On Thu 13 Dec 2018 at 23:32, Cong Wang <xiyou.wangc...@gmail.com> wrote: > On Tue, Dec 11, 2018 at 2:19 AM Vlad Buslov <vla...@mellanox.com> wrote: >> >> As a part of the effort to remove dependency on rtnl lock, cls API is being >> converted to use fine-grained locking mechanisms instead of global rtnl >> lock. However, chain_head_change callback for ingress Qdisc is a sleeping >> function and cannot be executed while holding a spinlock. > > > Why does it have to be a spinlock not a mutex? > > I've read your cover letter and this changelog, I don't find any > answer.
My initial implementation used mutex. However, it was changed to spinlock by Jiri's request during internal review. > >> >> Extend cls API with new workqueue intended to be used for tcf_proto >> lifetime management. Modify tcf_proto_destroy() to deallocate proto >> asynchronously on workqueue in order to ensure that all chain_head_change >> callbacks involving the proto complete before it is freed. Convert >> mini_qdisc_pair_swap(), that is used as a chain_head_change callback for >> ingress and clsact Qdiscs, to use a workqueue. Move Qdisc deallocation to >> tc_proto_wq ordered workqueue that is used to destroy tcf proto instances. >> This is necessary to ensure that Qdisc is destroyed after all instances of >> chain/proto that it contains in order to prevent use-after-free error in >> tc_chain_notify_delete(). > > > Please avoid async unless you have to, there are almost always bugs > when playing with deferred workqueue or any other callbacks. Indeed, async Qdisc and tp deallocation introduces additional complexity. What approach would you recommend to make chain_head_change callback atomic? > > Thanks. Thank you for reviewing my code!
