> Roland, I see this in kernel:
>
> ((struct mthca_next_seg *) prev_wqe)->nda_op =
> cpu_to_be32((ind << qp->rq.wqe_shift) | 1);
> wmb();
> ((struct mthca_next_seg *) prev_wqe)->ee_nds =
> cpu_to_be32(MTHCA_NEXT_DBD | size);
>
> but userspace does not have wmb here.
> Is it needed?
It does seem that way -- otherwise the hardware might read prev_wqe
and see the ee_nds field as set before the nda_op field has the right
variable. Does this look right to you as a libmthca fix?
diff --git a/src/qp.c b/src/qp.c
index 2d03d49..2ea9dc0 100644
--- a/src/qp.c
+++ b/src/qp.c
@@ -292,7 +292,10 @@ int mthca_tavor_post_send(struct ibv_qp *ibqp, struct
ibv_send_wr *wr,
htonl(((ind << qp->sq.wqe_shift) +
qp->send_wqe_offset) |
mthca_opcode[wr->opcode]);
-
+ /*
+ * Make sure that nda_op is written before setting ee_nds.
+ */
+ wmb();
((struct mthca_next_seg *) prev_wqe)->ee_nds =
htonl((size0 ? 0 : MTHCA_NEXT_DBD) | size |
((wr->send_flags & IBV_SEND_FENCE) ?
_______________________________________________
general mailing list
[email protected]
http://lists.openfabrics.org/cgi-bin/mailman/listinfo/general
To unsubscribe, please visit http://openib.org/mailman/listinfo/openib-general
