Re: [PATCH] mellanox: mlx5: Use logging functions to reduce text ~10k/5%

[Joe Perches <joe@perches.com>]

On Wed, 2016-06-22 at 14:40 -0600, Jason Gunthorpe wrote:
> On Wed, Jun 22, 2016 at 11:23:59AM -0700, Joe Perches wrote:
> > The output changes now do not include line #, but do include the
> > function offset.
> I've been using a technique like this in some code with good results:
> 
> struct source_location
> {
>    const char *file;
>    const char *func;
>    const char *format;
>    uint16_t line;
> };
> #define _LOCATION(format) ({static const source_location __location__
> = {\
>              __FILE__,__PRETTY_FUNCTION__,format,__LINE__};\
> 	     &__location__;})
> 
> void _mlx5_core_err(const struct source_location *loc,struct
> mlx5_core_dev *dev, ...);
> #define mlx5_core_err(dev,format,...)
> _mlx_core_err(_LOCATION(format),dev,__VA_ARGS__)
> 
> The call site .text overhead is the about same as what you have, but
> this still retains the function and line number information in
> .rodata.

Hello Jason.

As far as I know, no kernel code currently uses a _LOCATION
like macro.

I think your proposal is nearly identical code size to the
existing call.  Also, compiler format/argument checking is
eliminated and I think that is a significant negative.

Using the kernel vsprintf %pS or %ps extension is pretty common.

Using printk("%pS", __builtin_return_address(0)); in the called
function is no overhead at all and returns almost exactly
the same information.

Using more expressive messages is generally better than using
printk("%d", __LINE__);