Re: Feature: zero-copy splice API/opcode

[Pavel Begunkov <asml.silence@gmail.com>]

[-- Attachment #1.1: Type: text/plain, Size: 1151 bytes --]

On 09/01/2020 17:58, Mark Papadakis wrote:
> Greetings,
> 
> I ‘ve been trying to replicate the benefits provided by sendfile() using e.g O_DIRECT access, together with IOSQE_IO_LINK in SQE flags and MSG_ZEROCOPY, but it doesn’t appear to work. Other ideas didn’t work either.
> 
> I would really appreciate a sendfile like SQE opcode, but maybe some sort of generic DMA/zero-copy based opcode based on splice semantics could be implemented, so that e.g a vmsplice() like alternative could also work.
> 
Sounds interesting, I'll look into this. First, I want to dig a bit into
kernel's zero-copy and get an idea how to cover all use cases with a single
opcode, and that's probably keeping in mind p2p.

The obvious nuisance from  perspective is that it needs 2 fds. Any other concerns?

> 
> (That would be the last remaining bit of functionality missing from io_uring, now that Jens has implemented support for IOSQE_ASYNC, IORING_REGISTER_EVENTFD_ASYNC and for managing epoll FDs, for enabling support for io_uring on https://github.com/phaistos-networks/TANK ).
> 
> Thank you,
> @markpapadakis
> 

-- 
Pavel Begunkov


[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]