Re: [PATCH 2/2] x86: Add generic CPUID data dumper to ld.so --list-diagnostics

[Adhemerval Zanella Netto <adhemerval.zanella@linaro.org>]

On 11/09/23 13:19, Florian Weimer wrote:
> * Adhemerval Zanella Netto:
> 
>>> +void
>>> +_dl_diagnostics_cpu_kernel (void)
>>> +{
>>> +#if !HAS_CPUID
>>> +  /* CPUID is not supported, so there is nothing to dump.  */
>>> +  if (__get_cpuid_max (0, 0) == 0)
>>> +    return;
>>> +#endif
>>
>> I think we don't support __i486__ anymore, so we can just assume HAS_CPUID
>> at sysdeps/x86/include/cpu-features.h.
> 
> We still build an i486 variant in build-many-glibcs.py.

Indeed.

>> Why not use the interfaces to work on cpuset? 
>>
>>   if (length_reference > 0)
>>     {
>>       int cpu_count = CPU_COUNT_S (length_reference, mask_reference);
>>       for (int i = 0; i < cpu_count; i++)
>>         {
>>           if (CPU_ISSET_S (i, length_reference, mask_reference)
>>             {
>>               CPU_SET_S (i, length_reference, mask_request);
>>               if (INTERNAL_SYSCALL_CALL (sched_setaffinity, 0,
>> 					 length_reference, mask_request) == 0)
>>                 {
>>                   _dl_diagnostics_cpuid_collect (&ccd[i & 1]);
>>                   _dl_diagnostics_cpuid_report (processor_index, i, 
>> 						&ccd[processor_index & 1],
>> 						&ccd[!(processor_index & 1)]);
>>                   ++processor_index;
>>                 }
>>               CPU_CLR_S (i, length_reference, mask_request);
>>             }
>>         }
>>     }
>>
>> I will iterate over the list twice, but I don't think this would really matter
>> here.
> 
> I think the macros are somewhat incompatible with direct system calls.
> CPU_COUNT_S requires that the tail is zeroed, which the system call
> doesn't do.  Maybe we can memset the whole thing before the sched_getcpu
> system call.  I haven't tried if we can directly rebuild
> __sched_cpucount for ld.so, either.

Afaiu the setsize is exactly to avoid the tail zeroing and to use the syscall
return code, isn't?  It does not work outside libc because sched_getaffinity
does not return the bytes set from the kernel, but it does work with direct
syscalls:

$ cat > test.c <<EOF
#define _GNU_SOURCE
#include <assert.h>
#include <sched.h>
#include <stdio.h>
#include <string.h>
#include <sys/syscall.h>
#include <unistd.h>

int
main (int argc, char *argv[])
{
  size_t cpusetsize = CPU_ALLOC_SIZE (1024);
  cpu_set_t cpuset[cpusetsize];
  memset (cpuset, 0xff, cpusetsize * sizeof (cpu_set_t));

  long int sz = syscall (SYS_sched_getaffinity, getpid (), cpusetsize, cpuset);

  printf ("%ld %d\n", sz, CPU_COUNT_S (sz, cpuset));

  return 0;
}
EOF
$ gcc -Wall test.c -o test && ./test
8 24