On 5/25/21 17:45, Keith Busch wrote:

quoted

The bad controller should be fixed.

In the past, I've sent patches that check that sqid match in nvme cqe to
protect faulty drives that might send
the completion on a wrong msix.

My patch wasn't accepted since it added an additional "if" in the fast path.

Now we're adding much more operation in the fast path because of buggy ctrl
?

I shared the same performance concern on v1 on this series. I haven't
been able to test this one yet (only have emulation for two more weeks).

Hannes says the bug this catches happens frequently enough on TCP. If we
don't catch it, we get kernel panic or corruption, so we defintely need to
do something. Sagi correctly noted this type of bug is not unique to TCP
(or even NVMe, for that matter), but if there is a performance impact on
PCI, and no one so far reports such an issue, I would still recommend
this type of mitigation be isolated to transports that actually observe
invalid CQEs.
 

Please do that if possible.



_______________________________________________
Linux-nvme mailing list
Linux-nvme@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-nvme