Xen users - Nov 2008 - Xen GPL PV and undesirable requirement for APIC (irq > 15)

Hi James,

Thanks for all the hard work on the GPL PV project.

I''ve been having trouble with the more recent builds and then saw this:
> Making Windows do the right thing with IRQ''s etc was really hard,
I
tried quite a few different approaches and this was the best one I could
come up with. It is a pity that it doesn''t work with apic=0, but I
don''t
think that there is a valid reason these days why you would set apic=0.
Previously (xen 3.0.x) it was required to make some versions of windows
work.
>
> James
The APIC problems haven''t gone away, as far as I''m aware...


Like a number of people I run my Windows HVMs with apic=0 (or rather with
the "Standard PC" hal.dll or halacpi.dll) to avoid performance
problems.

Earlier releases of the GPL PV drivers worked well with "Standard PC"
and
Windows XP halacpi.dll , but in the last few weeks they have been
unusable.

I just wanted to express my concern about the idea that the xennet driver
should crash if used with a non APIC HAL.  The xennet driver should work
with any HAL, just like the xenvbd driver.

It would be nice if the apic problems were fixed, too.  Any idea why the
APIC related problem?  I think the pre-requisites for the symptoms are AMD
CPU and 64 bit hypervisor and 32 bit Windows HVM.

reference:
http://forums.novell.com/novell-product-support-forums/suse-linux-enterprise-server-sles/sles-virtualization/337603-win2k3-xp-poor-performance-eats-10-cpu-while-idle.html



_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

> Hi James,
> 
> Thanks for all the hard work on the GPL PV project.
> 
> I''ve been having trouble with the more recent builds and then saw
this:
> 
> > Making Windows do the right thing with IRQ''s etc was really
hard, I
> tried quite a few different approaches and this was the best one I
could
> come up with. It is a pity that it doesn''t work with apic=0, but I
don''t
> think that there is a valid reason these days why you would set
apic=0.
> Previously (xen 3.0.x) it was required to make some versions of
windows
> work.
> >
> > James
> 
> The APIC problems haven''t gone away, as far as I''m
aware...
> 
> 
> Like a number of people I run my Windows HVMs with apic=0 (or rather
with
> the "Standard PC" hal.dll or halacpi.dll) to avoid performance
problems.
> 
> Earlier releases of the GPL PV drivers worked well with "Standard
PC"
and
> Windows XP halacpi.dll , but in the last few weeks they have been
> unusable.
> 
> I just wanted to express my concern about the idea that the xennet
driver
> should crash if used with a non APIC HAL.  The xennet driver should
work
> with any HAL, just like the xenvbd driver.
> 
> It would be nice if the apic problems were fixed, too.  Any idea why
the
> APIC related problem?  I think the pre-requisites for the symptoms are
AMD
> CPU and 64 bit hypervisor and 32 bit Windows HVM.
> 
> reference:
> http://forums.novell.com/novell-product-support-forums/suse-linux-
> enterprise-server-sles/sles-virtualization/337603-win2k3-xp-poor-
> performance-eats-10-cpu-while-idle.html
I reversed the bit that made it fail when it couldn''t find an IRQ >
16.
I had incorrectly assumed that all IRQ''s < 16 were ISA anyway. I
think I
just need to detect the type of IRQ (ISA or PCI) and act
appropriately...

I will revisit this when I''ve gotten the pvscsi working again (I
desperately need that for internal testing, but it involves quite a few
changes).

James

_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users

I concur on seeing the issue. I changed the HAL to Standard PC (Non-apic)
and the Xennet driver strangely got an IP address from the DHCP, but could
not communicate after that. The regular APIC Multiprocessor HAL works fine
with the drivers, but a few machines kill the server. In fact, the idle
usage is 0.9% and if you add more virtual cpus it shoots up to 3.5%, making
impossible to use.
Looking at in perspective, if we are forced to use a single VCPU windows
machine, then Virtual Box gives a substantial advantage in performance, and
you may run many more machines. They don''t support SMP, but it seems
that
XEN does not either.
 

-----Original Message-----
From: xen-users-bounces@lists.xensource.com
[mailto:xen-users-bounces@lists.xensource.com] On Behalf Of James Harper
Sent: Tuesday, November 25, 2008 4:18 AM
To: Oliver Wilcock; xen-users@lists.xensource.com
Subject: RE: [Xen-users] Xen GPL PV and undesirable requirement for APIC
(irq >15)
> Hi James,
> 
> Thanks for all the hard work on the GPL PV project.
> 
> I''ve been having trouble with the more recent builds and then saw
this:
> 
> > Making Windows do the right thing with IRQ''s etc was really
hard, I
> tried quite a few different approaches and this was the best one I
could
> come up with. It is a pity that it doesn''t work with apic=0, but I
don''t
> think that there is a valid reason these days why you would set
apic=0.
> Previously (xen 3.0.x) it was required to make some versions of
windows
> work.
> >
> > James
> 
> The APIC problems haven''t gone away, as far as I''m
aware...
> 
> 
> Like a number of people I run my Windows HVMs with apic=0 (or rather
with
> the "Standard PC" hal.dll or halacpi.dll) to avoid performance
problems.
> 
> Earlier releases of the GPL PV drivers worked well with "Standard
PC"
and
> Windows XP halacpi.dll , but in the last few weeks they have been
> unusable.
> 
> I just wanted to express my concern about the idea that the xennet
driver
> should crash if used with a non APIC HAL.  The xennet driver should
work
> with any HAL, just like the xenvbd driver.
> 
> It would be nice if the apic problems were fixed, too.  Any idea why
the
> APIC related problem?  I think the pre-requisites for the symptoms are
AMD
> CPU and 64 bit hypervisor and 32 bit Windows HVM.
> 
> reference:
> http://forums.novell.com/novell-product-support-forums/suse-linux-
> enterprise-server-sles/sles-virtualization/337603-win2k3-xp-poor-
> performance-eats-10-cpu-while-idle.html
I reversed the bit that made it fail when it couldn''t find an IRQ >
16.
I had incorrectly assumed that all IRQ''s < 16 were ISA anyway. I
think I
just need to detect the type of IRQ (ISA or PCI) and act
appropriately...

I will revisit this when I''ve gotten the pvscsi working again (I
desperately need that for internal testing, but it involves quite a few
changes).

James

_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users


_______________________________________________
Xen-users mailing list
Xen-users@lists.xensource.com
http://lists.xensource.com/xen-users