Re: [PATCH v6 2/2] sched/uclamp: Protect uclamp fast path code with static key

[Peter Zijlstra]

On Tue, Jun 30, 2020 at 06:55:02PM +0100, Qais Yousef wrote:
> On 06/30/20 19:07, Peter Zijlstra wrote:

> > There's a fun race described in 9107c89e269d ("perf: Fix race between
> > event install and jump_labels"), are we sure this isn't also susceptible
> > to something similar?
> > 
> > I suspect not, but I just wanted to make sure.
> 
> IIUC, the worry is that not all CPUs might have observed the change in the
> static key state; hence could not be running the patched
> enqueue/dequeue_task(), so we could end up with some CPUs accounting for
> uclamp in the enqueue/dequeue path but not others?
> 
> I was hoping this synchronization is guaranteed by the static_branch_*() call.

It is, that isn't quite the the problem. Looking at it more I think
commit 1dbb6704de91 ("jump_label: Fix concurrent
static_key_enable/disable()") fixed some of it.

>From what I can remember there were two parts to this problem, one being
fixed by the above commit, the other being that if we enable while a
task is running we miss the switch-in event (exactly how in this patch
we miss the enqueue).

Due to the missing switch-in, the state is 'weird' and the subsequent
IPI to install a remote event didn't quite work.

So I put that sync_sched() call in to guarantee all CPUs have done a
schedule() cycle after having the key switched. This makes sure that
every running task has seen the switch-in and thus the state is as
expected.

But like I said, I think we're good, that one extra branch deals with
the half-state.