[LTP] [PATCH] syscalls/getcpu: Add libc sched_getcpu() detection && fix compiler errors

[xuyang <xuyang2018.jy@cn.fujitsu.com>]

on 2019/4/12 17:43, Cyril Hrubis wrote:

> Hi!
>> 1)sched_getcpu() isn't defined on some old distros, so we can add dection and
>> fix compiler error.
>>
>> 2)If CPU_ALLOC isn't defined on some old distros, it will used "include/lapi/cpuset"
>> CPU_ALLOC macro. But getcpu01.c has converted into new library, so it will report
>> "cleanup" undefine error. We should ensure CPU_ALLOC can be used both new cases and
>> old cases.
>>
>> Signed-off-by: Yang Xu<xuyang2018.jy@cn.fujitsu.com>
>> ---
>>   configure.ac                                |  1 +
>>   include/lapi/cpuset.h                       | 10 +++++++++-
>>   testcases/kernel/syscalls/getcpu/getcpu01.c |  6 +++---
>>   3 files changed, 13 insertions(+), 4 deletions(-)
>>
>> diff --git a/configure.ac b/configure.ac
>> index de12499..3ac0f28 100644
>> --- a/configure.ac
>> +++ b/configure.ac
>> @@ -77,6 +77,7 @@ AC_CHECK_FUNCS([ \
>>       readlinkat \
>>       renameat \
>>       renameat2 \
>> +    sched_getcpu \
>>       sigpending \
>>       splice \
>>       stime \
>> diff --git a/include/lapi/cpuset.h b/include/lapi/cpuset.h
>> index b946650..959462b 100644
>> --- a/include/lapi/cpuset.h
>> +++ b/include/lapi/cpuset.h
>> @@ -33,11 +33,19 @@
>>   #define LTP_CPUSET_H
>>
>>   #ifndef CPU_ALLOC
>> +#ifndef TST_TEST_H__
>>   #define CPU_ALLOC(ncpus) malloc(sizeof(cpu_set_t)); \
>>   if (ncpus>  CPU_SETSIZE) { \
>>   	tst_brkm(TCONF, cleanup, \
>> -		"Your libc does not support masks with %ld cpus", ncpus); \
>> +		"Your libc does not support masks with %d cpus", ncpus); \
>>   }
>> +#else
>> +#define CPU_ALLOC(ncpus) malloc(sizeof(cpu_set_t)); \
>> +if (ncpus>  CPU_SETSIZE) { \
>> +	tst_brk(TCONF, \
>> +		"Your libc does not support masks with %d cpus", ncpus); \
>> +}
>> +#endif
>>   #endif
> Good catch.
>
> I would be happier if we converted the sched_setaffinity01 to new
> library then we could simply change the tst_brkm() to tst_brk(), but we
> can always do that later on.
>
Hi Cyril

     I will converted the sched_setaffinity01 to new library in V2 patch.

>>   #ifndef CPU_FREE
>> diff --git a/testcases/kernel/syscalls/getcpu/getcpu01.c b/testcases/kernel/syscalls/getcpu/getcpu01.c
>> index eb6ded8..d033119 100644
>> --- a/testcases/kernel/syscalls/getcpu/getcpu01.c
>> +++ b/testcases/kernel/syscalls/getcpu/getcpu01.c
>> @@ -16,15 +16,15 @@
>>   #include<stdlib.h>
>>   #include<sys/types.h>
>>   #include "lapi/syscalls.h"
>> -#include "lapi/cpuset.h"
>>   #include "tst_test.h"
>> +#include "lapi/cpuset.h"
>>
>>   static inline int get_cpu(unsigned *cpu_id,
>>   			  unsigned *node_id LTP_ATTRIBUTE_UNUSED,
>>   			  void *cache_struct LTP_ATTRIBUTE_UNUSED)
>>   {
>> -#if defined(__i386__)
>> -	return syscall(__NR_getcpu, cpu_id, node_id, cache_struct);
>> +#ifndef HAVE_SCHED_GETCPU
>> +	return tst_syscall(__NR_getcpu, cpu_id, node_id, cache_struct);
>>   #else
>>   	*cpu_id = sched_getcpu();
>>   #endif
> So when __NR_getcpu is not implemented we end up with tst_brk(TCONF,
> ...) called from the tst_syscall, right? Since AFAIK the getcpu syscall
> is not implemented on all architectures...
Yes. It will report TCONF if __NR_getcpu is not implemented.

 From getcpu manpage, it is not implemented on all architectures as you
said, I think we can remove sched_getcpu and use __NR_getcpu directly.