[linux-elitists] Virtualization

Greg KH greg@kroah.com
Wed Feb 23 09:00:38 PST 2005


On Wed, Feb 23, 2005 at 12:46:03AM -0800, J. Paul Reed wrote:
> On 22 Feb 2005 at 22:36:41, Greg KH arranged the bits on my disk to say:
> > > > And if we do that, I see no reason why vmware can't use those drivers
> > > > too, so they can finally drop their weird add-on kernel module that is a
> > > > giant pain-in-the-butt for their users.
> > > 
> > > Uhm. Which one?
> > 
> > Heh, you are right, that should have been "modules", not "module".
> 
> There are multiple modules required for the Linux host installation and
> multiple optional modules for a Linux guest installation.
> 
> And the "standard modules" you refer to Xen building have nothing to do
> with the modules you're complaining about; they're building standard
> drivers for guest OS installations, and you're talking about the host
> installations.

Not true.  I'm talking about standard drivers on both sides.  I know
vmware is interested in this standardization based on some conversations
that I've had with some important people there in the past.

> > I mean the kernel modules that are necessary to load if you are using
> > Linux as the host os.  The ones that taint the kernel when loaded, and
> > drive you all crazy with every new kernel release as the different
> > kernel apis change all around you, with you all being helpless and
> > constantly playing catch up.  Quite fun to watch if you ask me.
> 
> Ahh... and Xen is different how? Oh right... they just modify the whole
> damn kernel (and require you to patch your entire kernel every rev).

No, Xen is going to make it into the mainline kernel soon, and then no
patching will be needed.  When will vmware achieve this goal?  :)

> The fact of the matter is, if you're not running a virtualization platform
> (a la ESX), and you're running a "hosted" virtualization product, you have
> to do this. UML does it. Xen does it. VMware does it.
> 
> Lord only knows why you're complaining about it only when VMware does it
> (and of course, glossing over the fact that all you have to do is recompile
> the kernel modules when you install a new kernel... you know... LIKE EVERY
> OTHER KERNEL MODULE YOU HAD TO RECOMPILE).

No, I complain as the vmware module taints my kenel, requires a separate
step to build from my normal kernel build process, and if I load it into
my kernel, I invalidate any and all support contracts I might have paid
for, and I can no longer ask for help with any kernel problems from
anyone in the open source kernel community.

I also complain that the module is fragile in that sometimes it doesn't
build properly due to the kernel apis changing all the time (and yes,
that's sometimes due to work that I do, see
http://www.kroah.com/log/linux/stable_api_nonsense.html for details as
to why this will not change.)

Xen will not have this same problem in a few months.  Nor does UML today
(yes, UML has other issues, I know about them too...)

> > ESX is the hacked up kernel that turns it into a hypervisor, right?
> > 2.4.9 based from what I recall last I looked.
> 
> Thank you for proving my point.
> 
> The correct answer is ESX is its own, proprietary kernel that was developed
> entirely in house, and has nothing to do with a 2.4.9 kernel.

Then why is a hacked up 2.4.9 kernel included in the ESX product?  See
http://vmware-svca.www.conxion.com/secured/esx/VMware-esx-public-source-2.5.0-11343.tar.gz
for details.

Is this kernel the guest os?  And if so, why can't someone use any
kernel they want as the guest os?  Or is it the host os as vmware is
taking advantage of Linux drivers?

Anyway, I'm not trying to argue here.  vmware is a great product, I use
it all the time to reverse engineer windows drivers :)

I just don't like the fact that their kernel modules are not included in
the main kernel tree.  I view Linux closed source modules as derivative
works, and feel they should be required to be released under the GPL.  I
know a lot of IP lawyers who also feel the same way.  See my, and other
kernel developers effort to help fix this issue in the recent kernel
releases (marking non-needed functions static, use of EXPORT_SYMBOL_GPL,
etc.)

I also begrudge the fact that the amount of development effort that
vmware gives back to the kernel community is pretty much non-existant,
in comparison to everything they are taking from it.  But for more
details on that topic, see my OLS paper for this year...

thanks,

greg k-h



More information about the linux-elitists mailing list