'perf stat' shows a mismatch in perf stat regarding counter names on
s390:
Run command:
[root@s35lp76 perf]# ./perf stat -e tx_nc_tend -v --
~/mytesttx 1 >/tmp/111
tx_nc_tend: 1 573146 573146
tx_nc_tend: 1 573146 573146
Performance counter stats for '/root/mytesttx 1':
3 tx_nc_tend
0.001037252 seconds time elapsed
[root@s35lp76 perf]#
shows transaction counter tx_nc_tend with value 3 but it was triggered
only once as seen by the output of mytesttx.
When looking up the event name tx_nc_tend the following function
sequence is called:
parse_events_multi_pmu_add()
+--> perf_pmu__scan() being called with NULL argument
+--> pmu_read_sysfs() scans directory ../devices/ for
all PMUs
+--> perf_pmu__find() tries to find a PMU in the
global pmu list.
+--> pmu_lookup() called to read all file
entries when not in global
list.
pmu_lookup() causes the issue. It calls
+---> pmu_aliases() to read all the entries in the PMU directory.
On s390 this is named
/sys/devices/cpum_cf/events.
+--> pmu_aliases_parse() reads all files and creates an
alias for each file name.
So we end up with first entry created by
reading the sysfs file
[root@s35lp76 perf]# cat /sys/devices/cpum_cf
/events/TX_NC_TEND
event=0x008d
[root@s35lp76 perf]#
Debug output shows this entry
tx_nc_tend -> 'cpum_cf'/'event=0x008d
'/
After all files in this directory have been
read and aliases created this function is called:
+--> pmu_add_cpu_aliases()
This function looks up the CPU tables
created by the json files.
With json files for s390 now available all
the aliases are added to
the PMU alias list a second time.
The second entry is added by
reading the json file converted by jevent
resulting in file pmu-events/pmu-events.c:
{
.name = "tx_nc_tend",
.event = "event=0x8d",
.desc = "Unit: cpum_cf Completed TEND \
instructions \
in non-constrained TX mode",
.topic = "extended",
.long_desc = "A TEND instruction has \
completed in a \
non-constrained \
transactional-execution mode",
.pmu = "cpum_cf",
},
Debug output shows this entry
tx_nc_tend -> 'cpum_cf'/'event=0x8d'/
Function pmu_aliases_parse() and pmu_add_cpu_aliases() both use
__perf_pmu__new_alias() to add an alias to the PMU alias list. There is
no check if an alias already exist
So we end up with 2 entries for tx_nc_tend in the PMU alias list.
Having set up the PMU alias list for this PMU now
parse_events_multi_add_pmu() reads the complete alias list and adds each
alias with parse_events_add_pmu() to the global perfev_list. This
causes the alias to be added multiple times to the event list.
Fix this by making __perf_pmu__new_alias() to merge alias definitions if
an alias is already on the alias list. Also print a debug message when
the alias has mismatches in some fields.
Output before:
[root@s35lp76 perf]# ./perf stat -e tx_nc_tend -v \
-- ~/mytesttx 1 >/tmp/111
tx_nc_tend: 1 551446 551446
Performance counter stats for '/root/mytesttx 1':
3 tx_nc_tend
0.000961134 seconds time elapsed
[root@s35lp76 perf]#
Output after:
[root@s35lp76 perf]# ./perf stat -e tx_nc_tend -v \
-- ~/mytesttx 1 >/tmp/111
tx_nc_tend: 1 551446 551446
Performance counter stats for '/root/mytesttx 1':
1 tx_nc_tend
0.000961134 seconds time elapsed
[root@s35lp76 perf]#
Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Reviewed-by: Hendrik Brueckner <brueckner@linux.ibm.com>
Reviewed-by: Jiri Olsa <jolsa@redhat.com>
Cc: Heiko Carstens <heiko.carstens@de.ibm.com>
Cc: Martin Schwidefsky <schwidefsky@de.ibm.com>
Link: http://lkml.kernel.org/r/20180615101105.47047-3-tmricht@linux.ibm.com
Signed-off-by: Arnaldo Carvalho de Melo <acme@redhat.com>
6dde6429c5