Re: [PATCH v3 2/4] ethdev: move error checking macros to header

[Adrien Mazarguil <adrien.mazarguil@6wind.com>]

On Mon, Nov 09, 2015 at 02:39:05PM +0100, Adrien Mazarguil wrote:
> On Fri, Nov 06, 2015 at 05:22:27PM +0000, Bruce Richardson wrote:
> > On Fri, Nov 06, 2015 at 05:10:07PM +0000, Bruce Richardson wrote:
> > > On Thu, Nov 05, 2015 at 04:09:18PM +0100, Adrien Mazarguil wrote:
> > > > 
> > > > I won't argue against this as it's obviously more complex than the original
> > > > method, however note that users of the RTE_PMD_DEBUG_TRACE() macro do not
> > > > have to modify their code. They shouldn't care about the implementation.
> > > > 
> > > > Also note that we can do much cleaner code if we drop the all macros
> > > > implementation using a (much easier to debug) static inline function,
> > > > only perhaps with a wrapper macro that provides __LINE__, __func__ and
> > > > __FILE__ as arguments. Nontrival code shouldn't be done in macros anyway.
> > > > 
> > > Getting something working with __FILE__ and probably __LINE__ would be easy enough
> > > with a helper macro, but __func__ is not so easy as it's not a preprocessor symbol
> > > [since the pre-processor has no idea what function you are in].
> > > 
> > > However, using func, here is the best I've come up with so far. It's not that
> > > pretty, but it's probably easier to work with than the macro version.
> > > 
> > >  #ifdef RTE_LIBRTE_ETHDEV_DEBUG
> > >  -#define RTE_PMD_DEBUG_TRACE(fmt, args...) \
> > >  -               RTE_LOG(ERR, PMD, "%s: " fmt, __func__, ## args)
> > >  +#define RTE_PMD_DEBUG_TRACE(...) \
> > >  +               rte_pmd_debug_trace(__func__, __VA_ARGS__)
> > >  +
> > >  +static inline void
> > >  +rte_pmd_debug_trace(const char *func_name, const char *fmt, ...)
> > >  +{
> > >  +       static __thread char buffer[128];
> > >  +       char *out_buf = buffer;
> > >  +       unsigned count;
> > >  +       va_list ap;
> > >  +
> > >  +       count = snprintf(buffer, sizeof(buffer), "%s: %s", func_name, fmt);
> > >  +       if (count >= sizeof(buffer)) { // truncated output
> > >  +               char *new_buf = malloc(count + 1);
> > >  +               if (new_buf == NULL) // no memory, just print 128 chars
> > >  +                       goto print_buffer;
> > >  +               snprintf(new_buf, count + 1, "%s: %s", func_name, fmt);
> > >  +               va_start(ap, fmt);
> > >  +               rte_vlog(RTE_LOG_ERR, RTE_LOGTYPE_PMD, buffer, ap);
> > >  +               va_end(ap);
> > >  +               free(new_buf);
> > >  +               return;
> > >  +       }
> > >  +
> > >  +print_buffer:
> > >  +       va_start(ap, fmt);
> > >  +       rte_vlog(RTE_LOG_ERR, RTE_LOGTYPE_PMD, out_buf, ap);
> > >  +       va_end(ap);
> > >  +}
> > >   #else
> > >   #define RTE_PMD_DEBUG_TRACE(fmt, args...)
> > >   #endif
> > > 
> > > Comments or improvements?
> 
> Such a function shouldn't malloc() anything. The entire line should fit on
> the stack (crashing is fine if it does not, then it's probably a bug). We
> did something in two passes along these lines in mlx5_defs.h (not pretty but
> quite useful):
> 
>  /* Allocate a buffer on the stack and fill it with a printf format string. */
>  #define MKSTR(name, ...) \
>          char name[snprintf(NULL, 0, __VA_ARGS__) + 1]; \
>          \
>          snprintf(name, sizeof(name), __VA_ARGS__)
> 
> Untested but I guess modifying that function accordingly would look like:
> 
>  static inline void
>  rte_pmd_debug_trace(const char *func_name, const char *fmt, ...)
>  {
>          va_list ap;
>          va_start(ap, fmt);
> 
>          static __thread char buffer[vsnprintf(NULL, 0, fmt, ap)];

Of course this line should read:

         static __thread char buffer[vsnprintf(NULL, 0, fmt, ap) + 1];

>          va_end(ap);
>          va_start(ap, fmt);
>          vsnprintf(buffer, sizeof(buffer), fmt, ap);
>          va_end(ap);
>          rte_log(RTE_LOG_ERR, RTE_LOGTYPE_PMD, "%s: %s", func_name, buffer);
>  }
> 
> > And here's the version if we are happy to have file and line number instead of
> > function name. I think this might be the best option.
> > 
> > /Bruce
> > 
> >  #ifdef RTE_LIBRTE_ETHDEV_DEBUG
> >  -#define RTE_PMD_DEBUG_TRACE(fmt, args...) \
> >  -               RTE_LOG(ERR, PMD, "%s: " fmt, __func__, ## args)
> >  +#define RTE_PMD_DEBUG_TRACE(...) \
> >  +       RTE_LOG(ERR, PMD, __FILE__", " RTE_STR(__LINE__) ": " __VA_ARGS__)
> >   #else
> >  -#define RTE_PMD_DEBUG_TRACE(fmt, args...)
> >  +#define RTE_PMD_DEBUG_TRACE(...)
> >   #endif
> 
> Much cleaner indeed, however __func__ might be useful when comparing log
> outputs from different source code versions. I think we should keep it.

-- 
Adrien Mazarguil
6WIND