Pause & Resume

[YoSTEALTH]

Are there any plans to add Pause and Resume features?

[axboe]

No plans to add that, because I don't know what that would refer to. What is being paused and resumed?

[YoSTEALTH]

Hmm... lets see, in co-routine environment you can pause/suspend an task and resume it later. In terms for io_uring hmm... It wouldn't make much sense to pause sqe task since its normally one-shot io task, and it either does that task or not (error, cancel, ...). Maybe in terms of something like file/socket fd you can open -> pause this way it can not read, write, close that fd till you resume.

Lets say you are running a http server and it can only handle 100 jobs before CPU is maxed... You have received 200 connections... What you could do is accept and pause rest of the tasks till resource is freed and resume. Most cases it could be an "bad actor", now you just pause their connections only, freeing system resources and than you can deal with that problem.

Another way to think about it would to move paused task into ttl this way its "paused" till timer runs out and it "resumes". Note: This can already be done using io_uring io_uring_prep_timeout*

I noticed there was io_uring_sqe.ioprio I am not really sure what its for. I am just assuming you have some kind of plan to prioritize one sqe over other?

[YoSTEALTH]

This is very easy to do in Python, you just remove the co-routine object from being used in event-manager, thus giving it that suspended effect, later you inject is back to resume. I am not sure how all this would work in C world.

[axboe]

I see - yes I don't think this fits with the OS/C model at all, as you can't really pause access to an fd. I suspect this is a higher level primitive that needs to be applied as such, not something io_uring can directly sppport.

ioprio is passed down the stack and already works for things like storage, it's not meant to be used in io_uring core.

[YoSTEALTH]

I see - yes I don't think this fits with the OS/C model at all, as you can't really pause access to an fd.

Ok, dealing with fd might have been too much of a stress.

ioprio is passed down the stack and already works for things like storage, it's not meant to be used in io_uring core.

I see.

I suspect this is a higher level primitive that needs to be applied as such, not something io_uring can directly support.

Its not something I would bug you to implement since its easy to do this in higher languages.

That being said I don't think its impossible either. Lets think about it... now that fd method is no go, how about dealing with sqe/cqe

---

Lets say you submit 5 entries:

sqe->user_data = [1, 2, 3, 4, 5]

You want to pause 4 but sqe is already been submitted.

The task is either not processed (still in sqe) or finished (already in cqe).

In io_uring you would remove 4 either from sqe or cqe and put it in temp holder.

Later you resume 4, now io_uring take 4 from temp holder and puts it back into sqe queue or returns the already finished cqe.

[axboe]

My point is that since it's easy to do elsewhere, adding extra code and tracking for this in io_uring would then just a cost that everyone else would have to pay. Hence I don't think that's a worthy tradeoff. The features that belong in io_uring are things you can't easily do higher up.