Re: [PATCH 3/3] timer: Reduce unnecessary sighand lock contention

[Jason Low <jason.low2@hp.com>]

On Thu, 2015-08-27 at 00:31 +0200, Frederic Weisbecker wrote:
> On Wed, Aug 26, 2015 at 10:53:35AM -0700, Linus Torvalds wrote:
> > On Tue, Aug 25, 2015 at 8:17 PM, Jason Low <jason.low2@hp.com> wrote:
> > >
> > > This patch addresses this by having the thread_group_cputimer structure
> > > maintain a boolean to signify when a thread in the group is already
> > > checking for process wide timers, and adds extra logic in the fastpath
> > > to check the boolean.
> > 
> > It is not at all obvious why the unlocked read of that variable is
> > safe, and why there is no race with another thread just about to end
> > its check_process_timers().
> 
> The risk is when a next timer is going to expire soon after we relaxed
> the "checking" variable due to a recent expiration. The thread which
> expires the next timer may still see a stale value on the "checking"
> state and therefore delay the timer firing until the new value is seen.
> So the worst that can happen is that the timer firing gets delayed for
> X jiffies (I guess in practice it's only 1 jiffy).
> 
> That said, posix cpu timers already suffer such race because
> sig->cputimer.running itself is checked outside the sighand lock anyway.

Right, in the worst case, the next thread that comes along will handle
it. The current code already expects that level of granularity, so this
shouldn't change anything much. For example, there is almost always a
delay between a timer expiring and when a thread actually calls into
check_process_timers().

> > I can well imagine that this is all perfectly safe and fine, but I'd
> > really like to see that comment about _why_ that's the case, and why a
> > completely unlocked access without even memory barriers is fine.
> 
> Agreed, there should be a comment about that in the code (that is already full
> of undocumented subtleties).

Okay, I will add more comments about this.