Re: [PATCH v2 0/3] Deferrable timers support for timerfd API

From: Thomas Gleixner
Date: Wed Feb 05 2014 - 17:16:57 EST


On Wed, 5 Feb 2014, John Stultz wrote:
> On 02/05/2014 01:41 PM, Thomas Gleixner wrote:
> > On Wed, 5 Feb 2014, Alexey Perevalov wrote:
> >> On 02/04/2014 08:10 PM, Thomas Gleixner wrote:
> >>> On Mon, 27 Jan 2014, Alexey Perevalov wrote:
> >>>> On 01/21/2014 11:12 PM, John Stultz wrote:
> >>>>> Thomas: Any thought here? Should we be trying to unify the timerfd flags
> >>>>> and the posix timer flags (specifically things like TIMER_CANCEL_ON_SET,
> >>>>> which is currently timerfd-only)? Should a deferrable flag be added to
> >>>>> the hrtimer core or left to the timer wheel?
> >>> The timer cancel on set was added only to timerfd because timerfd is a
> >>> non posix interface and we are halfways free to add stuff to
> >>> it. Adding extra flags to the real posix timer interfaces is a
> >>> different story.
> >> And what about "deferrable" possibility for hrtimers, do you consider it
> >> reasonable?
> > In principle, I have no objections, but we need a proper technical
> > solution. Just adding a flag and keeping the timers in the same rbtree
> > like we do for the timer wheel timers is not going to happen.
> >
> > The only feasible solution is to have separate clock ids,
> > e.g. CLOCK_*_DEFERRABLE, which would enable the deferrable
> > functionality for all user space interfaces. No need for magic flags
> > and complex search for non deferrable timers.
>
> So of course, I was actually arguing against having a new clockid (which
> was Alexey's first approach).

Mooo.

> My reasoning was that the deferrablity isn't a clock domain, and is more
> of a modifier. Thus to keep the interfaces somewhat sane (and avoiding
> having to add N new clockids for each new modifier), we should utilize
> the flag arguments to timers. So instead of just having TIMER_ABSTIME,
> we could add TIMER_DEFER, etc, which we could utilize instead.

I can see the point. I have no objections against that approach as
long as we map that against separate internal bases.

> Internally we can still keep separate bases, much as your patch does, to
> keep the next-event searching overhead more limited.

It's not only more limited, it's bound.

> I mainly wanted to get your thoughts on extending the flags, and doing
> so in a consistent manner between the timerfd and other timer interfaces.

So the only interface which does not support that is sys_nanosleep()
but that's not really an issue. sys_nanosleep() should die anyway :)

> Of course, all this is after I added the _ALARM clockids... so you can
> decide if its hypocrisy or experience.
> (The "old wisdom comes from experience and experience comes from bad
> decisions" bit ;).

Well, you have a valid point about the clock ids. I did not realize in
the first place that we can avoid that business if we use the flags to
select the internal representation.

Either way is preferred over reintroducing the timer wheel mess....

Thanks,

tglx
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/