Re: [External] : Re: [PATCH 1/1] IB/mlx5: Add a signature check to received EQEs and CQEs

[Leon Romanovsky]

On Fri, Oct 28, 2022 at 04:48:53PM -0700, Rohit Nair wrote:
> On 10/27/22 5:23 AM, Leon Romanovsky wrote:
> > On Tue, Oct 25, 2022 at 10:44:12AM -0700, Rohit Nair wrote:
> > > Hey Leon,
> > > 
> > > Please find my replies to your comments here below:
> > 
> > <...>
> > 
> > > > 
> > > > > This patch does not introduce any significant performance degradations
> > > > > and has been tested using qperf.
> > > > What does it mean? You made changes in kernel verbs flow, they are not
> > > > executed through qperf.
> > > We also conducted several extensive performance tests using our test-suite
> > > which utilizes rds-stress and also saw no significant performance
> > > degrdations in those results.
> > 
> > What does it mean "also"? Your change is applicable ONLY for kernel path.
> > 
> > Anyway, I'm not keen adding rare debug code to performance critical path.
> > 
> > Thanks
> 
> rds-stress exercises the codepath we are modifying here. rds-stress didn't
> show much of performance degrade when we ran internally. We also requested
> our DB team for performance regression testing and this change passed their
> test suite. This motivated us to submit this to upstream.
> 
> If there is any other test that is better suited for this change, I am
> willing to test it. Please let me know if you have something in mind. We can
> revisit this patch after such a test may be.
> 
> I agree that, this was a rare debug scenario, but it took lot more than
> needed to narrow down[engaged vendor on live sessions]. We are adding this
> in the hope to finding the cause at the earliest or at least point us which
> direction to look at. We also requested the vendor[mlx] to include some
> diagnostics[HW counter], which can help us narrow it faster next time. This
> is our attempt to add kernel side of diagnostics.

The thing is that "vendor" failed to explain internally if this debug
code is useful. Like I said, extremely rare debug code shouldn't be part
of main data path.

Thanks

> 
> Feel free to share your suggestions
> 
> Thanks
>