On Mon, Aug 26, 2024 at 01:22:13PM -0700, Jakub Kicinski wrote:

On Thu, 22 Aug 2024 15:41:19 -0500 Wei Huang wrote:

quoted

+		rtnl_lock();
+		bnxt_close_nic(irq->bp, false, false);
+		bnxt_open_nic(irq->bp, false, false);
+		rtnl_unlock();

This code means that under memory pressure changing CPU affinity
can take the machine offline. The entire machine, even if container
orchestration is trying to just move a few IRQs in place for a new
container.

We can't let you do this, it will set a bad precedent. I can't think
of any modern driver with reconfiguration safety as bad as bnxt.
Technical debt coming due.

Jakub,

We have not said this on the list, but we agree.  We plan to replace these
calls with calls to stop and start only that ring via netdev_rx_queue_restart
as soon as these calls all land in the same tree.  Since this set is
[presumably] coming through linux-pci we didn't think we could do that yet.

Thoughts?

-andy