ZFS corruption using zvols as backingstore for hvm VM's

[Roger Pau Monné]

On Thu, Nov 12, 2020 at 09:18:59PM -0800, Brian Buhrow wrote:
> 	hello Roger.  thanks for engaging with me on this issue.  I think
> I've made progress on the issue and have a better handle on what's going
> wrong.  There seem to be a cascade of bugs here, which I'll try to
> enumerate.  
> 
> 1.  The disk corruption issue seems to be a bug in qemu whereby the
> emulated IDE disk controller issues partial writes instead of full writes
> or no writes with appropriate failure to the disk.  The IDE driver in
> NetBSD-5.2 doesn't play well with this behavior, in fact, NetBSD until May
> of 2020, doesn't play well with this behavior
> See: 
> http://mail-index.NetBSD.org/source-changes/2020/05/24/msg117668.html

Oh great, so it's something specific to NetBSD. This means you are
running NetBSD in HVM mode?

> 2.  This causes memory corruption in the OS itself, which can trigger a xen
> server crash!  (In my view, no matter how badly behaved the guest OS is, it
> shouldn't be able to bring down the xen server.)

Right, we really need the trace from this crash. Do you have a serial
hooked up to the box so that you can provide the panic message?

> 3.  Running NetBSD-5.2/i386 as a domu, which works flawlessly under xen3, gets a 
> panic: HYPERVISOR_mmu_update failed

This would imply that you are running NetBSD in PV mode, in which case
it won't be using the emulated hard disk drive, and hence the commit
you referenced above would be unrelated. Can you assert whether you
are running NetBSD in PV or HVM mode?

> 	I suspect this can be worked around using some of the command line
> options under Xen, xpti=true,domu=false, perhaps?  
> Are there others I should consider?

Maybe. I think you should report this to xen-devel and NetBSD/Xen
mailing lists, with a full trace of the crash and the guest config
file.

Roger.