On Tue, 2016-08-30 at 14:39 +0300, Amir Vadai wrote: > On Tue, Aug 30, 2016 at 02:03:08PM +0300, Amir Vadai wrote:
> > Regarding the specific action in this patchset, correct me if I'm wrong, > > but I think that the lock could be removed safely. Sure ;) > > > > When the action is modified during traffic, an existing tcf_enc_metadata > > is not changed, but a new metadata is allocated and the pointer is > > replaced to point to the new one. > > I just need to make sure that when changing an action from 'release' > > into 'set' - tcf_enc_metadata will be set before the action type is > > changed - change the order of operations and add a memory barrier. > > Here is a pseudo code to explain: > > > > metadata_new = new allocated metadata > > metadata_old = t->tcft_enc_metadata > > > > Oh - I had a typo here: > Need to set the metadata and only after that, set the action: > > t->tcft_enc_metadata = metadata_new > wmb() rcu_assign_pointer() is your friend, it auto documents the thing. Note that you probably need to store in the allocated object : dst, tcf_action (a copy of it, read in tunnel_key_act() tcft_action ( a copy of it, read in tunnel_key_act()) rcu_head rcu for kfree_rcu() > t->tcft_action = encapdecap > > > t->tcft_action = encapdecap > > > > /* make sure the compiler won't swap the setting of tcft_action with > > * tcft_enc_metadata > > */ > > wmb() > > > > t->tcft_enc_metadata = metadata_new > > release metadata_old > > > > > > This way, no need for lock between the init() and act() operations. > > > > Please let me know if you see a problem with this approach. > > I will also change the stats to be percpu. Right, check tcf_hash_create() last argument. (false -> true) Thanks.