Re: Properly quitting qemu immediately after failing migration

[Dr. David Alan Gilbert]

* Max Reitz (mreitz@redhat.com) wrote:
> Hi,
> 
> In an iotest, I’m trying to quit qemu immediately after a migration has
> failed.  Unfortunately, that doesn’t seem to be possible in a clean way:
> migrate_fd_cleanup() runs only at some point after the migration state
> is already “failed”, so if I just wait for that “failed” state and
> immediately quit, some cleanup functions may not have been run yet.

Yeh this is hard; I always take the end of migrate_fd_cleanup to be the
real end.
It always happens on the main thread I think (it's done as a bh in some
cases).

> This is a problem with dirty bitmap migration at least, because it
> increases the refcount on all block devices that are to be migrated, so
> if we don’t call the cleanup function before quitting, the refcount will
> stay elevated and bdrv_close_all() will hit an assertion because those
> block devices are still around after blk_remove_all_bs() and
> blockdev_close_all_bdrv_states().
> 
> In practice this particular issue might not be that big of a problem,
> because it just means qemu aborts when the user intended to let it quit
> anyway.  But on one hand I could imagine that there are other clean-up
> paths that should definitely run before qemu quits (although I don’t
> know), and on the other, it’s a problem for my test.

'quit' varies - there are a lot of incoming failures that just assert;
very few of them cause a clean exit (I think there are more clean ones
after Peter's work on restartable postcopy a year or two ago).

I do see the end of migrate_fd_cleanup calls the notifier list; but it's
not clear to me that it's alwyas going to see the first transition to
'failed' at that point.

> I tried working around the problem for my test by waiting on “Unable to
> write” appearing on stderr, because that indicates that
> migrate_fd_cleanup()’s error_report_err() has been reached.  But on one
> hand, that isn’t really nice, and on the other, it doesn’t even work
> when the failure is on the source side (because then there is no
> s->error for migrate_fd_cleanup() to report).
> 
> In all, I’m asking:
> (1) Is there a nice solution for me now to delay quitting qemu until the
> failed migration has been fully resolved, including the clean-up?

In vl.c, I added a call to migration_shutdown in qemu_cleanup - although
that seems to be mostly about cleaning up the *outgoing* side; you could
add some incoming cleanup there.

> (2) Isn’t it a problem if qemu crashes when you issue “quit” via QMP at
> the wrong time?  Like, maybe lingering subprocesses when using “exec”?

Yeh that should be cleaner, but isn't.

Dave

> 
> Thanks,
> 
> Max
> 



--
Dr. David Alan Gilbert / dgilbert@redhat.com / Manchester, UK