When configured for interrupts (property "chardev" given), we receive
the shared memory from an ivshmem server. We do so asynchronously
after realize() completes, by setting up callbacks with
qemu_chr_add_handlers().
Keeping server I/O out of realize() that way avoids delays due to a
slow server. This is probably relevant only for hot plug.
However, this funny "no shared memory, yet" state of the device also
causes a raft of issues that are hard or impossible to work around:
* The guest is exposed to this state: when we enter and leave it its
shared memory contents is apruptly replaced, and device register
IVPosition changes.
This is a known issue. We document that guests should not access
the shared memory after device initialization until the IVPosition
register becomes non-negative.
For cold plug, the funny state is unlikely to be visible in
practice, because we normally receive the shared memory long before
the guest gets around to mess with the device.
For hot plug, the timing is tighter, but the relative slowness of
PCI device configuration has a good chance to hide the funny state.
In either case, guests complying with the documented procedure are
safe.
* Migration becomes racy.
If migration completes before the shared memory setup completes on
the source, shared memory contents is silently lost. Fortunately,
migration is rather unlikely to win this race.
If the shared memory's ramblock arrives at the destination before
shared memory setup completes, migration fails.
There is no known way for a management application to wait for
shared memory setup to complete.
All you can do is retry failed migration. You can improve your
chances by leaving more time between running the destination QEMU
and the migrate command.
To mitigate silent memory loss, you need to ensure the server
initializes shared memory exactly the same on source and
destination.
These issues are entirely undocumented so far.
I'd expect the server to be almost always fast enough to hide these
issues. But then rare catastrophic races are in a way the worst kind.
This is way more trouble than I'm willing to take from any device.
Kill the funny state by receiving shared memory synchronously in
realize(). If your hot plug hangs, go kill your ivshmem server.
For easier review, this commit only makes the receive synchronous, it
doesn't add the necessary error propagation. Without that, the funny
state persists. The next commit will do that, and kill it off for
real.
Signed-off-by: Markus Armbruster <armbru@redhat.com>
Reviewed-by: Marc-André Lureau <marcandre.lureau@redhat.com>
Message-Id: <1458066895-20632-26-git-send-email-armbru@redhat.com>
3a55fc0f24