[PATCH V2][Ping] perf test: Test 17 fails with make LIBPFM4=1 on s390 z/VM

[PATCH V2][Ping] perf test: Test 17 fails with make LIBPFM4=1 on s390 z/VM

[Thomas Richter]

Version 2:
   As suggested by Ian Rogers make perf_event_attribute member
   exclude_hv more robust and accept value 0 or 1 to handle more
   test cases which might fail on s390 virtual machine z/VM.

This test case fails on s390 virtual machine z/VM which has no PMU support
when the perf tool is built with LIBPFM4=1.

Using make LIBPFM4=1 builds the perf tool with support for libpfm
event notation. The command line flag --pfm-events is valid:
 # ./perf record --pfm-events cycles -- true
 [ perf record: Woken up 1 times to write data ]
 [ perf record: Captured and wrote 0.001 MB perf.data (2 samples) ]
 #

However the command 'perf test -Fv 17' fails on s390 z/VM virtual machine
with LIBPFM4=1:
  # perf test -Fv 17
  17: Setup struct perf_event_attr                                    :
  --- start ---
  .....
  running './tests/attr/test-record-group2'
  unsupp  './tests/attr/test-record-group2'
  running './tests/attr/test-record-pfm-period'
  expected exclude_hv=0, got 1
 FAILED './tests/attr/test-record-pfm-period' - match failure
 ---- end ----
 Setup struct perf_event_attr: FAILED!

When --pfm-event system is not supported, the test returns unsupported
and continues. Here is an example using a virtual machine on x86 and
Fedora 34:
 [root@f33 perf]# perf test -Fv 17
 17: Setup struct perf_event_attr                                    :
 --- start ---
 .....
 running './tests/attr/test-record-group2'
 unsupp  './tests/attr/test-record-group2'
 running './tests/attr/test-record-pfm-period'
 unsupp  './tests/attr/test-record-pfm-period'
 ....

The issue is file ./tests/attr/test-record-pfm-period
which requires perf event attribute member exclude_hv to be zero.
This is not the case on s390 where the value of exclude_hv is one when
executing on a z/VM virtual machine without PMU hardware support.

Fix this by allowing value exlucde_hv to be zero or one.

Output before:
 # /usr/bin/python ./tests/attr.py -d ./tests/attr/ -t \
	test-record-pfm-period -p ./perf  -vvv 2>&1| fgrep match
    matching [event:base-record]
    match: [event:base-record] matches []
 FAILED './tests/attr//test-record-pfm-period' - match failure
 #

Output after:
 # /usr/bin/python ./tests/attr.py -d ./tests/attr/ -t \
	test-record-pfm-period -p ./perf  -vvv 2>&1| fgrep match
    matching [event:base-record]
    match: [event:base-record] matches ['event-1-0-6', 'event-1-0-5']
  matched

Background:
Using libpfm library ends up in this function call sequence

pfm_get_perf_event_encoding()
+-- pfm_get_os_event_encoding()
    +-- pfmlib_perf_event_encode()

is called when no hardware specific PMU unit can be detected
as in the s390 z/VM virtual machine case. This uses the
"perf_events generic PMU" data structure which sets exclude_hv
to 1 per default.  Using this PMU that test case always fails.

That is the reason why exclude_hv attribute setting varies.

Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
Suggested-by: Ian Rogers <irogers@google.com>
---
 tools/perf/tests/attr/base-record | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/tools/perf/tests/attr/base-record b/tools/perf/tests/attr/base-record
index 4a7b8deef3fd..8c10955eff93 100644
--- a/tools/perf/tests/attr/base-record
+++ b/tools/perf/tests/attr/base-record
@@ -16,7 +16,7 @@ pinned=0
 exclusive=0
 exclude_user=0
 exclude_kernel=0|1
-exclude_hv=0
+exclude_hv=0|1
 exclude_idle=0
 mmap=1
 comm=1
-- 
2.31.1

Re: [PATCH V2][Ping] perf test: Test 17 fails with make LIBPFM4=1 on s390 z/VM

[Ian Rogers]

On Fri, May 28, 2021 at 2:11 AM Thomas Richter <tmricht@linux.ibm.com> wrote:
>
> Version 2:
>    As suggested by Ian Rogers make perf_event_attribute member
>    exclude_hv more robust and accept value 0 or 1 to handle more
>    test cases which might fail on s390 virtual machine z/VM.
>
> This test case fails on s390 virtual machine z/VM which has no PMU support
> when the perf tool is built with LIBPFM4=1.
>
> Using make LIBPFM4=1 builds the perf tool with support for libpfm
> event notation. The command line flag --pfm-events is valid:
>  # ./perf record --pfm-events cycles -- true
>  [ perf record: Woken up 1 times to write data ]
>  [ perf record: Captured and wrote 0.001 MB perf.data (2 samples) ]
>  #
>
> However the command 'perf test -Fv 17' fails on s390 z/VM virtual machine
> with LIBPFM4=1:
>   # perf test -Fv 17
>   17: Setup struct perf_event_attr                                    :
>   --- start ---
>   .....
>   running './tests/attr/test-record-group2'
>   unsupp  './tests/attr/test-record-group2'
>   running './tests/attr/test-record-pfm-period'
>   expected exclude_hv=0, got 1
>  FAILED './tests/attr/test-record-pfm-period' - match failure
>  ---- end ----
>  Setup struct perf_event_attr: FAILED!
>
> When --pfm-event system is not supported, the test returns unsupported
> and continues. Here is an example using a virtual machine on x86 and
> Fedora 34:
>  [root@f33 perf]# perf test -Fv 17
>  17: Setup struct perf_event_attr                                    :
>  --- start ---
>  .....
>  running './tests/attr/test-record-group2'
>  unsupp  './tests/attr/test-record-group2'
>  running './tests/attr/test-record-pfm-period'
>  unsupp  './tests/attr/test-record-pfm-period'
>  ....
>
> The issue is file ./tests/attr/test-record-pfm-period
> which requires perf event attribute member exclude_hv to be zero.
> This is not the case on s390 where the value of exclude_hv is one when
> executing on a z/VM virtual machine without PMU hardware support.
>
> Fix this by allowing value exlucde_hv to be zero or one.
>
> Output before:
>  # /usr/bin/python ./tests/attr.py -d ./tests/attr/ -t \
>         test-record-pfm-period -p ./perf  -vvv 2>&1| fgrep match
>     matching [event:base-record]
>     match: [event:base-record] matches []
>  FAILED './tests/attr//test-record-pfm-period' - match failure
>  #
>
> Output after:
>  # /usr/bin/python ./tests/attr.py -d ./tests/attr/ -t \
>         test-record-pfm-period -p ./perf  -vvv 2>&1| fgrep match
>     matching [event:base-record]
>     match: [event:base-record] matches ['event-1-0-6', 'event-1-0-5']
>   matched
>
> Background:
> Using libpfm library ends up in this function call sequence
>
> pfm_get_perf_event_encoding()
> +-- pfm_get_os_event_encoding()
>     +-- pfmlib_perf_event_encode()
>
> is called when no hardware specific PMU unit can be detected
> as in the s390 z/VM virtual machine case. This uses the
> "perf_events generic PMU" data structure which sets exclude_hv
> to 1 per default.  Using this PMU that test case always fails.
>
> That is the reason why exclude_hv attribute setting varies.
>
> Signed-off-by: Thomas Richter <tmricht@linux.ibm.com>
> Suggested-by: Ian Rogers <irogers@google.com>

Reviewed-by: Ian Rogers <irogers@google.com>

Tested on x86 with and without hypervisor and test still passes. This
would be expected as the change is making a flag more permissive. As
we don't care about this flag in existing tests I believe there's no
impact to this change other than fixing the tests on s390.

Thanks,
Ian

> ---
>  tools/perf/tests/attr/base-record | 2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
>
> diff --git a/tools/perf/tests/attr/base-record b/tools/perf/tests/attr/base-record
> index 4a7b8deef3fd..8c10955eff93 100644
> --- a/tools/perf/tests/attr/base-record
> +++ b/tools/perf/tests/attr/base-record
> @@ -16,7 +16,7 @@ pinned=0
>  exclusive=0
>  exclude_user=0
>  exclude_kernel=0|1
> -exclude_hv=0
> +exclude_hv=0|1
>  exclude_idle=0
>  mmap=1
>  comm=1
> --
> 2.31.1
>