Re: [PATCH v4] perf stat: Fix wrong skipping for per-die aggregation

[Jin, Yao]

Hi Jiri,

On 1/12/2021 6:04 PM, Jiri Olsa wrote:
> On Tue, Jan 05, 2021 at 10:36:15AM +0800, Jin Yao wrote:
>
> SNIP
>
> > diff --git a/tools/perf/util/stat.c b/tools/perf/util/stat.c
> > index 8ce1479c98f0..a658e0ffaf2a 100644
> > --- a/tools/perf/util/stat.c
> > +++ b/tools/perf/util/stat.c
> > @@ -13,6 +13,7 @@
> >   #include "evlist.h"
> >   #include "evsel.h"
> >   #include "thread_map.h"
> > +#include "hashmap.h"
> >   #include <linux/zalloc.h>
> >   
> >   void update_stats(struct stats *stats, u64 val)
> > @@ -276,15 +277,29 @@ void evlist__save_aggr_prev_raw_counts(struct evlist *evlist)
> >   static void zero_per_pkg(struct evsel *counter)
> >   {
> >   	if (counter->per_pkg_mask)
> > -		memset(counter->per_pkg_mask, 0, cpu__max_cpu());
> > +		hashmap__clear(counter->per_pkg_mask);
> > +}
> > +
> > +static size_t pkg_id_hash(const void *key, void *ctx __maybe_unused)
> > +{
> > +	int socket = (int64_t)key >> 32;
> > +
> > +	return socket;
>
> could be just simple return right, why the variable?

Sure, just simple return should be OK.

> > +}
> > +
> > +static bool pkg_id_equal(const void *key1, const void *key2,
> > +			 void *ctx __maybe_unused)
> > +{
> > +	return (int64_t)key1 == (int64_t)key2;
> >   }
> >   
> >   static int check_per_pkg(struct evsel *counter,
> >   			 struct perf_counts_values *vals, int cpu, bool *skip)
> >   {
> > -	unsigned long *mask = counter->per_pkg_mask;
> > +	struct hashmap *mask = counter->per_pkg_mask;
> >   	struct perf_cpu_map *cpus = evsel__cpus(counter);
> > -	int s;
> > +	int s, d, ret;
> > +	uint64_t key;
> >   
> >   	*skip = false;
> >   
> > @@ -295,7 +310,7 @@ static int check_per_pkg(struct evsel *counter,
> >   		return 0;
> >   
> >   	if (!mask) {
> > -		mask = zalloc(cpu__max_cpu());
> > +		mask = hashmap__new(pkg_id_hash, pkg_id_equal, NULL);
> >   		if (!mask)
> >   			return -ENOMEM;
> >   
> > @@ -317,7 +332,23 @@ static int check_per_pkg(struct evsel *counter,
> >   	if (s < 0)
> >   		return -1;
> >   
> > -	*skip = test_and_set_bit(s, mask) == 1;
> > +	/*
> > +	 * On multi-die system, 0 < die_id < 256. On no-die system, die_id = 0.
> > +	 * We use hashmap(socket, die) to check the used socket+die pair.
> > +	 */
> > +	d = cpu_map__get_die(cpus, cpu, NULL).die;
> > +	if (d < 0)
> > +		return -1;
> > +
> > +	key = (uint64_t)s << 32 | (d & 0xff);
>
> why the masking for d? also I thought it could be bigger than 256

I remember in previous perf tool it would show a warning "The die id number is too big" if die_id > 
256 and cpu_map__get_die() returned -1. See commit db5742b6849e ("perf stat: Support per-die 
aggregation").

But now the die_id checking is removed.

> would it better to do it the other way around?
>
>     die << 32 | socket

I think this way is OK, thanks! I will use it in v5.

> > +	if (hashmap__find(mask, (void *)key, NULL)) {
> > +		*skip = true;
> > +	} else {
> > +		ret = hashmap__add(mask, (void *)key, (void *)1);
> > +		if (ret)
> > +			return -1;
>
> you could 'return ret' at the end and skip this extra check

Yes, define int ret = 0 and just 'return ret' at the end of function.

Thanks
Jin Yao

> thanks,
> jirka
>
> > +	}
> > +
> >   	return 0;
> >   }
> >   
> > --
> > 2.17.1