Re: proposal for bus_dma(9) change

[David Brownlee <abs%NetBSD.org@localhost>]

On Wed, 4 Mar 2009, Izumi Tsutsui wrote:

> Christoph_Egger%gmx.de@localhost wrote:
>
> > > - Is there any benefit to change all bus_dma(9) implementation
> > >   instead of all drivers?  I'm afraid the former is more complicated
> > >   because there are many hidden indirect calls via function pointers
> > >   in bus_dma structures/macro among machine/bus dependent implementation.
>
> > It shouldn't be hard to change bus_dmamap_create() implementation
> > to return a NULL dmamap pointer in error case.
>
> How can you find all MD bus_dmamap_create() implementation in whole
> MI/MD source tree?

        I would suspect by a similar process of finding the calls in
        the source tree :)

        There are 698 references to bus_dmamap_create through the
        tree, 29 are defines, 486 are of the form 'if (bus_dmamap_create'
        or 'foo = bus_dmamap_create', and a fair chunk of the rest are
        in printfs, comments, or other misc areas.

        I'd estimate there would be fewer references to check in the
        implementations than the callers.

> Is it really easy for you than fixing all (but only problematic) drivers
> with grep(1)?

        It involves checking every caller, or every implementation. Once
        done we will tend to have fewer new bus_dmamap_create
        implementations than callers added...

> Changing API is really pain than fixing bugs, IMO.

        I can understand a reluctance to change bus_dma(9) if it made
        it more complicated, or added extra overhead in a common path,
        but this would appear to only change behaviour in the error case,
        and does appear to be a genuine (small) enhancement to the API.

        Making APIs easier and less error prone to use, particularly
        in the uncommon case of error handling should be something
        we should be encouraging.

        If this API was being written from scratch now, and both options
        were open - the easier way to dermine an error case or not, which
        would we pick?

--
                David/absolute       -- www.NetBSD.org: No hype required --