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

From: Avi Kivity
Date: Sun Oct 22 2006 - 04:40:47 EST


Arnd Bergmann wrote:
This looks _a_lot_ like what we're doing for the SPUs in the cell processor,
except that we're using different calls into the kernel. Have you looked
into what we have implemented there? The code is in
arch/powerpc/platforms/cell/spufs. I think it would be a good abstraction
to use for you as well, maybe we could even move to a common infrastructure,
as I have heard from a few other projects that want to do similar things.

The main differences to your interface are:

- A file system is used instead of a character device
- Directories, not open file descriptors represent contexts
- Two new syscalls were introduced (spu_create/spu_run)
- instead of ioctls, files represent different bits of information,
you can read/write, poll or mmap them.

Your example above could translate to something like:

int kvm_fd = kvm_create("/kvm/my_vcpu")
int mem_fd = openat(kvm_fd, "mem", O_RDWR);
void *mem = mmap(mem_fd, ...); // main memory
void *fbmem = mmap(mem_fd, ...); // frame buffer memory
int regs_fd = openat(kvm_fd, "regs", O_RDWR);
int irq_fd = openat(kvm_fd, "regs", O_WRONLY);

if (debugger) {
int fd = openat(fvm_fd, "debug", O_WRONLY);
write(fd, "1", 1);
close(fd);
}
while (1) {
int exit_reason = kvm_run(kvm_fd, &kvm_descriptor);
switch (exit reason) {
handle mmio, I/O etc. might call
write(irq_fd, &interrupt_packet, sizeof (interrupt_packet));
pread(regs_fd, &rax, sizeof rax, KVM_REG_RAX);
}

[cc'ing some others to solicit their opinion]


I like this. Since we plan to support multiple vcpus per vm, the fs structure might look like:

/kvm/my_vm
|
+----memory # mkdir to create memory slot.
| | # how to set size and offset?
| |
| +---0 # guest physical memory slot
| |
| +-- dirty_bitmap # read to get and atomically reset
| # the changed pages log
|
|
+----cpu # mkdir/rmdir to create/remove vcpu
|
+----0
| |
| +--- irq # write to inject an irq
| |
| +--- regs # read/write to get/set registers
| |
| +--- debugger # write to set breakpoints/singlestep mode
|
+----1
[...]

It's certainly a lot more code though, and requires new syscalls. Since this is a little esoteric does it warrant new syscalls?

--
error compiling committee.c: too many arguments to function

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/