[Buildroot] [PATCH 1/1] stress-ng: Depend on kernel headers >= 3.3

[Romain Naour <romain.naour@openwide.fr>]

Hi Thomas,

Le 11/09/2015 19:59, Thomas Petazzoni a ?crit :
> Julien,
> 
> On Fri, 11 Sep 2015 18:19:11 +0200, Julien Corjon wrote:
>> perf.c need PERF_COUNT_HW_REF_CPU_CYCLES instruction which is available
>> only since 3.3 kernel version
>>
>> Signed-off-by: Julien Corjon <corjon.j@ecagroup.com>
>> ---
>>  package/stress-ng/Config.in | 8 ++++----
>>  1 file changed, 4 insertions(+), 4 deletions(-)
>>
>> diff --git a/package/stress-ng/Config.in b/package/stress-ng/Config.in
>> index cdf9984..1e475b1 100644
>> --- a/package/stress-ng/Config.in
>> +++ b/package/stress-ng/Config.in
>> @@ -4,8 +4,8 @@ config BR2_PACKAGE_STRESS_NG
>>  	# disabled on musl: stress-malloc.c needs mallopt() and M_MMAP_THRESHOLD
>>  	# disabled on uClibc: stress-aio.c needs aio.h
>>  	depends on BR2_TOOLCHAIN_USES_GLIBC
>> -	# perf.c needs PERF_COUNT_HW_STALLED_CYCLES_FRONTEND
>> -	depends on BR2_TOOLCHAIN_HEADERS_AT_LEAST_3_0
>> +	# perf.c needs PERF_COUNT_HW_REF_CPU_CYCLES
>> +	depends on BR2_TOOLCHAIN_HEADERS_AT_LEAST_3_3
> 
> This seems wrong, because stress-ng.h has the definitions for all the
> PERF_COUNT_* values:
> 
> #if defined(__linux__) && defined(__NR_perf_event_open)
> #define STRESS_PERF_STATS       (1)
> #define STRESS_PERF_INVALID     (~0ULL)
> enum {
>         STRESS_PERF_HW_CPU_CYCLES = 0,
>         STRESS_PERF_HW_INSTRUCTIONS,
>         STRESS_PERF_HW_CACHE_REFERENCES,
>         STRESS_PERF_HW_CACHE_MISSES,
>         STRESS_PERF_HW_BRANCH_INSTRUCTIONS,
>         STRESS_PERF_HW_BRANCH_MISSES,
>         STRESS_PERF_HW_BUS_CYCLES,
>         STRESS_PERF_HW_REF_CPU_CYCLES,
>         STRESS_PERF_HW_STALLED_CYCLES_FRONTEND,
>         STRESS_PERF_HW_STALLED_CYCLES_BACKEND,
> 
>         STRESS_PERF_SW_PAGE_FAULTS_MIN,
>         STRESS_PERF_SW_PAGE_FAULTS_MAJ,
>         STRESS_PERF_SW_CONTEXT_SWITCHES,
>         STRESS_PERF_SW_CPU_MIGRATIONS,
>         STRESS_PERF_SW_ALIGNMENT_FAULTS,
> 
> Can you figure out why it doesn't get used?

Here is the complete build failure with a toolchain using 3.2 kernel headers:
[...]i586-buildroot-linux-gnu-gcc -D_LARGEFILE_SOURCE -D_LARGEFILE64_SOURCE
-D_FILE_OFFSET_BITS=64 -Os -Wall -Wextra -DVERSION="0.04.16" -O2 -c -o perf.o perf.c
perf.c:77:48: erreur: ?PERF_COUNT_HW_REF_CPU_CYCLES? undeclared here (not in a
function)
{ STRESS_PERF_ ## config, PERF_TYPE_ ## type, PERF_COUNT_ ## config, label }
^
perf.c:110:2: note: in expansion of macro ?PERF_INFO?
PERF_INFO(HARDWARE, HW_REF_CPU_CYCLES, "Total Cycles"),
^
Makefile:164: recipe for target 'perf.o' failed

This error is due to PERF_INFO macro from perf.c using PERF_COUNT_* values directly.

During my testing, I used an very old toolchain with 2.6 headers so I had an
error with PERF_COUNT_HW_STALLED_CYCLES_FRONTEND and I missed
PERF_COUNT_HW_REF_CPU_CYCLES.

I'm agree with the patch proposed by Julien, so
Acked-by: Romain Naour <romain.naour@openwide.fr>

Best regards,
Romain

> 
> Thanks,
> 
> Thomas
>