On Thu, Oct 19, 2017 at 8:34 AM, John Fastabend <john.fastab...@gmail.com> wrote: > > My take on this would be to stay with the current RCU callbacks and try > to simplify the implementation. Falling back to sync operations seems > like a step backwards to me. I know update/delete of filters is currently > a pain point for some use cases so getting the RTNL out of the way may > become a requirement to support those (alternatively maybe batching is > good enough).
For me it looks like very hard to make tc action destroy code completely race-free in RCU callbacks, at least looks harder than getting rid of RCU callbacks. > > I guess at a high level with Cris' patches actions are now doing reference > counting correctly. If shared filters also do reference counting similarly > we should be OK right? (yes I know simplifying maybe too much to be > meaningful) I don't know what you mean by "doing reference counting correctly", if you mean making them atomic, as I already explained to Chris, it is not necessary at all if we remove RCU callbacks. Refcnt doesn't have to be atomic if it is always serialized with a lock. > > Are we aware of any outstanding problem areas? > Potentially many problems, since tc action destroy code could be called either with a RTNL lock (fine) or in a RCU callback without RTNL lock (buggy), these two paths race with each other and RCU callbacks race among themselves too.
