feat: USDT probes for tokio task events

[conradludgate]

Motivation

As discussed on discord:

My team isn't comfortable with the performance concerns of enabling tokio's tracing feature since we do actually use tracing in our application and such there is potentially a measurable cost of evaluating our EnvFilter every time in our performance critical services.

Solution

Using Userspace Statically Defined Tracing (USDT) we expose lightweight probes that can be attached to at runtime with tools like bpftrace or dtrace. This is inspired by https://github.com/oxidecomputer/usdt.

The new functionality is behind a new unstable feature flag. Currently it only exposes some basic task events and not yet any resource events.

See USDT in the wild:

• https://github.com/oxidecomputer/tokio-dtrace
• https://github.com/aws/s2n-quic/blob/main/quic/s2n-quic-core/src/event/generated/metrics/probe.rs

[hds]

Nice idea to represent waker as wake + drop and remove the need for wake_by_ref.

[conradludgate]

I actually wasn't sure if I liked that 😅

[conradludgate]

Musings regarding usdt performance: oxidecomputer/usdt#490

[conradludgate]

I think I'm going to vendor the minimal subset of the code needed to support USDT on the platforms supported by the usdt crate. Removing any extra dependencies in the process.

I'd like to work with the oxide devs to improve this but currently usdt brings in a lot of unnecessary dependencies, and it also doesn't allow cross compilation. Lastly I want to rework the generated code to reduce any performance impact.

Because I will remove all dependencies, I'll move it back to a cfg tokio_usdt and remove the feature