Re: Regression, host crash with 4.5rc1

[Don Slutz <dslutz@verizon.com>]

On 11/03/14 16:46, Steve Freitas wrote:
> Hi all,
>
> I've got a Windows 7 x64 VM that is stable on 4.4.1 but crashes the 
> host after a few hours on 4.5rc1. The machine is a ThinkStation D20, 
> an X5660 with 5500-series chipset with an Nvidia Quadro FX4800 
> (genuine) passed through. Distro is Debian Jessie running stock distro 
> kernel and seabios. Both 4.4.1 and 4.5rc1 were built from source. 
> Apologies if this issue has already been spotted but I can't keep up 
> with the traffic on this list! :-)
>
> It looks as if something is stepping on PCI devices when the crash 
> happens. In the log included below, it's the SATA system that's 
> complaining but I've seen it hit the ethernet chip first, then SATA 
> after that. I'm happy to troubleshoot, apply patches, give more 
> information, etc.
>
> I've seen the crash when running Windows Update, when running the 
> Unigine graphics benchmark, when running an Avast anti-virus scan. 
> Haven't found a common thread.
>
> Don, one curious thing I noted which may be of no value whatsoever: 
> Under 4.4.1, I can't give the VM > 3.5 gigs of RAM without breaking 
> VGA passthrough. Under 4.5rc1 I can give the VM more than 3.5 gigs, 
> yet I *don't* need to use the "mmio_hole" settings in the VM config. 
> Not sure why or what that might signify, if anything.
>

First off, I do not know much about pci-passthru.  The fact that it now 
boots with
more then 3.5 GiB just tells me that things have changed (like what 
physical addresses
are used get changed).  Without a clear way to cause this issue, getting 
to the root
cause is much harder.  For example, there is no clear way to say that 
4.4.1 will never
have this issue (it just has not yet done so).

There are 3 mmio_size values that make sense to try.
1) mmio_hole = 1033
    This is what Konrad Wilk suggested (hole based on bare metal bios).  
This has been
     rounded up to the nearest MB.
2) mmio_hole = 1024
     This might be a good size (512 may be just a little too small).
3) mmio_size = 512
     This is the size that you think has worked in the past.

I would also see if less guest ram (like 3.0 GiB) can still reproduce 
the issue.

    -Don Slutz

> Possibly useful information follows.
>
> Thanks,
>
> Steve
>