Marek Lindner wrote:
Hi,
I continued improving the the performance patches which results in the first usable version. I tested it in a VM environment as well as on real hardware and did not encounter problematic behavior anymore.
I looked through your patches and have some small suggestions. First I would like to move the hashtable-functions related patches to the front. This would make it possible to send a new version of batman-adv to David S. Miller without adding to many changes (so maybe he doesn't get too angry this time).
Then you missed hashtable_t::elements of the hash implementation which are only needed for the resize implementation (now it is only a wom :) ). The hash_iterate function is also still there.... I would like to remove it completely. hash_remove only rcu locked the list - but we must spinlock here to prevent double frees. kref_put of data directly after call_rcu for bucket could lead to invalid memory access when a rcu protected region still accesses the bucket which isn't removed yet (but the data in the bucket was already removed). Unneeded(?) hlist_empty checks. Missing conversation of neigh_node hlist operations to _rcu. Missing rcu_read_unlock's. Missing correct refcnt for neigh_list
I would rework some of those parts and resend all patches again to the mailing list. This doesn't mean that I have checked everything.... some of the patches were just too "wirr" and this is a first try to get it in the right direction. It still needs a lot of cleanup and testing. For example neigh_list stuff and the kref_put after call_rcu is complete unfixed and I also didn't finish the complete patchset.... maybe will continue with it when the patchset is cleaned up.
Best regards, Sven