Messages in this thread | | | From | Linus Torvalds <> | Date | Tue, 3 Sep 2019 08:44:40 -0700 | Subject | Re: [PATCH 0/3] task: Making tasks on the runqueue rcu protected |
| |
On Mon, Sep 2, 2019 at 9:50 PM Eric W. Biederman <ebiederm@xmission.com> wrote: > > I have split this work into 3 simple patches, so the code is straight > forward to review and so that if any mistakes slip in it is easy to > bisect them. In the process of review what it takes to remove > task_rcu_dereference I found yet another user of tasks on the > runqueue in rcu context; the rcuwait_event code. That code only needs > it now unnecessary limits removed.
Looks very good to me.
I think PeterZ is right that the rcu_assign_pointer() in [PATCH 2/3] could be a RCU_INIT_POINTER() due to condition #3 in the RCU_INIT_POINTER rules. The initialization of the pointer value simply has nothing to do with what the pointer points to - we're just switching it to another case.
That said, it won't affect any of the core architectures much, because smp_store_release() isn't that expensive (it's just a compiler barrier on x86, it's a cheap instruction on arm64, and it should be very cheap on any other architecture too unless they do insane things - even on powerpc, which is about the worst case for any barriers, it's just an lwsync).
It might be good to have Paul _look_ at it, and because of the minimal performance impact I don't worry about it too much if it happens later, but it should be something we keep in mind.
Linus
| |