Re: [patch] new, scalable timer implementation, smptimers-2.4.0-B1

[Ingo Molnar]

On Wed, 31 Jan 2001, Daniel Phillips wrote:

>   b) The current API looks like it was designed primarly with one-shot
> timers in mind.  Most timers events are multishot (because
> sleep_on_timeout is better for most one-shot applications [...]

sleep_on_timeout() uses a one-shot timer internally.

but for 2.4, the changing of the timer interface is out of question. My
main goal was to achieve good SMP scalability with the existing interface.

i do not agree with passing the timer address instead of the ->data field.
It's one more dereference to use, for no particular reason. If you want to
get at the timer structure you can still do it by embedding it into a
structure:

	struct foo {
		...
		timer_t timer;
		...
	}

and ->data will point to &foo.

with the timerlist lock being per-CPU, basically all lock contention has
been eliminated. So it's not a problem anymore to drop/reaquire the lock,
it's not more than a nicely cached, CPU-local variable.

	Ingo