Re: [Xen-devel] xl vcpu-set <guest> <nr> does not work properly.

[Anthony PERARD <anthony.perard@xxxxxxxxxx>]

On Thu, Dec 05, 2013 at 12:27:21PM +0000, Anthony PERARD wrote:
> On Wed, Dec 04, 2013 at 02:57:24PM -0500, Konrad Rzeszutek Wilk wrote:
> > When I try this with the "SeaBIOS" QEMU it does not hotplug
> > all of the CPUs and it complains about the existing ones:
> > 

> [...]

> 
> > -bash-4.1# xl vcpu-list
> > Name                                ID  VCPU   CPU State   Time(s) CPU
> > Affinity
> > Domain-0                             0     0    0   -b-      87.4  any
> > cpu
> > Domain-0                             0     1    3   -b-      14.6  any
> > cpu
> > Domain-0                             0     2    2   -b-      11.1  any
> > cpu
> > Domain-0                             0     3    1   r--      13.7  any
> > cpu
> > latest                               7     0    2   -b-      11.7  any
> > cpu
> > latest                               7     1    1   -b-       1.7  any
> > cpu
> > latest                               7     2    3   -b-       0.0  any
> > cpu
> > 
> > And sure enough - in the guest it only added one CPU:
> > 
> > [   85.720139] CPU2 has been hot-added
> > [   85.726409] installing Xen timer for CPU 2
> > [   85.735674] smpboot: Booting Node 0 Processor 2 APIC 0x4
> > [   85.755080] cpu 2 spinlock event irq 86
> > [   85.763671] microcode: CPU2 sig=0x206a7, pf=0x2, revision=0x28
> > [   85.770669] Will online and init hotplugged CPU: 2
> > 
> > Perhaps the SeaBIOS/new qemu needs variants of:
> > 
> > f62079c piix4acpi, xen, hotplug: Fix race with ACPI AML code and hotplug.
> > e28e062 piix4acpi, xen: Clarify that the qemu_set_irq calls just do an IRQ 
> > pulse.
> > 2133c38 piix4acpi, xen, vcpu hotplug: Split the notification from the 
> > changes
> 
> Yes, we should fix that in QEMU.
> 
> Thanks for reporting, I will work on it.

Instead of porting those patch, I think we can fix the issue by a small
change in the ACPI table. I tried to reproduce the issue with KVM, but I
could not, so I tried to compare the ACPI table and there is one detail
that could be interesting, they use Edge-Triggered even, where we have
Level-Triggered even. After a quick test, it looks like it works (to use
Edge-Triggered even).

-- 
Anthony PERARD

_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel