Skip to contentSkip to conversation

VMI i386 Linux virtualization interface proposal

RFC PATCH In review Zachary Amsden 11 participants 30 messages Mar 13, 2006 – Mar 16, 2006

lore
Share & tools
lore
#0 lore raw

cover VMI i386 Linux virtualization interface proposal

In OLS 2005, we described the work that we have been doing in VMware with respect a common interface for paravirtualization of Linux. We shared the general vision in Rik's virtualization BoF.

This note is an update on our further work on the Virtual Machine Interface, VMI. The patches provided have been tested on 2.6.16-rc6. We are currently recollecting performance information for the new -rc6 kernel, but expect our numbers to match previous results, which showed no impact whatsoever on macro benchmarks, and nearly neglible impact on microbenchmarks.

Unlike the full-virtualization techniques used in the traditional VMware products, paravirtualization is a technique where the operating system is modified to enlighten the hypervisor with timely knowledge about the operating system's activities. Since the hypervisor now depends on the kernel to tell it about common idioms etc, it does not need to write protect OS objects such as page and descriptor tables as a solution based on full-virtualization needs. This has two important effects (a) it shortens the critical path, since faulting is expensive on modern processors (b) by eliminating complex heuristics the hypervisor is simplified. While the former delivers performance, the latter is quite important too.

Not surprisingly, paravirtualization's strength, ie that it encourages tighter communication between the kernel and the hypervisor, is also its weakness. Unless the changes to the operating system are moderated, you can very quickly find yourself with a kernel that (a) looks and feels like a brand new kernel or (b) cannot run on native machines or on newer versions of the hypervisor without a full recompile. The former can impede innovation in the Linux kernel, and the latter can be a problem for software vendors.

VMware proposes VMI as a paravirtualization interface for Linux that solves these problems.
- A VMI'fied Linux kernel runs unmodified on native hardware, and on many hypervisors, while simultaneously delivering on the performance promise of paravirtualization.
- VMI has a rich and low level interface, which allows the kernel to cope with future hardware evolution by querying for hardware capability. It is our expectation that a single kernel will run unmodified on both today's processors with limited hardware virtualization support and also keep up with any evolution on the processor front
- VMI Linux is a fairly clean interface, with distinct name spaces for objects from the kernel and the hypervisor. Nowhere do we mingle names from the hypervisor with that of the kernel. This separation allows innovation in the kernel to proceed at the same speed as always. For most kernel developers, a VMI kernel looks and feels like a regular Linux kernel.
- VMI Linux still supports "native" hypervisor device drivers, for example a hypervisor vendor's own private network or block device drivers which are free to use any interface desired to communicate with the hypervisor.

At present, we are sharing a working implementation of the VMI for 2.6.16-rc6 version of Linux. We have verified that VMI Linux does indeed run well on native machines (both P4 and Opterons), and on VMware style hypervisors. VMI Linux has negligible overheads on native machines, so much so, that we are confident that VMI Linux can, in the long run, be the default Linux for i386. We believe that this interface is both cleaner and more powerful than other proposals that have been made towards virtualization of Linux, and can easily be adapted to work with other hypervisors.

This is by no means finished work. A few of the areas that need more attention and exploration are (a) 64bit support is still lacking, but we feel a port of VMI to the 64 bit Linux can be done without too much trouble (b) the Xen compatibility layer needs some work to bring it up to the Xen 3.0 interfaces. Work is underway on this already, and no major issues are expected at this time.

Two final notes. This is not an attempt to force a proprietary interface into the Linux kernel. This is an attempt to find a common interface that can be used by many hypervisors by isolating hypervisor specific idioms into a neutral layer. This new layer is just what is claims to be - a virtual machine interface, which allows hypervisor dependent code to be abstracted in a way that benefits both Linux and hypervisor development.

This is also not an attempt to define an exact and final specification of how virtualization should be done in Linux. This is very much a work in progress, and it is understood that the interfaces proposed here will change in time to accommodate the needs of all interested parties. We hope to find a common solution that can eventually become part of the Linux kernel and serve as a model for other operating systems as well.

We appreciate your feedback on this design and the patches to Linux, and welcome working with anyone who is interested in making virtualization in Linux a friendly environment to innovate in. If you find the ideas here interesting, please volunteer to help improve them.

#1 lore raw
Quoted 7 lines

Two final notes. This is not an attempt to force a proprietary interface into the Linux kernel. This is an attempt to find a common interface that can be used by many hypervisors by isolating hypervisor specific idioms into a neutral layer. This new layer is just what is claims to be - a virtual machine interface, which allows hypervisor dependent code to be abstracted in a way that benefits both Linux and hypervisor development.

such an interface should be defined with source visibility of both sides though. At least of one user. Can XEN or any of the other open hypervisors use this? What does it look like? And if not, why not, wouldn't that make VMA a VMwareInterface instead ? ;)

Why can't vmware use the Xen interface instead?

#2 lore raw
Quoted 10 lines · Arjan van de Ven
Quoted 8 lines

Two final notes. This is not an attempt to force a proprietary interface into the Linux kernel. This is an attempt to find a common interface that can be used by many hypervisors by isolating hypervisor specific idioms into a neutral layer. This new layer is just what is claims to be - a virtual machine interface, which allows hypervisor dependent code to be abstracted in a way that benefits both Linux and hypervisor development.

Quoted 5 lines

such an interface should be defined with source visibility of both sides though. At least of one user. Can XEN or any of the other open hypervisors use this? What does it look like? And if not, why not, wouldn't that make VMA a VMwareInterface instead ? ;)

Yes, Xen can use this interface, even without modification to Xen. The interface was used successfully to run a VMI kernel on Xen 2.0. As it stands now, the interface does need to change a bit to accomodate Xen 3.0 - but it is possible to do. Rather than wait until we have a working prototype of that, we thought the interface itself warrants discussion now.

Quoted 2 lines

Why can't vmware use the Xen interface instead?

We could. But it is our opinion that the Xen interface is unnecessarily complicated, without a clean separation between the layer of interaction with the hypervisor and the kernel proper. The interface we propose we believe is more powerful, and more conducive to performance optimizations while providing significant advantages - most specifically, a single binary image that is properly virtualizable on multiple hypervisors and capable of running on native hardware.

Zach

#3 lore raw
Quoted 5 lines

The interface we propose we
believe is more powerful, and more conducive to performance optimizations while providing significant advantages - most specifically, a single binary image that is properly virtualizable on multiple hypervisors and capable of running on native hardware.

that is mostly an advantage in the binary would though.. less so in the open source world.

#4 lore raw
Quoted 7 lines · Arjan van de Ven
Quoted 3 lines

The interface we propose we
believe is more powerful, and more conducive to performance optimizations while providing significant advantages - most specifically, a single binary image that is properly virtualizable on multiple hypervisors and capable of running on native hardware.

that is mostly an advantage in the binary would though.. less so in the open source world.

It is an advantage for everyone. It cuts support and certification costs for Linux distributors, software vendors, makes debugging and development easier, and gives hypervisors room to grow while maintaining binary compatibility with already released kernels.

Zach

#5 lore raw
Quoted 8 lines · Zachary Amsden
Quoted 7 lines · Arjan van de Ven
Quoted 6 lines

The interface we propose we
believe is more powerful, and more conducive to performance optimizations while providing significant advantages - most specifically, a single binary image that is properly virtualizable on multiple hypervisors and capable of running on native hardware.

Quoted 6 lines
Quoted 3 lines

that is mostly an advantage in the binary would though.. less so in the open source world.

It is an advantage for everyone. It cuts support and certification costs for Linux distributors,

that I'll buy

Quoted 1 line

software vendors,

that I'll buy a lot less except those with kernel modules (which is evil ;)

Quoted 2 lines

makes debugging and
development easier,

that I don't buy; a fixed interface tends to make debugging harder not easier since you can't change it to add more information

Quoted 2 lines

and gives hypervisors room to grow while maintaining
binary compatibility with already released kernels.

that I buy for binary only hypervisors. But in an open source world I'll buy this a LOT less as being relevant.

#6 lore raw
Quoted 6 lines · Arjan van de Ven
Quoted 3 lines

makes debugging and development easier,

that I don't buy; a fixed interface tends to make debugging harder not easier since you can't change it to add more information

This we find to be quite true. Now, you can use a VMI kernel, make changes to it, run it on native hardware, and be confident that it will run properly in a VM as well. And you can develop in a VM, with confidence that you can run on native hardware. You can even replace the entire "ROM" image with your own custom debugging image to add any type of debugging or performance monitoring facility you want - and you have some very, very interesting hook points into the kernel that make that task much more achievable.

Quoted 3 lines

that I buy for binary only hypervisors. But in an open source world I'll buy this a LOT less as being relevant.

This is not about the open source versus the closed source world. It is about the real world, where customers want to make as few changes as possible to a working and already deployed system. If they have to recompile a kernel just to get their system to run again, that is a pain point that is easily avoided.

Zach

#8 lore raw

VMI interface documentation

In reply to Arjan van de Ven

Sorry for the spam if you got multiples - vger doesn't seem to like my documentation. I've attached it here.

Paravirtualization API Version 2.0

Zachary Amsden, Daniel Arai, Daniel Hecht, Pratap Subrahmanyam Copyright (C) 2005, 2006, VMware, Inc.
All rights reserved

Revision history:
1.0: Initial version
1.1: arai 2005-11-15
Added SMP-related sections: AP startup and Local APIC support
1.2: dhecht 2006-02-23
Added Time Interface section and Time related VMI calls

Contents

1) Motivations
2) Overview
Initialization
Privilege model
Memory management
Segmentation
Interrupt and I/O subsystem
IDT management
Transparent Paravirtualization
3rd Party Extensions
AP Startup
State Synchronization in SMP systems
Local APIC Support
Time Interface
3) Architectural Differences from Native Hardware
4) ROM Implementation
Detection
Data layout
Call convention
PCI implementation

Appendix A - VMI ROM low level ABI
Appendix B - VMI C prototypes
Appendix C - Sensitive x86 instructions

1) Motivations

There are several high level goals which must be balanced in designing an API for paravirtualization. The most general concerns are:

Portability - it should be easy to port a guest OS to use the API
High performance - the API must not obstruct a high performance
hypervisor implementation
Maintainability - it should be easy to maintain and upgrade the guest
OS
Extensibility - it should be possible for future expansion of the
API

Portability.

The general approach to paravirtualization rather than full virtualization is to modify the guest operating system. This means there is implicitly some code cost to port a guest OS to run in a paravirtual environment. The closer the API resembles a native platform which the OS supports, the lower the cost of porting. Rather than provide an alternative, high level interface for this API, the approach is to provide a low level interface which encapsulates the sensitive and performance critical parts of the system. Thus, we have direct parallels to most privileged instructions, and the process of converting a guest OS to use these instructions is in many cases a simple replacement of one function for another. Although this is sufficient for CPU virtualization, performance concerns have forced us to add additional calls for memory management, and notifications about updates to certain CPU data structures. Support for this in the Linux operating system has proved to be very minimal in cost because of the already somewhat portable and modular design of the memory management layer.

High Performance.

Providing a low level API that closely resembles hardware does not provide any support for compound operations; indeed, typical compound operations on hardware can be updating of many page table entries, flushing system TLBs, or providing floating point safety. Since these operations may require several privileged or sensitive operations, it becomes important to defer some of these operations until explicit flushes are issued, or to provide higher level operations around some of these functions. In order to keep with the goal of portability, this has been done only when deemed necessary for performance reasons, and we have tried to package these compound operations into methods that are typically used in guest operating systems. In the future, we envision that additional higher level abstractions will be added as an adjunct to the low-level API. These higher level abstractions will target large bulk operations such as creation, and destruction of address spaces, context switches, thread creation and control.

Maintainability.

In the course of development with a virtualized environment, it is not uncommon for support of new features or higher performance to require radical changes to the operation of the system. If these changes are visible to the guest OS in a paravirtualized system, this will require updates to the guest kernel, which presents a maintenance problem. In the Linux world, the rapid pace of development on the kernel means new kernel versions are produced every few months. This rapid pace is not always appropriate for end users, so it is not uncommon to have dozens of different versions of the Linux kernel in use that must be actively supported. To keep this many versions in sync with potentially radical changes in the paravirtualized system is not a scalable solution. To reduce the maintenance burden as much as possible, while still allowing the implementation to accommodate changes, the design provides a stable ABI with semantic invariants. The underlying implementation of the ABI and details of what data or how it communicates with the hypervisor are not visible to the guest OS. As a result, in most cases, the guest OS need not even be recompiled to work with a newer hypervisor. This allows performance optimizations, bug fixes, debugging, or statistical instrumentation to be added to the API implementation without any impact on the guest kernel. This is achieved by publishing a block of code from the hypervisor in the form of a ROM. The guest OS makes calls into this ROM to perform privileged or sensitive actions in the system.

Extensibility.

In order to provide a vehicle for new features, new device support, and general evolution, the API uses feature compartmentalization with controlled versioning. The API is split into sections, with each section having independent versions. Each section has a top level version which is incremented for each major revision, with a minor version indicating incremental level. Version compatibility is based on matching the major version field, and changes of the major version are assumed to break compatibility. This allows accurate matching of compatibility. In the event of incompatible API changes, multiple APIs may be advertised by the hypervisor if it wishes to support older versions of guest kernels. This provides the most general forward / backward compatibility possible. Currently, the API has a core section for CPU / MMU virtualization support, with additional sections provided for each supported device class.

2) Overview

Initialization.

Initialization is done with a bootstrap loader that creates the "start of day" state. This is a known state, running 32-bit protected mode code with paging enabled. The guest has all the standard structures in memory that are provided by a native ROM boot environment, including a memory map and ACPI tables. For the native hardware, this bootstrap loader can be run before the kernel code proper, and this environment can be created readily from within the hypervisor for the virtual case. At some point, the bootstrap loader or the kernel itself invokes the initialization call to enter paravirtualized mode.

Privilege Model.

The guest kernel must be modified to run at a dynamic privilege level, since if entry to paravirtual mode is successful, the kernel is no longer allowed to run at the highest hardware privilege level. On the IA-32 architecture, this means the kernel will be running at CPL 1-2, and with the hypervisor running at CPL0, and user code at CPL3. The IOPL will be lowered as well to avoid giving the guest direct access to hardware ports and control of the interrupt flag.

This change causes certain IA-32 instructions to become "sensitive", so additional support for clearing and setting the hardware interrupt flag are present. Since the switch into paravirtual mode may happen dynamically, the guest OS must not rely on testing for a specific privilege level by checking the RPL field of segment selectors, but should check for privileged execution by performing an (RPL != 3 && !EFLAGS_VM) comparison. This means the DPL of kernel ring descriptors in the GDT or LDT may be raised to match the CPL of the kernel. This change is visible by inspecting the segments registers while running in privileged code, and by using the LAR instruction.

The system also cannot be allowed to write directly to the hardware GDT, LDT, IDT, or TSS, so these data structures are maintained by the hypervisor, and may be shadowed or guest visible structures. These structures are required to be page aligned to support non-shadowed operation.

Currently, the system only provides for two guest security domains, kernel (which runs at the equivalent of virtual CPL-0), and user (which runs at the equivalent of virtual CPL-3, with no hardware access). Typically, this is not a problem, but if a guest OS relies on using multiple hardware rings for privilege isolation, this interface would need to be expanded to support that.

Memory Management.

Since a virtual machine typically does not have access to all the physical memory on the machine, there is a need to redefine the physical address space layout for the virtual machine. The spectrum of possibilities ranges from presenting the guest with a view of a physically contiguous memory of a boot-time determined size, exactly what the guest would see when running on hardware, to the opposite, which presents the guest with the actual machine pages which the hypervisor has allocated for it. Using this approach requires the guest to obtain information about the pages it has from the hypervisor; this can be done by using the memory map which would normally be passed to the guest by the BIOS.

The interface is designed to support either mode of operation. This allows the implementation to use either direct page tables or shadow page tables, or some combination of both. All writes to page table entries are done through calls to the hypervisor interface layer. The guest notifies the hypervisor about page tables updates, flushes, and invalidations through API calls.

The guest OS is also responsible for notifying the hypervisor about which pages in its physical memory are going to be used to hold page tables or page directories. Both PAE and non-PAE paging modes are supported. When the guest is finished using pages as page tables, it should release them promptly to allow the hypervisor to free the page table shadows. Using a page as both a page table and a page directory for linear page table access is possible, but currently not supported by our implementation.

The hypervisor lives concurrently in the same address space as the guest operating system. Although this is not strictly necessary on IA-32 hardware, performance would be severely degraded if that were not the case. The hypervisor must therefore reserve some portion of linear address space for its own use. The implementation currently reserves the top 64 megabytes of linear space for the hypervisor. This requires the guest to relocate any data in high linear space down by 64 megabytes. For non-paging mode guests, this means the high 64 megabytes of physical memory should be reserved. Because page tables are not sensitive to CPL, only to user/supervisor level, the hypervisor must combine segment protection to ensure that the guest can not access this 64 megabyte region.

An experimental patch is available to enable boot-time sizing of the hypervisor hole.

Segmentation.

The IA-32 architecture provides segmented virtual memory, which can be used as another form of privilege separation. Each segment contains a base, limit, and properties. The base is added to the virtual address to form a linear address. The limit determines the length of linear space which is addressable through the segment. The properties determine read/write, code and data size of the region, as well as the direction in which segments grow. Segments are loaded from descriptors in one of two system tables, the GDT or the LDT, and the values loaded are cached until the next load of the segment. This property, known as segment caching, allows the machine to be put into a non-reversible state by writing over the descriptor table entry from which a segment was loaded. There is no efficient way to extract the base field of the segment after it is loaded, as it is hidden by the processor. In a hypervisor environment, the guest OS can be interrupted at any point in time by interrupts and NMIs which must be serviced by the hypervisor. The hypervisor must be able to recreate the original guest state when it is done servicing the external event.

To avoid creating non-reversible segments, the hypervisor will forcibly reload any live segment registers that are updated by writes to the descriptor tables. *N.B - in the event that a segment is put into an invalid or not present state by an update to the descriptor table, the segment register must be forced to NULL so that reloading it will not cause a general protection fault (#GP) when restoring the guest state. This may require the guest to save the segment register value before issuing a hypervisor API call which will update the descriptor table.*

Because the hypervisor must protect its own memory space from privileged code running in the guest at CPL1-2, descriptors may not provide access to the 64 megabyte region of high linear space. To achieve this, the hypervisor will truncate descriptors in the descriptor tables. This means that attempts by the guest to access through negative offsets to the segment base will fault, so this is highly discouraged (some TLS implementations on Linux do this). In addition, this causes the truncated length of the segment to become visible to the guest through the LSL instruction.

Interrupt and I/O Subsystem.

For security reasons, the guest operating system is not given control over the hardware interrupt flag. We provide a virtual interrupt flag that is under guest control. The virtual operating system always runs with hardware interrupts enabled, but hardware interrupts are transparent to the guest. The API provides calls for all instructions which modify the interrupt flag.

The paravirtualization environment provides a legacy programmable interrupt controller (PIC) to the virtual machine. Future releases will provide a virtual interrupt controller (VIC) that provides more advanced features.

In addition to a virtual interrupt flag, there is also a virtual IOPL field which the guest can use to enable access to port I/O from userspace for privileged applications.

Generic PCI based device probing is available to detect virtual devices. The use of PCI is pragmatic, since it allows a vendor ID, class ID, and device ID to identify the appropriate driver for each virtual device.

IDT Management.

The paravirtual operating environment provides the traditional x86 interrupt descriptor table for handling external interrupts, software interrupts, and exceptions. The interrupt descriptor table provides the destination code selector and EIP for interruptions. The current task state structure (TSS) provides the new stack address to use for interruptions that result in a privilege level change. The guest OS is responsible for notifying the hypervisor when it updates the stack address in the TSS.

Two types of indirect control flow are of critical importance to the performance of an operating system. These are system calls and page faults. The guest is also responsible for calling out to the hypervisor when it updates gates in the IDT. Making IDT and TSS updates known to the hypervisor in this fashion allows efficient delivery through these performance critical gates.

Transparent Paravirtualization.

The guest operating system may provide an alternative implementation of the VMI option rom compiled in. This implementation should provide implementations of the VMI calls that are suitable for running on native x86 hardware. This code may be used by the guest operating system while it is being loaded, and may also be used if the operating system is loaded on hardware that does not support paravirtualization.

When the guest detects that the VMI option rom is available, it replaces the compiled-in version of the rom with the rom provided by the platform. This can be accomplished by copying the rom contents, or by remapping the virtual address containing the compiled-in rom to point to the platform's ROM. When booting on a platform that does not provide a VMI rom, the operating system can continue to use the compiled-in version to run in a non-paravirtualized fashion.

3rd Party Extensions.

If desired, it should be possible for 3rd party virtual machine monitors to implement a paravirtualization environment that can run guests written to this specification.

The general mechanism for providing customized features and capabilities is to provide notification of these feature through the CPUID call, and allowing configuration of CPU features through RDMSR / WRMSR instructions. This allows a hypervisor vendor ID to be published, and the kernel may enable or disable specific features based on this id. This has the advantage of following closely the boot time logic of many operating systems that enables certain performance enhancements or bugfixes based on processor revision, using exactly the same mechanism.

An exact formal specification of the new CPUID functions and which functions are vendor specific is still needed.

AP Startup.

Application Processor startup in paravirtual SMP systems works a bit differently than in a traditional x86 system.

APs will launch directly in paravirtual mode with initial state provided by the BSP. Rather than the traditional init/startup IPI sequence, the BSP must issue the init IPI, a set application processor state hypercall, followed by the startup IPI.

The initial state contains the AP's control registers, general purpose registers and segment registers, as well as the IDTR, GDTR, LDTR and EFER. Any processor state not included in the initial AP state (including x87 FPRs, SSE register states, and MSRs other than EFER), are left in the poweron state.

The BSP must construct the initial GDT used by each AP. The segment register hidden state will be loaded from the GDT specified in the initial AP state. The IDT and (if used) LDT may either be constructed by the BSP or by the AP.

Similarly, the initial page tables used by each AP must also be constructed by the BSP.

If an AP's initial state is invalid, or no initial state is provided before a start IPI is received by that AP, then the AP will fail to start. It is therefore advisable to have a timeout for waiting for AP's to start, as is recommended for traditional x86 systems.

See VMI_SetInitialAPState in Appendix A for a description of the VMI_SetInitialAPState hypercall and the associated APState data structure.

State Synchronization In SMP Systems.

Some in-memory data structures that may require no special synchronization on a traditional x86 systems need special handling when run on a hypervisor. Two of particular note are the descriptor tables and page tables.

Each processor in an SMP system should have its own GDT and LDT. Changes to each processor's descriptor tables must be made on that processor via the appropriate VMI calls. There is no VMI interface for updating another CPU's descriptor tables (aside from VMI_SetInitialAPState), and the result of memory writes to other processors' descriptor tables are undefined.

Page tables have slightly different semantics than in a traditional x86 system. As in traditional x86 systems, page table writes may not be respected by the current CPU until a TLB flush or invlpg is issued. In a paravirtual system, the hypervisor implementation is free to provide either shared or private caches of the guest's page tables. Page table updates must therefore be propagated to the other CPUs before they are guaranteed to be noticed.

In particular, when doing TLB shootdown, the initiating processor must ensure that all deferred page table updates are flushed to the hypervisor, to ensure that the receiving processor has the most up-to-date mapping when it performs its invlpg.

Local APIC Support.

A traditional x86 local APIC is provided by the hypervisor. The local APIC is enabled and its address is set via the IA32_APIC_BASE MSR, as usual. APIC registers may be read and written via ordinary memory operations.

For performance reasons, higher performance APIC read and write interfaces are provided. If possible, these interfaces should be used to access the local APIC.

The IO-APIC is not included in this spec, as it is typically not performance critical, and used mainly for initial wiring of IRQ pins. Currently, we implement a fully functional IO-APIC with all the capabilities of real hardware. This may seem like an unnecessary burden, but if the goal is transparent paravirtualization, the kernel must provide fallback support for an IO-APIC anyway. In addition, the hypervisor must support an IO-APIC for SMP non-paravirtualized guests. The net result is less code on both sides, and an already well defined interface between the two. This avoids the complexity burden of having to support two different interfaces to achieve the same task.

One shortcut we have found most helpful is to simply disable NMI delivery to the paravirtualized kernel. There is no reason NMIs can't be supported, but typical uses for them are not as productive in a virtualized environment. Watchdog NMIs are of limited use if the OS is already correct and running on stable hardware; profiling NMIs are similarly of less use, since this task is accomplished with more accuracy in the VMM itself; and NMIs for machine check errors should be handled outside of the VM. The addition of NMI support does create additional complexity for the trap handling code in the VM, and although the task is surmountable, the value proposition is debatable. Here, again, feedback is desired.

Time Interface.

In a virtualized environment, virtual machines (VM) will time share the system with each other and with other processes running on the host system. Therefore, a VM's virtual CPUs (VCPUs) will be executing on the host's physical CPUs (PCPUs) for only some portion of time. This section of the VMI exposes a paravirtual view of time to the guest operating systems so that they may operate more effectively in a virtual environment. The interface also provides a way for the VCPUs to set alarms in this paravirtual view of time.

Time Domains:

a) Wallclock Time:

Wallclock time exposed to the VM through this interface indicates the number of nanoseconds since epoch, 1970-01-01T00:00:00Z (ISO 8601 date format). If the host's wallclock time changes (say, when an error in the host's clock is corrected), so does the wallclock time as viewed through this interface.

b) Real Time:

Another view of time accessible through this interface is real time. Real time always progresses except for when the VM is stopped or suspended. Real time is presented to the guest as a counter which increments at a constant rate defined (and presented) by the hypervisor. All the VCPUs of a VM share the same real time counter.

The unit of the counter is called "cycles". The unit and initial value (corresponding to the time the VM enters para-virtual mode) are chosen by the hypervisor so that the real time counter will not rollover in any practical length of time. It is expected that the frequency (cycles per second) is chosen such that this clock provides a "high-resolution" view of time. The unit can only change when the VM (re)enters paravirtual mode.

c) Stolen time and Available time:

A VCPU is always in one of three states: running, halted, or ready. The VCPU is in the 'running' state if it is executing. When the VCPU executes the HLT interface, the VCPU enters the 'halted' state and remains halted until there is some work pending for the VCPU (e.g. an alarm expires, host I/O completes on behalf of virtual I/O). At this point, the VCPU enters the 'ready' state (waiting for the hypervisor to reschedule it). Finally, at any time when the VCPU is not in the 'running' state nor the 'halted' state, it is in the 'ready' state.

For example, consider the following sequence of events, with times given in real time:

(Example 1)

At 0 ms, VCPU executing guest code.
At 1 ms, VCPU requests virtual I/O.
At 2 ms, Host performs I/O for virtual I/0.
At 3 ms, VCPU executes VMI_Halt.
At 4 ms, Host completes I/O for virtual I/O request.
At 5 ms, VCPU begins executing guest code, vectoring to the interrupt
handler for the device initiating the virtual I/O.
At 6 ms, VCPU preempted by hypervisor.
At 9 ms, VCPU begins executing guest code.

From 0 ms to 3 ms, VCPU is in the 'running' state. At 3 ms, VCPU enters the 'halted' state and remains in this state until the 4 ms mark. From 4 ms to 5 ms, the VCPU is in the 'ready' state. At 5 ms, the VCPU re-enters the 'running' state until it is preempted by the hypervisor at the 6 ms mark. From 6 ms to 9 ms, VCPU is again in the 'ready' state, and finally 'running' again after 9 ms.

Stolen time is defined per VCPU to progress at the rate of real time when the VCPU is in the 'ready' state, and does not progress otherwise. Available time is defined per VCPU to progress at the rate of real time when the VCPU is in the 'running' and 'halted' states, and does not progress when the VCPU is in the 'ready' state.

So, for the above example, the following table indicates these time values for the VCPU at each ms boundary:

Real time Stolen time Available time
0 0 0
1 0 1
2 0 2
3 0 3
4 0 4
5 1 4
6 1 5
7 2 5
8 3 5
9 4 5
10 4 6

Notice that at any point:
real_time == stolen_time + available_time

Stolen time and available time are also presented as counters in "cycles" units. The initial value of the stolen time counter is 0. This implies the initial value of the available time counter is the same as the real time counter.

Alarms:

Alarms can be set (armed) against the real time counter or the available time counter. Alarms can be programmed to expire once (one-shot) or on a regular period (periodic). They are armed by indicating an absolute counter value expiry, and in the case of a periodic alarm, a non-zero relative period counter value. [TBD:
The method of wiring the alarms to an interrupt vector is dependent upon the virtual interrupt controller portion of the interface. Currently, the alarms may be wired as if they are attached to IRQ0 or the vector in the local APIC LVTT. This way, the alarms can be used as drop in replacements for the PIT or local APIC timer.]

Alarms are per-vcpu mechanisms. An alarm set by vcpu0 will fire only on vcpu0, while an alarm set by vcpu1 will only fire on vcpu1. If an alarm is set relative to available time, its expiry is a value relative to the available time counter of the vcpu that set it.

The interface includes a method to cancel (disarm) an alarm. On each vcpu, one alarm can be set against each of the two counters (real time and available time). A vcpu in the 'halted' state becomes 'ready' when any of its alarm's counters reaches the expiry.

An alarm "fires" by signaling the virtual interrupt controller. An alarm will fire as soon as possible after the counter value is greater than or equal to the alarm's current expiry. However, an alarm can fire only when its vcpu is in the 'running' state.

If the alarm is periodic, a sequence of expiry values,

E(i) = e0 + p * i , i = 0, 1, 2, 3, ...

where 'e0' is the expiry specified when setting the alarm and 'p' is the period of the alarm, is used to arm the alarm. Initially, E(0) is used as the expiry. When the alarm fires, the next expiry value in the sequence that is greater than the current value of the counter is used as the alarm's new expiry.

One-shot alarms have only one expiry. When a one-shot alarm fires, it is automatically disarmed.

Suppose an alarm is set relative to real time with expiry at the 3 ms mark and a period of 2 ms. It will expire on these real time
marks: 3, 5, 7, 9. Note that even if the alarm does not fire during the 5 ms to 7 ms interval, the alarm can fire at most once during the 7 ms to 9 ms interval (unless, of course, it is reprogrammed).

If an alarm is set relative to available time with expiry at the 1 ms mark (in available time) and with a period of 2 ms, then it will expire on these available time marks: 1, 3, 5. In the scenario described in example 1, those available time values correspond to these values in real time: 1, 3, 6.

3) Architectural Differences from Native Hardware.

For the sake of performance, some requirements are imposed on kernel fault handlers which are not present on real hardware. Most modern operating systems should have no trouble meeting these requirements. Failure to meet these requirements may prevent the kernel from working properly.

1) The hardware flags on entry to a fault handler may not match the EFLAGS image on the fault handler stack. The stack image is correct, and will have the correct state of the interrupt and arithmetic flags.

2) The stack used for kernel traps must be flat - that is, zero base, segment limit determined by the hypervisor.

3) On entry to any fault handler, the stack must have sufficient space to hold 32 bytes of data, or the guest may be terminated.

4) When calling VMI functions, the kernel must be running on a flat 32-bit stack and code segment.

5) Most VMI functions require flat data and extra segment (DS and ES) segments as well; notable exceptions are IRET and SYSEXIT. XXXPara - may need to add STI and CLI to this list.

6) Interrupts must always be enabled when running code in userspace.

7) IOPL semantics for userspace are changed; although userspace may be granted port access, it can not affect the interrupt flag.

8) The EIPs at which faults may occur in VMI calls may not match the original native instruction EIP; this is a bug in the system today, as many guests do rely on lazy fault handling.

9) On entry to V8086 mode, MSR_SYSENTER_CS is cleared to zero.

10) Todo - we would like to support these features, but they are not fully tested and / or implemented:

Userspace 16-bit stack support
Proper handling of faulting IRETs

4) ROM Implementation

Modularization

Originally, we envisioned modularizing the ROM API into several subsections, but the close coupling between the initial layers and the requirement to support native PCI bus devices has made ROM components for network or block devices unnecessary to this point in time.

VMI - the virtual machine interface. This is the core CPU, I/O
and MMU virtualization layer. I/O is currently limited to port access to emulated devices.

Detection

The presence of hypervisor ROMs can be recognized by scanning the upper region of the first megabyte of physical memory. Multiple ROMs may be provided to support older API versions for legacy guest OS support. ROM detection is done in the traditional manner, by scanning the memory region from C8000h - DFFFFh in 2 kilobyte increments. The romSignature bytes must be '0x55, 0xAA', and the checksum of the region indicated by the romLength field must be zero. The checksum is a simple 8-bit addition of all bytes in the ROM region.

Data layout

typedef struct HyperRomHeader {
uint16_t romSignature;
int8_t romLength;
unsigned char romEntry[4];
uint8_t romPad0;
uint32_t hyperSignature;
uint8_t APIVersionMinor;
uint8_t APIVersionMajor;
uint8_t reserved0;
uint8_t reserved1;
uint32_t reserved2;
uint32_t reserved3;
uint16_t pciHeaderOffset;
uint16_t pnpHeaderOffset;
uint32_t romPad3;
char reserved[32];
char elfHeader[64];
} HyperRomHeader;

The first set of fields is defined by the BIOS:

romSignature - fixed 0xAA55, BIOS ROM signature
romLength - the length of the ROM, in 512 byte chunks.
Determines the area to be checksummed.
romEntry - 16-bit initialization code stub used by BIOS.
romPad0 - reserved

The next set of fields is defined by this API:

hyperSignature - a 4 byte signature providing recognition of the
device class represented by this ROM. Each device class defines its own unique signature.
APIVersionMinor - the revision level of this device class' API.
This indicates incremental changes to the API.
APIVersionMajor - the major version. Used to indicates large
revisions or additions to the API which break compatibility with the previous version.
reserved0,1,2,3 - for future expansion

The next set of fields is defined by the PCI / PnP BIOS spec:

pciHeaderOffset - relative offset to the PCI device header from
the start of this ROM.
pnpHeaderOffset - relative offset to the PnP boot header from the
start of this ROM.
romPad3 - reserved by PCI spec.

Finally, there is space for future header fields, and an area reserved for an ELF header to point to symbol information.

Appendix A - VMI ROM Low Level ABI

OS writers intending to port their OS to the paravirtualizable x86 processor being modeled by this hypervisor need to access the hypervisor through the VMI layer. It is possible although it is currently unimplemented to add or replace the functionality of individual hypervisor calls by providing your own ROM images. This is intended to allow third party customizations.

VMI compatible ROMs user the signature "cVmi" in the hyperSignature field of the ROM header.

Many of these calls are compatible with the SVR4 C call ABI, using up
to three register arguments. Some calls are not, due to restrictions
of the native instruction set. Calls which diverge from this ABI are noted. In GNU terms, this means most of the calls are compatible with regparm(3) argument passing.

Most of these calls behave as standard C functions, and as such, may clobber registers EAX, EDX, ECX, flags. Memory clobbers are noted explicitly, since many of them may be inlined without a memory clobber.

Most of these calls require well defined segment conventions - that is, flat full size 32-bit segments for all the general segments, CS, SS, DS, ES. Exceptions in some cases are noted.

The net result of these choices is that most of the calls are very easy to make from C-code, and calls that are likely to be required in
low level trap handling code are easy to call from assembler. Most
of these calls are also very easily implemented by the hypervisor vendor in C code, and only the performance critical calls from assembler paths require custom assembly implementations.

CORE INTERFACE CALLS

This set of calls provides the base functionality to establish running the kernel in VMI mode.

The interface will be expanded to include feature negotiation, more explicit control over call bundling and flushing, and hypervisor notifications to allow inline code patching.

VMI_Init

VMICALL void VMI_Init(void);

Initializes the hypervisor environment. Returns zero on success, or -1 if the hypervisor could not be initialized. Note that this is a recoverable error if the guest provides the requisite native code to support transparent paravirtualization.

Inputs: None
Outputs: EAX = result
Clobbers: Standard
Segments: Standard

PROCESSOR STATE CALLS

This set of calls controls the online status of the processor. It include interrupt control, reboot, halt, and shutdown functionality. Future expansions may include deep sleep and hotplug CPU capabilities.

VMI_DisableInterrupts

VMICALL void VMI_DisableInterrupts(void);

Disable maskable interrupts on the processor.

Inputs: None
Outputs: None
Clobbers: Flags only
Segments: As this is both performance critical and likely to
be called from low level interrupt code, this call does not require flat DS/ES segments, but uses the stack segment for data access. Therefore only CS/SS must be well defined.

VMI_EnableInterrupts

VMICALL void VMI_EnableInterrupts(void);

Enable maskable interrupts on the processor. Note that the current implementation always will deliver any pending interrupts on a call which enables interrupts, for compatibility with kernel code which expects this behavior. Whether this should be required is open for debate.

Inputs: None
Outputs: None
Clobbers: Flags only
Segments: CS/SS only

VMI_GetInterruptMask

VMICALL VMI_UINT VMI_GetInterruptMask(void);

Returns the current interrupt state mask of the processor. The mask is defined to be 0x200 (matching processor flag IF) to indicate interrupts are enabled.

Inputs: None
Outputs: EAX = mask
Clobbers: Flags only
Segments: CS/SS only

VMI_SetInterruptMask

VMICALL void VMI_SetInterruptMask(VMI_UINT mask);

Set the current interrupt state mask of the processor. Also delivers any pending interrupts if the mask is set to allow them.

Inputs: EAX = mask
Outputs: None
Clobbers: Flags only
Segments: CS/SS only

VMI_DeliverInterrupts (For future debate)

Enable and deliver any pending interrupts. This would remove the implicit delivery semantic from the SetInterruptMask and EnableInterrupts calls.

VMI_Pause

VMICALL void VMI_Pause(void);

Pause the processor temporarily, to allow a hypertwin or remote CPU to continue operation without lock or cache contention.

Inputs: None
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_Halt

VMICALL void VMI_Halt(void);

Put the processor into interruptible halt mode. This is defined to be a non-running mode where maskable interrupts are enabled, not a deep low power sleep mode.

Inputs: None
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_Shutdown

VMICALL void VMI_Shutdown(void);

Put the processor into non-interruptible halt mode. This is defined to be a non-running mode where maskable interrupts are disabled, indicates a power-off event for this CPU.

Inputs: None
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_Reboot:

VMICALL void VMI_Reboot(VMI_INT how);

Reboot the virtual machine, using a hard or soft reboot. A soft reboot corresponds to the effects of an INIT IPI, and preserves some APIC and CR state. A hard reboot corresponds to a hardware reset.

Inputs: EAX = reboot mode
#define VMI_REBOOT_SOFT 0x0
#define VMI_REBOOT_HARD 0x1
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_SetInitialAPState:

void VMI_SetInitialAPState(APState *apState, VMI_UINT32 apicID);

Sets the initial state of the application processor with local APIC ID "apicID" to the state in apState. apState must be the page-aligned linear address of the APState structure describing the initial state of the specified application processor.

Control register CR0 must have both PE and PG set; the result of either of these bits being cleared is undefined. It is recommended that for best performance, all processors in the system have the same setting of the CR4 PAE bit. LME and LMA in EFER are both currently unsupported. The result of setting either of these bits is undefined.

Inputs: EAX = pointer to APState structure for new co-processor
EDX = APIC ID of processor to initialize
Outputs: None
Clobbers: Standard
Segments: Standard

DESCRIPTOR RELATED CALLS

VMI_SetGDT

VMICALL void VMI_SetGDT(VMI_DTR *gdtr);

Load the global descriptor table limit and base registers. In addition to the straightforward load of the hardware registers, this has the additional side effect of reloading all segment registers in a virtual machine. The reason is that otherwise, the hidden part of segment registers (the base field) may be put into a non-reversible state. Non-reversible segments are problematic because they can not be reloaded - any subsequent loads of the segment will load the new descriptor state. In general, is not possible to resume direct execution of the virtual machine if certain segments become non-reversible.

A load of the GDTR may cause the guest visible memory image of the GDT to be changed. This allows the hypervisor to share the GDT pages with the guest, but also continue to maintain appropriate protections on the GDT page by transparently adjusting the DPL and RPL of descriptors in the GDT.

Inputs: EAX = pointer to descriptor limit / base
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_SetIDT

VMICALL void VMI_SetIDT(VMI_DTR *idtr);

Load the interrupt descriptor table limit and base registers. The IDT format is defined to be the same as native hardware.

A load of the IDTR may cause the guest visible memory image of the IDT to be changed. This allows the hypervisor to rewrite the IDT pages in a format more suitable to the hypervisor, which may include adjusting the DPL and RPL of descriptors in the guest IDT.

Inputs: EAX = pointer to descriptor limit / base
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_SetLDT

VMICALL void VMI_SetLDT(VMI_SELECTOR ldtSel);

Load the local descriptor table. This has the additional side effect of of reloading all segment registers. See VMI_SetGDT for an explanation of why this is required. A load of the LDT may cause the guest visible memory image of the LDT to be changed, just as GDT and IDT loads.

Inputs: EAX = GDT selector of LDT descriptor
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_SetTR

VMICALL void VMI_SetTR(VMI_SELECTOR ldtSel);

Load the task register. Functionally equivalent to the LTR instruction.

Inputs: EAX = GDT selector of TR descriptor
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_GetGDT

VMICALL void VMI_GetGDT(VMI_DTR *gdtr);

Copy the GDT limit and base fields into the provided pointer. This is equivalent to the SGDT instruction, which is non-virtualizable.

Inputs: EAX = pointer to descriptor limit / base
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_GetIDT

VMICALL void VMI_GetIDT(VMI_DTR *idtr);

Copy the IDT limit and base fields into the provided pointer. This is equivalent to the SIDT instruction, which is non-virtualizable.

Inputs: EAX = pointer to descriptor limit / base
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_GetLDT

VMICALL VMI_SELECTOR VMI_GetLDT(void);

Load the task register. Functionally equivalent to the SLDT instruction, which is non-virtualizable.

Inputs: None
Outputs: EAX = selector of LDT descriptor
Clobbers: Standard, Memory
Segments: Standard

VMI_GetTR

VMICALL VMI_SELECTOR VMI_GetTR(void);

Load the task register. Functionally equivalent to the STR instruction, which is non-virtualizable.

Inputs: None
Outputs: EAX = selector of TR descriptor
Clobbers: Standard, Memory
Segments: Standard

VMI_WriteGDTEntry

VMICALL void VMI_WriteGDTEntry(void *gdt, VMI_UINT entry,
VMI_UINT32 descLo, VMI_UINT32 descHi);

Write a descriptor to a GDT entry. Note that writes to the GDT itself may be disallowed by the hypervisor, in which case this call must be converted into a hypercall. In addition, since the descriptor may need to be modified to change limits and / or permissions, the guest kernel should not assume the update will be binary identical to the passed input.

Inputs: EAX = pointer to GDT base
EDX = GDT entry number
ECX = descriptor low word
ST(1) = descriptor high word
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_WriteLDTEntry

VMICALL void VMI_WriteLDTEntry(void *gdt, VMI_UINT entry,
VMI_UINT32 descLo, VMI_UINT32 descHi);

Write a descriptor to a LDT entry. Note that writes to the LDT itself may be disallowed by the hypervisor, in which case this call must be converted into a hypercall. In addition, since the descriptor may need to be modified to change limits and / or permissions, the guest kernel should not assume the update will be binary identical to the passed input.

Inputs: EAX = pointer to LDT base
EDX = LDT entry number
ECX = descriptor low word
ST(1) = descriptor high word
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_WriteIDTEntry

VMICALL void VMI_WriteIDTEntry(void *gdt, VMI_UINT entry,
VMI_UINT32 descLo, VMI_UINT32 descHi);

Write a descriptor to a IDT entry. Since the descriptor may need to be modified to change limits and / or permissions, the guest kernel should not assume the update will be binary identical to the passed input.

Inputs: EAX = pointer to IDT base
EDX = IDT entry number
ECX = descriptor low word
ST(1) = descriptor high word
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

CPU CONTROL CALLS

These calls encapsulate the set of privileged instructions used to manipulate the CPU control state. These instructions are all properly virtualizable using trap and emulate, but for performance reasons, a direct call may be more efficient. With hardware virtualization capabilities, many of these calls can be left as IDENT translations, that is, inline implementations of the native instructions, which are not rewritten by the hypervisor. Some of these calls are performance critical during context switch paths, and some are not, but they are all included for completeness, with the exceptions of the obsoleted LMSW and SMSW instructions.

VMI_WRMSR

VMICALL void VMI_WRMSR(VMI_UINT64 val, VMI_UINT32 reg);

Write to a model specific register. This functions identically to the hardware WRMSR instruction. Note that a hypervisor may not implement the full set of MSRs supported by native hardware, since many of them are not useful in the context of a virtual machine.

Inputs: ECX = model specific register index
EAX = low word of register
EDX = high word of register
Outputs: None
Clobbers: Standard, Memory
Segments: Standard

VMI_RDMSR

VMICALL VMI_UINT64 VMI_RDMSR(VMI_UINT64 dummy, VMI_UINT32 reg);

Read from a model specific register. This functions identically to the hardware RDMSR instruction. Note that a hypervisor may not implement the full set of MSRs supported by native hardware, since many of them are not useful in the context of a virtual machine.

Inputs: ECX = machine specific register index
Outputs: EAX = low word of register
EDX = high word of register
Clobbers: Standard
Segments: Standard

VMI_SetCR0

VMICALL void VMI_SetCR0(VMI_UINT val);

Write to control register zero. This can cause TLB flush and FPU handling side effects. The set of features available to the kernel depend on the completeness of the hypervisor. An explicit list of supported functionality or required settings may need to be negotiated by the hypervisor and kernel during bootstrapping. This is likely to be implementation or vendor specific, and the precise restrictions are not yet worked out. Our implementation in general supports turning on additional functionality - enabling protected mode, paging, page write protections; however, once those features have been enabled, they may not be disabled on the virtual hardware.

Inputs: EAX = input to control register
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_SetCR2

VMICALL void VMI_SetCR2(VMI_UINT val);

Write to control register two. This has no side effects other than updating the CR2 register value.

Inputs: EAX = input to control register
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_SetCR3

VMICALL void VMI_SetCR3(VMI_UINT val);

Write to control register three. This causes a TLB flush on the local processor. In addition, this update may be queued as part of a lazy call invocation, which allows multiple hypercalls to be issued during the context switch path. The queuing convention is to be negotiated with the hypervisor during bootstrapping, but the interfaces for this negotiation are currently vendor specific.

Inputs: EAX = input to control register
Outputs: None
Clobbers: Standard
Segments: Standard
Queue Class: MMU

VMI_SetCR4

VMICALL void VMI_SetCR3(VMI_UINT val);

Write to control register four. This can cause TLB flush and many other CPU side effects. The set of features available to the kernel depend on the completeness of the hypervisor. An explicit list of supported functionality or required settings may need to be negotiated by the hypervisor and kernel during bootstrapping. This is likely to be implementation or vendor specific, and the precise restrictions are not yet worked out. Our implementation in general supports turning on additional MMU functionality - enabling global pages, large pages, PAE mode, and other features - however, once those features have been enabled, they may not be disabled on the virtual hardware. The remaining CPU control bits of CR4 remain active and behave identically to real hardware.

Inputs: EAX = input to control register
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_GetCR0
VMI_GetCR2
VMI_GetCR3
VMI_GetCR4

VMICALL VMI_UINT32 VMI_GetCR0(void);
VMICALL VMI_UINT32 VMI_GetCR2(void);
VMICALL VMI_UINT32 VMI_GetCR3(void);
VMICALL VMI_UINT32 VMI_GetCR4(void);

Read the value of a control register into EAX. The register contents are identical to the native hardware control registers; CR0 contains the control bits and task switched flag, CR2 contains the last page fault address, CR3 contains the page directory base pointer, and CR4 contains various feature control bits.

Inputs: None
Outputs: EAX = value of control register
Clobbers: Standard
Segments: Standard

VMI_CLTS

VMICALL void VMI_CLTS(void);

Used to clear the task switched (TS) flag in control register zero. A replacement for the CLTS instruction.

Inputs: None
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_SetDR

VMICALL void VMI_SetDR(VMI_UINT32 num, VMI_UINT32 val);

Set the debug register to the given value. If a hypervisor implementation supports debug registers, this functions equivalently to native hardware move to DR instructions.

Inputs: EAX = debug register number
EDX = debug register value
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_GetDR

VMICALL VMI_UINT32 VMI_GetDR(VMI_UINT32 num);

Read a debug register. If debug registers are not supported, the implementation is free to return zero values.

Inputs: EAX = debug register number
Outputs: EAX = debug register value
Clobbers: Standard
Segments: Standard

PROCESSOR INFORMATION CALLS

These calls provide access to processor identification, performance and cycle data, which may be inaccurate due to the nature of running on
virtual hardware. This information may be visible in a non-virtualizable
way to applications running outside of the kernel. As such, both RDTSC and RDPMC should be disabled by kernels or hypervisors where information leakage is a concern, and the accuracy of data retrieved by these functions is up to the individual hypervisor vendor.

VMI_CPUID

/* Not expressible as a C function */

The CPUID instruction provides processor feature identification in a vendor specific manner. The instruction itself is non-virtualizable without hardware support, requiring a hypervisor assisted CPUID call that emulates the effect of the native instruction, while masking any unsupported CPU feature bits.

Inputs: EAX = CPUID number
ECX = sub-level query (nonstandard)
Outputs: EAX = CPUID dword 0
EBX = CPUID dword 1
ECX = CPUID dword 2
EDX = CPUID dword 3
Clobbers: Flags only
Segments: Standard

VMI_RDTSC

VMICALL VMI_UINT64 VMI_RDTSC(void);

The RDTSC instruction provides a cycles counter which may be made visible to userspace. For better or worse, many applications have made use of this feature to implement userspace timers, database indices, or for micro-benchmarking of performance. This instruction is extremely problematic for virtualization, because even though it is selectively virtualizable using trap and emulate, it is much more expensive to virtualize it in this fashion. On the other hand, if this instruction is allowed to execute without trapping, the cycle counter provided could be wrong in any number of circumstances due to hardware drift, migration, suspend/resume, CPU hotplug, and other unforeseen consequences of running inside of a virtual machine. There is no standard specification for how this instruction operates when issued from userspace programs, but the VMI call here provides a proper interface for the kernel to read this cycle counter.

Inputs: None
Outputs: EAX = low word of TSC cycle counter
EDX = high word of TSC cycle counter
Clobbers: Standard
Segments: Standard

VMI_RDPMC

VMICALL VMI_UINT64 VMI_RDPMC(VMI_UINT64 dummy, VMI_UINT32 counter);

Similar to RDTSC, this call provides the functionality of reading processor performance counters. It also is selectively visible to userspace, and maintaining accurate data for the performance counters is an extremely difficult task due to the side effects introduced by the hypervisor.

Inputs: ECX = performance counter index
Outputs: EAX = low word of counter
EDX = high word of counter
Clobbers: Standard
Segments: Standard

STACK / PRIVILEGE TRANSITION CALLS

This set of calls encapsulates mechanisms required to transfer between higher privileged kernel tasks and userspace. The stack switching and return mechanisms are also used to return from interrupt handlers into the kernel, which may involve atomic interrupt state and stack transitions.

VMI_UpdateKernelStack

VMICALL void VMI_UpdateKernelStack(void *tss, VMI_UINT32 esp0);

Inform the hypervisor that a new kernel stack pointer has been loaded in the TSS structure. This new kernel stack pointer will be used for entry into the kernel on interrupts from userspace.

Inputs: EAX = pointer to TSS structure
EDX = new kernel stack top
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_IRET

/* No C prototype provided */

Perform a near equivalent of the IRET instruction, which atomically switches off the current stack and restore the interrupt mask. This may return to userspace or back to the kernel from an interrupt or exception handler. The VMI_IRET call does not restore IOPL from the stack image, as the native hardware equivalent would. Instead, IOPL must be explicitly restored using a VMI_SetIOPL call. The VMI_IRET call does, however, restore the state of the EFLAGS_VM bit from the stack image in the event that the hypervisor and kernel both support V8086 execution mode. If the hypervisor does not support V8086 mode, this can be silently ignored, generating an error that the guest must deal with. Note this call is made using a CALL instruction, just as all other VMI calls, so the EIP of the call site is available to the VMI layer. This allows faults during the sequence to be properly passed back to the guest kernel with the correct EIP.

Note that returning to userspace with interrupts disabled is an invalid operation in a paravirtualized kernel, and the results of an attempt to do so are undefined.

Also note that when issuing the VMI_IRET call, the userspace data segments may have already been restored, so only the stack and code segments can be assumed valid.

There is currently no support for IRET calls from a 16-bit stack segment, which poses a problem for supporting certain userspace applications which make use of high bits of ESP on a 16-bit stack. How to best resolve this is an open question. One possibility is to introduce a new VMI call which can operate on 16-bit segments, since it is desirable to make the common case here as fast as possible.

Inputs: ST(0) = New EIP
ST(1) = New CS
ST(2) = New Flags (including interrupt mask) ST(3) = New ESP (for userspace returns) ST(4) = New SS (for userspace returns) ST(5) = New ES (for v8086 returns) ST(6) = New DS (for v8086 returns) ST(7) = New FS (for v8086 returns) ST(8) = New GS (for v8086 returns)
Outputs: None (does not return)
Clobbers: None (does not return)
Segments: CS / SS only

VMI_SYSEXIT

/* No C prototype provided */

For hypervisors and processors which support SYSENTER / SYSEXIT, the VMI_SYSEXIT call is provided as a binary equivalent to the native SYSENTER instruction. Since interrupts must always be enabled in userspace, the VMI version of this function always combines atomically enabling interrupts with the return to userspace.

Inputs: EDX = New EIP
ECX = New ESP
Outputs: None (does not return)
Clobbers: None (does not return)
Segments: CS / SS only

I/O CALLS

This set of calls incorporates I/O related calls - PIO, setting I/O privilege level, and forcing memory writeback for device coherency.

VMI_INB
VMI_INW
VMI_INL

VMICALL VMI_UINT8 VMI_INB(VMI_UINT dummy, VMI_UINT port); VMICALL VMI_UINT16 VMI_INW(VMI_UINT dummy, VMI_UINT port); VMICALL VMI_UINT32 VMI_INL(VMI_UINT dummy, VMI_UINT port);

Input a byte, word, or doubleword from an I/O port. These instructions have binary equivalent semantics to native instructions.

Inputs: EDX = port number
EDX, rather than EAX is used, because the native encoding of the instruction may use this register implicitly.
Outputs: EAX = port value
Clobbers: Memory only
Segments: Standard

VMI_OUTB
VMI_OUTW
VMI_OUTL

VMICALL void VMI_OUTB(VMI_UINT value, VMI_UINT port); VMICALL void VMI_OUTW(VMI_UINT value, VMI_UINT port); VMICALL void VMI_OUTL(VMI_UINT value, VMI_UINT port);

Output a byte, word, or doubleword to an I/O port. These instructions have binary equivalent semantics to native instructions.

Inputs: EAX = port value
EDX = port number
Outputs: None
Clobbers: None
Segments: Standard

VMI_INSB
VMI_INSW
VMI_INSL

/* Not expressible as C functions */

Input a string of bytes, words, or doublewords from an I/O port. These instructions have binary equivalent semantics to native instructions. They do not follow a C calling convention, and clobber only the same registers as native instructions.

Inputs: EDI = destination address
EDX = port number
ECX = count
Outputs: None
Clobbers: ESI, ECX, Memory
Segments: Standard

VMI_OUTSB
VMI_OUTSW
VMI_OUTSL

/* Not expressible as C functions */

Output a string of bytes, words, or doublewords to an I/O port. These instructions have binary equivalent semantics to native instructions. They do not follow a C calling convention, and clobber only the same registers as native instructions.

Inputs: ESI = source address
EDX = port number
ECX = count
Outputs: None
Clobbers: ESI, ECX
Segments: Standard

VMI_IODelay

VMICALL void VMI_IODelay(void);

Delay the processor by time required to access a bus register. This is easily implemented on native hardware by an access to a bus scratch register, but is typically not useful in a virtual machine. It is paravirtualized to remove the overhead implied by executing the native delay.

Inputs: None
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_SetIOPLMask

VMICALL void VMI_SetIOPLMask(VMI_UINT32 mask);

Set the IOPL mask of the processor to allow userspace to access I/O ports. Note the mask is pre-shifted, so an IOPL of 3 would be expressed as (3 << 12). If the guest chooses to use IOPL to allow CPL-3 access to I/O ports, it must explicitly set and restore IOPL using these calls; attempting to set the IOPL flags with popf or iret may produce no result.

Inputs: EAX = Mask
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_WBINVD

VMICALL void VMI_WBINVD(void);

Write back and invalidate the data cache. This is used to synchronize I/O memory.

Inputs: None
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_INVD

This instruction is deprecated. It is invalid to execute in a virtual machine. It is documented here only because it is still declared in the interface, and dropping it required a version change.

APIC CALLS

APIC virtualization is currently quite simple. These calls support the functionality of the hardware APIC in a form that allows for more efficient implementation in a hypervisor, by avoiding trapping access to APIC memory. The calls are kept simple to make the implementation compatible with native hardware. The APIC must be mapped at a page boundary in the processor virtual address space.

VMI_APICWrite

VMICALL void VMI_APICWrite(void *reg, VMI_UINT32 value);

Write to a local APIC register. Side effects are the same as native hardware APICs.

Inputs: EAX = APIC register address
EDX = value to write
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_APICRead

VMICALL VMI_UINT32 VMI_APICRead(void *reg);

Read from a local APIC register. Side effects are the same as native hardware APICs.

Inputs: EAX = APIC register address
Outputs: EAX = APIC register value
Clobbers: Standard
Segments: Standard

TIMER CALLS

The VMI interfaces define a highly accurate and efficient timer interface that is available when running inside of a hypervisor. This is an optional but highly recommended feature which avoids many of the problems presented by classical timer virtualization. It provides notions of stolen time, counters, and wall clock time which allows the VM to get the most accurate information in a way which is free of races and legacy hardware dependence.

VMI_GetWallclockTime

VMI_NANOSECS VMICALL VMI_GetWallclockTime(void);

VMI_GetWallclockTime returns the current wallclock time as the number of nanoseconds since the epoch. Nanosecond resolution along with the 64-bit unsigned type provide over 580 years from epoch until rollover. The wallclock time is relative to the host's wallclock time.

Inputs: None
Outputs: EAX = low word, wallclock time in nanoseconds
EDX = high word, wallclock time in nanoseconds
Clobbers: Standard
Segments: Standard

VMI_WallclockUpdated

VMI_BOOL VMICALL VMI_WallclockUpdated(void);

VMI_WallclockUpdated returns TRUE if the wallclock time has changed relative to the real cycle counter since the previous time that VMI_WallclockUpdated was polled. For example, while a VM is suspended, the real cycle counter will halt, but wallclock time will continue to advance. Upon resuming the VM, the first call to VMI_WallclockUpdated will return TRUE.

Inputs: None
Outputs: EAX = 0 for FALSE, 1 for TRUE
Clobbers: Standard
Segments: Standard

VMI_GetCycleFrequency

VMICALL VMI_CYCLES VMI_GetCycleFrequency(void);

VMI_GetCycleFrequency returns the number of cycles in one second. This value can be used by the guest to convert between cycles and other time units.

Inputs: None
Outputs: EAX = low word, cycle frequency
EDX = high word, cycle frequency
Clobbers: Standard
Segments: Standard

VMI_GetCycleCounter

VMICALL VMI_CYCLES VMI_GetCycleCounter(VMI_UINT32 whichCounter);

VMI_GetCycleCounter returns the current value, in cycles units, of the counter corresponding to 'whichCounter' if it is one of VMI_CYCLES_REAL, VMI_CYCLES_AVAILABLE or VMI_CYCLES_STOLEN. VMI_GetCycleCounter returns 0 for any other value of 'whichCounter'.

Inputs: EAX = counter index, one of
#define VMI_CYCLES_REAL 0
#define VMI_CYCLES_AVAILABLE 1
#define VMI_CYCLES_STOLEN 2
Outputs: EAX = low word, cycle counter
EDX = high word, cycle counter
Clobbers: Standard
Segments: Standard

VMI_SetAlarm

VMICALL void VMI_SetAlarm(VMI_UINT32 flags, VMI_CYCLES expiry,
VMI_CYCLES period);

VMI_SetAlarm is used to arm the vcpu's alarms. The 'flags' parameter is used to specify which counter's alarm is being set (VMI_CYCLES_REAL or VMI_CYCLES_AVAILABLE), how to deliver the alarm to the vcpu (VMI_ALARM_WIRED_IRQ0 or VMI_ALARM_WIRED_LVTT), and the mode (VMI_ALARM_IS_ONESHOT or VMI_ALARM_IS_PERIODIC). If the alarm is set against the VMI_ALARM_STOLEN counter or an undefined counter number, the call is a nop. The 'expiry' parameter indicates the expiry of the alarm, and for periodic alarms, the 'period' parameter indicates the period of the alarm. If the value of 'period' is zero, the alarm is armed as a one-shot alarm regardless of the mode specified by 'flags'. Finally, a call to VMI_SetAlarm for an alarm that is already armed is equivalent to first calling VMI_CancelAlarm and then calling VMI_SetAlarm, except that the value returned by VMI_CancelAlarm is not accessible.

/* The alarm interface 'flags' bits. [TBD: exact format of 'flags'] */

Inputs: EAX = flags value, cycle counter number or'ed with
#define VMI_ALARM_WIRED_IRQ0 0x00000000
#define VMI_ALARM_WIRED_LVTT 0x00010000
#define VMI_ALARM_IS_ONESHOT 0x00000000
#define VMI_ALARM_IS_PERIODIC 0x00000100
EDX = low word, alarm expiry
ECX = high word, alarm expiry
ST(0) = low word, alarm expiry ST(1) = high word, alarm expiry
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_CancelAlarm

VMICALL VMI_BOOL VMI_CancelAlarm(VMI_UINT32 flags);

VMI_CancelAlarm is used to disarm an alarm. The 'flags' parameter indicates which alarm to cancel (VMI_CYCLES_REAL or VMI_CYCLES_AVAILABLE). The return value indicates whether or not the cancel succeeded. A return value of FALSE indicates that the alarm was already disarmed either because a) the alarm was never set or b) it was a one-shot alarm and has already fired (though perhaps not yet delivered to the guest). TRUE indicates that the alarm was armed and either a) the alarm was one-shot and has not yet fired (and will no longer fire until it is rearmed) or b) the alarm was periodic.

Inputs: EAX = cycle counter number
Outputs: EAX = 0 for FALSE, 1 for TRUE
Clobbers: Standard
Segments: Standard

MMU CALLS

The MMU plays a large role in paravirtualization due to the large performance opportunities realized by gaining insight into the guest machine's use of page tables. These calls are designed to accommodate the existing MMU functionality in the guest OS while providing the hypervisor with hints that can be used to optimize performance to a large degree.

VMI_SetLinearMapping

VMICALL void VMI_SetLinearMapping(int slot, VMI_UINT32 va,
VMI_UINT32 pages, VMI_UINT32 ppn);

/* The number of VMI address translation slot */
#define VMI_LINEAR_MAP_SLOTS 4

Register a virtual to physical translation of virtual address range to physical pages. This may be used to register single pages or to register large ranges. There is an upper limit on the number of active mappings, which should be sufficient to allow the hypervisor and VMI layer to perform page translation without requiring dynamic storage. Translations are only required to be registered for addresses used to access page table entries through the VMI page table access functions. The guest is free to use the provided linear map slots in a manner that it finds most convenient. Kernels which linearly map a large chunk of physical memory and use page tables in this linear region will only need to register one such region after initialization of the VMI. Hypervisors which do not require linear to physical conversion hints are free to leave these calls as NOPs, which is the default when inlined into the native kernel.

Inputs: EAX = linear map slot
EDX = virtual address start of mapping
ECX = number of pages in mapping
ST(0) = physical frame number to which pages are mapped
Outputs: None
Clobbers: Standard
Segments: Standard

VMI_FlushTLB

VMICALL void VMI_FlushTLB(int how);

Flush all non-global mappings in the TLB, optionally flushing global mappings as well. The VMI_FLUSH_TLB flag should always be specified, optionally or'ed with the VMI_FLUSH_GLOBAL flag.

Inputs: EAX = flush type
#define VMI_FLUSH_TLB 0x01
#define VMI_FLUSH_GLOBAL 0x02
Outputs: None
Clobbers: Standard, memory (implied)
Segments: Standard

VMI_InvalPage

VMICALL void VMI_InvalPage(VMI_UINT32 va);

Invalidate the TLB mapping for a single page or large page at the given virtual address.

Inputs: EAX = virtual address
Outputs: None
Clobbers: Standard, memory (implied)
Segments: Standard

The remaining documentation here needs updating when the PTE accessors are simplified.

70) VMI_SetPte

void VMI_SetPte(VMI_PTE pte, VMI_PTE *ptep);

Assigns a new value to a page table / directory entry. It is a requirement that ptep points to a page that has already been registered with the hypervisor as a page of the appropriate type using the VMI_RegisterPageUsage function.

71) VMI_SwapPte

VMI_PTE VMI_SwapPte(VMI_PTE pte, VMI_PTE *ptep);

Write 'pte' into the page table entry pointed by 'ptep', and returns the old value in 'ptep'. This function acts atomically on the PTE to provide up to date A/D bit information in the returned value.

72) VMI_TestAndSetPteBit

VMI_BOOL VMI_TestAndSetPteBit(VMI_INT bit, VMI_PTE *ptep);

Atomically set a bit in a page table entry. Returns zero if the bit was not set, and non-zero if the bit was set.

73) VMI_TestAndClearPteBit

VMI_BOOL VMI_TestAndSetClearBit(VMI_INT bit, VMI_PTE *ptep);

Atomically clear a bit in a page table entry. Returns zero if the bit was not set, and non-zero if the bit was set.

74) VMI_SetPteLong
75) VMI_SwapPteLong
76) VMI_TestAndSetPteBitLong
77) VMI_TestAndClearPteBitLong

void VMI_SetPteLong(VMI_PAE_PTE pte, VMI_PAE_PTE *ptep); VMI_PAE_PTE VMI_SwapPteLong(VMI_UINT64 pte, VMI_PAE_PTE *ptep); VMI_BOOL VMI_TestAndSetPteBitLong(VMI_INT bit, VMI_PAE_PTE *ptep); VMI_BOOL VMI_TestAndSetClearBitLong(VMI_INT bit, VMI_PAE_PTE *ptep);

These functions act identically to the 32-bit PTE update functions, but provide support for PAE mode. The calls are guaranteed to never create a temporarily invalid but present page mapping that could be accidentally prefetched by another processor, and all returned bits are guaranteed to be atomically up to date.

One special exception is the VMI_SwapPteLong function only provides synchronization against A/D bits from other processors, not against other invocations of VMI_SwapPteLong.

78) VMI_ClonePageTable
VMI_ClonePageDirectory

#define VMI_MKCLONE(start, count) (((start) << 16) | (count))

void VMI_ClonePageTable(VMI_UINT32 dstPPN, VMI_UINT32 srcPPN,
VMI_UINT32 flags);
void VMI_ClonePageDirectory(VMI_UINT32 dstPPN, VMI_UINT32 srcPPN,
VMI_UINT32 flags);

These functions tell the hypervisor to allocate a page shadow at the PT or PD level using a shadow template. Because of the availability of bits in the flags, these calls may be merged together as well as flag the PAE-ness of the shadows.

80) VMI_RegisterPageUsage
81) VMI_ReleasePage

#define VMI_PAGE_PT 0x01
#define VMI_PAGE_PD 0x02
#define VMI_PAGE_PDP 0x04
#define VMI_PAGE_PML4 0x08
#define VMI_PAGE_GDT 0x10
#define VMI_PAGE_LDT 0x20
#define VMI_PAGE_IDT 0x40
#define VMI_PAGE_TSS 0x80

void VMI_RegisterPageUsage(VMI_UINT32 ppn, int flags); void VMI_ReleasePage(VMI_UINT32 ppn, int flags);

These are used to register a page with the hypervisor as being of a particular type, for instance, VMI_PAGE_PT says it is a page table page.

85) VMI_SetDeferredMode

void VMI_SetDeferredMode(VMI_UINT32 deferBits);

Set the lazy state update mode to the specified set of bits. This allows the processor, hypervisor, or VMI layer to lazily update certain CPU and MMU state. When setting this to a more permissive setting, no flush is implied, but when clearing bits in the current defer mask, all pending state will be flushed.

The 'deferBits' is a mask specifying how to flush.

#define VMI_DEFER_NONE 0x00

Disallow all asynchronous state updates. This is the default state.

#define VMI_DEFER_MMU 0x01

Flush all pending page table updates. Note that page faults, invalidations and TLB flushes will implicitly flush all pending updates.

#define VMI_DEFER_CPU 0x02

Allow CPU state updates to control registers to be deferred, with the exception of updates that change FPU state. This is useful for combining a reload of the page table base in CR3 with other updates, such as the current kernel stack.

#define VMI_DEFER_DT 0x04

Allow descriptor table updates to be delayed. This allows the VMI_UpdateGDT / IDT / LDT calls to be asynchronously queued.

86) VMI_FlushDeferredCalls

void VMI_FlushDeferredCalls(void);

Flush all asynchronous state updates which may be queued as a result of setting deferred update mode.

Appendix B - VMI C prototypes

Most of the VMI calls are properly callable C functions. Note that for the absolute best performance, assembly calls are preferable in some cases, as they do not imply all of the side effects of a C function call, such as register clobber and memory access. Nevertheless, these wrappers serve as a useful interface definition for higher level languages.

In some cases, a dummy variable is passed as an unused input to force proper alignment of the remaining register values.

The call convention for these is defined to be standard GCC convention with register passing. The regparm call interface is documented at:

http://gcc.gnu.org/onlinedocs/gcc/Function-Attributes.html

Types used by these calls:

   VMI_UINT64   64 bit unsigned integer
   VMI_UINT32   32 bit unsigned integer
   VMI_UINT16   16 bit unsigned integer
   VMI_UINT8    8 bit unsigned integer
   VMI_INT      32 bit integer
   VMI_UINT     32 bit unsigned integer
   VMI_DTR      6 byte compressed descriptor table limit/base
   VMI_PTE      4 byte page table entry (or page directory)

VMI_LONG_PTE 8 byte page table entry (or PDE or PDPE) VMI_SELECTOR 16 bit segment selector
VMI_BOOL 32 bit unsigned integer
VMI_CYCLES 64 bit unsigned integer
VMI_NANOSECS 64 bit unsigned integer

   #ifndef VMI_PROTOTYPES_H
   #define VMI_PROTOTYPES_H

   /* Insert local type definitions here */
   typedef struct VMI_DTR {
      uint16 limit;
      uint32 offset __attribute__ ((packed));
   } VMI_DTR;

   typedef struct APState {
      VMI_UINT32 cr0;
      VMI_UINT32 cr2;
      VMI_UINT32 cr3;
      VMI_UINT32 cr4;
      VMI_UINT64 efer;

      VMI_UINT32 eip;
      VMI_UINT32 eflags;
      VMI_UINT32 eax;
      VMI_UINT32 ebx;
      VMI_UINT32 ecx;
      VMI_UINT32 edx;
      VMI_UINT32 esp;
      VMI_UINT32 ebp;
      VMI_UINT32 esi;
      VMI_UINT32 edi;
      VMI_UINT16 cs;
      VMI_UINT16 ss;

      VMI_UINT16 ds;
      VMI_UINT16 es;
      VMI_UINT16 fs;
      VMI_UINT16 gs;
      VMI_UINT16 ldtr;

      VMI_UINT16 gdtrLimit;
      VMI_UINT32 gdtrBase;
      VMI_UINT32 idtrBase;
      VMI_UINT16 idtrLimit;

} APState;

   #define VMICALL __attribute__((regparm(3)))

   /* CORE INTERFACE CALLS */
   VMICALL void VMI_Init(void);

   /* PROCESSOR STATE CALLS */
   VMICALL void     VMI_DisableInterrupts(void);
   VMICALL void     VMI_EnableInterrupts(void);

   VMICALL VMI_UINT VMI_GetInterruptMask(void);
   VMICALL void     VMI_SetInterruptMask(VMI_UINT mask);

   VMICALL void     VMI_Pause(void);
   VMICALL void     VMI_Halt(void);
   VMICALL void     VMI_Shutdown(void);
   VMICALL void     VMI_Reboot(VMI_INT how);

   #define VMI_REBOOT_SOFT 0x0
   #define VMI_REBOOT_HARD 0x1

   void VMI_SetInitialAPState(APState *apState, VMI_UINT32 apicID);

   /* DESCRIPTOR RELATED CALLS */
   VMICALL void         VMI_SetGDT(VMI_DTR *gdtr);
   VMICALL void         VMI_SetIDT(VMI_DTR *idtr);
   VMICALL void         VMI_SetLDT(VMI_SELECTOR ldtSel);
   VMICALL void         VMI_SetTR(VMI_SELECTOR ldtSel);

   VMICALL void         VMI_GetGDT(VMI_DTR *gdtr);
   VMICALL void         VMI_GetIDT(VMI_DTR *idtr);
   VMICALL VMI_SELECTOR VMI_GetLDT(void);
   VMICALL VMI_SELECTOR VMI_GetTR(void);

VMICALL void VMI_WriteGDTEntry(void *gdt,

                                          VMI_UINT entry,
                                          VMI_UINT32 descLo,
                                          VMI_UINT32 descHi);
   VMICALL void         VMI_WriteLDTEntry(void *gdt,
                                          VMI_UINT entry,
                                          VMI_UINT32 descLo,
                                          VMI_UINT32 descHi);
   VMICALL void         VMI_WriteIDTEntry(void *gdt,
                                          VMI_UINT entry,
                                          VMI_UINT32 descLo,
                                          VMI_UINT32 descHi);

/* CPU CONTROL CALLS */
VMICALL void VMI_WRMSR(VMI_UINT64 val, VMI_UINT32 reg);
VMICALL void VMI_WRMSR_SPLIT(VMI_UINT32 valLo, VMI_UINT32 valHi,

                                      VMI_UINT32 reg);

/* Not truly a proper C function; use dummy to align reg in ECX */ VMICALL VMI_UINT64 VMI_RDMSR(VMI_UINT64 dummy, VMI_UINT32 reg);

   VMICALL void VMI_SetCR0(VMI_UINT val);
   VMICALL void VMI_SetCR2(VMI_UINT val);
   VMICALL void VMI_SetCR3(VMI_UINT val);
   VMICALL void VMI_SetCR4(VMI_UINT val);

   VMICALL VMI_UINT32 VMI_GetCR0(void);
   VMICALL VMI_UINT32 VMI_GetCR2(void);
   VMICALL VMI_UINT32 VMI_GetCR3(void);
   VMICALL VMI_UINT32 VMI_GetCR4(void);

   VMICALL void       VMI_CLTS(void);

   VMICALL void       VMI_SetDR(VMI_UINT32 num, VMI_UINT32 val);
   VMICALL VMI_UINT32 VMI_GetDR(VMI_UINT32 num);

/* PROCESSOR INFORMATION CALLS */

   VMICALL VMI_UINT64 VMI_RDTSC(void);
   VMICALL VMI_UINT64 VMI_RDPMC(VMI_UINT64 dummy, VMI_UINT32 counter);

   /* STACK / PRIVILEGE TRANSITION CALLS */
   VMICALL void VMI_UpdateKernelStack(void *tss, VMI_UINT32 esp0);

   /* I/O CALLS */
   /* Native port in EDX - use dummy */
   VMICALL VMI_UINT8  VMI_INB(VMI_UINT dummy, VMI_UINT port);
   VMICALL VMI_UINT16 VMI_INW(VMI_UINT dummy, VMI_UINT port);
   VMICALL VMI_UINT32 VMI_INL(VMI_UINT dummy, VMI_UINT port);

   VMICALL void VMI_OUTB(VMI_UINT value, VMI_UINT port);
   VMICALL void VMI_OUTW(VMI_UINT value, VMI_UINT port);
   VMICALL void VMI_OUTL(VMI_UINT value, VMI_UINT port);

   VMICALL void VMI_IODelay(void);
   VMICALL void VMI_WBINVD(void);
   VMICALL void VMI_SetIOPLMask(VMI_UINT32 mask);

   /* APIC CALLS */
   VMICALL void       VMI_APICWrite(void *reg, VMI_UINT32 value);
   VMICALL VMI_UINT32 VMI_APICRead(void *reg);

   /* TIMER CALLS */
   VMICALL VMI_NANOSECS VMI_GetWallclockTime(void);
   VMICALL VMI_BOOL     VMI_WallclockUpdated(void);

/* Predefined rate of the wallclock. */
#define VMI_WALLCLOCK_HZ 1000000000

   VMICALL VMI_CYCLES VMI_GetCycleFrequency(void);
   VMICALL VMI_CYCLES VMI_GetCycleCounter(VMI_UINT32 whichCounter);

   /* Defined cycle counters */
   #define VMI_CYCLES_REAL        0
   #define VMI_CYCLES_AVAILABLE   1
   #define VMI_CYCLES_STOLEN      2

VMICALL void VMI_SetAlarm(VMI_UINT32 flags, VMI_CYCLES expiry,

                                 VMI_CYCLES period);

VMICALL VMI_BOOL VMI_CancelAlarm(VMI_UINT32 flags);

/* The alarm interface 'flags' bits. [TBD: exact format of 'flags'] */ #define VMI_ALARM_COUNTER_MASK 0x000000ff

   #define VMI_ALARM_WIRED_IRQ0   0x00000000
   #define VMI_ALARM_WIRED_LVTT   0x00010000

   #define VMI_ALARM_IS_ONESHOT   0x00000000
   #define VMI_ALARM_IS_PERIODIC  0x00000100

/* MMU CALLS */
VMICALL void VMI_SetLinearMapping(int slot, VMI_UINT32 va,

                                     VMI_UINT32 pages, VMI_UINT32 ppn);

/* The number of VMI address translation slot */
#define VMI_LINEAR_MAP_SLOTS 4

   VMICALL void VMI_InvalPage(VMI_UINT32 va);
   VMICALL void VMI_FlushTLB(int how);
   
   /* Flags used by VMI_FlushTLB call */
   #define VMI_FLUSH_TLB            0x01
   #define VMI_FLUSH_GLOBAL         0x02

   #endif

Appendix C - Sensitive x86 instructions in the paravirtual environment

This is a list of x86 instructions which may operate in a different manner when run inside of a paravirtual environment.

ARPL - continues to function as normal, but kernel segment registers
may be different, so parameters to this instruction may need to be modified. (System)

IRET - the IRET instruction will be unable to change the IOPL, VM,
VIF, VIP, or IF fields. (System)

the IRET instruction may #GP if the return CS/SS RPL are below the CPL, or are not equal. (System)

LAR - the LAR instruction will reveal changes to the DPL field of
descriptors in the GDT and LDT tables. (System, User)

LSL - the LSL instruction will reveal changes to the segment limit
of descriptors in the GDT and LDT tables. (System, User)

LSS - the LSS instruction may #GP if the RPL is not set properly.
(System)

MOV - the mov %seg, %reg instruction may reveal a different RPL
on the segment register. (System)

The mov %reg, %ss instruction may #GP if the RPL is not set to the current CPL. (System)

POP - the pop %ss instruction may #GP if the RPL is not set to
the appropriate CPL. (System)

POPF - the POPF instruction will be unable to set the hardware
interrupt flag. (System)

PUSH - the push %seg instruction may reveal a different RPL on the
segment register. (System)

PUSHF- the PUSHF instruction will reveal a possible different IOPL,
and the value of the hardware interrupt flag, which is always set. (System, User)

SGDT - the SGDT instruction will reveal the location and length of
the GDT shadow instead of the guest GDT. (System, User)

SIDT - the SIDT instruction will reveal the location and length of
the IDT shadow instead of the guest IDT. (System, User)

SLDT - the SLDT instruction will reveal the selector used for
the shadow LDT rather than the selector loaded by the guest. (System, User).

STR - the STR instruction will reveal the selector used for the
shadow TSS rather than the selector loaded by the guest. (System, User).

_______________________________________________
Xen-devel mailing list
[email protected]
http://lists.xensource.com/xen-devel

#9 lore raw
In reply to Arjan van de Ven
Quoted 10 lines · Arjan van de Ven
Quoted 2 lines · Zachary Amsden

and gives hypervisors room to grow while maintaining
binary compatibility with already released kernels.

that I buy for binary only hypervisors. But in an open source world I'll buy this a LOT less as being relevant.

Binary compatibility to Linux is pretty important for applications. Even though Apache is open source, I don't want to recompile it for every new Linux kernel. Fortunately I don't have to, because glibc abstracts the Linux kernel interface. Consider VMI in the same role as glibc -- when the hypervisor changes, VMI maintains compatibility with your pre-existing infrastructure, while letting you have some of the benefits of the new hypervisor. The upgrade and recompile game can quickly end in a stalemate when you have packages with conflicting dependencies (one package requires the old version, and the other package requires the new version).

Josh

#11 lore raw
In reply to Zachary Amsden
Quoted 4 lines · Zachary Amsden

It is an advantage for everyone.  It cuts support and certification costs for Linux distributors, software vendors, makes debugging and development easier, and gives hypervisors room to grow while maintaining binary compatibility with already released kernels.

It certainly is good for kernel developers and end-users.

However, it would be a foolish distributor or ISV who tests with one hypervisor and decides that covers all hypervisors which implement the same interface. So I'm not sure there's any advantage w.r.t. support and certification costs.

Signature · 2 lines
Hollis Blanchard
IBM Linux Technology Center
#7 lore raw
In reply to Zachary Amsden
Quoted 8 lines · * Zachary Amsden ([email protected]) wrote:

This we find to be quite true. Now, you can use a VMI kernel, make changes to it, run it on native hardware, and be confident that it will run properly in a VM as well. And you can develop in a VM, with confidence that you can run on native hardware. You can even replace the entire "ROM" image with your own custom debugging image to add any type of debugging or performance monitoring facility you want - and you have some very, very interesting hook points into the kernel that make that task much more achievable.

Replacing a ROM image is easier in terms package management, but still requires full validation and certification process. Swapping out the underlying core is a major change and needs to be re-validated.

thanks,
-chris

#12 lore raw
In reply to Hollis Blanchard
Quoted 9 lines · Hollis Blanchard
Quoted 2 lines · Zachary Amsden

It is an advantage for everyone. It cuts support and certification costs for Linux distributors, software vendors, makes debugging and development easier, and gives hypervisors room to grow while maintaining binary compatibility with already released kernels.

It certainly is good for kernel developers and end-users.

However, it would be a foolish distributor or ISV who tests with one hypervisor and decides that covers all hypervisors which implement the same interface. So I'm not sure there's any advantage w.r.t. support and certification costs.

Your point is well noted. I'm not arguing that it would be smart to test with just one hypervisor (or worse, yet, test only on native hardware), and proudly declare your kernel virtualization compatible. There are some things you can do (instrument a torture test verification module in a native VMI ROM) to help with that test load.

But in the end, having a single binary reduces the complexity and work that goes into a certification, which does simplify the process - even if you still have to validate against the list of all supported vendors / hardware.

Zach

#18 lore raw
In reply to Zachary Amsden
Quoted 7 lines · Zachary Amsden

In OLS 2005, we described the work that we have been doing in VMware with respect a common interface for paravirtualization of Linux. We shared the general vision in Rik's virtualization BoF.
[...]
Unlike the full-virtualization techniques used in the traditional VMware products, paravirtualization is a technique where the operating system is modified to enlighten the hypervisor with timely knowledge about the

Quoted 10 lines

operating system's activities. Since the hypervisor now depends on the kernel to tell it about common idioms etc, it does not need to write protect OS objects such as page and descriptor tables as a solution based on full-virtualization needs. This has two important effects (a) it shortens the critical path, since faulting is expensive on modern processors (b) by eliminating complex heuristics the hypervisor is simplified. While the former delivers performance, the latter is quite important too.

An interesting vision, especially if it merges the current VMWare / Xen techno-social rift.

I think there will still be a place for the "complete" (eg, QEMU, or of course your own product) and the "ultimately lightweight" (eg, vserver/openvz/jails/containers) approaches to virtualisation, though.

While the same kernel may be able to run in these different situations, having a "real" hardware emulator like QEMU/VMWare will allow you to test all of those alternate code paths. As time goes on this will no doubt seem a more and more superfluous requirement, especially if the actual code in those places is minimal as you suggest.

Looking the other way, there is no way that a system like this will ever approach the performance of fork(), as vserver and related technology does. No doubt for certain common applications of virtualisation, such as providing "complete" virtual servers, this will be seen as less and less important as time goes on. However, for other applications - such as jailing services, or systems that make use of advantages of single kernel virtualisation (such as shared VFS/network, visibility into other systems' processes, etc) - the extra kernel that a virtualisation context implies is simply unwanted. No doubt still other users will simply prefer the simplicity of system call level virtualisation, such as only having one set of routing tables/iptables rules/VGs to manage, etc.

There are currently two debates on virtualisation happening here, on and off. We have Xen/VMI, and vserver/openvz/jails/containers. Let's just try not to get them confused :-). From the perspective of the vserver project, I consider your work orthogonal and complementary and wish you well in the success of your gracious offering to the Linux community.

Sam.

#19 lore raw
In reply to Zachary Amsden
Quoted 7 lines · Zachary Amsden

This is by no means finished work. A few of the areas that need more attention and exploration are (a) 64bit support is still lacking, but we feel a port of VMI to the 64 bit Linux can be done without too much trouble (b) the Xen compatibility layer needs some work to bring it up to the Xen 3.0 interfaces. Work is underway on this already, and no major issues are expected at this time.

Hi Zach,

Can you please post the Xen compatibility layer (even if it is for 2.0.x). I think it's important to see that code to understand the advantages/disadvantages compared to the existing Xen paravirtualization interface. Likewise, any Xen performance data would be useful as there has been some discussion about whether VMI would negatively impact Xen performance[1].

Thanks,

Anthony Liguori

Quoted 26 lines

Two final notes. This is not an attempt to force a proprietary interface into the Linux kernel. This is an attempt to find a common interface that can be used by many hypervisors by isolating hypervisor specific idioms into a neutral layer. This new layer is just what is claims to be - a virtual machine interface, which allows hypervisor dependent code to be abstracted in a way that benefits both Linux and hypervisor development.

This is also not an attempt to define an exact and final specification
of how virtualization should be done in Linux. This is very much a work
in progress, and it is understood that the interfaces proposed here will
change in time to accommodate the needs of all interested parties. We hope to find a common solution that can eventually become part of the Linux kernel and serve as a model for other operating systems as well.

We appreciate your feedback on this design and the patches to Linux, and welcome working with anyone who is interested in making virtualization in Linux a friendly environment to innovate in. If you find the ideas here interesting, please volunteer to help improve them.

------------------------------------------------------------------------

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/virtualization

#20 lore raw
Quoted 10 lines · Anthony Liguori
Quoted 6 lines · Zachary Amsden

This is by no means finished work. A few of the areas that need more attention and exploration are (a) 64bit support is still lacking, but we feel a port of VMI to the 64 bit Linux can be done without too much trouble (b) the Xen compatibility layer needs some work to bring it up to the Xen 3.0 interfaces. Work is underway on this already, and no major issues are expected at this time.

Hi Zach,

Can you please post the Xen compatibility layer (even if it is for 2.0.x). I think it's important to see that code to understand the advantages/disadvantages compared to the existing Xen paravirtualization interface. Likewise, any Xen performance data would be useful as there has been some discussion about whether VMI would negatively impact Xen performance[1].

About performance - I actually believe that it is possible to implement VMI Linux in such a way that it actually has _better_ performance on Xen than the current XenoLinux kernels.

I'm working on getting together the older interface pieces now.

Zach

#21 lore raw
Quoted 3 lines · Zachary Amsden

About performance - I actually believe that it is possible to implement VMI Linux in such a way that it actually has _better_ performance on Xen than the current XenoLinux kernels.

How would VMI allow page table batching at fault time? (one of the future optimizations that are probably worth making for Xen)

Signature · 1 line
All Rights Reversed
#23 lore raw
In reply to Zachary Amsden

Hi Zach,

A number of the files you posted (including the vmi_spec.txt) have the phrase 'All rights reserved'. That seems incompatible with the GPL. In particular, it makes it unclear about how one can use the actual vmi spec.

In your next round of patches, could you clarify the actual licensing of the files?

Regards,

Anthony Liguori

Quoted 97 lines · Zachary Amsden

In OLS 2005, we described the work that we have been doing in VMware with respect a common interface for paravirtualization of Linux. We shared the general vision in Rik's virtualization BoF.

This note is an update on our further work on the Virtual Machine Interface, VMI. The patches provided have been tested on 2.6.16-rc6. We are currently recollecting performance information for the new -rc6 kernel, but expect our numbers to match previous results, which showed no impact whatsoever on macro benchmarks, and nearly neglible impact on microbenchmarks.

Unlike the full-virtualization techniques used in the traditional VMware products, paravirtualization is a technique where the operating system is modified to enlighten the hypervisor with timely knowledge about the operating system's activities. Since the hypervisor now depends on the kernel to tell it about common idioms etc, it does not need to write protect OS objects such as page and descriptor tables as a solution based on full-virtualization needs. This has two important effects (a) it shortens the critical path, since faulting is expensive on modern processors (b) by eliminating complex heuristics the hypervisor is simplified. While the former delivers performance, the latter is quite important too.

Not surprisingly, paravirtualization's strength, ie that it encourages tighter communication between the kernel and the hypervisor, is also its weakness. Unless the changes to the operating system are moderated, you can very quickly find yourself with a kernel that (a) looks and feels like a brand new kernel or (b) cannot run on native machines or on newer versions of the hypervisor without a full recompile. The former can impede innovation in the Linux kernel, and the latter can be a problem for software vendors.

VMware proposes VMI as a paravirtualization interface for Linux that solves these problems.
- A VMI'fied Linux kernel runs unmodified on native hardware, and on many hypervisors, while simultaneously delivering on the performance promise of paravirtualization.
- VMI has a rich and low level interface, which allows the kernel to cope with future hardware evolution by querying for hardware capability. It is our expectation that a single kernel will run unmodified on both today's processors with limited hardware virtualization support and also keep up with any evolution on the processor front
- VMI Linux is a fairly clean interface, with distinct name spaces for objects from the kernel and the hypervisor. Nowhere do we mingle names from the hypervisor with that of the kernel. This separation allows innovation in the kernel to proceed at the same speed as always. For most kernel developers, a VMI kernel looks and feels like a regular Linux kernel.
- VMI Linux still supports "native" hypervisor device drivers, for example a hypervisor vendor's own private network or block device drivers which are free to use any interface desired to communicate with the hypervisor.

At present, we are sharing a working implementation of the VMI for 2.6.16-rc6 version of Linux. We have verified that VMI Linux does indeed run well on native machines (both P4 and Opterons), and on VMware style hypervisors. VMI Linux has negligible overheads on native machines, so much so, that we are confident that VMI Linux can, in the long run, be the default Linux for i386. We believe that this interface is both cleaner and more powerful than other proposals that have been made towards virtualization of Linux, and can easily be adapted to work with other hypervisors.

This is by no means finished work. A few of the areas that need more attention and exploration are (a) 64bit support is still lacking, but we feel a port of VMI to the 64 bit Linux can be done without too much trouble (b) the Xen compatibility layer needs some work to bring it up to the Xen 3.0 interfaces. Work is underway on this already, and no major issues are expected at this time.

Two final notes. This is not an attempt to force a proprietary interface into the Linux kernel. This is an attempt to find a common interface that can be used by many hypervisors by isolating hypervisor specific idioms into a neutral layer. This new layer is just what is claims to be - a virtual machine interface, which allows hypervisor dependent code to be abstracted in a way that benefits both Linux and hypervisor development.

This is also not an attempt to define an exact and final specification
of how virtualization should be done in Linux. This is very much a work
in progress, and it is understood that the interfaces proposed here will
change in time to accommodate the needs of all interested parties. We hope to find a common solution that can eventually become part of the Linux kernel and serve as a model for other operating systems as well.

We appreciate your feedback on this design and the patches to Linux, and welcome working with anyone who is interested in making virtualization in Linux a friendly environment to innovate in. If you find the ideas here interesting, please volunteer to help improve them.

------------------------------------------------------------------------

_______________________________________________
Virtualization mailing list
[email protected]
https://lists.osdl.org/mailman/listinfo/virtualization

#24 lore raw
Quoted 5 lines · Anthony Liguori

Hi Zach,

A number of the files you posted (including the vmi_spec.txt) have the phrase 'All rights reserved'. That seems incompatible with the GPL. In particular, it makes it unclear about how one can use the actual vmi spec.

In your next round of patches, could you clarify the actual licensing of the files?

I'm sorry about the legalese. The patches are patches to the Linux kernel, and therefore under GPL v2 by default. I thought that would be implicit.

Zach

#25 lore raw
Quoted 7 lines · Zachary Amsden
Quoted 2 lines · Anthony Liguori

In your next round of patches, could you clarify the actual licensing of the files?

I'm sorry about the legalese. The patches are patches to the Linux kernel, and therefore under GPL v2 by default. I thought that would be implicit.

It would be very bad if Linus started applying code with a dubious license to the kernel, if we want to keep the kernel GPL v2.

Having an explicit license and a Signed-off-by: line are things to remember with big patch sets. At the very least a Signed-off-by: line..

Signature · 1 line
All Rights Reversed
#22 lore raw
In reply to Rik van Riel
Quoted 10 lines · Rik van Riel
Quoted 2 lines · Zachary Amsden

About performance - I actually believe that it is possible to implement VMI Linux in such a way that it actually has _better_ performance on Xen than the current XenoLinux kernels.

How would VMI allow page table batching at fault time? (one of the future optimizations that are probably worth making for Xen)

This is exactly what we do. All page table transitions from P->NP or P->P already require a flushing call (FlushTLB or InvalPage). The remaining transitions, NP->P require explicit flushing, and we have added the appropriate call sites to do so. It turns out, the external MMU cache on Sparc provided exactly the required hook point in this case - update_mmu_cache().

Zach

#26 lore raw
In reply to Rik van Riel
Quoted 4 lines · Rik van Riel

It would be very bad if Linus started applying code with a dubious license to the kernel, if we want to keep the kernel GPL v2.

I believe it says explicitly in our patches that they are licensed under GPL v2.

Quoted 4 lines

Having an explicit license and a Signed-off-by: line are things to remember with big patch sets. At the very least a Signed-off-by: line.

There is a Signed-off-by line on every patch I send out, with full knowledge that this constitutes the work of the author of the said line, and full knowledge that this commits the patch into the domain of the GPL license. Sorry for sounding like a lawyer here. IANAL, but I thought that was completely implicit in all patches made to GPL'd software. The signed off by provides accountability and open licensing simultaneously.

But most importantly, I really don't understand how it is possible to make a patch to the Linux kernel and not release it under GPL.

Zach

#27 lore raw
Quoted 1 line · Zachary Amsden

There is a Signed-off-by line on every patch I send out,

You're right. It was just the first 1/24 that was missing it, it was there in the second copy.

Quoted 2 lines

But most importantly, I really don't understand how it is possible to make a patch to the Linux kernel and not release it under GPL.

This can really only be done if the person posting the patch does not have the right to release the code. This is what the Signed-off-by lines are for, IIRC.

Signature · 1 line
All Rights Reversed
#28 lore raw
Quoted 9 lines · Rik van Riel
Quoted 2 lines · Zachary Amsden

There is a Signed-off-by line on every patch I send out,

You're right. It was just the first 1/24 that was missing it, it was there in the second copy.

BTW, I have no idea why the first 1/24 was missing it. I checked right before sending, and it was there - perhaps I forgot to save my changes. The second copy turned out fine, but didn't make it to LKML. Everyone cc'd directly got it, but the LKML filter has a ban on the word propasition, and being blackholed by it, I merely assumed the patch was too large - so I split it up, and actually ended up binary searching down to the problematic section before finding the taboo list.

_Every_ problem eventually turns into a binary search.

Zach

#13 lore raw
In reply to Zachary Amsden
Quoted 10 lines · Zachary Amsden
Quoted 2 lines

Why can't vmware use the Xen interface instead?

We could. But it is our opinion that the Xen interface is unnecessarily complicated, without a clean separation between the layer of interaction with the hypervisor and the kernel proper. The interface we propose we believe is more powerful, and more conducive to performance optimizations while providing significant advantages - most specifically, a single binary image that is properly virtualizable on multiple hypervisors and capable of running on native hardware.

I agree with Zach here, the Xen hypervisor <-> kernel interface is not very nice. This proposal seems like a step forward althogh it'll probably need to go through a few iterations. Without and actually useable opensource hypevisor reference implementation it's totally unacceptable, though.

#14 lore raw
Quoted 6 lines · Christoph Hellwig

I agree with Zach here, the Xen hypervisor <-> kernel interface is not very nice. This proposal seems like a step forward althogh it'll probably need to go through a few iterations. Without and actually useable opensource hypevisor reference implementation it's totally unacceptable, though.

Which is why our top priority is getting VMI Linux to run on Xen. The churn rate on both ends has been very high, and we really wanted to release our patches with Xen support, but we also didn't want to wait some unknown number of weeks more to release them - and we're actually looking for volunteers to help with the port if anyone is interested.

What we are hoping for, in the end, is a Linux kernel with a clean virtualization interface, that is maintainable, does not slow hypervisor exploitation of new technologies, still offers all of the same performance advantages, works for the open source community, and allows hypervisor vendors of many creeds to benefit from cross-kernel binary compatibility.

Zach

#15 lore raw
In reply to Christoph Hellwig
Quoted 12 lines · Christoph Hellwig
Quoted 10 lines · Zachary Amsden
Quoted 2 lines

Why can't vmware use the Xen interface instead?

We could. But it is our opinion that the Xen interface is unnecessarily complicated, without a clean separation between the layer of interaction with the hypervisor and the kernel proper. The interface we propose we believe is more powerful, and more conducive to performance optimizations while providing significant advantages - most specifically, a single binary image that is properly virtualizable on multiple hypervisors and capable of running on native hardware.

Quoted 6 lines

I agree with Zach here, the Xen hypervisor <-> kernel interface is not very nice. This proposal seems like a step forward althogh it'll probably need to go through a few iterations. Without and actually useable opensource hypevisor reference implementation it's totally unacceptable, though.

As part of our pre-virtualization work, we developed a virtualization solution similar to VMI. We support Xen v2 and v3 with high performance. We added support for the first generation of VMI to our project, and are currently adding support for the latest VMI patch. Our work is open source. We'll announce when we finish the VMI updates.

We also experimented with other architectures and found the approach highly suitable, such as for Itanium.

Joshua

#16 lore raw
Quoted 6 lines · Joshua LeVasseur

As part of our pre-virtualization work, we developed a virtualization solution similar to VMI. We support Xen v2 and v3 with high performance. We added support for the first generation of VMI to our project, and are currently adding support for the latest VMI patch. Our work is open source. We'll announce when we finish the VMI updates.

Who is "we" and what product are you referring to?

(I think an important part of this discussion is getting an understanding of which virtualisation products (current or planned) could use a VMI).

Thanks.

#17 lore raw
Quoted 8 lines · Andrew Morton
Quoted 6 lines · Joshua LeVasseur

As part of our pre-virtualization work, we developed a virtualization solution similar to VMI. We support Xen v2 and v3 with high performance. We added support for the first generation of VMI to our project, and are currently adding support for the latest VMI patch. Our work is open source. We'll announce when we finish the VMI updates.

Quoted 6 lines

Who is "we" and what product are you referring to?

(I think an important part of this discussion is getting an understanding of which virtualisation products (current or planned) could use a VMI).

Thanks.

This is a project at the University of Karlsruhe. The project web page is:
http://l4ka.org/projects/virtualization/afterburn/
and there you'll find documentation and source code. The source code supports our pre-virtualization project on L4 and Xen (and we have some nascent implementations for Linux-as-hypervisor and Windows-as-hypervisor). We automate the transformations to the Linux code, thus minimizing the number of manual modifications. Due to the similarities of our approach and VMI, our virtualization runtime supports VMI with some minor additions.

Joshua

#10 lore raw
In reply to Joshua LeVasseur
Quoted 11 lines
Quoted 5 lines
Quoted 2 lines

and gives hypervisors room to grow while maintaining binary compatibility with already released kernels.

that I buy for binary only hypervisors. But in an open source world I'll buy this a LOT less as being relevant.

Binary compatibility to Linux is pretty important for applications. Even though Apache is open source, I don't want to recompile it for every new Linux kernel. Fortunately I don't have to, because glibc abstracts the Linux kernel interface. Consider VMI in the same role as glibc -- when the hypervisor changes, VMI maintains compatibility with your pre-existing infrastructure,

VMI = kernel code (AFAIU)

I would rather like a user-space-based compat layer.

Jan Engelhardt

#29 lore raw
In reply to Zachary Amsden
Quoted 4 lines

But most importantly, I really don't understand how it is possible to make a patch to the Linux kernel and not release it under GPL.

If the patch is so ultimatively trivial that there is only a few solutions (one or two), then there is no use in gpl'ing that flock of patchcode, in which case I think, it is (or at best should be) public domain. In conjunction with the patched function, they will/should become GPL.

Jan Engelhardt