Re: Another SIGFPE in display code, now in cirrus

[Avi Kivity <avi@redhat.com>]

On 05/12/2010 03:20 PM, Stefano Stabellini wrote:
> On Mon, 10 May 2010, Avi Kivity wrote:
>    
>> On 05/10/2010 10:41 AM, Avi Kivity wrote:
>>      
>>> On 05/06/2010 11:07 PM, Michael Tokarev wrote:
>>>        
>>>> There was a bug recently fixed in vnc code.  Apparently
>>>> there's something similar in the cirrus emulation as well.
>>>> Here it triggers _always_ (including old versions of kvm)
>>>> when running windows NT and hitting "test" button in its
>>>> display resolution dialog.  Here's what gdb is to say:
>>>>
>>>> Program received signal SIGFPE, Arithmetic exception.
>>>> [Switching to Thread 0xf76cab70 (LWP 580)]
>>>> 0x080c5e45 in cirrus_do_copy (s=0x86134dc, dst=960000, src=0, w=2, h=9)
>>>>      at hw/cirrus_vga.c:687
>>>> 687        sx = (src % ABS(s->cirrus_blt_srcpitch)) / depth;
>>>> (gdb) p depth
>>>> $1 = 2
>>>> (gdb) p s->cirrus_blt_srcpitch
>>>> $2 = 0
>>>>
>>>>
>>>> This qemu-kvm-0.12.3 - actually a debian package of it,
>>>> but there's no patches relevant to video applied.
>>>>
>>>> Anything can be done with it?
>>>>          
>>> Well, it's trivial to check for the condition, but how to handle it?
>>>
>>>        
>> That code is wrong.  It shouldn't be using the bitblt source pitch, but
>> the actual display pitch.  If the two are different, then the blt
>> doesn't actually affect a rectangular region, but instead a parallelogram.
>>
>>      
> Reading some documention about the Cirrus card we are emulating, it
> seems that the source pitch is ignored in video to video operations, so
> I think you are right here.
>    

 From what I've read I don't think it's ignored.  Rather there are two 
separate pitches, one for bitblt and one for display.

I think the code should be something like

     if bitblt destination intersects display memory:
         if bitblt pitch == display pitch
            use console_copy
        else
            invalidate entire display

> The Windows NT driver probably relies on this behaviour and doesn't set
> the source pitch properly before the bitblt.
>    

Note it's just during mode changes.  During normal operation I'm sure 
the pitches are equal.

>> 31c05501c says this breaks bitblt, but I can't see why this is true.
>> The copy should update the display.  This is probably due to a
>> miscalculation of the affected region, and now we have two invalidates
>> instead of one, reducing performance.
>>
>>      
> I agree with you: qemu_console_copy does imply that the copied portion
> of the screen changed, so there is no reason to invalidate it again if
> qemu_console_copy is called.
>    

Well, we can't just revert 31c05501c.  There's probably another bug 
involved.

-- 
Do not meddle in the internals of kernels, for they are subtle and quick to panic.