Re: [PATCH] mmc_spi: do propper retry managment in the block layer - 3rd try

[Matt Fleming <matt@console-pimps.org>]

On Thu, May 28, 2009 at 10:28:29AM +0200, Wolfgang Mües wrote:
> Pierre,
> 
> Am Mittwoch, 27. Mai 2009 schrieb Pierre Ossman:
>

[...]

> > > +		/* Do retries for all sort of transmission errors */
> > > +		switch (error) {
> > >
> > > -		/*
> > > -		 * A block was successfully transferred.
> > > +		case 0:	/* no error: continue, reset error variables */
> > > +			disable_multi = 0;
> > > +			retries = 3;
> > > +			break;
> > > +
> > > +		/* Card has not understand command. As we do only send
> > > +		 * valid commands, this must be a transmission error. */
> > > +		case -EPROTO:	/* fall through */
> >
> > This indicates a layering problem. The host driver should not be aware
> > of anything but pure bit errors.
> >
> > Also, this special meaning should be documented in core.h should we
> > decide to keep it.
> 
> This is not MY error code. EPROTO is send from mmc_spi_writeblock(), and I 
> have listed it here because it is a transmission error.
> 
> So there seems to be contrary objections: Matt Flemming has requested that the 
> exact cause of error is reported by the driver (because otherwise the caller 
> will loose information), and you requested to distinguish only between 
> transmission errors and the rest.
> 
> Matt Flemming has pointed out that further changes in the code will request to 
> get the exact cause of error in the block layer, and that error codes might 
> have interpreted different according to the command class, so IMHO it is 
> better to transport the exact error cause into block.c and do the error 
> handling here, according to the type of request.
> 

When I said "report the correct error" I was objecting to using EILSEQ
as The One True Error Code based on the fact that a transmission error
may, or may not, have occurred.

That is not contrary to Pierre's comments. I was not advocating
layering violations ;-)