Re: [PATCH v1 1/1] migration: Unregister yank if migration setup fails

[Peter Xu]

On Thu, Jun 24, 2021 at 06:14:39PM +0100, Dr. David Alan Gilbert wrote:
> * Leonardo Bras (leobras@redhat.com) wrote:
> > Currently, if a qemu instance is started with "-incoming defer" and
> > an incorect parameter is passed to "migrate_incoming", it will print the
> > expected error and reply with "duplicate yank instance" for any upcoming
> > "migrate_incoming" command.
> > 
> > This renders current qemu process unusable, and requires a new qemu
> > process to be started before accepting a migration.
> > 
> > This is caused by a yank_register_instance() that happens in
> > qemu_start_incoming_migration() but is never reverted if any error
> > happens.
> > 
> > Solves this by unregistering the instance if anything goes wrong
> > in the function, allowing a new "migrate_incoming" command to be
> > accepted.
> > 
> > Fixes: b5eea99ec2f ("migration: Add yank feature", 2021-01-13)
> > Buglink: https://bugzilla.redhat.com/show_bug.cgi?id=1974366
> > Signed-off-by: Leonardo Bras <leobras@redhat.com>
> > 
> > ---
> >  migration/migration.c | 6 +++++-
> >  1 file changed, 5 insertions(+), 1 deletion(-)
> > 
> > diff --git a/migration/migration.c b/migration/migration.c
> > index 4228635d18..ddcf9e1868 100644
> > --- a/migration/migration.c
> > +++ b/migration/migration.c
> > @@ -474,9 +474,13 @@ static void qemu_start_incoming_migration(const char 
> > *uri, Error **errp)
> >      } else if (strstart(uri, "fd:", &p)) {
> >          fd_start_incoming_migration(p, errp);
> >      } else {
> > -        yank_unregister_instance(MIGRATION_YANK_INSTANCE);
> >          error_setg(errp, "unknown migration protocol: %s", uri);
> >      }
> > +
> > +    if (*errp) {
> > +        yank_unregister_instance(MIGRATION_YANK_INSTANCE);
> > +    }
> 
> My understanding is that testing *errp isn't allowed, because
> it's legal to pass NULL to ignore errors, or legal to pass
> &error_abort to mean that any error you do hit will cause the
> process to assert; so you need to have something separate you can test.

Per my understanding error_abort should be fine, as the value of error_abort is
still NULL (in error_setg() we only check against &error_abort as the pointer,
and its value seems to be better always be NULL..).

But indeed at least we need "errp && *errp", but that won't capture the case
when errp==NULL.

So I think we may need to define a local error, check here when unregister
yank, and do error_propagate() before return..

-- 
Peter Xu