Re: [Qemu-devel] [RFC 29/29] migration: reset migrate thread vars when resumed

[Peter Xu]

On Thu, Aug 03, 2017 at 02:54:35PM +0100, Dr. David Alan Gilbert wrote:
> * Peter Xu (address@hidden) wrote:
> > Firstly, MigThrError enumeration is introduced to describe the error in
> > migration_detect_error() better. This gives the migration_thread() a
> > chance to know whether a recovery has happened.
> > 
> > Then, if a recovery is detected, migration_thread() will reset its local
> > variables to prepare for that.
> > 
> > Signed-off-by: Peter Xu <address@hidden>
> > ---
> >  migration/migration.c | 40 +++++++++++++++++++++++++++++-----------
> >  1 file changed, 29 insertions(+), 11 deletions(-)
> > 
> > diff --git a/migration/migration.c b/migration/migration.c
> > index ecebe30..439bc22 100644
> > --- a/migration/migration.c
> > +++ b/migration/migration.c
> > @@ -2159,6 +2159,15 @@ static bool postcopy_should_start(MigrationState *s)
> >      return atomic_read(&s->start_postcopy) || s->start_postcopy_fast;
> >  }
> >  
> > +typedef enum MigThrError {
> > +    /* No error detected */
> > +    MIG_THR_ERR_NONE = 0,
> > +    /* Detected error, but resumed successfully */
> > +    MIG_THR_ERR_RECOVERED = 1,
> > +    /* Detected fatal error, need to exit */
> > +    MIG_THR_ERR_FATAL = 2,
> > +} MigThrError;
> > +
> 
> Could you move this patch earlier to when postcopy_pause is created
> so it's created with this enum?

Sure.

[...]

> > @@ -2319,6 +2327,7 @@ static void *migration_thread(void *opaque)
> >      /* The active state we expect to be in; ACTIVE or POSTCOPY_ACTIVE */
> >      enum MigrationStatus current_active_state = MIGRATION_STATUS_ACTIVE;
> >      bool enable_colo = migrate_colo_enabled();
> > +    MigThrError thr_error;
> >  
> >      rcu_register_thread();
> >  
> > @@ -2395,8 +2404,17 @@ static void *migration_thread(void *opaque)
> >           * Try to detect any kind of failures, and see whether we
> >           * should stop the migration now.
> >           */
> > -        if (migration_detect_error(s)) {
> > +        thr_error = migration_detect_error(s);
> > +        if (thr_error == MIG_THR_ERR_FATAL) {
> > +            /* Stop migration */
> >              break;
> > +        } else if (thr_error == MIG_THR_ERR_RECOVERED) {
> > +            /*
> > +             * Just recovered from a e.g. network failure, reset all
> > +             * the local variables.
> > +             */
> > +            initial_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
> > +            initial_bytes = 0;
> 
> They don't seem that important to reset?

The problem is that we have this in migration_thread():

        if (current_time >= initial_time + BUFFER_DELAY) {
            uint64_t transferred_bytes = qemu_ftell(s->to_dst_file) -
                                         initial_bytes;
            uint64_t time_spent = current_time - initial_time;
            double bandwidth = (double)transferred_bytes / time_spent;
            threshold_size = bandwidth * s->parameters.downtime_limit;
            ...
        }

Here qemu_ftell() would possibly be very small since we have just
resumed... and then transferred_bytes will be extremely huge since
"qemu_ftell(s->to_dst_file) - initial_bytes" is actually negative...
Then, with luck, we'll got extremely huge "bandwidth" as well.

-- 
Peter Xu