Re: [PATCH 2/2] perf tools: Fix record failure when mixed with ARM SPE event

[Mathieu Poirier]

Hi Li,

On Tue, Jun 23, 2020 at 08:31:41PM +0800, Wei Li wrote:
> When recording with cache-misses and arm_spe_x event, i found that
> it will just fail without showing any error info if i put cache-misses
> after arm_spe_x event.
> 
> [root@localhost 0620]# perf record -e cache-misses -e \
> arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,load_filter=1,\
> jitter=1,store_filter=1,min_latency=0/ sleep 1
> [ perf record: Woken up 1 times to write data ]
> [ perf record: Captured and wrote 0.067 MB perf.data ]
> [root@localhost 0620]# perf record -e \
> arm_spe_0/ts_enable=1,pct_enable=1,pa_enable=1,load_filter=1,jitter=1,\
> store_filter=1,min_latency=0/ -e cache-misses sleep 1
> [root@localhost 0620]#
> 
> Finally, i found the reason is that the parameter 'arm_spe_pmu' passed to
> arm_spe_recording_init() in auxtrace_record__init() is wrong. When the
> arm_spe_x event is not the last event, 'arm_spe_pmus[i]' will be out of
> bounds.

Yes, this indeed broken.  

The current code can only work if the only event to be
traced is an arm_spe_X, or if it is the last event to be specified.
Otherwise the last event type will be checked against all the
arm_spe_pmus[i]->types, none will match and an out of bound i index will be
used in arm_spc_recording_init().

Since this problem is not easy to figure out please include the above
explanation in the changelog.

> 
> It seems that the code can't support concurrent multiple different
> arm_spe_x events currently. So add the code to check and record the
> found 'arm_spe_pmu' to fix this issue.
> 
> In fact, we don't support concurrent multiple same arm_spe_x events either,
> that is checked in arm_spe_recording_options(), and it will show the
> relevant info.
> 
> Fixes: ffd3d18c20b8d ("perf tools: Add ARM Statistical Profiling Extensions (SPE) support")
> Signed-off-by: Wei Li <liwei391@xxxxxxxxxx>
> ---
>  tools/perf/arch/arm/util/auxtrace.c | 10 +++++++++-
>  1 file changed, 9 insertions(+), 1 deletion(-)
> 
> diff --git a/tools/perf/arch/arm/util/auxtrace.c b/tools/perf/arch/arm/util/auxtrace.c
> index 62b7b03d691a..7bb6f29e766c 100644
> --- a/tools/perf/arch/arm/util/auxtrace.c
> +++ b/tools/perf/arch/arm/util/auxtrace.c
> @@ -58,6 +58,7 @@ struct auxtrace_record
>  	bool found_etm = false;
>  	bool found_spe = false;
>  	static struct perf_pmu **arm_spe_pmus;
> +	static struct perf_pmu *arm_spe_pmu;

As far as I can tell the "static" doesn't do anything.

>  	static int nr_spes = 0;
>  	int i = 0;
>  
> @@ -77,6 +78,13 @@ struct auxtrace_record
>  
>  		for (i = 0; i < nr_spes; i++) {
>  			if (evsel->core.attr.type == arm_spe_pmus[i]->type) {
> +				if (found_spe && (arm_spe_pmu != arm_spe_pmus[i])) {
> +					pr_err("Concurrent multiple SPE operation not currently supported\n");
> +					*err = -EOPNOTSUPP;
> +					return NULL;
> +				}

Instead of the above, which as you rightly pointed out, is also done in
arm_spe_recording_options() it might be best to just fix the "if (!nr_spes)"
condition:
                if (!nr_spes || arm_spe_pmu)
                        continue

Furthermore, instead of having a new arm_spe_pmu variable you could simply make
found_spe a struct perf_pmu.  That would be one less variable to take care of.

> +
> +				arm_spe_pmu = arm_spe_pmus[i];
>  				found_spe = true;

Last but not least do you know where the memory allocated for array arm_spe_pmus
is released?  If you can't find it either then we have a memory leak and it
would be nice to have that fixed.

Regards
Mathieu

PS: Leo Yan has spent a fair amount of time in the SPE code - please CC him on
your next revision.


>  				break;
>  			}
> @@ -94,7 +102,7 @@ struct auxtrace_record
>  
>  #if defined(__aarch64__)
>  	if (found_spe)
> -		return arm_spe_recording_init(err, arm_spe_pmus[i]);
> +		return arm_spe_recording_init(err, arm_spe_pmu);
>  #endif
>  
>  	/*
> -- 
> 2.17.1
>