Currently the gitdb will memory map pack indices, and loose objects for reading. If there are too many mapped files, which could happen on large databases, one has to gracefully unload existing memory maps and reload them later.
This could be as easy as deleting the data caches, which are memory maps most of the time.

This task could be done by the PackedDB, and additionally by the git-db which handles multiple object databases.