Re: [PATCH 1/10] Trivial: /dev/kvm interface is no longer experimental.

[Arnd Bergmann <arnd-r2nGTMty4D4@public.gmane.org>]

On Tuesday 17 July 2007, Rusty Russell wrote:
> KVM interface is no longer experimental.
> 
> Signed-off-by: Rusty Russell <rusty-8n+1lVoiYb80n/F98K4Iww@public.gmane.org>
> 
> diff -r 4e57f5c6d4a9 include/linux/kvm.h
> --- a/include/linux/kvm.h       Tue Jul 17 13:04:58 2007 +1000
> +++ b/include/linux/kvm.h       Tue Jul 17 13:09:54 2007 +1000
> @@ -4,8 +4,7 @@
>  /*
>   * Userspace interface for /dev/kvm - kernel based virtual machine
>   *
> - * Note: this interface is considered experimental and may change without
> - *       notice.
> + * Note: you must update KVM_API_VERSION if you change this interface.
>   */

I don't like the idea of making this interface official. If this is the
point where we think it's perfect, it should become a set of real system
calls instead.

At that point, incompatible API changes are simply not acceptable any
more, and the bar should be raised as high as possible. With a system
call, you basically have to allocate a new syscall number if you need
to add another parameter or similar, and yuo need to keep the behaviour
of the old number for eternity.

The equivalent of the current set of ioctls comes down to roughtly this
set of syscalls:

int kvm_create_vm(void);
int kvm_get_msr_index_list(int fd, sttuct kvm_msr_list);
size_t kvm_get_vcpu_mmap_size(int fd);
int kvm_set_memory_region(int fd, unsigned slot, unsigned flags,
	__u64 guest_phys_addr, __u64 size);
int kvm_create_vcpu(int fd);
int kvm_get_dirty_log(int fd, unsigned slot, void *dirty_bitmap);
int kvm_get_memory_alias(int fd, struct kvm_memory_region *region);
int kvm_run(int fd);
int kvm_get_regs(int fd, struct kvm_regs *regs);
int kvm_set_regs(int fd, const struct kvm_regs *regs);
int kvm_get_sregs(int fd, struct kvm_sregs *sregs);
int kvm_set_sregs(int fd, const struct kvm_sregs *sregs);
int kvm_translate(int fd, __u64 linear_address, __u64 *physical_address,
		__u8  *valid, __u8 *writeable, __u8 *usemode);
int kvm_interrupt(int fd, __u32 irq);
int kvm_debug_guest(...);
int kvm_get_msrs(...);
int kvm_set_msrs(...);
int kvm_set_cpuid(...);
int kvm_set_signal_mask(...);
int kvm_get_fpu(...);
int kvm_set_fpu(...);

That's a lot of system calls! The only ioctl calls that can immediately
go away are KVM_GET_API_VERSION and KVM_CHECK_EXTENSION, if the API
is fixed.

Before moving to the final syscall interface, I'd wait for the memory
model to have moved to using a region of the host process address instead
of a separate address range. As I understood Carsten, that's what was
discussed in Ottawa anyway. It can probably remove a few of the
existing ioctl calls.

Some more can be saved by changing the interface for the kvm_get/set_*
calls. One way I can see this done at the syscall level is to replace
kvm_get_regs() with

regsfd = openat(vcpu, "regs", O_RDWR);
(void)read(regsfs, &regs, sizeof (regs);

Some of the others can also be done with files like this, e.g. writing
to an "interrupt" file. Of course, you would keep the file descriptors
open for the life time of the guest.

	Arnd <><

-------------------------------------------------------------------------
This SF.net email is sponsored by DB2 Express
Download DB2 Express C - the FREE version of DB2 express and take
control of your XML. No limits. Just data. Click to get it now.
http://sourceforge.net/powerbar/db2/