Linux Kernel to Include KVM Virtualization

It looks like the newest version of the Linux kernel (2.6.20) will include KVM, the relatively new virtualization environment. From the article: "Thanks to its approach KVM already runs in the current kernel, without any extensive bouts of patching and compiling being required, after the fairly simple compilation of a module. Virtual machines that run unmodified operating systems are meant to appear in the host as a simple process and work independently of the host kernel. In a fashion comparable to that of Xen a modified QEMU is used for the supportive emulation of typical PC components of the virtual machines."

Not frist psot!

[Aladrin]

Not first post, but at least I can be the first not to kvetch about them not integrating a physical object with a piece of software. -sigh-

The article talks about a news article mistakenly stating it was for Intel processors only. I imagine it said that because the official site says it's for Intel only. http://kvm.sourceforge.net/howto.html [sourceforge.net]

It does also say elsewhere on the site http://kvm.sourceforge.net/faq.html [sourceforge.net] that it's for certain AMDs also.

It claims it can run 32-bit windows inside the virtualization. Does this mean Windows can directly access the hardware, and provide true 3D support and such? Or is it simply another hardware emulator with all the associated problems? Too bad 'windows guest' installation is broken at the moment.

Excellent

[October_30th]

But is this going to let me run 32-bit Windows under 64-bit Linux? Apparently Xen can't do it and that really bugs me.

Instant kernel switching?

[Anonymous Coward]

Is this related to that new experimental Linux system call in the kernel that will allow instantaneous switching to another operating system/kernel?

Not everything, just video

[brunes69]

One of the main barriers to Linux adoptoin is the fact that you can't ru Windows games in Linux, unless you reboot into windows. If LVM / Xen / QEMU / VMWare started realizing this and made video driver performance a priority, they could have a real market leader on their hands.

I know if there was a VM out there that coudl run Windows games with full native windows video accelleration, I woudl pay very good money for it.

Sound / disk / CPU performance has been there in VMs for years, at least froma desktop users standpoint. The one area that lags behind all other sis video support. Even with VMWare (arguable the fastest VM out there right now), running a full scrteen Windows session under Linux feels sluggish at best...a nd there isno Direct3D support at all.

And as far as your comment - there is absolutely nothing stopping them from doing this. Just look at X, it interfaces direct with the kernel via DRI, and it's secure.. a crashing X session won't bring your whole machine down.

VMotion/HA?

[Stile 65]

The company I work for now is virtualizing on RedHat boxes running VMWare, and one of the neat features that it has is called VMotion, which lets you nearly instantly move a virtual machine from one box to another without interrupting its execution (except a slight delay). The high availability (HA) feature, which they also have but we have not yet configured, allows this to happen automatically if a host box goes down. There are rules about which VMs may not run on the same machines, etc. (for redundancy purposes, you don't want all your web servers running on the same host, for example).

Is this at all possible with KVM? If not, are they planning it? Is it possible to approximate it with something like OpenMosix, since (IIRC) OpenMosix can move processes around dynamically when nodes fail or get bogged down, and a VM is just a process (assuming a central SAN that all the host boxes connect to)?

Re:how many KVMs

[lintux]

And it's not even an KDE app! I used to think it was some KDE front-end to an existing virtualization program...

A W E S O M E ! ! !

[Anonymous Coward]

According to http://kvm.sourceforge.net/faq.html [sourceforge.net] is will support VMWare images and it does run win32.

Now turn that kernel into a BIOS http://linuxbios.org/Welcome_to_LinuxBIOS [linuxbios.org] and you will be able to use the same images for all your machines.

SVM for .net / Java

[0xABADC0DA]

Probably a joke but shouldn't be. The kernel developers could make it possible to accelerate the JVM/CLR by giving faster access to the actual hardware pages. What I mean is, the JVM has a 'scratch' area where recently created objects are allocated and then after garbage collecting this area the leftover objects are moved out. They do this because the vast majority of objects last only a tiny amount of time. So to be able to do a GC *only* of this scratch area the JVM actually replaces reference assignment with code that ads to a big list of all objects that took references to objects in this scratch area. Even though this is obviously slow it means the JVM doesn't have to garbage collect the entire contents of memory to make sure nothing has a ref to a new object.

If the JVM could get access to the hardware's dirty page bit that says if a page has been modified since last checked then the JVM could do direct reference assignment and then when garbage collecting only search the modified pages for references into the 'scratch' area. I expect this would be many times faster than the pointer write barrier used now.

Maybe a system call could take a mmap region and return a bitmask of page dirty flags? I think in any case there are plenty of things the kernel developers could do to make software virtual machines better if they tried. I think they just don't care to since that world is alien to them.