Re: [PATCH 18/21] Remove host_alarm_timer hacks.

[Anthony Liguori <aliguori@us.ibm.com>]

Avi Kivity wrote:
> Anthony Liguori wrote:
>> Signed-off-by: Anthony Liguori <aliguori@us.ibm.com>
>> ---
>>  vl.c |    3 +--
>>  1 files changed, 1 insertions(+), 2 deletions(-)
>>
>> diff --git a/vl.c b/vl.c
>> index 3b0e3dc..848a8f8 100644
>> --- a/vl.c
>> +++ b/vl.c
>> @@ -1367,8 +1367,7 @@ static void host_alarm_handler(int host_signum)
>>          last_clock = ti;
>>      }
>>  #endif
>> -    if (1 ||
>> -        alarm_has_dynticks(alarm_timer) ||
>> +    if (alarm_has_dynticks(alarm_timer) ||
>>          (!use_icount &&
>>              qemu_timer_expired(active_timers[QEMU_TIMER_VIRTUAL],
>>                                 qemu_get_clock(vm_clock))) ||
>>   
>
> This was added to fix a problem.  Have you tested it?

Do you know what problem it fixes?

This goes back a very long time.  IIUC, this was added prior to the IO 
thread as an "optimization".  This ensures that any time there's a 
timer, the vcpu is interrupted to allow IO to run.  With non-dynticks, 
there can be spurious timer signals because we problem the timer with a 
fixed frequency.  It's necessary to take this path with dynticks because 
we need to rearm the timer which happens in the IO path.  It's not 
necessary to take this path with a non-dynticks timer unless there's 
been an expiration.

In modern KVM, the IO thread is capable of interrupting the CPU whenever 
it needs to process IO.  Therefore this "problem" no longer exists.

Regards,

Anthony Liguori


-- 
Regards,

Anthony Liguori