From: Roland Dreier <rolandd@cisco.com>
Date: Mon, 18 Jun 2007 16:23:47 +0000 (-0700)
Subject: IB/mlx4: Make sure inline data segments don't cross a 64 byte boundary
X-Git-Tag: v2.6.22-rc6~70^2
X-Git-Url: https://err.no/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=e61ef2416b0b92828512b6cfcd0104a02b6431fe;p=linux-2.6

IB/mlx4: Make sure inline data segments don't cross a 64 byte boundary

Inline data segments in send WQEs are not allowed to cross a 64 byte
boundary.  We use inline data segments to hold the UD headers for MLX
QPs (QP0 and QP1).  A send with GRH on QP1 will have a UD header that
is too big to fit in a single inline data segment without crossing a
64 byte boundary, so split the header into two inline data segments.

Signed-off-by: Roland Dreier <rolandd@cisco.com>
---

diff --git a/drivers/infiniband/hw/mlx4/qp.c b/drivers/infiniband/hw/mlx4/qp.c
index 355a31f9c0..28a08bdd18 100644
--- a/drivers/infiniband/hw/mlx4/qp.c
+++ b/drivers/infiniband/hw/mlx4/qp.c
@@ -192,6 +192,8 @@ static int send_wqe_overhead(enum ib_qp_type type)
 	case IB_QPT_GSI:
 		return sizeof (struct mlx4_wqe_ctrl_seg) +
 			ALIGN(MLX4_IB_UD_HEADER_SIZE +
+			      DIV_ROUND_UP(MLX4_IB_UD_HEADER_SIZE,
+					   MLX4_INLINE_ALIGN) *
 			      sizeof (struct mlx4_wqe_inline_seg),
 			      sizeof (struct mlx4_wqe_data_seg)) +
 			ALIGN(4 +
@@ -1049,6 +1051,7 @@ static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
 	u16 pkey;
 	int send_size;
 	int header_size;
+	int spc;
 	int i;
 
 	send_size = 0;
@@ -1124,10 +1127,43 @@ static int build_mlx_header(struct mlx4_ib_sqp *sqp, struct ib_send_wr *wr,
 		printk("\n");
 	}
 
-	inl->byte_count = cpu_to_be32(1 << 31 | header_size);
-	memcpy(inl + 1, sqp->header_buf, header_size);
+	/*
+	 * Inline data segments may not cross a 64 byte boundary.  If
+	 * our UD header is bigger than the space available up to the
+	 * next 64 byte boundary in the WQE, use two inline data
+	 * segments to hold the UD header.
+	 */
+	spc = MLX4_INLINE_ALIGN -
+		((unsigned long) (inl + 1) & (MLX4_INLINE_ALIGN - 1));
+	if (header_size <= spc) {
+		inl->byte_count = cpu_to_be32(1 << 31 | header_size);
+		memcpy(inl + 1, sqp->header_buf, header_size);
+		i = 1;
+	} else {
+		inl->byte_count = cpu_to_be32(1 << 31 | spc);
+		memcpy(inl + 1, sqp->header_buf, spc);
+
+		inl = (void *) (inl + 1) + spc;
+		memcpy(inl + 1, sqp->header_buf + spc, header_size - spc);
+		/*
+		 * Need a barrier here to make sure all the data is
+		 * visible before the byte_count field is set.
+		 * Otherwise the HCA prefetcher could grab the 64-byte
+		 * chunk with this inline segment and get a valid (!=
+		 * 0xffffffff) byte count but stale data, and end up
+		 * generating a packet with bad headers.
+		 *
+		 * The first inline segment's byte_count field doesn't
+		 * need a barrier, because it comes after a
+		 * control/MLX segment and therefore is at an offset
+		 * of 16 mod 64.
+		 */
+		wmb();
+		inl->byte_count = cpu_to_be32(1 << 31 | (header_size - spc));
+		i = 2;
+	}
 
-	return ALIGN(sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
+	return ALIGN(i * sizeof (struct mlx4_wqe_inline_seg) + header_size, 16);
 }
 
 static int mlx4_wq_overflow(struct mlx4_ib_wq *wq, int nreq, struct ib_cq *ib_cq)
diff --git a/include/linux/mlx4/qp.h b/include/linux/mlx4/qp.h
index 9eeb61adf6..10c57d2791 100644
--- a/include/linux/mlx4/qp.h
+++ b/include/linux/mlx4/qp.h
@@ -269,6 +269,10 @@ struct mlx4_wqe_data_seg {
 	__be64			addr;
 };
 
+enum {
+	MLX4_INLINE_ALIGN	= 64,
+};
+
 struct mlx4_wqe_inline_seg {
 	__be32			byte_count;
 };