[PATCH for-next 0/6] Perf and debug fixes for hfi

From: Dennis Dalessandro <dennis.dalessandro@cornelisnetworks.com>
To: jgg@ziepe.ca, dledford@redhat.com
Cc: linux-rdma@vger.kernel.org
Subject: [PATCH for-next 0/6] Perf and debug fixes for hfi
Date: Mon, 13 Sep 2021 09:28:20 -0400	[thread overview]
Message-ID: <20210913132317.131370.54825.stgit@awfm-01.cornelisnetworks.com> (raw)

Here is a series of perf improvements and debug/trace fixes from Mike,
who has this to say about the patches...

The AIP SDMA interrupt handling is inefficient:

- A slab entry is allocated for each sent packet

  This is despite the fact that there is a ring for each possible send slot
  that could be occupied by a tx descriptor

- The interrupt handling/NAPI is lock happy has a mixed up notion of
  producer and consumer

  The ring should be a ring of tx descriptors vs. a ring of pointers

  The consumer of descriptors should be the xmit side of the TX

  The producer of the descriptors is the SDMA interrupt handling and NAPI
  tx completion

  There is certainly no locking required in the interrupt/TX napi tx queue

  There is no locking required in the xmit side since that is held off by NAPI
  code

Note that these patches are also staged publicly on our GitHub site for easy
browsing in context.

https://github.com/cornelisnetworks/linux

---

Mike Marciniszyn (6):
      IB/hfi1: Remove cache and embed txreq in ring
      IB/hfi1: Get rid of hot path divide
      IB/hfi1: Get rid of tx priv backpointer
      IB/hfi1: Tune netdev xmit cachelines
      IB/hfi1: Remove atomic completion count
      IB/hfi1: Add ring consumer and producers traces

 drivers/infiniband/hw/hfi1/ipoib.h    |   76 +++++---
 drivers/infiniband/hw/hfi1/ipoib_tx.c |  314 ++++++++++++++-------------------
 drivers/infiniband/hw/hfi1/trace_tx.h |   71 +++++++
 3 files changed, 246 insertions(+), 215 deletions(-)

--
-Denny