Downplaying a serious issue

[Frank Leonhardt]

On 03/04/2014 12:36, Glen Barber wrote:
> On Thu, Apr 03, 2014 at 09:47:02AM +0200, Torbjorn Granlund wrote:
>> Glen Barber <gjb at FreeBSD.org> writes:
>>
>>    It is not a doc problem.
>>    
>>    The issue is specific to certain hardware configurations, and unless
>>    anyone has made any breakthroughs that I am unaware of, the cause is
>>    still unknown.
>>    
>> It happens on:
>>
>> AMD piledriver running Linux+KVM
>> AMD piledriver running Linux+Xen
>> Intel Nehalem running NetBSD+Xen
>> Intel Sandybridge running NetBSD+Xen
>> Intel Haswell running NetBSD+Xen
>> AMD K10 Barcelona running NetBSD+Xen
>> AMD Bulldozer running NetBSD+Xen
>>
> We need more specifics.
>
>> I've seen the laughable claim that this is a "bug in Virtualbox", and now
>> the major downplay at http://www.freebsd.org/releases/10.0R/errata.html,
>> where this is a minor hardware specific problem.
>>
>> I have not found one piece of PC hardware where it does not happen under
>> virtualisation.  Please let me know some configuration where FreeBSD/i386
>> works under a type 1 virtualiser?  Perhaps Bhyve is FreeBSD-compatible?
>>
> Does not happen on my VirtualBox host.
>
> Glen
>
I've been following this discussion with some alarm, but have now looked 
at the Errata:

------------------------------------------

"FreeBSD/i386 10.0-RELEASE running as a guest operating system 
onVirtualBoxcan have a problem with disk I/O access. It depends on some 
specific hardware configuration and does not depend on a specific 
version ofVirtualBoxor host operating system.

It causes various errors and makes FreeBSD quite unstable. Although the 
cause is still unclear, disabling unmapped I/O works as a workaround." 
etcetera

-------------------------------------

I don't read this as "down-playing" - it's up front about saying that 
there's a problem with every version of VirtualBox. It would, of course, 
be useful to add that it doesn't work with other named emulators too 
(for a virtual machine IS emulating the I/O hardware).

My concern was that this bug may be present on the real hardware too. I 
suspect more people would be running an i386 version as a VM than on 
real metal these days.

Does the problem exist on previous releases?

It seems to me that, after research, the list of confirmed incompatible 
configurations need to be expanded, especially to encompass other 
known-to-fail emulations. A list of "confirmed" problem environments 
would make readers wary about untested emulators too.

Incidentally, I don't see this as a bug in FreeBSD. A hypervisor is 
supposed to transparent to the OS, emulating the hardware that the OS 
thinks it has, to perfection. This is a broken VM, as clearly it's not 
behaving as the real hardware would. Or is it?

Regards, Frank.