On Mon, 17 Apr 2023 18:19:50 -0700, Jakub Kicinski <kuba at kernel.org>
wrote:> On Tue, 18 Apr 2023 09:07:30 +0800 Jason Wang wrote:
> > > > Would you mind explaining this a bit more to folks like me
who are not
> > > > familiar with VirtIO? DMA API is supposed to hide the DMA
mapping
> > > > details from the stack, why is it not sufficient here.
> >
> > The reason is that legacy virtio device don't use
DMA(vring_use_dma_api()).
> >
> > The AF_XDP assumes DMA for netdev doesn't work in this case. We
need a
> > way to make it work.
>
> Can we not push this down to be bus level? virtio has its own bus it
> can plug in whatever magic it wants into dma ops.
It is actually not possible.
[1] https://lore.kernel.org/virtualization/ZDUCDeYLqAwQVJe7 at infradead.org/
>
> Doesn't have to be super fast for af_xdp's sake - for af_xdp dma
mapping
> is on the control path. You can keep using the if (vring_use_dma_api())
> elsewhere for now if there is a perf concern.
Sorry, I don't particularly understand this passage.
Now, the question is if vring_use_dma_api() is false, then we cannot use DMA
API in AF_XDP.
The good news is that except for some of sync's operations, they are in the
control path. I think it is very small effect on performance. Because in most
case the sync is unnecessary.
>
> Otherwise it really seems like we're bubbling up a virtio hack into
> generic code :(
Can we understand the purpose of this matter to back the DMA operation to the
driver? Although I don't know if there are other drivers with similar
requirements.
Thanks.