Nim 3.0 and asynchronous programming with Passive Procs

[anderflash]

Hi,

It's not an issue, but a discussion (I think you may want enabling the "Discussion" tab in order to place those type of threads.

I see Araq is developing a new compiler "Nimony" for Nim 3.0, and one of the new features is a way to unify asynchronous programming techniques. Its documentation (in Passive Procs) briefly refers to io_uring along with poll and other APIs wrapped into some "ioRing".

How do you see this feature being incorporated into the language? If so, are you still working on this package?

[blackmius]

why asking it here? it is more a question about nimony and passive procs nor the library.

and one of the new features is a way to unify asynchronous programming techniques. Its documentation (in Passive Procs) briefly refers to io_uring along with poll and other APIs wrapped into some "ioRing".

as i understand unify means that passive procs can be called within synchronous code and avoid "function color" problem. combining poll and io_uring is left for a scheduler (event loop).

how to write a scheduler? i like the way python did that: you have a global eventloop object and all system modules uses it. then you can replace eventloop with your own that just need to satisfy a specific interface. like having write(fd, buf) commands and so on. and you can create different types of event loops based on poll, kqueue, io_uring, iocp and so on

[blackmius]

why does it sound so easy, but in reality nim2.0 makes it hard to support different types of event loops? The main reason is that the event loop abstraction is broken in both std/asyncloop and chronos. they only abstract epoll calls from the user, while the remaining syscalls are still left to the user.
for example, consider how async write is done in nim: you call asyncloop to await until a file descriptor is ready, and then manually call write on it. as a result, all the code depends on this pattern, and you cant simply swap out event loops. you end up with a mess of manual syscalls, and the code is not reusable

[blackmius]

another good example is libuv: it also replaces syscalls with its own functions for every operation, which made supporting io_uring easy for them