Re: [PATCH 0/7] KVM: Kernel-based Virtual Machine

[Arnd Bergmann <arnd@arndb.de>]

On Sunday 22 October 2006 19:39, Anthony Liguori wrote:
> I like the idea of a filesystem.  In particular, if you exposed the CPU
> state as a mmap()'able file, you could read/write from userspace without
> any syscall overhead.

Right. It's a little tricky though regarding what happens when you
write to the register mapping of a running guest, without stopping
it first.

> There are some clever ways that you could get around need that many
> syscalls.  For instance, you could have a "paused" file that you could
> write a "1" into in order to run the guest (assuming that the memory/CPU
> state is setup properly).

what for? writing 1, then 0 to that file is two full syscalls.
Calling kvm_run and returning from it is just one.

You can also just send SIGSTOP/SIGCONT to the task to stop it.

> You could then have an "event" file that you could select() for read
> on.  When "event" became readable, you could read the exit reason, do
> whatever is needed, and then write a "1" into "paused" again.

It's very handy to stay inside of a single process context for both
the hypervisor and the guest, and to simply block in a kvm_run syscall
for the time the guest executes.

This syscall can then simply return the exit reason as its return
value so you don't need another syscall to read it.

> Perhaps an ioctl is better for pausing/unpausing but I do think it's
> necessary to select() on something to wait for the next exit reason to
> occur.

I would not mix ioctls with a new file system. ioctl is fine on
a character device, but with a new file system, you should be able
to express everything as read/write, or one of the new syscalls.

	Arnd <><