Iofthetiger

[FObersteiner]

@dude_the_builder

The Go (and I belive Erlang / Elixir) model

to be fair, they all use a runtime (IIRC)

[dee0xeed]

@dude_the_builder

As I see it, Zig’s first attempt at async IO was via the async / await model, albeit implemented in a a novel way using the function’s frame directly with suspend and resume points. It seems although it was a cool and novel idea

Hmm… novel? Coroutines (subroutines with multiple entry points) were invented in 1958…

@dude_the_builder

it clashes in critical ways with other aspects of the compiler which have higher priority in terms of achieving Zig’s goals

As I see it… If a compiler when generating code for async/await based (async) I/O has to use specific system calls (which are different across OSes)… well, this is awful. This is ok for languages like python/tcl, but not for a language like C/Zig. Just let people use OS directly and they will someday come up with such an abstraction, but do not put this abstraction into compiler proper.

[dude_the_builder]

True, but when you think about it, all these async / await implementations need an event loop to drive them and many times a worker thread pool to improve performance even more. So in the end, they end up having a runtime too. Take Rust for example; the language itself doesn’t have a runtime but does provide the async / await semantics (as syntax sugar over Futures). So you must add a runtime like Tokio or async-std to make it work.

[Sze]

@dude_the_builder

As I see it, Zig’s first attempt at async IO was via the async / await model, albeit implemented in a a novel way using the function’s frame directly with suspend and resume points. It seems although it was a cool and novel idea, it clashes in critical ways with other aspects of the compiler which have higher priority in terms of achieving Zig’s goals. So it was sacrificed and removed.

I think we should add “for now” here:

[#36] Zig Roadmap 2024 - Andrew Kelley

Question: Can we have a couple of words about async? timestamp: https://youtu.be/5eL_LcxwwHg?t=1842 Paraphrased: Andrew explains that it currently isn’t practical, because there are a lot of difficult problems, that are in the way of a good implementation of async, these other things need to be addressed first. Summary And I tried to transcribe the summary in the video: timestamp of async/await summary Ultimately, I think it can be recovered, I think we can have async/await. But it is just…

[gonzo]

Thanks for the detailed explanation. I know this song, it was also very popular where I come from (they call it “the Rocky song” over there). But iofthetiger blocks my brain into interpreting the first two characters as I/O (as you mentioned), and I am left with fthetiger, which I can only read as… <expletive> the tiger. LOL.

[slonik-az]

@gonzo

and I am left with fthetiger, which I can only read as… <expletive> the tiger. LOL.

This double interpretation only makes the pun spicier and jucier

[gonzo]

Was Linus foreshadowing io_uring in those thoughts, back in 2016?

[dude_the_builder]

@gonzo

fthetiger

ROTFL! Definitely a whole new meaning to the name. And hey, I call it the Rocky song too.

[dude_the_builder]

@gonzo

blocks my brain

This must be a bug. It’s supposed to be non-blocking.

[dee0xeed]

@gonzo

Was Linus foreshadowing io_uring in those thoughts, back in 2016?

I guess he was, look at this phrase.

If you want to do arbitrary asynchronous system calls…

There was/is libaio, which Linus did not like (charitably speaking).
Also there was a citation in man 2 open… it seems it has been removed from this manual page, but in older Linux distros (deb8 for instance) it’s still there (it’s about O_DIRECT flag).

As far as I could understand in that message Linus is saying that there must be
a single general mechanism for doing things asynchronously, but not scrappy “solutions” for every special case (libaio works only with file i/o).

[dee0xeed]

@AndrewCodeDev

There was a famous song called “eye of the tiger” by the band Survivor

The song is dull, no drive at all in it.
How about rock’n’roll dev ?

ReadOnly DevOps, why ain’t should be?

[dude_the_builder]

Hey man, Survivor was good, but you’re bringing in the heavy guns with AC/DC!

P.S.: Now we’re going off-topic but hey, it’s AC/DC, so it’s allowed. lol

[dee0xeed]

@dude_the_builder

Now we’re going off-topic

We are not:

Ain't No Fun (Waiting Round to Be a Millionaire)

[mperillo]

@dude_the_builder

…
Take Rust for example; the language itself doesn’t have a runtime but does provide the async / await semantics (as syntax sugar over Futures). So you must add a runtime like Tokio or async-std to make it work.

Starting from go1.14, goroutines are asynchronously preemptible in order to avoid, as example, loops without function to deadlock the scheduler.

Does languages supporting async/await like Rust support this feature?

[Anthon]

Thanks for pointing this out. I had only read the name in the overview of the posts and thought it was some strange german name starting with an ‘L’ ( lofthetiger ) that I hadn’t seen before.

[dude_the_builder]

That would depend on the async runtime you choose. I’m not sure if Tokio or async-std have this feature, but they could implement it at the runtime level without needing a change in the language.

[dude_the_builder]

Yeah, Discourse automatically capitalizes topic titles so it turned iofthetiger into Iofthetiger.

[dee0xeed]

@dude_the_builder

Yes, they implemented a uniform callback-based async IO abstraction layer over each operating system’s specific async IO interface

Sorry for bumping/off-topic/etc. I am just wondering how do timers and signals fit into io_uring mechanism/abstraction?

Both timers and signals are not I/O per se, no data transfer between peripherals of a “computer” and CPU, no data transfer between nodes in a network, it’s just some interaction
between user/kernel space inside a single “device”/“gadget”.

In Linux, after you got readiness notification concerning timers/signals you must read some data from underlying fds, otherwise you will be over-flooded with these notifications.

In FreeBSD, timers are different beasts, someone may like it, someone do not, I personally don’t - “everything in Unix is a file”, after all

So, both with FreeBSD kqueue and with LInux epoll, both timers and signals more or less fit into “reactor pattern”. How do they fit into (really cool) io_uring facility?

[dude_the_builder]

io_uring is really totally centered around IO, be it network sockets or files. It does have facilities for settings timeouts and the submission / completion queue entries could be used as a signaling mechanism. But the main focus is just IO. I don’t know much about epoll but I believe it’s a much more generalized async system that can handle other things besides just IO, no?

[dee0xeed]

@dude_the_builder

I don’t know much about epoll but I believe it’s a much more generalized async system that can handle other things besides just IO, no?

epoll is not “async” (that’s why io_uring appeared), it’s main purpose is to deliver readiness (operation won’t “block”, a process won’t go into sleep state) notifications, not operation complete notifications.

epoll can easily be used for timers, created with timerfd_create and with signals, created with signalfd, epoll_wait() returns when a timer expired or signal arrived. File system events fits too.

epoll is not suited for file I/O, it would just flood an application with read/write readiness notifications, io_uring is.

But what if want to use all the stuff (network I/O, file I/O, timers, signals, FS events) at once, how would I deal with them using io_uring?

Of course, it is not a question to zig lang developers, it’s more a question to os kernels developers.