[ofa-general] [PATCH 1 of 2] libmlx4: Handle new FW requirement for send request prefetching, for WQE sg lists

[Jack Morgenstein]

This is an addendum to Roland's commit 561da8d10e419ffb333fe6faf05004d9a3670e7a
(June 13). This addendum adds prefetch headroom marking processing for s/g segments.

We write s/g segments in reverse order into the WQE, in order to guarantee
that the first dword of all cachelines containing s/g segments is written last
(overwriting the headroom invalidation pattern). The entire cacheline will thus
contain valid data when the invalidation pattern is overwritten.

Signed-off-by: Jack Morgenstein <jackm at dev.mellanox.co.il>

Index: libmlx4/src/qp.c
===================================================================
--- libmlx4.orig/src/qp.c	2007-09-04 10:03:38.264742000 +0300
+++ libmlx4/src/qp.c	2007-09-04 10:04:35.536784000 +0300
@@ -312,10 +312,19 @@ int mlx4_post_send(struct ibv_qp *ibqp, 
 		} else {
 			struct mlx4_wqe_data_seg *seg = wqe;
 
-			for (i = 0; i < wr->num_sge; ++i) {
-				seg[i].byte_count = htonl(wr->sg_list[i].length);
+			/*
+			 * Write the s/g entries in reverse order, so that the
+			 * first dword of all cachelines is written last.
+			 */
+			for (i = wr->num_sge - 1; i >= 0; --i) {
 				seg[i].lkey       = htonl(wr->sg_list[i].lkey);
 				seg[i].addr       = htonll(wr->sg_list[i].addr);
+				/*
+				 * This entry may start a new cacheline.
+				 * See barrier comment above.
+				 */
+				wmb();
+				seg[i].byte_count = htonl(wr->sg_list[i].length);
 			}
 
 			size += wr->num_sge * (sizeof *seg / 16);