On Tue, Feb 20, 2024 at 01:01:36AM -0800, Christoph Hellwig wrote:
> On Mon, Feb 19, 2024 at 06:06:06PM +0100, Maxime Coquelin wrote:
> > Since commit 295525e29a5b ("virtio_net: merge dma
> > operations when filling mergeable buffers"), VDUSE device
> > require support for DMA's .sync_single_for_cpu() operation
> > as the memory is non-coherent between the device and CPU
> > because of the use of a bounce buffer.
> >
> > This patch implements both .sync_single_for_cpu() and
> > sync_single_for_device() callbacks, and also skip bounce
> > buffer copies during DMA map and unmap operations if the
> > DMA_ATTR_SKIP_CPU_SYNC attribute is set to avoid extra
> > copies of the same buffer.
>
> vduse really needs to get out of implementing fake DMA operations for
> something that is not DMA.
In a sense ... but on the other hand, the "fake DMA" metaphor seems to
work surprisingly well, like in this instance - internal bounce buffer
looks a bit like non-coherent DMA. A way to make this all prettier
would I guess be to actually wrap all of DMA with virtio wrappers which
would all go if () dma_... else vduse_...; or something to this end. A
lot of work for sure, and is it really worth it? if the only crazy
driver is vduse I'd maybe rather keep the crazy hacks local there ...
--
MST
