Section 4.11.7.1 of rev 1.0 of the xhci specification states that a link TRB
can only occur at a boundary between underlying USB frames (512 bytes for 480M).

If this isn't done the USB frames aren't formatted correctly and, for example,
the USB3 ethernet ax88179_178a card will stop sending (while still receiving)
when running a netperf tcp transmit test with (say) and 8k buffer.

This should be a candidate for stable, the ax88179_178a driver defaults to
gso and tso enabled so it passes a lot of fragmented skb to the USB stack.

Signed-off-by: David Laight <david.laight@xxxxxxxxxx>
---

Changes for v2:

1) Only act on bulk endpoints.
   While isoc endpoints could suffer from the same problem it is much less
   likely and can't be fixed by adding NOP TRBs (they would stop data being
   sent in the poll interval).

2) When writing the NOP TRB use the count of TRBs instead of scanning for
   the link TRB.

 drivers/usb/host/xhci-ring.c | 53 ++++++++++++++++++++++++++++++++++++++++++--
 1 file changed, 51 insertions(+), 2 deletions(-)

diff --git a/drivers/usb/host/xhci-ring.c b/drivers/usb/host/xhci-ring.c
index 5480215..c1342dc 100644
--- a/drivers/usb/host/xhci-ring.c
+++ b/drivers/usb/host/xhci-ring.c
@@ -2929,8 +2929,57 @@
 	}
 
 	while (1) {
-		if (room_on_ring(xhci, ep_ring, num_trbs))
-			break;
+		if (room_on_ring(xhci, ep_ring, num_trbs)) {
+			union xhci_trb *trb = ep_ring->enqueue;
+			unsigned int usable = ep_ring->enq_seg->trbs +
+					TRBS_PER_SEGMENT - 1 - trb;
+			u32 nop_cmd;
+
+			/*
+			 * Section 4.11.7.1 TD Fragments states that a link
+			 * TRB must only occur at the boundary between
+			 * data bursts (eg 512 bytes for 480M).
+			 * While it is possible to split a large fragment
+			 * we don't know the size yet.
+			 * Simplest solution is to fill the trb before the
+			 * LINK with nop commands.
+			 */
+			if (num_trbs == 1 || num_trbs <= usable || usable == 0)
+				break;
+
+			if (ep_ring->type != TYPE_BULK)
+				/*
+				 * While isoc transfers might have a buffer that
+				 * crosses a 64k boundary it is unlikely.
+				 * Since we can't add NOPs without generating
+				 * gaps in the traffic just hope it never
+				 * happens at the end of the ring.
+				 * This could be fixed by writing a LINK TRB
+				 * instead of the first NOP - however the
+				 * TRB_TYPE_LINK_LE32() calls would all need
+				 * changing to check the ring length. */
+				break;
+
+			if (num_trbs >= TRBS_PER_SEGMENT) {
+				xhci_err(xhci, "Too many fragments %d, max %d\n",
+						num_trbs, TRBS_PER_SEGMENT - 1);
+				return -ENOMEM;
+			}
+
+			nop_cmd = cpu_to_le32(TRB_TYPE(TRB_TR_NOOP) |
+					ep_ring->cycle_state);
+			ep_ring->num_trbs_free -= usable;
+			do {
+				trb->generic.field[0] = 0;
+				trb->generic.field[1] = 0;
+				trb->generic.field[2] = 0;
+				trb->generic.field[3] = nop_cmd;
+				trb++;
+			} while (--usable);
+			ep_ring->enqueue = trb;
+			if (room_on_ring(xhci, ep_ring, num_trbs))
+				break;
+		}
 
 		if (ep_ring == xhci->cmd_ring) {
 			xhci_err(xhci, "Do not support expand command ring\n");