Use rb_io_wait function and cache io instance.

[ioquatix]

https://github.com/github/activerecord-trilogy-adapter/issues/28

[matthewd]

Thanks for following up on this @ioquatix!

[matthewd]

Unnecessary?

[ioquatix]

Mark moveable doesn't check if the field is NULL, IIRC without this it can crash.

[matthewd]

I was looking at https://github.com/ruby/ruby/blob/1fdaa0666086529b3aae2d509a2e71c4247c3a12/gc.c#L7108 and https://github.com/ruby/ruby/blob/1fdaa0666086529b3aae2d509a2e71c4247c3a12/gc.c#L10395 -- I'm not 100% sure it's part of their defined contract, but I imagine it must be true by ossification, even if not originally intended?

Intuitively I just prefer "this slot contains a a possible GC ref", rather than an [even trivial] encoding of "we might need to mark this". 🤷🏻‍♂️

[ioquatix]

I agree it's possibly not the greatest interface.

[matthewd]

I acknowledge it's what's already happening on sufficiently-new rubies (along with the IO construction), but rb_fiber_scheduler_make_timeout sounds like it might allocate. Is that true? Something we can/should avoid?

[ioquatix]

There is a small chance I the future it might allocate but for the moment it doesn't, it just embeds a float into the VALUE. I hope at some point in the future to migrate to fixed point timeouts.

[matthewd]

Bonus question: is this going to affect our min ruby version (which I note we don't seem to currently declare)? I'm guessing this is moderately recent.

[ioquatix]

No I'll add feature detection, this PR is just what I could manage in the 30 minutes I had before bed time.

[eregon]

it just embeds a float into the VALUE

Flonums only work for a subset of all double, if it doesn't fit it will be an allocation.

[casperisfine]

I think it also always allocate on 32bits?

[ioquatix]

@matthewd can you please approve and run the workflows.

[ioquatix]

There is one more piece we should consider exposing, which is the "resolve DNS" to use the fiber scheduler hook.

[ioquatix]

So, writing this PR has made me realise we need a set of shims for the blocking C APIs, e.g. open, addrinfo, etc. We have the code, it's just not exposed. I supposed we can do that for Ruby 3.3.

We should also extend the test suite to run within a fiber scheduler (probably Async).

[matthewd]

@ioquatix is this good to go? It seems fine to me, not sure if you've kept it in draft for a reason, or just not hit the button?

[ioquatix]

It's probably as good as we can make it.. but I want to review it this weekend. This week has been pretty busy, have not had much time for open source stuff.

[casperisfine]

trilogy_data_type is marked with RUBY_TYPED_WB_PROTECTED, so you have to trigger a write barrier here.

Problem is you don't have access to the VALUE here.

[ioquatix]

Okay, so we should pass around the VALUE then?

[casperisfine]

Either that, or create the IO eagerly in a place where you have access to it.

[kirs]

👋 just wanted to say that I appreciate this WIP PR by Samuel and I'd really really love to see this happen in the upstream, and to be able to use trilogy with fiber scheduler for IO-heavy workloads.

[casperisfine]

I'd really really love to see this happen in the upstream, and to be able to use trilogy with fiber scheduler for IO-heavy workloads.

So my knowledge of the fiber scheduler is limited, but AFAIK you can already use it today? My understanding is that this PR make it a bit more efficient, but rb_wait_for_single_fd is already enough to be usable.

As for it happening, there's not much work needed, all the blockers are listed in review comments.

[ioquatix]

I've rebased and updated the PR.

rb_wait_for_single_fd is okay but it's a slow path. rb_io_wait is the fast path.

@matthewd can you please run CI.

[ioquatix]

I recall why I gave up on this PR for a while, we were missing rb_io_open_descriptor which is needed to create a "temporary non-owning FMODE_EXTERNAL IO instance". Since that's now available in recent CRuby, it is possible to implement it correctly.