IB/srp: Fix completion vector assignment algorithm
authorBart Van Assche <bart.vanassche@wdc.com>
Mon, 12 Feb 2018 17:50:25 +0000 (09:50 -0800)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Tue, 24 Apr 2018 07:36:32 +0000 (09:36 +0200)
commit 3a148896b24adf8688dc0c59af54531931677a40 upstream.

Ensure that cv_end is equal to ibdev->num_comp_vectors for the
NUMA node with the highest index. This patch improves spreading
of RDMA channels over completion vectors and thereby improves
performance, especially on systems with only a single NUMA node.
This patch drops support for the comp_vector login parameter by
ignoring the value of that parameter since I have not found a
good way to combine support for that parameter and automatic
spreading of RDMA channels over completion vectors.

Fixes: d92c0da71a35 ("IB/srp: Add multichannel support")
Reported-by: Alexander Schmid <alex@modula-shop-systems.de>
Signed-off-by: Bart Van Assche <bart.vanassche@wdc.com>
Cc: Alexander Schmid <alex@modula-shop-systems.de>
Cc: stable@vger.kernel.org
Signed-off-by: Jason Gunthorpe <jgg@mellanox.com>
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
drivers/infiniband/ulp/srp/ib_srp.c

index 02a3169b2fd640d64d1fceea19cfc4a50f04c2a3..299a97b7e17fff2982e8762e4620afb904683870 100644 (file)
@@ -3430,12 +3430,10 @@ static ssize_t srp_create_target(struct device *dev,
                                      num_online_nodes());
                const int ch_end = ((node_idx + 1) * target->ch_count /
                                    num_online_nodes());
-               const int cv_start = (node_idx * ibdev->num_comp_vectors /
-                                     num_online_nodes() + target->comp_vector)
-                                    % ibdev->num_comp_vectors;
-               const int cv_end = ((node_idx + 1) * ibdev->num_comp_vectors /
-                                   num_online_nodes() + target->comp_vector)
-                                  % ibdev->num_comp_vectors;
+               const int cv_start = node_idx * ibdev->num_comp_vectors /
+                                    num_online_nodes();
+               const int cv_end = (node_idx + 1) * ibdev->num_comp_vectors /
+                                  num_online_nodes();
                int cpu_idx = 0;
 
                for_each_online_cpu(cpu) {