net: macb: fix dropped RX frames due to a race
authorAnssi Hannula <anssi.hannula@bitwise.fi>
Mon, 17 Dec 2018 13:05:40 +0000 (15:05 +0200)
committerGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Sun, 13 Jan 2019 09:01:00 +0000 (10:01 +0100)
[ Upstream commit 8159ecab0db9095902d4c73605fb8787f5c7d653 ]

Bit RX_USED set to 0 in the address field allows the controller to write
data to the receive buffer descriptor.

The driver does not ensure the ctrl field is ready (cleared) when the
controller sees the RX_USED=0 written by the driver. The ctrl field might
only be cleared after the controller has already updated it according to
a newly received frame, causing the frame to be discarded in gem_rx() due
to unexpected ctrl field contents.

A message is logged when the above scenario occurs:

  macb ff0b0000.ethernet eth0: not whole frame pointed by descriptor

Fix the issue by ensuring that when the controller sees RX_USED=0 the
ctrl field is already cleared.

This issue was observed on a ZynqMP based system.

Fixes: 4df95131ea80 ("net/macb: change RX path for GEM")
Signed-off-by: Anssi Hannula <anssi.hannula@bitwise.fi>
Tested-by: Claudiu Beznea <claudiu.beznea@microchip.com>
Cc: Nicolas Ferre <nicolas.ferre@microchip.com>
Signed-off-by: David S. Miller <davem@davemloft.net>
Signed-off-by: Sasha Levin <sashal@kernel.org>
drivers/net/ethernet/cadence/macb_main.c

index 3aabc36eec0bfd77259bfe771509b9e7136378e6..9046993947cc1e2d4ec5461b43c6f85910397cd7 100644 (file)
@@ -923,14 +923,19 @@ static void gem_rx_refill(struct macb *bp)
 
                        if (entry == bp->rx_ring_size - 1)
                                paddr |= MACB_BIT(RX_WRAP);
-                       macb_set_addr(bp, desc, paddr);
                        desc->ctrl = 0;
+                       /* Setting addr clears RX_USED and allows reception,
+                        * make sure ctrl is cleared first to avoid a race.
+                        */
+                       dma_wmb();
+                       macb_set_addr(bp, desc, paddr);
 
                        /* properly align Ethernet header */
                        skb_reserve(skb, NET_IP_ALIGN);
                } else {
-                       desc->addr &= ~MACB_BIT(RX_USED);
                        desc->ctrl = 0;
+                       dma_wmb();
+                       desc->addr &= ~MACB_BIT(RX_USED);
                }
        }