Hi,
On Thu, Dec 15, 2011 at 06:16:39PM -0500, Alan Stern wrote:
> On Thu, 15 Dec 2011, Felipe Balbi wrote:
>
> > > > + if (req->dma == DMA_ADDR_INVALID) {
> > >
> > > What's the purpose of this test? Just to catch UDC drivers that try to
> > > map a request twice in a row? If that's the case, then why not provide
> > > an error message when it happens?
> >
> > The thing is that there has never been an agreement of who will do the
> > final mapping on the Gadget Framework. Look at all UDCs and you'll see
> > that they all have the same checks.
> >
> > We can add such a requirement now, but I'm afraid we could have
> > out-of-tree gadget-drivers relying on dma_alloc_coherent() which would
> > already give a request with a valid dma address. Or gadget drivers
> > actually deciding to call dma_map_*() APIs.
>
> Okay, I see.
>
> > If more people would be willing to accept that possibility, I'm very
> > eager to change the patch so that it would assume we have an unmapped
> > request always. What do you say ?
>
> I'm not too concerned about breaking out-of-tree drivers. After all,
> the kernel doesn't have a stable internal ABI. Have you checked the
> in-tree drivers? I haven't.
>
> You could start off easy by leaving in the test and doing WARN_ONCE
> if the DMA address is already set. Then if nothing shows up in
> kerneloops, take it out entirely.
aaaa haven't thought about that :-) Sounds like a great plan, will do
:-)
will check in tree drivers anyway :-D
--
balbi
Attachment:
signature.asc
Description: Digital signature
