Re: time_after - what on earth???

From: Björn Steinbrink
Date: Tue Sep 11 2007 - 19:48:53 EST


On 2007.09.12 00:10:19 +0100, Adrian McMenamin wrote:
> On 12/09/2007, Björn Steinbrink <B.Steinbrink@xxxxxx> wrote:
> > On 2007.09.12 00:19:09 +0200, Rene Herman wrote:
> > > On 09/12/2007 12:15 AM, Adrian McMenamin wrote:
> > >
> > >> On 11/09/2007, Rene Herman <rene.herman@xxxxxxxxx> wrote:
> > >>> On 09/12/2007 12:05 AM, Adrian McMenamin wrote:
> > >>>
> > >>>> OK, why does this line occasionally return true:
> >
> > What exactly is "occassionally"? Does it happen more than once per
> > boot? If not, and it happens after a certain time after booting, it
> > might be wrapping of the jiffie counter (see below).
> >
> > >>>>
> > >>>> if ((maple_dev->interval > 0) && (jiffies >maple_dev->when))
> > >>>>
> > >>>> while this one never does (no other changes made):
> > >>>>
> > >>>> if ((maple_dev->interval > 0) && (time_after(jiffies,
> > >>>> maple_dev->when)))
> > >>> Is maple_dev->when an unsigned long?
> > >>>
> > >> Yes. Does that make a difference?
> > >
> > > If it had been a signed type, it could've wrapped to something you didn't
> > > expect, explaining the difference at least...
> > >
> > > With an unsigned long, the only diference should be that time_after() deals
> > > with jiffie wrapping which I assume is not an actual problem here. I'll
> > > retreat into the shades again... ;-(
> >
> > If "occasionally" is limited to once per boot, it might be jiffie
> > wrapping. IIRC jiffies are initialized so that they wrap after about 5
> > minutes of uptime to reveal such bugs without forcing you to wait for
> > ages just to have the counter wrap for the first time.
> >
>
> No, I mean "works properly" - ie occasionally evaluates as true

Ehrm, yeah, I somehow parsed that as if it had a negation in there.

Anyway, I looked up the patches you posted. "when" is initialized to 0
and only changed if the above condition evaluates to true. Now,
time_after and "<" have different points at which "future" and "past"
are separated. time_after splits (about) equally between future and
past, so 0 can be either, depending on the value of jiffies. But for "<"
0 is almost always in the past, except for the seldom event of jiffies
being 0.

Now, given that jiffies start out at a huge value to make the counter
wrap around early, 0 happens to be in the "future" for time_after, until
the wrap around occurs. So in your case, you just might have to wait
those 5 minutes to get the working behaviour, instead of the common case
in which it breaks after that time ;-)

A fix would likely initialize "when" to jiffies.

Björn
-
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/