Re: [LTP] [PATCH] getcpu01: Reinstate node_id test

[Richard Palethorpe <rpalethorpe@suse.de>]

Hello,

Cyril Hrubis <chrubis@suse.cz> writes:

> Hi!
>> Presently the node_id is only checked on i386 and it is broken. The
>> sched_getcpu call was substituted for getcpu when
>> available. sched_getcpu does not have the node_id parameter and does
>> not even call SYS_getcpu if it can be completed by vDSO.
>> 
>> Also we can at least check the node_id on x86_64 as well.
>
> I suppose that the getcpu manual page is a bit confusing, the function
> is implemented on most of the major achitectures and as VDSO on most of
> them too. It was only added first to x86 architectures.

I suppose I overlooked that there is a libc implementation of
getcpu. IDK why sched_getcpu is here at all and my vDSO comment is just
about how sched_getcpu would not result in a call to getcpu. However I
agree we should just call the libc version.

>
>> Signed-off-by: Richard Palethorpe <rpalethorpe@suse.com>
>> ---
>>  testcases/kernel/syscalls/getcpu/getcpu01.c | 22 +++++++++------------
>>  1 file changed, 9 insertions(+), 13 deletions(-)
>> 
>> diff --git a/testcases/kernel/syscalls/getcpu/getcpu01.c b/testcases/kernel/syscalls/getcpu/getcpu01.c
>> index fcc273e29..21c67f412 100644
>> --- a/testcases/kernel/syscalls/getcpu/getcpu01.c
>> +++ b/testcases/kernel/syscalls/getcpu/getcpu01.c
>> @@ -15,20 +15,16 @@
>>  #include <stdio.h>
>>  #include <stdlib.h>
>>  #include <sys/types.h>
>> +#include "tst_test.h"
>>  #include "lapi/syscalls.h"
>>  #include "lapi/cpuset.h"
>> -#include "tst_test.h"
>>  #include "config.h"
>>  
>>  static inline int get_cpu(unsigned *cpu_id,
>> -			  unsigned *node_id LTP_ATTRIBUTE_UNUSED,
>> -			  void *cache_struct LTP_ATTRIBUTE_UNUSED)
>> +			  unsigned *node_id)
>>  {
>> -#ifndef HAVE_SCHED_GETCPU
>> -	return tst_syscall(__NR_getcpu, cpu_id, node_id, cache_struct);
>> -#else
>> -	*cpu_id = sched_getcpu();
>> -#endif
>> +	return tst_syscall(__NR_getcpu, cpu_id, node_id, NULL);
>
> The call is mostly implemneted as VDSO so it would make much more sense
> to test the libc function instead and test the implementation that is
> actually used in production.
>
>>  	return 0;
>>  }
>>  
>> @@ -78,7 +74,7 @@ realloc:
>>  	return cpu_max;
>>  }
>>  
>> -#ifdef __i386__
>> +#if  defined(__i386__) || defined(__x86_64__)
>>  static unsigned int get_nodeid(unsigned int cpu_id)
>>  {
>>  	DIR *directory_parent, *directory_node;
>> @@ -125,22 +121,22 @@ static void run(void)
>>  {
>>  	unsigned int cpu_id, node_id = 0;
>>  	unsigned int cpu_set;
>
>
> The get_nodeid() just parses sysfs, that should be portable, can't we
> just get rid of the ifdefs completely instead?

That's what I thought too. I Will remove them.

>
>> -#ifdef __i386__
>> +#if defined(__i386__) || defined(__x86_64__)
>>  	unsigned int node_set;
>>  #endif
>>  
>>  	cpu_set = set_cpu_affinity();
>> -#ifdef __i386__
>> +#if defined(__i386__) || defined(__x86_64__)
>>  	node_set = get_nodeid(cpu_set);
>>  #endif
>>  
>> -	TEST(get_cpu(&cpu_id, &node_id, NULL));
>> +	TEST(get_cpu(&cpu_id, &node_id));
>>  	if (TST_RET == 0) {
>>  		if (cpu_id != cpu_set)
>>  			tst_res(TFAIL, "getcpu() returned wrong value"
>>  				" expected cpuid:%d, returned value cpuid: %d",
>>  				cpu_set, cpu_id);
>> -#ifdef __i386__
>> +#if defined(__i386__) || defined(__x86_64__)
>>  		else if (node_id != node_set)
>>  			tst_res(TFAIL, "getcpu() returned wrong value"
>>  				" expected  node id:%d returned  node id:%d",
>> -- 
>> 2.38.1
>> 
>> 
>> -- 
>> Mailing list info: https://lists.linux.it/listinfo/ltp

Also I guess the kver check for 2.6 can be removed.

-- 
Thank you,
Richard.

-- 
Mailing list info: https://lists.linux.it/listinfo/ltp