Avi Kivity is the lead developer and maintainer of the Kernel-based Virtual Machine project, better known as kvm. The project was started in mid-2006, and has been part of the Linux kernel since the 2.6.20 release in February of 2007. kvm is a full virtualization system for x86-based Linux hosts, allowing users to run isolated x86 guest operating systems in virtual machines.
In this interview, Avi talks about the current status of the kvm project, detailing what kvm is, how it works, and how it differs from other virtualization projects. He goes on to examine future plans for improving stability and performance.
Jeremy Andrews: Please share a little about yourself and your background.
Avi Kivity: I’m a software developer living in Israel. I am 36 years old. Somewhat surprisingly, I studied aeronautical engineering, but I’ve designed very few aircraft, being instead swept away to the fast-paced field of software development.
I currently work at Qumranet Technologies, a start up company which develops kvm as well as a still-undisclosed proprietary product.
JA: Is the undisclosed product related to kvm and virtualization?
Avi Kivity: Yes, it uses kvm as the lowest layer in the stack.
JA: When did you get started with Linux?
Avi Kivity: About 1996 or 1997 our sysadmin installed Linux 2.0.something on a disused desktop, to act as a firewall. Out of curiosity I telnetted in and learned everything I could about it. I was already somewhat familiar with Unix, so it wasn’t that strange for me.
Some months later that little box took over a few services from our big NT server, including our all-important version control services. Visual SourceSafe decided to eat our files while I was branching them for an important demo, so I wrote a Perl script to check out the surviving files from SourceSafe and check them in to CVS. I babysat four or five workstations which were performing the conversion over the weekend, regularly restarting the NT server when it died. That little Linux box didn’t even break up a sweat.
JA: What is your involvement in the kvm (Kernel-based Virtual Machine) project?
Avi Kivity: I am the lead developer and maintainer of kvm. I also did what little design work was necessary (though most of kvm’s design is dictated by the hardware specs).
JA: What hardware is necessary to use kvm, and how prevalent is this hardware?
Avi Kivity: kvm currently requires an x86 processor that supports hardware virtualization. This means an Intel processor manufactured in the last two years, or a year-old AMD processor. I don’t know what the installed base of these processors is, but one thing is certain – it keeps increasing.
I’m hoping that a port of kvm to powerpc will materialize some day so that kvm will not be x86-only.
JA: How can a person tell if their hardware supports kvm?
Avi Kivity: You can easily tell if your processor supports hardware extensions by running the command
$ egrep '^flags.*(vmx|svm)' /proc/cpuinfo
If there’s any output at all, your processor supports kvm. Unfortunately, some early laptops that support kvm have VT extensions disabled in the BIOS, so a BIOS upgrade is required. There is also a comprehensive list on the Xen site, http://wiki.xensource.com/xenwiki/HVM_Compatible_Processors.
JA: What other architectures are you aware of that support hardware virtualization?
Avi Kivity: Xen supports ia64 (Itanium) and PowerPC in addition to x86, so at least these two architectures have the necessary hardware support.
JA: Does Qumranet Technologies have any intention to port kvm to other architectures?
Avi Kivity: No. These ports are best done by the hardware vendors who have the necessary expertise. Of course, we will cooperate and assist the ports in any way we can.
JA: What are the similarities and differences between Intel’s virtualization technology and AMD’s virtualization technology?
Avi Kivity: In general, the two technologies are similar. There are some differences, probably attributable to AMD being second to the market and thus being able to learn from Intel’s experience:
- AMD SVM is generally much simpler to program
- AMD SVM contains support for virtualizing real mode (which is used in the boot sequence); therefore, AMD hosts have better support for boot loaders
- AMD SVM allows the guest TLB to coexist with the host TLB, which improves performance; in contrast, on an Intel host a guest/host switch flushes the TLB
We will probably see a game of leapfrog being played in this area, as each vendor will try to increase its attractiveness by adding features and improving performance.
Compared to other virtualization projects:
JA: What is kvm, and how does it compare to other virtualization projects?
Avi Kivity: kvm is (now) a full virtualization system for x86-based Linux hosts. It allows users to create virtual machines and run various x86 operating systems as isolated guests, similar to other virtualization and emulation solutions, for example VMware, Xen, and Qemu.
To compare kvm to other solutions:
- In many ways, VMware is a groundbreaking technology. VMware manages to fully virtualize the notoriously complex x86 architecture using software techniques only, and to achieve very good performance and stability. As a result, VMware is a very large and complex piece of software.kvm, on the other hand, relies on the new hardware virtualization technologies that have appeared recently. As such, it is very small (about 10,000 lines) and relatively simple. Another big difference is that VMware is proprietary while kvm is open source.
- Xen is a fairly large project, providing both paravirtualization and full virtualization. It is designed as a standalone kernel, which only requires Linux to perform I/O. This makes it rather large, as it has its own scheduler, memory manager, timer handling, and machine initialization.kvm, in contrast, uses the standard Linux scheduler, memory management, and other services. This allows the kvm developers to concentrate on virtualization, building on the core kernel instead of replacing it.
- Qemu is a userspace emulator. It is a fairly amazing project, emulating a variety of guest processors on several host processors, with fairly decent performance. However, the userspace architecture does not allow it to approach native speeds without a kernel accelerator. kvm recognizes the utility of qemu by using it for I/O hardware emulation. Although kvm is not tied to any particular userspace, the qemu code was too good not to use — so we used it.
JA: Why is kvm able to use the standard Linux scheduler, memory management and other services, while the Xen project had to write its own?
Avi Kivity: Xen started out as a paravirtualization system, without hardware support, back when 32-bit x86 was the predominant architecture. Now, if you want to fit a host kernel, a guest kernel, and guest userspace into a 4GB address space, you’ve got to make the host kernel very small – a microkernel. The Xen developers chose to fit the host kernel into the top 64MB of memory. Since Linux won’t fit into such a small space, they had to write a new kernel.
In contrast, the hardware virtualization support utilized by kvm has explicit support for switching between a host address space and a guest address space, making a specialized kernel unnecessary.
JA: Is kvm a complete replacement for Xen, or are there still things that Xen can do that kvm does not and may never do?
Avi Kivity: At this point in time, Xen has some features that kvm does not: guest SMP support, running without hardware extensions, more paravirtualized device drivers for example. However, we expect to close the gap rapidly; in fact in some areas kvm is ahead of Xen: kvm can do live migration of fully virtualized guests whereas Xen is still developing this feature. Laptop users also appreciate the ability to suspend-to-disk a machine with several guests running and have it resume.
It is quite possible that one day kvm will gain the ability to run without hardware support; Rusty Russell’s lguest has shown that it can be done.
JA: What I/O hardware emulation does kvm require from qemu?
Avi Kivity: We use qemu for emulating the motherboard hardware: the pci bus, memory controller, ROM BIOS, VGA graphics, sound card, IDE controller, and network interface card. We also use qemu for emulating the processor’s local APIC.
JA: What user interfaces are available for kvm?
Avi Kivity: Qemu provides a fairly simple user interface. More elaborate user interfaces are works in progress: for example, Fedora’s virt-manager is slated to support kvm in Fedora 7.
JA: How long has kvm been under development?
Avi Kivity: kvm development started the first half of 2006. We got Linux x86_64 to boot about two months into development, and 32-bit Windows about a week later.
The remaining time was spent on adding AMD host support, performance improvements, and stability and correctness.
JA: What sorts of feedback did you get from kernel developers when you first submitted patches against the kernel?
Avi Kivity: I generally received very positive feedback. Virtualization is an important topic these days, and kvm addresses it in the Linux way, as a minimally intrusive driver.
JA: Is kvm’s full functionality merged into the 2.6.20 kernel, or just some parts of it?
Avi Kivity: All of kvm is continuously merged into the kernel. There’s usually some lag between our source repository and Linus’ tree, but until now it was pretty short. I expect the lag to grow a little as kvm matures and starts to fit into the Linux development cycle, however we will keep it as short as possible so users can enjoy the latest and greatest in virtualization.
JA: What future plans do you have for kvm?
Avi Kivity: Besides the usual performance and stability improvements, I’d like to see paravirtualization support (as started by Ingo Molnar) become a major feature of kvm.
Initially I felt that kvm and paravirtualization had too little in common to be merged, however after looking at Rusty Russell’s lguest I was convinced that duplicating all the delicate mmu work was wasted effort and that the projects should merge.
JA: What is lguest, and how is it related to kvm?
Avi Kivity: lguest (born lhype) is the paravirtualized equivalent of kvm. Where kvm virtualizes an OS-agnostic hardware platform, lguest is specifically oriented to virtualize Linux guests. Because the guest and host cooperate, lguest can potentially be very fast, though at present it has not yet realized this potential.
JA: Can you describe what paravirtualization is, and what’s involved in getting kvm to support paravirtualization?
Avi Kivity: Paravirtualization is a technique that is used to improve virtualization performance. By making the guest aware that it is running on a virtualized platform, the guest and host can cooperate, using more efficient methods of communication.
One example is using paravirtualized drivers. Instead of emulating a real network interface card, we use a software interface designed for performance. Preliminary numbers from Ingo Molnar show a > 20X improvement in bandwidth over the qemu emulated NIC.
JA: How stable is kvm at this time?
Avi Kivity: I find kvm to be very stable: if it runs your guest, it won’t crash. There are still issues with guest support (e.g. some guests won’t boot, mostly due to real-mode emulation problems on Intel hosts), and, rarely, some hosts refuse to work with kvm.
JA: In what ways can kvm’s performance be improved?
Avi Kivity: kvm has not yet undergone major performance tuning. Some issues that I’d like to address are:
- avoid saving and loading the floating point unit’s state on every guest/host switch, using a technique similar to the kernel’s lazy fpu mode
- use a better page replacement algorithm when evicting shadow page tables
- when the guest modifies a page table entry, do the corresponding modification to the shadow page table entry
JA: What limitations does kvm impose upon guest operating systems?
Avi Kivity: kvm tries not to impose any limitations on the guest other than the type of virtual hardware provided. Right now we can only virtualize a single processor, which is a fairly significant limitation. Bugs in real-mode emulation mean that some guests may not boot at all.
JA: How long before kvm will be able to virtualize multiple processors?
Avi Kivity: As with most open source projects, it is difficult to give a timeframe. I’d like to start work on guest SMP when processors that support nested page tables appear; that allows work on the core SMP support to be done without regard to the mmu (which is the most complex code in kvm). Once that’s done, we can add the necessary locking and cross-cpu TLB management to the mmu and support older processors.
JA: What are some guests that are not able to boot?
Avi Kivity: Typically guests that use a boot splashscreen in grub fail on Intel hosts; Ubuntu and opensuse are examples. In addition, ACPI support is incomplete, so 64-bit Windows does not work right now.
In some cases, workarounds are available (for example, disabling the boot splashscreen).
JA: What are some examples of bugs that prevent guests from being able to boot?
Avi Kivity: The major cause of problems is the lack of real mode support in Intel’s virtualization extensions. When an x86 processor starts up, it executes an instruction set and processing environment that is compatible with the 16-bit 8086 family; this is called “real mode” to distinguish it from the “protected mode” in which modern operating systems such as Linux execute. Real mode is only used briefly during the boot stage.
As the hardware doesn’t support virtualized real mode fully, kvm has to emulate it. Unfortunately, our emulator is incomplete and thus there are some guests that cannot boot.
I am happy to say that this issue is being addressed and we hope to have it resolved in the near future.
I should also note that this issue does not affect AMD hosts, which support real mode virtualization in hardware.
JA: What other Linux kernel projects have you worked on, before kvm?
Avi Kivity: Most of my previous work was in userspace. I did some hacking on Linux asynchronous I/O as part of a larger project a few years ago.
JA: How do you enjoy spending time when you’re not hacking on the Linux kernel?
Avi Kivity: I like going to the movies, listening to music, reading, playing backgammon (online and in person), table tennis (in person only!). I’ve also taken up snowboarding recently.
JA: What advice would you offer to readers who are only begging to get interested in the inner workings of the Linux Kernel.
Avi Kivity: Read lkml! There is lots of insight to be gained from the posts there.
Also, pick an area you have a personal interest in and study it.
JA: Is there anything else you’d like to add?
Avi Kivity: One of the benefits of working on the Linux kernel is that you get to work with some very brilliant developers. It’s quite an enjoyable experience (barring the occasional flame), and I learned quite a lot from it. I’d like to thank the kvm user and developer community for the testing, feedback, and of course, patches.
JA: Thank you for your time answering these questions, and for all your efforts on the kvm project!