[uml-user] Strange behavior when trying to use UML with Yocto

[Marcelo E. M. <mar...@hp...>]

Hi,

I apologize in advance for the fuzzy details, but I'm hoping someone can 
ask the right questions to provide that missing information.

A couple of people are trying to integrate UML with a Yocto-based 
distribution called OpenSwitch (http://openswitch.net/). The relevant 
change request can be seen here:

https://review.openswitch.net/#/c/6079/

The OPS build process builds a full kernel and userspace, and Christian 
did a proof of concept integrating UML with that build process. 
Everything *seems* to be working fine, except that we are seeing strange 
behaviors at runtime.

Christian build the whole thing inside a VM, which has Debian 8 
installed. When he tries to boot the image, everything works fine.

When I try to reproduce Christian's results, I use a computer with 
Debian 8 installed in it (no VM), generate the image and boot. That 
alone seems to work, but several components inside the system exhibit 
strange behavior. OPS uses systemd, and journald starts spewing out 
messages about not being able to write a log entry. There's another 
component (OVSDB) which has a helper program create a database for it at 
boot time, and that component is behaving in a strange manner: it reads 
a file to translate it into a different format, and although strace does 
not reveal any errors, the output file is truncated (filesystem is 
ext4). There's also an issue with OpenSSH generating keys at boot time, 
and those keys are also truncated.

We have even tried with Christian's generated image on the computer 
where things fail, and it stills fails. I have generate an image out of 
a chroot to eliminate inconsistencies between the two build 
environments, and as far as we can tell, everything is the same, but it 
still fails. Christian tried building the same thing using actual 
hardware (no VM), and it fails. Weird thing is that if you take that 
image to the VM, it works.

Perhaps we are misunderstanding how UML works. Is it possible to run 
kernel version X (UML) on top of kernel version Y (X > Y)? How are 
system calls handled? I'm asking because we found a thread where someone 
asked the kernel headers to be installed so that a toolchain for the 
target environment can be generated, but the reply back then is that 
"that doesn't make sense". In our case the host is running 3.16 and the 
UML kernel is 4.5. The VM where things work is the same.

One thing we have noticed is that when the host is not a VM, systemd 
says the virtualization technology is UML. When the host is the VM, 
systemd says the virtualization technology is vmware. Scanning the code 
it does not seem to be doing anything different based on that.

I can provide the binaries that I compiled and are failing upon request 
if you want to take a look at the error yourself.

I'm looking forward to your questions regarding this,

Marcelo