Re: infinite loop in read_hpet from ktime_get_boot_fast_ns

[Peter Zijlstra]

On Wed, Jun 12, 2019 at 11:44:35AM +0200, Jason A. Donenfeld wrote:
> Hey Peter,
> 
> On Wed, Jun 12, 2019 at 11:03 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > How quasi? Do the comments in kernel/sched/clock.c look like something
> > you could use?
> >
> > As already mentioned in the other tasks, anything ktime will be
> > horrifically crap when it ends up using the HPET, the code in
> > kernel/sched/clock.c is a best effort to keep using TSC even when it is
> > deemed unusable for timekeeping.
> 
> Thanks for pointing that out. Indeed the HPET path is a bummer and I'd
> like to just escape using ktime all together.
> 
> In fact, my accuracy requirements are very lax. I could probably even
> deal with an inaccuracy as huge as ~200 milliseconds. But what I do
> need is 64-bit, so that it doesn't wrap, allowing me to compare two
> stamps taken a long time apart, and for it to take into account sleep
> time, like CLOCK_BOOTTIME does, which means get_jiffies_64() doesn't
> fit the bill. I was under the impression that I could only get this
> with ktime_get_boot & co, because those add the sleep offset.
> 
> It looks like, though, kernel/sched/clock.c keeps track of some
> offsets too -- __sched_clock_offset and __gtod_offset,

Right, those are used to keep the clock values coherent (as best as
possible) when we switch modes.

When the TSC is stable sched_clock_cpu() is mapped directly to
sched_clock() for performance reasons. The moment the TSC is detected to
be unsuitable, we switch to the unstable mode, where we take a GTOD
timestamp every tick and add resolution with the CPU local TSC (plus
filters etc..).

To make this mode-switch as smooth as possible, we track those offsets.

> and the comment at the top mentions explicit sleep hooks. I wasn't
> sure which function to use from here, though. 

Either local_clock() or cpu_clock(cpu). The sleep hooks are not
something the consumer has to worry about.

> sched_clock() seems based on jiffies, which
> has the 32-bit wraparound issue, and the base implementation doesn't
> seem to take into account sleeptime. The x86 implementation seems use
> rdtsc and then adds cyc2ns_offset which looks to be based on
> cyc2ns_suspend, which I assume is what I want. 

Yes.

> But there's still the
> issue of the 32-bit wraparound on the base implementation.

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.

(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)

AFAICT only: alpha, h8300, hexagon, m68knommu, nds32, nios2, openrisc
are lacking any form of sched_clock(), the rest has it either natively
or through sched_clock_register().

> I guess you know this code better than my quick perusal. Is there some
> clock in here that doesn't have a wrap around issue and takes into
> account sleeptime, without being super slow like ktime/hpet?

You probably want to use local_clock().