[Qemu-devel] i8259 defaults wrong?

[Qemu-devel] i8259 defaults wrong?

[Herbert Poetzl]

Greetings!

looking at the i8259 implementation found in qemu
as well as in the in-kernel kvm implementation, I
see that on pic_reset() special_fully_nested_mode
is set to zero, but the intel(r) 8259A manual says
on page 15:

  Fully Nested Mode
 
  This mode is entered after initialization unless
  another mode is programmed.

any comments are appreciated, a patch should be
trivial, but I'd gladly submit one if requested.

TIA,
Herbert

[Qemu-devel] Re: i8259 defaults wrong?

[Avi Kivity]

On 08/17/2009 05:20 AM, Herbert Poetzl wrote:
> Greetings!
>
> looking at the i8259 implementation found in qemu
> as well as in the in-kernel kvm implementation, I
> see that on pic_reset() special_fully_nested_mode
> is set to zero, but the intel(r) 8259A manual says
> on page 15:
>
>    Fully Nested Mode
>
>    This mode is entered after initialization unless
>    another mode is programmed.
>
> any comments are appreciated, a patch should be
> trivial, but I'd gladly submit one if requested.
>
>    

Well, does the BIOS set fully nested mode on reset?  If not, a kvm 
change is difficult, since the bios and kvm are updated independently.

-- 
error compiling committee.c: too many arguments to function

[Qemu-devel] Re: i8259 defaults wrong?

[Herbert Poetzl]

>On Mon, Aug 17, 2009 at 04:20:24AM +0200, Herbert Poetzl wrote:
>> 
>> Greetings!
>> 
>> looking at the i8259 implementation found in qemu
>> as well as in the in-kernel kvm implementation, I
>> see that on pic_reset() special_fully_nested_mode
>> is set to zero, but the intel(r) 8259A manual says
>> on page 15:
>> 
>>   Fully Nested Mode
>>  
>>   This mode is entered after initialization unless
>>   another mode is programmed.
>> 
>> any comments are appreciated, a patch should be
>> trivial, but I'd gladly submit one if requested.
>
> Well, does the BIOS set fully nested mode on reset? 

no idea, but as far as I tested, it doesn't matter
for Linux guests, but makes some other operating
systems - which seem to heavily rely on this default 
(e.g. OpenStep, darwin) - work like expected.

> If not, a kvm change is difficult, since the bios 
> and kvm are updated independently.

nevertheless IMHO it should be fixed ASAP, and the 
required change is trivial, i.e. only a single line 
in the kvm/qemu code needs to be changed to make it
work correctly (see patches below)

best,
Herbert

please CC me as I'm not subscribed to either list


; fix i8259 reset behaviour 
; qemu/kvm (outside the kernel)
;
--- ./hw/i8259.c.orig	2009-08-12 19:45:45.000000000 +0200
+++ ./hw/i8259.c	2009-08-17 03:48:15.000000000 +0200
@@ -295,7 +295,7 @@ static void pic_reset(void *opaque)
     s->init_state = 0;
     s->auto_eoi = 0;
     s->rotate_on_auto_eoi = 0;
-    s->special_fully_nested_mode = 0;
+    s->special_fully_nested_mode = 1;
     s->init4 = 0;
     s->single_mode = 0;
     /* Note: ELCR is not reset */


; fix i8259 reset behaviour
; kvm (inside the kernel)

--- ./arch/x86/kvm/i8259.c.orig	2009-06-11 17:12:23.000000000 +0200
+++ ./arch/x86/kvm/i8259.c	2009-08-18 02:50:00.000000000 +0200
@@ -279,7 +279,7 @@ void kvm_pic_reset(struct kvm_kpic_state
 	s->init_state = 0;
 	s->auto_eoi = 0;
 	s->rotate_on_auto_eoi = 0;
-	s->special_fully_nested_mode = 0;
+	s->special_fully_nested_mode = 1;
 	s->init4 = 0;
 }
 

>> TIA,
>> Herbert

[Qemu-devel] Re: i8259 defaults wrong?

[Avi Kivity]

On 08/18/2009 04:02 AM, Herbert Poetzl wrote:
>>
>> Well, does the BIOS set fully nested mode on reset?
>>      
> no idea, but as far as I tested, it doesn't matter
> for Linux guests, but makes some other operating
> systems - which seem to heavily rely on this default
> (e.g. OpenStep, darwin) - work like expected.
>
>    

Please send the patches then separately to their lists with appropriate 
signoffs.  I'll give it an autotest run and see.

> please CC me as I'm not subscribed to either list
>    

But you have

     Mail-Followup-To: kvm@vger.kernel.org, qemu-devel@nongnu.org,
   	    Avi Kivity<avi@redhat.com>


This is incredibly annoying.


-- 
error compiling committee.c: too many arguments to function

Re: [Qemu-devel] i8259 defaults wrong?

[Bernhard Kauer]

On Mon, Aug 17, 2009 at 04:20:24AM +0200, Herbert Poetzl wrote:
> looking at the i8259 implementation found in qemu
> as well as in the in-kernel kvm implementation, I
> see that on pic_reset() special_fully_nested_mode
> is set to zero, but the intel(r) 8259A manual says
> on page 15:
> 
>   Fully Nested Mode
>  
>   This mode is entered after initialization unless
>   another mode is programmed.
> 
> any comments are appreciated,

You confuse two modes of the PIC: the FullyNestedMode
and the SpecialFullyNestedMode. See page 18 in the
manual.

The first ist the "default" mode, where an IRQ in
service from a slave masks further interrupts from this
slave.  In SFNM a slave PIC can still issue another
IRQ.


Greetings,

	Bernhard Kauer