Re: [PATCH] RFC: abstract access to xtime_lock into a set of inline functions

[Thomas Gleixner <tglx@linutronix.de>]

[-- Attachment #1: Type: TEXT/PLAIN, Size: 1612 bytes --]

On Thu, 20 Jan 2011, john stultz wrote:

> On Thu, Jan 20, 2011 at 9:30 AM, Christoph Hellwig <hch@infradead.org> wrote:
> > On Thu, Jan 20, 2011 at 06:14:08PM +0100, Torben Hohn wrote:
> >> the -rt patches change the xtime_lock to a raw_seqlock_t
> >> so a pretty huge portion of the patch deals with changing
> >> the locking functions.
> >>
> >> this commit uses inline functions, to hide the type
> >> of the lock.
> >
> > That's not how kernel code usually works.
> 
> Yea, I'm not a fan of this patch either.
> 
> 
> >> -     write_seqlock(&xtime_lock);
> >> +     xtime_write_seqlock();
> >>       do_timer(1);
> >> -     write_sequnlock(&xtime_lock);
> >> +     xtime_write_sequnlock();
> >
> > However there's a pretty clear pattern of taking xtime_lock, calling
> > do_timer and then releasing.  A useful thing you could do is to rename
> > do_timer to do_timer_locked and make do_timer take and release
> > xtime_lock in one place.
> 
> Seems like a reasonable suggestion.  I suspect there's still quite a
> bit of stuff done under the same lock right around do_timer on a

I looked through the few remaining usage sites in arch/ and there is
really no code inside the locked region which relies on xtime_lock,
AFAICT.

> number of arches, but  having a locked call would cut down on how
> widely xtime is used.

I'd suggest to move do_timer() to kernel/time/timekeeping.c, then add
xtime_update() which takes xtime_lock and calls do_timer(). After that
convert each arch with a separate patch. The last step makes
xtime_lock and do_timer() local to the timekeeping / clockevents code.

Thanks,

	tglx