Re: [coLinux-devel] Re: Posix Subsystem for ReactOS - Project 2010

[Ian C. B. <ia...@bl...>]

On Thu, Jan 29, 2004 at 12:42:30AM +0200, Shachar Shemesh wrote:
> While it's certainly an exiting project, I understood that it was NOT 
> based on the Windows subsystem mechanism, but rather hacking your way 
> into Windows NT's ring 0 using a device driver. While I'm a great fan of 
> "if it works", wouldn't it be nicer, long run, to have a proper 
> subsystem of it?

Why work within Microsoft's cludgy POSIX subsystem shenanigans? There's
something very alluring to having a Linux kernel running at ring 0 in
parallel with OTHER kernels. Think of Linux X86 images running on a 
Linux X86 host without UML or VMWare, running Linux PPC images on a Darwin
box, or Linux Sparc images on a Solaris box - why limit this
architecture purely to Microsoft?

> Also, what are you planning on doing with graphic applications? CoLinux 
> works by running a X server on the windows machine. That's probably not 
> the best solution there is. How is the ReactOS windowing back-end 
> implemented?

I'm more interested in running headless servers myself. If I *really* need
a head, I have ssh tunnelled X11 or tight VNC sessions. 

To truely grasp the importance what CoLinux offers, think hugely
scalable hosting farm. The closest thing to CoLinux at the moment is
really the Xen Hypervisor:

	http://www.cl.cam.ac.uk/Research/SRG/netos/xen/

Unfortunately, Xen isn't cross architecture, and you need to retrofit
other kernels to run under it (see the Windows XP port limited to
academic source licensees, or the recent NetBSD port to the same).

At the moment, UML is the hosting platform of choice for virtual Linux
servers - but there are real speed and minor irritating stability issues.
Xen has its own LVM like virtual disk system in 1.2 - there is only
planned support for UBD style devices, and all domains must access the
disk simultaneously.

Both Plex86 and Fabrice Bellard's QEMU have slightly modified kernels
that can run across platforms and virtualize kernels in a somewhat 
similar way (though on a simulated or vitualized system without a ring 0
driver hack).

I would LOVE to have a Linux self-hosted, ring 0 device driver driven,
generic virtual machine subsystem for Linux that does not require
Jeff Dyke's user-space hacks to work (no offense intended toward Jeff
Dyke's incredible accomplishments with UML).

There are a suprising number of people quietly following this train of
thought. The early adopters are already running farms of hundreds (soon
to be thousands) of virtual machines in clustered Linux farms. If there
were a way to run CoLinux on an OpenSSI Linux cluster right now,
you better believe we would already be doing it.

- Ian C. Blenke <ia...@bl...>