Re: [Qemu-devel] [PATCH V2] migration: expose qemu_announce_self() via qmp

[Dr. David Alan Gilbert]

* Vlad Yasevich (address@hidden) wrote:
> On 02/20/2017 07:16 PM, Germano Veit Michel wrote:
> > qemu_announce_self() is triggered by qemu at the end of migrations
> > to update the network regarding the path to the guest l2addr.
> > 
> > however it is also useful when there is a network change such as
> > an active bond slave swap. Essentially, it's the same as a migration
> > from a network perspective - the guest moves to a different point
> > in the network topology.
> > 
> > this exposes the function via qmp.
> > 
> > Signed-off-by: Germano Veit Michel <address@hidden>
> > ---
> >  include/migration/vmstate.h |  5 +++++
> >  migration/savevm.c          | 30 +++++++++++++++++++-----------
> >  qapi-schema.json            | 18 ++++++++++++++++++
> >  3 files changed, 42 insertions(+), 11 deletions(-)
> > 
> > diff --git a/include/migration/vmstate.h b/include/migration/vmstate.h
> > index 63e7b02..a08715c 100644
> > --- a/include/migration/vmstate.h
> > +++ b/include/migration/vmstate.h
> > @@ -1042,6 +1042,11 @@ int64_t self_announce_delay(int round)
> >      return 50 + (SELF_ANNOUNCE_ROUNDS - round - 1) * 100;
> >  }
> > 
> > +struct AnnounceRound {
> > +    QEMUTimer *timer;
> > +    int count;
> > +};
> > +
> >  void dump_vmstate_json_to_file(FILE *out_fp);
> > 
> >  #endif
> > diff --git a/migration/savevm.c b/migration/savevm.c
> > index 5ecd264..44e196b 100644
> > --- a/migration/savevm.c
> > +++ b/migration/savevm.c
> > @@ -118,29 +118,37 @@ static void qemu_announce_self_iter(NICState
> > *nic, void *opaque)
> >      qemu_send_packet_raw(qemu_get_queue(nic), buf, len);
> >  }
> > 
> > -
> >  static void qemu_announce_self_once(void *opaque)
> >  {
> > -    static int count = SELF_ANNOUNCE_ROUNDS;
> > -    QEMUTimer *timer = *(QEMUTimer **)opaque;
> > +    struct AnnounceRound *round = opaque;
> > 
> >      qemu_foreach_nic(qemu_announce_self_iter, NULL);
> > 
> > -    if (--count) {
> > +    round->count--;
> > +    if (round->count) {
> >          /* delay 50ms, 150ms, 250ms, ... */
> > -        timer_mod(timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) +
> > -                  self_announce_delay(count));
> > +        timer_mod(round->timer, qemu_clock_get_ms(QEMU_CLOCK_REALTIME) +
> > +                  self_announce_delay(round->count));
> >      } else {
> > -            timer_del(timer);
> > -            timer_free(timer);
> > +            timer_del(round->timer);
> > +            timer_free(round->timer);
> > +            g_free(round);
> >      }
> >  }
> > 
> >  void qemu_announce_self(void)
> >  {
> > -    static QEMUTimer *timer;
> > -    timer = timer_new_ms(QEMU_CLOCK_REALTIME, qemu_announce_self_once, 
> > &timer);
> > -    qemu_announce_self_once(&timer);
> > +    struct AnnounceRound *round = g_malloc(sizeof(struct AnnounceRound));
> > +    if (!round)
> > +        return;
> > +    round->count = SELF_ANNOUNCE_ROUNDS;
> > +    round->timer = timer_new_ms(QEMU_CLOCK_REALTIME,
> > qemu_announce_self_once, round);
> > +    qemu_announce_self_once(round);
> > +}
> 
> So, I've been looking and this code and have been playing with it and with 
> David's
> patches and my patches to include virtio self announcements as well.  What 
> I've discovered
> is what I think is a possible packet amplification issue here.
> 
> This creates a new timer every time we do do a announce_self.  With just 
> migration,
> this is not an issue since you only migrate once at a time, so there is only 
> 1 timer.
> With exposing this as an API, a user can potentially call it in a tight loop
> and now you have a ton of timers being created.  Add in David's patches 
> allowing timeouts
> and retries to be configurable, and you may now have a ton of long lived 
> timers.
> Add in the patches I am working on to let virtio do self announcements too 
> (to really fix
> bonding issues), and now you add in a possibility of a lot of packets being 
> sent for
> each timeout (RARP, GARP, NA, IGMPv4 Reports, IGMPv6 Reports [even worse if 
> MLD1 is used]).
> 
> As you can see, this can get rather ugly...
> 
> I think we need timer user here.  Migration and QMP being two to begin with.  
> Each
> one would get a single timer to play with.  If a given user already has a 
> timer running,
> we could return an error or just not do anything.

If you did have specific timers, then you could add to/reset the counts
rather than doing nothing.  That way it's less racy; if you issue the
command just as you reconfigure your network, there's no chance the
command would fail, you will send the packets out.

Dave

> -vlad
> 
> > +
> > +void qmp_announce_self(Error **errp)
> > +{
> > +    qemu_announce_self();
> >  }
> > 
> >  /***********************************************************/
> > diff --git a/qapi-schema.json b/qapi-schema.json
> > index baa0d26..0d9bffd 100644
> > --- a/qapi-schema.json
> > +++ b/qapi-schema.json
> > @@ -6080,3 +6080,21 @@
> >  #
> >  ##
> >  { 'command': 'query-hotpluggable-cpus', 'returns': ['HotpluggableCPU'] }
> > +
> > +##
> > +# @announce-self:
> > +#
> > +# Trigger generation of broadcast RARP frames to update network switches.
> > +# This can be useful when network bonds fail-over the active slave.
> > +#
> > +# Arguments: None.
> > +#
> > +# Example:
> > +#
> > +# -> { "execute": "announce-self" }
> > +# <- { "return": {} }
> > +#
> > +# Since: 2.9
> > +##
> > +{ 'command': 'announce-self' }
> > +
> > 
> 
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK