Re: [Qemu-block] [PATCH v10 05/12] migration: introduce postcopy-only pending

[Vladimir Sementsov-Ogievskiy]

12.03.2018 18:30, Dr. David Alan Gilbert wrote:
> * Vladimir Sementsov-Ogievskiy (address@hidden) wrote:
> > There would be savevm states (dirty-bitmap) which can migrate only in
> > postcopy stage. The corresponding pending is introduced here.
> >
> > Signed-off-by: Vladimir Sementsov-Ogievskiy <address@hidden>
> > ---

[...]

> >   static MigIterateState migration_iteration_run(MigrationState *s)
> >   {
> > -    uint64_t pending_size, pend_post, pend_nonpost;
> > +    uint64_t pending_size, pend_pre, pend_compat, pend_post;
> >       bool in_postcopy = s->state == MIGRATION_STATUS_POSTCOPY_ACTIVE;
> >   
> > -    qemu_savevm_state_pending(s->to_dst_file, s->threshold_size,
> > -                              &pend_nonpost, &pend_post);
> > -    pending_size = pend_nonpost + pend_post;
> > +    qemu_savevm_state_pending(s->to_dst_file, s->threshold_size, &pend_pre,
> > +                              &pend_compat, &pend_post);
> > +    pending_size = pend_pre + pend_compat + pend_post;
> >   
> >       trace_migrate_pending(pending_size, s->threshold_size,
> > -                          pend_post, pend_nonpost);
> > +                          pend_pre, pend_compat, pend_post);
> >   
> >       if (pending_size && pending_size >= s->threshold_size) {
> >           /* Still a significant amount to transfer */
> >           if (migrate_postcopy() && !in_postcopy &&
> > -            pend_nonpost <= s->threshold_size &&
> > -            atomic_read(&s->start_postcopy)) {
> > +            pend_pre <= s->threshold_size &&
> > +            (atomic_read(&s->start_postcopy) ||
> > +             (pend_pre + pend_compat <= s->threshold_size)))
> This change does something different from the description;
> it causes a postcopy_start even if the user never ran the postcopy-start
> command; so sorry, we can't do that; because postcopy for RAM is
> something that users can enable but only switch into when they've given
> up on it completing normally.
>
> However, I guess that leaves you with a problem; which is what happens
> to the system when you've run out of pend_pre+pend_compat but can't
> complete because pend_post is non-0; so I don't know the answer to that.

Hmm. Here, we go to postcopy only if "pend_pre + pend_compat <= 
s->threshold_size". Pre-patch, in this case we will go to 
migration_completion(). So, precopy stage is finishing anyway. So, we 
want in this case to finish ram migration like it was finished by 
migration_completion(), and then, run postcopy, which will handle only 
dirty bitmaps, yes?

Hmm2. Looked through migration_completion(), I don't understand, how it 
finishes ram migration without postcopy. It calls 
qemu_savevm_state_complete_precopy(), which skips states with 
has_postcopy=true, which is ram...

--
Best regards,
Vladimir