>> >> 1) Use refs for buckets (this seems unnecessarily heavyweight) >> 2) Always try bucket_lookup when adding stats to a bucket via the cache. >> If the bucket is not there in the list, we know its been removed via >> modify_group() and we simply don't update stats. >> 3) Use ofgroup->rw_lock to protect writing to stats via the xlate cache >> (this seems to break modularity in the code but it would be the simplest >> to do). >> 4) Similar to netdev, remove the group then re-add the group to ofproto >> instead of modifying the group in place. Thus, the bucket will only be >> invalid when its group parent is also invalid. > >On balance, #4 seems to be the most straight forward way of dealing with >it.
Sweet, I'm working on implementing this. Its taking a surprising amount of refactoring to make it look somewhat clean, so I'll send the new patch out tomorrow. >> >> Let me know if you have any other ideas. I don't particularly like any >>of >> these, but these are what I could come up with on the top of my head. > >How about we remove the rwlock and only use ref count? We need the ref count to determine when we can free the group / bucket. The rwlock is to serialize access to the ofgroup's properties. For example, 2 threads could be accessing ofgroup's properties, but a ref count doesn't prevent them from both editing its properties simultaneously. Unfortunately, I think we still need both here. _______________________________________________ dev mailing list dev@openvswitch.org http://openvswitch.org/mailman/listinfo/dev
