On Thu, 2019-01-17 at 10:30 +0100, hch@xxxxxx wrote: > On Wed, Jan 16, 2019 at 10:24:36AM -0700, Jason Gunthorpe wrote: > > The fact is there is 0 industry interest in using RDMA on platforms > > that can't do HW DMA cache coherency - the kernel syscalls required > > to > > do the cache flushing on the IO path would just destroy performance > > to > > the point of making RDMA pointless. Better to use netdev on those > > platforms. > > In general there is no syscall required for doing cache flushing, you > just issue the proper instructions directly from userspace. But what if there are other coherence issues? Like bounce-buffers? I'd like to +1 on what Jason says about industry interest: FWIW, vmwgfx is probably one of the graphics drivers that would lend itself best to do a fully-dma-interface compliant graphics stack experiment. But being a paravirtual driver, all platforms we can ever run on are fully coherent unless someone introduces a fake incoherency by forcing swiotlb. Putting many man-months of effort into supporting systems on which we would never run on and can never test on can never make more than academic sense. > > > > The reality is that *all* the subsytems doing DMA kernel bypass are > > ignoring the DMA mapping rules, I think we should support this > > better, > > and just accept that user space DMA will not be using syncing. > > Block > > access in cases when this is required, otherwise let it work as is > > today. > > In that case we just need to block userspace DMA access entirely. > Which given the amount of problems it creates sounds like a pretty > good idea anyway. I'm not sure I'm following you here. Are you suggesting scratching support for all (GP)GPU- and RDMA drivers? Thanks, Thomas
