Re: [Qemu-devel] [PATCH] migration: allow clearing migration string parameters

[Markus Armbruster <armbru@redhat.com>]

"Daniel P. Berrange" <berrange@redhat.com> writes:

> On Fri, Mar 03, 2017 at 03:44:12PM +0100, Markus Armbruster wrote:
>> "Daniel P. Berrange" <berrange@redhat.com> writes:
>> 
>> > On Wed, Mar 01, 2017 at 08:36:03AM -0600, Eric Blake wrote:
>> >> On 03/01/2017 06:32 AM, Daniel P. Berrange wrote:
>> >> >      }
>> >> >      if (params->has_tls_creds) {
>> >> >          g_free(s->parameters.tls_creds);
>> >> > -        s->parameters.tls_creds = g_strdup(params->tls_creds);
>> >> > +        if (*params->tls_creds == '\0') {
>> >> > +            s->parameters.tls_creds = NULL;
>> >> 
>> >> I'm wondering if you should also do s->parameters.has_tls_creds = false
>> >> at this point?  The visitors expect that if has_tls_creds is true, then
>> >> the string is non-NULL.
>> >
>> > The fact that s->parameters contains has_* fields is completely ignored
>> > by the migration code afaict. IOW the code behaves as if all the has_*
>> > fields are hardwired to true in s->parameters, even though that is not
>> > the case :-) The has_* fields are only used when the various migration
>> > QMP methods are executed, and those all use a separate MigrationParameters
>> > struct instance.
>> 
>> Not keeping the has_ members up-to-date is harmless as long as you don't
>> pass the thing to visitors, including the one hiding in qapi_free_FOO().
>> That one ignores scalars, though.
>> 
>> 
>> From a more abstract point of view, we have two related data types: one
>> for the state, and one for state changes requests.
>> 
>> In state, members are always present.
>> 
>> A state change request is a bag of state member change requests, and
>> each request can either specify the new value or ask for a reset to
>> default.  Absent member means no change.
>> 
>> We press the same QAPI type into service for both by making all members
>> optional.
>> 
>> For the state case, we hardwire the has_ to true.  Or even ignore them
>> completely.
>> 
>> For the state change request, we use has_ = false for "no change", has_
>> = true with a special value for "reset to default" (new in this patch)
>> and has_ = true with a non-special value for "set to this value".
>> 
>> Requires a special value outside the set of non-special values.  The
>> obvious one is JSON null, but the QAPI generator doesn't quite support
>> that, yet.  "" works here, but is not general.
>> 
>> I think I can get you null support in 2.10.  Would that work for you?
>
> Not really - we need this fix so libvirt can enable TLS migration,
> in addition we'll be wanting to add it to stable branches upstream
> and downstream.

Alternatives:

(a) We support null in 2.9, and use it to solve this problem.

(b) We use "", support null in 2.10, immediately deprecate "" in favor
    of null.

PRO (a): Less churn, less variation, less overall work.

PRO (b): Might be easier to backport to really old versions.  Soft
freeze.

Non-alternative: use "" forever.  Sorry, I hate it.