Re: [BUG] Paravirtual time accounting / IRQ time accounting

[lwcheng]

Quoting Mike Galbraith <umgwanakikbuti@xxxxxxxxx>:

> On Thu, 2014-03-20 at 12:01 -0300, Glauber Costa wrote:
> > On Wed, Mar 19, 2014 at 6:42 AM,  <lwcheng@xxxxxxxxx> wrote:
> > > In consolidated environments, when there are multiple virtual  
> > machines (VMs)
> > > running on one CPU core, timekeeping will be a problem to the guest OS.
> > > Here, I report my findings about Linux process scheduler.
> > >
> > >
> > > Description
> > > ------------
> > > Linux CFS relies on rq->clock_task to charge each task, determine  
> > vruntime,
> > > etc.
> > >
> > > When CONFIG_IRQ_TIME_ACCOUNTING is enabled, the time spent on serving IRQ
> > > will be excluded from updating rq->clock_task.
> > > When CONFIG_PARAVIRT_TIME_ACCOUNTING is enabled, the time stolen by the
> > > hypervisor
> > > will also be excluded from updating rq->clock_task.
> > >
> > > With "both" CONFIG_IRQ_TIME_ACCOUNTING and CONFIG_PARAVIRT_TIME_ACCOUNTING
> > > enabled,
> > > I put three KVM guests on one core and run hackbench in each guest. I find
> > > that
> > > in the guests, rq->clock_task stays *unchanged*. The malfunction  
> > embarrasses
> > > CFS.
> > > ------------
> > >
> > >
> > > Analysis
> > > ------------
> > > [src/kernel/sched/core.c]
> > > static void update_rq_clock_task(struct rq *rq, s64 delta)
> > > {
> > >     ... ...
> > > #ifdef CONFIG_IRQ_TIME_ACCOUNTING
> > >     irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
> > >     ... ...
> > >     rq->prev_irq_time += irq_delta;
> > >     delta -= irq_delta;
> > > #endif
> > >
> > > #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
> > >     if (static_key_false((&paravirt_steal_rq_enabled))) {
> > >         steal = paravirt_steal_clock(cpu_of(rq));
> > >         steal -= rq->prev_steal_time_rq;
> > >         ... ...
> > >         rq->prev_steal_time_rq += steal;
> > >         delta -= steal;
> > >     }
> > > #endif
> > >
> > >     rq->clock_task += delta;
> > >     ... ...
> > > }
> > > --
> > > "delta" -> the intended increment to rq->clock_task
> > > "irq_delta" -> the time spent on serving IRQ (hard + soft)
> > > "steal" -> the time stolen by the underlying hypervisor
> > > --
> > > "irq_delta" is calculated based on sched_clock_cpu(), which is vulnerable
> > > to VM scheduling delays.
> >
> > This looks like a real problem indeed. The main problem in searching
> > for a solution, is that of course not all of the irq time is steal
> > time and vice versa. In this case, we could subtract irq_time from
> > steal, and add only the steal part time that is in excess. I don't
> > think this is 100 % guaranteed, but maybe it is a good approximation.
> >
> > Rik, do you have an opinion on this ?
>
> Hrm, on my little Q6600 box, I'm running 3 VMS all pinned to CPU3, all
> running hackbench -l zillion, one of them also running crash, staring at
> it's sole rq->clock_task as I write this, with kernels (3.11.10) on both
> host and guest configured as reported.
>
>   clock_task = 631322187004,
>   clock_task = 631387807452,
>   clock_task = 631474214294,
>   clock_task = 631523864893,
>   clock_task = 631604646268,
>   clock_task = 631643276025,
>
> Maybe 3 VMs isn't enough overload for such a beastly CPU.  Top reports
> some very funky utilization numbers, but other than that, the things
> seem to work fine here.  perf thinks scheduling work too.
>
> -Mike

I checked the source code again. I forgot to mention that I commented out
steal_ticks() in my experiments (sorry):
[kernel/sched/core.c]
-------------------------------
#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
    if (static_key_false((&paravirt_steal_rq_enabled))) {
        /*
        u64 st;
        */
        steal = paravirt_steal_clock(cpu_of(rq));
        steal -= rq->prev_steal_time_rq;

        if (unlikely(steal > delta))
            steal = delta;
        /*
        st = steal_ticks(steal);
        steal = st * TICK_NSEC;
        */
        rq->prev_steal_time_rq += steal;

        delta -= steal;
    }
#endif

    rq->clock_task += delta;
-------------------------------

I do it just for "accuracy", because I fully trust "steal" reported
by the hypervisor. I do not quite understand why it is trimmed again
using steal_ticks(). Please enlighten me.

Even when "steal == delta", as long as "steal" is not exactly
(N * TICK_NSEC), after steal_ticks(), it will be always smaller
than "delta".
So, rq->clock_task can still progress, but may not in the precise way.
Each time, the error is within the range (0, TICK_NSEC).

When CONFIG_IRQ_TIME_ACCOUNTING is disabled, deleting steal_ticks()
does not affect the update to rq->clock_task.

I tested both 3.10.0 and 3.13.5. The results are consistent.

-Luwei




--
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/