Re: [PATCH v2] mc146818rtc: fix timer interrupt reinjection

["Philippe Mathieu-Daudé" <philmd@redhat.com>]

On 10/24/19 1:22 PM, Philippe Mathieu-Daudé wrote:
> On 10/10/19 2:30 PM, Marcelo Tosatti wrote:
>>
>> commit 369b41359af46bded5799c9ef8be2b641d92e043 broke timer interrupt
>> reinjection when there is no period change by the guest.
>>
>> In that case, old_period is 0, which ends up zeroing irq_coalesced
>> (counter of reinjected interrupts).
>>
>> The consequence is Windows 7 is unable to synchronize time via NTP.
>> Easily reproducible by playing a fullscreen video with cirrus and VNC.
>>
>> Fix by not updating s->irq_coalesced when old_period is 0.
>>
>> V2: reorganize code (Paolo Bonzini)
> 
> ^ This line shouldn't be part of git history,
> Paolo if possible can you drop it?
> 
>>
>> Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
>>
>> diff --git a/hw/timer/mc146818rtc.c b/hw/timer/mc146818rtc.c
>> index 6cb378751b..0e7cf97042 100644
>> --- a/hw/timer/mc146818rtc.c
>> +++ b/hw/timer/mc146818rtc.c
>> @@ -203,24 +203,28 @@ periodic_timer_update(RTCState *s, int64_t 
>> current_time, uint32_t old_period)
>>       period = rtc_periodic_clock_ticks(s);
>> -    if (period) {
>> -        /* compute 32 khz clock */
>> -        cur_clock =
>> -            muldiv64(current_time, RTC_CLOCK_RATE, 
>> NANOSECONDS_PER_SECOND);
>> +    if (!period) {
>> +        s->irq_coalesced = 0;
>> +        timer_del(s->periodic_timer);
>> +        return;
>> +    }
> 
> This is the first change, simplify the if statement with a return.
> 
>> -        /*
>> -        * if the periodic timer's update is due to period 
>> re-configuration,
>> -        * we should count the clock since last interrupt.
>> -        */
>> -        if (old_period) {
>> -            int64_t last_periodic_clock, next_periodic_clock;
>> -
>> -            next_periodic_clock = muldiv64(s->next_periodic_time,
>> -                                    RTC_CLOCK_RATE, 
>> NANOSECONDS_PER_SECOND);
>> -            last_periodic_clock = next_periodic_clock - old_period;
>> -            lost_clock = cur_clock - last_periodic_clock;
>> -            assert(lost_clock >= 0);
>> -        }
>> +    /* compute 32 khz clock */
>> +    cur_clock =
>> +        muldiv64(current_time, RTC_CLOCK_RATE, NANOSECONDS_PER_SECOND);
>> +
>> +    /*
>> +     * if the periodic timer's update is due to period re-configuration,
>> +     * we should count the clock since last interrupt.
>> +     */
>> +    if (old_period) {
>> +        int64_t last_periodic_clock, next_periodic_clock;
>> +
>> +        next_periodic_clock = muldiv64(s->next_periodic_time,
>> +                                RTC_CLOCK_RATE, NANOSECONDS_PER_SECOND);
>> +        last_periodic_clock = next_periodic_clock - old_period;
>> +        lost_clock = cur_clock - last_periodic_clock;
>> +        assert(lost_clock >= 0);
>>           /*
>>            * s->irq_coalesced can change for two reasons:
>> @@ -251,22 +255,19 @@ periodic_timer_update(RTCState *s, int64_t 
>> current_time, uint32_t old_period)
>>                   rtc_coalesced_timer_update(s);
>>               }
>>           } else {
>> -           /*
>> +            /*
>>                * no way to compensate the interrupt if 
>> LOST_TICK_POLICY_SLEW
>>                * is not used, we should make the time progress anyway.
>>                */
>>               lost_clock = MIN(lost_clock, period);
>>           }
>> +    }
> 
> This is the second change, changing the logic and fixing the bug.
> 
>> -        assert(lost_clock >= 0 && lost_clock <= period);
>> +    assert(lost_clock >= 0 && lost_clock <= period);
>> -        next_irq_clock = cur_clock + period - lost_clock;
>> -        s->next_periodic_time = periodic_clock_to_ns(next_irq_clock) 
>> + 1;
>> -        timer_mod(s->periodic_timer, s->next_periodic_time);
>> -    } else {
>> -        s->irq_coalesced = 0;
>> -        timer_del(s->periodic_timer);
>> -    }
>> +    next_irq_clock = cur_clock + period - lost_clock;
>> +    s->next_periodic_time = periodic_clock_to_ns(next_irq_clock) + 1;
>> +    timer_mod(s->periodic_timer, s->next_periodic_time);
> 
> This is part of the 1st change.
> 
> I'd rather see this patch split in 2 logical changes...

Since this split makes sense to me (easier to understand the patch diff 
when reviewing), I can quickly do it if you don't mind.

>>   }
>>   static void rtc_periodic_timer(void *opaque)
>>
>>