Re: [Patch v3 16/22] perf/core: Support to capture higher width vector registers

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Apr 15, 2025 at 11:44:22AM +0000, Dapeng Mi wrote:
>  extern unsigned long perf_arch_instruction_pointer(struct pt_regs *regs);
> diff --git a/arch/x86/include/uapi/asm/perf_regs.h b/arch/x86/include/uapi/asm/perf_regs.h
> index f9c5b16b1882..5e2d9796b2cc 100644
> --- a/arch/x86/include/uapi/asm/perf_regs.h
> +++ b/arch/x86/include/uapi/asm/perf_regs.h
> @@ -33,7 +33,7 @@ enum perf_event_x86_regs {
>  	PERF_REG_X86_32_MAX = PERF_REG_X86_GS + 1,
>  	PERF_REG_X86_64_MAX = PERF_REG_X86_SSP + 1,
>  
> -	/* These all need two bits set because they are 128bit */
> +	/* These all need two bits set because they are 128 bits */
>  	PERF_REG_X86_XMM0  = 32,
>  	PERF_REG_X86_XMM1  = 34,
>  	PERF_REG_X86_XMM2  = 36,
> @@ -53,6 +53,83 @@ enum perf_event_x86_regs {
>  
>  	/* These include both GPRs and XMMX registers */
>  	PERF_REG_X86_XMM_MAX = PERF_REG_X86_XMM15 + 2,
> +
> +	/* Leave bits[127:64] for other GP registers, like R16 ~ R31.*/
> +
> +	/*
> +	 * Each YMM register need 4 bits to represent because they are 256 bits.
> +	 * PERF_REG_X86_YMMH0 = 128
> +	 */
> +	PERF_REG_X86_YMM0	= 128,
> +	PERF_REG_X86_YMM1	= PERF_REG_X86_YMM0 + 4,
> +	PERF_REG_X86_YMM2	= PERF_REG_X86_YMM1 + 4,
> +	PERF_REG_X86_YMM3	= PERF_REG_X86_YMM2 + 4,
> +	PERF_REG_X86_YMM4	= PERF_REG_X86_YMM3 + 4,
> +	PERF_REG_X86_YMM5	= PERF_REG_X86_YMM4 + 4,
> +	PERF_REG_X86_YMM6	= PERF_REG_X86_YMM5 + 4,
> +	PERF_REG_X86_YMM7	= PERF_REG_X86_YMM6 + 4,
> +	PERF_REG_X86_YMM8	= PERF_REG_X86_YMM7 + 4,
> +	PERF_REG_X86_YMM9	= PERF_REG_X86_YMM8 + 4,
> +	PERF_REG_X86_YMM10	= PERF_REG_X86_YMM9 + 4,
> +	PERF_REG_X86_YMM11	= PERF_REG_X86_YMM10 + 4,
> +	PERF_REG_X86_YMM12	= PERF_REG_X86_YMM11 + 4,
> +	PERF_REG_X86_YMM13	= PERF_REG_X86_YMM12 + 4,
> +	PERF_REG_X86_YMM14	= PERF_REG_X86_YMM13 + 4,
> +	PERF_REG_X86_YMM15	= PERF_REG_X86_YMM14 + 4,
> +	PERF_REG_X86_YMM_MAX	= PERF_REG_X86_YMM15 + 4,
> +
> +	/*
> +	 * Each ZMM register needs 8 bits to represent because they are 512 bits
> +	 * PERF_REG_X86_ZMMH0 = 192
> +	 */
> +	PERF_REG_X86_ZMM0	= PERF_REG_X86_YMM_MAX,
> +	PERF_REG_X86_ZMM1	= PERF_REG_X86_ZMM0 + 8,
> +	PERF_REG_X86_ZMM2	= PERF_REG_X86_ZMM1 + 8,
> +	PERF_REG_X86_ZMM3	= PERF_REG_X86_ZMM2 + 8,
> +	PERF_REG_X86_ZMM4	= PERF_REG_X86_ZMM3 + 8,
> +	PERF_REG_X86_ZMM5	= PERF_REG_X86_ZMM4 + 8,
> +	PERF_REG_X86_ZMM6	= PERF_REG_X86_ZMM5 + 8,
> +	PERF_REG_X86_ZMM7	= PERF_REG_X86_ZMM6 + 8,
> +	PERF_REG_X86_ZMM8	= PERF_REG_X86_ZMM7 + 8,
> +	PERF_REG_X86_ZMM9	= PERF_REG_X86_ZMM8 + 8,
> +	PERF_REG_X86_ZMM10	= PERF_REG_X86_ZMM9 + 8,
> +	PERF_REG_X86_ZMM11	= PERF_REG_X86_ZMM10 + 8,
> +	PERF_REG_X86_ZMM12	= PERF_REG_X86_ZMM11 + 8,
> +	PERF_REG_X86_ZMM13	= PERF_REG_X86_ZMM12 + 8,
> +	PERF_REG_X86_ZMM14	= PERF_REG_X86_ZMM13 + 8,
> +	PERF_REG_X86_ZMM15	= PERF_REG_X86_ZMM14 + 8,
> +	PERF_REG_X86_ZMM16	= PERF_REG_X86_ZMM15 + 8,
> +	PERF_REG_X86_ZMM17	= PERF_REG_X86_ZMM16 + 8,
> +	PERF_REG_X86_ZMM18	= PERF_REG_X86_ZMM17 + 8,
> +	PERF_REG_X86_ZMM19	= PERF_REG_X86_ZMM18 + 8,
> +	PERF_REG_X86_ZMM20	= PERF_REG_X86_ZMM19 + 8,
> +	PERF_REG_X86_ZMM21	= PERF_REG_X86_ZMM20 + 8,
> +	PERF_REG_X86_ZMM22	= PERF_REG_X86_ZMM21 + 8,
> +	PERF_REG_X86_ZMM23	= PERF_REG_X86_ZMM22 + 8,
> +	PERF_REG_X86_ZMM24	= PERF_REG_X86_ZMM23 + 8,
> +	PERF_REG_X86_ZMM25	= PERF_REG_X86_ZMM24 + 8,
> +	PERF_REG_X86_ZMM26	= PERF_REG_X86_ZMM25 + 8,
> +	PERF_REG_X86_ZMM27	= PERF_REG_X86_ZMM26 + 8,
> +	PERF_REG_X86_ZMM28	= PERF_REG_X86_ZMM27 + 8,
> +	PERF_REG_X86_ZMM29	= PERF_REG_X86_ZMM28 + 8,
> +	PERF_REG_X86_ZMM30	= PERF_REG_X86_ZMM29 + 8,
> +	PERF_REG_X86_ZMM31	= PERF_REG_X86_ZMM30 + 8,
> +	PERF_REG_X86_ZMM_MAX	= PERF_REG_X86_ZMM31 + 8,
> +
> +	/*
> +	 * OPMASK Registers
> +	 * PERF_REG_X86_OPMASK0 = 448
> +	 */
> +	PERF_REG_X86_OPMASK0	= PERF_REG_X86_ZMM_MAX,
> +	PERF_REG_X86_OPMASK1	= PERF_REG_X86_OPMASK0 + 1,
> +	PERF_REG_X86_OPMASK2	= PERF_REG_X86_OPMASK1 + 1,
> +	PERF_REG_X86_OPMASK3	= PERF_REG_X86_OPMASK2 + 1,
> +	PERF_REG_X86_OPMASK4	= PERF_REG_X86_OPMASK3 + 1,
> +	PERF_REG_X86_OPMASK5	= PERF_REG_X86_OPMASK4 + 1,
> +	PERF_REG_X86_OPMASK6	= PERF_REG_X86_OPMASK5 + 1,
> +	PERF_REG_X86_OPMASK7	= PERF_REG_X86_OPMASK6 + 1,
> +
> +	PERF_REG_X86_VEC_MAX	= PERF_REG_X86_OPMASK7 + 1,
>  };

> diff --git a/include/uapi/linux/perf_event.h b/include/uapi/linux/perf_event.h
> index 5fc753c23734..78aae0464a54 100644
> --- a/include/uapi/linux/perf_event.h
> +++ b/include/uapi/linux/perf_event.h
> @@ -379,6 +379,10 @@ enum perf_event_read_format {
>  #define PERF_ATTR_SIZE_VER6	120	/* add: aux_sample_size */
>  #define PERF_ATTR_SIZE_VER7	128	/* add: sig_data */
>  #define PERF_ATTR_SIZE_VER8	136	/* add: config3 */
> +#define PERF_ATTR_SIZE_VER9	168	/* add: sample_regs_intr_ext[PERF_EXT_REGS_ARRAY_SIZE] */
> +
> +#define PERF_EXT_REGS_ARRAY_SIZE	7
> +#define PERF_NUM_EXT_REGS		(PERF_EXT_REGS_ARRAY_SIZE * 64)
>  
>  /*
>   * Hardware event_id to monitor via a performance monitoring event:
> @@ -533,6 +537,13 @@ struct perf_event_attr {
>  	__u64	sig_data;
>  
>  	__u64	config3; /* extension of config2 */
> +
> +	/*
> +	 * Extension sets of regs to dump for each sample.
> +	 * See asm/perf_regs.h for details.
> +	 */
> +	__u64	sample_regs_intr_ext[PERF_EXT_REGS_ARRAY_SIZE];
> +	__u64   sample_regs_user_ext[PERF_EXT_REGS_ARRAY_SIZE];
>  };
>  
>  /*

I still utterly hate this interface. This is a giant waste of bits.

What makes it even worse is that XMMn is the lower half of YMMn which in
turn is the lower half of ZMMn.

So by exposing only ZMMn you already expose all of them. The interface
explicitly allows asking for sub-words.

But most importantly of all, last time I asked if there are users that
actually care about the whole per-register thing and I don't see an
answer here.

Can we please find a better interface? Ideally one that scales up to
1024 and 2048 bit vector width, because I'd hate to have to rev this
again.

Perhaps add sample_vec_regs_*[] with a saner format, and if that is !0
then the XMM regs dissapear from sample_regs_*[] and we get to use that
space to extended GPs.