Re: [bug-report] possible s64 overflow in max_vruntime()

[Roman Kagan]

On Thu, Feb 09, 2023 at 02:44:49PM +0100, Vincent Guittot wrote:
> On Thu, 9 Feb 2023 at 14:33, Roman Kagan <rkagan@xxxxxxxxx> wrote:
> >
> > On Thu, Feb 09, 2023 at 12:26:12PM +0100, Vincent Guittot wrote:
> > > On Wed, 8 Feb 2023 at 19:09, Roman Kagan <rkagan@xxxxxxxxx> wrote:
> > > > On Wed, Feb 08, 2023 at 11:13:35AM +0100, Vincent Guittot wrote:
> > > > > On Tue, 7 Feb 2023 at 20:37, Roman Kagan <rkagan@xxxxxxxxx> wrote:
> > > > > >
> > > > > > On Tue, Jan 31, 2023 at 12:10:29PM +0100, Vincent Guittot wrote:
> > > > > > > On Tue, 31 Jan 2023 at 11:00, Roman Kagan <rkagan@xxxxxxxxx> wrote:
> > > > > > > > On Tue, Jan 31, 2023 at 11:21:17AM +0800, Chen Yu wrote:
> > > > > > > > > On 2023-01-27 at 17:18:56 +0100, Vincent Guittot wrote:
> > > > > > > > > > On Fri, 27 Jan 2023 at 12:44, Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > > > > > > > > >
> > > > > > > > > > > On Thu, Jan 26, 2023 at 07:31:02PM +0100, Roman Kagan wrote:
> > > > > > > > > > >
> > > > > > > > > > > > > All that only matters for small sleeps anyway.
> > > > > > > > > > > > >
> > > > > > > > > > > > > Something like:
> > > > > > > > > > > > >
> > > > > > > > > > > > >         sleep_time = U64_MAX;
> > > > > > > > > > > > >         if (se->avg.last_update_time)
> > > > > > > > > > > > >           sleep_time = cfs_rq_clock_pelt(cfs_rq) - se->avg.last_update_time;
> > > > > > > > > > > >
> > > > > > > > > > > > Interesting, why not rq_clock_task(rq_of(cfs_rq)) - se->exec_start, as
> > > > > > > > > > > > others were suggesting?  It appears to better match the notion of sleep
> > > > > > > > > > > > wall-time, no?
> > > > > > > > > > >
> > > > > > > > > > > Should also work I suppose. cfs_rq_clock takes throttling into account,
> > > > > > > > > > > but that should hopefully also not be *that* long, so either should
> > > > > > > > > > > work.
> > > > > > > > > >
> > > > > > > > > > yes rq_clock_task(rq_of(cfs_rq)) should be fine too
> > > > > > > > > >
> > > > > > > > > > Another thing to take into account is the sleeper credit that the
> > > > > > > > > > waking task deserves so the detection should be done once it has been
> > > > > > > > > > subtracted from vruntime.
> > > > > > > > > >
> > > > > > > > > > Last point, when a nice -20 task runs on a rq, it will take a bit more
> > > > > > > > > > than 2 seconds for the vruntime to be increased by more than 24ms (the
> > > > > > > > > > maximum credit that a waking task can get) so threshold must be
> > > > > > > > > > significantly higher than 2 sec. On the opposite side, the lowest
> > > > > > > > > > possible weight of a cfs rq is 2 which means that the problem appears
> > > > > > > > > > for a sleep longer or equal to 2^54 = 2^63*2/1024. We should use this
> > > > > > > > > > value instead of an arbitrary 200 days
> > > > > > > > > Does it mean any threshold between 2 sec and 2^54 nsec should be fine? Because
> > > > > > > > > 1. Any task sleeps longer than 2 sec will get at most 24 ms(sysctl_sched_latency)
> > > > > > > > >    'vruntime bonus' when enqueued.
> > > > > > >
> > > > > > > This means that if a task nice -20 runs on cfs rq while your task is
> > > > > > > sleeping 2seconds, the min vruntime of the cfs rq will increase by
> > > > > > > 24ms. If there are 2 nice -20 tasks then the min vruntime will
> > > > > > > increase by 24ms after 4 seconds and so on ...
> > > > > > >
> > > > > > > On the other side, a task nice 19 that runs 1ms will increase its
> > > > > > > vruntime by around 68ms.
> > > > > > >
> > > > > > > So if there is 1 task nice 19 with 11 tasks nice -20 on the same cfs
> > > > > > > rq, the nice -19 one should run 1ms every 65 seconds and this also
> > > > > > > means that the vruntime of task nice -19 should still be above
> > > > > > > min_vruntime after sleeping 60 seconds. Of course this is even worse
> > > > > > > with a child cgroup with the lowest weight (weight of 2 instead of 15)
> > > > > > >
> > > > > > > Just to say that 60 seconds is not so far away and 2^54 should be better IMHO
> > > > > >
> > > > > > If we go this route, what would be the proper way to infer this value?
> > > > > > Looks like
> > > > > >
> > > > > >   (1ull << 63) / NICE_0_LOAD * scale_load(MIN_SHARES)
> > > > >
> > > > > (1ull << 63) / NICE_0_LOAD * MIN_SHARES
> > > >
> > > > On 64bit platforms NICE_0_LOAD == 1L << 20 (i.e. it's also scaled) for
> > > > better precision.  So this will yield 2^63 / 2^20 * 2 = 2^44.  Good
> > > > enough probably but confusing.
> > >
> > > Something like the below should be enough to explain the value
> > >
> > > /*
> > >  * min_vruntime can move forward much faster than real time. The worst case
> > >  * happens when an entity with the min weight always runs on the cfs rq. In this
> > >  * case, the max comparison between vruntime and min_vruntime can fail after a
> > >  * sleep greater than :
> > >  *     (1ull << 63) / NICE_0_LOAD) * MIN_SHARES
> >
> > Sorry if I'm being dense, but aren't NICE_0_LOAD and MIN_SHARES measured
> > in different units: the former is scaled while the latter is not?
> 
> There are 2 usages of MIN_SHARES:
> - one when setting cgroup weight in __sched_group_set_shares() which
> uses  scale_load(MIN_SHARES)
> - one when sharing this weight between the cfs of the group in
> calc_group_shares() : clamp_t(long, shares, MIN_SHARES, tg_shares)
> 
> The 2nd one is the most important in our case that's why I use
> MIN_SHARES and not scale_load(MIN_SHARES)

I see now, thanks a lot for explaining!

I'll post an updated patch later today.

Thanks,
Roman.



Amazon Development Center Germany GmbH
Krausenstr. 38
10117 Berlin
Geschaeftsfuehrung: Christian Schlaeger, Jonathan Weiss
Eingetragen am Amtsgericht Charlottenburg unter HRB 149173 B
Sitz: Berlin
Ust-ID: DE 289 237 879