core: fix missed register_callsite error

[hds]

Note: This is a change against the v0.1.x branch.

Motivation

There are 2 triggers which will cause a subscriber to receive a call to
Subscriber::register_callsite for a specific callsite.

1. The first time the event or span at that callsite is executed.
2. When a new subscriber is added or removed (for example, calls to
   set_default or with_default)

It is trigger (2) that will cause a new subscriber to receive
Subscriber::register_callsite for all the callsites which had already
been registered before it became active.

When a callsite is registered for trigger (1), the callsite starts in
state UNREGISTERED.

The first thread to encounter the callsite will transition it to
REGISTERING and determine the overall interest for the callsite by
registering with all known dispatchers (which will call into
Subscriber::register_callsite).

Once that is complete, the callsite is added to the list of all known
callsites and its state is transitioned to REGISTERED.

is (re)built for all known dispatchers. The callsite starts in state
UNREGISTERED. The This calls down into
Subscriber::register_callsite for each subscriber. Once that is
complete, the callsite is added to the global list of known callsites.

While the callsite interest is being rebuilt, other threads that
encounter the callsite will be given Interest::sometimes() until the
registration is complete. However, if a new subscriber is added during
this window, all the interest for all callsites will be rebuilt, but
because the new callsite (in state REGISTERING) won't be included
because it isn't yet in the global list of callsites.

This can cause a case where that new subscriber being added won't
receive Subscriber::register_callsite before it receives the subsequent
call to Subscriber::event or Subscriber::new_span.

The documentation on Registering Callsites is not very explicit on
this point, but it does suggest that Subscriber::register_callsite
will be called before the call to either Subscriber::event or
Subscriber::new_span, and the current behavior can break this implicit
contract.

Solution

This change swaps the order of rebuilding the callsite interest and
adding the callsite to the global list so that the callsite gets pushed
first, avoiding this window in which a subscriber won't get a call to
register_callsite.

As such, a callsite may have its interest read before it is set. In this
case, the existing implementation will return Interest::sometimes()
for the DefaultCallsite implementation. Other implementations (outside
of the tracing project) may perform this differently, but in this
case, there is no documented guarantee regarding the ordering.

A regression test is included which provokes the race condition 100% of
the time before the changes in this fix.

Fixes: #2743

[hds]

@hawkw I'd love to get your opinion on the change I've made in the callsite module. It seems fine for me (and certainly no tests failing), but I think you're probably the only person in the world who would know.

The test in tracing can probably go, and I might need another test in tracing-core to cover the dyn Callsite case. Also, interestingly, I couldn't reproduce this issue on master.

[hawkw]

@hawkw I'd love to get your opinion on the change I've made in the callsite module. It seems fine for me (and certainly no tests failing), but I think you're probably the only person in the world who would know.

Thanks for the ping! I'll try to give it a look today or tomorrow.

[mladedav]

@hds you mentioned the docs not being explicit on the topic, could you maybe clarify them? You just might be the second most knowledgeable person on the topic now.