On Tue, 2017-10-10 at 21:02 -0700, Paul E. McKenney wrote: > Linus and Ingo will ask me how users decide how they should set that > additional build flag. Especially given that if there is code that > requires non-strict checking, isn't everyone required to set up non-strict > checking to avoid false positives? Unless you can also configure out > all the code that requires non-strict checking, I suppose, but how > would you keep track of what to configure out?
I'm working to a new version using a single compile flag - without additional strict option. I don't know of any other subsytem that stores rcu pointer in datastructures for a longer amount of time. That having said, I wonder if the tests should completely move to the networking subsystem for the time being. The Kconfig option would thus be called NET_DEBUG or something along the lines. For abstraction it would be possible to add an atomic_notifier_chain to the rcu_read/unlock methods, where multiple users or checkers could register for. That way we keep the users seperate from the implementation for the cost of a bit more layering and more indirect calls. But given that this will anyway slow down execution a lot, it will anyway only be suitable in testing/verification/debugging environments. > OK. There will probably be some discussion about the API in that case. I'll drop noref parameter, the key will became mandatory - the exact position of where the reference of RCU managed object is stored. In the case of noref dst it is &skb->_skb_refdst. With this kind of API it should suite more subsystems. > True enough. Except that if people were good about always clearing the > pointer, then the pointer couldn't leak, right? Or am I missing something > in your use cases? This is correct. The dst_entry checking in skb, which this patch series implements there are strict brackets in the sense of skb_dst_set, skb_dst_set_noref, skb_dst_force, etc., which form brackets around the safe uses of those dst_entries. This patch series validates that the correct skb_dst_* functions are being called before the sk_buff leaves the rcu protected section. Thus we don't need to modify and review a lot of code but we can just patch into those helpers already. > Or to put it another way -- have you been able to catch any real > pointer-leak bugs with thister-leak bugs with this? The other RCU > debug options have had pretty long found-bug lists before we accepted > them. There have been two problems found so far, one is a rather minor one while the other one seems like a normal bug. The patches for those are part of this series (3/4 and 4/4). Regards, Paolo
