On date Saturday 2011-05-28 11:23:50 +0200, Anton Khirnov wrote: 
> On Sat, 28 May 2011 09:54:27 +0200, Stefano Sabatini 
> <[email protected]> wrote:
> > On date Saturday 2011-05-28 07:32:47 +0200, Anton Khirnov encoded:
> > > 
> > > On Fri, 27 May 2011 21:25:21 -0400, "Ronald S. Bultje" 
> > > <[email protected]> wrote:
> > > > Hi,
> > > > 
> > > > On Fri, May 27, 2011 at 7:53 PM, Stefano Sabatini
> > > > <[email protected]> wrote:
> > > > > +int av_parse_number(double *res, const char *numstr, enum 
> > > > > AVParseNumberType type,
> > > > > +                    double min, double max,
> > > > > +                    int log_offset, void *log_ctx)
> > > > > +{
> > > > > +    ParseUtils parseutils = { &parseutils_class, log_offset, log_ctx 
> > > > > };
> > > > > +
> > > > > +    char *tail;
> > > > > +    double d = av_strtod(numstr, &tail);
> > > > > +    if ((unsigned)type >= AV_PARSE_NUM_TYPE_NB) {
> > > > > +        av_log(&parseutils, AV_LOG_ERROR, "Unknown parse number type 
> > > > > '%d'\n", type);
> > > > 
> > > > This is a little bit of an ugly way to set up a logging context each
> > > > time this is called. I admit this code isn't performance-critical,
> > > > just looks a little ... Awkward. Can't we just pass it an existing
> > > > logging context? The point of the log context is generally not to know
> > > > where the error occurred (grep -r rocks), but rather to see what
> > > > object it happened with.
> > > 
> > 
> > > I think those functions have way too many parameters to be considered
> > > developer-friendly. If the problem is that the function has nothing
> > > to log to -- well, then it shouldn't log at all. Just return
> > > an appropriate error code and have the caller deal with it.
> > 
> > Sometimes you may want to log, sometimes you need just to skip logging
> > and avoid useless log spamming.
> > 
> > In this specific case providing a logging context allows the
> > application code to avoid to define a log message each time the
> > function is called.
> > 
> > The log_ctx+log_level_offset is the mechanism employed for such a
> > scenario (e.g. libavutil/imgutils.c:av_image_check_size(),
> > libavutil/eval.h), I agree that is a little awkward but at least from
> > the API POV doesn't look that bad.
> > 
> > And if the number of parameters is a problem you can either define a
> > macro/convenience function in the application, either define them in
> > the lib, e.g.
> > av_parse_number_no_log(double *res, const char *numstr, enum 
> > AVParseNumberType type, double min, double max);
> > 
> > but then you're trading function complexity with API complexity.
> 

> I think that such a utility function shouldn't log anything
> at all ever. It's simply not its place to decide such things.
> I wouldn't expect av_malloc to log anything either.

Logging in the called function code has two advantages:

* you don't have to log the same message again and again in
  application code, in case of failure

* the function code knows exactly why the failure happened, and can
  provide useful hints for spotting the problem, which can't be done
  in the calling code.

libav* adopts both logging and non-logging policy, in general it makes
sense to define a log context when there is something meaningful to
say and/or you don't want to repeat the message in application code.
_______________________________________________
libav-devel mailing list
[email protected]
https://lists.libav.org/mailman/listinfo/libav-devel

Reply via email to