Re: [RFC/PATCH] dma: dma_{un}map_{single|sg}_attrs() interface

[Roland Dreier <rdreier@cisco.com>]

 > --- a/include/linux/dma-attrs.h
 > +++ b/include/linux/dma-attrs.h
 > @@ -0,0 +1,33 @@
 > +#ifndef _DMA_ATTR_H
 > +#define _DMA_ATTR_H
 > +
 > +#include <linux/bug.h>
 > +
 > +enum {
 > +	DMA_ATTR_INVALID,
 > +	DMA_ATTR_BARRIER,
 > +	DMA_ATTR_FOO,
 > +	DMA_ATTR_GOO,
 > +	DMA_ATTR_MAX,
 > +};
 > +
 > +struct dma_attrs {
 > +	unsigned flags;
 > +};
 > +
 > +static inline int dma_set_attr(struct dma_attrs *attrs, unsigned attr) {

maybe this would be cleaner if you named the DMA_ATTR enum and used
that instead of unsigned here (and below)?

 > +	BUG_ON(attrs == NULL);

does this BUG_ON() buy us much?  It seems the only thing we would fail
to oops on is if someone did dma_set_attr(NULL, INVALID) and I'm not
sure it's worth it to BUG here.

 > +	if (attr > DMA_ATTR_INVALID && attr < DMA_ATTR_MAX) {
 > +		attrs->flags = (1 << attr);
 > +		return 0;
 > +	}
 > +	return 1;

returning -EINVAL here instead of 1 would probably be more "kernelish".

 > +}
 > +
 > +static inline int dma_get_attr(struct dma_attrs *attrs, unsigned attr) {
 > +	if (attrs) 
 > +		 return attrs->flags & (1 << attr);

so it's OK to pass attrs == NULL into dma_get_attr() but not into
dma_set_attr()?  seems kind of odd.

 > +	return 0;
 > +}

It seems you're missing a way to initialize a struct dma_attrs.  How
do I clear the flags field to start with?

A macro like DEFINE_DMA_ATTRS() that initializes things for you (like
LIST_HEAD or DEFINE_SPIN_LOCK) would probably be a good thing to have
as well.

Also I guess you could test ARCH_USES_DMA_ATTRS in this file and stub
everything out and define an empty structure if it's not defined.
save a few bytes of stack etc.

 > +
 > +#endif /* _DMA_ATTR_H */