On 11/21/2011 02:01 PM, Richard Laager wrote:
I'm not an expert on the architecture of KVM, so perhaps this is a QEMU
question. If so, please let me know and I'll ask on a different list.
It is a qemu question, yes (though fork()ing a guest also relates to kvm).
Background:
Assuming the block layer can make instantaneous snapshots of a guest's
disk (e.g. lvcreate -s), one can get "crash consistent" (i.e. as if the
guest crashed) snapshots. To get a "fully consistent" snapshot, you need
to shutdown the guest. For production VMs, this is obviously not ideal.
Idea:
What if KVM/QEMU was to fork() the guest and shutdown one copy?
KVM/QEMU would momentarily halt the execution of the guest and take a
writable, instantaneous snapshot of each block device. Then it would
fork(). The parent would resume execution as normal. The child would
redirect disk writes to the snapshot(s). The RAM should have
copy-on-write behavior as with any other fork()ed process. Other
resources like the network, display, sound, serial, etc. would simply be
disconnected/bit-bucketed. Finally, the child would resume guest
execution and send the guest an ACPI power button press event. This
would cause the guest OS to perform an orderly shutdown.
I believe this would provide consistent snapshots in the vast majority
of real-world scenarios in a guest OS and application-independent way.
Interesting idea. Will the guest actually shut down nicely without a
network? Things like NFS mounts will break.
Implementation Nits:
* A timeout on the child process would likely be a good idea.
* It'd probably be best to disconnect the network (i.e. tell the
guest the cable is unplugged) to avoid long timeouts. Likewise
for the hardware flow-control lines on the serial port.
This is actually critical, otherwise the guest will shutdown(2) all
sockets and confuse the clients.
* For correctness, fdatasync()ing or similar might be necessary
after halting execution and before creating the snapshots.
Microsoft guests have an API to quiesce storage prior to a snapshot, and
I think there is work to bring this to Linux guests. So it should be
possible to get consistent snapshots even without this, but it takes
more integration.