Re: cpr-transfer with caveats

[Peter Xu <peterx@redhat.com>]

On Fri, Oct 25, 2024 at 11:01:27AM -0400, Steven Sistare wrote:
> Hi Peter, are you OK if we proceed with cpr-transfer as is, without the
> precreate phase?  Here are the problems that motivated it:
> 
> * migration test wants to enable migration events on the dest.
>   fix: enable on dest qemu using -global.  only for the test.

If it's to be documented that cpr-transfer allows no migration parameter /
cap setup, then the test case should follow, IMHO, rather than enabling
anything from cmdline.

I hope that's trivial if we can have MigrateCommon.disable_events, for
example, then only cpr selects it.

> 
> * migration test needs to fetch the dynamically assigned migration
>     listen port number
>   Fix: require unix domain socket for cpr-transfer, or a fixed port
>   number. Document it.

Fixed port is probably not a good idea.. requires unix sockets looks fine.

But then again the whole point previously having -cpr-uri together with
-incoming (IIRC, from your previous email) is to have that flexibility to
use non-unix too.  Now I am not sure whether it is still needed to be
separate just that you'll still need to rely on SIGHUP just because
precreate is gone.

> 
> * migration test hangs connecting to the qtest socket.
>   fix: in the qtest code, defer connection.

Not sure how much change here, hopefully still manageable.

From use case POV shouldn't be a huge issue, if we don't use -qtest in
production anyway.

> 
> Document that one cannot set migration caps or params on the dest
> for cpr-transfer.

This also needs to be reviewed by Dan, on impact of Libvirt initiating the
dest QEMU instance with QMP all dead before a "migrate" on src.  I'm not
sure whether Libvirt will also do early setup like what our qtest does to
enable events and so on, assuming QMP available since the start.

> 
> Document that for -incoming defer, mgmt must send the migrate command
> to the src first (so dest reads cpr state and progresses to start the
> monitor), then send the hotplug monitor commands to the dest.
> 
> Daniel, are you OK with that last bit?
> 
> - Steve
> 

-- 
Peter Xu