Re: [PATCH 2/3] perf/x86/intel: fix for MSR_LAST_BRANCH_FROM_x quirk when no TSX

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jun 01, 2016 at 07:42:02PM -0700, David Carrillo-Cisneros wrote:

> diff --git a/arch/x86/events/intel/lbr.c b/arch/x86/events/intel/lbr.c
> index 2dca66c..6aa2d8a 100644
> --- a/arch/x86/events/intel/lbr.c
> +++ b/arch/x86/events/intel/lbr.c
> @@ -80,6 +80,7 @@ static enum {
>  #define LBR_FROM_FLAG_MISPRED  (1ULL << 63)
>  #define LBR_FROM_FLAG_IN_TX    (1ULL << 62)
>  #define LBR_FROM_FLAG_ABORT    (1ULL << 61)
> +#define LBR_FROM_SIGNEXT_MSB   (1ULL << 60)
>  
>  /*
>   * x86control flow change classification
> @@ -235,6 +236,62 @@ enum {
>  	LBR_VALID,
>  };
>  
> +/*
> + * For formats with LBR_TSX flags (e.g. LBR_FORMAT_EIP_FLAGS2), bits 61:62 in
> + * MSR_LAST_BRANCH_FROM_x are the TSX flags when TSX is supported.
> + * When TSX support is disabled the behavior differs as follows:
> + *   - For wrmsr, bits 61:62 are considered part of the sign-extension.
> + *   - HW updates to the MSR (no through wrmsr) will clear bits 61:62,
> + *   regardless of the sign of bit at position 47, i.e. bit 61:62 are not part
> + *   of the sign-extension.
> + *
> + * Therefore, if the conditions:
> + *   1) LBR has TSX format.
> + *   2) CPU has no TSX support enabled.
> + *   3) data in MSR (bits 0:48) is negative.

I take issue with 3; sign extension has nothing to do with being
negative or not.

> + * are all true, then any value passed to wrmsr must be signed-extended to
> + * 63 bits and any value from rdmsr must be converted to 61 bits, ignoring
> + * the TSX flags.
> + */
> +

> +static inline bool lbr_from_signext_quirk_on(void)

This function name is bad; it doesn't return if the quirk is on at all.
It detects if the quirk is relevant and should be turned on, look at the
bloody usage site.

> +{
> +	int lbr_format = x86_pmu.intel_cap.lbr_format;
> +	bool tsx_support = boot_cpu_has(X86_FEATURE_HLE) ||
> +			   boot_cpu_has(X86_FEATURE_RTM);
> +
> +	return !tsx_support && (lbr_desc[lbr_format] & LBR_TSX);
> +}
> +
> +DEFINE_STATIC_KEY_FALSE(lbr_from_quirk_key);
> +
> +static inline bool lbr_from_signext_quirk_test(u64 val)
> +{
> +	return static_branch_unlikely(&lbr_from_quirk_key) &&
> +	       (val & LBR_FROM_SIGNEXT_MSB);
> +}

No, that's just wrong.

> +
> +/*
> + * If quirk is needed, do sign extension to 63 bits.
> + */
> +inline u64 lbr_from_signext_quirk_wr(u64 val)
> +{
> +	if (lbr_from_signext_quirk_test(val))
> +		val |= (LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);

That too, that's horrible. You don't set IN_TX and ABORT.

What's wrong with:

	if (static_branch_unlikely(&lbr_from_quirk)) {
		val <<= 2;
		(s64)val >>= 2;
	}

That actually sign extends and without branches.

> +	return val;
> +}
> +
> +/*
> + * If quirk is needed, ensure sign extension is 61 bits.

You cannot extend shorter. Extend means to make longer/more.

You can truncate sign bits, or just say you need to clear the 2 MSBs.

> + */
> +
> +u64 lbr_from_signext_quirk_rd(u64 val)
> +{
> +	if (lbr_from_signext_quirk_test(val))

Again, no point in wasting an extra branch, just clear the bits
unconditional:

	if (static_branch_unlikely(&lbr_from_quirk))
> +		val &= ~(LBR_FROM_FLAG_IN_TX | LBR_FROM_FLAG_ABORT);

> +	return val;
> +}
> +
>  static void __intel_pmu_lbr_restore(struct x86_perf_task_context *task_ctx)
>  {
>  	int i;