Add tracing spans for input and output gates hold and wait

[shrima-cf]

No description provided.

[danlapid]

Suggested change

	#include "workerd/io/io-gate.h"
	#include <workerd/io/io-gate.h>

[danlapid]

I think we should always have an active IoContext shouldn't we?
So

    auto& context = IoContext::current();
    auto inputGateHoldSpan = context.makeTraceSpan("actor_input_gate_hold"_kjc);

Should be better

[shrima-cf]

Based on my testing, we don't have IoContext when we invoke the input gate or hold them. I am trying to see if I can use the parent's context.

[danlapid]

That would really limit the effectiveness and reliability of this PR.

I don't understand input gates well enough but I would love to learn more about them to understand how to best approach this.
If we need to instead iterate over all of the actor's current inflight requests and open spans in all of them we could do that instead of relying on the IoContext

[a-robinson]

This is naively overwriting any existing span, but it's possible for an OutputGate to have multiple waiters. I'll look more closely in a minute, but I'd expect us to want to attach a span to each (traced) waiter

[a-robinson]

Yeah, I'd think we could hold these WaitSpan variables in the InputGate::Waiter for InputGates, and for OutputGates we could attach a span on the branch promise returned from OutputGate::wait (similar to the defer for hooks.outputGateWaiterRemoved())

[shrima-cf]

I tested with one request and this works, but or input gates the IoContext is not set at the point where the gates are held/waited, so have to use the parent's IoContext or pass in a parent span. I am working on that change now, will mostly switch to same model for outpute gate too

[justin-mp]

Why is this the right layer to create the spans? It certainly seems awkward to have to keep track of the spans at the hook layer.

It seems to me that a better approach might be to take SpanParent parameters to InputGate::wait() and OutputGate::lockWhile(). Each of those could create a child span that has the lifetime of the promise that they return. (This would be particularly trivial in OutputGate::lockWhile because it's a co-routine.) Moreover, it would let us trace each call into those functions as necessary.

(I think @a-robinson said more or less the same thing in #5563 (comment).)

[shrima-cf]

Why is this the right layer to create the spans? It certainly seems awkward to have to keep track of the spans at the hook layer.

It seems to me that a better approach might be to take SpanParent parameters to InputGate::wait() and OutputGate::lockWhile(). Each of those could create a child span that has the lifetime of the promise that they return. (This would be particularly trivial in OutputGate::lockWhile because it's a co-routine.) Moreover, it would let us trace each call into those functions as necessary.

(I think @a-robinson said more or less the same thing in #5563 (comment).)

Yes, see my reply above for Alex's comment