[RFC 2/4] dma: add dmaengine driver for Samsung s3c24xx SoCs

[Tomasz Figa]

On Thursday 16 of May 2013 00:45:03 Heiko Stübner wrote:
> Am Donnerstag, 16. Mai 2013, 00:02:40 schrieb Tomasz Figa:
> > On Wednesday 15 of May 2013 23:48:31 Heiko Stübner wrote:
> > > Am Mittwoch, 15. Mai 2013, 23:20:08 schrieb Sylwester Nawrocki:
> > > > On 05/15/2013 10:31 PM, Heiko Stübner wrote:
> > > > >>> +       BUG();
> > > > >>> 
> > > > >> >  Isn't that a bit nasty. This macro should be used with care
> > > > >> >  and
> > > > >> >  we
> > > > >> >  should recover if possible. dev_err()?
> > > > > 
> > > > > runtime_config already denies any settings not in the 1,2 or
> > > > > 4bytes
> > > > > range - the default-part should therefore never be reached. So
> > > > > if
> > > > > any other value magically appears in the register and triggers
> > > > > the
> > > > > default-part, something is seriously wrong. So my guess is, the
> > > > > BUG
> > > > > might be appropriate.
> > > > > 
> > > > > On the other hand the whole default+BUG part could also simply
> > > > > go
> > > > > away,
> > > > > for the same reasons.
> > > > 
> > > > IMHO BUG() is not needed at all. As Linus suggested dev_err() is
> > > > such
> > > > case or WARN_ON() would be more appropriate. This has been
> > > > discussed
> > > > in the past extensively, not sure if you are aware of the other
> > > > Linus' opinion on BUG()/BUG_ON() proliferation:
> > > > https://lkml.org/lkml/2012/9/27/461
> > > 
> > > Very interesting read and I'll keep this in mind in the future. What
> > > about the other option ... i.e. simply getting rid of the whole
> > > "error
> > > handling", as the other code paths should already make sure that
> > > only
> > > valid values get written into the register.
> > > 
> > > Can the value change in the register somehow on its own without
> > > kernel
> > > intervention, or does this not happen?
> > 
> > Hmm, it depends on hardware, I guess. Not sure how it works on this
> > particular IP.
> > 
> > Still, the mentioned BUG() was about a value in a driver-filled
> > struct,
> > wasn't it?
> > 
> > /* Quoting the the code for reference */
> > 
> > > +static u32 s3c24xx_dma_getbytes_chan(struct s3c24xx_dma_chan
> > > *s3cchan)
> > > +{
> > > +       struct s3c24xx_dma_phy *phy = s3cchan->phy;
> > > +       struct s3c24xx_txd *txd = s3cchan->at;
> > > +       u32 tc = readl(phy->base + DSTAT) & DSTAT_CURRTC_MASK;
> > > +
> > > +       switch (txd->dcon & DCON_DSZ_MASK) {
> > > +       case DCON_DSZ_BYTE:
> > > +               return tc;
> > > +       case DCON_DSZ_HALFWORD:
> > > +               return tc * 2;
> > > +       case DCON_DSZ_WORD:
> > > +               return tc * 4;
> > > +       default:
> > > +               break;
> > > +       }
> > > +
> > > +       BUG();
> > 
> > (Btw. I don't see anything setting the DCON_DSZ bits in this field. Am
> > I missing something?)
> 
> this is for calculating the remaining bytes of the transaction. which is
> used in s3c24xx_dma_tx_status.
> 
> And when looking at it again, I can't really fathom why I did it this
> way with decoding the DSZ from the dcon value of the s3c24xx_txd again
> instead of simply using the width value of the same struct ....
> 
> So it can be much simpler as
> 	(...)
>      u32 tc = readl(phy->base + DSTAT) & DSTAT_CURRTC_MASK;
> 	return tc * txd->width;
> 
> getting rid of this stuff alltogether
> 
> 
> still puzzled how I came up with this strangeness in the first place

Hehe, happens.

I'm still yet to review the whole series, but I'm failing to find enough 
time. Hopefully will get to do it soon.

Best regards,
Tomasz