Re: [PATCH 13/14] tools lib traceevent: Get rid of die() in some string conversion funcitons

[Namhyung Kim <namhyung@kernel.org>]

Hi Arnaldo,

On Tue, 17 Dec 2013 17:02:39 -0300, Arnaldo Carvalho de Melo wrote:
> Em Tue, Dec 17, 2013 at 09:02:36AM +0900, Namhyung Kim escreveu:
>> On Mon, 16 Dec 2013 09:40:51 -0300, Arnaldo Carvalho de Melo wrote:
>> > Em Mon, Dec 16, 2013 at 01:49:11PM +0900, Namhyung Kim escreveu:
>> >> On Fri, 13 Dec 2013 11:52:04 -0300, Arnaldo Carvalho de Melo wrote:
>> >> > All the rest is ok, so its just the malloc + strcpy that remains to be
>> >> > converted, do you want me to do it?
>
>> >> Hmm.. did you mean like this?
>
>> >> 		str = NULL;
>> >>                 if (val)
>> >>                 	asprintf(&str, "TRUE");
>> >>                 else
>> >>                 	asprintf(&str, "FALSE");
>> >>                 return str;
>
>> > More compact:
>
>> > 	if (asprintf(&str, "%s", val ? "TRUE" : "FALSE") < 0)
>> > 		// error handling path
>
>> > At that point str already is set to NULL.
>
>> Okay, this is a new one:
>
> Thanks, it all seems now, but just prior to applying this I noticed:
>
>> Those functions are for stringify filter arguments.  As caller of
>> those functions handles NULL string properly, it seems that it's
>> enough to return NULL rather than calling die().
>
> It handles NULL in what way? This comment:
>
>> @@ -2369,7 +2340,7 @@ static char *arg_to_str(struct event_filter *filter, struct filter_arg *arg)
>>   * Returns a string that displays the filter contents.
>>   *  This string must be freed with free(str).
>> - *  NULL is returned if no filter is found.
>> + *  NULL is returned if no filter is found or allocation failed.
>>   */
>>  char *
>>  pevent_filter_make_string(struct event_filter *filter, int event_id)
>
> Made me a bit unconfortable, so if it handles NULL as a filter not
> found, how will it figure out what happened?
>
> /me looks at the callers...
>
> From just a quick look I couldn't see cases where NULL could cause
> segfaults, but saw some cases where allocation errors would not be
> notified in any way to the user :-\

Right.  I just wanted to keep the existing interface as long as possible.

>
> Anyway, applying this patch, those are other kinds of problems, i.e. further
> fallout from converting from the previous panic()-at-alloc-failure approach.

Thanks!  But there's one more patch (14/14) left from the series.
Please also consider merging it too. :)

Thanks,
Namhyung