RegexURLResolver.reverse() is not threadsafe

[tdterry]

I have found an incomplete-initialization bug in RegexURLResolver.reverse() when running under Apache mpm_worker with mod_wsgi.

RegexURLResolver objects are stored in process local memory using @memoize. The _reverse_dict is built on demand, so two threads trying to build it for the first time can run into an incomplete initialization. The attached patch changes RegexURLResolver._get_reverse_dict() to do an atomic set of self._reverse_dict, preventing multiple threads from seeing an inconsistent _reverse_dict.

Testing requires a multi-threaded setup and a good load test. These are not available in the framework, so I have not included a specific test with the patch.

[tdterry]

Patch against r10030 to solve threadsafe issues with the urlresolvers.RegexURLResolver

[mrts]

Pushing to accepted, the bug is real (strangely I haven't been hit by this running under the same setup) and the patch looks OK.

[mrts]

Looking around a bit more thoroughly, I didn't find no places where RegexURLResolver instances are shared between threads. Main use is via BaseHandler.get_response() where the BaseHandler instance creates the resolver instance anew for each request. The ones created within urls.py (conf/urls/defaults.py) are initialized before requests come in and not modified after (correct me in the latter).

Pushing back to unreviewed, IMHO can be closed as invalid.

[mrts]

tdterry, I assumed you were actually hit by a bug, not doing code review. If you indeed were, please provide more context.

[Malcolm Tredinnick]

(In [10037]) Fixed #10472 -- Fixed a race condition in reverse URL resolving.

This only shows up in for reverse() (not forwards resolving), since that
path uses a globally shared resolver object. Based on a patch from
Travis Terry.

[Malcolm Tredinnick]

(In [10039]) [1.0.X] Fixed #10472 -- Fixed a race condition in reverse URL resolving.

This only shows up in for reverse() (not forwards resolving), since that
path uses a globally shared resolver object. Based on a patch from
Travis Terry.

Backport of r10037 from trunk.

[mrts]

tdterry, sorry, I missed the memoize on line 73, should have read the ticket description more carefully.