVM patches/series.  That would allow throwing everything into
one series without losing track of things.  Hopefully.

On Mon, Sep 05, 2022, Like Xu wrote:
> diff --git a/lib/x86/processor.h b/lib/x86/processor.h
> index 9c490d9..b9592c4 100644
> --- a/lib/x86/processor.h
> +++ b/lib/x86/processor.h
> @@ -796,8 +796,12 @@ static inline void flush_tlb(void)
>  
>  static inline u8 pmu_version(void)
>  {
> -	if (!is_intel())
> +	if (!is_intel()) {
> +		/* Performance Monitoring Version 2 Supported */
> +		if (cpuid(0x80000022).a & 0x1)

Add an X86_FEATURE_*, that way this is self-documenting.

> +			return 2;
>  		return 0;
> +	}
>  
>  	return cpuid(10).a & 0xff;
>  }
> @@ -824,6 +828,9 @@ static inline u8 pmu_nr_gp_counters(void)
>  {
>  	if (is_intel()) {
>  		return (cpuid(10).a >> 8) & 0xff;
> +	} else if (this_cpu_has_perf_global_ctrl()) {

Eww.  Took me too long to connect the dots to understand how this guarantees that
leaf 0x80000022 is available.  With an X86_FEATURE_* this can simply be:

	} else if (this_cpu_has(X86_FEATURE_AMD_PMU_V2) {

or whatever name is appropriate.

> +		/* Number of Core Performance Counters. */
> +		return cpuid(0x80000022).b & 0xf;
>  	} else if (!has_amd_perfctr_core()) {
>  		return AMD64_NUM_COUNTERS;
>  	}
> diff --git a/x86/pmu.c b/x86/pmu.c
> index 11607c0..6d5363b 100644
> --- a/x86/pmu.c
> +++ b/x86/pmu.c
> @@ -72,6 +72,9 @@ struct pmu_event {
>  #define PMU_CAP_FW_WRITES	(1ULL << 13)
>  static u32 gp_counter_base;
>  static u32 gp_select_base;
> +static u32 global_status_msr;
> +static u32 global_ctl_msr;
> +static u32 global_status_clr_msr;

What do you think about naming these like MSR #defines?  E.g.

  MSR_PERF_GLOBAL_CTRL
  MSR_PERF_GLOBAL_STATUS
  MSR_PERF_GLOBAL_STATUS_CLR

There's a little risk of causing confusing, but I think it would make the code
easier to read.

>  static unsigned int gp_events_size;
>  static unsigned int nr_gp_counters;
>  
> @@ -150,8 +153,7 @@ static void global_enable(pmu_counter_t *cnt)
>  		return;
>  
>  	cnt->idx = event_to_global_idx(cnt);
> -	wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, rdmsr(MSR_CORE_PERF_GLOBAL_CTRL) |
> -			(1ull << cnt->idx));
> +	wrmsr(global_ctl_msr, rdmsr(global_ctl_msr) | (1ull << cnt->idx));

Opportunistically use BIT_ULL().

>  }
>  
>  static void global_disable(pmu_counter_t *cnt)
> @@ -159,8 +161,7 @@ static void global_disable(pmu_counter_t *cnt)
>  	if (pmu_version() < 2)
>  		return;
>  
> -	wrmsr(MSR_CORE_PERF_GLOBAL_CTRL, rdmsr(MSR_CORE_PERF_GLOBAL_CTRL) &
> -			~(1ull << cnt->idx));
> +	wrmsr(global_ctl_msr, rdmsr(global_ctl_msr) & ~(1ull << cnt->idx));

BIT_ULL()

>  }
>  
>  static inline uint32_t get_gp_counter_msr(unsigned int i)
> @@ -326,6 +327,23 @@ static void check_counters_many(void)
>  	report(i == n, "all counters");
>  }
>  
> +static bool is_the_count_reproducible(pmu_counter_t *cnt)
> +{
> +	unsigned int i;
> +	uint64_t count;
> +
> +	__measure(cnt, 0);
> +	count = cnt->count;
> +
> +	for (i = 0; i < 10; i++) {
> +		__measure(cnt, 0);
> +		if (count != cnt->count)
> +			return false;
> +	}
> +
> +	return true;
> +}
> +
>  static void check_counter_overflow(void)
>  {
>  	uint64_t count;
> @@ -334,13 +352,14 @@ static void check_counter_overflow(void)
>  		.ctr = gp_counter_base,
>  		.config = EVNTSEL_OS | EVNTSEL_USR | (*gp_events)[1].unit_sel /* instructions */,
>  	};
> +	bool precise_event = is_the_count_reproducible(&cnt);
> +
>  	__measure(&cnt, 0);
>  	count = cnt.count;
>  
>  	/* clear status before test */
>  	if (pmu_version() > 1) {

Please provide helper(s) to replace the myriad open coded "pmu_version() > ???"
checks.  E.g. this one appears to be:

	if (this_cpu_has_perf_global_status_clr())

I obviously don't care about the verbosity, it's that people like me might not
know what the PMU version has to do with an MSR being accessible.