Re: [PATCH v1 9/9] perf arm-spe: Dump metadata with version 2

[James Clark <james.clark@linaro.org>]

On 27/08/2024 5:44 pm, Leo Yan wrote:
> This commit dumps metadata with version 2. It uses two string arrays
> metadata_hdr_fmts and metadata_per_cpu_fmts as string formats for the
> header and per CPU data respectively, and the arm_spe_print_info()
> function is enhanced to support dumping metadata with the version 2
> format.
> 
> After:
> 
>    0 0 0x4a8 [0x170]: PERF_RECORD_AUXTRACE_INFO type: 4
>    PMU Type           :13
>    Version            :2
>    Num of CPUs        :8
>      CPU #            :0
>      MIDR             :0x410fd801
>      Bound PMU Type   :-1
>      Min Interval     :0
>      Load Data Source :0
>      CPU #            :1
>      MIDR             :0x410fd801
>      Bound PMU Type   :-1
>      Min Interval     :0
>      Load Data Source :0
>      CPU #            :2
>      MIDR             :0x410fd870
>      Bound PMU Type   :13
>      Min Interval     :1024
>      Load Data Source :1
>      CPU #            :3
>      MIDR             :0x410fd870
>      Bound PMU Type   :13
>      Min Interval     :1024
>      Load Data Source :1
>      CPU #            :4
>      MIDR             :0x410fd870
>      Bound PMU Type   :13
>      Min Interval     :1024
>      Load Data Source :1
>      CPU #            :5
>      MIDR             :0x410fd870
>      Bound PMU Type   :13
>      Min Interval     :1024
>      Load Data Source :1
>      CPU #            :6
>      MIDR             :0x410fd850
>      Bound PMU Type   :14
>      Min Interval     :1024
>      Load Data Source :1
>      CPU #            :7
>      MIDR             :0x410fd850
>      Bound PMU Type   :14
>      Min Interval     :1024
>      Load Data Source :1
> 
> Signed-off-by: Leo Yan <leo.yan@arm.com>
> ---
>   tools/perf/util/arm-spe.c | 43 ++++++++++++++++++++++++++++++++++-----
>   1 file changed, 38 insertions(+), 5 deletions(-)
> 
> diff --git a/tools/perf/util/arm-spe.c b/tools/perf/util/arm-spe.c
> index 87cf06db765b..be34d4c4306a 100644
> --- a/tools/perf/util/arm-spe.c
> +++ b/tools/perf/util/arm-spe.c
> @@ -1067,16 +1067,49 @@ static bool arm_spe_evsel_is_auxtrace(struct perf_session *session __maybe_unuse
>   	return strstarts(evsel->name, ARM_SPE_PMU_NAME);
>   }
>   
> -static const char * const arm_spe_info_fmts[] = {
> -	[ARM_SPE_PMU_TYPE]		= "  PMU Type           %"PRId64"\n",
> +static const char * const metadata_hdr_fmts[] = {
> +	[ARM_SPE_PMU_TYPE]		= "  PMU Type           :%"PRId64"\n",
> +	[ARM_SPE_HEADER_VERSION]	= "  Version            :%"PRId64"\n",
> +	[ARM_SPE_CPU_NUM]		= "  Num of CPUs        :%"PRId64"\n",
>   };
>   
> -static void arm_spe_print_info(__u64 *arr)
> +static const char * const metadata_per_cpu_fmts[] = {
> +	[ARM_SPE_CPU]			= "    CPU #            :%"PRId64"\n",
> +	[ARM_SPE_CPU_MIDR]		= "    MIDR             :0x%"PRIx64"\n",
> +	[ARM_SPE_CPU_PMU_TYPE]		= "    Bound PMU Type   :%"PRId64"\n",
> +	[ARM_SPE_CAP_MIN_IVAL]		= "    Min Interval     :%"PRId64"\n",
> +	[ARM_SPE_CAP_LDS]		= "    Load Data Source :%"PRId64"\n",
> +};
> +
> +static void arm_spe_print_info(struct arm_spe *spe, __u64 *arr)
>   {
> +	unsigned int i, cpu, header_size, cpu_num, per_cpu_size;
> +
>   	if (!dump_trace)
>   		return;
>   
> -	fprintf(stdout, arm_spe_info_fmts[ARM_SPE_PMU_TYPE], arr[ARM_SPE_PMU_TYPE]);
> +	if (spe->metadata_ver == 1) {
> +		cpu_num = 0;
> +		header_size = ARM_SPE_AUXTRACE_V1_PRIV_MAX;
> +		per_cpu_size = 0;
> +	} else if (spe->metadata_ver == 2) {

Assuming future version updates are backwards compatible and only add 
new info this should be spe->metadata_ver >= 2, otherwise version bumps 
end up causing errors when files get passed around.

I know there are arguments about what should and shouldn't be supported 
when opening new files on old perfs, but in this case it's easy to only 
add new info to the aux header and leave the old stuff intact.

> +		cpu_num = arr[ARM_SPE_CPU_NUM];
> +		header_size = ARM_SPE_AUXTRACE_V2_PRIV_MAX;
> +		per_cpu_size = ARM_SPE_AUXTRACE_V2_PRIV_PER_CPU_MAX;

I think for coresight we also save the size of each per-cpu block rather 
than use a constant, that way new items can be appended without breaking 
readers.

That kind of leads to another point that this mechanism is mostly 
duplicated from coresight. It saves a main header version, then per-cpu 
groups of variable size with named elements. I'm not saying we should 
definitely try to share the code, but it's worth keeping in mind.

> +	} else {
> +		pr_err("Cannot support metadata ver: %ld\n", spe->metadata_ver);
> +		return;
> +	}
> +
> +	for (i = 0; i < header_size; i++)
> +		fprintf(stdout, metadata_hdr_fmts[i], arr[i]);
> +
> +	arr += header_size;
> +	for (cpu = 0; cpu < cpu_num; cpu++) {
> +		for (i = 0; i < per_cpu_size; i++)
> +			fprintf(stdout, metadata_per_cpu_fmts[i], arr[i]);
> +		arr += per_cpu_size;
> +	}
>   }
>   
>   static void arm_spe_set_event_name(struct evlist *evlist, u64 id,
> @@ -1383,7 +1416,7 @@ int arm_spe_process_auxtrace_info(union perf_event *event,
>   	spe->auxtrace.evsel_is_auxtrace = arm_spe_evsel_is_auxtrace;
>   	session->auxtrace = &spe->auxtrace;
>   
> -	arm_spe_print_info(&auxtrace_info->priv[0]);
> +	arm_spe_print_info(spe, &auxtrace_info->priv[0]);
>   
>   	if (dump_trace)
>   		return 0;