On Tue, Dec 08, 2020 at 01:57:09PM +0100, Sebastian Andrzej Siewior wrote:

Yes, you are right. Even cross-CPU completion for single-queue was
already completing in softirq. So the only change is for multiqueue
devices which you just demonstrated that it does not happen.

It can happen if there are less hardware queues than CPUs. But I'd bet
that application which care about performance are well aware of this and
try to keep the work vertical aligned.