We observed some issues in a mod_worker environment where attempts to

reload extensions would fail, causing state to become corrupted, leading

to more and more errors. Some of these cases were made more

bullet-proof, but we didn't fix the core issue.
This change attempts to prevent what may be causing the issue by

allowing only one thread to reload the list of extensions at a time.
We have a lock set up for reloading extensions (in case two users try to

reload in the admin UI at the same time).
There's a separate lock in the middleware that's a bit smarter about

calling load(). It checks if the extension state has expired, tries to

grab a lock, and then checks again for expiration. The goal there is to

not have a bunch of threads that are waiting decide they're all going to

trigger their own reloads.
Unit tests have been written that hit issues both during uninit and

during init of extensions. These mirror the errors we've been getting

on reviews.reviewboard.org with Review Bot. With the fix, these tests

pass.

I don't have a real-world repro case yet (we'll see in a couple days on

reviews.reviewboard.org if we start to hit any problems), but I was able

to reproduce some of the errors we through unit tests.
Basically, there's a number of possible failure points, depending on what

instructions are being executed, but the tests manage to reliably trigger

one error during init and one during uninit. Those trigger only if I turn off

locking. Everything passes if locking is on.