Re: [PATCH] net: Revert "net_sched: no need to free qdisc in RCU callback"

Thu, Dec 21, 2017 at 09:59:56PM CET, xiyou.wangcong@gmail.com wrote:

On Thu, Dec 21, 2017 at 12:39 AM, Jiri Pirko [off-list ref] wrote:

quoted

Why just moving qdisc_free to rcu is not enough? It would resolve this
issue and also avoid using synchronize net. Something like:

If you mean Jakub's issue, apparently not:
https://www.kernel.org/pub/linux/kernel/people/paulmck/Answers/RCU/RCUCBordering.html

Jiri, you have to use a rcu barrier to wait for a rcu callback, not
queuing another rcu callback, the ordering is simply NOT guaranteed.

Sure. But the ordering does not matter in this case. You just need to
make sure the reader does not see freed qdisc struct.

What's more importantly, you already have one rcu barrier in the
same function. Why keep believing you don't need it?

The rcu barrier is there for a different reason. It is there to avoid
reuse of old miniq that readers still might see in scenario
miniq1
miniq2
miniq1 again