Fix deadlock in update_counts

[davidhewitt]

This fixes the deadlock reported in the discussion in #916, which became possible after incomplete test coverage of the change made in #899.

I wrote the test first and so can confirm the deadlock did indeed exist before this bugfix.

[kngwyu]

Looks good, but I feel the combination of Lock and UnsafeCell redundant here. How about using parking_lot::Mutex or std::sync::Mutex?
Also, it would be good if @IvanKuzavkov could check this patch works.

[IvanKuzavkov]

@kngwyu @davidhewitt I've checked this fix, it works well - thanks.

[davidhewitt]

I would be happy to use a mutex to reduce amount of unsafe code. For now we'll probably need to use parking_lot because std::sync::mutex cannot be easily constructed in a static without lazy_static. Also I think I read somewhere that parking_lot has better performance?

[davidhewitt]

Changed to remove Option. I gotta head to work; if you want this branch merged today and want any changes to it, please feel free to change as you like 😄

[kngwyu]

I don't think we need Option here, too. Zero-sized Vec is internally a constant and fast.

[davidhewitt]

Yep agreed 👍 if you want to change this before I get a chance later tonight, please push to this branch 😄

[kngwyu]

Done 👍

[kngwyu]

Thank you!