Re: BISECTED: perf test 'Miscellaneous Intel PT' failing on Intel hybrid machines

[Adrian Hunter]

On 9/04/24 18:46, Ian Rogers wrote:
> On Tue, Apr 9, 2024 at 8:34 AM Arnaldo Carvalho de Melo <acme@xxxxxxxxxx> wrote:
>>
>> On Tue, Apr 09, 2024 at 12:32:06PM -0300, Arnaldo Carvalho de Melo wrote:
>>> root@x1:~# perf test "Miscellaneous Intel PT testing"
>>> 112: Miscellaneous Intel PT testing                                  : FAILED!
>>> root@x1:~#
>>>
>>> then I revert:
>>>
>>> commit 642e1ac96aaa12aeb41402e68eac7faf5917a67a (HEAD -> perf-tools-next)
>>> Author: Arnaldo Carvalho de Melo <acme@xxxxxxxxxx>
>>> Date:   Tue Apr 9 12:28:49 2024 -0300
>>>
>>>     Revert "perf pmus: Check if we can encode the PMU number in perf_event_attr.type"
>>>
>>>     This reverts commit 82fe2e45cdb00de4fa648050ae33bdadf9b3294a.
>>> ⬢[acme@toolbox perf-tools-next]$
>>>
>>> It works now:
>>>
>>> root@x1:~# perf -v
>>> perf version 6.8.g642e1ac96aaa
>>> root@x1:~# perf test "Miscellaneous Intel PT testing"
>>> 117: Miscellaneous Intel PT testing                                  : Ok
>>> root@x1:~#
>>>
>>> Investigating, if you come up with ideas, lemme know.
>>
>> Some more context:
>>
>> When this patch was implemented/tested I had access only to an ARM64
>> hybrid machine, now my notebook is a Rocket Lake lenovo (13th gen), that
>> is hybrid and the test is failing with:
>>
>> root@x1:~# perf test -v "Miscellaneous Intel PT testing"
>> 112: Miscellaneous Intel PT testing                                  :
>> --- start ---
>> test child forked, pid 304355
>> --- Test system-wide sideband ---
>> Checking for CPU-wide recording on CPU 0
>> OK
>> Checking for CPU-wide recording on CPU 1
>> OK
>> Linux
>> [ perf record: Woken up 1 times to write data ]
>> [ perf record: Captured and wrote 1.934 MB /tmp/perf-test-intel-pt-sh.xACV6V7Hn4/test-perf.data ]
>> OK
>> --- Test per-thread recording ---
>> Workload PIDs are 304377 and 304378
>> perf PID is 304389
>> Waiting for "perf record has started" message
>> OK
>> pid 0 cpu -1 fd 5 : sys_perf_event_open: pid 0  cpu -1  group_fd -1  flags 0x8 = 5
>> pid 0 cpu -1 fd 6 : sys_perf_event_open: pid 0  cpu -1  group_fd -1  flags 0x8 = 6
>> pid 304377 cpu -1 fd 7 : sys_perf_event_open: pid 304377  cpu -1  group_fd -1  flags 0x8 = 7
>> pid 304380 cpu -1 fd 8 : sys_perf_event_open: pid 304380  cpu -1  group_fd -1  flags 0x8 = 8
>> pid 304378 cpu -1 fd 9 : sys_perf_event_open: pid 304378  cpu -1  group_fd -1  flags 0x8 = 9
>> pid 304381 cpu -1 fd 10 : sys_perf_event_open: pid 304381  cpu -1  group_fd -1  flags 0x8 = 10
>> pid 304377 cpu -1 fd 11 : sys_perf_event_open: pid 304377  cpu -1  group_fd -1  flags 0x8 = 11
>> pid 304380 cpu -1 fd 12 : sys_perf_event_open: pid 304380  cpu -1  group_fd -1  flags 0x8 = 12
>> pid 304378 cpu -1 fd 13 : sys_perf_event_open: pid 304378  cpu -1  group_fd -1  flags 0x8 = 13
>> pid 304381 cpu -1 fd 14 : sys_perf_event_open: pid 304381  cpu -1  group_fd -1  flags 0x8 = 14
>> fd 7 : idx 0: mmapping fd 7
>> fd 11 fd_to 7 : idx 0: set output fd 11 -> 7
>> fd 8 : idx 1: mmapping fd 8
>> fd 12 fd_to 8 : idx 1: set output fd 12 -> 8
>> fd 9 : idx 2: mmapping fd 9
>> fd 13 fd_to 9 : idx 2: set output fd 13 -> 9
>> fd 10 : idx 3: mmapping fd 10
>> fd 14 fd_to 10 : idx 3: set output fd 14 -> 10
>> Checking 10 fds
>> No mmap for fd 5
> 
> Thanks Arnaldo, so the reverted change is:
> ```
> --- a/tools/perf/util/pmus.c
> +++ b/tools/perf/util/pmus.c
> @@ -4,6 +4,7 @@
> #include <subcmd/pager.h>
> #include <sys/types.h>
> #include <dirent.h>
> +#include <pthread.h>
> #include <string.h>
> #include <unistd.h>
> #include "debug.h"
> @@ -492,9 +493,35 @@ int perf_pmus__num_core_pmus(void)
>        return count;
> }
> 
> +static bool __perf_pmus__supports_extended_type(void)
> +{
> +       struct perf_pmu *pmu = NULL;
> +
> +       if (perf_pmus__num_core_pmus() <= 1)
> +               return false;
> +
> +       while ((pmu = perf_pmus__scan_core(pmu)) != NULL) {
> +               if (!is_event_supported(PERF_TYPE_HARDWARE,
> PERF_COUNT_HW_CPU_CYCLES | ((__u64)pmu->
> type << PERF_PMU_TYPE_SHIFT)))
> +                       return false;
> +       }
> +
> +       return true;
> +}
> +
> +static bool perf_pmus__do_support_extended_type;
> +
> +static void perf_pmus__init_supports_extended_type(void)
> +{
> +       perf_pmus__do_support_extended_type =
> __perf_pmus__supports_extended_type();
> +}
> +
> bool perf_pmus__supports_extended_type(void)
> {
> -       return perf_pmus__num_core_pmus() > 1;
> +       static pthread_once_t extended_type_once = PTHREAD_ONCE_INIT;
> +
> +       pthread_once(&extended_type_once,
> perf_pmus__init_supports_extended_type);
> +
> +       return perf_pmus__do_support_extended_type;
> }
> 
> struct perf_pmu *evsel__find_pmu(const struct evsel *evsel)
> ```
> On your Intel this should have previously returned true as
> "perf_pmus__num_core_pmus() > 1", and with the new code presumably
> is_event_supported is returning false. Could you dump the PMU's name
> at that point? Is cpu_core or cpu_atom looking like it doesn't support
> the event? Is the test failing when run as root (ie is
> is_event_supported failing to have expected fallback paths)?

Problem is the test scrapes debug output and is_event_supported()
prints out debug that confuses the test.  Other probe functions
like in perf_api_probe.c use sys_perf_event_open() so the issue
has not arisen before.