On Sun, Aug 28, 2016 at 10:04:21PM -0700, Cong Wang wrote: > On Fri, Aug 26, 2016 at 12:16 PM, Eric Dumazet <eric.duma...@gmail.com> wrote: > > On Fri, 2016-08-26 at 11:26 -0700, Cong Wang wrote: > >> 1) Currently there are only a few actions using lockless, and they are > >> questionable, as we already discussed before, there could be some > >> race condition when you modify an existing action. > > > > There is no fundamental issue with a race condition. > > For mirred action, maybe. As we already discussed, the more > complex an action is, the harder to make it lockless in your > way (that is, not using RCU) > > > > > Sure, there are races, but they have no serious effect. > > > > Feel free to send a fix if you really have time to spare. > > It's because the code is written by you? > > I am surprised how you try to hide your own problem in > such a way... > > > > > >> > >> 2) We need to change the tc action API in order to fully support RCU, > >> which is what I have been working on these days. I should come up > >> with something next Monday (if not this weekend). > >> > >> So for this patchset, using spinlock is fine, just as many other actions. > >> I will take care of it later. > > > > This is _not_ fine. > > > OK, so where are your patches to make the rest actions > lockless? > > > > > > We are in 2016, not in 1995 anymore. > > > > Fair enough, sounds like all actions are already lockless in > fast path now in 2016, you know this is not true... > > > > We are not adding a spinlock in a hot path unless absolutely needed. > > If it is bug-free, yes, I am totally with you. I care about corretness > more than any performance. > > > > > > With multi queue NIC, this spinlock is going to hurt performance so much > > that this action wont be used by any serious user. > > We have used mirred action even before you make it lockless. > > > > > > Here, it is absolutely trivial to use RCU and/or percpu counters. > > Sounds like we don't need any API change, why not go ahead > and try it? Please do teach me how to modify an existing > action in a lockless way without changing any API (and of course > needs to be bug-free), I am very happy to learn your "trivial" way > to fix this, since I don't have any trivial fix. > > Please, stop bullsh*t, show me your trivial code.
Regarding the specific action in this patchset, correct me if I'm wrong, but I think that the lock could be removed safely. 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 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. Thanks, Amir
