On Wed, 19 Apr 2023 23:19:22 -0700, Christoph Hellwig <hch@xxxxxxxxxxxxx> wrote: > On Wed, Apr 19, 2023 at 09:45:06AM -0700, Jakub Kicinski wrote: > > > Can you explain what the actual use case is? > > > > > > From the original patchset I suspect it is dma mapping something very > > > long term and then maybe doing syncs on it as needed? > > > > In this case yes, pinned user memory, it gets sliced up into MTU sized > > chunks, fed into an Rx queue of a device, and user can see packets > > without any copies. > > How long is the life time of these mappings? Because dma_map_* > assumes a temporary mapping and not one that is pinned bascically > forever. > > > Quite similar use case #2 is upcoming io_uring / "direct placement" > > patches (former from Meta, latter for Google) which will try to receive > > just the TCP data into pinned user memory. > > I don't think we can just long term pin user memory here. E.g. for > confidential computing cases we can't even ever do DMA straight to > userspace. I had that conversation with Meta's block folks who > want to do something similar with io_uring and the only option is an > an allocator for memory that is known DMAable, e.g. through dma-bufs. > > You guys really all need to get together and come up with a scheme > that actually works instead of piling these hacks over hacks. I think that cases Jakub mentioned are new requirements. From the perspective of implementation, compared to this patch I submitted, if the DMA API can be expanded, compatible with some special hardware, I think it is a good solution. I know that the current design of DMA API only supports some physical hardware, but can it be modified or expanded? Still the previous idea, can we add a new ops (not dma_ops) in device? After the driver configuration, so that the DMA API can be compatible with some special hardware? Thanks.
