Re: [PATCH v7 1/3] x86/fpu: Measure the Latency of XSAVE and XRSTOR

[Ingo Molnar]

* Yi Sun <yi.sun@xxxxxxxxx> wrote:

> @@ -113,7 +116,7 @@ static inline u64 xfeatures_mask_independent(void)
>   * original instruction which gets replaced. We need to use it here as the
>   * address of the instruction where we might get an exception at.
>   */
> -#define XSTATE_XSAVE(st, lmask, hmask, err)				\
> +#define __XSTATE_XSAVE(st, lmask, hmask, err)				\
>  	asm volatile(ALTERNATIVE_3(XSAVE,				\
>  				   XSAVEOPT, X86_FEATURE_XSAVEOPT,	\
>  				   XSAVEC,   X86_FEATURE_XSAVEC,	\
> @@ -130,7 +133,7 @@ static inline u64 xfeatures_mask_independent(void)
>   * Use XRSTORS to restore context if it is enabled. XRSTORS supports compact
>   * XSAVE area format.
>   */
> -#define XSTATE_XRESTORE(st, lmask, hmask)				\
> +#define __XSTATE_XRESTORE(st, lmask, hmask)				\
>  	asm volatile(ALTERNATIVE(XRSTOR,				\
>  				 XRSTORS, X86_FEATURE_XSAVES)		\
>  		     "\n"						\
> @@ -140,6 +143,35 @@ static inline u64 xfeatures_mask_independent(void)
>  		     : "D" (st), "m" (*st), "a" (lmask), "d" (hmask)	\
>  		     : "memory")
>  
> +#if defined(CONFIG_X86_DEBUG_FPU)
> +#define XSTATE_XSAVE(fps, lmask, hmask, err)				\
> +	do {								\
> +		struct fpstate *f = fps;				\
> +		u64 tc = -1;						\
> +		if (tracepoint_enabled(x86_fpu_latency_xsave))		\
> +			tc = trace_clock();				\
> +		__XSTATE_XSAVE(&f->regs.xsave, lmask, hmask, err);	\
> +		if (tracepoint_enabled(x86_fpu_latency_xsave))		\
> +			trace_x86_fpu_latency_xsave(f, trace_clock() - tc);\
> +	} while (0)
> +
> +#define XSTATE_XRESTORE(fps, lmask, hmask)				\
> +	do {								\
> +		struct fpstate *f = fps;				\
> +		u64 tc = -1;						\
> +		if (tracepoint_enabled(x86_fpu_latency_xrstor))		\
> +			tc = trace_clock();				\
> +		__XSTATE_XRESTORE(&f->regs.xsave, lmask, hmask);	\
> +		if (tracepoint_enabled(x86_fpu_latency_xrstor))		\
> +			trace_x86_fpu_latency_xrstor(f, trace_clock() - tc);\

This v7 version does not adequately address the review feedback I gave for 
v6: it adds tracing overhead to potential hot paths, and putting it behind 
CONFIG_X86_DEBUG_FPU is not a solution either: it's default-y, so de-facto 
enabled on all major distros...

It seems unnecessarily complex: why does it have to measure latency 
directly? Tracepoints *by default* come with event timestamps. A latency 
measurement tool should be able to subtract two timestamps to extract the 
latency between two tracepoints...

In fact, function tracing is enabled on all major Linux distros:

  kepler:~/tip> grep FUNCTION_TRACER /boot/config-6.2.0-33-generic 
  CONFIG_HAVE_FUNCTION_TRACER=y
  CONFIG_FUNCTION_TRACER=y

Why not just enable function tracing for the affected FPU context switching 
functions?

The relevant functions are already standalone in a typical kernel:

   xsave:                # ffffffff8103cfe0 T save_fpregs_to_fpstate
   xrstor:               # ffffffff8103d160 T restore_fpregs_from_fpstate
   xrstor_supervisor:    # ffffffff8103dc50 T fpu__clear_user_states

... and FPU context switching overhead dominates the cost of these 
functions.

Thanks,

	Ingo