[RFC PATCH 06/11] dma: amba-pl08x: Keep LLIs aligned to 4-word boundary

[Tomasz Figa]

On Monday 17 of June 2013 15:51:20 Linus Walleij wrote:
> On Sun, Jun 16, 2013 at 10:54 PM, Tomasz Figa <tomasz.figa at gmail.com> 
wrote:
> > PL080 reference manual states that to LLI entries should be aligned
> > to 4-word boundary to make LLI fetches more efficient. This patch adds
> > a 3-word padding to the LLi struct to make this condition true.
> > 
> > Signed-off-by: Tomasz Figa <tomasz.figa at gmail.com>
> > ---
> > 
> >  drivers/dma/amba-pl08x.c | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/drivers/dma/amba-pl08x.c b/drivers/dma/amba-pl08x.c
> > index eb10eb8..0da5539 100644
> > --- a/drivers/dma/amba-pl08x.c
> > +++ b/drivers/dma/amba-pl08x.c
> > @@ -127,6 +127,7 @@ struct pl08x_lli {
> > 
> >         u32 lli;
> >         u32 cctl;
> >         u32 cctl1;
> > 
> > +       u32 dummy[3];
> 
> Atleast put a comment into the code explaining what this is all
> about. Or someone will add another member to the struct and
> all is lost. Call it "padding" rather than dummy.
> 
> >  };
> 
> So it used to be like this before you added cctl1:
> 
> struct pl08x_lli {
>         u32 src;
>         u32 dst;
>         u32 lli;
>         u32 cctl;
> };
> 
> Meaning it was 3 words.
> 
> And now you make it take 8 words for everyone.
> 
> Atleast this patch should be squashed into the patch
> adding cctl1.
> 
> But I really don't like this fragile way of casting structs right
> into memory, and I don't like that teh other PL080's also have
> to waste 8 words when their LLIs fit so nicely into 4.
> 
> I would have solved this problem by creating a
> marshalling function that just allocate the number of bytes the
> LLI entry shall have and fill it in by assigning directly to the
> precise target memory cell. This way the LLIs will take
> 4 words on the original variants and you can use some
> nice logic to pad out to 8 words on the PL080S variant.

Definitely a valid point. I'll see what I can do about it.

I was thinking about it originally, but I couldn't find any really good 
solution for this so I just went with this extremely simple approach as a 
proof of concept and to show the problem. :)

Best regards,
Tomasz