On 11/6/23 22:34, Stanislav Fomichev wrote:
On 11/06, Willem de Bruijn wrote:
IMHO, we need a better UAPI to receive the tokens and give them back to the kernel. CMSG + setsockopt(SO_DEVMEM_DONTNEED) get the job done, but look dated and hacky :-(
We should either do some kind of user/kernel shared memory queue to receive/return the tokens (similar to what Jonathan was doing in his proposal?)
Oops, missed the discussion. IMHO shared rings are more elegant here. With that the app -> kernel buffer return path doesn't need to setsockopt(), which will have to figure out how to return buffers to pp efficiently, and then potentially some sync on the pp allocation side. It just grabs entries from the ring in the napi context on allocation when necessary. But then you basically get the io_uring zc rx... just saying
I'll take a look at Jonathan's proposal, sorry, I'm not immediately familiar but I wanted to respond :-) But is the suggestion here to build a new kernel-user communication channel primitive for the purpose of passing the information in the devmem cmsg? IMHO that seems like an overkill. Why add 100-200 lines of code to the kernel to add something that can already be done with existing primitives? I don't see anything concretely wrong with cmsg & setsockopt approach, and if we switch to something I'd prefer to switch to an existing primitive for simplicity?
The only other existing primitive to pass data outside of the linear buffer is the MSG_ERRQUEUE that is used for zerocopy. Is that preferred? Any other suggestions or existing primitives I'm not aware of?
or bite the bullet and switch to io_uring.
IMO io_uring & socket support are orthogonal, and one doesn't preclude the other.
They don't preclude each other, but I wouldn't say they're orthogonal. Similar approaches, some different details. FWIW, we'll be posting a next iteration on top of the pp providers patches soon.
As you know we like to use sockets and I believe there are issues with io_uring adoption at Google that I'm not familiar with (and could be wrong). I'm interested in exploring io_uring support as a follow up but I think David Wei will be interested in io_uring support as well anyway.
Well, not exactly support of devmem, but true, we definitely want to have io_uring zerocopy, considering all the api differences. (at the same time not duplicating net bits).
I also disagree that we need to replace a standard socket interface with something "faster", in quotes.
This interface is not the bottleneck to the target workload.
Replacing the synchronous sockets interface with something more performant for workloads where it is, is an orthogonal challenge. However we do that, I think that traditional sockets should continue to be supported.
The feature may already even work with io_uring, as both recvmsg with cmsg and setsockopt have io_uring support now.
It should, in theory, but the api wouldn't suit io_uring, internals wouldn't be properly optimised, and we can't use it with some important features like multishot recv because of cmsg.
I'm not really concerned with faster. I would prefer something cleaner :-)
Or maybe we should just have it documented. With some kind of path towards beautiful world where we can create dynamic queues..