On Wed, 2017-12-13 at 13:32 +0100, Nicolas Morey-Chaisemartin wrote:
> SM traps are polled through poll_cq which waited for a CQ event
> before polling the CQ itself.
> However it may happens that multiple completions are attached
> to a single event. As stated by the ibv_get_cq_event man page,
> it is required to poll the the CQ to get those completions
> after the call to ibv_req_notify_cq.
>
> As completions need to be handled one by one in an outer function,
> start by polling the CQ and return the completion (if any) before
> waiting for the next completion event.
> This will allow emptying all pending completions, through multiple calls
> to poll_cq, before waiting for a new event.
>
> The buggy use case seems to appear when the master SM is switched multiple
> times between two nodes. As the number of ping-pong between the SMs increases,
> the number of traps sent to notify that the SM just became master increases
> too. This causes burst of completions linked to a single event.
> Note that the race condition is also possible in other scenario.
Reviewed-by: Bart Van Assche <bart.vanassche@xxxxxxx>
��.n��������+%������w��{.n�����{���fk��ܨ}���Ơz�j:+v�����w����ޙ��&�)ߡ�a����z�ޗ���ݢj��w�f
