Concurrency bug in LRUCache about jinja

[jvs]

[Markdown messed up my report. Let me try again.]

I think the LRUCache class in Jinja2 has a bug. The problem is that the "getitem" method may mutate the cache, but that method does not acquire the write lock, "_wlock".

For example, here's one scenario:

• Thread A calls "getitem" with the key "K".
• "K" is not the last element of the queue, so Thread A removes it.
• Thread A is suspended before it appends "K" to the queue.
• Thread B calls "setitem" with the key "K".
• Thread B finds "K" in the "mapping" dictionary.
• Thread B tries to remove "K" from the queue.
• The queue raises an exception, because it does not contain "K".

In my project I've seen the "setitem" method raise a KeyError. I assume that there's a similar execution scenario behind those exceptions, but I could be wrong.

In any case, I think "getitem" clearly needs to acquire "_wlock" before it removes and appends keys.

[tonytargonski]

To add to this 2 year old ticket... I have also observed the same exception in my system.

[Tue Feb 12 19:48:16 2013] [error]   File "/local/python_envs/www/local/lib/python2.7/site-packages/jinja2/utils.py", line 473, in __setitem__
[Tue Feb 12 19:48:16 2013] [error]     del self._mapping[self._popleft()]
[Tue Feb 12 19:48:16 2013] [error] KeyError: '_breadcrumbs.html