On Wed, Jun 10, 2026 at 09:16:11PM +0530, Sumit Saxena wrote:

The motivation for this change stems from performance issue we
encountered due to false sharing of the 'nr_active_requests_shared_tags'
counter
on certain CPU architectures. I initially submitted a patch to move that
counter to
its own cache line to avoid conflicts with 'nr_requests' and other hot
fields
(see:
https://patchwork.kernel.org/project/linux-scsi/patch/20260402074637.92417-3-sumit.saxena@broadcom.com/
).

During the review, Bart shared his work, which eliminates the
counter entirely by removing the fairness throttling. My testing confirmed
that
this approach resolved the performance issues and improved IOPS.
This patch is part of a larger set, and I have reported the cumulative
performance
improvements in the cover letter.

So the problem is just the atomic operation accounting overhead? I
previously thought the device just really needed to consume all the tags
to hit performance.