Xen users - Jan 2009 - GPLPV Network Driver causing VM crash

I''m running Xen 3.3.0 with AMD Opteron 2350s. Windows 2003 32bit
SP2+patches.

When I bring a Windows VM up with the GPLPV drivers, it runs until I try
to do anything with the network and then it crashes.  Trying to web
browse will lock the VM up, RDPing to it will lock the VM up.  I''ve
tried 0.9.11 and 0.9.12pre13, both with the same results.  I''ve removed
the type=ioemu from the vif line without any change as well.

Any ideas?


Thanks,
pace


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

> I''m running Xen 3.3.0 with AMD Opteron 2350s. Windows 2003 32bit
> SP2+patches.
> 
> When I bring a Windows VM up with the GPLPV drivers, it runs until I
try
> to do anything with the network and then it crashes.  Trying to web
browse
> will lock the VM up, RDPing to it will lock the VM up.  I''ve tried
0.9.11
> and 0.9.12pre13, both with the same results.  I''ve removed the
type=ioemu
> from the vif line without any change as well.
> 
> Any ideas?
> 
Add this to your domu config, all on one line:

cpuid
[''0x80000005:ecx=xxxxxxxxxxxxxxxxxxxxxxxxkkkkkkkk,edx=xxxxxxxxxxxxxxxxxx
xxxxxxkkkkkkkk'']

that will correct a bug in 3.3.0 where it wasn''t exporting the cache
line size to windows, which would cause Windows to hang the first time
it received a packet making use of rx csum offload. It only affects AMD
cpu''s.

James


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

Sweet. Thanks James. That did the trick.

Would this 3.3.0 bug affect the Windows VMs regardless of the PV drivers?


pace

James Harper wrote:
>> I''m running Xen 3.3.0 with AMD Opteron 2350s. Windows 2003
32bit
>> SP2+patches.
>>
>> When I bring a Windows VM up with the GPLPV drivers, it runs until I
>>     
> try
>   
>> to do anything with the network and then it crashes.  Trying to web
>>     
> browse
>   
>> will lock the VM up, RDPing to it will lock the VM up.  I''ve
tried
>>     
> 0.9.11
>   
>> and 0.9.12pre13, both with the same results.  I''ve removed the
>>     
> type=ioemu
>   
>> from the vif line without any change as well.
>>
>> Any ideas?
>>
>>     
>
> Add this to your domu config, all on one line:
>
> cpuid >
[''0x80000005:ecx=xxxxxxxxxxxxxxxxxxxxxxxxkkkkkkkk,edx=xxxxxxxxxxxxxxxxxx
> xxxxxxkkkkkkkk'']
>
> that will correct a bug in 3.3.0 where it wasn''t exporting the
cache
> line size to windows, which would cause Windows to hang the first time
> it received a packet making use of rx csum offload. It only affects AMD
> cpu''s.
>
> James
>   

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

> 
> Sweet.  Thanks James.  That did the trick.
> 
> Would this 3.3.0 bug affect the Windows VMs regardless of the PV
drivers?
> 
I''m not sure. I have only heard of it being a problem when using my PV
drivers. There is a kernel call in Windows that preloads a certain
amount of data into the cache. This is a gross oversimplification, but
the routine in question runs like:

bytes_loaded = 0
while (bytes_loaded < bytes_to_load)
{
  ...
  bytes_loaded += cache_line_size
}

Because the cpuid isn''t correct in 3.3.0 for AMD processors,
cache_line_size is always zero so bytes_loaded never increases and it
just spins forever.

NDIS (windows network stack) calls the above code when I give it a
packet and tell it that the checksum has already been calculated, but
that''s the only time I have ever seen it called. It''s a bit of
a
mystery.

James

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

> NDIS (windows network stack) calls the above code when I give it a
> packet and tell it that the checksum has already been calculated, but
> that''s the only time I have ever seen it called. It''s a
bit of a
> mystery.
>
>   
Sounds like the issue would only crop up if the Realtek driver did
checksum offloading.  From a brief search, you may be able to get newer
drivers that support that.  I''m not sure that warrants any research
since I see on xen-devel that they''ve fixed this for 3.3.1...


Thanks again,
pace

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