On Thu, Mar 30, 2023 at 2:06 PM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote: > On Wed, Mar 29, 2023 at 09:45:45PM -0500, Bob Pearson wrote: > > On 3/29/23 18:16, Jason Gunthorpe wrote: > > > On Wed, Mar 29, 2023 at 04:28:08PM +0200, Linus Walleij wrote: > > > > > >> I'm a bit puzzled: could the above code (which exist in > > >> three instances in the driver) even work as it is? Or is it not used? > > >> Or is there some failover from DMA to something else that is constantly > > >> happening? > > > > > > The physical address dma type IB_MR_TYPE_DMA is rarely used and maybe > > > nobody ever tested it, at least in a configuration where kva != pa > > > > > > Then again, maybe I got it wrong and it is still a kva in this case > > > because it is still ultimately DMA mapped when using IB_MR_TYPE_DMA? > > > > > > Bob? > > > > > > Jason > > > > In the amd64 environment I use to dev and test there is no difference AFAIK. > > I have to go on faith that other platforms work but have very little experience. > > Honestly, I would be much happier if the virtual wrappers used > physical not kva :\ But that is too much to ask > > So I suspsect I got it wrong and it is a kva still OK I'll respin the first version of the patch but with the quirks making it work on parisc. Yours, Linus Walleij
