On 6/28/22 1:09 PM, Stefan Metzmacher wrote: > Hi Jens, > > I'm wondering what happened to you patch passing file_ops->uring_cmd() > down to socket layers. > > It was part of you work in progress branches... > > The latest one I found was this: > https://git.kernel.dk/cgit/linux-block/commit/?h=nvme-passthru-wip.2&id=28b71b85831f5dd303acae12cfdc89e5aaae442b > > And this one just having the generic parts were in a separate commit > https://git.kernel.dk/cgit/linux-block/commit/?h=io_uring-fops.v7&id=c2ba3bd8940ef0b7d1c09adf4bed01acc8171407 > vs. > https://git.kernel.dk/cgit/linux-block/commit/?h=io_uring-fops.v7&id=542c38da58841097f97f710d1f05055c2f1039f0 > > I took this: > https://git.kernel.dk/cgit/linux-block/commit/?h=io_uring-fops.v7&id=c2ba3bd8940ef0b7d1c09adf4bed01acc8171407 > adapted it on top of v5.19-rc4 and removed stuff that was not really needed. > > Even if it's not used in tree, it would be good to have uring_cmd hooks in > struct proto_ops and struct proto, so that out of tree socket implementations > like my smbdirect driver are able to hook into it. > > What do you think? We need to just finalize a format for this, don't think it'll be too complicated. But for in-kernel users first and foremost, not for some out of tree code! I've got various things I'd like to use it for internally too. For example, returning number of bytes left in a socket post receive rather than just a flag telling you there's more data. -- Jens Axboe
