On Fri, Aug 30, 2013 at 01:40:18PM -0700, Jarno Rajahalme wrote: > The "groups" container is protected with a rwlock, and each group is > protected with their own rwlock. Whenever groups are looked up, the > container's lock is taken. If a group is found, the group is locked > before the container is released, as otherwise the group might be > deleted while the lock is being acquired. The group's hmap_node is > considered part of the container, and is governed by the container's > lock. This means that while a reader is still accessing the group, > the hmap itself may be rebalanced, or the group may be removed from > the container for deletion, but the group itself is not deleted as > long as there are any readers left. > > A groups lock is only locked for writing when the container's write > lock is already held. This prevents deadlocks where one thread is > holding a write lock on the group, and is waiting for a write lock > on the container, while another (reading) thread is holding a lock > on the container and is waiting for a lock for the group. > > Signed-off-by: Jarno Rajahalme <jrajaha...@nicira.com>
The design seems well thought out. Thanks. I folded this in. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
