Re: [PATCH 3/4] linux: Use /sys/devices/system/cpu/possible on __get_nprocs_conf

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

On 07/05/2021 08:07, Florian Weimer wrote:
> * Adhemerval Zanella via Libc-alpha:
> 
>> On 06/05/2021 10:51, Florian Weimer wrote:
>>> * Adhemerval Zanella via Libc-alpha:
>>>
>>>> On 05/05/2021 15:06, Florian Weimer wrote:
>>>>> * Florian Weimer:
>>>>>
>>>>>> * Adhemerval Zanella via Libc-alpha:
>>>>>>
>>>>>>> +      /* The entry is in the form of '[cpuX]-[cpuY]'.  */
>>>>>>> +      char buf[2 * INT_STRLEN_BOUND (unsigned int) + 1];
>>>>>>> +
>>>>>>> +      ssize_t n = __read_nocancel (fd, buf, sizeof (buf));
>>>>>>> +      if (n > 0)
>>>>>>> +	{
>>>>>>> +	  buf[n] = '\0';
>>>>>>> +
>>>>>>> +	  /* Start on the right, to find highest node number.  */
>>>>>>> +	  int m = 1;
>>>>>>> +	  while (--n)
>>>>>>> +	    {
>>>>>>> +	      if ((buf[n] == ',') || (buf[n] == '-'))
>>>>>>> +		break;
>>>>>>> +	      /* Ignore '\n'  */
>>>>>>> +	      if (! isdigit (buf[n]))
>>>>>>> +		continue;
>>>>>>> +	      result += (buf[n] - '0') * m;
>>>>>>> +	      m *= 10;
>>>>>>> +	    }
>>>>>>> +	}
>>>>>>> +
>>>>>>> +      __close_nocancel (fd);
>>>>>>> +      return result + 1;
>>>>>>>      }
>>>>>>
>>>>>> I think the /online and /possible files have the same layout, so you
>>>>>> could use both.
>>>>>
>>>>> Sorry, I meant to write: “so you could use *one parser for* both”
>>>>
>>>> I am not following, the second patch in this set *removed* the 
>>>> /sys/devices/system/cpu/online parsing in favor of the sched_getaffinity.
>>>> So now it only parses /sys/devices/system/cpu/possible and fallbacks
>>>> to sched_getaffinity.
>>>
>>> Oh, right.  So you removed the old implementation and bring bug a
>>> slightly different new one.  Got it.
>>
>> Right, are you ok with the patch then?
> 
> The new parser cannot handle gaps or ranges that do not start at 0, I
> think.  The old parser could cope with that.  The kernel data structures
> support gaps in the possible CPU mask.  I don't know if they occur in
> practice, but firmware quirks in this area aren't exactly rare (e.g.,
> single-socket systems which report hundreds of hot-pluggable CPU cores).

Indeed I did not take this in consideration and it seems that using
/possible is more complicated than just opendir the directory.  And 
we will need to handle the possible unlikely issue of a truncated
value from kernel, since the buffer is limited to PAGE_SIZE.

I will drop this patch.