Re: [Qemu-devel] [RESEND][PATCH 0/3] Fix guest time drift under heavy load.

[Dor Laor <dlaor@redhat.com>]

[-- Attachment #1: Type: text/plain, Size: 2243 bytes --]

Glauber Costa wrote:
> On Wed, Nov 12, 2008 at 9:42 AM, Gleb Natapov <gleb@redhat.com> wrote:
>   
>> On Tue, Nov 11, 2008 at 02:17:49PM -0600, Anthony Liguori wrote:
>>     
>>> Gleb Natapov wrote:
>>>       
>>>> On Mon, Nov 10, 2008 at 09:46:12AM -0600, Anthony Liguori wrote:
>>>>   -usbdevice tablet has nothing to do with it. Qemu misses interrupt
>>>> even
>>>> without this option and with SDL screen it misses them in bunches when
>>>> SDL redraws a screen. In case of vnc qemu misses interrupt because of
>>>> fsync() call in raw_flush(), or so my instrumentation shows.
>>>>
>>>>         
>>> Can you give this patch a spin?
>>>
>>>       
>> Doesn't compile for me. fd_pool_inuse and fd_inuse are used but not
>> defined.
>>
>>     
>>> This introduces a bdrv_aio_flush() which will wait for all existing AIO
>>> operations to complete before indicating completion.  It also fixes up
>>> IDE.  Fixing up SCSI will be a little more tricky but not much.  Since
>>> we now use O_DSYNC, it's unnecessary to do an fsync (or an fdatasync).
>>>
>>> Assuming you're using IDE, this should eliminate any delays from fsync.
>>>       
>> I am using IDE.
>>
>>     
>>> SDL delays are unavoidable because it's going to come down to SDL doing
>>> sychronous updates to the X server.  The proper long term solution here
>>> would be to put SDL in it's own thread but I'm not too worried about
>>>       
>> And probably time-keeping deserves its own thread. And CPU execution
>> too.
>>     
>
> It might well be a stupid idea, (would have to benchmark it), but the
> other day it occurred to me
> that we could keep timekeeping in a separate _process_, with a shm
> area, doing timekeeping
> for all running guests.
>
>   
The problem is that the right vcpu should be preempted in order us 
injecting time irq into it.
It will introduce latency (signal/IPI).
This is why the in-kernel pit performs better/ more accurate.

What we can do is add a kernel interface to schedule a timer on a 
specific cpu for preempting the vcpu.
The problem such interface is a bit awkward and Avi feels it's beginning 
to be too complex.

So either we fix it using Gleb's set_irq ack method or just fix the RTC 
using Andrzej suggestion for RTC irq
status bit.

[-- Attachment #2: Type: text/html, Size: 3041 bytes --]