Messages in this thread | | | Date | Mon, 12 Jan 2015 11:05:29 +0100 (CET) | From | Thomas Gleixner <> | Subject | Re: frequent lockups in 3.18rc4 |
| |
On Sun, 21 Dec 2014, Linus Torvalds wrote: > On Sun, Dec 21, 2014 at 2:32 PM, Dave Jones <davej@codemonkey.org.uk> wrote: > > On Sun, Dec 21, 2014 at 02:19:03PM -0800, Linus Torvalds wrote: > > > > > > And finally, and stupidly, is there any chance that you have anything > > > accessing /dev/hpet? > > > > Not knowingly at least, but who the hell knows what systemd has its > > fingers in these days. > > Actually, it looks like /dev/hpet doesn't allow write access. > > I can do the mmap(/dev/mem) thing and access the HPET by hand, and > when I write zero to it I immediately get something like this: > > Clocksource tsc unstable (delta = -284317725450 ns) > Switched to clocksource hpet > > just to confirm that yes, a jump in the HPET counter would indeed give > those kinds of symptoms:blaming the TSC with a negative delta in the > 0-300s range, even though it's the HPET that is broken. > > And if the HPET then occasionally jumps around afterwards, it would > show up as ktime_get() occasionally going backwards, which in turn > would - as far as I can tell - result in exactly that pseudo-infirnite > loop with timers. > > Anyway, any wild kernel pointer access *could* happen to just hit the > HPET and write to the main counter value, although I'd personally be > more inclined to blame BIOS/SMM kind of code playing tricks with > time.. We do have a few places where we explicitly write the value on > purpose, but they are in the HPET init code, and in the clocksource > resume code, so they should not be involved.
Right.
> Thomas - have you had reports of HPET breakage in RT circles, the same > way BIOSes have been tinkering with TSC?
Not that I'm aware of.
> Also, would it perhaps be a good idea to make "ktime_get()" save the > last time in a percpu variable, and warn if time ever goes backwards > on a particular CPU? A percpu thing should be pretty cheap, even if > we write to it every time somebody asks for time..
That should be simple enough to implement.
Thanks,
tglx
| |