On Wed, 19 May 2010 15:58:37 +0100 David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote: > On Wed, 2010-05-19 at 23:38 +0900, FUJITA Tomonori wrote: > > > > > I don't think we need to hide the fact that some platforms have > > > specific alignment restrictions for DMA. So if any drivers make use > > > of the alignment, I see no problem with __dma_aligned. > > > > IIRC, such was proposed several times: > > > > http://www.mail-archive.com/linux-scsi@xxxxxxxxxxxxxxx/msg12633.html > > > > I guess that we agreed that it's better to tell driver writers to just > > use kmalloc. > > Perhaps -- but only a few days ago in this thread, they were being > advised to use ____cacheline_aligned instead! Hmm, driver writers should read DMA-API and DMA-API-HOWTO docs. > And for this case it really does seem to make sense to keep the buffer > in the parent structure rather than allocating it separately. The DMA > buffers are tiny and on cache-coherent architectures it's _much_ more > efficient just to have them in the original structure and use > __dma_aligned. Yeah, I think that that's a valid point (which was also discussed in the past). However, I tend to agree on: http://lwn.net/Articles/2270/ I think that forcing kmalloc is not that bad. Surely more code, but the performance is not notable in most cases. IMHO, exporting cache line thing everywhere is worse than it. -- To unsubscribe from this list: send the line "unsubscribe linux-input" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
