Re: [Qemu-devel] [PATCH v3 5/5] hpet 'driftfix': add code in hpet_timer() to compensate delayed callbacks and coalesced interrupts

[Marcelo Tosatti]

On Tue, May 03, 2011 at 07:08:25PM -0300, Glauber Costa wrote:
> On Tue, 2011-05-03 at 16:03 -0300, Marcelo Tosatti wrote:
> > On Thu, Apr 28, 2011 at 04:25:00PM +0200, Ulrich Obergfell wrote:
> > > Loss of periodic timer interrupts caused by delayed callbacks and by
> > > interrupt coalescing is compensated by gradually injecting additional
> > > interrupts during subsequent timer intervals, starting at a rate of
> > > one additional interrupt per interval. The injection of additional
> > > interrupts is based on a backlog of unaccounted HPET clock periods
> > > (new HPETTimer field 'ticks_not_accounted'). The backlog increases
> > > due to delayed callbacks and coalesced interrupts, and it decreases
> > > if an interrupt was injected successfully. If the backlog increases
> > > while compensation is still in progress, the rate at which additional
> > > interrupts are injected is increased too. A limit is imposed on the
> > > backlog and on the rate.
> > > 
> > > Injecting additional timer interrupts to compensate lost interrupts
> > > can alleviate long term time drift. However, on a short time scale,
> > > this method can have the side effect of making virtual machine time
> > > intermittently pass slower and faster than real time (depending on
> > > the guest's time keeping algorithm). Compensation is disabled by
> > > default and can be enabled for guests where this behaviour may be
> > > acceptable.
> > > 
> > > Signed-off-by: Ulrich Obergfell <address@hidden>
> > > ---
> > >  hw/hpet.c |   63 
> > > +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++-
> > >  1 files changed, 61 insertions(+), 2 deletions(-)
> > > 
> > > diff --git a/hw/hpet.c b/hw/hpet.c
> > > index 35466ae..92d5f58 100644
> > > --- a/hw/hpet.c
> > > +++ b/hw/hpet.c
> > > @@ -32,6 +32,7 @@
> > >  #include "sysbus.h"
> > >  #include "mc146818rtc.h"
> > >  #include "sysemu.h"
> > > +#include <assert.h>
> > >  
> > >  //#define HPET_DEBUG
> > >  #ifdef HPET_DEBUG
> > > @@ -42,6 +43,9 @@
> > >  
> > >  #define HPET_MSI_SUPPORT        0
> > >  
> > > +#define MAX_TICKS_NOT_ACCOUNTED     (uint64_t)500000000 /* 5 sec */
> > > +#define MAX_IRQ_RATE                (uint32_t)10
> > > +
> > >  struct HPETState;
> > >  typedef struct HPETTimer {  /* timers */
> > >      uint8_t tn;             /*timer number*/
> > > @@ -326,28 +330,63 @@ static const VMStateDescription vmstate_hpet = {
> > >      }
> > >  };
> > >  
> > > +static bool hpet_timer_has_tick_backlog(HPETTimer *t)
> > > +{
> > > +    uint64_t backlog = t->ticks_not_accounted - (t->period + 
> > > t->prev_period);
> > > +    return (backlog >= t->period);
> > > +}
> > > +
> > >  /*
> > >   * timer expiration callback
> > >   */
> > >  static void hpet_timer(void *opaque)
> > >  {
> > >      HPETTimer *t = opaque;
> > > +    HPETState *s = t->state;
> > >      uint64_t diff;
> > > -
> > > +    int irq_delivered = 0;
> > > +    uint32_t irq_count = 0;
> > >      uint64_t period = t->period;
> > >      uint64_t cur_tick = hpet_get_ticks(t->state);
> > >  
> > > +    if (s->driftfix && !t->ticks_not_accounted) {
> > > +        t->ticks_not_accounted = t->prev_period = t->period;
> > > +    }
> > >      if (timer_is_periodic(t) && period != 0) {
> > >          if (t->config & HPET_TN_32BIT) {
> > >              while (hpet_time_after(cur_tick, t->cmp)) {
> > >                  t->cmp = (uint32_t)(t->cmp + t->period);
> > > +                t->ticks_not_accounted += t->period;
> > > +                irq_count++;
> > >              }
> > >          } else {
> > >              while (hpet_time_after64(cur_tick, t->cmp)) {
> > >                  t->cmp += period;
> > > +                t->ticks_not_accounted += period;
> > > +                irq_count++;
> > >              }
> > >          }
> > >          diff = hpet_calculate_diff(t, cur_tick);
> > > +        if (s->driftfix) {
> > > +            if (t->ticks_not_accounted > MAX_TICKS_NOT_ACCOUNTED) {
> > > +                t->ticks_not_accounted = t->period + t->prev_period;
> > > +            }
> > > +            if (hpet_timer_has_tick_backlog(t)) {
> > > +                if (t->irq_rate == 1 || irq_count > 1) {
> > > +                    t->irq_rate++;
> > > +                    t->irq_rate = MIN(t->irq_rate, MAX_IRQ_RATE);
> > > +                }
> > > +                if (t->divisor == 0) {
> > > +                    assert(irq_count);
> > > +                }
> > > +                if (irq_count) {
> > > +                    t->divisor = t->irq_rate;
> > > +                }
> > > +                diff /= t->divisor--;
> > > +            } else {
> > > +                t->irq_rate = 1;
> > > +            }
> > > +        }
> > >          qemu_mod_timer(t->qemu_timer,
> > >                         qemu_get_clock_ns(vm_clock) + 
> > > (int64_t)ticks_to_ns(diff));
> > >      } else if (t->config & HPET_TN_32BIT && !timer_is_periodic(t)) {
> > > @@ -358,7 +397,22 @@ static void hpet_timer(void *opaque)
> > >              t->wrap_flag = 0;
> > >          }
> > >      }
> > > -    update_irq(t, 1);
> > > +    if (s->driftfix && timer_is_periodic(t) && period != 0) {
> > > +        if (t->ticks_not_accounted >= t->period + t->prev_period) {
                                             ^^^^^^^^^^

> > > +            irq_delivered = update_irq(t, 1);
> > > +            if (irq_delivered) {
> > > +                t->ticks_not_accounted -= t->prev_period;
> > > +                t->prev_period = t->period;
> > > +            } else {
> > > +                if (irq_count) {
> > > +                    t->irq_rate++;
> > > +                    t->irq_rate = MIN(t->irq_rate, MAX_IRQ_RATE);
> > > +                }
> > > +            }
> > > +        }
> > > +    } else {
> > > +        update_irq(t, 1);
> > > +    }
> > >  }
> > 
> > Hi Ulrich,
> > 
> > Whats prev_period for, since in practice the period will not change
> > between interrupts (OS programs comparator once, or perhaps twice during
> > bootup) ?
> 
> Actually, some systems, like Windows 2000-and-something change this
> period quite heavily under multimedia workloads.

I see. Still, using the period at the previous timer handler instance in
the decision whether to inject an interrupt to the guest makes no sense
to me.

> > Other than that, shouldnt reset accounting variables to init state on
> > write to GLOBAL_ENABLE_CFG / writes to main counter?

What i meant to ask here is whether the reinjection state (including
ticks_not_accounted) should be reset whenever a different period is
programmed.