Re: infinite loop in read_hpet from ktime_get_boot_fast_ns

From: Peter Zijlstra
Date: Wed Jun 12 2019 - 11:41:12 EST


On Wed, Jun 12, 2019 at 02:58:21PM +0200, Jason A. Donenfeld wrote:
> Hi Peter,
>
> Thanks for the explanation.
>
> On Wed, Jun 12, 2019 at 2:29 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > Either local_clock() or cpu_clock(cpu). The sleep hooks are not
> > something the consumer has to worry about.
>
> Alright. Just so long as it *is* tracking sleep, then that's fine. If
> it isn't some important aspects of the protocol will be violated.

The scheduler also cares about how long a task has been sleeping, so
yes, that's automagic.

> > If an architecture doesn't provide a sched_clock(), you're on a
> > seriously handicapped arch. It wraps in ~500 days, and aside from
> > changing jiffies_lock to a latch, I don't think we can do much about it.
>
> Are you sure? The base definition I'm looking at uses jiffies:
>
> unsigned long long __weak sched_clock(void)
> {
> return (unsigned long long)(jiffies - INITIAL_JIFFIES)
> * (NSEC_PER_SEC / HZ);
> }
>
> On a CONFIG_HZ_1000 machine, jiffies wraps in ~49.7 days:
> >>> ((1<<32)-1)/1000/(60*60*24)
> 49.710269618055555

Bah, I must've done the math wrong (or assumed HZ=100).

> Why not just use get_jiffies_64()? The lock is too costly on 32bit?

Deadlocks when you do get_jiffies_64() from within an update. What would
be an easier update is forcing everyone to use the GENERIC_SCHED_CLOCK
fallback or something like that.

OTOH, changing jiffies_lock to a latch shouldn't be rocket science
either.

> > (the scheduler too expects sched_clock() to not wrap short of the u64
> > and so having those machines online for 500 days will get you 'funny'
> > results)
>
> Ahh. So if, on the other hand, the whole machine explodes at the wrap
> mark, I guess my silly protocol is the least of concerns, and so this
> shouldn't matter?

That was my thinking...