Re: [Qemu-devel] [PATCH 2/3] migration: create migration event

[Juan Quintela]

Jiri Denemark <address@hidden> wrote:
> On Wed, May 20, 2015 at 17:35:23 +0200, Juan Quintela wrote:
>> We have one argument that tells us what event has happened.
>> 
>> Signed-off-by: Juan Quintela <address@hidden>
>> ---
>>  docs/qmp/qmp-events.txt | 16 ++++++++++++++++
>>  migration/migration.c   | 12 ++++++++++++
>>  qapi/event.json         | 14 ++++++++++++++
>>  3 files changed, 42 insertions(+)
>
> Hi Juan,
>
> I patched libvirt to be able to consume this event and it all seems to
> work fine except for events sent after migrate_cancel command:
>
> {"execute":"migrate_cancel","id":"libvirt-33"}
> {"timestamp": {"seconds": 1432899178, "microseconds": 844907}, "event":
> "MIGRATION", "data": {"status": "cancelling"}}
> {"return": {}, "id": "libvirt-33"}
> {"timestamp": {"seconds": 1432899178, "microseconds": 845625}, "event":
> "MIGRATION", "data": {"status": "failed"}}
> {"timestamp": {"seconds": 1432899178, "microseconds": 846432}, "event":
> "MIGRATION", "data": {"status": "cancelled"}}
>
> In other words, the status first changes to "failed" and then it changes
> correctly "cancelled". Can you fix this weird behavior in QEMU or do we
> have to work around it in libvirt?
>
> Jirka

Found root cause of this, basically we had the code:

migrate_set_state(FOO, MIGRATION_STATUS_FAILED);
qapi_send_event(...., MIGRATION_STATUS_FAILED);

But the first call only change the state if the current state is FOO.
So, changed the event to inside migrate_set_state() and only sent it
when we change the state.

See new patch series, basically we were using incorrectly
atomic_cmpxchg() and that made things ... interesting.

Later, Juan.