Re: [PATCH v4 01/44] dma-mapping: Use unsigned long for dma_attrs

From: Krzysztof Kozlowski
Date: Fri Jun 10 2016 - 16:44:30 EST


On Fri, Jun 10, 2016 at 10:23:47PM +0200, Luis R. Rodriguez wrote:
> On Fri, Jun 10, 2016 at 10:16:00PM +0200, Krzysztof Kozlowski wrote:
> > On Fri, Jun 10, 2016 at 04:49:47PM +0200, Luis R. Rodriguez wrote:
> > > On Fri, Jun 10, 2016 at 12:11:18PM +0200, Krzysztof Kozlowski wrote:
> > > > The dma-mapping core and the implementations do not change the
> > > > DMA attributes passed by pointer. Thus the pointer can point to const
> > > > data. However the attributes do not have to be a bitfield. Instead
> > > > unsigned long will do fine:
> > > >
> > > > 1. This is just simpler. Both in terms of reading the code and setting
> > > > attributes. Instead of initializing local attributes on the stack
> > > > and passing pointer to it to dma_set_attr(), just set the bits.
> > > >
> > > > 2. It brings safeness and checking for const correctness because the
> > > > attributes are passed by value.
> > >
> > > Do we not expect the number of argument to grow ? This "cleanup" would
> > > do away with such possibilities, and then require adding the API later,
> > > and this requiring a full set of collateral evolutions again when this
> > > is needed. What was the original motivation for using this instead of
> > > the approach you are suggesting ?
> >
> > What do you mean by "possibilities of argument to grow"? Something like
> > adding new members to "struct dma_attrs" and changing its meaning?
>
> Yup that.
>
> > I think such growth is still constrained - you cannot put there anything
> > without changing the meaning of the argument.
>
> Obviously, however it would mean no needed collateral evolutions,
> just an extension to the struct and drivers that use the new member
> can make use of it.

For parts of the API, there is still possibility of adding new layer of
wrapping, just like it was done with dma_map_single_attrs():
#define dma_map_single(d, a, s, r) dma_map_single_attrs(d, a, s, r, NULL)

For the dma_map_ops not...

> > The dma-attrs in current form were added around 2008 in 74bc7ceebfa1
> > ("dma: add dma_*map*_attrs() interfaces"), I think. Since that time, for
> > example, the dma_map_*_attrs() did not change.
>
> So we don't expect this to change either?

I do not know, I am not aware of planned changes to that.

>
> > > If the concern is the const data, why not require const struct dma_attr
> > > for the APIs that we know can and should use const ?
> >
> > The const is one concern. Complicated (more than expected) usage of dma
> > attributes by the caller is second.
> >
> > Switching it to const would also reduce the possibilities of API
> > extension.
>
> My point was that const can be used for only APIs that we are sure of
> that need it.

As of now, dma_attrs should be const everywhere. That would be almost
the same patchset as current one. If you consider extending the
dma_attrs to something new and not yet known, then how will
differentiate between cases when 'const' is needed for sure?

I understand your concern. Sticking to current API for that reason might
be a good defensive API programming... or might be way of keeping this
function prototype for long...

Best regards,
Krzysztof