Re: Instability with Xen, interrupt routing frozen, HPET broadcast

[Jeremy Fitzhardinge <jeremy@goop.org>]

 On 09/30/2010 03:16 AM, Andreas Kinzler wrote:
> On 29.09.2010 21:50, Jeremy Fitzhardinge wrote:
>>> It is a Supermicro X8SIL-F, Intel Xeon 3450 system.
>> The big problem I had initially was instability with the integrated
>> ethernet until I disabled PCIe ASPM.  The symptom was that the ethernet
>> devices would disappear (ie, their PCI config space would start to read
>> all 0xff...)
>
> I know that this is a known problem of Intel 82574L chips (on X8SIL) -
> it is discussed on "Intel Wired Ethernet"
> (http://sourceforge.net/projects/e1000/).

Aha, specifically
http://sourceforge.net/tracker/index.php?func=detail&aid=2908463&group_id=42302&atid=447449,
in which several people invoke me, but nobody bothered to tell me that
this bug existed on sf :/

> That is why I tested different NICs (Intel ET Server Adapter (82576
> [igb]) and Realtek 8168) and the problem remained. So I can say with
> certainty that the NIC and/or its power management is not the problem.

OK.

>
> I also spend extensive time changing hardware components. I used a
> different mainboard (ASUS P7F-M), a different power supply, changed
> CPU, changed NICs (see above) - problems remained.
>
> > That's exactly what my main test/devel machine is.  It has been very
> > stable for me with xen-unstable.
>
> We have a second Supermicro X8SIL-F, Intel Xeon 3450 system which only
> runs Linux PVM domains and it is totally stable (without my HPET
> patch). So I think as with all timing/race/deadlock/... issues it
> depends on what you do on your system. Let me give you my crash
> "recipe" [quite reliable ;-)]

OK.  My machine is mostly running PV domains, with some low-intensity
hvm ones.

>
> Have two HVMs (called win1, win2) with Windows 7 x64 installed (do
> install everything twice, never clone, VM config attached). Install
> GPLPV 0.11.0.213, iometer 2006.07.27, prime95 25.11 x64. On both
> systems: start prime95 torture test (in-place large FFT) and using
> Windows task manager set CPU affinity on win1 of process prime95 to
> use only CPU1. On win2 do the same thing but to use only CPU0. Then
> start iometer on both VMs using the following parameters: have a
> second virtual disk in both VMs (so every windows has 2 virtual disks,
> one for Windows and one for iometer), use "# of outstanding I/Os" = 4,
> access spec = "All in one". Wait some minutes. Crash!

Yes, that's a very different workload from mine.

    J