Xen devel - Jun 2012 - [bug report] Windows HVM Hang when reboot/power off using special config

Hi,

Seems there''s a bug when booting HVM guest Windows 2003 using
special config (pin 2 VCPUs to same phy CPU). The guest OS will hange when
reboot or power off system
in guest. The CPU will be 100 percent when watching xentop. The config file
as following:

****************************************************
kernel="/usr/lib/xen-4.1/boot/hvmloader"
builder=''hvm''
name="windows_2003"
uuid="bb29f502-315a-488d-a234-c5651bcd6fbe"
memory=4096
vcpus=2
on_reboot=''restart''
on_crash=''restart''
sdl=0
vnc=1
vnclisten="0.0.0.0"
vncdisplay=29
stdvga=0
serial=''pty''
usbdevice=''tablet''
localtime=1
cpus=[''5'',''5'']
***************************************

Then I debug qemu-dm, find the OS never execute the ACPI register write, so
the QEMU can not catch the system reboot/power off event.
The normal case for guest OS poweroff will be : at first all system
process/driver quit, then OS write ACPI register to poweroff system power.

1. Qemu fetch the register memory map in shared page, judge if it''s
ACPI
register write.
2. If it''s a reset, reboot, poweroff ACPI register operation, then
call qemu_system_shutdown_request() or qemu_system_reset_request()
 to set a flag
3. If the flag be set, call destroy_hvm_domain()
4. Qemu process quit, xend clear other resource

In my case, the qemu_system_shutdown_request ( ACPI register write ) never
be triggered. And the VCPU usage be 100 percent.
So I think it must exist some spinlock-like code in guest OS which cause
the ACPI write never be executed.
If I pin one VCPU to another CPU like ''6'', then ACPI register
write be
called immediately, guest OS poweroff smoothly.

So anyone know why it''s not work when PIN 2 VCPUs to same physical CPU
when
booting HVM Windows 2003? Thanks.

Normal call stack:
qemu_system_reset_request () at /root/qemu/xen-4.1.2/qemu/vl.c:3673
#1  0x000000000047950a in cpu_ioreq_pio (req=0x7ff6d7dbd000, env=0x22a1c40)
at /root/qemu/xen-4.1.2/qemu/i386-dm/helper2.c:351
#2  __handle_ioreq (env=0x22a1c40, req=0x7ff6d7dbd000) at
/root/qemu/xen-4.1.2/qemu/i386-dm/helper2.c:446
#3  0x0000000000479d7b in cpu_handle_ioreq (opaque=0x22a1c40) at
/root/qemu/xen-4.1.2/qemu/i386-dm/helper2.c:515
#4  0x000000000040d81f in main_loop_wait (timeout=<optimized out>) at
/root/qemu/xen-4.1.2/qemu/vl.c:3794


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

On Tue, Jun 5, 2012 at 4:46 AM, Zhou Jacky <jackyzt98@gmail.com>
wrote:
> In my case, the qemu_system_shutdown_request ( ACPI register write ) never
> be triggered. And the VCPU usage be 100 percent.
> So I think it must exist some spinlock-like code in guest OS which cause
the
> ACPI write never be executed.
> If I pin one VCPU to another CPU like ''6'', then ACPI
register write be
> called immediately, guest OS poweroff smoothly.
>
> So anyone know why it''s not work when PIN 2 VCPUs to same physical
CPU when
> booting HVM Windows 2003? Thanks.
I think you''ve already given the most probable answer your own
question -- there must be some kind of synchronization that the guest
OS is doing, probably relating to gracefully bringing down secondary
cpus, that breaks when you pin them to a single core.

 -George